Line data Source code
1 : pub mod chaos_injector;
2 : pub mod feature_flag;
3 : pub(crate) mod safekeeper_reconciler;
4 : mod safekeeper_service;
5 : mod tenant_shard_iterator;
6 :
7 : use crate::hadron_token::HadronTokenGenerator;
8 : use std::borrow::Cow;
9 : use std::cmp::Ordering;
10 : use std::collections::{BTreeMap, HashMap, HashSet};
11 : use std::error::Error;
12 : use std::num::NonZeroU32;
13 : use std::ops::{Deref, DerefMut};
14 : use std::path::PathBuf;
15 : use std::str::FromStr;
16 : use std::sync::{Arc, OnceLock};
17 : use std::time::{Duration, Instant, SystemTime};
18 :
19 : use anyhow::Context;
20 : use control_plane::storage_controller::{
21 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
22 : };
23 : use diesel::result::DatabaseErrorKind;
24 : use futures::StreamExt;
25 : use futures::stream::FuturesUnordered;
26 : use http_utils::error::ApiError;
27 : use hyper::Uri;
28 : use itertools::Itertools;
29 : use pageserver_api::config::PostHogConfig;
30 : use pageserver_api::controller_api::{
31 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
32 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
33 : ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
34 : SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse, TenantCreateResponseShard,
35 : TenantDescribeResponse, TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
36 : TenantShardMigrateRequest, TenantShardMigrateResponse, TenantTimelineDescribeResponse,
37 : };
38 : use pageserver_api::models::{
39 : self, DetachBehavior, LocationConfig, LocationConfigListResponse, LocationConfigMode, LsnLease,
40 : PageserverUtilization, SecondaryProgress, ShardImportStatus, ShardParameters, TenantConfig,
41 : TenantConfigPatchRequest, TenantConfigRequest, TenantLocationConfigRequest,
42 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
43 : TenantShardSplitResponse, TenantSorting, TenantTimeTravelRequest,
44 : TimelineArchivalConfigRequest, TimelineCreateRequest, TimelineCreateResponseStorcon,
45 : TimelineInfo, TopTenantShardItem, TopTenantShardsRequest,
46 : };
47 : use pageserver_api::shard::{
48 : DEFAULT_STRIPE_SIZE, ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
49 : };
50 : use pageserver_api::upcall_api::{
51 : PutTimelineImportStatusRequest, ReAttachRequest, ReAttachResponse, ReAttachResponseTenant,
52 : TimelineImportStatusRequest, ValidateRequest, ValidateResponse, ValidateResponseTenant,
53 : };
54 : use pageserver_client::{BlockUnblock, mgmt_api};
55 : use reqwest::{Certificate, StatusCode};
56 : use safekeeper_api::models::SafekeeperUtilization;
57 : use safekeeper_reconciler::SafekeeperReconcilers;
58 : use tenant_shard_iterator::{TenantShardExclusiveIterator, create_shared_shard_iterator};
59 : use tokio::sync::TryAcquireError;
60 : use tokio::sync::mpsc::error::TrySendError;
61 : use tokio_util::sync::CancellationToken;
62 : use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
63 : use utils::completion::Barrier;
64 : use utils::env;
65 : use utils::generation::Generation;
66 : use utils::id::{NodeId, TenantId, TimelineId};
67 : use utils::lsn::Lsn;
68 : use utils::shard::ShardIndex;
69 : use utils::sync::gate::{Gate, GateGuard};
70 : use utils::{failpoint_support, pausable_failpoint};
71 :
72 : use crate::background_node_operations::{
73 : Delete, Drain, Fill, MAX_RECONCILES_PER_OPERATION, Operation, OperationError, OperationHandler,
74 : };
75 : use crate::compute_hook::{self, ComputeHook, NotifyError};
76 : use crate::heartbeater::{Heartbeater, PageserverState, SafekeeperState};
77 : use crate::id_lock_map::{
78 : IdLockMap, TracingExclusiveGuard, trace_exclusive_lock, trace_shared_lock,
79 : };
80 : use crate::leadership::Leadership;
81 : use crate::metrics;
82 : use crate::node::{AvailabilityTransition, Node};
83 : use crate::operation_utils::{self, TenantShardDrain, TenantShardDrainAction};
84 : use crate::pageserver_client::PageserverClient;
85 : use crate::peer_client::GlobalObservedState;
86 : use crate::persistence::split_state::SplitState;
87 : use crate::persistence::{
88 : AbortShardSplitStatus, ControllerPersistence, DatabaseError, DatabaseResult,
89 : MetadataHealthPersistence, Persistence, ShardGenerationState, TenantFilter,
90 : TenantShardPersistence,
91 : };
92 : use crate::reconciler::{
93 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder, ReconcilerPriority,
94 : attached_location_conf,
95 : };
96 : use crate::safekeeper::Safekeeper;
97 : use crate::scheduler::{
98 : AttachedShardTag, MaySchedule, ScheduleContext, ScheduleError, ScheduleMode, Scheduler,
99 : };
100 : use crate::tenant_shard::{
101 : IntentState, MigrateAttachment, ObservedState, ObservedStateDelta, ObservedStateLocation,
102 : ReconcileNeeded, ReconcileResult, ReconcileWaitError, ReconcilerStatus, ReconcilerWaiter,
103 : ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
104 : };
105 : use crate::timeline_import::{
106 : FinalizingImport, ImportResult, ShardImportStatuses, TimelineImport,
107 : TimelineImportFinalizeError, TimelineImportState, UpcallClient,
108 : };
109 :
110 : const WAITER_OPERATION_POLL_TIMEOUT: Duration = Duration::from_millis(500);
111 :
112 : // For operations that should be quick, like attaching a new tenant
113 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
114 :
115 : // For operations that might be slow, like migrating a tenant with
116 : // some data in it.
117 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
118 :
119 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
120 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
121 : // input generation from future requests as authoritative.
122 : const INITIAL_GENERATION: Generation = Generation::new(0);
123 :
124 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
125 : /// up on unresponsive pageservers and proceed.
126 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
127 :
128 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
129 : /// This must be long enough to cover node restarts as well as normal operations: in future
130 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
131 :
132 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
133 : /// offline.
134 : ///
135 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
136 : /// handling of the re-attach response may take a long time and blocks heartbeats from
137 : /// being handled on the pageserver side.
138 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
139 :
140 : /// How often to send heartbeats to registered nodes?
141 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
142 :
143 : /// How long is too long for a reconciliation?
144 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
145 :
146 : #[derive(Clone, strum_macros::Display)]
147 : enum TenantOperations {
148 : Create,
149 : LocationConfig,
150 : ConfigSet,
151 : ConfigPatch,
152 : TimeTravelRemoteStorage,
153 : Delete,
154 : UpdatePolicy,
155 : ShardSplit,
156 : SecondaryDownload,
157 : TimelineCreate,
158 : TimelineDelete,
159 : AttachHook,
160 : TimelineArchivalConfig,
161 : TimelineDetachAncestor,
162 : TimelineGcBlockUnblock,
163 : DropDetached,
164 : DownloadHeatmapLayers,
165 : TimelineLsnLease,
166 : TimelineSafekeeperMigrate,
167 : }
168 :
169 : #[derive(Clone, strum_macros::Display)]
170 : enum NodeOperations {
171 : Register,
172 : Configure,
173 : Delete,
174 : DeleteTombstone,
175 : }
176 :
177 : /// The leadership status for the storage controller process.
178 : /// Allowed transitions are:
179 : /// 1. Leader -> SteppedDown
180 : /// 2. Candidate -> Leader
181 : #[derive(
182 : Eq,
183 : PartialEq,
184 : Copy,
185 : Clone,
186 : strum_macros::Display,
187 : strum_macros::EnumIter,
188 : measured::FixedCardinalityLabel,
189 : )]
190 : #[strum(serialize_all = "snake_case")]
191 : pub(crate) enum LeadershipStatus {
192 : /// This is the steady state where the storage controller can produce
193 : /// side effects in the cluster.
194 : Leader,
195 : /// We've been notified to step down by another candidate. No reconciliations
196 : /// take place in this state.
197 : SteppedDown,
198 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
199 : #[allow(unused)]
200 : Candidate,
201 : }
202 :
203 : enum ShardGenerationValidity {
204 : Valid,
205 : Mismatched {
206 : claimed: Generation,
207 : actual: Option<Generation>,
208 : },
209 : }
210 :
211 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
212 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
213 : pub const SAFEKEEPER_RECONCILER_CONCURRENCY_DEFAULT: usize = 32;
214 :
215 : // Number of consecutive reconciliations that have occurred for one shard,
216 : // after which the shard is ignored when considering to run optimizations.
217 : const MAX_CONSECUTIVE_RECONCILES: usize = 10;
218 :
219 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
220 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
221 : // than they're being pushed onto the queue.
222 : const MAX_DELAYED_RECONCILES: usize = 10000;
223 :
224 : // Top level state available to all HTTP handlers
225 : struct ServiceState {
226 : leadership_status: LeadershipStatus,
227 :
228 : tenants: BTreeMap<TenantShardId, TenantShard>,
229 :
230 : nodes: Arc<HashMap<NodeId, Node>>,
231 :
232 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
233 :
234 : safekeeper_reconcilers: SafekeeperReconcilers,
235 :
236 : scheduler: Scheduler,
237 :
238 : /// Ongoing background operation on the cluster if any is running.
239 : /// Note that only one such operation may run at any given time,
240 : /// hence the type choice.
241 : ongoing_operation: Option<OperationHandler>,
242 :
243 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
244 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
245 :
246 : /// Tracks ongoing timeline import finalization tasks
247 : imports_finalizing: BTreeMap<(TenantId, TimelineId), FinalizingImport>,
248 : }
249 :
250 : /// Transform an error from a pageserver into an error to return to callers of a storage
251 : /// controller API.
252 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
253 0 : match e {
254 0 : mgmt_api::Error::SendRequest(e) => {
255 : // Presume errors sending requests are connectivity/availability issues
256 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
257 : }
258 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
259 : // Presume errors receiving body are connectivity/availability issues
260 0 : ApiError::ResourceUnavailable(
261 0 : format!("{node} error receiving error body: {str}").into(),
262 0 : )
263 : }
264 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
265 : // Return 500 for decoding errors.
266 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
267 : }
268 0 : mgmt_api::Error::ReceiveBody(err) => {
269 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
270 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
271 0 : ApiError::ResourceUnavailable(
272 0 : format!("{node} error receiving error body: {err} {src_str}").into(),
273 0 : )
274 : }
275 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
276 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
277 : }
278 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
279 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
280 : }
281 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
282 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
283 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
284 : // internal server errors, showing that something is wrong with the pageserver or
285 : // storage controller's auth configuration.
286 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
287 : }
288 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
289 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
290 : // turn ask our clients to wait + retry
291 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
292 : }
293 0 : mgmt_api::Error::ApiError(status, msg) => {
294 : // Presume general case of pageserver API errors is that we tried to do something
295 : // that can't be done right now.
296 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
297 : }
298 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
299 0 : mgmt_api::Error::Timeout(e) => ApiError::Timeout(e.into()),
300 : }
301 0 : }
302 :
303 : impl ServiceState {
304 0 : fn new(
305 0 : nodes: HashMap<NodeId, Node>,
306 0 : safekeepers: HashMap<NodeId, Safekeeper>,
307 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
308 0 : scheduler: Scheduler,
309 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
310 0 : initial_leadership_status: LeadershipStatus,
311 0 : reconcilers_cancel: CancellationToken,
312 0 : ) -> Self {
313 0 : metrics::update_leadership_status(initial_leadership_status);
314 :
315 0 : Self {
316 0 : leadership_status: initial_leadership_status,
317 0 : tenants,
318 0 : nodes: Arc::new(nodes),
319 0 : safekeepers: Arc::new(safekeepers),
320 0 : safekeeper_reconcilers: SafekeeperReconcilers::new(reconcilers_cancel),
321 0 : scheduler,
322 0 : ongoing_operation: None,
323 0 : delayed_reconcile_rx,
324 0 : imports_finalizing: Default::default(),
325 0 : }
326 0 : }
327 :
328 0 : fn parts_mut(
329 0 : &mut self,
330 0 : ) -> (
331 0 : &mut Arc<HashMap<NodeId, Node>>,
332 0 : &mut BTreeMap<TenantShardId, TenantShard>,
333 0 : &mut Scheduler,
334 0 : ) {
335 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
336 0 : }
337 :
338 : #[allow(clippy::type_complexity)]
339 0 : fn parts_mut_sk(
340 0 : &mut self,
341 0 : ) -> (
342 0 : &mut Arc<HashMap<NodeId, Node>>,
343 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
344 0 : &mut BTreeMap<TenantShardId, TenantShard>,
345 0 : &mut Scheduler,
346 0 : ) {
347 0 : (
348 0 : &mut self.nodes,
349 0 : &mut self.safekeepers,
350 0 : &mut self.tenants,
351 0 : &mut self.scheduler,
352 0 : )
353 0 : }
354 :
355 0 : fn get_leadership_status(&self) -> LeadershipStatus {
356 0 : self.leadership_status
357 0 : }
358 :
359 0 : fn step_down(&mut self) {
360 0 : self.leadership_status = LeadershipStatus::SteppedDown;
361 0 : metrics::update_leadership_status(self.leadership_status);
362 0 : }
363 :
364 0 : fn become_leader(&mut self) {
365 0 : self.leadership_status = LeadershipStatus::Leader;
366 0 : metrics::update_leadership_status(self.leadership_status);
367 0 : }
368 : }
369 :
370 : #[derive(Clone)]
371 : pub struct Config {
372 : // All pageservers managed by one instance of this service must have
373 : // the same public key. This JWT token will be used to authenticate
374 : // this service to the pageservers it manages.
375 : pub pageserver_jwt_token: Option<String>,
376 :
377 : // All safekeepers managed by one instance of this service must have
378 : // the same public key. This JWT token will be used to authenticate
379 : // this service to the safekeepers it manages.
380 : pub safekeeper_jwt_token: Option<String>,
381 :
382 : // This JWT token will be used to authenticate this service to the control plane.
383 : pub control_plane_jwt_token: Option<String>,
384 :
385 : // This JWT token will be used to authenticate with other storage controller instances
386 : pub peer_jwt_token: Option<String>,
387 :
388 : /// Prefix for storage API endpoints of the control plane. We use this prefix to compute
389 : /// URLs that we use to send pageserver and safekeeper attachment locations.
390 : /// If this is None, the compute hook will assume it is running in a test environment
391 : /// and try to invoke neon_local instead.
392 : pub control_plane_url: Option<String>,
393 :
394 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
395 : /// considered active. Once the grace period elapses, the next heartbeat failure will
396 : /// mark the pagseserver offline.
397 : pub max_offline_interval: Duration,
398 :
399 : /// Extended grace period within which pageserver may not respond to heartbeats.
400 : /// This extended grace period kicks in after the node has been drained for restart
401 : /// and/or upon handling the re-attach request from a node.
402 : pub max_warming_up_interval: Duration,
403 :
404 : /// How many normal-priority Reconcilers may be spawned concurrently
405 : pub reconciler_concurrency: usize,
406 :
407 : /// How many high-priority Reconcilers may be spawned concurrently
408 : pub priority_reconciler_concurrency: usize,
409 :
410 : /// How many safekeeper reconciles may happen concurrently (per safekeeper)
411 : pub safekeeper_reconciler_concurrency: usize,
412 :
413 : /// How many API requests per second to allow per tenant, across all
414 : /// tenant-scoped API endpoints. Further API requests queue until ready.
415 : pub tenant_rate_limit: NonZeroU32,
416 :
417 : /// If a tenant shard's largest timeline (max_logical_size) exceeds this value, all tenant
418 : /// shards will be split in 2 until they fall below split_threshold (up to max_split_shards).
419 : ///
420 : /// This will greedily split into as many shards as necessary to fall below split_threshold, as
421 : /// powers of 2: if a tenant shard is 7 times larger than split_threshold, it will split into 8
422 : /// immediately, rather than first 2 then 4 then 8.
423 : ///
424 : /// None or 0 disables auto-splitting.
425 : ///
426 : /// TODO: consider using total logical size of all timelines instead.
427 : pub split_threshold: Option<u64>,
428 :
429 : /// The maximum number of shards a tenant can be split into during autosplits. Does not affect
430 : /// manual split requests. 0 or 1 disables autosplits, as we already have 1 shard.
431 : pub max_split_shards: u8,
432 :
433 : /// The size at which an unsharded tenant should initially split. Ingestion is significantly
434 : /// faster with multiple shards, so eagerly splitting below split_threshold will typically speed
435 : /// up initial ingestion of large tenants.
436 : ///
437 : /// This should be below split_threshold, but it is not required. If both split_threshold and
438 : /// initial_split_threshold qualify, the largest number of target shards will be used.
439 : ///
440 : /// Does not apply to already sharded tenants: changing initial_split_threshold or
441 : /// initial_split_shards is not retroactive for already-sharded tenants.
442 : ///
443 : /// None or 0 disables initial splits.
444 : pub initial_split_threshold: Option<u64>,
445 :
446 : /// The number of shards to split into when reaching initial_split_threshold. Will
447 : /// be clamped to max_split_shards.
448 : ///
449 : /// 0 or 1 disables initial splits. Has no effect if initial_split_threshold is disabled.
450 : pub initial_split_shards: u8,
451 :
452 : // TODO: make this cfg(feature = "testing")
453 : pub neon_local_repo_dir: Option<PathBuf>,
454 :
455 : // Maximum acceptable download lag for the secondary location
456 : // while draining a node. If the secondary location is lagging
457 : // by more than the configured amount, then the secondary is not
458 : // upgraded to primary.
459 : pub max_secondary_lag_bytes: Option<u64>,
460 :
461 : pub heartbeat_interval: Duration,
462 :
463 : pub address_for_peers: Option<Uri>,
464 :
465 : pub start_as_candidate: bool,
466 :
467 : pub long_reconcile_threshold: Duration,
468 :
469 : pub use_https_pageserver_api: bool,
470 :
471 : pub use_https_safekeeper_api: bool,
472 :
473 : pub ssl_ca_certs: Vec<Certificate>,
474 :
475 : pub timelines_onto_safekeepers: bool,
476 :
477 : pub use_local_compute_notifications: bool,
478 :
479 : /// Number of safekeepers to choose for a timeline when creating it.
480 : /// Safekeepers will be choosen from different availability zones.
481 : pub timeline_safekeeper_count: usize,
482 :
483 : /// PostHog integration config
484 : pub posthog_config: Option<PostHogConfig>,
485 :
486 : /// When set, actively checks and initiates heatmap downloads/uploads.
487 : pub kick_secondary_downloads: bool,
488 :
489 : /// Timeout used for HTTP client of split requests. [`Duration::MAX`] if None.
490 : pub shard_split_request_timeout: Duration,
491 :
492 : // Feature flag: Whether the storage controller should act to rectify pageserver-reported local disk loss.
493 : pub handle_ps_local_disk_loss: bool,
494 : }
495 :
496 : impl From<DatabaseError> for ApiError {
497 0 : fn from(err: DatabaseError) -> ApiError {
498 0 : match err {
499 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
500 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
501 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
502 0 : ApiError::ShuttingDown
503 : }
504 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
505 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
506 : }
507 0 : DatabaseError::Cas(reason) => ApiError::Conflict(reason),
508 : }
509 0 : }
510 : }
511 :
512 : enum InitialShardScheduleOutcome {
513 : Scheduled(TenantCreateResponseShard),
514 : NotScheduled,
515 : ShardScheduleError(ScheduleError),
516 : }
517 :
518 : pub struct Service {
519 : inner: Arc<std::sync::RwLock<ServiceState>>,
520 : config: Config,
521 : persistence: Arc<Persistence>,
522 :
523 : // HadronTokenGenerator to generate (sign) JWTs during compute deployment and compute-spec generation.
524 : #[allow(unused)]
525 : token_generator: Option<HadronTokenGenerator>,
526 :
527 : compute_hook: Arc<ComputeHook>,
528 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
529 :
530 : heartbeater_ps: Heartbeater<Node, PageserverState>,
531 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
532 :
533 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
534 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
535 :
536 : // Locking on a tenant granularity (covers all shards in the tenant):
537 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
538 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
539 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
540 :
541 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
542 : // that transition it to/from Active.
543 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
544 :
545 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
546 : // and reconciliation on startup.
547 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
548 :
549 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
550 : // a human user might be waiting for.
551 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
552 :
553 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
554 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
555 : ///
556 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
557 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
558 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
559 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
560 :
561 : // Process shutdown will fire this token
562 : cancel: CancellationToken,
563 :
564 : // Child token of [`Service::cancel`] used by reconcilers
565 : reconcilers_cancel: CancellationToken,
566 :
567 : // Background tasks will hold this gate
568 : gate: Gate,
569 :
570 : // Reconcilers background tasks will hold this gate
571 : reconcilers_gate: Gate,
572 :
573 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
574 : /// passes, it isn't safe to do any actions that mutate tenants.
575 : pub(crate) startup_complete: Barrier,
576 :
577 : /// HTTP client with proper CA certs.
578 : http_client: reqwest::Client,
579 :
580 : /// Handle for the step down background task if one was ever requested
581 : step_down_barrier: OnceLock<tokio::sync::watch::Receiver<Option<GlobalObservedState>>>,
582 : }
583 :
584 : impl From<ReconcileWaitError> for ApiError {
585 0 : fn from(value: ReconcileWaitError) -> Self {
586 0 : match value {
587 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
588 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
589 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
590 : }
591 0 : }
592 : }
593 :
594 : impl From<OperationError> for ApiError {
595 0 : fn from(value: OperationError) -> Self {
596 0 : match value {
597 0 : OperationError::NodeStateChanged(err)
598 0 : | OperationError::FinalizeError(err)
599 0 : | OperationError::ImpossibleConstraint(err) => {
600 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
601 : }
602 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
603 : }
604 0 : }
605 : }
606 :
607 : #[allow(clippy::large_enum_variant)]
608 : enum TenantCreateOrUpdate {
609 : Create(TenantCreateRequest),
610 : Update(Vec<ShardUpdate>),
611 : }
612 :
613 : struct ShardSplitParams {
614 : old_shard_count: ShardCount,
615 : new_shard_count: ShardCount,
616 : new_stripe_size: Option<ShardStripeSize>,
617 : targets: Vec<ShardSplitTarget>,
618 : policy: PlacementPolicy,
619 : config: TenantConfig,
620 : shard_ident: ShardIdentity,
621 : preferred_az_id: Option<AvailabilityZone>,
622 : }
623 :
624 : // When preparing for a shard split, we may either choose to proceed with the split,
625 : // or find that the work is already done and return NoOp.
626 : enum ShardSplitAction {
627 : Split(Box<ShardSplitParams>),
628 : NoOp(TenantShardSplitResponse),
629 : }
630 :
631 : // A parent shard which will be split
632 : struct ShardSplitTarget {
633 : parent_id: TenantShardId,
634 : node: Node,
635 : child_ids: Vec<TenantShardId>,
636 : }
637 :
638 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
639 : /// might not be available. We therefore use a queue of abort operations processed in the background.
640 : struct TenantShardSplitAbort {
641 : tenant_id: TenantId,
642 : /// The target values from the request that failed
643 : new_shard_count: ShardCount,
644 : new_stripe_size: Option<ShardStripeSize>,
645 : /// Until this abort op is complete, no other operations may be done on the tenant
646 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
647 : /// The reconciler gate for the duration of the split operation, and any included abort.
648 : _gate: GateGuard,
649 : }
650 :
651 : #[derive(thiserror::Error, Debug)]
652 : enum TenantShardSplitAbortError {
653 : #[error(transparent)]
654 : Database(#[from] DatabaseError),
655 : #[error(transparent)]
656 : Remote(#[from] mgmt_api::Error),
657 : #[error("Unavailable")]
658 : Unavailable,
659 : }
660 :
661 : /// Inputs for computing a target shard count for a tenant.
662 : struct ShardSplitInputs {
663 : /// Current shard count.
664 : shard_count: ShardCount,
665 : /// Total size of largest timeline summed across all shards.
666 : max_logical_size: u64,
667 : /// Size-based split threshold. Zero if size-based splits are disabled.
668 : split_threshold: u64,
669 : /// Upper bound on target shards. 0 or 1 disables splits.
670 : max_split_shards: u8,
671 : /// Initial split threshold. Zero if initial splits are disabled.
672 : initial_split_threshold: u64,
673 : /// Number of shards for initial splits. 0 or 1 disables initial splits.
674 : initial_split_shards: u8,
675 : }
676 :
677 : struct ShardUpdate {
678 : tenant_shard_id: TenantShardId,
679 : placement_policy: PlacementPolicy,
680 : tenant_config: TenantConfig,
681 :
682 : /// If this is None, generation is not updated.
683 : generation: Option<Generation>,
684 :
685 : /// If this is None, scheduling policy is not updated.
686 : scheduling_policy: Option<ShardSchedulingPolicy>,
687 : }
688 :
689 : enum StopReconciliationsReason {
690 : ShuttingDown,
691 : SteppingDown,
692 : }
693 :
694 : impl std::fmt::Display for StopReconciliationsReason {
695 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
696 0 : let s = match self {
697 0 : Self::ShuttingDown => "Shutting down",
698 0 : Self::SteppingDown => "Stepping down",
699 : };
700 0 : write!(writer, "{s}")
701 0 : }
702 : }
703 :
704 : pub(crate) enum ReconcileResultRequest {
705 : ReconcileResult(ReconcileResult),
706 : Stop,
707 : }
708 :
709 : #[derive(Clone)]
710 : pub(crate) struct MutationLocation {
711 : pub(crate) node: Node,
712 : pub(crate) generation: Generation,
713 : }
714 :
715 : #[derive(Clone)]
716 : pub(crate) struct ShardMutationLocations {
717 : pub(crate) latest: MutationLocation,
718 : pub(crate) other: Vec<MutationLocation>,
719 : }
720 :
721 : #[derive(Default, Clone)]
722 : pub(crate) struct TenantMutationLocations(pub BTreeMap<TenantShardId, ShardMutationLocations>);
723 :
724 : struct ReconcileAllResult {
725 : spawned_reconciles: usize,
726 : stuck_reconciles: usize,
727 : has_delayed_reconciles: bool,
728 : }
729 :
730 : impl ReconcileAllResult {
731 0 : fn new(
732 0 : spawned_reconciles: usize,
733 0 : stuck_reconciles: usize,
734 0 : has_delayed_reconciles: bool,
735 0 : ) -> Self {
736 0 : assert!(
737 0 : spawned_reconciles >= stuck_reconciles,
738 0 : "It is impossible to have less spawned reconciles than stuck reconciles"
739 : );
740 0 : Self {
741 0 : spawned_reconciles,
742 0 : stuck_reconciles,
743 0 : has_delayed_reconciles,
744 0 : }
745 0 : }
746 :
747 : /// We can run optimizations only if we don't have any delayed reconciles and
748 : /// all spawned reconciles are also stuck reconciles.
749 0 : fn can_run_optimizations(&self) -> bool {
750 0 : !self.has_delayed_reconciles && self.spawned_reconciles == self.stuck_reconciles
751 0 : }
752 : }
753 :
754 : enum TenantIdOrShardId {
755 : TenantId(TenantId),
756 : TenantShardId(TenantShardId),
757 : }
758 :
759 : impl TenantIdOrShardId {
760 0 : fn tenant_id(&self) -> TenantId {
761 0 : match self {
762 0 : TenantIdOrShardId::TenantId(tenant_id) => *tenant_id,
763 0 : TenantIdOrShardId::TenantShardId(tenant_shard_id) => tenant_shard_id.tenant_id,
764 : }
765 0 : }
766 :
767 0 : fn matches(&self, tenant_shard_id: &TenantShardId) -> bool {
768 0 : match self {
769 0 : TenantIdOrShardId::TenantId(tenant_id) => tenant_shard_id.tenant_id == *tenant_id,
770 0 : TenantIdOrShardId::TenantShardId(this_tenant_shard_id) => {
771 0 : this_tenant_shard_id == tenant_shard_id
772 : }
773 : }
774 0 : }
775 : }
776 :
777 : impl Service {
778 0 : pub fn get_config(&self) -> &Config {
779 0 : &self.config
780 0 : }
781 :
782 0 : pub fn get_http_client(&self) -> &reqwest::Client {
783 0 : &self.http_client
784 0 : }
785 :
786 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
787 : /// view of the world, and determine which pageservers are responsive.
788 : #[instrument(skip_all)]
789 : async fn startup_reconcile(
790 : self: &Arc<Service>,
791 : current_leader: Option<ControllerPersistence>,
792 : leader_step_down_state: Option<GlobalObservedState>,
793 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
794 : Result<(), (TenantShardId, NotifyError)>,
795 : >,
796 : ) {
797 : // Startup reconciliation does I/O to other services: whether they
798 : // are responsive or not, we should aim to finish within our deadline, because:
799 : // - If we don't, a k8s readiness hook watching /ready will kill us.
800 : // - While we're waiting for startup reconciliation, we are not fully
801 : // available for end user operations like creating/deleting tenants and timelines.
802 : //
803 : // We set multiple deadlines to break up the time available between the phases of work: this is
804 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
805 : let start_at = Instant::now();
806 : let node_scan_deadline = start_at
807 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
808 : .expect("Reconcile timeout is a modest constant");
809 :
810 : let observed = if let Some(state) = leader_step_down_state {
811 : tracing::info!(
812 : "Using observed state received from leader at {}",
813 : current_leader.as_ref().unwrap().address
814 : );
815 :
816 : state
817 : } else {
818 : self.build_global_observed_state(node_scan_deadline).await
819 : };
820 :
821 : // Accumulate a list of any tenant locations that ought to be detached
822 : let mut cleanup = Vec::new();
823 :
824 : // Send initial heartbeat requests to all nodes loaded from the database
825 : let all_nodes = {
826 : let locked = self.inner.read().unwrap();
827 : locked.nodes.clone()
828 : };
829 : let (mut nodes_online, mut sks_online) =
830 : self.initial_heartbeat_round(all_nodes.keys()).await;
831 :
832 : // List of tenants for which we will attempt to notify compute of their location at startup
833 : let mut compute_notifications = Vec::new();
834 :
835 : // Populate intent and observed states for all tenants, based on reported state on pageservers
836 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
837 : let shard_count = {
838 : let mut locked = self.inner.write().unwrap();
839 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
840 :
841 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
842 : let mut new_nodes = (**nodes).clone();
843 : for (node_id, node) in new_nodes.iter_mut() {
844 : if let Some(utilization) = nodes_online.remove(node_id) {
845 : node.set_availability(NodeAvailability::Active(utilization));
846 : scheduler.node_upsert(node);
847 : }
848 : }
849 : *nodes = Arc::new(new_nodes);
850 :
851 : let mut new_sks = (**safekeepers).clone();
852 : for (node_id, node) in new_sks.iter_mut() {
853 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
854 : node.set_availability(SafekeeperState::Available {
855 : utilization,
856 : last_seen_at,
857 : });
858 : }
859 : }
860 : *safekeepers = Arc::new(new_sks);
861 :
862 : for (tenant_shard_id, observed_state) in observed.0 {
863 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
864 : for node_id in observed_state.locations.keys() {
865 : cleanup.push((tenant_shard_id, *node_id));
866 : }
867 :
868 : continue;
869 : };
870 :
871 : tenant_shard.observed = observed_state;
872 : }
873 :
874 : // Populate each tenant's intent state
875 : let mut schedule_context = ScheduleContext::default();
876 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
877 : if tenant_shard_id.shard_number == ShardNumber(0) {
878 : // Reset scheduling context each time we advance to the next Tenant
879 : schedule_context = ScheduleContext::default();
880 : }
881 :
882 : tenant_shard.intent_from_observed(scheduler);
883 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
884 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
885 : // not enough pageservers are available. The tenant may well still be available
886 : // to clients.
887 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
888 : } else {
889 : // If we're both intending and observed to be attached at a particular node, we will
890 : // emit a compute notification for this. In the case where our observed state does not
891 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
892 : if let Some(attached_at) = tenant_shard.stably_attached() {
893 : compute_notifications.push(compute_hook::ShardUpdate {
894 : tenant_shard_id: *tenant_shard_id,
895 : node_id: attached_at,
896 : stripe_size: tenant_shard.shard.stripe_size,
897 : preferred_az: tenant_shard
898 : .preferred_az()
899 0 : .map(|az| Cow::Owned(az.clone())),
900 : });
901 : }
902 : }
903 : }
904 :
905 : tenants.len()
906 : };
907 :
908 : // Before making any obeservable changes to the cluster, persist self
909 : // as leader in database and memory.
910 : let leadership = Leadership::new(
911 : self.persistence.clone(),
912 : self.config.clone(),
913 : self.cancel.child_token(),
914 : );
915 :
916 : if let Err(e) = leadership.become_leader(current_leader).await {
917 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
918 : std::process::exit(1);
919 : }
920 :
921 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
922 : let sk_schedule_requests =
923 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
924 : Ok(v) => v,
925 : Err(e) => {
926 : tracing::warn!(
927 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
928 : );
929 : // std::process::exit(1);
930 : Vec::new()
931 : }
932 : };
933 :
934 : {
935 : let mut locked = self.inner.write().unwrap();
936 : locked.become_leader();
937 :
938 : for (sk_id, _sk) in locked.safekeepers.clone().iter() {
939 : locked.safekeeper_reconcilers.start_reconciler(*sk_id, self);
940 : }
941 :
942 : locked
943 : .safekeeper_reconcilers
944 : .schedule_request_vec(sk_schedule_requests);
945 : }
946 :
947 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
948 : // generation_pageserver in the database.
949 :
950 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
951 : // will emit compute hook notifications when they reconcile.
952 : //
953 : // Ordering: our calls to notify_attach_background synchronously establish a relative order for these notifications vs. any later
954 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
955 : // calls will be correctly ordered wrt these.
956 : //
957 : // Concurrency: we call notify_attach_background for all tenants, which will create O(N) tokio tasks, but almost all of them
958 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
959 : // unit and start doing I/O.
960 : tracing::info!(
961 : "Sending {} compute notifications",
962 : compute_notifications.len()
963 : );
964 : self.compute_hook.notify_attach_background(
965 : compute_notifications,
966 : bg_compute_notify_result_tx.clone(),
967 : &self.cancel,
968 : );
969 :
970 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
971 : // which require it: under normal circumstances this should only include tenants that were in some
972 : // transient state before we restarted, or any tenants whose compute hooks failed above.
973 : tracing::info!("Checking for shards in need of reconciliation...");
974 : let reconcile_all_result = self.reconcile_all();
975 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
976 : // normal operations may proceed.
977 :
978 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
979 : // background because it does not need to complete in order to proceed with other work.
980 : if !cleanup.is_empty() {
981 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
982 : tokio::task::spawn({
983 : let cleanup_self = self.clone();
984 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
985 : });
986 : }
987 :
988 : // Reconcile the timeline imports:
989 : // 1. Mark each tenant shard of tenants with an importing timeline as importing.
990 : // 2. Finalize the completed imports in the background. This handles the case where
991 : // the previous storage controller instance shut down whilst finalizing imports.
992 : let imports = self.persistence.list_timeline_imports().await;
993 : match imports {
994 : Ok(mut imports) => {
995 : {
996 : let mut locked = self.inner.write().unwrap();
997 : for import in &imports {
998 : locked
999 : .tenants
1000 : .range_mut(TenantShardId::tenant_range(import.tenant_id))
1001 0 : .for_each(|(_id, shard)| {
1002 0 : shard.importing = TimelineImportState::Importing
1003 0 : });
1004 : }
1005 : }
1006 :
1007 0 : imports.retain(|import| import.is_complete());
1008 : tokio::task::spawn({
1009 : let finalize_imports_self = self.clone();
1010 0 : async move {
1011 0 : finalize_imports_self
1012 0 : .finalize_timeline_imports(imports)
1013 0 : .await
1014 0 : }
1015 : });
1016 : }
1017 : Err(err) => {
1018 : tracing::error!("Could not retrieve completed imports from database: {err}");
1019 : }
1020 : }
1021 :
1022 : let spawned_reconciles = reconcile_all_result.spawned_reconciles;
1023 : tracing::info!(
1024 : "Startup complete, spawned {spawned_reconciles} reconciliation tasks ({shard_count} shards total)"
1025 : );
1026 : }
1027 :
1028 0 : async fn initial_heartbeat_round<'a>(
1029 0 : &self,
1030 0 : node_ids: impl Iterator<Item = &'a NodeId>,
1031 0 : ) -> (
1032 0 : HashMap<NodeId, PageserverUtilization>,
1033 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
1034 0 : ) {
1035 0 : assert!(!self.startup_complete.is_ready());
1036 :
1037 0 : let all_nodes = {
1038 0 : let locked = self.inner.read().unwrap();
1039 0 : locked.nodes.clone()
1040 : };
1041 :
1042 0 : let mut nodes_to_heartbeat = HashMap::new();
1043 0 : for node_id in node_ids {
1044 0 : match all_nodes.get(node_id) {
1045 0 : Some(node) => {
1046 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
1047 0 : }
1048 : None => {
1049 0 : tracing::warn!("Node {node_id} was removed during start-up");
1050 : }
1051 : }
1052 : }
1053 :
1054 0 : let all_sks = {
1055 0 : let locked = self.inner.read().unwrap();
1056 0 : locked.safekeepers.clone()
1057 : };
1058 :
1059 0 : tracing::info!("Sending initial heartbeats...");
1060 0 : let (res_ps, res_sk) = tokio::join!(
1061 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
1062 0 : self.heartbeater_sk.heartbeat(all_sks)
1063 : );
1064 :
1065 0 : let mut online_nodes = HashMap::new();
1066 0 : if let Ok(deltas) = res_ps {
1067 0 : for (node_id, status) in deltas.0 {
1068 0 : match status {
1069 0 : PageserverState::Available { utilization, .. } => {
1070 0 : online_nodes.insert(node_id, utilization);
1071 0 : }
1072 0 : PageserverState::Offline => {}
1073 : PageserverState::WarmingUp { .. } => {
1074 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
1075 : }
1076 : }
1077 : }
1078 0 : }
1079 :
1080 0 : let mut online_sks = HashMap::new();
1081 0 : if let Ok(deltas) = res_sk {
1082 0 : for (node_id, status) in deltas.0 {
1083 0 : match status {
1084 : SafekeeperState::Available {
1085 0 : utilization,
1086 0 : last_seen_at,
1087 0 : } => {
1088 0 : online_sks.insert(node_id, (utilization, last_seen_at));
1089 0 : }
1090 0 : SafekeeperState::Offline => {}
1091 : }
1092 : }
1093 0 : }
1094 :
1095 0 : (online_nodes, online_sks)
1096 0 : }
1097 :
1098 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
1099 : ///
1100 : /// The result includes only nodes which responded within the deadline
1101 0 : async fn scan_node_locations(
1102 0 : &self,
1103 0 : deadline: Instant,
1104 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
1105 0 : let nodes = {
1106 0 : let locked = self.inner.read().unwrap();
1107 0 : locked.nodes.clone()
1108 : };
1109 :
1110 0 : let mut node_results = HashMap::new();
1111 :
1112 0 : let mut node_list_futs = FuturesUnordered::new();
1113 :
1114 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
1115 0 : for node in nodes.values() {
1116 0 : node_list_futs.push({
1117 0 : async move {
1118 0 : tracing::info!("Scanning shards on node {node}...");
1119 0 : let timeout = Duration::from_secs(5);
1120 0 : let response = node
1121 0 : .with_client_retries(
1122 0 : |client| async move { client.list_location_config().await },
1123 0 : &self.http_client,
1124 0 : &self.config.pageserver_jwt_token,
1125 : 1,
1126 : 5,
1127 0 : timeout,
1128 0 : &self.cancel,
1129 : )
1130 0 : .await;
1131 0 : (node.get_id(), response)
1132 0 : }
1133 : });
1134 : }
1135 :
1136 : loop {
1137 0 : let (node_id, result) = tokio::select! {
1138 0 : next = node_list_futs.next() => {
1139 0 : match next {
1140 0 : Some(result) => result,
1141 : None =>{
1142 : // We got results for all our nodes
1143 0 : break;
1144 : }
1145 :
1146 : }
1147 : },
1148 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
1149 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
1150 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
1151 0 : break;
1152 : }
1153 : };
1154 :
1155 0 : let Some(list_response) = result else {
1156 0 : tracing::info!("Shutdown during startup_reconcile");
1157 0 : break;
1158 : };
1159 :
1160 0 : match list_response {
1161 0 : Err(e) => {
1162 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
1163 : }
1164 0 : Ok(listing) => {
1165 0 : node_results.insert(node_id, listing);
1166 0 : }
1167 : }
1168 : }
1169 :
1170 0 : node_results
1171 0 : }
1172 :
1173 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
1174 0 : let node_listings = self.scan_node_locations(deadline).await;
1175 0 : let mut observed = GlobalObservedState::default();
1176 :
1177 0 : for (node_id, location_confs) in node_listings {
1178 0 : tracing::info!(
1179 0 : "Received {} shard statuses from pageserver {}",
1180 0 : location_confs.tenant_shards.len(),
1181 : node_id
1182 : );
1183 :
1184 0 : for (tid, location_conf) in location_confs.tenant_shards {
1185 0 : let entry = observed.0.entry(tid).or_default();
1186 0 : entry.locations.insert(
1187 0 : node_id,
1188 0 : ObservedStateLocation {
1189 0 : conf: location_conf,
1190 0 : },
1191 0 : );
1192 0 : }
1193 : }
1194 :
1195 0 : observed
1196 0 : }
1197 :
1198 : /// Used during [`Self::startup_reconcile`] and shard splits: detach a list of unknown-to-us
1199 : /// tenants from pageservers.
1200 : ///
1201 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1202 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1203 : /// other task trying to attach it.
1204 : #[instrument(skip_all)]
1205 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1206 : let nodes = self.inner.read().unwrap().nodes.clone();
1207 :
1208 : for (tenant_shard_id, node_id) in cleanup {
1209 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1210 : let Some(node) = nodes.get(&node_id) else {
1211 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1212 : // a location to clean up on a node that has since been removed.
1213 : tracing::info!(
1214 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1215 : );
1216 : continue;
1217 : };
1218 :
1219 : if self.cancel.is_cancelled() {
1220 : break;
1221 : }
1222 :
1223 : let client = PageserverClient::new(
1224 : node.get_id(),
1225 : self.http_client.clone(),
1226 : node.base_url(),
1227 : self.config.pageserver_jwt_token.as_deref(),
1228 : );
1229 : match client
1230 : .location_config(
1231 : tenant_shard_id,
1232 : LocationConfig {
1233 : mode: LocationConfigMode::Detached,
1234 : generation: None,
1235 : secondary_conf: None,
1236 : shard_number: tenant_shard_id.shard_number.0,
1237 : shard_count: tenant_shard_id.shard_count.literal(),
1238 : shard_stripe_size: 0,
1239 : tenant_conf: models::TenantConfig::default(),
1240 : },
1241 : None,
1242 : false,
1243 : )
1244 : .await
1245 : {
1246 : Ok(()) => {
1247 : tracing::info!(
1248 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1249 : );
1250 : }
1251 : Err(e) => {
1252 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1253 : // break anything.
1254 : tracing::error!(
1255 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1256 : );
1257 : }
1258 : }
1259 : }
1260 : }
1261 :
1262 : /// Long running background task that periodically wakes up and looks for shards that need
1263 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1264 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1265 : /// for those retries.
1266 : #[instrument(skip_all)]
1267 : async fn background_reconcile(self: &Arc<Self>) {
1268 : self.startup_complete.clone().wait().await;
1269 :
1270 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1271 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1272 : while !self.reconcilers_cancel.is_cancelled() {
1273 : tokio::select! {
1274 : _ = interval.tick() => {
1275 : let reconcile_all_result = self.reconcile_all();
1276 : if reconcile_all_result.can_run_optimizations() {
1277 : // Run optimizer only when we didn't find any other work to do
1278 : self.optimize_all().await;
1279 : }
1280 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1281 : // must be responsive when new projects begin ingesting and reach the threshold.
1282 : self.autosplit_tenants().await;
1283 : },
1284 : _ = self.reconcilers_cancel.cancelled() => return
1285 : }
1286 : }
1287 : }
1288 : /// Heartbeat all storage nodes once in a while.
1289 : #[instrument(skip_all)]
1290 : async fn spawn_heartbeat_driver(self: &Arc<Self>) {
1291 : self.startup_complete.clone().wait().await;
1292 :
1293 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1294 : while !self.cancel.is_cancelled() {
1295 : tokio::select! {
1296 : _ = interval.tick() => { }
1297 : _ = self.cancel.cancelled() => return
1298 : };
1299 :
1300 : let nodes = {
1301 : let locked = self.inner.read().unwrap();
1302 : locked.nodes.clone()
1303 : };
1304 :
1305 : let safekeepers = {
1306 : let locked = self.inner.read().unwrap();
1307 : locked.safekeepers.clone()
1308 : };
1309 :
1310 : let (res_ps, res_sk) = tokio::join!(
1311 : self.heartbeater_ps.heartbeat(nodes),
1312 : self.heartbeater_sk.heartbeat(safekeepers)
1313 : );
1314 :
1315 : if let Ok(deltas) = res_ps {
1316 : let mut to_handle = Vec::default();
1317 :
1318 : for (node_id, state) in deltas.0 {
1319 : let new_availability = match state {
1320 : PageserverState::Available { utilization, .. } => {
1321 : NodeAvailability::Active(utilization)
1322 : }
1323 : PageserverState::WarmingUp { started_at } => {
1324 : NodeAvailability::WarmingUp(started_at)
1325 : }
1326 : PageserverState::Offline => {
1327 : // The node might have been placed in the WarmingUp state
1328 : // while the heartbeat round was on-going. Hence, filter out
1329 : // offline transitions for WarmingUp nodes that are still within
1330 : // their grace period.
1331 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1332 : .get_node(node_id)
1333 : .await
1334 : .as_ref()
1335 0 : .map(|n| n.get_availability())
1336 : {
1337 : let now = Instant::now();
1338 : if now - *started_at >= self.config.max_warming_up_interval {
1339 : NodeAvailability::Offline
1340 : } else {
1341 : NodeAvailability::WarmingUp(*started_at)
1342 : }
1343 : } else {
1344 : NodeAvailability::Offline
1345 : }
1346 : }
1347 : };
1348 :
1349 : let node_lock = trace_exclusive_lock(
1350 : &self.node_op_locks,
1351 : node_id,
1352 : NodeOperations::Configure,
1353 : )
1354 : .await;
1355 :
1356 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1357 :
1358 : // This is the code path for geniune availability transitions (i.e node
1359 : // goes unavailable and/or comes back online).
1360 : let res = self
1361 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1362 : .await;
1363 :
1364 : match res {
1365 : Ok(transition) => {
1366 : // Keep hold of the lock until the availability transitions
1367 : // have been handled in
1368 : // [`Service::handle_node_availability_transitions`] in order avoid
1369 : // racing with [`Service::external_node_configure`].
1370 : to_handle.push((node_id, node_lock, transition));
1371 : }
1372 : Err(ApiError::NotFound(_)) => {
1373 : // This should be rare, but legitimate since the heartbeats are done
1374 : // on a snapshot of the nodes.
1375 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1376 : }
1377 : Err(ApiError::ShuttingDown) => {
1378 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1379 : }
1380 : Err(err) => {
1381 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1382 : // becomes unavailable again, we may get an error here.
1383 : tracing::error!(
1384 : "Failed to update node state {} after heartbeat round: {}",
1385 : node_id,
1386 : err
1387 : );
1388 : }
1389 : }
1390 : }
1391 :
1392 : // We collected all the transitions above and now we handle them.
1393 : let res = self.handle_node_availability_transitions(to_handle).await;
1394 : if let Err(errs) = res {
1395 : for (node_id, err) in errs {
1396 : match err {
1397 : ApiError::NotFound(_) => {
1398 : // This should be rare, but legitimate since the heartbeats are done
1399 : // on a snapshot of the nodes.
1400 : tracing::info!(
1401 : "Node {} was not found after heartbeat round",
1402 : node_id
1403 : );
1404 : }
1405 : err => {
1406 : tracing::error!(
1407 : "Failed to handle availability transition for {} after heartbeat round: {}",
1408 : node_id,
1409 : err
1410 : );
1411 : }
1412 : }
1413 : }
1414 : }
1415 : }
1416 : if let Ok(deltas) = res_sk {
1417 : let mut to_activate = Vec::new();
1418 : {
1419 : let mut locked = self.inner.write().unwrap();
1420 : let mut safekeepers = (*locked.safekeepers).clone();
1421 :
1422 : for (id, state) in deltas.0 {
1423 : let Some(sk) = safekeepers.get_mut(&id) else {
1424 : tracing::info!(
1425 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1426 : );
1427 : continue;
1428 : };
1429 : if sk.scheduling_policy() == SkSchedulingPolicy::Activating
1430 : && let SafekeeperState::Available { .. } = state
1431 : {
1432 : to_activate.push(id);
1433 : }
1434 : sk.set_availability(state);
1435 : }
1436 : locked.safekeepers = Arc::new(safekeepers);
1437 : }
1438 : for sk_id in to_activate {
1439 : // TODO this can race with set_scheduling_policy (can create disjoint DB <-> in-memory state)
1440 : tracing::info!("Activating safekeeper {sk_id}");
1441 : match self.persistence.activate_safekeeper(sk_id.0 as i64).await {
1442 : Ok(Some(())) => {}
1443 : Ok(None) => {
1444 : tracing::info!(
1445 : "safekeeper {sk_id} has been removed from db or has different scheduling policy than active or activating"
1446 : );
1447 : }
1448 : Err(e) => {
1449 : tracing::warn!("couldn't apply activation of {sk_id} to db: {e}");
1450 : continue;
1451 : }
1452 : }
1453 : if let Err(e) = self
1454 : .set_safekeeper_scheduling_policy_in_mem(sk_id, SkSchedulingPolicy::Active)
1455 : .await
1456 : {
1457 : tracing::info!("couldn't activate safekeeper {sk_id} in memory: {e}");
1458 : continue;
1459 : }
1460 : tracing::info!("Activation of safekeeper {sk_id} done");
1461 : }
1462 : }
1463 : }
1464 : }
1465 :
1466 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1467 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1468 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1469 : /// will indicate that reconciliation is not needed.
1470 : #[instrument(skip_all, fields(
1471 : seq=%result.sequence,
1472 : tenant_id=%result.tenant_shard_id.tenant_id,
1473 : shard_id=%result.tenant_shard_id.shard_slug(),
1474 : ))]
1475 : fn process_result(&self, result: ReconcileResult) {
1476 : let mut locked = self.inner.write().unwrap();
1477 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1478 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1479 : // A reconciliation result might race with removing a tenant: drop results for
1480 : // tenants that aren't in our map.
1481 : return;
1482 : };
1483 :
1484 : // Usually generation should only be updated via this path, so the max() isn't
1485 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1486 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1487 :
1488 : // If the reconciler signals that it failed to notify compute, set this state on
1489 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1490 : tenant.pending_compute_notification = result.pending_compute_notification;
1491 :
1492 : // Let the TenantShard know it is idle.
1493 : tenant.reconcile_complete(result.sequence);
1494 :
1495 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1496 : // make to the tenant
1497 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1498 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1499 : // make to the tenant
1500 0 : let node = nodes.get(delta.node_id())?;
1501 :
1502 0 : if node.is_available() {
1503 0 : return Some(delta);
1504 0 : }
1505 :
1506 : // In case a node became unavailable concurrently with the reconcile, observed
1507 : // locations on it are now uncertain. By convention, set them to None in order
1508 : // for them to get refreshed when the node comes back online.
1509 0 : Some(ObservedStateDelta::Upsert(Box::new((
1510 0 : node.get_id(),
1511 0 : ObservedStateLocation { conf: None },
1512 0 : ))))
1513 0 : });
1514 :
1515 : match result.result {
1516 : Ok(()) => {
1517 : tenant.apply_observed_deltas(deltas);
1518 : tenant.waiter.advance(result.sequence);
1519 : }
1520 : Err(e) => {
1521 : match e {
1522 : ReconcileError::Cancel => {
1523 : tracing::info!("Reconciler was cancelled");
1524 : }
1525 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1526 : // This might be due to the reconciler getting cancelled, or it might
1527 : // be due to the `Node` being marked offline.
1528 : tracing::info!("Reconciler cancelled during pageserver API call");
1529 : }
1530 : _ => {
1531 : tracing::warn!("Reconcile error: {}", e);
1532 : }
1533 : }
1534 :
1535 : // Ordering: populate last_error before advancing error_seq,
1536 : // so that waiters will see the correct error after waiting.
1537 : tenant.set_last_error(result.sequence, e);
1538 :
1539 : // If the reconciliation failed, don't clear the observed state for places where we
1540 : // detached. Instead, mark the observed state as uncertain.
1541 0 : let failed_reconcile_deltas = deltas.map(|delta| {
1542 0 : if let ObservedStateDelta::Delete(node_id) = delta {
1543 0 : ObservedStateDelta::Upsert(Box::new((
1544 0 : node_id,
1545 0 : ObservedStateLocation { conf: None },
1546 0 : )))
1547 : } else {
1548 0 : delta
1549 : }
1550 0 : });
1551 : tenant.apply_observed_deltas(failed_reconcile_deltas);
1552 : }
1553 : }
1554 :
1555 : tenant.consecutive_reconciles_count = tenant.consecutive_reconciles_count.saturating_add(1);
1556 :
1557 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1558 : if tenant.policy == PlacementPolicy::Detached {
1559 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1560 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1561 : // duration of the request.
1562 : let guard = self.tenant_op_locks.try_exclusive(
1563 : tenant.tenant_shard_id.tenant_id,
1564 : TenantOperations::DropDetached,
1565 : );
1566 : if let Some(guard) = guard {
1567 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1568 : }
1569 : }
1570 :
1571 : // Maybe some other work can proceed now that this job finished.
1572 : //
1573 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1574 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1575 : if self.reconciler_concurrency.available_permits() > 0 {
1576 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1577 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1578 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1579 : shard.delayed_reconcile = false;
1580 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1581 : }
1582 :
1583 : if self.reconciler_concurrency.available_permits() == 0 {
1584 : break;
1585 : }
1586 : }
1587 : }
1588 : }
1589 :
1590 0 : async fn process_results(
1591 0 : &self,
1592 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1593 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1594 0 : Result<(), (TenantShardId, NotifyError)>,
1595 0 : >,
1596 0 : ) {
1597 : loop {
1598 : // Wait for the next result, or for cancellation
1599 0 : tokio::select! {
1600 0 : r = result_rx.recv() => {
1601 0 : match r {
1602 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1603 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1604 : }
1605 : }
1606 0 : _ = async{
1607 0 : match bg_compute_hook_result_rx.recv().await {
1608 0 : Some(result) => {
1609 0 : if let Err((tenant_shard_id, notify_error)) = result {
1610 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1611 0 : let mut locked = self.inner.write().unwrap();
1612 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1613 0 : shard.pending_compute_notification = true;
1614 0 : }
1615 :
1616 0 : }
1617 : },
1618 : None => {
1619 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1620 0 : self.cancel.cancelled().await;
1621 : }
1622 : }
1623 0 : } => {},
1624 0 : _ = self.cancel.cancelled() => {
1625 0 : break;
1626 : }
1627 : };
1628 : }
1629 0 : }
1630 :
1631 0 : async fn process_aborts(
1632 0 : &self,
1633 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1634 0 : ) {
1635 : loop {
1636 : // Wait for the next result, or for cancellation
1637 0 : let op = tokio::select! {
1638 0 : r = abort_rx.recv() => {
1639 0 : match r {
1640 0 : Some(op) => {op},
1641 0 : None => {break;}
1642 : }
1643 : }
1644 0 : _ = self.cancel.cancelled() => {
1645 0 : break;
1646 : }
1647 : };
1648 :
1649 : // Retry until shutdown: we must keep this request object alive until it is properly
1650 : // processed, as it holds a lock guard that prevents other operations trying to do things
1651 : // to the tenant while it is in a weird part-split state.
1652 0 : while !self.reconcilers_cancel.is_cancelled() {
1653 0 : match self.abort_tenant_shard_split(&op).await {
1654 0 : Ok(_) => break,
1655 0 : Err(e) => {
1656 0 : tracing::warn!(
1657 0 : "Failed to abort shard split on {}, will retry: {e}",
1658 : op.tenant_id
1659 : );
1660 :
1661 : // If a node is unavailable, we hope that it has been properly marked Offline
1662 : // when we retry, so that the abort op will succeed. If the abort op is failing
1663 : // for some other reason, we will keep retrying forever, or until a human notices
1664 : // and does something about it (either fixing a pageserver or restarting the controller).
1665 0 : tokio::time::timeout(
1666 0 : Duration::from_secs(5),
1667 0 : self.reconcilers_cancel.cancelled(),
1668 0 : )
1669 0 : .await
1670 0 : .ok();
1671 : }
1672 : }
1673 : }
1674 : }
1675 0 : }
1676 :
1677 0 : pub async fn spawn(
1678 0 : config: Config,
1679 0 : persistence: Arc<Persistence>,
1680 0 : token_generator: Option<HadronTokenGenerator>,
1681 0 : ) -> anyhow::Result<Arc<Self>> {
1682 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1683 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1684 :
1685 0 : let leadership_cancel = CancellationToken::new();
1686 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1687 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1688 :
1689 : // Apply the migrations **after** the current leader has stepped down
1690 : // (or we've given up waiting for it), but **before** reading from the
1691 : // database. The only exception is reading the current leader before
1692 : // migrating.
1693 0 : persistence.migration_run().await?;
1694 :
1695 0 : tracing::info!("Loading nodes from database...");
1696 0 : let nodes = persistence
1697 0 : .list_nodes()
1698 0 : .await?
1699 0 : .into_iter()
1700 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1701 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1702 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1703 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1704 0 : metrics::METRICS_REGISTRY
1705 0 : .metrics_group
1706 0 : .storage_controller_pageserver_nodes
1707 0 : .set(nodes.len() as i64);
1708 0 : metrics::METRICS_REGISTRY
1709 0 : .metrics_group
1710 0 : .storage_controller_https_pageserver_nodes
1711 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
1712 :
1713 0 : tracing::info!("Loading safekeepers from database...");
1714 0 : let safekeepers = persistence
1715 0 : .list_safekeepers()
1716 0 : .await?
1717 0 : .into_iter()
1718 0 : .map(|skp| {
1719 0 : Safekeeper::from_persistence(
1720 0 : skp,
1721 0 : CancellationToken::new(),
1722 0 : config.use_https_safekeeper_api,
1723 : )
1724 0 : })
1725 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1726 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1727 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1728 0 : let count_policy = |policy| {
1729 0 : safekeepers
1730 0 : .iter()
1731 0 : .filter(|sk| sk.1.scheduling_policy() == policy)
1732 0 : .count()
1733 0 : };
1734 0 : let active_sk_count = count_policy(SkSchedulingPolicy::Active);
1735 0 : let activating_sk_count = count_policy(SkSchedulingPolicy::Activating);
1736 0 : let pause_sk_count = count_policy(SkSchedulingPolicy::Pause);
1737 0 : let decom_sk_count = count_policy(SkSchedulingPolicy::Decomissioned);
1738 0 : tracing::info!(
1739 0 : "Loaded {} safekeepers from database. Active {active_sk_count}, activating {activating_sk_count}, \
1740 0 : paused {pause_sk_count}, decomissioned {decom_sk_count}.",
1741 0 : safekeepers.len()
1742 : );
1743 0 : metrics::METRICS_REGISTRY
1744 0 : .metrics_group
1745 0 : .storage_controller_safekeeper_nodes
1746 0 : .set(safekeepers.len() as i64);
1747 0 : metrics::METRICS_REGISTRY
1748 0 : .metrics_group
1749 0 : .storage_controller_https_safekeeper_nodes
1750 0 : .set(safekeepers.values().filter(|s| s.has_https_port()).count() as i64);
1751 :
1752 0 : tracing::info!("Loading shards from database...");
1753 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1754 0 : tracing::info!(
1755 0 : "Loaded {} shards from database.",
1756 0 : tenant_shard_persistence.len()
1757 : );
1758 :
1759 : // If any shard splits were in progress, reset the database state to abort them
1760 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1761 0 : HashMap::new();
1762 0 : for tsp in &mut tenant_shard_persistence {
1763 0 : let shard = tsp.get_shard_identity()?;
1764 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1765 0 : let entry = tenant_shard_count_min_max
1766 0 : .entry(tenant_shard_id.tenant_id)
1767 0 : .or_insert_with(|| (shard.count, shard.count));
1768 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1769 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1770 : }
1771 :
1772 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1773 0 : if count_min != count_max {
1774 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1775 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1776 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1777 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1778 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1779 :
1780 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1781 : // identified this tenant has having mismatching min/max counts.
1782 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1783 :
1784 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1785 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1786 : // Set idle split state on those shards that we will retain.
1787 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1788 0 : if tsp_tenant_id == tenant_id
1789 0 : && tsp.get_shard_identity().unwrap().count == count_min
1790 0 : {
1791 0 : tsp.splitting = SplitState::Idle;
1792 0 : } else if tsp_tenant_id == tenant_id {
1793 : // Leave the splitting state on the child shards: this will be used next to
1794 : // drop them.
1795 0 : tracing::info!(
1796 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1797 : );
1798 0 : }
1799 0 : });
1800 :
1801 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1802 0 : tenant_shard_persistence.retain(|tsp| {
1803 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1804 0 : || tsp.splitting == SplitState::Idle
1805 0 : });
1806 0 : }
1807 : }
1808 :
1809 0 : let mut tenants = BTreeMap::new();
1810 :
1811 0 : let mut scheduler = Scheduler::new(nodes.values());
1812 :
1813 : #[cfg(feature = "testing")]
1814 : {
1815 : use pageserver_api::controller_api::AvailabilityZone;
1816 :
1817 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1818 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1819 : // after when pageservers start up and register.
1820 0 : let mut node_ids = HashSet::new();
1821 0 : for tsp in &tenant_shard_persistence {
1822 0 : if let Some(node_id) = tsp.generation_pageserver {
1823 0 : node_ids.insert(node_id);
1824 0 : }
1825 : }
1826 0 : for node_id in node_ids {
1827 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1828 0 : let node = Node::new(
1829 0 : NodeId(node_id as u64),
1830 0 : "".to_string(),
1831 : 123,
1832 0 : None,
1833 0 : "".to_string(),
1834 : 123,
1835 0 : None,
1836 0 : None,
1837 0 : AvailabilityZone("test_az".to_string()),
1838 : false,
1839 : )
1840 0 : .unwrap();
1841 :
1842 0 : scheduler.node_upsert(&node);
1843 : }
1844 : }
1845 0 : for tsp in tenant_shard_persistence {
1846 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1847 :
1848 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1849 : // it with what we can infer: the node for which a generation was most recently issued.
1850 0 : let mut intent = IntentState::new(
1851 0 : tsp.preferred_az_id
1852 0 : .as_ref()
1853 0 : .map(|az| AvailabilityZone(az.clone())),
1854 : );
1855 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1856 : {
1857 0 : if nodes.contains_key(&generation_pageserver) {
1858 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1859 0 : } else {
1860 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1861 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1862 : // on different pageservers.
1863 0 : tracing::warn!(
1864 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1865 : );
1866 : }
1867 0 : }
1868 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1869 :
1870 0 : tenants.insert(tenant_shard_id, new_tenant);
1871 : }
1872 :
1873 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1874 :
1875 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1876 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1877 0 : tokio::sync::mpsc::channel(512);
1878 :
1879 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1880 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1881 :
1882 0 : let cancel = CancellationToken::new();
1883 0 : let reconcilers_cancel = cancel.child_token();
1884 :
1885 0 : let mut http_client = reqwest::Client::builder();
1886 : // We intentionally disable the connection pool, so every request will create its own TCP connection.
1887 : // It's especially important for heartbeaters to notice more network problems.
1888 : //
1889 : // TODO: It makes sense to use this client only in heartbeaters and create a second one with
1890 : // connection pooling for everything else. But reqwest::Client may create a connection without
1891 : // ever using it (it uses hyper's Client under the hood):
1892 : // https://github.com/hyperium/hyper-util/blob/d51318df3461d40e5f5e5ca163cb3905ac960209/src/client/legacy/client.rs#L415
1893 : //
1894 : // Because of a bug in hyper0::Connection::graceful_shutdown such connections hang during
1895 : // graceful server shutdown: https://github.com/hyperium/hyper/issues/2730
1896 : //
1897 : // The bug has been fixed in hyper v1, so keep alive may be enabled only after we migrate to hyper1.
1898 0 : http_client = http_client.pool_max_idle_per_host(0);
1899 0 : for ssl_ca_cert in &config.ssl_ca_certs {
1900 0 : http_client = http_client.add_root_certificate(ssl_ca_cert.clone());
1901 0 : }
1902 0 : let http_client = http_client.build()?;
1903 :
1904 0 : let heartbeater_ps = Heartbeater::new(
1905 0 : http_client.clone(),
1906 0 : config.pageserver_jwt_token.clone(),
1907 0 : config.max_offline_interval,
1908 0 : config.max_warming_up_interval,
1909 0 : cancel.clone(),
1910 : );
1911 :
1912 0 : let heartbeater_sk = Heartbeater::new(
1913 0 : http_client.clone(),
1914 0 : config.safekeeper_jwt_token.clone(),
1915 0 : config.max_offline_interval,
1916 0 : config.max_warming_up_interval,
1917 0 : cancel.clone(),
1918 : );
1919 :
1920 0 : let initial_leadership_status = if config.start_as_candidate {
1921 0 : LeadershipStatus::Candidate
1922 : } else {
1923 0 : LeadershipStatus::Leader
1924 : };
1925 :
1926 0 : let this = Arc::new(Self {
1927 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1928 0 : nodes,
1929 0 : safekeepers,
1930 0 : tenants,
1931 0 : scheduler,
1932 0 : delayed_reconcile_rx,
1933 0 : initial_leadership_status,
1934 0 : reconcilers_cancel.clone(),
1935 : ))),
1936 0 : config: config.clone(),
1937 0 : persistence,
1938 0 : token_generator,
1939 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())?),
1940 0 : result_tx,
1941 0 : heartbeater_ps,
1942 0 : heartbeater_sk,
1943 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1944 0 : config.reconciler_concurrency,
1945 : )),
1946 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1947 0 : config.priority_reconciler_concurrency,
1948 : )),
1949 0 : delayed_reconcile_tx,
1950 0 : abort_tx,
1951 0 : startup_complete: startup_complete.clone(),
1952 0 : cancel,
1953 0 : reconcilers_cancel,
1954 0 : gate: Gate::default(),
1955 0 : reconcilers_gate: Gate::default(),
1956 0 : tenant_op_locks: Default::default(),
1957 0 : node_op_locks: Default::default(),
1958 0 : http_client,
1959 0 : step_down_barrier: Default::default(),
1960 : });
1961 :
1962 0 : let result_task_this = this.clone();
1963 0 : tokio::task::spawn(async move {
1964 : // Block shutdown until we're done (we must respect self.cancel)
1965 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1966 0 : result_task_this
1967 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1968 0 : .await
1969 0 : }
1970 0 : });
1971 :
1972 0 : tokio::task::spawn({
1973 0 : let this = this.clone();
1974 0 : async move {
1975 : // Block shutdown until we're done (we must respect self.cancel)
1976 0 : if let Ok(_gate) = this.gate.enter() {
1977 0 : this.process_aborts(abort_rx).await
1978 0 : }
1979 0 : }
1980 : });
1981 :
1982 0 : tokio::task::spawn({
1983 0 : let this = this.clone();
1984 0 : async move {
1985 0 : if let Ok(_gate) = this.gate.enter() {
1986 : loop {
1987 0 : tokio::select! {
1988 0 : _ = this.cancel.cancelled() => {
1989 0 : break;
1990 : },
1991 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1992 : };
1993 0 : this.tenant_op_locks.housekeeping();
1994 : }
1995 0 : }
1996 0 : }
1997 : });
1998 :
1999 0 : tokio::task::spawn({
2000 0 : let this = this.clone();
2001 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
2002 : // is done.
2003 0 : let startup_completion = startup_completion.clone();
2004 0 : async move {
2005 : // Block shutdown until we're done (we must respect self.cancel)
2006 0 : let Ok(_gate) = this.gate.enter() else {
2007 0 : return;
2008 : };
2009 :
2010 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
2011 0 : .await;
2012 :
2013 0 : drop(startup_completion);
2014 0 : }
2015 : });
2016 :
2017 0 : tokio::task::spawn({
2018 0 : let this = this.clone();
2019 0 : let startup_complete = startup_complete.clone();
2020 0 : async move {
2021 0 : startup_complete.wait().await;
2022 0 : this.background_reconcile().await;
2023 0 : }
2024 : });
2025 :
2026 0 : tokio::task::spawn({
2027 0 : let this = this.clone();
2028 0 : let startup_complete = startup_complete.clone();
2029 0 : async move {
2030 0 : startup_complete.wait().await;
2031 0 : this.spawn_heartbeat_driver().await;
2032 0 : }
2033 : });
2034 :
2035 : // Check that there is enough safekeepers configured that we can create new timelines
2036 0 : let test_sk_res_str = match this.safekeepers_for_new_timeline().await {
2037 0 : Ok(v) => format!("Ok({v:?})"),
2038 0 : Err(v) => format!("Err({v:})"),
2039 : };
2040 0 : tracing::info!(
2041 : timeline_safekeeper_count = config.timeline_safekeeper_count,
2042 : timelines_onto_safekeepers = config.timelines_onto_safekeepers,
2043 0 : "viability test result (test timeline creation on safekeepers): {test_sk_res_str}",
2044 : );
2045 :
2046 0 : Ok(this)
2047 0 : }
2048 :
2049 0 : pub(crate) async fn attach_hook(
2050 0 : &self,
2051 0 : attach_req: AttachHookRequest,
2052 0 : ) -> anyhow::Result<AttachHookResponse> {
2053 0 : let _tenant_lock = trace_exclusive_lock(
2054 0 : &self.tenant_op_locks,
2055 0 : attach_req.tenant_shard_id.tenant_id,
2056 0 : TenantOperations::AttachHook,
2057 0 : )
2058 0 : .await;
2059 :
2060 : // This is a test hook. To enable using it on tenants that were created directly with
2061 : // the pageserver API (not via this service), we will auto-create any missing tenant
2062 : // shards with default state.
2063 0 : let insert = {
2064 0 : match self
2065 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
2066 0 : .await
2067 : {
2068 0 : Ok(_) => false,
2069 0 : Err(ApiError::NotFound(_)) => true,
2070 0 : Err(e) => return Err(e.into()),
2071 : }
2072 : };
2073 :
2074 0 : if insert {
2075 0 : let config = attach_req.config.clone().unwrap_or_default();
2076 0 : let tsp = TenantShardPersistence {
2077 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
2078 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
2079 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
2080 0 : shard_stripe_size: 0,
2081 0 : generation: attach_req.generation_override.or(Some(0)),
2082 0 : generation_pageserver: None,
2083 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
2084 0 : config: serde_json::to_string(&config).unwrap(),
2085 0 : splitting: SplitState::default(),
2086 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2087 0 : .unwrap(),
2088 0 : preferred_az_id: None,
2089 0 : };
2090 :
2091 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
2092 0 : Err(e) => match e {
2093 : DatabaseError::Query(diesel::result::Error::DatabaseError(
2094 : DatabaseErrorKind::UniqueViolation,
2095 : _,
2096 : )) => {
2097 0 : tracing::info!(
2098 0 : "Raced with another request to insert tenant {}",
2099 : attach_req.tenant_shard_id
2100 : )
2101 : }
2102 0 : _ => return Err(e.into()),
2103 : },
2104 : Ok(()) => {
2105 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
2106 :
2107 0 : let mut shard = TenantShard::new(
2108 0 : attach_req.tenant_shard_id,
2109 0 : ShardIdentity::unsharded(),
2110 0 : PlacementPolicy::Attached(0),
2111 0 : None,
2112 : );
2113 0 : shard.config = config;
2114 :
2115 0 : let mut locked = self.inner.write().unwrap();
2116 0 : locked.tenants.insert(attach_req.tenant_shard_id, shard);
2117 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
2118 : }
2119 : }
2120 0 : }
2121 :
2122 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
2123 0 : let maybe_tenant_conf = {
2124 0 : let locked = self.inner.write().unwrap();
2125 0 : locked
2126 0 : .tenants
2127 0 : .get(&attach_req.tenant_shard_id)
2128 0 : .map(|t| t.config.clone())
2129 : };
2130 :
2131 0 : match maybe_tenant_conf {
2132 0 : Some(conf) => {
2133 0 : let new_generation = self
2134 0 : .persistence
2135 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
2136 0 : .await?;
2137 :
2138 : // Persist the placement policy update. This is required
2139 : // when we reattaching a detached tenant.
2140 0 : self.persistence
2141 0 : .update_tenant_shard(
2142 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
2143 0 : Some(PlacementPolicy::Attached(0)),
2144 0 : Some(conf),
2145 0 : None,
2146 0 : None,
2147 0 : )
2148 0 : .await?;
2149 0 : Some(new_generation)
2150 : }
2151 : None => {
2152 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
2153 : }
2154 : }
2155 : } else {
2156 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
2157 0 : None
2158 : };
2159 :
2160 0 : let mut locked = self.inner.write().unwrap();
2161 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2162 :
2163 0 : let tenant_shard = tenants
2164 0 : .get_mut(&attach_req.tenant_shard_id)
2165 0 : .expect("Checked for existence above");
2166 :
2167 0 : if let Some(new_generation) = new_generation {
2168 0 : tenant_shard.generation = Some(new_generation);
2169 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
2170 0 : } else {
2171 : // This is a detach notification. We must update placement policy to avoid re-attaching
2172 : // during background scheduling/reconciliation, or during storage controller restart.
2173 0 : assert!(attach_req.node_id.is_none());
2174 0 : tenant_shard.policy = PlacementPolicy::Detached;
2175 : }
2176 :
2177 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
2178 0 : tracing::info!(
2179 : tenant_id = %attach_req.tenant_shard_id,
2180 : ps_id = %attaching_pageserver,
2181 : generation = ?tenant_shard.generation,
2182 0 : "issuing",
2183 : );
2184 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
2185 0 : tracing::info!(
2186 : tenant_id = %attach_req.tenant_shard_id,
2187 : %ps_id,
2188 : generation = ?tenant_shard.generation,
2189 0 : "dropping",
2190 : );
2191 : } else {
2192 0 : tracing::info!(
2193 : tenant_id = %attach_req.tenant_shard_id,
2194 0 : "no-op: tenant already has no pageserver");
2195 : }
2196 0 : tenant_shard
2197 0 : .intent
2198 0 : .set_attached(scheduler, attach_req.node_id);
2199 :
2200 0 : tracing::info!(
2201 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}, config {:?}",
2202 : attach_req.tenant_shard_id,
2203 : tenant_shard.generation,
2204 : // TODO: this is an odd number of 0xf's
2205 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff)),
2206 : attach_req.config,
2207 : );
2208 :
2209 : // Trick the reconciler into not doing anything for this tenant: this helps
2210 : // tests that manually configure a tenant on the pagesrever, and then call this
2211 : // attach hook: they don't want background reconciliation to modify what they
2212 : // did to the pageserver.
2213 : #[cfg(feature = "testing")]
2214 : {
2215 0 : if let Some(node_id) = attach_req.node_id {
2216 0 : tenant_shard.observed.locations = HashMap::from([(
2217 0 : node_id,
2218 0 : ObservedStateLocation {
2219 0 : conf: Some(attached_location_conf(
2220 0 : tenant_shard.generation.unwrap(),
2221 0 : &tenant_shard.shard,
2222 0 : &tenant_shard.config,
2223 0 : &PlacementPolicy::Attached(0),
2224 0 : tenant_shard.intent.get_secondary().len(),
2225 0 : )),
2226 0 : },
2227 0 : )]);
2228 0 : } else {
2229 0 : tenant_shard.observed.locations.clear();
2230 0 : }
2231 : }
2232 :
2233 : Ok(AttachHookResponse {
2234 0 : generation: attach_req
2235 0 : .node_id
2236 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
2237 : })
2238 0 : }
2239 :
2240 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
2241 0 : let locked = self.inner.read().unwrap();
2242 :
2243 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
2244 :
2245 : InspectResponse {
2246 0 : attachment: tenant_shard.and_then(|s| {
2247 0 : s.intent
2248 0 : .get_attached()
2249 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
2250 0 : }),
2251 : }
2252 0 : }
2253 :
2254 : // When the availability state of a node transitions to active, we must do a full reconciliation
2255 : // of LocationConfigs on that node. This is because while a node was offline:
2256 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
2257 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
2258 : //
2259 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
2260 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
2261 : // this function.
2262 : //
2263 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
2264 : // for written for a single node rather than as a batch job for all nodes.
2265 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
2266 : async fn node_activate_reconcile(
2267 : &self,
2268 : mut node: Node,
2269 : _lock: &TracingExclusiveGuard<NodeOperations>,
2270 : ) -> Result<(), ApiError> {
2271 : // This Node is a mutable local copy: we will set it active so that we can use its
2272 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
2273 : // later.
2274 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
2275 :
2276 : let configs = match node
2277 : .with_client_retries(
2278 0 : |client| async move { client.list_location_config().await },
2279 : &self.http_client,
2280 : &self.config.pageserver_jwt_token,
2281 : 1,
2282 : 5,
2283 : SHORT_RECONCILE_TIMEOUT,
2284 : &self.cancel,
2285 : )
2286 : .await
2287 : {
2288 : None => {
2289 : // We're shutting down (the Node's cancellation token can't have fired, because
2290 : // we're the only scope that has a reference to it, and we didn't fire it).
2291 : return Err(ApiError::ShuttingDown);
2292 : }
2293 : Some(Err(e)) => {
2294 : // This node didn't succeed listing its locations: it may not proceed to active state
2295 : // as it is apparently unavailable.
2296 : return Err(ApiError::PreconditionFailed(
2297 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2298 : ));
2299 : }
2300 : Some(Ok(configs)) => configs,
2301 : };
2302 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2303 :
2304 : let mut cleanup = Vec::new();
2305 : let mut mismatched_locations = 0;
2306 : {
2307 : let mut locked = self.inner.write().unwrap();
2308 :
2309 : for (tenant_shard_id, reported) in configs.tenant_shards {
2310 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2311 : cleanup.push(tenant_shard_id);
2312 : continue;
2313 : };
2314 :
2315 : let on_record = &mut tenant_shard
2316 : .observed
2317 : .locations
2318 : .entry(node.get_id())
2319 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2320 : .conf;
2321 :
2322 : // If the location reported by the node does not match our observed state,
2323 : // then we mark it as uncertain and let the background reconciliation loop
2324 : // deal with it.
2325 : //
2326 : // Note that this also covers net new locations reported by the node.
2327 : if *on_record != reported {
2328 : mismatched_locations += 1;
2329 : *on_record = None;
2330 : }
2331 : }
2332 : }
2333 :
2334 : if mismatched_locations > 0 {
2335 : tracing::info!(
2336 : "Set observed state to None for {mismatched_locations} mismatched locations"
2337 : );
2338 : }
2339 :
2340 : for tenant_shard_id in cleanup {
2341 : tracing::info!("Detaching {tenant_shard_id}");
2342 : match node
2343 : .with_client_retries(
2344 0 : |client| async move {
2345 0 : let config = LocationConfig {
2346 0 : mode: LocationConfigMode::Detached,
2347 0 : generation: None,
2348 0 : secondary_conf: None,
2349 0 : shard_number: tenant_shard_id.shard_number.0,
2350 0 : shard_count: tenant_shard_id.shard_count.literal(),
2351 0 : shard_stripe_size: 0,
2352 0 : tenant_conf: models::TenantConfig::default(),
2353 0 : };
2354 0 : client
2355 0 : .location_config(tenant_shard_id, config, None, false)
2356 0 : .await
2357 0 : },
2358 : &self.http_client,
2359 : &self.config.pageserver_jwt_token,
2360 : 1,
2361 : 5,
2362 : SHORT_RECONCILE_TIMEOUT,
2363 : &self.cancel,
2364 : )
2365 : .await
2366 : {
2367 : None => {
2368 : // We're shutting down (the Node's cancellation token can't have fired, because
2369 : // we're the only scope that has a reference to it, and we didn't fire it).
2370 : return Err(ApiError::ShuttingDown);
2371 : }
2372 : Some(Err(e)) => {
2373 : // Do not let the node proceed to Active state if it is not responsive to requests
2374 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2375 : // detach completing: we should not let this node back into the set of nodes considered
2376 : // okay for scheduling.
2377 : return Err(ApiError::Conflict(format!(
2378 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2379 : )));
2380 : }
2381 : Some(Ok(_)) => {}
2382 : };
2383 : }
2384 :
2385 : Ok(())
2386 : }
2387 :
2388 0 : pub(crate) async fn re_attach(
2389 0 : &self,
2390 0 : reattach_req: ReAttachRequest,
2391 0 : ) -> Result<ReAttachResponse, ApiError> {
2392 0 : if let Some(register_req) = reattach_req.register {
2393 0 : self.node_register(register_req).await?;
2394 0 : }
2395 :
2396 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2397 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2398 :
2399 0 : tracing::info!(
2400 : node_id=%reattach_req.node_id,
2401 0 : "Incremented {} tenant shards' generations",
2402 0 : incremented_generations.len()
2403 : );
2404 :
2405 : // Apply the updated generation to our in-memory state, and
2406 : // gather discover secondary locations.
2407 0 : let mut locked = self.inner.write().unwrap();
2408 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2409 :
2410 0 : let mut response = ReAttachResponse {
2411 0 : tenants: Vec::new(),
2412 0 : };
2413 :
2414 : // [Hadron] If the pageserver reports in the reattach message that it has an empty disk, it's possible that it just
2415 : // recovered from a local disk failure. The response of the reattach request will contain a list of tenants but it
2416 : // will not be honored by the pageserver in this case (disk failure). We should make sure we clear any observed
2417 : // locations of tenants attached to the node so that the reconciler will discover the discrpancy and reconfigure the
2418 : // missing tenants on the node properly.
2419 0 : if self.config.handle_ps_local_disk_loss && reattach_req.empty_local_disk.unwrap_or(false) {
2420 0 : tracing::info!(
2421 0 : "Pageserver {node_id} reports empty local disk, clearing observed locations referencing the pageserver for all tenants",
2422 : node_id = reattach_req.node_id
2423 : );
2424 0 : let mut num_tenant_shards_affected = 0;
2425 0 : for (tenant_shard_id, shard) in tenants.iter_mut() {
2426 0 : if shard
2427 0 : .observed
2428 0 : .locations
2429 0 : .remove(&reattach_req.node_id)
2430 0 : .is_some()
2431 : {
2432 0 : tracing::info!("Cleared observed location for tenant shard {tenant_shard_id}");
2433 0 : num_tenant_shards_affected += 1;
2434 0 : }
2435 : }
2436 0 : tracing::info!(
2437 0 : "Cleared observed locations for {num_tenant_shards_affected} tenant shards"
2438 : );
2439 0 : }
2440 :
2441 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2442 : // to call location_conf API with an old generation. Wait for cancellation to complete
2443 : // before responding to this request. Requires well implemented CancellationToken logic
2444 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2445 : // request in flight over the network: TODO handle that by making location_conf API refuse
2446 : // to go backward in generations.
2447 :
2448 : // Scan through all shards, applying updates for ones where we updated generation
2449 : // and identifying shards that intend to have a secondary location on this node.
2450 0 : for (tenant_shard_id, shard) in tenants {
2451 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2452 0 : let new_gen = *new_gen;
2453 0 : response.tenants.push(ReAttachResponseTenant {
2454 0 : id: *tenant_shard_id,
2455 0 : r#gen: Some(new_gen.into().unwrap()),
2456 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2457 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2458 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2459 0 : // the stale/multi states at this point.
2460 0 : mode: LocationConfigMode::AttachedSingle,
2461 0 : stripe_size: shard.shard.stripe_size,
2462 0 : });
2463 :
2464 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2465 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2466 : // Why can we update `observed` even though we're not sure our response will be received
2467 : // by the pageserver? Because the pageserver will not proceed with startup until
2468 : // it has processed response: if it loses it, we'll see another request and increment
2469 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2470 0 : if let Some(conf) = observed.conf.as_mut() {
2471 0 : conf.generation = new_gen.into();
2472 0 : }
2473 0 : } else {
2474 0 : // This node has no observed state for the shard: perhaps it was offline
2475 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2476 0 : // will be prompted to learn the location's state before it makes changes.
2477 0 : shard
2478 0 : .observed
2479 0 : .locations
2480 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2481 0 : }
2482 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2483 0 : // Ordering: pageserver will not accept /location_config requests until it has
2484 0 : // finished processing the response from re-attach. So we can update our in-memory state
2485 0 : // now, and be confident that we are not stamping on the result of some later location config.
2486 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2487 0 : // so we might update observed state here, and then get over-written by some racing
2488 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2489 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2490 0 :
2491 0 : response.tenants.push(ReAttachResponseTenant {
2492 0 : id: *tenant_shard_id,
2493 0 : r#gen: None,
2494 0 : mode: LocationConfigMode::Secondary,
2495 0 : stripe_size: shard.shard.stripe_size,
2496 0 : });
2497 0 :
2498 0 : // We must not update observed, because we have no guarantee that our
2499 0 : // response will be received by the pageserver. This could leave it
2500 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2501 0 : }
2502 : }
2503 :
2504 : // We consider a node Active once we have composed a re-attach response, but we
2505 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2506 : // implicitly synchronizes the LocationConfigs on the node.
2507 : //
2508 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2509 : // but those requests will not be accepted by the node until it has finished processing
2510 : // the re-attach response.
2511 : //
2512 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2513 : // in [`Persistence::re_attach`].
2514 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2515 0 : let reset_scheduling = matches!(
2516 0 : node.get_scheduling(),
2517 : NodeSchedulingPolicy::PauseForRestart
2518 : | NodeSchedulingPolicy::Draining
2519 : | NodeSchedulingPolicy::Filling
2520 : | NodeSchedulingPolicy::Deleting
2521 : );
2522 :
2523 0 : let mut new_nodes = (**nodes).clone();
2524 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2525 0 : if reset_scheduling {
2526 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2527 0 : }
2528 :
2529 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2530 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2531 :
2532 0 : scheduler.node_upsert(node);
2533 0 : let new_nodes = Arc::new(new_nodes);
2534 0 : *nodes = new_nodes;
2535 : } else {
2536 0 : tracing::error!(
2537 0 : "Reattaching node {} was removed while processing the request",
2538 : reattach_req.node_id
2539 : );
2540 : }
2541 0 : }
2542 :
2543 0 : Ok(response)
2544 0 : }
2545 :
2546 0 : pub(crate) async fn validate(
2547 0 : &self,
2548 0 : validate_req: ValidateRequest,
2549 0 : ) -> Result<ValidateResponse, DatabaseError> {
2550 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2551 : // in-memory generation for a shard
2552 0 : let in_memory_result = {
2553 0 : let mut in_memory_result = Vec::new();
2554 0 : let locked = self.inner.read().unwrap();
2555 0 : for req_tenant in validate_req.tenants {
2556 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2557 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2558 0 : tracing::info!(
2559 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2560 : req_tenant.id,
2561 : req_tenant.r#gen,
2562 : tenant_shard.generation
2563 : );
2564 :
2565 0 : in_memory_result.push((
2566 0 : req_tenant.id,
2567 0 : Generation::new(req_tenant.r#gen),
2568 0 : valid,
2569 0 : ));
2570 : } else {
2571 : // This is legal: for example during a shard split the pageserver may still
2572 : // have deletions in its queue from the old pre-split shard, or after deletion
2573 : // of a tenant that was busy with compaction/gc while being deleted.
2574 0 : tracing::info!(
2575 0 : "Refusing deletion validation for missing shard {}",
2576 : req_tenant.id
2577 : );
2578 : }
2579 : }
2580 :
2581 0 : in_memory_result
2582 : };
2583 :
2584 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2585 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2586 0 : let db_generations = self
2587 0 : .persistence
2588 0 : .shard_generations(
2589 0 : in_memory_result
2590 0 : .iter()
2591 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2592 : )
2593 0 : .await?;
2594 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2595 :
2596 0 : let mut response = ValidateResponse {
2597 0 : tenants: Vec::new(),
2598 0 : };
2599 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2600 0 : let valid = if valid {
2601 0 : let db_generation = db_generations.get(&tenant_shard_id);
2602 0 : db_generation == Some(&Some(validate_generation))
2603 : } else {
2604 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2605 : // this prevents a pageserver from cleaning up an object in S3.
2606 0 : false
2607 : };
2608 :
2609 0 : response.tenants.push(ValidateResponseTenant {
2610 0 : id: tenant_shard_id,
2611 0 : valid,
2612 0 : })
2613 : }
2614 :
2615 0 : Ok(response)
2616 0 : }
2617 :
2618 0 : pub(crate) async fn tenant_create(
2619 0 : &self,
2620 0 : create_req: TenantCreateRequest,
2621 0 : ) -> Result<TenantCreateResponse, ApiError> {
2622 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2623 :
2624 : // Exclude any concurrent attempts to create/access the same tenant ID
2625 0 : let _tenant_lock = trace_exclusive_lock(
2626 0 : &self.tenant_op_locks,
2627 0 : create_req.new_tenant_id.tenant_id,
2628 0 : TenantOperations::Create,
2629 0 : )
2630 0 : .await;
2631 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2632 :
2633 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2634 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2635 : // accept compute notifications while it is in the process of creating. Reconciliation will
2636 : // be retried in the background.
2637 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2638 0 : }
2639 0 : Ok(response)
2640 0 : }
2641 :
2642 0 : pub(crate) async fn do_tenant_create(
2643 0 : &self,
2644 0 : create_req: TenantCreateRequest,
2645 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2646 0 : let placement_policy = create_req
2647 0 : .placement_policy
2648 0 : .clone()
2649 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2650 0 : .unwrap_or(PlacementPolicy::Attached(0));
2651 :
2652 : // This service expects to handle sharding itself: it is an error to try and directly create
2653 : // a particular shard here.
2654 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2655 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2656 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2657 0 : )));
2658 : } else {
2659 0 : create_req.new_tenant_id.tenant_id
2660 : };
2661 :
2662 0 : tracing::info!(
2663 0 : "Creating tenant {}, shard_count={:?}",
2664 : create_req.new_tenant_id,
2665 : create_req.shard_parameters.count,
2666 : );
2667 :
2668 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2669 0 : .map(|i| TenantShardId {
2670 0 : tenant_id,
2671 0 : shard_number: ShardNumber(i),
2672 0 : shard_count: create_req.shard_parameters.count,
2673 0 : })
2674 0 : .collect::<Vec<_>>();
2675 :
2676 : // If the caller specifies a None generation, it means "start from default". This is different
2677 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2678 : // an incompletely-onboarded tenant.
2679 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2680 0 : tracing::info!(
2681 0 : "tenant_create: secondary mode, generation is_some={}",
2682 0 : create_req.generation.is_some()
2683 : );
2684 0 : create_req.generation.map(Generation::new)
2685 : } else {
2686 0 : tracing::info!(
2687 0 : "tenant_create: not secondary mode, generation is_some={}",
2688 0 : create_req.generation.is_some()
2689 : );
2690 0 : Some(
2691 0 : create_req
2692 0 : .generation
2693 0 : .map(Generation::new)
2694 0 : .unwrap_or(INITIAL_GENERATION),
2695 0 : )
2696 : };
2697 :
2698 0 : let preferred_az_id = {
2699 0 : let locked = self.inner.read().unwrap();
2700 : // Idempotency: take the existing value if the tenant already exists
2701 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2702 0 : shard.preferred_az().cloned()
2703 : } else {
2704 0 : locked.scheduler.get_az_for_new_tenant()
2705 : }
2706 : };
2707 :
2708 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2709 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2710 : // during the creation, rather than risking leaving orphan objects in S3.
2711 0 : let persist_tenant_shards = create_ids
2712 0 : .iter()
2713 0 : .map(|tenant_shard_id| TenantShardPersistence {
2714 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2715 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2716 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2717 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2718 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2719 : // The pageserver is not known until scheduling happens: we will set this column when
2720 : // incrementing the generation the first time we attach to a pageserver.
2721 0 : generation_pageserver: None,
2722 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2723 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2724 0 : splitting: SplitState::default(),
2725 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2726 0 : .unwrap(),
2727 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2728 0 : })
2729 0 : .collect();
2730 :
2731 0 : match self
2732 0 : .persistence
2733 0 : .insert_tenant_shards(persist_tenant_shards)
2734 0 : .await
2735 : {
2736 0 : Ok(_) => {}
2737 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2738 : DatabaseErrorKind::UniqueViolation,
2739 : _,
2740 : ))) => {
2741 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2742 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2743 : // creation's shard count.
2744 0 : tracing::info!(
2745 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2746 : );
2747 : }
2748 : // Any other database error is unexpected and a bug.
2749 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2750 : };
2751 :
2752 0 : let mut schedule_context = ScheduleContext::default();
2753 0 : let mut schedule_error = None;
2754 0 : let mut response_shards = Vec::new();
2755 0 : for tenant_shard_id in create_ids {
2756 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2757 :
2758 0 : let outcome = self
2759 0 : .do_initial_shard_scheduling(
2760 0 : tenant_shard_id,
2761 0 : initial_generation,
2762 0 : create_req.shard_parameters,
2763 0 : create_req.config.clone(),
2764 0 : placement_policy.clone(),
2765 0 : preferred_az_id.as_ref(),
2766 0 : &mut schedule_context,
2767 0 : )
2768 0 : .await;
2769 :
2770 0 : match outcome {
2771 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2772 0 : InitialShardScheduleOutcome::NotScheduled => {}
2773 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2774 0 : schedule_error = Some(err);
2775 0 : }
2776 : }
2777 : }
2778 :
2779 : // If we failed to schedule shards, then they are still created in the controller,
2780 : // but we return an error to the requester to avoid a silent failure when someone
2781 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2782 : // are present in the system. We do this here rather than in the above loop, to
2783 : // avoid situations where we only create a subset of shards in the tenant.
2784 0 : if let Some(e) = schedule_error {
2785 0 : return Err(ApiError::Conflict(format!(
2786 0 : "Failed to schedule shard(s): {e}"
2787 0 : )));
2788 0 : }
2789 :
2790 0 : let waiters = {
2791 0 : let mut locked = self.inner.write().unwrap();
2792 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2793 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2794 0 : .tenant_creation_hint(true)
2795 0 : .build();
2796 0 : tenants
2797 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2798 0 : .filter_map(|(_shard_id, shard)| {
2799 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2800 0 : })
2801 0 : .collect::<Vec<_>>()
2802 : };
2803 :
2804 0 : Ok((
2805 0 : TenantCreateResponse {
2806 0 : shards: response_shards,
2807 0 : },
2808 0 : waiters,
2809 0 : ))
2810 0 : }
2811 :
2812 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2813 : /// case of a new tenant and a pre-existing one.
2814 : #[allow(clippy::too_many_arguments)]
2815 0 : async fn do_initial_shard_scheduling(
2816 0 : &self,
2817 0 : tenant_shard_id: TenantShardId,
2818 0 : initial_generation: Option<Generation>,
2819 0 : shard_params: ShardParameters,
2820 0 : config: TenantConfig,
2821 0 : placement_policy: PlacementPolicy,
2822 0 : preferred_az_id: Option<&AvailabilityZone>,
2823 0 : schedule_context: &mut ScheduleContext,
2824 0 : ) -> InitialShardScheduleOutcome {
2825 0 : let mut locked = self.inner.write().unwrap();
2826 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2827 :
2828 : use std::collections::btree_map::Entry;
2829 0 : match tenants.entry(tenant_shard_id) {
2830 0 : Entry::Occupied(mut entry) => {
2831 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2832 :
2833 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2834 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2835 0 : }
2836 :
2837 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2838 0 : let generation = entry
2839 0 : .get()
2840 0 : .generation
2841 0 : .expect("Generation is set when in attached mode");
2842 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2843 0 : shard_id: tenant_shard_id,
2844 0 : node_id: *node_id,
2845 0 : generation: generation.into().unwrap(),
2846 0 : })
2847 : } else {
2848 0 : InitialShardScheduleOutcome::NotScheduled
2849 : }
2850 : }
2851 0 : Entry::Vacant(entry) => {
2852 0 : let state = entry.insert(TenantShard::new(
2853 0 : tenant_shard_id,
2854 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2855 0 : placement_policy,
2856 0 : preferred_az_id.cloned(),
2857 : ));
2858 :
2859 0 : state.generation = initial_generation;
2860 0 : state.config = config;
2861 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2862 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2863 0 : }
2864 :
2865 : // Only include shards in result if we are attaching: the purpose
2866 : // of the response is to tell the caller where the shards are attached.
2867 0 : if let Some(node_id) = state.intent.get_attached() {
2868 0 : let generation = state
2869 0 : .generation
2870 0 : .expect("Generation is set when in attached mode");
2871 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2872 0 : shard_id: tenant_shard_id,
2873 0 : node_id: *node_id,
2874 0 : generation: generation.into().unwrap(),
2875 0 : })
2876 : } else {
2877 0 : InitialShardScheduleOutcome::NotScheduled
2878 : }
2879 : }
2880 : }
2881 0 : }
2882 :
2883 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2884 : /// wait for reconciliation to complete before responding.
2885 0 : async fn await_waiters(
2886 0 : &self,
2887 0 : waiters: Vec<ReconcilerWaiter>,
2888 0 : timeout: Duration,
2889 0 : ) -> Result<(), ReconcileWaitError> {
2890 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2891 0 : for waiter in waiters {
2892 0 : let timeout = deadline.duration_since(Instant::now());
2893 0 : waiter.wait_timeout(timeout).await?;
2894 : }
2895 :
2896 0 : Ok(())
2897 0 : }
2898 :
2899 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2900 : /// in progress
2901 0 : async fn await_waiters_remainder(
2902 0 : &self,
2903 0 : waiters: Vec<ReconcilerWaiter>,
2904 0 : timeout: Duration,
2905 0 : ) -> Vec<ReconcilerWaiter> {
2906 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2907 0 : for waiter in waiters.iter() {
2908 0 : let timeout = deadline.duration_since(Instant::now());
2909 0 : let _ = waiter.wait_timeout(timeout).await;
2910 : }
2911 :
2912 0 : waiters
2913 0 : .into_iter()
2914 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2915 0 : .collect::<Vec<_>>()
2916 0 : }
2917 :
2918 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2919 : /// and transform it into either a tenant creation of a series of shard updates.
2920 : ///
2921 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2922 : /// still be returned.
2923 0 : fn tenant_location_config_prepare(
2924 0 : &self,
2925 0 : tenant_id: TenantId,
2926 0 : req: TenantLocationConfigRequest,
2927 0 : ) -> TenantCreateOrUpdate {
2928 0 : let mut updates = Vec::new();
2929 0 : let mut locked = self.inner.write().unwrap();
2930 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2931 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2932 :
2933 : // Use location config mode as an indicator of policy.
2934 0 : let placement_policy = match req.config.mode {
2935 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2936 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2937 : LocationConfigMode::AttachedMulti
2938 : | LocationConfigMode::AttachedSingle
2939 : | LocationConfigMode::AttachedStale => {
2940 0 : if nodes.len() > 1 {
2941 0 : PlacementPolicy::Attached(1)
2942 : } else {
2943 : // Convenience for dev/test: if we just have one pageserver, import
2944 : // tenants into non-HA mode so that scheduling will succeed.
2945 0 : PlacementPolicy::Attached(0)
2946 : }
2947 : }
2948 : };
2949 :
2950 : // Ordinarily we do not update scheduling policy, but when making major changes
2951 : // like detaching or demoting to secondary-only, we need to force the scheduling
2952 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2953 0 : let scheduling_policy = match req.config.mode {
2954 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2955 : // Special case: when making major changes like detaching or demoting to secondary-only,
2956 : // we need to force the scheduling mode to Active, or nothing will happen.
2957 0 : Some(ShardSchedulingPolicy::Active)
2958 : }
2959 : LocationConfigMode::AttachedMulti
2960 : | LocationConfigMode::AttachedSingle
2961 : | LocationConfigMode::AttachedStale => {
2962 : // While attached, continue to respect whatever the existing scheduling mode is.
2963 0 : None
2964 : }
2965 : };
2966 :
2967 0 : let mut create = true;
2968 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2969 : // Saw an existing shard: this is not a creation
2970 0 : create = false;
2971 :
2972 : // Shards may have initially been created by a Secondary request, where we
2973 : // would have left generation as None.
2974 : //
2975 : // We only update generation the first time we see an attached-mode request,
2976 : // and if there is no existing generation set. The caller is responsible for
2977 : // ensuring that no non-storage-controller pageserver ever uses a higher
2978 : // generation than they passed in here.
2979 : use LocationConfigMode::*;
2980 0 : let set_generation = match req.config.mode {
2981 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2982 0 : req.config.generation.map(Generation::new)
2983 : }
2984 0 : _ => None,
2985 : };
2986 :
2987 0 : updates.push(ShardUpdate {
2988 0 : tenant_shard_id: *shard_id,
2989 0 : placement_policy: placement_policy.clone(),
2990 0 : tenant_config: req.config.tenant_conf.clone(),
2991 0 : generation: set_generation,
2992 0 : scheduling_policy,
2993 0 : });
2994 : }
2995 :
2996 0 : if create {
2997 : use LocationConfigMode::*;
2998 0 : let generation = match req.config.mode {
2999 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
3000 : // If a caller provided a generation in a non-attached request, ignore it
3001 : // and leave our generation as None: this enables a subsequent update to set
3002 : // the generation when setting an attached mode for the first time.
3003 0 : _ => None,
3004 : };
3005 :
3006 0 : TenantCreateOrUpdate::Create(
3007 0 : // Synthesize a creation request
3008 0 : TenantCreateRequest {
3009 0 : new_tenant_id: tenant_shard_id,
3010 0 : generation,
3011 0 : shard_parameters: ShardParameters {
3012 0 : count: tenant_shard_id.shard_count,
3013 0 : // We only import un-sharded or single-sharded tenants, so stripe
3014 0 : // size can be made up arbitrarily here.
3015 0 : stripe_size: DEFAULT_STRIPE_SIZE,
3016 0 : },
3017 0 : placement_policy: Some(placement_policy),
3018 0 : config: req.config.tenant_conf,
3019 0 : },
3020 0 : )
3021 : } else {
3022 0 : assert!(!updates.is_empty());
3023 0 : TenantCreateOrUpdate::Update(updates)
3024 : }
3025 0 : }
3026 :
3027 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
3028 : /// the tenant is present in memory. If not, load it from the database. If it is found
3029 : /// in neither location, return a NotFound error.
3030 : ///
3031 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
3032 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
3033 0 : async fn maybe_load_tenant(
3034 0 : &self,
3035 0 : tenant_id: TenantId,
3036 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
3037 0 : ) -> Result<(), ApiError> {
3038 : // Check if the tenant is present in memory, and select an AZ to use when loading
3039 : // if we will load it.
3040 0 : let load_in_az = {
3041 0 : let locked = self.inner.read().unwrap();
3042 0 : let existing = locked
3043 0 : .tenants
3044 0 : .range(TenantShardId::tenant_range(tenant_id))
3045 0 : .next();
3046 :
3047 : // If the tenant is not present in memory, we expect to load it from database,
3048 : // so let's figure out what AZ to load it into while we have self.inner locked.
3049 0 : if existing.is_none() {
3050 0 : locked
3051 0 : .scheduler
3052 0 : .get_az_for_new_tenant()
3053 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
3054 0 : "No AZ with nodes found to load tenant"
3055 0 : )))?
3056 : } else {
3057 : // We already have this tenant in memory
3058 0 : return Ok(());
3059 : }
3060 : };
3061 :
3062 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
3063 0 : if tenant_shards.is_empty() {
3064 0 : return Err(ApiError::NotFound(
3065 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3066 0 : ));
3067 0 : }
3068 :
3069 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
3070 0 : self.persistence
3071 0 : .set_tenant_shard_preferred_azs(
3072 0 : tenant_shards
3073 0 : .iter()
3074 0 : .map(|t| {
3075 0 : (
3076 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
3077 0 : Some(load_in_az.clone()),
3078 0 : )
3079 0 : })
3080 0 : .collect(),
3081 : )
3082 0 : .await?;
3083 :
3084 0 : let mut locked = self.inner.write().unwrap();
3085 0 : tracing::info!(
3086 0 : "Loaded {} shards for tenant {}",
3087 0 : tenant_shards.len(),
3088 : tenant_id
3089 : );
3090 :
3091 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
3092 0 : let intent = IntentState::new(Some(load_in_az.clone()));
3093 0 : let shard =
3094 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
3095 :
3096 : // Sanity check: when loading on-demand, we should always be loaded something Detached
3097 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
3098 0 : if shard.policy != PlacementPolicy::Detached {
3099 0 : tracing::error!(
3100 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
3101 : shard.tenant_shard_id,
3102 : shard.policy
3103 : );
3104 0 : }
3105 :
3106 0 : (shard.tenant_shard_id, shard)
3107 0 : }));
3108 :
3109 0 : Ok(())
3110 0 : }
3111 :
3112 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
3113 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
3114 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
3115 : ///
3116 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
3117 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
3118 0 : fn maybe_drop_tenant(
3119 0 : &self,
3120 0 : tenant_id: TenantId,
3121 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
3122 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
3123 0 : ) {
3124 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
3125 0 : if tenant_shards.all(|(_id, shard)| {
3126 0 : shard.policy == PlacementPolicy::Detached
3127 0 : && shard.reconciler.is_none()
3128 0 : && shard.observed.is_empty()
3129 0 : }) {
3130 0 : let keys = locked
3131 0 : .tenants
3132 0 : .range(TenantShardId::tenant_range(tenant_id))
3133 0 : .map(|(id, _)| id)
3134 0 : .copied()
3135 0 : .collect::<Vec<_>>();
3136 0 : for key in keys {
3137 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
3138 0 : locked.tenants.remove(&key);
3139 : }
3140 0 : }
3141 0 : }
3142 :
3143 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
3144 : /// directly with pageservers into this service.
3145 : ///
3146 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
3147 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
3148 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
3149 : /// tenant's source of generation numbers.
3150 : ///
3151 : /// The mode in this request coarse-grained control of tenants:
3152 : /// - Call with mode Attached* to upsert the tenant.
3153 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
3154 : /// to set an existing tenant to PolicyMode::Secondary
3155 : /// - Call with mode Detached to switch to PolicyMode::Detached
3156 0 : pub(crate) async fn tenant_location_config(
3157 0 : &self,
3158 0 : tenant_shard_id: TenantShardId,
3159 0 : req: TenantLocationConfigRequest,
3160 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
3161 : // We require an exclusive lock, because we are updating both persistent and in-memory state
3162 0 : let _tenant_lock = trace_exclusive_lock(
3163 0 : &self.tenant_op_locks,
3164 0 : tenant_shard_id.tenant_id,
3165 0 : TenantOperations::LocationConfig,
3166 0 : )
3167 0 : .await;
3168 :
3169 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
3170 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3171 0 : "This API is for importing single-sharded or unsharded tenants"
3172 0 : )));
3173 : } else {
3174 0 : tenant_shard_id.tenant_id
3175 : };
3176 :
3177 : // In case we are waking up a Detached tenant
3178 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
3179 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
3180 0 : // This is a creation or an update
3181 0 : }
3182 0 : Err(e) => {
3183 0 : return Err(e);
3184 : }
3185 : };
3186 :
3187 : // First check if this is a creation or an update
3188 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
3189 :
3190 0 : let mut result = TenantLocationConfigResponse {
3191 0 : shards: Vec::new(),
3192 0 : stripe_size: None,
3193 0 : };
3194 0 : let waiters = match create_or_update {
3195 0 : TenantCreateOrUpdate::Create(create_req) => {
3196 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
3197 0 : result.shards = create_resp
3198 0 : .shards
3199 0 : .into_iter()
3200 0 : .map(|s| TenantShardLocation {
3201 0 : node_id: s.node_id,
3202 0 : shard_id: s.shard_id,
3203 0 : })
3204 0 : .collect();
3205 0 : waiters
3206 : }
3207 0 : TenantCreateOrUpdate::Update(updates) => {
3208 : // Persist updates
3209 : // Ordering: write to the database before applying changes in-memory, so that
3210 : // we will not appear time-travel backwards on a restart.
3211 :
3212 0 : let mut schedule_context = ScheduleContext::default();
3213 : for ShardUpdate {
3214 0 : tenant_shard_id,
3215 0 : placement_policy,
3216 0 : tenant_config,
3217 0 : generation,
3218 0 : scheduling_policy,
3219 0 : } in &updates
3220 : {
3221 0 : self.persistence
3222 0 : .update_tenant_shard(
3223 0 : TenantFilter::Shard(*tenant_shard_id),
3224 0 : Some(placement_policy.clone()),
3225 0 : Some(tenant_config.clone()),
3226 0 : *generation,
3227 0 : *scheduling_policy,
3228 0 : )
3229 0 : .await?;
3230 : }
3231 :
3232 : // Apply updates in-memory
3233 0 : let mut waiters = Vec::new();
3234 : {
3235 0 : let mut locked = self.inner.write().unwrap();
3236 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3237 :
3238 : for ShardUpdate {
3239 0 : tenant_shard_id,
3240 0 : placement_policy,
3241 0 : tenant_config,
3242 0 : generation: update_generation,
3243 0 : scheduling_policy,
3244 0 : } in updates
3245 : {
3246 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
3247 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
3248 0 : continue;
3249 : };
3250 :
3251 : // Update stripe size
3252 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
3253 0 : result.stripe_size = Some(shard.shard.stripe_size);
3254 0 : }
3255 :
3256 0 : shard.policy = placement_policy;
3257 0 : shard.config = tenant_config;
3258 0 : if let Some(generation) = update_generation {
3259 0 : shard.generation = Some(generation);
3260 0 : }
3261 :
3262 0 : if let Some(scheduling_policy) = scheduling_policy {
3263 0 : shard.set_scheduling_policy(scheduling_policy);
3264 0 : }
3265 :
3266 0 : shard.schedule(scheduler, &mut schedule_context)?;
3267 :
3268 0 : let maybe_waiter =
3269 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3270 0 : if let Some(waiter) = maybe_waiter {
3271 0 : waiters.push(waiter);
3272 0 : }
3273 :
3274 0 : if let Some(node_id) = shard.intent.get_attached() {
3275 0 : result.shards.push(TenantShardLocation {
3276 0 : shard_id: tenant_shard_id,
3277 0 : node_id: *node_id,
3278 0 : })
3279 0 : }
3280 : }
3281 : }
3282 0 : waiters
3283 : }
3284 : };
3285 :
3286 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3287 : // Do not treat a reconcile error as fatal: we have already applied any requested
3288 : // Intent changes, and the reconcile can fail for external reasons like unavailable
3289 : // compute notification API. In these cases, it is important that we do not
3290 : // cause the cloud control plane to retry forever on this API.
3291 0 : tracing::warn!(
3292 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
3293 : );
3294 0 : }
3295 :
3296 : // Logging the full result is useful because it lets us cross-check what the cloud control
3297 : // plane's tenant_shards table should contain.
3298 0 : tracing::info!("Complete, returning {result:?}");
3299 :
3300 0 : Ok(result)
3301 0 : }
3302 :
3303 0 : pub(crate) async fn tenant_config_patch(
3304 0 : &self,
3305 0 : req: TenantConfigPatchRequest,
3306 0 : ) -> Result<(), ApiError> {
3307 0 : let _tenant_lock = trace_exclusive_lock(
3308 0 : &self.tenant_op_locks,
3309 0 : req.tenant_id,
3310 0 : TenantOperations::ConfigPatch,
3311 0 : )
3312 0 : .await;
3313 :
3314 0 : let tenant_id = req.tenant_id;
3315 0 : let patch = req.config;
3316 :
3317 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3318 :
3319 0 : let base = {
3320 0 : let locked = self.inner.read().unwrap();
3321 0 : let shards = locked
3322 0 : .tenants
3323 0 : .range(TenantShardId::tenant_range(req.tenant_id));
3324 :
3325 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
3326 :
3327 0 : let first = match configs.peek() {
3328 0 : Some(first) => (*first).clone(),
3329 : None => {
3330 0 : return Err(ApiError::NotFound(
3331 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3332 0 : ));
3333 : }
3334 : };
3335 :
3336 0 : if !configs.all_equal() {
3337 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3338 : // This can't happen because we atomically update the database records
3339 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3340 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3341 0 : "Tenant configs for {} are mismatched",
3342 0 : req.tenant_id
3343 0 : )));
3344 0 : }
3345 :
3346 0 : first
3347 : };
3348 :
3349 0 : let updated_config = base
3350 0 : .apply_patch(patch)
3351 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3352 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3353 0 : .await
3354 0 : }
3355 :
3356 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3357 : // We require an exclusive lock, because we are updating persistent and in-memory state
3358 0 : let _tenant_lock = trace_exclusive_lock(
3359 0 : &self.tenant_op_locks,
3360 0 : req.tenant_id,
3361 0 : TenantOperations::ConfigSet,
3362 0 : )
3363 0 : .await;
3364 :
3365 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3366 :
3367 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3368 0 : .await
3369 0 : }
3370 :
3371 0 : async fn set_tenant_config_and_reconcile(
3372 0 : &self,
3373 0 : tenant_id: TenantId,
3374 0 : config: TenantConfig,
3375 0 : ) -> Result<(), ApiError> {
3376 0 : self.persistence
3377 0 : .update_tenant_shard(
3378 0 : TenantFilter::Tenant(tenant_id),
3379 0 : None,
3380 0 : Some(config.clone()),
3381 0 : None,
3382 0 : None,
3383 0 : )
3384 0 : .await?;
3385 :
3386 0 : let waiters = {
3387 0 : let mut waiters = Vec::new();
3388 0 : let mut locked = self.inner.write().unwrap();
3389 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3390 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3391 0 : shard.config = config.clone();
3392 0 : if let Some(waiter) =
3393 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3394 0 : {
3395 0 : waiters.push(waiter);
3396 0 : }
3397 : }
3398 0 : waiters
3399 : };
3400 :
3401 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3402 : // Treat this as success because we have stored the configuration. If e.g.
3403 : // a node was unavailable at this time, it should not stop us accepting a
3404 : // configuration change.
3405 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3406 0 : }
3407 :
3408 0 : Ok(())
3409 0 : }
3410 :
3411 0 : pub(crate) fn tenant_config_get(
3412 0 : &self,
3413 0 : tenant_id: TenantId,
3414 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3415 0 : let config = {
3416 0 : let locked = self.inner.read().unwrap();
3417 :
3418 0 : match locked
3419 0 : .tenants
3420 0 : .range(TenantShardId::tenant_range(tenant_id))
3421 0 : .next()
3422 : {
3423 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3424 : None => {
3425 0 : return Err(ApiError::NotFound(
3426 0 : anyhow::anyhow!("Tenant not found").into(),
3427 0 : ));
3428 : }
3429 : }
3430 : };
3431 :
3432 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3433 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3434 : // and `effective_config` in the response is meaningless, but we retain that syntax
3435 : // in order to remain compatible with the pageserver API.
3436 :
3437 0 : let response = HashMap::from([
3438 : (
3439 : "tenant_specific_overrides",
3440 0 : serde_json::to_value(&config)
3441 0 : .context("serializing tenant specific overrides")
3442 0 : .map_err(ApiError::InternalServerError)?,
3443 : ),
3444 : (
3445 0 : "effective_config",
3446 0 : serde_json::to_value(&config)
3447 0 : .context("serializing effective config")
3448 0 : .map_err(ApiError::InternalServerError)?,
3449 : ),
3450 : ]);
3451 :
3452 0 : Ok(response)
3453 0 : }
3454 :
3455 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3456 0 : &self,
3457 0 : time_travel_req: &TenantTimeTravelRequest,
3458 0 : tenant_id: TenantId,
3459 0 : timestamp: Cow<'_, str>,
3460 0 : done_if_after: Cow<'_, str>,
3461 0 : ) -> Result<(), ApiError> {
3462 0 : let _tenant_lock = trace_exclusive_lock(
3463 0 : &self.tenant_op_locks,
3464 0 : tenant_id,
3465 0 : TenantOperations::TimeTravelRemoteStorage,
3466 0 : )
3467 0 : .await;
3468 :
3469 0 : let node = {
3470 0 : let mut locked = self.inner.write().unwrap();
3471 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3472 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3473 : // but only at the start of the process, so it's really just to prevent operator
3474 : // mistakes.
3475 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3476 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3477 : {
3478 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3479 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3480 0 : )));
3481 0 : }
3482 0 : let maybe_attached = shard
3483 0 : .observed
3484 0 : .locations
3485 0 : .iter()
3486 0 : .filter_map(|(node_id, observed_location)| {
3487 0 : observed_location
3488 0 : .conf
3489 0 : .as_ref()
3490 0 : .map(|loc| (node_id, observed_location, loc.mode))
3491 0 : })
3492 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3493 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3494 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3495 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3496 0 : )));
3497 0 : }
3498 : }
3499 0 : let scheduler = &mut locked.scheduler;
3500 : // Right now we only perform the operation on a single node without parallelization
3501 : // TODO fan out the operation to multiple nodes for better performance
3502 0 : let node_id = scheduler.any_available_node()?;
3503 0 : let node = locked
3504 0 : .nodes
3505 0 : .get(&node_id)
3506 0 : .expect("Pageservers may not be deleted while lock is active");
3507 0 : node.clone()
3508 : };
3509 :
3510 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3511 0 : let mut counts = time_travel_req
3512 0 : .shard_counts
3513 0 : .iter()
3514 0 : .copied()
3515 0 : .collect::<HashSet<_>>()
3516 0 : .into_iter()
3517 0 : .collect::<Vec<_>>();
3518 0 : counts.sort_unstable();
3519 :
3520 0 : for count in counts {
3521 0 : let shard_ids = (0..count.count())
3522 0 : .map(|i| TenantShardId {
3523 0 : tenant_id,
3524 0 : shard_number: ShardNumber(i),
3525 0 : shard_count: count,
3526 0 : })
3527 0 : .collect::<Vec<_>>();
3528 0 : for tenant_shard_id in shard_ids {
3529 0 : let client = PageserverClient::new(
3530 0 : node.get_id(),
3531 0 : self.http_client.clone(),
3532 0 : node.base_url(),
3533 0 : self.config.pageserver_jwt_token.as_deref(),
3534 : );
3535 :
3536 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3537 :
3538 0 : client
3539 0 : .tenant_time_travel_remote_storage(
3540 0 : tenant_shard_id,
3541 0 : ×tamp,
3542 0 : &done_if_after,
3543 0 : )
3544 0 : .await
3545 0 : .map_err(|e| {
3546 0 : ApiError::InternalServerError(anyhow::anyhow!(
3547 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3548 0 : node
3549 0 : ))
3550 0 : })?;
3551 : }
3552 : }
3553 0 : Ok(())
3554 0 : }
3555 :
3556 0 : pub(crate) async fn tenant_secondary_download(
3557 0 : &self,
3558 0 : tenant_id: TenantId,
3559 0 : wait: Option<Duration>,
3560 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3561 0 : let _tenant_lock = trace_shared_lock(
3562 0 : &self.tenant_op_locks,
3563 0 : tenant_id,
3564 0 : TenantOperations::SecondaryDownload,
3565 0 : )
3566 0 : .await;
3567 :
3568 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3569 0 : let targets = {
3570 0 : let locked = self.inner.read().unwrap();
3571 0 : let mut targets = Vec::new();
3572 :
3573 0 : for (tenant_shard_id, shard) in
3574 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3575 : {
3576 0 : for node_id in shard.intent.get_secondary() {
3577 0 : let node = locked
3578 0 : .nodes
3579 0 : .get(node_id)
3580 0 : .expect("Pageservers may not be deleted while referenced");
3581 0 :
3582 0 : targets.push((*tenant_shard_id, node.clone()));
3583 0 : }
3584 : }
3585 0 : targets
3586 : };
3587 :
3588 : // Issue concurrent requests to all shards' locations
3589 0 : let mut futs = FuturesUnordered::new();
3590 0 : for (tenant_shard_id, node) in targets {
3591 0 : let client = PageserverClient::new(
3592 0 : node.get_id(),
3593 0 : self.http_client.clone(),
3594 0 : node.base_url(),
3595 0 : self.config.pageserver_jwt_token.as_deref(),
3596 : );
3597 0 : futs.push(async move {
3598 0 : let result = client
3599 0 : .tenant_secondary_download(tenant_shard_id, wait)
3600 0 : .await;
3601 0 : (result, node, tenant_shard_id)
3602 0 : })
3603 : }
3604 :
3605 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3606 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3607 : // well as more general cases like 503s, 500s, or timeouts.
3608 0 : let mut aggregate_progress = SecondaryProgress::default();
3609 0 : let mut aggregate_status: Option<StatusCode> = None;
3610 0 : let mut error: Option<mgmt_api::Error> = None;
3611 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3612 0 : match result {
3613 0 : Err(e) => {
3614 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3615 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3616 : // than they had hoped for.
3617 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3618 0 : error = Some(e)
3619 : }
3620 0 : Ok((status_code, progress)) => {
3621 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3622 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3623 0 : aggregate_progress.layers_total += progress.layers_total;
3624 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3625 0 : aggregate_progress.bytes_total += progress.bytes_total;
3626 0 : aggregate_progress.heatmap_mtime =
3627 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3628 0 : aggregate_status = match aggregate_status {
3629 0 : None => Some(status_code),
3630 0 : Some(StatusCode::OK) => Some(status_code),
3631 0 : Some(cur) => {
3632 : // Other status codes (e.g. 202) -- do not overwrite.
3633 0 : Some(cur)
3634 : }
3635 : };
3636 : }
3637 : }
3638 : }
3639 :
3640 : // If any of the shards return 202, indicate our result as 202.
3641 0 : match aggregate_status {
3642 : None => {
3643 0 : match error {
3644 0 : Some(e) => {
3645 : // No successes, and an error: surface it
3646 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3647 : }
3648 : None => {
3649 : // No shards found
3650 0 : Err(ApiError::NotFound(
3651 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3652 0 : ))
3653 : }
3654 : }
3655 : }
3656 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3657 : }
3658 0 : }
3659 :
3660 0 : pub(crate) async fn tenant_delete(
3661 0 : self: &Arc<Self>,
3662 0 : tenant_id: TenantId,
3663 0 : ) -> Result<StatusCode, ApiError> {
3664 0 : let _tenant_lock =
3665 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3666 :
3667 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3668 :
3669 : // Detach all shards. This also deletes local pageserver shard data.
3670 0 : let (detach_waiters, node) = {
3671 0 : let mut detach_waiters = Vec::new();
3672 0 : let mut locked = self.inner.write().unwrap();
3673 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3674 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3675 : // Update the tenant's intent to remove all attachments
3676 0 : shard.policy = PlacementPolicy::Detached;
3677 0 : shard
3678 0 : .schedule(scheduler, &mut ScheduleContext::default())
3679 0 : .expect("De-scheduling is infallible");
3680 0 : debug_assert!(shard.intent.get_attached().is_none());
3681 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3682 :
3683 0 : if let Some(waiter) =
3684 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3685 0 : {
3686 0 : detach_waiters.push(waiter);
3687 0 : }
3688 : }
3689 :
3690 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3691 : // was attached, just has to be able to see the S3 content)
3692 0 : let node_id = scheduler.any_available_node()?;
3693 0 : let node = nodes
3694 0 : .get(&node_id)
3695 0 : .expect("Pageservers may not be deleted while lock is active");
3696 0 : (detach_waiters, node.clone())
3697 : };
3698 :
3699 : // This reconcile wait can fail in a few ways:
3700 : // A there is a very long queue for the reconciler semaphore
3701 : // B some pageserver is failing to handle a detach promptly
3702 : // C some pageserver goes offline right at the moment we send it a request.
3703 : //
3704 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3705 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3706 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3707 : // deleting the underlying data).
3708 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3709 0 : .await?;
3710 :
3711 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3712 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3713 : // the tenant ID prefix, including all shards (even possibly stale ones).
3714 0 : match node
3715 0 : .with_client_retries(
3716 0 : |client| async move {
3717 0 : client
3718 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3719 0 : .await
3720 0 : },
3721 0 : &self.http_client,
3722 0 : &self.config.pageserver_jwt_token,
3723 : 1,
3724 : 3,
3725 : RECONCILE_TIMEOUT,
3726 0 : &self.cancel,
3727 : )
3728 0 : .await
3729 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3730 : {
3731 0 : Ok(_) => {}
3732 : Err(mgmt_api::Error::Cancelled) => {
3733 0 : return Err(ApiError::ShuttingDown);
3734 : }
3735 0 : Err(e) => {
3736 : // This is unexpected: remote deletion should be infallible, unless the object store
3737 : // at large is unavailable.
3738 0 : tracing::error!("Error deleting via node {node}: {e}");
3739 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3740 : }
3741 : }
3742 :
3743 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3744 : // our in-memory state and database state.
3745 :
3746 : // Ordering: we delete persistent state first: if we then
3747 : // crash, we will drop the in-memory state.
3748 :
3749 : // Drop persistent state.
3750 0 : self.persistence.delete_tenant(tenant_id).await?;
3751 :
3752 : // Drop in-memory state
3753 : {
3754 0 : let mut locked = self.inner.write().unwrap();
3755 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3756 :
3757 : // Dereference Scheduler from shards before dropping them
3758 0 : for (_tenant_shard_id, shard) in
3759 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3760 0 : {
3761 0 : shard.intent.clear(scheduler);
3762 0 : }
3763 :
3764 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3765 0 : tracing::info!(
3766 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3767 0 : locked.tenants.len()
3768 : );
3769 : };
3770 :
3771 : // Delete the tenant from safekeepers (if needed)
3772 0 : self.tenant_delete_safekeepers(tenant_id)
3773 0 : .instrument(tracing::info_span!("tenant_delete_safekeepers", %tenant_id))
3774 0 : .await?;
3775 :
3776 : // Success is represented as 404, to imitate the existing pageserver deletion API
3777 0 : Ok(StatusCode::NOT_FOUND)
3778 0 : }
3779 :
3780 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3781 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3782 : /// the tenant's policies (configuration) within the storage controller
3783 0 : pub(crate) async fn tenant_update_policy(
3784 0 : &self,
3785 0 : tenant_id: TenantId,
3786 0 : req: TenantPolicyRequest,
3787 0 : ) -> Result<(), ApiError> {
3788 : // We require an exclusive lock, because we are updating persistent and in-memory state
3789 0 : let _tenant_lock = trace_exclusive_lock(
3790 0 : &self.tenant_op_locks,
3791 0 : tenant_id,
3792 0 : TenantOperations::UpdatePolicy,
3793 0 : )
3794 0 : .await;
3795 :
3796 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3797 :
3798 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3799 :
3800 : let TenantPolicyRequest {
3801 0 : placement,
3802 0 : mut scheduling,
3803 0 : } = req;
3804 :
3805 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3806 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3807 : // this to take effect.
3808 0 : if scheduling.is_none() {
3809 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3810 0 : }
3811 0 : }
3812 :
3813 0 : self.persistence
3814 0 : .update_tenant_shard(
3815 0 : TenantFilter::Tenant(tenant_id),
3816 0 : placement.clone(),
3817 0 : None,
3818 0 : None,
3819 0 : scheduling,
3820 0 : )
3821 0 : .await?;
3822 :
3823 0 : let mut schedule_context = ScheduleContext::default();
3824 0 : let mut locked = self.inner.write().unwrap();
3825 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3826 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3827 0 : if let Some(placement) = &placement {
3828 0 : shard.policy = placement.clone();
3829 :
3830 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3831 0 : "Updated placement policy to {placement:?}");
3832 0 : }
3833 :
3834 0 : if let Some(scheduling) = &scheduling {
3835 0 : shard.set_scheduling_policy(*scheduling);
3836 :
3837 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3838 0 : "Updated scheduling policy to {scheduling:?}");
3839 0 : }
3840 :
3841 : // In case scheduling is being switched back on, try it now.
3842 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3843 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3844 : }
3845 :
3846 0 : Ok(())
3847 0 : }
3848 :
3849 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3850 0 : &self,
3851 0 : tenant_id: TenantId,
3852 0 : mut create_req: TimelineCreateRequest,
3853 0 : ) -> Result<TimelineInfo, ApiError> {
3854 0 : tracing::info!(
3855 0 : "Creating timeline {}/{}",
3856 : tenant_id,
3857 : create_req.new_timeline_id,
3858 : );
3859 :
3860 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3861 0 : if targets.0.is_empty() {
3862 0 : return Err(ApiError::NotFound(
3863 0 : anyhow::anyhow!("Tenant not found").into(),
3864 0 : ));
3865 0 : };
3866 :
3867 0 : let (shard_zero_tid, shard_zero_locations) =
3868 0 : targets.0.pop_first().expect("Must have at least one shard");
3869 0 : assert!(shard_zero_tid.is_shard_zero());
3870 :
3871 0 : async fn create_one(
3872 0 : tenant_shard_id: TenantShardId,
3873 0 : locations: ShardMutationLocations,
3874 0 : http_client: reqwest::Client,
3875 0 : jwt: Option<String>,
3876 0 : mut create_req: TimelineCreateRequest,
3877 0 : ) -> Result<TimelineInfo, ApiError> {
3878 0 : let latest = locations.latest.node;
3879 :
3880 0 : tracing::info!(
3881 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3882 : tenant_shard_id,
3883 : create_req.new_timeline_id,
3884 : locations.latest.generation
3885 : );
3886 :
3887 0 : let client =
3888 0 : PageserverClient::new(latest.get_id(), http_client.clone(), latest.base_url(), jwt.as_deref());
3889 :
3890 0 : let timeline_info = client
3891 0 : .timeline_create(tenant_shard_id, &create_req)
3892 0 : .await
3893 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3894 :
3895 : // If we are going to create the timeline on some stale locations for shard 0, then ask them to re-use
3896 : // the initdb generated by the latest location, rather than generating their own. This avoids racing uploads
3897 : // of initdb to S3 which might not be binary-identical if different pageservers have different postgres binaries.
3898 0 : if tenant_shard_id.is_shard_zero() {
3899 0 : if let models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } = &mut create_req.mode {
3900 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id);
3901 0 : }
3902 0 : }
3903 :
3904 : // We propagate timeline creations to all attached locations such that a compute
3905 : // for the new timeline is able to start regardless of the current state of the
3906 : // tenant shard reconciliation.
3907 0 : for location in locations.other {
3908 0 : tracing::info!(
3909 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3910 : tenant_shard_id,
3911 : create_req.new_timeline_id,
3912 : location.node,
3913 : location.generation
3914 : );
3915 :
3916 0 : let client = PageserverClient::new(
3917 0 : location.node.get_id(),
3918 0 : http_client.clone(),
3919 0 : location.node.base_url(),
3920 0 : jwt.as_deref(),
3921 : );
3922 :
3923 0 : let res = client
3924 0 : .timeline_create(tenant_shard_id, &create_req)
3925 0 : .await;
3926 :
3927 0 : if let Err(e) = res {
3928 0 : match e {
3929 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3930 0 : // Tenant might have been detached from the stale location,
3931 0 : // so ignore 404s.
3932 0 : },
3933 : _ => {
3934 0 : return Err(passthrough_api_error(&location.node, e));
3935 : }
3936 : }
3937 0 : }
3938 : }
3939 :
3940 0 : Ok(timeline_info)
3941 0 : }
3942 :
3943 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3944 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3945 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3946 : //
3947 : // This also enables non-zero shards to use the initdb that shard 0 generated and uploaded to S3, rather than
3948 : // independently generating their own initdb. This guarantees that shards cannot end up with different initial
3949 : // states if e.g. they have different postgres binary versions.
3950 0 : let timeline_info = create_one(
3951 0 : shard_zero_tid,
3952 0 : shard_zero_locations,
3953 0 : self.http_client.clone(),
3954 0 : self.config.pageserver_jwt_token.clone(),
3955 0 : create_req.clone(),
3956 0 : )
3957 0 : .await?;
3958 :
3959 : // Update the create request for shards >= 0
3960 0 : match &mut create_req.mode {
3961 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3962 0 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3963 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3964 0 : },
3965 0 : models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } => {
3966 : // For shards >= 0, do not run initdb: use the one that shard 0 uploaded to S3
3967 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id)
3968 : }
3969 0 : _ => {}
3970 : }
3971 :
3972 : // Create timeline on remaining shards with number >0
3973 0 : if !targets.0.is_empty() {
3974 : // If we had multiple shards, issue requests for the remainder now.
3975 0 : let jwt = &self.config.pageserver_jwt_token;
3976 0 : self.tenant_for_shards(
3977 0 : targets
3978 0 : .0
3979 0 : .iter()
3980 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3981 0 : .collect(),
3982 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3983 0 : let create_req = create_req.clone();
3984 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3985 0 : Box::pin(create_one(
3986 0 : tenant_shard_id,
3987 0 : mutation_locations,
3988 0 : self.http_client.clone(),
3989 0 : jwt.clone(),
3990 0 : create_req,
3991 0 : ))
3992 0 : },
3993 : )
3994 0 : .await?;
3995 0 : }
3996 :
3997 0 : Ok(timeline_info)
3998 0 : })
3999 0 : .await?
4000 0 : }
4001 :
4002 0 : pub(crate) async fn tenant_timeline_create(
4003 0 : self: &Arc<Self>,
4004 0 : tenant_id: TenantId,
4005 0 : create_req: TimelineCreateRequest,
4006 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
4007 0 : let safekeepers = self.config.timelines_onto_safekeepers;
4008 0 : let timeline_id = create_req.new_timeline_id;
4009 :
4010 0 : tracing::info!(
4011 0 : mode=%create_req.mode_tag(),
4012 : %safekeepers,
4013 0 : "Creating timeline {}/{}",
4014 : tenant_id,
4015 : timeline_id,
4016 : );
4017 :
4018 0 : let _tenant_lock = trace_shared_lock(
4019 0 : &self.tenant_op_locks,
4020 0 : tenant_id,
4021 0 : TenantOperations::TimelineCreate,
4022 0 : )
4023 0 : .await;
4024 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
4025 0 : let is_import = create_req.is_import();
4026 0 : let read_only = matches!(
4027 0 : create_req.mode,
4028 : models::TimelineCreateRequestMode::Branch {
4029 : read_only: true,
4030 : ..
4031 : }
4032 : );
4033 :
4034 0 : if is_import {
4035 : // Ensure that there is no split on-going.
4036 : // [`Self::tenant_shard_split`] holds the exclusive tenant lock
4037 : // for the duration of the split, but here we handle the case
4038 : // where we restarted and the split is being aborted.
4039 0 : let locked = self.inner.read().unwrap();
4040 0 : let splitting = locked
4041 0 : .tenants
4042 0 : .range(TenantShardId::tenant_range(tenant_id))
4043 0 : .any(|(_id, shard)| shard.splitting != SplitState::Idle);
4044 :
4045 0 : if splitting {
4046 0 : return Err(ApiError::Conflict("Tenant is splitting shard".to_string()));
4047 0 : }
4048 0 : }
4049 :
4050 0 : let timeline_info = self
4051 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
4052 0 : .await?;
4053 :
4054 0 : let selected_safekeepers = if is_import {
4055 0 : let shards = {
4056 0 : let locked = self.inner.read().unwrap();
4057 0 : locked
4058 0 : .tenants
4059 0 : .range(TenantShardId::tenant_range(tenant_id))
4060 0 : .map(|(ts_id, _)| ts_id.to_index())
4061 0 : .collect::<Vec<_>>()
4062 : };
4063 :
4064 0 : if !shards
4065 0 : .iter()
4066 0 : .map(|shard_index| shard_index.shard_count)
4067 0 : .all_equal()
4068 : {
4069 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4070 0 : "Inconsistent shard count"
4071 0 : )));
4072 0 : }
4073 :
4074 0 : let import = TimelineImport {
4075 0 : tenant_id,
4076 0 : timeline_id,
4077 0 : shard_statuses: ShardImportStatuses::new(shards),
4078 0 : };
4079 :
4080 0 : let inserted = self
4081 0 : .persistence
4082 0 : .insert_timeline_import(import.to_persistent())
4083 0 : .await
4084 0 : .context("timeline import insert")
4085 0 : .map_err(ApiError::InternalServerError)?;
4086 :
4087 : // Set the importing flag on the tenant shards
4088 0 : self.inner
4089 0 : .write()
4090 0 : .unwrap()
4091 0 : .tenants
4092 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4093 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Importing);
4094 :
4095 0 : match inserted {
4096 : true => {
4097 0 : tracing::info!(%tenant_id, %timeline_id, "Inserted timeline import");
4098 : }
4099 : false => {
4100 0 : tracing::info!(%tenant_id, %timeline_id, "Timeline import entry already present");
4101 : }
4102 : }
4103 :
4104 0 : None
4105 0 : } else if safekeepers || read_only {
4106 : // Note that for imported timelines, we do not create the timeline on the safekeepers
4107 : // straight away. Instead, we do it once the import finalized such that we know what
4108 : // start LSN to provide for the safekeepers. This is done in
4109 : // [`Self::finalize_timeline_import`].
4110 0 : let res = self
4111 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
4112 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
4113 0 : .await?;
4114 0 : Some(res)
4115 : } else {
4116 0 : None
4117 : };
4118 :
4119 0 : Ok(TimelineCreateResponseStorcon {
4120 0 : timeline_info,
4121 0 : safekeepers: selected_safekeepers,
4122 0 : })
4123 0 : }
4124 :
4125 : #[instrument(skip_all, fields(
4126 : tenant_id=%req.tenant_shard_id.tenant_id,
4127 : shard_id=%req.tenant_shard_id.shard_slug(),
4128 : timeline_id=%req.timeline_id,
4129 : ))]
4130 : pub(crate) async fn handle_timeline_shard_import_progress(
4131 : self: &Arc<Self>,
4132 : req: TimelineImportStatusRequest,
4133 : ) -> Result<ShardImportStatus, ApiError> {
4134 : let validity = self
4135 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4136 : .await?;
4137 : match validity {
4138 : ShardGenerationValidity::Valid => {
4139 : // fallthrough
4140 : }
4141 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4142 : tracing::info!(
4143 : claimed=?claimed.into(),
4144 0 : actual=?actual.and_then(|g| g.into()),
4145 : "Rejecting import progress fetch from stale generation"
4146 : );
4147 :
4148 : return Err(ApiError::BadRequest(anyhow::anyhow!("Invalid generation")));
4149 : }
4150 : }
4151 :
4152 : let maybe_import = self
4153 : .persistence
4154 : .get_timeline_import(req.tenant_shard_id.tenant_id, req.timeline_id)
4155 : .await?;
4156 :
4157 0 : let import = maybe_import.ok_or_else(|| {
4158 0 : ApiError::NotFound(
4159 0 : format!(
4160 0 : "import for {}/{} not found",
4161 0 : req.tenant_shard_id.tenant_id, req.timeline_id
4162 0 : )
4163 0 : .into(),
4164 0 : )
4165 0 : })?;
4166 :
4167 : import
4168 : .shard_statuses
4169 : .0
4170 : .get(&req.tenant_shard_id.to_index())
4171 : .cloned()
4172 0 : .ok_or_else(|| {
4173 0 : ApiError::NotFound(
4174 0 : format!("shard {} not found", req.tenant_shard_id.shard_slug()).into(),
4175 0 : )
4176 0 : })
4177 : }
4178 :
4179 : #[instrument(skip_all, fields(
4180 : tenant_id=%req.tenant_shard_id.tenant_id,
4181 : shard_id=%req.tenant_shard_id.shard_slug(),
4182 : timeline_id=%req.timeline_id,
4183 : ))]
4184 : pub(crate) async fn handle_timeline_shard_import_progress_upcall(
4185 : self: &Arc<Self>,
4186 : req: PutTimelineImportStatusRequest,
4187 : ) -> Result<(), ApiError> {
4188 : let validity = self
4189 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4190 : .await?;
4191 : match validity {
4192 : ShardGenerationValidity::Valid => {
4193 : // fallthrough
4194 : }
4195 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4196 : tracing::info!(
4197 : claimed=?claimed.into(),
4198 0 : actual=?actual.and_then(|g| g.into()),
4199 : "Rejecting import progress update from stale generation"
4200 : );
4201 :
4202 : return Err(ApiError::PreconditionFailed("Invalid generation".into()));
4203 : }
4204 : }
4205 :
4206 : let res = self
4207 : .persistence
4208 : .update_timeline_import(req.tenant_shard_id, req.timeline_id, req.status)
4209 : .await;
4210 : let timeline_import = match res {
4211 : Ok(Ok(Some(timeline_import))) => timeline_import,
4212 : Ok(Ok(None)) => {
4213 : // Idempotency: we've already seen and handled this update.
4214 : return Ok(());
4215 : }
4216 : Ok(Err(logical_err)) => {
4217 : return Err(logical_err.into());
4218 : }
4219 : Err(db_err) => {
4220 : return Err(db_err.into());
4221 : }
4222 : };
4223 :
4224 : tracing::info!(
4225 : tenant_id=%req.tenant_shard_id.tenant_id,
4226 : timeline_id=%req.timeline_id,
4227 : shard_id=%req.tenant_shard_id.shard_slug(),
4228 : "Updated timeline import status to: {timeline_import:?}");
4229 :
4230 : if timeline_import.is_complete() {
4231 : tokio::task::spawn({
4232 : let this = self.clone();
4233 0 : async move { this.finalize_timeline_import(timeline_import).await }
4234 : });
4235 : }
4236 :
4237 : Ok(())
4238 : }
4239 :
4240 : /// Check that a provided generation for some tenant shard is the most recent one.
4241 : ///
4242 : /// Validate with the in-mem state first, and, if that passes, validate with the
4243 : /// database state which is authoritative.
4244 0 : async fn validate_shard_generation(
4245 0 : self: &Arc<Self>,
4246 0 : tenant_shard_id: TenantShardId,
4247 0 : generation: Generation,
4248 0 : ) -> Result<ShardGenerationValidity, ApiError> {
4249 : {
4250 0 : let locked = self.inner.read().unwrap();
4251 0 : let tenant_shard =
4252 0 : locked
4253 0 : .tenants
4254 0 : .get(&tenant_shard_id)
4255 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4256 0 : "{} shard not found",
4257 0 : tenant_shard_id
4258 0 : )))?;
4259 :
4260 0 : if tenant_shard.generation != Some(generation) {
4261 0 : return Ok(ShardGenerationValidity::Mismatched {
4262 0 : claimed: generation,
4263 0 : actual: tenant_shard.generation,
4264 0 : });
4265 0 : }
4266 : }
4267 :
4268 0 : let mut db_generations = self
4269 0 : .persistence
4270 0 : .shard_generations(std::iter::once(&tenant_shard_id))
4271 0 : .await?;
4272 0 : let (_tid, db_generation) =
4273 0 : db_generations
4274 0 : .pop()
4275 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4276 0 : "{} shard not found",
4277 0 : tenant_shard_id
4278 0 : )))?;
4279 :
4280 0 : if db_generation != Some(generation) {
4281 0 : return Ok(ShardGenerationValidity::Mismatched {
4282 0 : claimed: generation,
4283 0 : actual: db_generation,
4284 0 : });
4285 0 : }
4286 :
4287 0 : Ok(ShardGenerationValidity::Valid)
4288 0 : }
4289 :
4290 : /// Finalize the import of a timeline
4291 : ///
4292 : /// This method should be called once all shards have reported that the import is complete.
4293 : /// Firstly, it polls the post import timeline activation endpoint exposed by the pageserver.
4294 : /// Once the timeline is active on all shards, the timeline also gets created on the
4295 : /// safekeepers. Finally, notify cplane of the import completion (whether failed or
4296 : /// successful), and remove the import from the database and in-memory.
4297 : ///
4298 : /// If this method gets pre-empted by shut down, it will be called again at start-up (on-going
4299 : /// imports are stored in the database).
4300 : ///
4301 : /// # Cancel-Safety
4302 : /// Not cancel safe.
4303 : /// If the caller stops polling, the import will not be removed from
4304 : /// [`ServiceState::imports_finalizing`].
4305 : #[instrument(skip_all, fields(
4306 : tenant_id=%import.tenant_id,
4307 : timeline_id=%import.timeline_id,
4308 : ))]
4309 :
4310 : async fn finalize_timeline_import(
4311 : self: &Arc<Self>,
4312 : import: TimelineImport,
4313 : ) -> Result<(), TimelineImportFinalizeError> {
4314 : let tenant_timeline = (import.tenant_id, import.timeline_id);
4315 :
4316 : let (_finalize_import_guard, cancel) = {
4317 : let mut locked = self.inner.write().unwrap();
4318 : let gate = Gate::default();
4319 : let cancel = CancellationToken::default();
4320 :
4321 : let guard = gate.enter().unwrap();
4322 :
4323 : locked.imports_finalizing.insert(
4324 : tenant_timeline,
4325 : FinalizingImport {
4326 : gate,
4327 : cancel: cancel.clone(),
4328 : },
4329 : );
4330 :
4331 : (guard, cancel)
4332 : };
4333 :
4334 : let res = tokio::select! {
4335 : res = self.finalize_timeline_import_impl(import) => {
4336 : res
4337 : },
4338 : _ = cancel.cancelled() => {
4339 : Err(TimelineImportFinalizeError::Cancelled)
4340 : }
4341 : };
4342 :
4343 : let mut locked = self.inner.write().unwrap();
4344 : locked.imports_finalizing.remove(&tenant_timeline);
4345 :
4346 : res
4347 : }
4348 :
4349 0 : async fn finalize_timeline_import_impl(
4350 0 : self: &Arc<Self>,
4351 0 : import: TimelineImport,
4352 0 : ) -> Result<(), TimelineImportFinalizeError> {
4353 0 : tracing::info!("Finalizing timeline import");
4354 :
4355 0 : pausable_failpoint!("timeline-import-pre-cplane-notification");
4356 :
4357 0 : let tenant_id = import.tenant_id;
4358 0 : let timeline_id = import.timeline_id;
4359 :
4360 0 : let import_error = import.completion_error();
4361 0 : match import_error {
4362 0 : Some(err) => {
4363 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Err(err))
4364 0 : .await?;
4365 0 : tracing::warn!("Timeline import completed with shard errors");
4366 0 : Ok(())
4367 : }
4368 0 : None => match self.activate_timeline_post_import(&import).await {
4369 0 : Ok(timeline_info) => {
4370 0 : tracing::info!("Post import timeline activation complete");
4371 :
4372 0 : if self.config.timelines_onto_safekeepers {
4373 : // Now that we know the start LSN of this timeline, create it on the
4374 : // safekeepers.
4375 0 : self.tenant_timeline_create_safekeepers_until_success(
4376 0 : import.tenant_id,
4377 0 : timeline_info,
4378 0 : )
4379 0 : .await?;
4380 0 : }
4381 :
4382 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Ok(()))
4383 0 : .await?;
4384 :
4385 0 : tracing::info!("Timeline import completed successfully");
4386 0 : Ok(())
4387 : }
4388 : Err(TimelineImportFinalizeError::ShuttingDown) => {
4389 : // We got pre-empted by shut down and will resume after the restart.
4390 0 : Err(TimelineImportFinalizeError::ShuttingDown)
4391 : }
4392 0 : Err(err) => {
4393 : // Any finalize error apart from shut down is permanent and requires us to notify
4394 : // cplane such that it can clean up.
4395 0 : tracing::error!("Import finalize failed with permanent error: {err}");
4396 0 : self.notify_cplane_and_delete_import(
4397 0 : tenant_id,
4398 0 : timeline_id,
4399 0 : Err(err.to_string()),
4400 0 : )
4401 0 : .await?;
4402 0 : Err(err)
4403 : }
4404 : },
4405 : }
4406 0 : }
4407 :
4408 0 : async fn notify_cplane_and_delete_import(
4409 0 : self: &Arc<Self>,
4410 0 : tenant_id: TenantId,
4411 0 : timeline_id: TimelineId,
4412 0 : import_result: ImportResult,
4413 0 : ) -> Result<(), TimelineImportFinalizeError> {
4414 0 : let import_failed = import_result.is_err();
4415 0 : tracing::info!(%import_failed, "Notifying cplane of import completion");
4416 :
4417 0 : let client = UpcallClient::new(self.get_config(), self.cancel.child_token());
4418 0 : client
4419 0 : .notify_import_complete(tenant_id, timeline_id, import_result)
4420 0 : .await
4421 0 : .map_err(|_err| TimelineImportFinalizeError::ShuttingDown)?;
4422 :
4423 0 : if let Err(err) = self
4424 0 : .persistence
4425 0 : .delete_timeline_import(tenant_id, timeline_id)
4426 0 : .await
4427 : {
4428 0 : tracing::warn!("Failed to delete timeline import entry from database: {err}");
4429 0 : }
4430 :
4431 0 : self.inner
4432 0 : .write()
4433 0 : .unwrap()
4434 0 : .tenants
4435 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4436 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Idle);
4437 :
4438 0 : Ok(())
4439 0 : }
4440 :
4441 : /// Activate an imported timeline on all shards once the import is complete.
4442 : /// Returns the [`TimelineInfo`] reported by shard zero.
4443 0 : async fn activate_timeline_post_import(
4444 0 : self: &Arc<Self>,
4445 0 : import: &TimelineImport,
4446 0 : ) -> Result<TimelineInfo, TimelineImportFinalizeError> {
4447 : const TIMELINE_ACTIVATE_TIMEOUT: Duration = Duration::from_millis(128);
4448 :
4449 0 : let mut shards_to_activate: HashSet<ShardIndex> =
4450 0 : import.shard_statuses.0.keys().cloned().collect();
4451 0 : let mut shard_zero_timeline_info = None;
4452 :
4453 0 : while !shards_to_activate.is_empty() {
4454 0 : if self.cancel.is_cancelled() {
4455 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4456 0 : }
4457 :
4458 0 : let targets = {
4459 0 : let locked = self.inner.read().unwrap();
4460 0 : let mut targets = Vec::new();
4461 :
4462 0 : for (tenant_shard_id, shard) in locked
4463 0 : .tenants
4464 0 : .range(TenantShardId::tenant_range(import.tenant_id))
4465 : {
4466 0 : if !import
4467 0 : .shard_statuses
4468 0 : .0
4469 0 : .contains_key(&tenant_shard_id.to_index())
4470 : {
4471 0 : return Err(TimelineImportFinalizeError::MismatchedShards(
4472 0 : tenant_shard_id.to_index(),
4473 0 : ));
4474 0 : }
4475 :
4476 0 : if let Some(node_id) = shard.intent.get_attached() {
4477 0 : let node = locked
4478 0 : .nodes
4479 0 : .get(node_id)
4480 0 : .expect("Pageservers may not be deleted while referenced");
4481 0 : targets.push((*tenant_shard_id, node.clone()));
4482 0 : }
4483 : }
4484 :
4485 0 : targets
4486 : };
4487 :
4488 0 : let targeted_tenant_shards: Vec<_> = targets.iter().map(|(tid, _node)| *tid).collect();
4489 :
4490 0 : let results = self
4491 0 : .tenant_for_shards_api(
4492 0 : targets,
4493 0 : |tenant_shard_id, client| async move {
4494 0 : client
4495 0 : .activate_post_import(
4496 0 : tenant_shard_id,
4497 0 : import.timeline_id,
4498 0 : TIMELINE_ACTIVATE_TIMEOUT,
4499 0 : )
4500 0 : .await
4501 0 : },
4502 : 1,
4503 : 1,
4504 : SHORT_RECONCILE_TIMEOUT,
4505 0 : &self.cancel,
4506 : )
4507 0 : .await;
4508 :
4509 0 : let mut failed = 0;
4510 0 : for (tid, (_, result)) in targeted_tenant_shards.iter().zip(results.into_iter()) {
4511 0 : match result {
4512 0 : Ok(ok) => {
4513 0 : if tid.is_shard_zero() {
4514 0 : shard_zero_timeline_info = Some(ok);
4515 0 : }
4516 :
4517 0 : shards_to_activate.remove(&tid.to_index());
4518 : }
4519 0 : Err(_err) => {
4520 0 : failed += 1;
4521 0 : }
4522 : }
4523 : }
4524 :
4525 0 : if failed > 0 {
4526 0 : tracing::info!(
4527 0 : "Failed to activate timeline on {failed} shards post import. Will retry"
4528 : );
4529 0 : }
4530 :
4531 0 : tokio::select! {
4532 0 : _ = tokio::time::sleep(Duration::from_millis(250)) => {},
4533 0 : _ = self.cancel.cancelled() => {
4534 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4535 : }
4536 : }
4537 : }
4538 :
4539 0 : Ok(shard_zero_timeline_info.expect("All shards replied"))
4540 0 : }
4541 :
4542 0 : async fn finalize_timeline_imports(self: &Arc<Self>, imports: Vec<TimelineImport>) {
4543 0 : futures::future::join_all(
4544 0 : imports
4545 0 : .into_iter()
4546 0 : .map(|import| self.finalize_timeline_import(import)),
4547 : )
4548 0 : .await;
4549 0 : }
4550 :
4551 : /// Delete a timeline import if it exists
4552 : ///
4553 : /// Firstly, delete the entry from the database. Any updates
4554 : /// from pageservers after the update will fail with a 404, so the
4555 : /// import cannot progress into finalizing state if it's not there already.
4556 : /// Secondly, cancel the finalization if one is in progress.
4557 0 : pub(crate) async fn maybe_delete_timeline_import(
4558 0 : self: &Arc<Self>,
4559 0 : tenant_id: TenantId,
4560 0 : timeline_id: TimelineId,
4561 0 : ) -> Result<(), DatabaseError> {
4562 0 : let tenant_has_ongoing_import = {
4563 0 : let locked = self.inner.read().unwrap();
4564 0 : locked
4565 0 : .tenants
4566 0 : .range(TenantShardId::tenant_range(tenant_id))
4567 0 : .any(|(_tid, shard)| shard.importing == TimelineImportState::Importing)
4568 : };
4569 :
4570 0 : if !tenant_has_ongoing_import {
4571 0 : return Ok(());
4572 0 : }
4573 :
4574 0 : self.persistence
4575 0 : .delete_timeline_import(tenant_id, timeline_id)
4576 0 : .await?;
4577 :
4578 0 : let maybe_finalizing = {
4579 0 : let mut locked = self.inner.write().unwrap();
4580 0 : locked.imports_finalizing.remove(&(tenant_id, timeline_id))
4581 : };
4582 :
4583 0 : if let Some(finalizing) = maybe_finalizing {
4584 0 : finalizing.cancel.cancel();
4585 0 : finalizing.gate.close().await;
4586 0 : }
4587 :
4588 0 : Ok(())
4589 0 : }
4590 :
4591 0 : pub(crate) async fn tenant_timeline_archival_config(
4592 0 : &self,
4593 0 : tenant_id: TenantId,
4594 0 : timeline_id: TimelineId,
4595 0 : req: TimelineArchivalConfigRequest,
4596 0 : ) -> Result<(), ApiError> {
4597 0 : tracing::info!(
4598 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
4599 : req.state
4600 : );
4601 :
4602 0 : let _tenant_lock = trace_shared_lock(
4603 0 : &self.tenant_op_locks,
4604 0 : tenant_id,
4605 0 : TenantOperations::TimelineArchivalConfig,
4606 0 : )
4607 0 : .await;
4608 :
4609 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4610 0 : if targets.0.is_empty() {
4611 0 : return Err(ApiError::NotFound(
4612 0 : anyhow::anyhow!("Tenant not found").into(),
4613 0 : ));
4614 0 : }
4615 0 : async fn config_one(
4616 0 : tenant_shard_id: TenantShardId,
4617 0 : timeline_id: TimelineId,
4618 0 : node: Node,
4619 0 : http_client: reqwest::Client,
4620 0 : jwt: Option<String>,
4621 0 : req: TimelineArchivalConfigRequest,
4622 0 : ) -> Result<(), ApiError> {
4623 0 : tracing::info!(
4624 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4625 : );
4626 :
4627 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4628 :
4629 0 : client
4630 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
4631 0 : .await
4632 0 : .map_err(|e| match e {
4633 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
4634 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
4635 : }
4636 0 : _ => passthrough_api_error(&node, e),
4637 0 : })
4638 0 : }
4639 :
4640 : // no shard needs to go first/last; the operation should be idempotent
4641 : // TODO: it would be great to ensure that all shards return the same error
4642 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4643 0 : let results = self
4644 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4645 0 : futures::FutureExt::boxed(config_one(
4646 0 : tenant_shard_id,
4647 0 : timeline_id,
4648 0 : node,
4649 0 : self.http_client.clone(),
4650 0 : self.config.pageserver_jwt_token.clone(),
4651 0 : req.clone(),
4652 0 : ))
4653 0 : })
4654 0 : .await?;
4655 0 : assert!(!results.is_empty(), "must have at least one result");
4656 :
4657 0 : Ok(())
4658 0 : }).await?
4659 0 : }
4660 :
4661 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
4662 0 : &self,
4663 0 : tenant_id: TenantId,
4664 0 : timeline_id: TimelineId,
4665 0 : behavior: Option<DetachBehavior>,
4666 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
4667 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
4668 :
4669 0 : let _tenant_lock = trace_shared_lock(
4670 0 : &self.tenant_op_locks,
4671 0 : tenant_id,
4672 0 : TenantOperations::TimelineDetachAncestor,
4673 0 : )
4674 0 : .await;
4675 :
4676 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4677 0 : if targets.0.is_empty() {
4678 0 : return Err(ApiError::NotFound(
4679 0 : anyhow::anyhow!("Tenant not found").into(),
4680 0 : ));
4681 0 : }
4682 :
4683 0 : async fn detach_one(
4684 0 : tenant_shard_id: TenantShardId,
4685 0 : timeline_id: TimelineId,
4686 0 : node: Node,
4687 0 : http_client: reqwest::Client,
4688 0 : jwt: Option<String>,
4689 0 : behavior: Option<DetachBehavior>,
4690 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
4691 0 : tracing::info!(
4692 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4693 : );
4694 :
4695 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4696 :
4697 0 : client
4698 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id, behavior)
4699 0 : .await
4700 0 : .map_err(|e| {
4701 : use mgmt_api::Error;
4702 :
4703 0 : match e {
4704 : // no ancestor (ever)
4705 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
4706 0 : "{node}: {}",
4707 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
4708 0 : )),
4709 : // too many ancestors
4710 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
4711 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
4712 : }
4713 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
4714 : // avoid turning these into conflicts to remain compatible with
4715 : // pageservers, 500 errors are sadly retryable with timeline ancestor
4716 : // detach
4717 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
4718 : }
4719 : // rest can be mapped as usual
4720 0 : other => passthrough_api_error(&node, other),
4721 : }
4722 0 : })
4723 0 : .map(|res| (tenant_shard_id.shard_number, res))
4724 0 : }
4725 :
4726 : // no shard needs to go first/last; the operation should be idempotent
4727 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4728 0 : let mut results = self
4729 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4730 0 : futures::FutureExt::boxed(detach_one(
4731 0 : tenant_shard_id,
4732 0 : timeline_id,
4733 0 : node,
4734 0 : self.http_client.clone(),
4735 0 : self.config.pageserver_jwt_token.clone(),
4736 0 : behavior,
4737 0 : ))
4738 0 : })
4739 0 : .await?;
4740 :
4741 0 : let any = results.pop().expect("we must have at least one response");
4742 :
4743 0 : let mismatching = results
4744 0 : .iter()
4745 0 : .filter(|(_, res)| res != &any.1)
4746 0 : .collect::<Vec<_>>();
4747 0 : if !mismatching.is_empty() {
4748 : // this can be hit by races which should not happen because operation lock on cplane
4749 0 : let matching = results.len() - mismatching.len();
4750 0 : tracing::error!(
4751 : matching,
4752 : compared_against=?any,
4753 : ?mismatching,
4754 0 : "shards returned different results"
4755 : );
4756 :
4757 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
4758 0 : }
4759 :
4760 0 : Ok(any.1)
4761 0 : }).await?
4762 0 : }
4763 :
4764 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
4765 0 : &self,
4766 0 : tenant_id: TenantId,
4767 0 : timeline_id: TimelineId,
4768 0 : dir: BlockUnblock,
4769 0 : ) -> Result<(), ApiError> {
4770 0 : let _tenant_lock = trace_shared_lock(
4771 0 : &self.tenant_op_locks,
4772 0 : tenant_id,
4773 0 : TenantOperations::TimelineGcBlockUnblock,
4774 0 : )
4775 0 : .await;
4776 :
4777 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4778 0 : if targets.0.is_empty() {
4779 0 : return Err(ApiError::NotFound(
4780 0 : anyhow::anyhow!("Tenant not found").into(),
4781 0 : ));
4782 0 : }
4783 :
4784 0 : async fn do_one(
4785 0 : tenant_shard_id: TenantShardId,
4786 0 : timeline_id: TimelineId,
4787 0 : node: Node,
4788 0 : http_client: reqwest::Client,
4789 0 : jwt: Option<String>,
4790 0 : dir: BlockUnblock,
4791 0 : ) -> Result<(), ApiError> {
4792 0 : let client = PageserverClient::new(
4793 0 : node.get_id(),
4794 0 : http_client,
4795 0 : node.base_url(),
4796 0 : jwt.as_deref(),
4797 : );
4798 :
4799 0 : client
4800 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
4801 0 : .await
4802 0 : .map_err(|e| passthrough_api_error(&node, e))
4803 0 : }
4804 :
4805 : // no shard needs to go first/last; the operation should be idempotent
4806 0 : let locations = targets
4807 0 : .0
4808 0 : .iter()
4809 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
4810 0 : .collect();
4811 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
4812 0 : futures::FutureExt::boxed(do_one(
4813 0 : tenant_shard_id,
4814 0 : timeline_id,
4815 0 : node,
4816 0 : self.http_client.clone(),
4817 0 : self.config.pageserver_jwt_token.clone(),
4818 0 : dir,
4819 0 : ))
4820 0 : })
4821 0 : .await
4822 0 : })
4823 0 : .await??;
4824 0 : Ok(())
4825 0 : }
4826 :
4827 0 : pub(crate) fn is_tenant_not_found_error(body: &str, tenant_id: TenantId) -> bool {
4828 0 : body.contains(&format!("tenant {tenant_id}"))
4829 0 : }
4830 :
4831 0 : fn process_result_and_passthrough_errors<T>(
4832 0 : &self,
4833 0 : tenant_id: TenantId,
4834 0 : results: Vec<(Node, Result<T, mgmt_api::Error>)>,
4835 0 : ) -> Result<Vec<(Node, T)>, ApiError> {
4836 0 : let mut processed_results: Vec<(Node, T)> = Vec::with_capacity(results.len());
4837 0 : for (node, res) in results {
4838 0 : match res {
4839 0 : Ok(res) => processed_results.push((node, res)),
4840 0 : Err(mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, body))
4841 0 : if Self::is_tenant_not_found_error(&body, tenant_id) =>
4842 : {
4843 : // If there's a tenant not found, we are still in the process of attaching the tenant.
4844 : // Return 503 so that the client can retry.
4845 0 : return Err(ApiError::ResourceUnavailable(
4846 0 : format!(
4847 0 : "Timeline is not attached to the pageserver {} yet, please retry",
4848 0 : node.get_id()
4849 0 : )
4850 0 : .into(),
4851 0 : ));
4852 : }
4853 0 : Err(e) => return Err(passthrough_api_error(&node, e)),
4854 : }
4855 : }
4856 0 : Ok(processed_results)
4857 0 : }
4858 :
4859 0 : pub(crate) async fn tenant_timeline_lsn_lease(
4860 0 : &self,
4861 0 : tenant_id: TenantId,
4862 0 : timeline_id: TimelineId,
4863 0 : lsn: Lsn,
4864 0 : ) -> Result<LsnLease, ApiError> {
4865 0 : let _tenant_lock = trace_shared_lock(
4866 0 : &self.tenant_op_locks,
4867 0 : tenant_id,
4868 0 : TenantOperations::TimelineLsnLease,
4869 0 : )
4870 0 : .await;
4871 :
4872 0 : self.tenant_remote_mutation(tenant_id, |locations| async move {
4873 0 : if locations.0.is_empty() {
4874 0 : return Err(ApiError::NotFound(
4875 0 : anyhow::anyhow!("Tenant not found").into(),
4876 0 : ));
4877 0 : }
4878 :
4879 0 : let results = self
4880 0 : .tenant_for_shards_api(
4881 0 : locations
4882 0 : .0
4883 0 : .iter()
4884 0 : .map(|(tenant_shard_id, ShardMutationLocations { latest, .. })| {
4885 0 : (*tenant_shard_id, latest.node.clone())
4886 0 : })
4887 0 : .collect(),
4888 0 : |tenant_shard_id, client| async move {
4889 0 : client
4890 0 : .timeline_lease_lsn(tenant_shard_id, timeline_id, lsn)
4891 0 : .await
4892 0 : },
4893 : 1,
4894 : 1,
4895 : SHORT_RECONCILE_TIMEOUT,
4896 0 : &self.cancel,
4897 : )
4898 0 : .await;
4899 :
4900 0 : let leases = self.process_result_and_passthrough_errors(tenant_id, results)?;
4901 0 : let mut valid_until = None;
4902 0 : for (_, lease) in leases {
4903 0 : if let Some(ref mut valid_until) = valid_until {
4904 0 : *valid_until = std::cmp::min(*valid_until, lease.valid_until);
4905 0 : } else {
4906 0 : valid_until = Some(lease.valid_until);
4907 0 : }
4908 : }
4909 0 : Ok(LsnLease {
4910 0 : valid_until: valid_until.unwrap_or_else(SystemTime::now),
4911 0 : })
4912 0 : })
4913 0 : .await?
4914 0 : }
4915 :
4916 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4917 0 : &self,
4918 0 : tenant_shard_id: TenantShardId,
4919 0 : timeline_id: TimelineId,
4920 0 : concurrency: Option<usize>,
4921 0 : recurse: bool,
4922 0 : ) -> Result<(), ApiError> {
4923 0 : let _tenant_lock = trace_shared_lock(
4924 0 : &self.tenant_op_locks,
4925 0 : tenant_shard_id.tenant_id,
4926 0 : TenantOperations::DownloadHeatmapLayers,
4927 0 : )
4928 0 : .await;
4929 :
4930 0 : let targets = {
4931 0 : let locked = self.inner.read().unwrap();
4932 0 : let mut targets = Vec::new();
4933 :
4934 : // If the request got an unsharded tenant id, then apply
4935 : // the operation to all shards. Otherwise, apply it to a specific shard.
4936 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4937 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4938 : } else {
4939 0 : tenant_shard_id.range()
4940 : };
4941 :
4942 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4943 0 : if let Some(node_id) = shard.intent.get_attached() {
4944 0 : let node = locked
4945 0 : .nodes
4946 0 : .get(node_id)
4947 0 : .expect("Pageservers may not be deleted while referenced");
4948 0 :
4949 0 : targets.push((*tenant_shard_id, node.clone()));
4950 0 : }
4951 : }
4952 0 : targets
4953 : };
4954 :
4955 0 : self.tenant_for_shards_api(
4956 0 : targets,
4957 0 : |tenant_shard_id, client| async move {
4958 0 : client
4959 0 : .timeline_download_heatmap_layers(
4960 0 : tenant_shard_id,
4961 0 : timeline_id,
4962 0 : concurrency,
4963 0 : recurse,
4964 0 : )
4965 0 : .await
4966 0 : },
4967 : 1,
4968 : 1,
4969 : SHORT_RECONCILE_TIMEOUT,
4970 0 : &self.cancel,
4971 : )
4972 0 : .await;
4973 :
4974 0 : Ok(())
4975 0 : }
4976 :
4977 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4978 : ///
4979 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`
4980 : /// and returned element at index `i` is the result for `req_fn(op(locations[i])`.
4981 0 : async fn tenant_for_shards<F, R>(
4982 0 : &self,
4983 0 : locations: Vec<(TenantShardId, Node)>,
4984 0 : mut req_fn: F,
4985 0 : ) -> Result<Vec<R>, ApiError>
4986 0 : where
4987 0 : F: FnMut(
4988 0 : TenantShardId,
4989 0 : Node,
4990 0 : )
4991 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4992 0 : {
4993 0 : let mut futs = FuturesUnordered::new();
4994 0 : let mut results = Vec::with_capacity(locations.len());
4995 :
4996 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4997 0 : let fut = req_fn(tenant_shard_id, node);
4998 0 : futs.push(async move { (idx, fut.await) });
4999 : }
5000 :
5001 0 : while let Some((idx, r)) = futs.next().await {
5002 0 : results.push((idx, r?));
5003 : }
5004 :
5005 0 : results.sort_by_key(|(idx, _)| *idx);
5006 0 : Ok(results.into_iter().map(|(_, r)| r).collect())
5007 0 : }
5008 :
5009 : /// Concurrently invoke a pageserver API call on many shards at once.
5010 : ///
5011 : /// The returned Vec has the same length as the `locations` Vec,
5012 : /// and returned element at index `i` is the result for `op(locations[i])`.
5013 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
5014 0 : &self,
5015 0 : locations: Vec<(TenantShardId, Node)>,
5016 0 : op: O,
5017 0 : warn_threshold: u32,
5018 0 : max_retries: u32,
5019 0 : timeout: Duration,
5020 0 : cancel: &CancellationToken,
5021 0 : ) -> Vec<(Node, mgmt_api::Result<T>)>
5022 0 : where
5023 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
5024 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
5025 0 : {
5026 0 : let mut futs = FuturesUnordered::new();
5027 0 : let mut results = Vec::with_capacity(locations.len());
5028 :
5029 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
5030 0 : futs.push(async move {
5031 0 : let r = node
5032 0 : .with_client_retries(
5033 0 : |client| op(tenant_shard_id, client),
5034 0 : &self.http_client,
5035 0 : &self.config.pageserver_jwt_token,
5036 0 : warn_threshold,
5037 0 : max_retries,
5038 0 : timeout,
5039 0 : cancel,
5040 : )
5041 0 : .await;
5042 0 : (idx, node, r)
5043 0 : });
5044 : }
5045 :
5046 0 : while let Some((idx, node, r)) = futs.next().await {
5047 0 : results.push((idx, node, r.unwrap_or(Err(mgmt_api::Error::Cancelled))));
5048 0 : }
5049 :
5050 0 : results.sort_by_key(|(idx, _, _)| *idx);
5051 0 : results.into_iter().map(|(_, node, r)| (node, r)).collect()
5052 0 : }
5053 :
5054 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
5055 : /// when creating and deleting timelines:
5056 : /// - Makes sure shards are attached somewhere if they weren't already
5057 : /// - Looks up the shards and the nodes where they were most recently attached
5058 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
5059 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
5060 0 : pub(crate) async fn tenant_remote_mutation<R, O, F>(
5061 0 : &self,
5062 0 : tenant_id: TenantId,
5063 0 : op: O,
5064 0 : ) -> Result<R, ApiError>
5065 0 : where
5066 0 : O: FnOnce(TenantMutationLocations) -> F,
5067 0 : F: std::future::Future<Output = R>,
5068 0 : {
5069 0 : self.tenant_remote_mutation_inner(TenantIdOrShardId::TenantId(tenant_id), op)
5070 0 : .await
5071 0 : }
5072 :
5073 0 : pub(crate) async fn tenant_shard_remote_mutation<R, O, F>(
5074 0 : &self,
5075 0 : tenant_shard_id: TenantShardId,
5076 0 : op: O,
5077 0 : ) -> Result<R, ApiError>
5078 0 : where
5079 0 : O: FnOnce(TenantMutationLocations) -> F,
5080 0 : F: std::future::Future<Output = R>,
5081 0 : {
5082 0 : self.tenant_remote_mutation_inner(TenantIdOrShardId::TenantShardId(tenant_shard_id), op)
5083 0 : .await
5084 0 : }
5085 :
5086 0 : async fn tenant_remote_mutation_inner<R, O, F>(
5087 0 : &self,
5088 0 : tenant_id_or_shard_id: TenantIdOrShardId,
5089 0 : op: O,
5090 0 : ) -> Result<R, ApiError>
5091 0 : where
5092 0 : O: FnOnce(TenantMutationLocations) -> F,
5093 0 : F: std::future::Future<Output = R>,
5094 0 : {
5095 0 : let mutation_locations = {
5096 0 : let mut locations = TenantMutationLocations::default();
5097 :
5098 : // Load the currently attached pageservers for the latest generation of each shard. This can
5099 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
5100 : // will still be the latest when we're done: we will check generations again at the end of
5101 : // this function to handle that.
5102 0 : let generations = self
5103 0 : .persistence
5104 0 : .tenant_generations(tenant_id_or_shard_id.tenant_id())
5105 0 : .await?
5106 0 : .into_iter()
5107 0 : .filter(|i| tenant_id_or_shard_id.matches(&i.tenant_shard_id))
5108 0 : .collect::<Vec<_>>();
5109 :
5110 0 : if generations
5111 0 : .iter()
5112 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
5113 : {
5114 0 : let shard_generations = generations
5115 0 : .into_iter()
5116 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
5117 0 : .collect::<HashMap<_, _>>();
5118 :
5119 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
5120 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
5121 0 : let locked = self.inner.read().unwrap();
5122 0 : let tenant_shards = locked
5123 0 : .tenants
5124 0 : .range(TenantShardId::tenant_range(
5125 0 : tenant_id_or_shard_id.tenant_id(),
5126 : ))
5127 0 : .filter(|(shard_id, _)| tenant_id_or_shard_id.matches(shard_id))
5128 0 : .collect::<Vec<_>>();
5129 0 : for (shard_id, shard) in tenant_shards {
5130 0 : match shard.policy {
5131 : PlacementPolicy::Attached(_) => {
5132 : // This shard is meant to be attached: the caller is not wrong to try and
5133 : // use this function, but we can't service the request right now.
5134 0 : let Some(generation) = shard_generations.get(shard_id) else {
5135 : // This can only happen if there is a split brain controller modifying the database. This should
5136 : // never happen when testing, and if it happens in production we can only log the issue.
5137 0 : debug_assert!(false);
5138 0 : tracing::error!(
5139 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
5140 : );
5141 0 : continue;
5142 : };
5143 0 : let (generation, generation_pageserver) = generation;
5144 0 : if let Some(generation) = generation {
5145 0 : if generation_pageserver.is_none() {
5146 : // This is legitimate only in a very narrow window where the shard was only just configured into
5147 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
5148 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
5149 0 : tracing::warn!(
5150 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
5151 : );
5152 0 : }
5153 : } else {
5154 : // This should never happen: a shard with no generation is only permitted when it was created in some state
5155 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
5156 0 : debug_assert!(false);
5157 0 : tracing::error!(
5158 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
5159 : );
5160 0 : continue;
5161 : }
5162 : }
5163 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
5164 0 : return Err(ApiError::Conflict(format!(
5165 0 : "Shard {shard_id} tenant has policy {:?}",
5166 0 : shard.policy
5167 0 : )));
5168 : }
5169 : }
5170 : }
5171 :
5172 0 : return Err(ApiError::ResourceUnavailable(
5173 0 : "One or more shards in tenant is not yet attached".into(),
5174 0 : ));
5175 0 : }
5176 :
5177 0 : let locked = self.inner.read().unwrap();
5178 : for ShardGenerationState {
5179 0 : tenant_shard_id,
5180 0 : generation,
5181 0 : generation_pageserver,
5182 0 : } in generations
5183 : {
5184 0 : let node_id = generation_pageserver.expect("We checked for None above");
5185 0 : let node = locked
5186 0 : .nodes
5187 0 : .get(&node_id)
5188 0 : .ok_or(ApiError::Conflict(format!(
5189 0 : "Raced with removal of node {node_id}"
5190 0 : )))?;
5191 0 : let generation = generation.expect("Checked above");
5192 :
5193 0 : let tenant = locked.tenants.get(&tenant_shard_id);
5194 :
5195 : // TODO(vlad): Abstract the logic that finds stale attached locations
5196 : // from observed state into a [`Service`] method.
5197 0 : let other_locations = match tenant {
5198 0 : Some(tenant) => {
5199 0 : let mut other = tenant.attached_locations();
5200 0 : let latest_location_index =
5201 0 : other.iter().position(|&l| l == (node.get_id(), generation));
5202 0 : if let Some(idx) = latest_location_index {
5203 0 : other.remove(idx);
5204 0 : }
5205 :
5206 0 : other
5207 : }
5208 0 : None => Vec::default(),
5209 : };
5210 :
5211 0 : let location = ShardMutationLocations {
5212 0 : latest: MutationLocation {
5213 0 : node: node.clone(),
5214 0 : generation,
5215 0 : },
5216 0 : other: other_locations
5217 0 : .into_iter()
5218 0 : .filter_map(|(node_id, generation)| {
5219 0 : let node = locked.nodes.get(&node_id)?;
5220 :
5221 0 : Some(MutationLocation {
5222 0 : node: node.clone(),
5223 0 : generation,
5224 0 : })
5225 0 : })
5226 0 : .collect(),
5227 : };
5228 0 : locations.0.insert(tenant_shard_id, location);
5229 : }
5230 :
5231 0 : locations
5232 : };
5233 :
5234 0 : let result = op(mutation_locations.clone()).await;
5235 :
5236 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
5237 : // our remote operation executed on the latest generation and is therefore persistent.
5238 : {
5239 0 : let latest_generations = self
5240 0 : .persistence
5241 0 : .tenant_generations(tenant_id_or_shard_id.tenant_id())
5242 0 : .await?
5243 0 : .into_iter()
5244 0 : .filter(|i| tenant_id_or_shard_id.matches(&i.tenant_shard_id))
5245 0 : .collect::<Vec<_>>();
5246 :
5247 0 : if latest_generations
5248 0 : .into_iter()
5249 0 : .map(
5250 : |ShardGenerationState {
5251 : tenant_shard_id,
5252 : generation,
5253 : generation_pageserver: _,
5254 0 : }| (tenant_shard_id, generation),
5255 : )
5256 0 : .collect::<Vec<_>>()
5257 0 : != mutation_locations
5258 0 : .0
5259 0 : .into_iter()
5260 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
5261 0 : .collect::<Vec<_>>()
5262 : {
5263 : // We raced with something that incremented the generation, and therefore cannot be
5264 : // confident that our actions are persistent (they might have hit an old generation).
5265 : //
5266 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
5267 0 : return Err(ApiError::ResourceUnavailable(
5268 0 : "Tenant attachment changed, please retry".into(),
5269 0 : ));
5270 0 : }
5271 : }
5272 :
5273 0 : Ok(result)
5274 0 : }
5275 :
5276 0 : pub(crate) async fn tenant_timeline_delete(
5277 0 : self: &Arc<Self>,
5278 0 : tenant_id: TenantId,
5279 0 : timeline_id: TimelineId,
5280 0 : ) -> Result<StatusCode, ApiError> {
5281 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
5282 0 : let _tenant_lock = trace_shared_lock(
5283 0 : &self.tenant_op_locks,
5284 0 : tenant_id,
5285 0 : TenantOperations::TimelineDelete,
5286 0 : )
5287 0 : .await;
5288 :
5289 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
5290 0 : if targets.0.is_empty() {
5291 0 : return Err(ApiError::NotFound(
5292 0 : anyhow::anyhow!("Tenant not found").into(),
5293 0 : ));
5294 0 : }
5295 :
5296 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
5297 0 : assert!(shard_zero_tid.is_shard_zero());
5298 :
5299 0 : async fn delete_one(
5300 0 : tenant_shard_id: TenantShardId,
5301 0 : timeline_id: TimelineId,
5302 0 : node: Node,
5303 0 : http_client: reqwest::Client,
5304 0 : jwt: Option<String>,
5305 0 : ) -> Result<StatusCode, ApiError> {
5306 0 : tracing::info!(
5307 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
5308 : );
5309 :
5310 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
5311 0 : let res = client
5312 0 : .timeline_delete(tenant_shard_id, timeline_id)
5313 0 : .await;
5314 :
5315 0 : match res {
5316 0 : Ok(ok) => Ok(ok),
5317 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
5318 0 : Err(mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg)) if msg.contains("Requested tenant is missing") => {
5319 0 : Err(ApiError::ResourceUnavailable("Tenant migration in progress".into()))
5320 : },
5321 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
5322 0 : Err(e) => {
5323 0 : Err(
5324 0 : ApiError::InternalServerError(anyhow::anyhow!(
5325 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
5326 0 : ))
5327 0 : )
5328 : }
5329 : }
5330 0 : }
5331 :
5332 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
5333 0 : let statuses = self
5334 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
5335 0 : Box::pin(delete_one(
5336 0 : tenant_shard_id,
5337 0 : timeline_id,
5338 0 : node,
5339 0 : self.http_client.clone(),
5340 0 : self.config.pageserver_jwt_token.clone(),
5341 0 : ))
5342 0 : })
5343 0 : .await?;
5344 :
5345 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
5346 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
5347 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
5348 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
5349 0 : return Ok(StatusCode::CONFLICT);
5350 0 : }
5351 :
5352 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
5353 0 : return Ok(StatusCode::ACCEPTED);
5354 0 : }
5355 :
5356 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
5357 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
5358 0 : let shard_zero_status = delete_one(
5359 0 : shard_zero_tid,
5360 0 : timeline_id,
5361 0 : shard_zero_locations.latest.node,
5362 0 : self.http_client.clone(),
5363 0 : self.config.pageserver_jwt_token.clone(),
5364 0 : )
5365 0 : .await?;
5366 0 : Ok(shard_zero_status)
5367 0 : }).await?;
5368 :
5369 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
5370 0 : .await?;
5371 :
5372 0 : status_code
5373 0 : }
5374 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
5375 : ///
5376 : /// Returns the node, tenant shard id, and whether it is consistent with the observed state.
5377 0 : pub(crate) async fn tenant_shard0_node(
5378 0 : &self,
5379 0 : tenant_id: TenantId,
5380 0 : ) -> Result<(Node, TenantShardId), ApiError> {
5381 0 : let tenant_shard_id = {
5382 0 : let locked = self.inner.read().unwrap();
5383 0 : let Some((tenant_shard_id, _shard)) = locked
5384 0 : .tenants
5385 0 : .range(TenantShardId::tenant_range(tenant_id))
5386 0 : .next()
5387 : else {
5388 0 : return Err(ApiError::NotFound(
5389 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
5390 0 : ));
5391 : };
5392 :
5393 0 : *tenant_shard_id
5394 : };
5395 :
5396 0 : self.tenant_shard_node(tenant_shard_id)
5397 0 : .await
5398 0 : .map(|node| (node, tenant_shard_id))
5399 0 : }
5400 :
5401 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
5402 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
5403 : ///
5404 : /// Returns the intent node and whether it is consistent with the observed state.
5405 0 : pub(crate) async fn tenant_shard_node(
5406 0 : &self,
5407 0 : tenant_shard_id: TenantShardId,
5408 0 : ) -> Result<Node, ApiError> {
5409 : // Look up in-memory state and maybe use the node from there.
5410 : {
5411 0 : let locked = self.inner.read().unwrap();
5412 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5413 0 : return Err(ApiError::NotFound(
5414 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
5415 0 : ));
5416 : };
5417 :
5418 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
5419 0 : tracing::warn!(
5420 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
5421 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
5422 : shard.policy
5423 : );
5424 0 : return Err(ApiError::Conflict(
5425 0 : "Cannot call timeline API on non-attached tenant".to_string(),
5426 0 : ));
5427 : };
5428 :
5429 0 : if shard.reconciler.is_none() {
5430 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
5431 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
5432 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
5433 : // This should never happen
5434 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5435 0 : "Shard refers to nonexistent node"
5436 0 : )));
5437 : };
5438 0 : return Ok(node.clone());
5439 0 : }
5440 : };
5441 :
5442 : // Look up the latest attached pageserver location from the database
5443 : // generation state: this will reflect the progress of any ongoing migration.
5444 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
5445 : // the case where they call through to the pageserver and get a 404.
5446 0 : let db_result = self
5447 0 : .persistence
5448 0 : .tenant_generations(tenant_shard_id.tenant_id)
5449 0 : .await?;
5450 : let Some(ShardGenerationState {
5451 : tenant_shard_id: _,
5452 : generation: _,
5453 0 : generation_pageserver: Some(node_id),
5454 0 : }) = db_result
5455 0 : .into_iter()
5456 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
5457 : else {
5458 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
5459 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
5460 : // or a conflict response (case where tenant was detached in background)
5461 0 : return Err(ApiError::ResourceUnavailable(
5462 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
5463 0 : ));
5464 : };
5465 0 : let locked = self.inner.read().unwrap();
5466 0 : let Some(node) = locked.nodes.get(&node_id) else {
5467 : // This should never happen
5468 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5469 0 : "Shard refers to nonexistent node"
5470 0 : )));
5471 : };
5472 : // As a reconciliation is in flight, we do not have the observed state yet, and therefore we assume it is always inconsistent.
5473 0 : Ok(node.clone())
5474 0 : }
5475 :
5476 0 : pub(crate) fn tenant_locate(
5477 0 : &self,
5478 0 : tenant_id: TenantId,
5479 0 : ) -> Result<TenantLocateResponse, ApiError> {
5480 0 : let locked = self.inner.read().unwrap();
5481 0 : tracing::info!("Locating shards for tenant {tenant_id}");
5482 :
5483 0 : let mut result = Vec::new();
5484 0 : let mut shard_params: Option<ShardParameters> = None;
5485 :
5486 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5487 : {
5488 0 : let node_id =
5489 0 : shard
5490 0 : .intent
5491 0 : .get_attached()
5492 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
5493 0 : "Cannot locate a tenant that is not attached"
5494 0 : )))?;
5495 :
5496 0 : let node = locked
5497 0 : .nodes
5498 0 : .get(&node_id)
5499 0 : .expect("Pageservers may not be deleted while referenced");
5500 :
5501 0 : result.push(node.shard_location(*tenant_shard_id));
5502 :
5503 0 : match &shard_params {
5504 0 : None => {
5505 0 : shard_params = Some(ShardParameters {
5506 0 : stripe_size: shard.shard.stripe_size,
5507 0 : count: shard.shard.count,
5508 0 : });
5509 0 : }
5510 0 : Some(params) => {
5511 0 : if params.stripe_size != shard.shard.stripe_size {
5512 : // This should never happen. We enforce at runtime because it's simpler than
5513 : // adding an extra per-tenant data structure to store the things that should be the same
5514 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5515 0 : "Inconsistent shard stripe size parameters!"
5516 0 : )));
5517 0 : }
5518 : }
5519 : }
5520 : }
5521 :
5522 0 : if result.is_empty() {
5523 0 : return Err(ApiError::NotFound(
5524 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
5525 0 : ));
5526 0 : }
5527 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
5528 0 : tracing::info!(
5529 0 : "Located tenant {} with params {:?} on shards {}",
5530 : tenant_id,
5531 : shard_params,
5532 0 : result
5533 0 : .iter()
5534 0 : .map(|s| format!("{s:?}"))
5535 0 : .collect::<Vec<_>>()
5536 0 : .join(",")
5537 : );
5538 :
5539 0 : Ok(TenantLocateResponse {
5540 0 : shards: result,
5541 0 : shard_params,
5542 0 : })
5543 0 : }
5544 :
5545 : /// Returns None if the input iterator of shards does not include a shard with number=0
5546 0 : fn tenant_describe_impl<'a>(
5547 0 : &self,
5548 0 : shards: impl Iterator<Item = &'a TenantShard>,
5549 0 : ) -> Option<TenantDescribeResponse> {
5550 0 : let mut shard_zero = None;
5551 0 : let mut describe_shards = Vec::new();
5552 :
5553 0 : for shard in shards {
5554 0 : if shard.tenant_shard_id.is_shard_zero() {
5555 0 : shard_zero = Some(shard);
5556 0 : }
5557 :
5558 0 : describe_shards.push(TenantDescribeResponseShard {
5559 0 : tenant_shard_id: shard.tenant_shard_id,
5560 0 : node_attached: *shard.intent.get_attached(),
5561 0 : node_secondary: shard.intent.get_secondary().to_vec(),
5562 0 : last_error: shard
5563 0 : .last_error
5564 0 : .lock()
5565 0 : .unwrap()
5566 0 : .as_ref()
5567 0 : .map(|e| format!("{e}"))
5568 0 : .unwrap_or("".to_string())
5569 0 : .clone(),
5570 0 : is_reconciling: shard.reconciler.is_some(),
5571 0 : is_pending_compute_notification: shard.pending_compute_notification,
5572 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
5573 0 : is_importing: shard.importing == TimelineImportState::Importing,
5574 0 : scheduling_policy: shard.get_scheduling_policy(),
5575 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
5576 : })
5577 : }
5578 :
5579 0 : let shard_zero = shard_zero?;
5580 :
5581 0 : Some(TenantDescribeResponse {
5582 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
5583 0 : shards: describe_shards,
5584 0 : stripe_size: shard_zero.shard.stripe_size,
5585 0 : policy: shard_zero.policy.clone(),
5586 0 : config: shard_zero.config.clone(),
5587 0 : })
5588 0 : }
5589 :
5590 0 : pub(crate) fn tenant_describe(
5591 0 : &self,
5592 0 : tenant_id: TenantId,
5593 0 : ) -> Result<TenantDescribeResponse, ApiError> {
5594 0 : let locked = self.inner.read().unwrap();
5595 :
5596 0 : self.tenant_describe_impl(
5597 0 : locked
5598 0 : .tenants
5599 0 : .range(TenantShardId::tenant_range(tenant_id))
5600 0 : .map(|(_k, v)| v),
5601 : )
5602 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
5603 0 : }
5604 :
5605 : /* BEGIN_HADRON */
5606 0 : pub(crate) async fn tenant_timeline_describe(
5607 0 : &self,
5608 0 : tenant_id: TenantId,
5609 0 : timeline_id: TimelineId,
5610 0 : ) -> Result<TenantTimelineDescribeResponse, ApiError> {
5611 0 : self.tenant_remote_mutation(tenant_id, |locations| async move {
5612 0 : if locations.0.is_empty() {
5613 0 : return Err(ApiError::NotFound(
5614 0 : anyhow::anyhow!("Tenant not found").into(),
5615 0 : ));
5616 0 : };
5617 :
5618 0 : let locations: Vec<(TenantShardId, Node)> = locations
5619 0 : .0
5620 0 : .iter()
5621 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
5622 0 : .collect();
5623 0 : let mut futs = FuturesUnordered::new();
5624 :
5625 0 : for (shard_id, node) in locations {
5626 0 : futs.push({
5627 0 : async move {
5628 0 : let result = node
5629 0 : .with_client_retries(
5630 0 : |client| async move {
5631 0 : client
5632 0 : .tenant_timeline_describe(&shard_id, &timeline_id)
5633 0 : .await
5634 0 : },
5635 0 : &self.http_client,
5636 0 : &self.config.pageserver_jwt_token,
5637 : 3,
5638 : 3,
5639 0 : Duration::from_secs(30),
5640 0 : &self.cancel,
5641 : )
5642 0 : .await;
5643 0 : (result, shard_id, node.get_id())
5644 0 : }
5645 : });
5646 : }
5647 :
5648 0 : let mut results: Vec<TimelineInfo> = Vec::new();
5649 0 : while let Some((result, tenant_shard_id, node_id)) = futs.next().await {
5650 0 : match result {
5651 0 : Some(Ok(timeline_info)) => results.push(timeline_info),
5652 0 : Some(Err(e)) => {
5653 0 : tracing::warn!(
5654 0 : "Failed to describe tenant {} timeline {} for pageserver {}: {e}",
5655 : tenant_shard_id,
5656 : timeline_id,
5657 : node_id,
5658 : );
5659 0 : return Err(ApiError::ResourceUnavailable(format!("{e}").into()));
5660 : }
5661 0 : None => return Err(ApiError::Cancelled),
5662 : }
5663 : }
5664 0 : let mut image_consistent_lsn: Option<Lsn> = Some(Lsn::MAX);
5665 0 : for timeline_info in &results {
5666 0 : if let Some(tline_image_consistent_lsn) = timeline_info.image_consistent_lsn {
5667 0 : image_consistent_lsn = Some(std::cmp::min(
5668 0 : image_consistent_lsn.unwrap(),
5669 0 : tline_image_consistent_lsn,
5670 0 : ));
5671 0 : } else {
5672 0 : tracing::warn!(
5673 0 : "Timeline {} on shard {} does not have image consistent lsn",
5674 : timeline_info.timeline_id,
5675 : timeline_info.tenant_id
5676 : );
5677 0 : image_consistent_lsn = None;
5678 0 : break;
5679 : }
5680 : }
5681 :
5682 0 : Ok(TenantTimelineDescribeResponse {
5683 0 : shards: results,
5684 0 : image_consistent_lsn,
5685 0 : })
5686 0 : })
5687 0 : .await?
5688 0 : }
5689 : /* END_HADRON */
5690 :
5691 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
5692 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
5693 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
5694 : /// in our external API.
5695 0 : pub(crate) fn tenant_list(
5696 0 : &self,
5697 0 : limit: Option<usize>,
5698 0 : start_after: Option<TenantId>,
5699 0 : ) -> Vec<TenantDescribeResponse> {
5700 0 : let locked = self.inner.read().unwrap();
5701 :
5702 : // Apply start_from parameter
5703 0 : let shard_range = match start_after {
5704 0 : None => locked.tenants.range(..),
5705 0 : Some(tenant_id) => locked.tenants.range(
5706 0 : TenantShardId {
5707 0 : tenant_id,
5708 0 : shard_number: ShardNumber(u8::MAX),
5709 0 : shard_count: ShardCount(u8::MAX),
5710 0 : }..,
5711 : ),
5712 : };
5713 :
5714 0 : let mut result = Vec::new();
5715 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
5716 0 : result.push(
5717 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
5718 0 : .expect("Groups are always non-empty"),
5719 : );
5720 :
5721 : // Enforce `limit` parameter
5722 0 : if let Some(limit) = limit {
5723 0 : if result.len() >= limit {
5724 0 : break;
5725 0 : }
5726 0 : }
5727 : }
5728 :
5729 0 : result
5730 0 : }
5731 :
5732 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
5733 : async fn abort_tenant_shard_split(
5734 : &self,
5735 : op: &TenantShardSplitAbort,
5736 : ) -> Result<(), TenantShardSplitAbortError> {
5737 : // Cleaning up a split:
5738 : // - Parent shards are not destroyed during a split, just detached.
5739 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
5740 : // just the children attached, or both.
5741 : //
5742 : // Therefore our work to do is to:
5743 : // 1. Clean up storage controller's internal state to just refer to parents, no children
5744 : // 2. Call out to pageservers to ensure that children are detached
5745 : // 3. Call out to pageservers to ensure that parents are attached.
5746 : //
5747 : // Crash safety:
5748 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
5749 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
5750 : // and detach them.
5751 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
5752 : // from our database, then we will re-enter this cleanup routine on startup.
5753 :
5754 : let TenantShardSplitAbort {
5755 : tenant_id,
5756 : new_shard_count,
5757 : new_stripe_size,
5758 : ..
5759 : } = op;
5760 :
5761 : // First abort persistent state, if any exists.
5762 : match self
5763 : .persistence
5764 : .abort_shard_split(*tenant_id, *new_shard_count)
5765 : .await?
5766 : {
5767 : AbortShardSplitStatus::Aborted => {
5768 : // Proceed to roll back any child shards created on pageservers
5769 : }
5770 : AbortShardSplitStatus::Complete => {
5771 : // The split completed (we might hit that path if e.g. our database transaction
5772 : // to write the completion landed in the database, but we dropped connection
5773 : // before seeing the result).
5774 : //
5775 : // We must update in-memory state to reflect the successful split.
5776 : self.tenant_shard_split_commit_inmem(
5777 : *tenant_id,
5778 : *new_shard_count,
5779 : *new_stripe_size,
5780 : );
5781 : return Ok(());
5782 : }
5783 : }
5784 :
5785 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
5786 : let detach_locations: Vec<(Node, TenantShardId)> = {
5787 : let mut detach_locations = Vec::new();
5788 : let mut locked = self.inner.write().unwrap();
5789 : let (nodes, tenants, scheduler) = locked.parts_mut();
5790 :
5791 : for (tenant_shard_id, shard) in
5792 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
5793 : {
5794 : if shard.shard.count == op.new_shard_count {
5795 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
5796 : // is infallible, so if we got an error we shouldn't have got that far.
5797 : tracing::warn!(
5798 : "During split abort, child shard {tenant_shard_id} found in-memory"
5799 : );
5800 : continue;
5801 : }
5802 :
5803 : // Add the children of this shard to this list of things to detach
5804 : if let Some(node_id) = shard.intent.get_attached() {
5805 : for child_id in tenant_shard_id.split(*new_shard_count) {
5806 : detach_locations.push((
5807 : nodes
5808 : .get(node_id)
5809 : .expect("Intent references nonexistent node")
5810 : .clone(),
5811 : child_id,
5812 : ));
5813 : }
5814 : } else {
5815 : tracing::warn!(
5816 : "During split abort, shard {tenant_shard_id} has no attached location"
5817 : );
5818 : }
5819 :
5820 : tracing::info!("Restoring parent shard {tenant_shard_id}");
5821 :
5822 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
5823 : // if the original attachment location is offline.
5824 : if let Some(node_id) = shard.intent.get_attached() {
5825 : if !nodes.get(node_id).unwrap().is_available() {
5826 : tracing::info!(
5827 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
5828 : );
5829 : shard.intent.demote_attached(scheduler, *node_id);
5830 : }
5831 : }
5832 : for node_id in shard.intent.get_secondary().clone() {
5833 : if !nodes.get(&node_id).unwrap().is_available() {
5834 : tracing::info!(
5835 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
5836 : );
5837 : shard.intent.remove_secondary(scheduler, node_id);
5838 : }
5839 : }
5840 :
5841 : shard.splitting = SplitState::Idle;
5842 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
5843 : // If this shard can't be scheduled now (perhaps due to offline nodes or
5844 : // capacity issues), that must not prevent us rolling back a split. In this
5845 : // case it should be eventually scheduled in the background.
5846 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
5847 : }
5848 :
5849 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
5850 : }
5851 :
5852 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
5853 : tenants
5854 0 : .retain(|id, s| !(id.tenant_id == *tenant_id && s.shard.count == *new_shard_count));
5855 :
5856 : detach_locations
5857 : };
5858 :
5859 : for (node, child_id) in detach_locations {
5860 : if !node.is_available() {
5861 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
5862 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
5863 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
5864 : // them from the node.
5865 : tracing::warn!(
5866 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
5867 : );
5868 : continue;
5869 : }
5870 :
5871 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
5872 : // a 503 and retry, up to our limit.
5873 : tracing::info!("Detaching {child_id} on {node}...");
5874 : match node
5875 : .with_client_retries(
5876 0 : |client| async move {
5877 0 : let config = LocationConfig {
5878 0 : mode: LocationConfigMode::Detached,
5879 0 : generation: None,
5880 0 : secondary_conf: None,
5881 0 : shard_number: child_id.shard_number.0,
5882 0 : shard_count: child_id.shard_count.literal(),
5883 0 : // Stripe size and tenant config don't matter when detaching
5884 0 : shard_stripe_size: 0,
5885 0 : tenant_conf: TenantConfig::default(),
5886 0 : };
5887 :
5888 0 : client.location_config(child_id, config, None, false).await
5889 0 : },
5890 : &self.http_client,
5891 : &self.config.pageserver_jwt_token,
5892 : 1,
5893 : 10,
5894 : Duration::from_secs(5),
5895 : &self.reconcilers_cancel,
5896 : )
5897 : .await
5898 : {
5899 : Some(Ok(_)) => {}
5900 : Some(Err(e)) => {
5901 : // We failed to communicate with the remote node. This is problematic: we may be
5902 : // leaving it with a rogue child shard.
5903 : tracing::warn!(
5904 : "Failed to detach child {child_id} from node {node} during abort"
5905 : );
5906 : return Err(e.into());
5907 : }
5908 : None => {
5909 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
5910 : // clean up on restart. The node going offline requires a retry.
5911 : return Err(TenantShardSplitAbortError::Unavailable);
5912 : }
5913 : };
5914 : }
5915 :
5916 : tracing::info!("Successfully aborted split");
5917 : Ok(())
5918 : }
5919 :
5920 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
5921 : /// of the tenant map to reflect the child shards that exist after the split.
5922 0 : fn tenant_shard_split_commit_inmem(
5923 0 : &self,
5924 0 : tenant_id: TenantId,
5925 0 : new_shard_count: ShardCount,
5926 0 : new_stripe_size: Option<ShardStripeSize>,
5927 0 : ) -> (
5928 0 : TenantShardSplitResponse,
5929 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
5930 0 : Vec<ReconcilerWaiter>,
5931 0 : ) {
5932 0 : let mut response = TenantShardSplitResponse {
5933 0 : new_shards: Vec::new(),
5934 0 : };
5935 0 : let mut child_locations = Vec::new();
5936 0 : let mut waiters = Vec::new();
5937 :
5938 : {
5939 0 : let mut locked = self.inner.write().unwrap();
5940 :
5941 0 : let parent_ids = locked
5942 0 : .tenants
5943 0 : .range(TenantShardId::tenant_range(tenant_id))
5944 0 : .map(|(shard_id, _)| *shard_id)
5945 0 : .collect::<Vec<_>>();
5946 :
5947 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5948 0 : for parent_id in parent_ids {
5949 0 : let child_ids = parent_id.split(new_shard_count);
5950 :
5951 : let (
5952 0 : pageserver,
5953 0 : generation,
5954 0 : policy,
5955 0 : parent_ident,
5956 0 : config,
5957 0 : preferred_az,
5958 0 : secondary_count,
5959 : ) = {
5960 0 : let mut old_state = tenants
5961 0 : .remove(&parent_id)
5962 0 : .expect("It was present, we just split it");
5963 :
5964 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
5965 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
5966 : // nothing else can clear this.
5967 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
5968 :
5969 0 : let old_attached = old_state.intent.get_attached().unwrap();
5970 0 : old_state.intent.clear(scheduler);
5971 0 : let generation = old_state.generation.expect("Shard must have been attached");
5972 0 : (
5973 0 : old_attached,
5974 0 : generation,
5975 0 : old_state.policy.clone(),
5976 0 : old_state.shard,
5977 0 : old_state.config.clone(),
5978 0 : old_state.preferred_az().cloned(),
5979 0 : old_state.intent.get_secondary().len(),
5980 0 : )
5981 : };
5982 :
5983 0 : let mut schedule_context = ScheduleContext::default();
5984 0 : for child in child_ids {
5985 0 : let mut child_shard = parent_ident;
5986 0 : child_shard.number = child.shard_number;
5987 0 : child_shard.count = child.shard_count;
5988 0 : if let Some(stripe_size) = new_stripe_size {
5989 0 : child_shard.stripe_size = stripe_size;
5990 0 : }
5991 :
5992 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5993 0 : child_observed.insert(
5994 0 : pageserver,
5995 0 : ObservedStateLocation {
5996 0 : conf: Some(attached_location_conf(
5997 0 : generation,
5998 0 : &child_shard,
5999 0 : &config,
6000 0 : &policy,
6001 0 : secondary_count,
6002 0 : )),
6003 0 : },
6004 : );
6005 :
6006 0 : let mut child_state =
6007 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
6008 0 : child_state.intent =
6009 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
6010 0 : child_state.observed = ObservedState {
6011 0 : locations: child_observed,
6012 0 : };
6013 0 : child_state.generation = Some(generation);
6014 0 : child_state.config = config.clone();
6015 :
6016 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
6017 : // as at this point in the split process we have succeeded and this part is infallible:
6018 : // we will never need to do any special recovery from this state.
6019 :
6020 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
6021 :
6022 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
6023 : // This is not fatal, because we've implicitly already got an attached
6024 : // location for the child shard. Failure here just means we couldn't
6025 : // find a secondary (e.g. because cluster is overloaded).
6026 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
6027 0 : }
6028 : // In the background, attach secondary locations for the new shards
6029 0 : if let Some(waiter) = self.maybe_reconcile_shard(
6030 0 : &mut child_state,
6031 0 : nodes,
6032 0 : ReconcilerPriority::High,
6033 0 : ) {
6034 0 : waiters.push(waiter);
6035 0 : }
6036 :
6037 0 : tenants.insert(child, child_state);
6038 0 : response.new_shards.push(child);
6039 : }
6040 : }
6041 0 : (response, child_locations, waiters)
6042 : }
6043 0 : }
6044 :
6045 0 : async fn tenant_shard_split_start_secondaries(
6046 0 : &self,
6047 0 : tenant_id: TenantId,
6048 0 : waiters: Vec<ReconcilerWaiter>,
6049 0 : ) {
6050 : // Wait for initial reconcile of child shards, this creates the secondary locations
6051 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
6052 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
6053 : // their secondaries couldn't be attached.
6054 0 : tracing::warn!("Failed to reconcile after split: {e}");
6055 0 : return;
6056 0 : }
6057 :
6058 : // Take the state lock to discover the attached & secondary intents for all shards
6059 0 : let (attached, secondary) = {
6060 0 : let locked = self.inner.read().unwrap();
6061 0 : let mut attached = Vec::new();
6062 0 : let mut secondary = Vec::new();
6063 :
6064 0 : for (tenant_shard_id, shard) in
6065 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
6066 : {
6067 0 : let Some(node_id) = shard.intent.get_attached() else {
6068 : // Unexpected. Race with a PlacementPolicy change?
6069 0 : tracing::warn!(
6070 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
6071 : );
6072 0 : continue;
6073 : };
6074 :
6075 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
6076 : // No secondary location. Nothing for us to do.
6077 0 : continue;
6078 : };
6079 :
6080 0 : let attached_node = locked
6081 0 : .nodes
6082 0 : .get(node_id)
6083 0 : .expect("Pageservers may not be deleted while referenced");
6084 :
6085 0 : let secondary_node = locked
6086 0 : .nodes
6087 0 : .get(secondary_node_id)
6088 0 : .expect("Pageservers may not be deleted while referenced");
6089 :
6090 0 : attached.push((*tenant_shard_id, attached_node.clone()));
6091 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
6092 : }
6093 0 : (attached, secondary)
6094 : };
6095 :
6096 0 : if secondary.is_empty() {
6097 : // No secondary locations; nothing for us to do
6098 0 : return;
6099 0 : }
6100 :
6101 0 : for (_, result) in self
6102 0 : .tenant_for_shards_api(
6103 0 : attached,
6104 0 : |tenant_shard_id, client| async move {
6105 0 : client.tenant_heatmap_upload(tenant_shard_id).await
6106 0 : },
6107 : 1,
6108 : 1,
6109 : SHORT_RECONCILE_TIMEOUT,
6110 0 : &self.cancel,
6111 : )
6112 0 : .await
6113 : {
6114 0 : if let Err(e) = result {
6115 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
6116 0 : return;
6117 0 : }
6118 : }
6119 :
6120 0 : for (_, result) in self
6121 0 : .tenant_for_shards_api(
6122 0 : secondary,
6123 0 : |tenant_shard_id, client| async move {
6124 0 : client
6125 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
6126 0 : .await
6127 0 : },
6128 : 1,
6129 : 1,
6130 : SHORT_RECONCILE_TIMEOUT,
6131 0 : &self.cancel,
6132 : )
6133 0 : .await
6134 : {
6135 0 : if let Err(e) = result {
6136 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
6137 0 : return;
6138 0 : }
6139 : }
6140 0 : }
6141 :
6142 0 : pub(crate) async fn tenant_shard_split(
6143 0 : &self,
6144 0 : tenant_id: TenantId,
6145 0 : split_req: TenantShardSplitRequest,
6146 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
6147 : // TODO: return 503 if we get stuck waiting for this lock
6148 : // (issue https://github.com/neondatabase/neon/issues/7108)
6149 0 : let _tenant_lock = trace_exclusive_lock(
6150 0 : &self.tenant_op_locks,
6151 0 : tenant_id,
6152 0 : TenantOperations::ShardSplit,
6153 0 : )
6154 0 : .await;
6155 :
6156 0 : let _gate = self
6157 0 : .reconcilers_gate
6158 0 : .enter()
6159 0 : .map_err(|_| ApiError::ShuttingDown)?;
6160 :
6161 : // Timeline imports on the pageserver side can't handle shard-splits.
6162 : // If the tenant is importing a timeline, dont't shard split it.
6163 0 : match self
6164 0 : .persistence
6165 0 : .is_tenant_importing_timeline(tenant_id)
6166 0 : .await
6167 : {
6168 0 : Ok(importing) => {
6169 0 : if importing {
6170 0 : return Err(ApiError::Conflict(
6171 0 : "Cannot shard split during timeline import".to_string(),
6172 0 : ));
6173 0 : }
6174 : }
6175 0 : Err(err) => {
6176 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6177 0 : "Failed to check for running imports: {err}"
6178 0 : )));
6179 : }
6180 : }
6181 :
6182 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
6183 0 : let new_stripe_size = split_req.new_stripe_size;
6184 :
6185 : // Validate the request and construct parameters. This phase is fallible, but does not require
6186 : // rollback on errors, as it does no I/O and mutates no state.
6187 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
6188 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
6189 0 : ShardSplitAction::Split(params) => params,
6190 : };
6191 :
6192 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
6193 : // we must roll back.
6194 0 : let r = self
6195 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
6196 0 : .await;
6197 :
6198 0 : let (response, waiters) = match r {
6199 0 : Ok(r) => r,
6200 0 : Err(e) => {
6201 : // Split might be part-done, we must do work to abort it.
6202 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
6203 0 : self.abort_tx
6204 0 : .send(TenantShardSplitAbort {
6205 0 : tenant_id,
6206 0 : new_shard_count,
6207 0 : new_stripe_size,
6208 0 : _tenant_lock,
6209 0 : _gate,
6210 0 : })
6211 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
6212 0 : .ok();
6213 0 : return Err(e);
6214 : }
6215 : };
6216 :
6217 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
6218 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
6219 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
6220 : // in [`Self::optimize_all`]
6221 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
6222 0 : .await;
6223 0 : Ok(response)
6224 0 : }
6225 :
6226 0 : fn prepare_tenant_shard_split(
6227 0 : &self,
6228 0 : tenant_id: TenantId,
6229 0 : split_req: TenantShardSplitRequest,
6230 0 : ) -> Result<ShardSplitAction, ApiError> {
6231 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
6232 0 : anyhow::anyhow!("failpoint")
6233 0 : )));
6234 :
6235 0 : let mut policy = None;
6236 0 : let mut config = None;
6237 0 : let mut shard_ident = None;
6238 0 : let mut preferred_az_id = None;
6239 : // Validate input, and calculate which shards we will create
6240 0 : let (old_shard_count, targets) =
6241 : {
6242 0 : let locked = self.inner.read().unwrap();
6243 :
6244 0 : let pageservers = locked.nodes.clone();
6245 :
6246 0 : let mut targets = Vec::new();
6247 :
6248 : // In case this is a retry, count how many already-split shards we found
6249 0 : let mut children_found = Vec::new();
6250 0 : let mut old_shard_count = None;
6251 :
6252 0 : for (tenant_shard_id, shard) in
6253 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
6254 : {
6255 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
6256 : Ordering::Equal => {
6257 : // Already split this
6258 0 : children_found.push(*tenant_shard_id);
6259 0 : continue;
6260 : }
6261 : Ordering::Greater => {
6262 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6263 0 : "Requested count {} but already have shards at count {}",
6264 0 : split_req.new_shard_count,
6265 0 : shard.shard.count.count()
6266 0 : )));
6267 : }
6268 0 : Ordering::Less => {
6269 0 : // Fall through: this shard has lower count than requested,
6270 0 : // is a candidate for splitting.
6271 0 : }
6272 : }
6273 :
6274 0 : match old_shard_count {
6275 0 : None => old_shard_count = Some(shard.shard.count),
6276 0 : Some(old_shard_count) => {
6277 0 : if old_shard_count != shard.shard.count {
6278 : // We may hit this case if a caller asked for two splits to
6279 : // different sizes, before the first one is complete.
6280 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
6281 : // of shard_count=1 and shard_count=2 shards in the map.
6282 0 : return Err(ApiError::Conflict(
6283 0 : "Cannot split, currently mid-split".to_string(),
6284 0 : ));
6285 0 : }
6286 : }
6287 : }
6288 0 : if policy.is_none() {
6289 0 : policy = Some(shard.policy.clone());
6290 0 : }
6291 0 : if shard_ident.is_none() {
6292 0 : shard_ident = Some(shard.shard);
6293 0 : }
6294 0 : if config.is_none() {
6295 0 : config = Some(shard.config.clone());
6296 0 : }
6297 0 : if preferred_az_id.is_none() {
6298 0 : preferred_az_id = shard.preferred_az().cloned();
6299 0 : }
6300 :
6301 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
6302 0 : tracing::info!(
6303 0 : "Tenant shard {} already has shard count {}",
6304 : tenant_shard_id,
6305 : split_req.new_shard_count
6306 : );
6307 0 : continue;
6308 0 : }
6309 :
6310 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
6311 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
6312 0 : ))?;
6313 :
6314 0 : let node = pageservers
6315 0 : .get(&node_id)
6316 0 : .expect("Pageservers may not be deleted while referenced");
6317 :
6318 0 : targets.push(ShardSplitTarget {
6319 0 : parent_id: *tenant_shard_id,
6320 0 : node: node.clone(),
6321 0 : child_ids: tenant_shard_id
6322 0 : .split(ShardCount::new(split_req.new_shard_count)),
6323 0 : });
6324 : }
6325 :
6326 0 : if targets.is_empty() {
6327 0 : if children_found.len() == split_req.new_shard_count as usize {
6328 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
6329 0 : new_shards: children_found,
6330 0 : }));
6331 : } else {
6332 : // No shards found to split, and no existing children found: the
6333 : // tenant doesn't exist at all.
6334 0 : return Err(ApiError::NotFound(
6335 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
6336 0 : ));
6337 : }
6338 0 : }
6339 :
6340 0 : (old_shard_count, targets)
6341 : };
6342 :
6343 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
6344 0 : let old_shard_count = old_shard_count.unwrap();
6345 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
6346 : // This ShardIdentity will be used as the template for all children, so this implicitly
6347 : // applies the new stripe size to the children.
6348 0 : let mut shard_ident = shard_ident.unwrap();
6349 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
6350 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6351 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
6352 0 : shard_ident.stripe_size
6353 0 : )));
6354 0 : }
6355 :
6356 0 : shard_ident.stripe_size = new_stripe_size;
6357 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
6358 0 : shard_ident
6359 : } else {
6360 0 : shard_ident.unwrap()
6361 : };
6362 0 : let policy = policy.unwrap();
6363 0 : let config = config.unwrap();
6364 :
6365 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
6366 0 : old_shard_count,
6367 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
6368 0 : new_stripe_size: split_req.new_stripe_size,
6369 0 : targets,
6370 0 : policy,
6371 0 : config,
6372 0 : shard_ident,
6373 0 : preferred_az_id,
6374 0 : })))
6375 0 : }
6376 :
6377 0 : async fn do_tenant_shard_split(
6378 0 : &self,
6379 0 : tenant_id: TenantId,
6380 0 : params: Box<ShardSplitParams>,
6381 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
6382 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
6383 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
6384 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
6385 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
6386 : // (https://github.com/neondatabase/neon/issues/6676)
6387 :
6388 : let ShardSplitParams {
6389 0 : old_shard_count,
6390 0 : new_shard_count,
6391 0 : new_stripe_size,
6392 0 : mut targets,
6393 0 : policy,
6394 0 : config,
6395 0 : shard_ident,
6396 0 : preferred_az_id,
6397 0 : } = *params;
6398 :
6399 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
6400 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
6401 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
6402 : // at the time of split.
6403 0 : let waiters = {
6404 0 : let mut locked = self.inner.write().unwrap();
6405 0 : let mut waiters = Vec::new();
6406 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6407 0 : for target in &mut targets {
6408 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
6409 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6410 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6411 0 : "Shard {} not found",
6412 0 : target.parent_id
6413 0 : )));
6414 : };
6415 :
6416 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
6417 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6418 0 : return Err(ApiError::Conflict(format!(
6419 0 : "Shard {} unexpectedly rescheduled during split",
6420 0 : target.parent_id
6421 0 : )));
6422 0 : }
6423 :
6424 : // Irrespective of PlacementPolicy, clear secondary locations from intent
6425 0 : shard.intent.clear_secondary(scheduler);
6426 :
6427 : // Run Reconciler to execute detach fo secondary locations.
6428 0 : if let Some(waiter) =
6429 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6430 0 : {
6431 0 : waiters.push(waiter);
6432 0 : }
6433 : }
6434 0 : waiters
6435 : };
6436 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
6437 :
6438 : // Before creating any new child shards in memory or on the pageservers, persist them: this
6439 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
6440 : // acts as the protection against two concurrent attempts to split: one of them will get a database
6441 : // error trying to insert the child shards.
6442 0 : let mut child_tsps = Vec::new();
6443 0 : for target in &targets {
6444 0 : let mut this_child_tsps = Vec::new();
6445 0 : for child in &target.child_ids {
6446 0 : let mut child_shard = shard_ident;
6447 0 : child_shard.number = child.shard_number;
6448 0 : child_shard.count = child.shard_count;
6449 :
6450 0 : tracing::info!(
6451 0 : "Create child shard persistence with stripe size {}",
6452 : shard_ident.stripe_size.0
6453 : );
6454 :
6455 0 : this_child_tsps.push(TenantShardPersistence {
6456 0 : tenant_id: child.tenant_id.to_string(),
6457 0 : shard_number: child.shard_number.0 as i32,
6458 0 : shard_count: child.shard_count.literal() as i32,
6459 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
6460 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
6461 : // populate the correct generation as part of its transaction, to protect us
6462 : // against racing with changes in the state of the parent.
6463 0 : generation: None,
6464 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
6465 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
6466 0 : config: serde_json::to_string(&config).unwrap(),
6467 0 : splitting: SplitState::Splitting,
6468 :
6469 : // Scheduling policies and preferred AZ do not carry through to children
6470 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
6471 0 : .unwrap(),
6472 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
6473 : });
6474 : }
6475 :
6476 0 : child_tsps.push((target.parent_id, this_child_tsps));
6477 : }
6478 :
6479 0 : if let Err(e) = self
6480 0 : .persistence
6481 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
6482 0 : .await
6483 : {
6484 0 : match e {
6485 : DatabaseError::Query(diesel::result::Error::DatabaseError(
6486 : DatabaseErrorKind::UniqueViolation,
6487 : _,
6488 : )) => {
6489 : // Inserting a child shard violated a unique constraint: we raced with another call to
6490 : // this function
6491 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
6492 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
6493 : }
6494 0 : _ => return Err(ApiError::InternalServerError(e.into())),
6495 : }
6496 0 : }
6497 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
6498 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6499 : ));
6500 :
6501 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
6502 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
6503 : // is not set in memory, then it was not persisted.
6504 : {
6505 0 : let mut locked = self.inner.write().unwrap();
6506 0 : for target in &targets {
6507 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
6508 0 : parent_shard.splitting = SplitState::Splitting;
6509 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
6510 0 : // split operation.
6511 0 : parent_shard
6512 0 : .observed
6513 0 : .locations
6514 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
6515 0 : }
6516 : }
6517 : }
6518 :
6519 : // TODO: issue split calls concurrently (this only matters once we're splitting
6520 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
6521 :
6522 : // HADRON: set a timeout for splitting individual shards on page servers.
6523 : // Currently we do not perform any retry because it's not clear if page server can handle
6524 : // partially split shards correctly.
6525 0 : let shard_split_timeout =
6526 0 : if let Some(env::DeploymentMode::Local) = env::get_deployment_mode() {
6527 0 : Duration::from_secs(30)
6528 : } else {
6529 0 : self.config.shard_split_request_timeout
6530 : };
6531 0 : let mut http_client_builder = reqwest::ClientBuilder::new()
6532 0 : .pool_max_idle_per_host(0)
6533 0 : .timeout(shard_split_timeout);
6534 :
6535 0 : for ssl_ca_cert in &self.config.ssl_ca_certs {
6536 0 : http_client_builder = http_client_builder.add_root_certificate(ssl_ca_cert.clone());
6537 0 : }
6538 0 : let http_client = http_client_builder
6539 0 : .build()
6540 0 : .expect("Failed to construct HTTP client");
6541 0 : for target in &targets {
6542 : let ShardSplitTarget {
6543 0 : parent_id,
6544 0 : node,
6545 0 : child_ids,
6546 0 : } = target;
6547 :
6548 0 : let client = PageserverClient::new(
6549 0 : node.get_id(),
6550 0 : http_client.clone(),
6551 0 : node.base_url(),
6552 0 : self.config.pageserver_jwt_token.as_deref(),
6553 : );
6554 :
6555 0 : let response = client
6556 0 : .tenant_shard_split(
6557 0 : *parent_id,
6558 0 : TenantShardSplitRequest {
6559 0 : new_shard_count: new_shard_count.literal(),
6560 0 : new_stripe_size,
6561 0 : },
6562 0 : )
6563 0 : .await
6564 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {parent_id}: {e}")))?;
6565 :
6566 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
6567 0 : "failpoint".to_string()
6568 0 : )));
6569 :
6570 0 : failpoint_support::sleep_millis_async!(
6571 : "shard-split-post-remote-sleep",
6572 0 : &self.reconcilers_cancel
6573 : );
6574 :
6575 0 : tracing::info!(
6576 0 : "Split {} into {}",
6577 : parent_id,
6578 0 : response
6579 0 : .new_shards
6580 0 : .iter()
6581 0 : .map(|s| format!("{s:?}"))
6582 0 : .collect::<Vec<_>>()
6583 0 : .join(",")
6584 : );
6585 :
6586 0 : if &response.new_shards != child_ids {
6587 : // This should never happen: the pageserver should agree with us on how shard splits work.
6588 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6589 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
6590 0 : parent_id,
6591 0 : response.new_shards,
6592 0 : child_ids
6593 0 : )));
6594 0 : }
6595 : }
6596 :
6597 0 : fail::fail_point!("shard-split-pre-complete", |_| Err(ApiError::Conflict(
6598 0 : "failpoint".to_string()
6599 0 : )));
6600 :
6601 0 : pausable_failpoint!("shard-split-pre-complete-pause");
6602 :
6603 : // TODO: if the pageserver restarted concurrently with our split API call,
6604 : // the actual generation of the child shard might differ from the generation
6605 : // we expect it to have. In order for our in-database generation to end up
6606 : // correct, we should carry the child generation back in the response and apply it here
6607 : // in complete_shard_split (and apply the correct generation in memory)
6608 : // (or, we can carry generation in the request and reject the request if
6609 : // it doesn't match, but that requires more retry logic on this side)
6610 :
6611 0 : self.persistence
6612 0 : .complete_shard_split(tenant_id, old_shard_count, new_shard_count)
6613 0 : .await?;
6614 :
6615 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
6616 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6617 : ));
6618 :
6619 : // Replace all the shards we just split with their children: this phase is infallible.
6620 0 : let (response, child_locations, waiters) =
6621 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
6622 :
6623 : // Notify all page servers to detach and clean up the old shards because they will no longer
6624 : // be needed. This is best-effort: if it fails, it will be cleaned up on a subsequent
6625 : // Pageserver re-attach/startup.
6626 0 : let shards_to_cleanup = targets
6627 0 : .iter()
6628 0 : .map(|target| (target.parent_id, target.node.get_id()))
6629 0 : .collect();
6630 0 : self.cleanup_locations(shards_to_cleanup).await;
6631 :
6632 : // Send compute notifications for all the new shards
6633 0 : let mut failed_notifications = Vec::new();
6634 0 : for (child_id, child_ps, stripe_size) in child_locations {
6635 0 : if let Err(e) = self
6636 0 : .compute_hook
6637 0 : .notify_attach(
6638 0 : compute_hook::ShardUpdate {
6639 0 : tenant_shard_id: child_id,
6640 0 : node_id: child_ps,
6641 0 : stripe_size,
6642 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
6643 0 : },
6644 0 : &self.reconcilers_cancel,
6645 0 : )
6646 0 : .await
6647 : {
6648 0 : tracing::warn!(
6649 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
6650 : child_id,
6651 : child_ps
6652 : );
6653 0 : failed_notifications.push(child_id);
6654 0 : }
6655 : }
6656 :
6657 : // If we failed any compute notifications, make a note to retry later.
6658 0 : if !failed_notifications.is_empty() {
6659 0 : let mut locked = self.inner.write().unwrap();
6660 0 : for failed in failed_notifications {
6661 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
6662 0 : shard.pending_compute_notification = true;
6663 0 : }
6664 : }
6665 0 : }
6666 :
6667 0 : Ok((response, waiters))
6668 0 : }
6669 :
6670 : /// A graceful migration: update the preferred node and let optimisation handle the migration
6671 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
6672 : ///
6673 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
6674 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
6675 0 : fn tenant_shard_migrate_with_prewarm(
6676 0 : &self,
6677 0 : migrate_req: &TenantShardMigrateRequest,
6678 0 : shard: &mut TenantShard,
6679 0 : scheduler: &mut Scheduler,
6680 0 : schedule_context: ScheduleContext,
6681 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
6682 0 : shard.set_preferred_node(Some(migrate_req.node_id));
6683 :
6684 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
6685 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
6686 : // e.g. by checking secondary is warm enough.
6687 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
6688 0 : }
6689 :
6690 : /// Immediate migration: directly update the intent state and kick off a reconciler
6691 0 : fn tenant_shard_migrate_immediate(
6692 0 : &self,
6693 0 : migrate_req: &TenantShardMigrateRequest,
6694 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6695 0 : shard: &mut TenantShard,
6696 0 : scheduler: &mut Scheduler,
6697 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
6698 : // Non-graceful migration: update the intent state immediately
6699 0 : let old_attached = *shard.intent.get_attached();
6700 0 : match shard.policy {
6701 0 : PlacementPolicy::Attached(n) => {
6702 : // If our new attached node was a secondary, it no longer should be.
6703 0 : shard
6704 0 : .intent
6705 0 : .remove_secondary(scheduler, migrate_req.node_id);
6706 :
6707 0 : shard
6708 0 : .intent
6709 0 : .set_attached(scheduler, Some(migrate_req.node_id));
6710 :
6711 : // If we were already attached to something, demote that to a secondary
6712 0 : if let Some(old_attached) = old_attached {
6713 0 : if n > 0 {
6714 : // Remove other secondaries to make room for the location we'll demote
6715 0 : while shard.intent.get_secondary().len() >= n {
6716 0 : shard.intent.pop_secondary(scheduler);
6717 0 : }
6718 :
6719 0 : shard.intent.push_secondary(scheduler, old_attached);
6720 0 : }
6721 0 : }
6722 : }
6723 0 : PlacementPolicy::Secondary => {
6724 0 : shard.intent.clear(scheduler);
6725 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6726 0 : }
6727 : PlacementPolicy::Detached => {
6728 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6729 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
6730 0 : )));
6731 : }
6732 : }
6733 :
6734 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
6735 0 : shard.sequence = shard.sequence.next();
6736 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
6737 0 : Ok(self.maybe_configured_reconcile_shard(
6738 0 : shard,
6739 0 : nodes,
6740 0 : (&migrate_req.migration_config).into(),
6741 0 : ))
6742 0 : }
6743 :
6744 0 : pub(crate) async fn tenant_shard_migrate(
6745 0 : &self,
6746 0 : tenant_shard_id: TenantShardId,
6747 0 : migrate_req: TenantShardMigrateRequest,
6748 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6749 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
6750 : // get a different outcome to handle
6751 : enum MigrationOutcome {
6752 : Optimization(Option<ScheduleOptimization>),
6753 : Reconcile(Option<ReconcilerWaiter>),
6754 : }
6755 :
6756 0 : let outcome = {
6757 0 : let mut locked = self.inner.write().unwrap();
6758 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6759 :
6760 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6761 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6762 0 : "Node {} not found",
6763 0 : migrate_req.node_id
6764 0 : )));
6765 : };
6766 :
6767 : // Migration to unavavailable node requires force flag
6768 0 : if !node.is_available() {
6769 0 : if migrate_req.migration_config.override_scheduler {
6770 : // Warn but proceed: the caller may intend to manually adjust the placement of
6771 : // a shard even if the node is down, e.g. if intervening during an incident.
6772 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
6773 : } else {
6774 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
6775 0 : return Err(ApiError::PreconditionFailed(
6776 0 : format!("Node {node} is unavailable").into_boxed_str(),
6777 0 : ));
6778 : }
6779 0 : }
6780 :
6781 : // Calculate the ScheduleContext for this tenant
6782 0 : let mut schedule_context = ScheduleContext::default();
6783 0 : for (_shard_id, shard) in
6784 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
6785 0 : {
6786 0 : schedule_context.avoid(&shard.intent.all_pageservers());
6787 0 : }
6788 :
6789 : // Look up the specific shard we will migrate
6790 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6791 0 : return Err(ApiError::NotFound(
6792 0 : anyhow::anyhow!("Tenant shard not found").into(),
6793 0 : ));
6794 : };
6795 :
6796 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
6797 : // be migrated back by the optimiser.
6798 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
6799 0 : scheduler,
6800 0 : &schedule_context,
6801 0 : migrate_req.node_id,
6802 0 : &[],
6803 0 : ) {
6804 0 : if !migrate_req.migration_config.override_scheduler {
6805 0 : return Err(ApiError::PreconditionFailed(
6806 0 : "Migration to a worse-scoring node".into(),
6807 0 : ));
6808 : } else {
6809 0 : tracing::info!(
6810 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
6811 : migrate_req.node_id
6812 : );
6813 : }
6814 0 : }
6815 :
6816 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
6817 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
6818 0 : return Err(ApiError::PreconditionFailed(
6819 0 : format!(
6820 0 : "Migration expected to originate from {} but shard is on {:?}",
6821 0 : origin_node_id,
6822 0 : shard.intent.get_attached()
6823 0 : )
6824 0 : .into(),
6825 0 : ));
6826 0 : }
6827 0 : }
6828 :
6829 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6830 : // No-op case: we will still proceed to wait for reconciliation in case it is
6831 : // incomplete from an earlier update to the intent.
6832 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
6833 :
6834 : // An instruction to migrate to the currently attached node should
6835 : // cancel any pending graceful migration
6836 0 : shard.set_preferred_node(None);
6837 :
6838 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
6839 0 : shard,
6840 0 : nodes,
6841 0 : (&migrate_req.migration_config).into(),
6842 0 : ))
6843 0 : } else if migrate_req.migration_config.prewarm {
6844 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
6845 0 : &migrate_req,
6846 0 : shard,
6847 0 : scheduler,
6848 0 : schedule_context,
6849 0 : )?)
6850 : } else {
6851 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
6852 0 : &migrate_req,
6853 0 : nodes,
6854 0 : shard,
6855 0 : scheduler,
6856 0 : )?)
6857 : }
6858 : };
6859 :
6860 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
6861 0 : let waiter = match outcome {
6862 0 : MigrationOutcome::Optimization(Some(optimization)) => {
6863 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
6864 : // we might as well do it more promptly as this is a direct external request.
6865 0 : let mut validated = self
6866 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
6867 0 : .await;
6868 0 : if let Some((_shard_id, optimization)) = validated.pop() {
6869 0 : let mut locked = self.inner.write().unwrap();
6870 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6871 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6872 : // Rare but possible: tenant is removed between generating optimisation and validating it.
6873 0 : return Err(ApiError::NotFound(
6874 0 : anyhow::anyhow!("Tenant shard not found").into(),
6875 0 : ));
6876 : };
6877 :
6878 0 : if !shard.apply_optimization(scheduler, optimization) {
6879 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
6880 : // and therefore not applied.
6881 0 : tracing::warn!(
6882 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
6883 : );
6884 0 : }
6885 0 : self.maybe_configured_reconcile_shard(
6886 0 : shard,
6887 0 : nodes,
6888 0 : (&migrate_req.migration_config).into(),
6889 : )
6890 : } else {
6891 0 : None
6892 : }
6893 : }
6894 0 : MigrationOutcome::Optimization(None) => None,
6895 0 : MigrationOutcome::Reconcile(waiter) => waiter,
6896 : };
6897 :
6898 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
6899 : // locations, this has a good chance of timing out.
6900 0 : if let Some(waiter) = waiter {
6901 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6902 : } else {
6903 0 : tracing::info!("Migration is a no-op");
6904 : }
6905 :
6906 0 : Ok(TenantShardMigrateResponse {})
6907 0 : }
6908 :
6909 0 : pub(crate) async fn tenant_shard_migrate_secondary(
6910 0 : &self,
6911 0 : tenant_shard_id: TenantShardId,
6912 0 : migrate_req: TenantShardMigrateRequest,
6913 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6914 0 : let waiter = {
6915 0 : let mut locked = self.inner.write().unwrap();
6916 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6917 :
6918 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6919 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6920 0 : "Node {} not found",
6921 0 : migrate_req.node_id
6922 0 : )));
6923 : };
6924 :
6925 0 : if !node.is_available() {
6926 : // Warn but proceed: the caller may intend to manually adjust the placement of
6927 : // a shard even if the node is down, e.g. if intervening during an incident.
6928 0 : tracing::warn!("Migrating to unavailable node {node}");
6929 0 : }
6930 :
6931 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6932 0 : return Err(ApiError::NotFound(
6933 0 : anyhow::anyhow!("Tenant shard not found").into(),
6934 0 : ));
6935 : };
6936 :
6937 0 : if shard.intent.get_secondary().len() == 1
6938 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
6939 : {
6940 0 : tracing::info!(
6941 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
6942 : shard.intent
6943 : );
6944 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6945 0 : tracing::info!(
6946 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
6947 : );
6948 : } else {
6949 0 : let old_secondaries = shard.intent.get_secondary().clone();
6950 0 : for secondary in old_secondaries {
6951 0 : shard.intent.remove_secondary(scheduler, secondary);
6952 0 : }
6953 :
6954 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6955 0 : shard.sequence = shard.sequence.next();
6956 0 : tracing::info!(
6957 0 : "Migrating secondary to {node}: new intent {:?}",
6958 : shard.intent
6959 : );
6960 : }
6961 :
6962 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6963 : };
6964 :
6965 0 : if let Some(waiter) = waiter {
6966 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6967 : } else {
6968 0 : tracing::info!("Migration is a no-op");
6969 : }
6970 :
6971 0 : Ok(TenantShardMigrateResponse {})
6972 0 : }
6973 :
6974 : /// 'cancel' in this context means cancel any ongoing reconcile
6975 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
6976 0 : &self,
6977 0 : tenant_shard_id: TenantShardId,
6978 0 : ) -> Result<(), ApiError> {
6979 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
6980 0 : let waiter = {
6981 0 : let locked = self.inner.write().unwrap();
6982 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6983 0 : return Err(ApiError::NotFound(
6984 0 : anyhow::anyhow!("Tenant shard not found").into(),
6985 0 : ));
6986 : };
6987 :
6988 0 : let waiter = shard.get_waiter();
6989 0 : match waiter {
6990 : None => {
6991 0 : tracing::info!("Shard does not have an ongoing Reconciler");
6992 0 : return Ok(());
6993 : }
6994 0 : Some(waiter) => {
6995 0 : tracing::info!("Cancelling Reconciler");
6996 0 : shard.cancel_reconciler();
6997 0 : waiter
6998 : }
6999 : }
7000 : };
7001 :
7002 : // Cancellation should be prompt. If this fails we have still done our job of firing the
7003 : // cancellation token, but by returning an ApiError we will indicate to the caller that
7004 : // the Reconciler is misbehaving and not respecting the cancellation token
7005 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
7006 0 : .await?;
7007 :
7008 0 : Ok(())
7009 0 : }
7010 :
7011 : /// This is for debug/support only: we simply drop all state for a tenant, without
7012 : /// detaching or deleting it on pageservers.
7013 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
7014 0 : self.persistence.delete_tenant(tenant_id).await?;
7015 :
7016 0 : let mut locked = self.inner.write().unwrap();
7017 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
7018 0 : let mut shards = Vec::new();
7019 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
7020 0 : shards.push(*tenant_shard_id);
7021 0 : }
7022 :
7023 0 : for shard_id in shards {
7024 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
7025 0 : shard.intent.clear(scheduler);
7026 0 : }
7027 : }
7028 :
7029 0 : Ok(())
7030 0 : }
7031 :
7032 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
7033 : /// tenant with a very high generation number so that it will see the existing data.
7034 : /// It does not create timelines on safekeepers, because they might already exist on some
7035 : /// safekeeper set. So, the timelines are not storcon-managed after the import.
7036 0 : pub(crate) async fn tenant_import(
7037 0 : &self,
7038 0 : tenant_id: TenantId,
7039 0 : ) -> Result<TenantCreateResponse, ApiError> {
7040 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
7041 0 : let maybe_node = {
7042 0 : self.inner
7043 0 : .read()
7044 0 : .unwrap()
7045 0 : .nodes
7046 0 : .values()
7047 0 : .find(|n| n.is_available())
7048 0 : .cloned()
7049 : };
7050 0 : let Some(node) = maybe_node else {
7051 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
7052 : };
7053 :
7054 0 : let client = PageserverClient::new(
7055 0 : node.get_id(),
7056 0 : self.http_client.clone(),
7057 0 : node.base_url(),
7058 0 : self.config.pageserver_jwt_token.as_deref(),
7059 : );
7060 :
7061 0 : let scan_result = client
7062 0 : .tenant_scan_remote_storage(tenant_id)
7063 0 : .await
7064 0 : .map_err(|e| passthrough_api_error(&node, e))?;
7065 :
7066 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
7067 0 : let Some(shard_count) = scan_result
7068 0 : .shards
7069 0 : .iter()
7070 0 : .map(|s| s.tenant_shard_id.shard_count)
7071 0 : .max()
7072 : else {
7073 0 : return Err(ApiError::NotFound(
7074 0 : anyhow::anyhow!("No shards found").into(),
7075 0 : ));
7076 : };
7077 :
7078 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
7079 : // to
7080 0 : let generation = scan_result
7081 0 : .shards
7082 0 : .iter()
7083 0 : .map(|s| s.generation)
7084 0 : .max()
7085 0 : .expect("We already validated >0 shards");
7086 :
7087 : // Find the tenant's stripe size. This wasn't always persisted in the tenant manifest, so
7088 : // fall back to the original default stripe size of 32768 (256 MB) if it's not specified.
7089 : const ORIGINAL_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(32768);
7090 0 : let stripe_size = scan_result
7091 0 : .shards
7092 0 : .iter()
7093 0 : .find(|s| s.tenant_shard_id.shard_count == shard_count && s.generation == generation)
7094 0 : .expect("we validated >0 shards above")
7095 : .stripe_size
7096 0 : .unwrap_or_else(|| {
7097 0 : if shard_count.count() > 1 {
7098 0 : warn!("unknown stripe size, assuming {ORIGINAL_STRIPE_SIZE}");
7099 0 : }
7100 0 : ORIGINAL_STRIPE_SIZE
7101 0 : });
7102 :
7103 0 : let (response, waiters) = self
7104 0 : .do_tenant_create(TenantCreateRequest {
7105 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
7106 0 : generation,
7107 0 :
7108 0 : shard_parameters: ShardParameters {
7109 0 : count: shard_count,
7110 0 : stripe_size,
7111 0 : },
7112 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
7113 0 : config: TenantConfig::default(),
7114 0 : })
7115 0 : .await?;
7116 :
7117 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
7118 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
7119 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
7120 : // reconcile, as reconciliation includes notifying compute.
7121 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
7122 0 : }
7123 :
7124 0 : Ok(response)
7125 0 : }
7126 :
7127 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
7128 : /// we don't have to make TenantShard clonable in the return path.
7129 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7130 0 : let serialized = {
7131 0 : let locked = self.inner.read().unwrap();
7132 0 : let result = locked.tenants.values().collect::<Vec<_>>();
7133 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
7134 : };
7135 :
7136 0 : hyper::Response::builder()
7137 0 : .status(hyper::StatusCode::OK)
7138 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7139 0 : .body(hyper::Body::from(serialized))
7140 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7141 0 : }
7142 :
7143 : /// Check the consistency of in-memory state vs. persistent state, and check that the
7144 : /// scheduler's statistics are up to date.
7145 : ///
7146 : /// These consistency checks expect an **idle** system. If changes are going on while
7147 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
7148 : /// checks, but not suitable for running continuously in the background in the field.
7149 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
7150 0 : let (mut expect_nodes, mut expect_shards) = {
7151 0 : let locked = self.inner.read().unwrap();
7152 :
7153 0 : locked
7154 0 : .scheduler
7155 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
7156 0 : .context("Scheduler checks")
7157 0 : .map_err(ApiError::InternalServerError)?;
7158 :
7159 0 : let expect_nodes = locked
7160 0 : .nodes
7161 0 : .values()
7162 0 : .map(|n| n.to_persistent())
7163 0 : .collect::<Vec<_>>();
7164 :
7165 0 : let expect_shards = locked
7166 0 : .tenants
7167 0 : .values()
7168 0 : .map(|t| t.to_persistent())
7169 0 : .collect::<Vec<_>>();
7170 :
7171 : // This method can only validate the state of an idle system: if a reconcile is in
7172 : // progress, fail out early to avoid giving false errors on state that won't match
7173 : // between database and memory under a ReconcileResult is processed.
7174 0 : for t in locked.tenants.values() {
7175 0 : if t.reconciler.is_some() {
7176 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7177 0 : "Shard {} reconciliation in progress",
7178 0 : t.tenant_shard_id
7179 0 : )));
7180 0 : }
7181 : }
7182 :
7183 0 : (expect_nodes, expect_shards)
7184 : };
7185 :
7186 0 : let mut nodes = self.persistence.list_nodes().await?;
7187 0 : expect_nodes.sort_by_key(|n| n.node_id);
7188 0 : nodes.sort_by_key(|n| n.node_id);
7189 :
7190 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
7191 0 : let node_result = if nodes != expect_nodes {
7192 0 : tracing::error!("Consistency check failed on nodes.");
7193 0 : tracing::error!(
7194 0 : "Nodes in memory: {}",
7195 0 : serde_json::to_string(&expect_nodes)
7196 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7197 : );
7198 0 : tracing::error!(
7199 0 : "Nodes in database: {}",
7200 0 : serde_json::to_string(&nodes)
7201 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7202 : );
7203 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
7204 0 : "Node consistency failure"
7205 0 : )))
7206 : } else {
7207 0 : Ok(())
7208 : };
7209 :
7210 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
7211 0 : persistent_shards
7212 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7213 :
7214 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7215 :
7216 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
7217 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
7218 : // fields are added, before doing a comparison.
7219 0 : for tsp in &mut persistent_shards {
7220 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
7221 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
7222 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
7223 : }
7224 :
7225 0 : if persistent_shards != expect_shards {
7226 0 : tracing::error!("Consistency check failed on shards.");
7227 :
7228 0 : tracing::error!(
7229 0 : "Shards in memory: {}",
7230 0 : serde_json::to_string(&expect_shards)
7231 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7232 : );
7233 0 : tracing::error!(
7234 0 : "Shards in database: {}",
7235 0 : serde_json::to_string(&persistent_shards)
7236 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7237 : );
7238 :
7239 : // The total dump log lines above are useful in testing but in the field grafana will
7240 : // usually just drop them because they're so large. So we also do some explicit logging
7241 : // of just the diffs.
7242 0 : let persistent_shards = persistent_shards
7243 0 : .into_iter()
7244 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7245 0 : .collect::<HashMap<_, _>>();
7246 0 : let expect_shards = expect_shards
7247 0 : .into_iter()
7248 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7249 0 : .collect::<HashMap<_, _>>();
7250 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
7251 0 : match expect_shards.get(tenant_shard_id) {
7252 : None => {
7253 0 : tracing::error!(
7254 0 : "Shard {} found in database but not in memory",
7255 : tenant_shard_id
7256 : );
7257 : }
7258 0 : Some(expect_tsp) => {
7259 0 : if expect_tsp != persistent_tsp {
7260 0 : tracing::error!(
7261 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
7262 : tenant_shard_id,
7263 0 : serde_json::to_string(expect_tsp).unwrap(),
7264 0 : serde_json::to_string(&persistent_tsp).unwrap()
7265 : );
7266 0 : }
7267 : }
7268 : }
7269 : }
7270 :
7271 : // Having already logged any differences, log any shards that simply aren't present in the database
7272 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
7273 0 : if !persistent_shards.contains_key(tenant_shard_id) {
7274 0 : tracing::error!(
7275 0 : "Shard {} found in memory but not in database: {}",
7276 : tenant_shard_id,
7277 0 : serde_json::to_string(memory_tsp)
7278 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7279 : );
7280 0 : }
7281 : }
7282 :
7283 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7284 0 : "Shard consistency failure"
7285 0 : )));
7286 0 : }
7287 :
7288 0 : node_result
7289 0 : }
7290 :
7291 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
7292 : /// we don't have to make TenantShard clonable in the return path.
7293 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7294 0 : let serialized = {
7295 0 : let locked = self.inner.read().unwrap();
7296 0 : serde_json::to_string(&locked.scheduler)
7297 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7298 : };
7299 :
7300 0 : hyper::Response::builder()
7301 0 : .status(hyper::StatusCode::OK)
7302 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7303 0 : .body(hyper::Body::from(serialized))
7304 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7305 0 : }
7306 :
7307 : /// This is for debug/support only: we simply drop all state for a tenant, without
7308 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
7309 : /// tenants that were on this node.
7310 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
7311 0 : self.persistence.set_tombstone(node_id).await?;
7312 :
7313 0 : let mut locked = self.inner.write().unwrap();
7314 :
7315 0 : for shard in locked.tenants.values_mut() {
7316 0 : shard.deref_node(node_id);
7317 0 : shard.observed.locations.remove(&node_id);
7318 0 : }
7319 :
7320 0 : let mut nodes = (*locked.nodes).clone();
7321 0 : nodes.remove(&node_id);
7322 0 : locked.nodes = Arc::new(nodes);
7323 0 : metrics::METRICS_REGISTRY
7324 0 : .metrics_group
7325 0 : .storage_controller_pageserver_nodes
7326 0 : .set(locked.nodes.len() as i64);
7327 0 : metrics::METRICS_REGISTRY
7328 0 : .metrics_group
7329 0 : .storage_controller_https_pageserver_nodes
7330 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7331 :
7332 0 : locked.scheduler.node_remove(node_id);
7333 :
7334 0 : Ok(())
7335 0 : }
7336 :
7337 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
7338 : /// that we don't leave any bad state behind in the storage controller, but unclean
7339 : /// in the sense that we are not carefully draining the node.
7340 0 : pub(crate) async fn node_delete_old(&self, node_id: NodeId) -> Result<(), ApiError> {
7341 0 : let _node_lock =
7342 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
7343 :
7344 : // 1. Atomically update in-memory state:
7345 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
7346 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
7347 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
7348 : // re-insert references to this node into the ObservedState of shards
7349 : // - drop the node from the scheduler
7350 : {
7351 0 : let mut locked = self.inner.write().unwrap();
7352 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7353 :
7354 : {
7355 0 : let mut nodes_mut = (*nodes).deref().clone();
7356 0 : match nodes_mut.get_mut(&node_id) {
7357 0 : Some(node) => {
7358 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
7359 0 : // if we crash it would not be desirable to leave the node paused after a restart.
7360 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
7361 0 : }
7362 : None => {
7363 0 : tracing::info!(
7364 0 : "Node not found: presuming this is a retry and returning success"
7365 : );
7366 0 : return Ok(());
7367 : }
7368 : }
7369 :
7370 0 : *nodes = Arc::new(nodes_mut);
7371 : }
7372 :
7373 0 : for (_tenant_id, mut schedule_context, shards) in
7374 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7375 : {
7376 0 : for shard in shards {
7377 0 : if shard.deref_node(node_id) {
7378 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
7379 : // TODO: implement force flag to remove a node even if we can't reschedule
7380 : // a tenant
7381 0 : tracing::error!(
7382 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7383 : shard.tenant_shard_id
7384 : );
7385 0 : return Err(e.into());
7386 : } else {
7387 0 : tracing::info!(
7388 0 : "Rescheduled shard {} away from node during deletion",
7389 : shard.tenant_shard_id
7390 : )
7391 : }
7392 :
7393 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
7394 0 : }
7395 :
7396 : // Here we remove an existing observed location for the node we're removing, and it will
7397 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7398 : // process_result.
7399 : //
7400 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
7401 : // means any reconciles we spawned will know about the node we're deleting, enabling them
7402 : // to do live migrations if it's still online.
7403 0 : shard.observed.locations.remove(&node_id);
7404 : }
7405 : }
7406 :
7407 0 : scheduler.node_remove(node_id);
7408 :
7409 : {
7410 0 : let mut nodes_mut = (**nodes).clone();
7411 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7412 0 : // Ensure that any reconciler holding an Arc<> to this node will
7413 0 : // drop out when trying to RPC to it (setting Offline state sets the
7414 0 : // cancellation token on the Node object).
7415 0 : removed_node.set_availability(NodeAvailability::Offline);
7416 0 : }
7417 0 : *nodes = Arc::new(nodes_mut);
7418 0 : metrics::METRICS_REGISTRY
7419 0 : .metrics_group
7420 0 : .storage_controller_pageserver_nodes
7421 0 : .set(nodes.len() as i64);
7422 0 : metrics::METRICS_REGISTRY
7423 0 : .metrics_group
7424 0 : .storage_controller_https_pageserver_nodes
7425 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7426 : }
7427 : }
7428 :
7429 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
7430 : // the removed node, as this column means "The pageserver to which this generation was issued", and
7431 : // their generations won't get updated until the reconcilers moving them away from this node complete.
7432 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
7433 : // that exists.
7434 :
7435 : // 2. Actually delete the node from in-memory state and set tombstone to the database
7436 : // for preventing the node to register again.
7437 0 : tracing::info!("Deleting node from database");
7438 0 : self.persistence.set_tombstone(node_id).await?;
7439 :
7440 0 : Ok(())
7441 0 : }
7442 :
7443 0 : pub(crate) async fn delete_node(
7444 0 : self: &Arc<Self>,
7445 0 : node_id: NodeId,
7446 0 : policy_on_start: NodeSchedulingPolicy,
7447 0 : force: bool,
7448 0 : cancel: CancellationToken,
7449 0 : ) -> Result<(), OperationError> {
7450 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal).build();
7451 :
7452 0 : let mut waiters: Vec<ReconcilerWaiter> = Vec::new();
7453 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
7454 :
7455 0 : let reset_node_policy_on_cancel = || async {
7456 0 : match self
7457 0 : .node_configure(node_id, None, Some(policy_on_start))
7458 0 : .await
7459 : {
7460 0 : Ok(()) => OperationError::Cancelled,
7461 0 : Err(err) => {
7462 0 : OperationError::FinalizeError(
7463 0 : format!(
7464 0 : "Failed to finalise delete cancel of {} by setting scheduling policy to {}: {}",
7465 0 : node_id, String::from(policy_on_start), err
7466 0 : )
7467 0 : .into(),
7468 0 : )
7469 : }
7470 : }
7471 0 : };
7472 :
7473 0 : while !tid_iter.finished() {
7474 0 : if cancel.is_cancelled() {
7475 0 : return Err(reset_node_policy_on_cancel().await);
7476 0 : }
7477 :
7478 0 : operation_utils::validate_node_state(
7479 0 : &node_id,
7480 0 : self.inner.read().unwrap().nodes.clone(),
7481 0 : NodeSchedulingPolicy::Deleting,
7482 0 : )?;
7483 :
7484 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7485 0 : let tid = match tid_iter.next() {
7486 0 : Some(tid) => tid,
7487 : None => {
7488 0 : break;
7489 : }
7490 : };
7491 :
7492 0 : let mut locked = self.inner.write().unwrap();
7493 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7494 :
7495 : // Calculate a schedule context here to avoid borrow checker issues.
7496 0 : let mut schedule_context = ScheduleContext::default();
7497 0 : for (_, shard) in tenants.range(TenantShardId::tenant_range(tid.tenant_id)) {
7498 0 : schedule_context.avoid(&shard.intent.all_pageservers());
7499 0 : }
7500 :
7501 0 : let tenant_shard = match tenants.get_mut(&tid) {
7502 0 : Some(tenant_shard) => tenant_shard,
7503 : None => {
7504 : // Tenant shard was deleted by another operation. Skip it.
7505 0 : continue;
7506 : }
7507 : };
7508 :
7509 0 : match tenant_shard.get_scheduling_policy() {
7510 0 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {
7511 0 : // A migration during delete is classed as 'essential' because it is required to
7512 0 : // uphold our availability goals for the tenant: this shard is elegible for migration.
7513 0 : }
7514 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
7515 : // If we have been asked to avoid rescheduling this shard, then do not migrate it during a deletion
7516 0 : tracing::warn!(
7517 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7518 0 : "Skip migration during deletion because shard scheduling policy {:?} disallows it",
7519 0 : tenant_shard.get_scheduling_policy(),
7520 : );
7521 0 : continue;
7522 : }
7523 : }
7524 :
7525 0 : if tenant_shard.deref_node(node_id) {
7526 0 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
7527 0 : tracing::error!(
7528 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7529 : tenant_shard.tenant_shard_id
7530 : );
7531 0 : return Err(OperationError::ImpossibleConstraint(e.to_string().into()));
7532 : } else {
7533 0 : tracing::info!(
7534 0 : "Rescheduled shard {} away from node during deletion",
7535 : tenant_shard.tenant_shard_id
7536 : )
7537 : }
7538 :
7539 0 : let waiter = self.maybe_configured_reconcile_shard(
7540 0 : tenant_shard,
7541 0 : nodes,
7542 0 : reconciler_config,
7543 0 : );
7544 :
7545 0 : if force {
7546 0 : // Here we remove an existing observed location for the node we're removing, and it will
7547 0 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7548 0 : // process_result.
7549 0 : //
7550 0 : // Note that we update the shard's observed state _after_ calling maybe_configured_reconcile_shard:
7551 0 : // that means any reconciles we spawned will know about the node we're deleting,
7552 0 : // enabling them to do live migrations if it's still online.
7553 0 : tenant_shard.observed.locations.remove(&node_id);
7554 0 : } else if let Some(waiter) = waiter {
7555 0 : waiters.push(waiter);
7556 0 : }
7557 0 : }
7558 : }
7559 :
7560 0 : waiters = self
7561 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
7562 0 : .await;
7563 :
7564 0 : failpoint_support::sleep_millis_async!("sleepy-delete-loop", &cancel);
7565 : }
7566 :
7567 0 : while !waiters.is_empty() {
7568 0 : if cancel.is_cancelled() {
7569 0 : return Err(reset_node_policy_on_cancel().await);
7570 0 : }
7571 :
7572 0 : tracing::info!("Awaiting {} pending delete reconciliations", waiters.len());
7573 :
7574 0 : waiters = self
7575 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7576 0 : .await;
7577 : }
7578 :
7579 0 : let pf = pausable_failpoint!("delete-node-after-reconciles-spawned", &cancel);
7580 0 : if pf.is_err() {
7581 : // An error from pausable_failpoint indicates the cancel token was triggered.
7582 0 : return Err(reset_node_policy_on_cancel().await);
7583 0 : }
7584 :
7585 0 : self.persistence
7586 0 : .set_tombstone(node_id)
7587 0 : .await
7588 0 : .map_err(|e| OperationError::FinalizeError(e.to_string().into()))?;
7589 :
7590 : {
7591 0 : let mut locked = self.inner.write().unwrap();
7592 0 : let (nodes, _, scheduler) = locked.parts_mut();
7593 :
7594 0 : scheduler.node_remove(node_id);
7595 :
7596 0 : let mut nodes_mut = (**nodes).clone();
7597 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7598 0 : // Ensure that any reconciler holding an Arc<> to this node will
7599 0 : // drop out when trying to RPC to it (setting Offline state sets the
7600 0 : // cancellation token on the Node object).
7601 0 : removed_node.set_availability(NodeAvailability::Offline);
7602 0 : }
7603 0 : *nodes = Arc::new(nodes_mut);
7604 :
7605 0 : metrics::METRICS_REGISTRY
7606 0 : .metrics_group
7607 0 : .storage_controller_pageserver_nodes
7608 0 : .set(nodes.len() as i64);
7609 0 : metrics::METRICS_REGISTRY
7610 0 : .metrics_group
7611 0 : .storage_controller_https_pageserver_nodes
7612 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7613 : }
7614 :
7615 0 : Ok(())
7616 0 : }
7617 :
7618 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
7619 0 : let nodes = {
7620 0 : self.inner
7621 0 : .read()
7622 0 : .unwrap()
7623 0 : .nodes
7624 0 : .values()
7625 0 : .cloned()
7626 0 : .collect::<Vec<_>>()
7627 : };
7628 :
7629 0 : Ok(nodes)
7630 0 : }
7631 :
7632 0 : pub(crate) async fn tombstone_list(&self) -> Result<Vec<Node>, ApiError> {
7633 0 : self.persistence
7634 0 : .list_tombstones()
7635 0 : .await?
7636 0 : .into_iter()
7637 0 : .map(|np| Node::from_persistent(np, false))
7638 0 : .collect::<Result<Vec<_>, _>>()
7639 0 : .map_err(ApiError::InternalServerError)
7640 0 : }
7641 :
7642 0 : pub(crate) async fn tombstone_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
7643 0 : let _node_lock = trace_exclusive_lock(
7644 0 : &self.node_op_locks,
7645 0 : node_id,
7646 0 : NodeOperations::DeleteTombstone,
7647 0 : )
7648 0 : .await;
7649 :
7650 0 : if matches!(self.get_node(node_id).await, Err(ApiError::NotFound(_))) {
7651 0 : self.persistence.delete_node(node_id).await?;
7652 0 : Ok(())
7653 : } else {
7654 0 : Err(ApiError::Conflict(format!(
7655 0 : "Node {node_id} is in use, consider using tombstone API first"
7656 0 : )))
7657 : }
7658 0 : }
7659 :
7660 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
7661 0 : self.inner
7662 0 : .read()
7663 0 : .unwrap()
7664 0 : .nodes
7665 0 : .get(&node_id)
7666 0 : .cloned()
7667 0 : .ok_or(ApiError::NotFound(
7668 0 : format!("Node {node_id} not registered").into(),
7669 0 : ))
7670 0 : }
7671 :
7672 0 : pub(crate) async fn get_node_shards(
7673 0 : &self,
7674 0 : node_id: NodeId,
7675 0 : ) -> Result<NodeShardResponse, ApiError> {
7676 0 : let locked = self.inner.read().unwrap();
7677 0 : let mut shards = Vec::new();
7678 0 : for (tid, tenant) in locked.tenants.iter() {
7679 0 : let is_intended_secondary = match (
7680 0 : tenant.intent.get_attached() == &Some(node_id),
7681 0 : tenant.intent.get_secondary().contains(&node_id),
7682 0 : ) {
7683 : (true, true) => {
7684 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7685 0 : "{} attached as primary+secondary on the same node",
7686 0 : tid
7687 0 : )));
7688 : }
7689 0 : (true, false) => Some(false),
7690 0 : (false, true) => Some(true),
7691 0 : (false, false) => None,
7692 : };
7693 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
7694 0 : tenant.observed.locations.get(&node_id)
7695 : {
7696 0 : Some(conf.secondary_conf.is_some())
7697 : } else {
7698 0 : None
7699 : };
7700 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
7701 0 : shards.push(NodeShard {
7702 0 : tenant_shard_id: *tid,
7703 0 : is_intended_secondary,
7704 0 : is_observed_secondary,
7705 0 : });
7706 0 : }
7707 : }
7708 0 : Ok(NodeShardResponse { node_id, shards })
7709 0 : }
7710 :
7711 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
7712 0 : self.persistence.get_leader().await
7713 0 : }
7714 :
7715 0 : pub(crate) async fn node_register(
7716 0 : &self,
7717 0 : register_req: NodeRegisterRequest,
7718 0 : ) -> Result<(), ApiError> {
7719 0 : let _node_lock = trace_exclusive_lock(
7720 0 : &self.node_op_locks,
7721 0 : register_req.node_id,
7722 0 : NodeOperations::Register,
7723 0 : )
7724 0 : .await;
7725 :
7726 : #[derive(PartialEq)]
7727 : enum RegistrationStatus {
7728 : UpToDate,
7729 : NeedUpdate,
7730 : Mismatched,
7731 : New,
7732 : }
7733 :
7734 0 : let registration_status = {
7735 0 : let locked = self.inner.read().unwrap();
7736 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
7737 0 : if node.registration_match(®ister_req) {
7738 0 : if node.need_update(®ister_req) {
7739 0 : RegistrationStatus::NeedUpdate
7740 : } else {
7741 0 : RegistrationStatus::UpToDate
7742 : }
7743 : } else {
7744 0 : RegistrationStatus::Mismatched
7745 : }
7746 : } else {
7747 0 : RegistrationStatus::New
7748 : }
7749 : };
7750 :
7751 0 : match registration_status {
7752 : RegistrationStatus::UpToDate => {
7753 0 : tracing::info!(
7754 0 : "Node {} re-registered with matching address and is up to date",
7755 : register_req.node_id
7756 : );
7757 :
7758 0 : return Ok(());
7759 : }
7760 : RegistrationStatus::Mismatched => {
7761 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
7762 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
7763 : // a fixed address through the lifetime of a node.
7764 0 : tracing::warn!(
7765 0 : "Node {} tried to register with different address",
7766 : register_req.node_id
7767 : );
7768 0 : return Err(ApiError::Conflict(
7769 0 : "Node is already registered with different address".to_string(),
7770 0 : ));
7771 : }
7772 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
7773 0 : // fallthrough
7774 0 : }
7775 : }
7776 :
7777 : // We do not require that a node is actually online when registered (it will start life
7778 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
7779 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
7780 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
7781 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
7782 0 : if tokio::net::lookup_host(format!(
7783 0 : "{}:{}",
7784 : register_req.listen_http_addr, register_req.listen_http_port
7785 : ))
7786 0 : .await
7787 0 : .is_err()
7788 : {
7789 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
7790 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
7791 : // we return a soft 503 error, to encourage callers to retry past transient issues.
7792 0 : return Err(ApiError::ResourceUnavailable(
7793 0 : format!(
7794 0 : "Node {} tried to register with unknown DNS name '{}'",
7795 0 : register_req.node_id, register_req.listen_http_addr
7796 0 : )
7797 0 : .into(),
7798 0 : ));
7799 0 : }
7800 :
7801 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
7802 0 : return Err(ApiError::PreconditionFailed(
7803 0 : format!(
7804 0 : "Node {} has no https port, but use_https is enabled",
7805 0 : register_req.node_id
7806 0 : )
7807 0 : .into(),
7808 0 : ));
7809 0 : }
7810 :
7811 0 : if register_req.listen_grpc_addr.is_some() != register_req.listen_grpc_port.is_some() {
7812 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
7813 0 : "must specify both gRPC address and port"
7814 0 : )));
7815 0 : }
7816 :
7817 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
7818 : // This ensures that before we use it for anything or expose it via any external
7819 : // API, it is guaranteed to be available after a restart.
7820 0 : let new_node = Node::new(
7821 0 : register_req.node_id,
7822 0 : register_req.listen_http_addr,
7823 0 : register_req.listen_http_port,
7824 0 : register_req.listen_https_port,
7825 0 : register_req.listen_pg_addr,
7826 0 : register_req.listen_pg_port,
7827 0 : register_req.listen_grpc_addr.clone(),
7828 0 : register_req.listen_grpc_port,
7829 0 : register_req.availability_zone_id.clone(),
7830 0 : self.config.use_https_pageserver_api,
7831 : );
7832 0 : let new_node = match new_node {
7833 0 : Ok(new_node) => new_node,
7834 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
7835 : };
7836 :
7837 0 : match registration_status {
7838 : RegistrationStatus::New => {
7839 0 : self.persistence.insert_node(&new_node).await.map_err(|e| {
7840 0 : if matches!(
7841 0 : e,
7842 : crate::persistence::DatabaseError::Query(
7843 : diesel::result::Error::DatabaseError(
7844 : diesel::result::DatabaseErrorKind::UniqueViolation,
7845 : _,
7846 : )
7847 : )
7848 : ) {
7849 : // The node can be deleted by tombstone API, and not show up in the list of nodes.
7850 : // If you see this error, check tombstones first.
7851 0 : ApiError::Conflict(format!("Node {} is already exists", new_node.get_id()))
7852 : } else {
7853 0 : ApiError::from(e)
7854 : }
7855 0 : })?;
7856 : }
7857 : RegistrationStatus::NeedUpdate => {
7858 0 : self.persistence
7859 0 : .update_node_on_registration(
7860 0 : register_req.node_id,
7861 0 : register_req.listen_https_port,
7862 0 : register_req.listen_grpc_addr,
7863 0 : register_req.listen_grpc_port,
7864 0 : )
7865 0 : .await?
7866 : }
7867 0 : _ => unreachable!("Other statuses have been processed earlier"),
7868 : }
7869 :
7870 0 : let mut locked = self.inner.write().unwrap();
7871 0 : let mut new_nodes = (*locked.nodes).clone();
7872 :
7873 0 : locked.scheduler.node_upsert(&new_node);
7874 0 : new_nodes.insert(register_req.node_id, new_node);
7875 :
7876 0 : locked.nodes = Arc::new(new_nodes);
7877 :
7878 0 : metrics::METRICS_REGISTRY
7879 0 : .metrics_group
7880 0 : .storage_controller_pageserver_nodes
7881 0 : .set(locked.nodes.len() as i64);
7882 0 : metrics::METRICS_REGISTRY
7883 0 : .metrics_group
7884 0 : .storage_controller_https_pageserver_nodes
7885 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7886 :
7887 0 : match registration_status {
7888 : RegistrationStatus::New => {
7889 0 : tracing::info!(
7890 0 : "Registered pageserver {} ({}), now have {} pageservers",
7891 : register_req.node_id,
7892 : register_req.availability_zone_id,
7893 0 : locked.nodes.len()
7894 : );
7895 : }
7896 : RegistrationStatus::NeedUpdate => {
7897 0 : tracing::info!(
7898 0 : "Re-registered and updated node {} ({})",
7899 : register_req.node_id,
7900 : register_req.availability_zone_id,
7901 : );
7902 : }
7903 0 : _ => unreachable!("Other statuses have been processed earlier"),
7904 : }
7905 0 : Ok(())
7906 0 : }
7907 :
7908 : /// Configure in-memory and persistent state of a node as requested
7909 : ///
7910 : /// Note that this function does not trigger any immediate side effects in response
7911 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
7912 0 : async fn node_state_configure(
7913 0 : &self,
7914 0 : node_id: NodeId,
7915 0 : availability: Option<NodeAvailability>,
7916 0 : scheduling: Option<NodeSchedulingPolicy>,
7917 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
7918 0 : ) -> Result<AvailabilityTransition, ApiError> {
7919 0 : if let Some(scheduling) = scheduling {
7920 : // Scheduling is a persistent part of Node: we must write updates to the database before
7921 : // applying them in memory
7922 0 : self.persistence
7923 0 : .update_node_scheduling_policy(node_id, scheduling)
7924 0 : .await?;
7925 0 : }
7926 :
7927 : // If we're activating a node, then before setting it active we must reconcile any shard locations
7928 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
7929 : // by calling [`Self::node_activate_reconcile`]
7930 : //
7931 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
7932 : // nothing else can mutate its availability while we run.
7933 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
7934 0 : let (activate_node, availability_transition) = {
7935 0 : let locked = self.inner.read().unwrap();
7936 0 : let Some(node) = locked.nodes.get(&node_id) else {
7937 0 : return Err(ApiError::NotFound(
7938 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7939 0 : ));
7940 : };
7941 :
7942 0 : (
7943 0 : node.clone(),
7944 0 : node.get_availability_transition(input_availability),
7945 0 : )
7946 : };
7947 :
7948 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
7949 0 : self.node_activate_reconcile(activate_node, node_lock)
7950 0 : .await?;
7951 0 : }
7952 0 : availability_transition
7953 : } else {
7954 0 : AvailabilityTransition::Unchanged
7955 : };
7956 :
7957 : // Apply changes from the request to our in-memory state for the Node
7958 0 : let mut locked = self.inner.write().unwrap();
7959 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
7960 :
7961 0 : let mut new_nodes = (**nodes).clone();
7962 :
7963 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
7964 0 : return Err(ApiError::NotFound(
7965 0 : anyhow::anyhow!("Node not registered").into(),
7966 0 : ));
7967 : };
7968 :
7969 0 : if let Some(availability) = availability {
7970 0 : node.set_availability(availability);
7971 0 : }
7972 :
7973 0 : if let Some(scheduling) = scheduling {
7974 0 : node.set_scheduling(scheduling);
7975 0 : }
7976 :
7977 : // Update the scheduler, in case the elegibility of the node for new shards has changed
7978 0 : scheduler.node_upsert(node);
7979 :
7980 0 : let new_nodes = Arc::new(new_nodes);
7981 0 : locked.nodes = new_nodes;
7982 :
7983 0 : Ok(availability_transition)
7984 0 : }
7985 :
7986 : /// Handle availability transition of one node
7987 : ///
7988 : /// Note that you should first call [`Self::node_state_configure`] to update
7989 : /// the in-memory state referencing that node. If you need to handle more than one transition
7990 : /// consider using [`Self::handle_node_availability_transitions`].
7991 0 : async fn handle_node_availability_transition(
7992 0 : &self,
7993 0 : node_id: NodeId,
7994 0 : transition: AvailabilityTransition,
7995 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
7996 0 : ) -> Result<(), ApiError> {
7997 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
7998 0 : match transition {
7999 : AvailabilityTransition::ToOffline => {
8000 0 : tracing::info!("Node {} transition to offline", node_id);
8001 :
8002 0 : let mut locked = self.inner.write().unwrap();
8003 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8004 :
8005 0 : let mut tenants_affected: usize = 0;
8006 :
8007 0 : for (_tenant_id, mut schedule_context, shards) in
8008 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
8009 : {
8010 0 : for tenant_shard in shards {
8011 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
8012 0 : if let Some(observed_loc) =
8013 0 : tenant_shard.observed.locations.get_mut(&node_id)
8014 0 : {
8015 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
8016 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
8017 0 : observed_loc.conf = None;
8018 0 : }
8019 :
8020 0 : if nodes.len() == 1 {
8021 : // Special case for single-node cluster: there is no point trying to reschedule
8022 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
8023 : // failures to schedule them.
8024 0 : continue;
8025 0 : }
8026 :
8027 0 : if !nodes
8028 0 : .values()
8029 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8030 : {
8031 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
8032 : // trying to reschedule since there's nowhere else to go. Without this
8033 : // branch we incorrectly detach tenants in response to node unavailability.
8034 0 : continue;
8035 0 : }
8036 :
8037 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
8038 0 : tenant_shard.sequence = tenant_shard.sequence.next();
8039 :
8040 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
8041 0 : Err(e) => {
8042 : // It is possible that some tenants will become unschedulable when too many pageservers
8043 : // go offline: in this case there isn't much we can do other than make the issue observable.
8044 : // TODO: give TenantShard a scheduling error attribute to be queried later.
8045 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
8046 : }
8047 : Ok(()) => {
8048 0 : if self
8049 0 : .maybe_reconcile_shard(
8050 0 : tenant_shard,
8051 0 : nodes,
8052 0 : ReconcilerPriority::Normal,
8053 0 : )
8054 0 : .is_some()
8055 0 : {
8056 0 : tenants_affected += 1;
8057 0 : };
8058 : }
8059 : }
8060 0 : }
8061 : }
8062 : }
8063 0 : tracing::info!(
8064 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
8065 : tenants_affected,
8066 : node_id
8067 : )
8068 : }
8069 : AvailabilityTransition::ToActive => {
8070 0 : tracing::info!("Node {} transition to active", node_id);
8071 :
8072 0 : let mut locked = self.inner.write().unwrap();
8073 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
8074 :
8075 : // When a node comes back online, we must reconcile any tenant that has a None observed
8076 : // location on the node.
8077 0 : for tenant_shard in tenants.values_mut() {
8078 : // If a reconciliation is already in progress, rely on the previous scheduling
8079 : // decision and skip triggering a new reconciliation.
8080 0 : if tenant_shard.reconciler.is_some() {
8081 0 : continue;
8082 0 : }
8083 :
8084 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
8085 0 : if observed_loc.conf.is_none() {
8086 0 : self.maybe_reconcile_shard(
8087 0 : tenant_shard,
8088 0 : nodes,
8089 0 : ReconcilerPriority::Normal,
8090 0 : );
8091 0 : }
8092 0 : }
8093 : }
8094 :
8095 : // TODO: in the background, we should balance work back onto this pageserver
8096 : }
8097 : // No action required for the intermediate unavailable state.
8098 : // When we transition into active or offline from the unavailable state,
8099 : // the correct handling above will kick in.
8100 : AvailabilityTransition::ToWarmingUpFromActive => {
8101 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
8102 : }
8103 : AvailabilityTransition::ToWarmingUpFromOffline => {
8104 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
8105 : }
8106 : AvailabilityTransition::Unchanged => {
8107 0 : tracing::debug!("Node {} no availability change during config", node_id);
8108 : }
8109 : }
8110 :
8111 0 : Ok(())
8112 0 : }
8113 :
8114 : /// Handle availability transition for multiple nodes
8115 : ///
8116 : /// Note that you should first call [`Self::node_state_configure`] for
8117 : /// all nodes being handled here for the handling to use fresh in-memory state.
8118 0 : async fn handle_node_availability_transitions(
8119 0 : &self,
8120 0 : transitions: Vec<(
8121 0 : NodeId,
8122 0 : TracingExclusiveGuard<NodeOperations>,
8123 0 : AvailabilityTransition,
8124 0 : )>,
8125 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
8126 0 : let mut errors = Vec::default();
8127 0 : for (node_id, node_lock, transition) in transitions {
8128 0 : let res = self
8129 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
8130 0 : .await;
8131 0 : if let Err(err) = res {
8132 0 : errors.push((node_id, err));
8133 0 : }
8134 : }
8135 :
8136 0 : if errors.is_empty() {
8137 0 : Ok(())
8138 : } else {
8139 0 : Err(errors)
8140 : }
8141 0 : }
8142 :
8143 0 : pub(crate) async fn node_configure(
8144 0 : &self,
8145 0 : node_id: NodeId,
8146 0 : availability: Option<NodeAvailability>,
8147 0 : scheduling: Option<NodeSchedulingPolicy>,
8148 0 : ) -> Result<(), ApiError> {
8149 0 : let node_lock =
8150 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
8151 :
8152 0 : let transition = self
8153 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
8154 0 : .await?;
8155 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
8156 0 : .await
8157 0 : }
8158 :
8159 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
8160 : /// operation for HTTP api.
8161 0 : pub(crate) async fn external_node_configure(
8162 0 : &self,
8163 0 : node_id: NodeId,
8164 0 : availability: Option<NodeAvailability>,
8165 0 : scheduling: Option<NodeSchedulingPolicy>,
8166 0 : ) -> Result<(), ApiError> {
8167 : {
8168 0 : let locked = self.inner.read().unwrap();
8169 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
8170 0 : return Err(ApiError::PreconditionFailed(
8171 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
8172 0 : ));
8173 0 : }
8174 : }
8175 :
8176 0 : self.node_configure(node_id, availability, scheduling).await
8177 0 : }
8178 :
8179 0 : pub(crate) async fn start_node_delete(
8180 0 : self: &Arc<Self>,
8181 0 : node_id: NodeId,
8182 0 : force: bool,
8183 0 : ) -> Result<(), ApiError> {
8184 0 : let (ongoing_op, node_policy, schedulable_nodes_count) = {
8185 0 : let locked = self.inner.read().unwrap();
8186 0 : let nodes = &locked.nodes;
8187 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8188 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8189 0 : ))?;
8190 0 : let schedulable_nodes_count = nodes
8191 0 : .iter()
8192 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8193 0 : .count();
8194 :
8195 : (
8196 0 : locked
8197 0 : .ongoing_operation
8198 0 : .as_ref()
8199 0 : .map(|ongoing| ongoing.operation),
8200 0 : node.get_scheduling(),
8201 0 : schedulable_nodes_count,
8202 : )
8203 : };
8204 :
8205 0 : if let Some(ongoing) = ongoing_op {
8206 0 : return Err(ApiError::PreconditionFailed(
8207 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8208 0 : ));
8209 0 : }
8210 :
8211 0 : if schedulable_nodes_count == 0 {
8212 0 : return Err(ApiError::PreconditionFailed(
8213 0 : "No other schedulable nodes to move shards".into(),
8214 0 : ));
8215 0 : }
8216 :
8217 0 : match node_policy {
8218 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
8219 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Deleting))
8220 0 : .await?;
8221 :
8222 0 : let cancel = self.cancel.child_token();
8223 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8224 0 : let policy_on_start = node_policy;
8225 :
8226 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8227 0 : operation: Operation::Delete(Delete { node_id }),
8228 0 : cancel: cancel.clone(),
8229 0 : });
8230 :
8231 0 : let span = tracing::info_span!(parent: None, "delete_node", %node_id);
8232 :
8233 0 : tokio::task::spawn(
8234 : {
8235 0 : let service = self.clone();
8236 0 : let cancel = cancel.clone();
8237 0 : async move {
8238 0 : let _gate_guard = gate_guard;
8239 :
8240 0 : scopeguard::defer! {
8241 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8242 :
8243 : if let Some(Operation::Delete(removed_delete)) = prev.map(|h| h.operation) {
8244 : assert_eq!(removed_delete.node_id, node_id, "We always take the same operation");
8245 : } else {
8246 : panic!("We always remove the same operation")
8247 : }
8248 : }
8249 :
8250 0 : tracing::info!("Delete background operation starting");
8251 0 : let res = service
8252 0 : .delete_node(node_id, policy_on_start, force, cancel)
8253 0 : .await;
8254 0 : match res {
8255 : Ok(()) => {
8256 0 : tracing::info!(
8257 0 : "Delete background operation completed successfully"
8258 : );
8259 : }
8260 : Err(OperationError::Cancelled) => {
8261 0 : tracing::info!("Delete background operation was cancelled");
8262 : }
8263 0 : Err(err) => {
8264 0 : tracing::error!(
8265 0 : "Delete background operation encountered: {err}"
8266 : )
8267 : }
8268 : }
8269 0 : }
8270 : }
8271 0 : .instrument(span),
8272 : );
8273 : }
8274 : NodeSchedulingPolicy::Deleting => {
8275 0 : return Err(ApiError::Conflict(format!(
8276 0 : "Node {node_id} has delete in progress"
8277 0 : )));
8278 : }
8279 0 : policy => {
8280 0 : return Err(ApiError::PreconditionFailed(
8281 0 : format!("Node {node_id} cannot be deleted due to {policy:?} policy").into(),
8282 0 : ));
8283 : }
8284 : }
8285 :
8286 0 : Ok(())
8287 0 : }
8288 :
8289 0 : pub(crate) async fn cancel_node_delete(
8290 0 : self: &Arc<Self>,
8291 0 : node_id: NodeId,
8292 0 : ) -> Result<(), ApiError> {
8293 : {
8294 0 : let locked = self.inner.read().unwrap();
8295 0 : let nodes = &locked.nodes;
8296 0 : nodes.get(&node_id).ok_or(ApiError::NotFound(
8297 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8298 0 : ))?;
8299 : }
8300 :
8301 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8302 0 : if let Operation::Delete(delete) = op_handler.operation {
8303 0 : if delete.node_id == node_id {
8304 0 : tracing::info!("Cancelling background delete operation for node {node_id}");
8305 0 : op_handler.cancel.cancel();
8306 0 : return Ok(());
8307 0 : }
8308 0 : }
8309 0 : }
8310 :
8311 0 : Err(ApiError::PreconditionFailed(
8312 0 : format!("Node {node_id} has no delete in progress").into(),
8313 0 : ))
8314 0 : }
8315 :
8316 0 : pub(crate) async fn start_node_drain(
8317 0 : self: &Arc<Self>,
8318 0 : node_id: NodeId,
8319 0 : ) -> Result<(), ApiError> {
8320 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
8321 0 : let locked = self.inner.read().unwrap();
8322 0 : let nodes = &locked.nodes;
8323 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8324 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8325 0 : ))?;
8326 0 : let schedulable_nodes_count = nodes
8327 0 : .iter()
8328 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8329 0 : .count();
8330 :
8331 : (
8332 0 : locked
8333 0 : .ongoing_operation
8334 0 : .as_ref()
8335 0 : .map(|ongoing| ongoing.operation),
8336 0 : node.is_available(),
8337 0 : node.get_scheduling(),
8338 0 : schedulable_nodes_count,
8339 : )
8340 : };
8341 :
8342 0 : if let Some(ongoing) = ongoing_op {
8343 0 : return Err(ApiError::PreconditionFailed(
8344 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8345 0 : ));
8346 0 : }
8347 :
8348 0 : if !node_available {
8349 0 : return Err(ApiError::ResourceUnavailable(
8350 0 : format!("Node {node_id} is currently unavailable").into(),
8351 0 : ));
8352 0 : }
8353 :
8354 0 : if schedulable_nodes_count == 0 {
8355 0 : return Err(ApiError::PreconditionFailed(
8356 0 : "No other schedulable nodes to drain to".into(),
8357 0 : ));
8358 0 : }
8359 :
8360 0 : match node_policy {
8361 : NodeSchedulingPolicy::Active => {
8362 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
8363 0 : .await?;
8364 :
8365 0 : let cancel = self.cancel.child_token();
8366 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8367 :
8368 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8369 0 : operation: Operation::Drain(Drain { node_id }),
8370 0 : cancel: cancel.clone(),
8371 0 : });
8372 :
8373 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
8374 :
8375 0 : tokio::task::spawn({
8376 0 : let service = self.clone();
8377 0 : let cancel = cancel.clone();
8378 0 : async move {
8379 0 : let _gate_guard = gate_guard;
8380 :
8381 0 : scopeguard::defer! {
8382 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8383 :
8384 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
8385 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
8386 : } else {
8387 : panic!("We always remove the same operation")
8388 : }
8389 : }
8390 :
8391 0 : tracing::info!("Drain background operation starting");
8392 0 : let res = service.drain_node(node_id, cancel).await;
8393 0 : match res {
8394 : Ok(()) => {
8395 0 : tracing::info!("Drain background operation completed successfully");
8396 : }
8397 : Err(OperationError::Cancelled) => {
8398 0 : tracing::info!("Drain background operation was cancelled");
8399 : }
8400 0 : Err(err) => {
8401 0 : tracing::error!("Drain background operation encountered: {err}")
8402 : }
8403 : }
8404 0 : }
8405 0 : }.instrument(span));
8406 : }
8407 : NodeSchedulingPolicy::Draining => {
8408 0 : return Err(ApiError::Conflict(format!(
8409 0 : "Node {node_id} has drain in progress"
8410 0 : )));
8411 : }
8412 0 : policy => {
8413 0 : return Err(ApiError::PreconditionFailed(
8414 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
8415 0 : ));
8416 : }
8417 : }
8418 :
8419 0 : Ok(())
8420 0 : }
8421 :
8422 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
8423 0 : let node_available = {
8424 0 : let locked = self.inner.read().unwrap();
8425 0 : let nodes = &locked.nodes;
8426 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8427 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8428 0 : ))?;
8429 :
8430 0 : node.is_available()
8431 : };
8432 :
8433 0 : if !node_available {
8434 0 : return Err(ApiError::ResourceUnavailable(
8435 0 : format!("Node {node_id} is currently unavailable").into(),
8436 0 : ));
8437 0 : }
8438 :
8439 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8440 0 : if let Operation::Drain(drain) = op_handler.operation {
8441 0 : if drain.node_id == node_id {
8442 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8443 0 : op_handler.cancel.cancel();
8444 0 : return Ok(());
8445 0 : }
8446 0 : }
8447 0 : }
8448 :
8449 0 : Err(ApiError::PreconditionFailed(
8450 0 : format!("Node {node_id} has no drain in progress").into(),
8451 0 : ))
8452 0 : }
8453 :
8454 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
8455 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
8456 0 : let locked = self.inner.read().unwrap();
8457 0 : let nodes = &locked.nodes;
8458 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8459 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8460 0 : ))?;
8461 :
8462 : (
8463 0 : locked
8464 0 : .ongoing_operation
8465 0 : .as_ref()
8466 0 : .map(|ongoing| ongoing.operation),
8467 0 : node.is_available(),
8468 0 : node.get_scheduling(),
8469 0 : nodes.len(),
8470 : )
8471 : };
8472 :
8473 0 : if let Some(ongoing) = ongoing_op {
8474 0 : return Err(ApiError::PreconditionFailed(
8475 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8476 0 : ));
8477 0 : }
8478 :
8479 0 : if !node_available {
8480 0 : return Err(ApiError::ResourceUnavailable(
8481 0 : format!("Node {node_id} is currently unavailable").into(),
8482 0 : ));
8483 0 : }
8484 :
8485 0 : if total_nodes_count <= 1 {
8486 0 : return Err(ApiError::PreconditionFailed(
8487 0 : "No other nodes to fill from".into(),
8488 0 : ));
8489 0 : }
8490 :
8491 0 : match node_policy {
8492 : NodeSchedulingPolicy::Active => {
8493 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
8494 0 : .await?;
8495 :
8496 0 : let cancel = self.cancel.child_token();
8497 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8498 :
8499 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8500 0 : operation: Operation::Fill(Fill { node_id }),
8501 0 : cancel: cancel.clone(),
8502 0 : });
8503 :
8504 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
8505 :
8506 0 : tokio::task::spawn({
8507 0 : let service = self.clone();
8508 0 : let cancel = cancel.clone();
8509 0 : async move {
8510 0 : let _gate_guard = gate_guard;
8511 :
8512 0 : scopeguard::defer! {
8513 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8514 :
8515 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
8516 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
8517 : } else {
8518 : panic!("We always remove the same operation")
8519 : }
8520 : }
8521 :
8522 0 : tracing::info!("Fill background operation starting");
8523 0 : let res = service.fill_node(node_id, cancel).await;
8524 0 : match res {
8525 : Ok(()) => {
8526 0 : tracing::info!("Fill background operation completed successfully");
8527 : }
8528 : Err(OperationError::Cancelled) => {
8529 0 : tracing::info!("Fill background operation was cancelled");
8530 : }
8531 0 : Err(err) => {
8532 0 : tracing::error!("Fill background operation encountered: {err}")
8533 : }
8534 : }
8535 0 : }
8536 0 : }.instrument(span));
8537 : }
8538 : NodeSchedulingPolicy::Filling => {
8539 0 : return Err(ApiError::Conflict(format!(
8540 0 : "Node {node_id} has fill in progress"
8541 0 : )));
8542 : }
8543 0 : policy => {
8544 0 : return Err(ApiError::PreconditionFailed(
8545 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
8546 0 : ));
8547 : }
8548 : }
8549 :
8550 0 : Ok(())
8551 0 : }
8552 :
8553 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
8554 0 : let node_available = {
8555 0 : let locked = self.inner.read().unwrap();
8556 0 : let nodes = &locked.nodes;
8557 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8558 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8559 0 : ))?;
8560 :
8561 0 : node.is_available()
8562 : };
8563 :
8564 0 : if !node_available {
8565 0 : return Err(ApiError::ResourceUnavailable(
8566 0 : format!("Node {node_id} is currently unavailable").into(),
8567 0 : ));
8568 0 : }
8569 :
8570 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8571 0 : if let Operation::Fill(fill) = op_handler.operation {
8572 0 : if fill.node_id == node_id {
8573 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8574 0 : op_handler.cancel.cancel();
8575 0 : return Ok(());
8576 0 : }
8577 0 : }
8578 0 : }
8579 :
8580 0 : Err(ApiError::PreconditionFailed(
8581 0 : format!("Node {node_id} has no fill in progress").into(),
8582 0 : ))
8583 0 : }
8584 :
8585 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
8586 : /// configuration
8587 0 : fn maybe_reconcile_shard(
8588 0 : &self,
8589 0 : shard: &mut TenantShard,
8590 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8591 0 : priority: ReconcilerPriority,
8592 0 : ) -> Option<ReconcilerWaiter> {
8593 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
8594 0 : }
8595 :
8596 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
8597 0 : fn get_reconciler_units(
8598 0 : &self,
8599 0 : priority: ReconcilerPriority,
8600 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
8601 0 : let units = match priority {
8602 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
8603 : ReconcilerPriority::High => {
8604 0 : match self
8605 0 : .priority_reconciler_concurrency
8606 0 : .clone()
8607 0 : .try_acquire_owned()
8608 : {
8609 0 : Ok(u) => Ok(u),
8610 : Err(TryAcquireError::NoPermits) => {
8611 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
8612 : // the normal priority semaphore.
8613 0 : self.reconciler_concurrency.clone().try_acquire_owned()
8614 : }
8615 0 : Err(e) => Err(e),
8616 : }
8617 : }
8618 : };
8619 :
8620 0 : units.map(ReconcileUnits::new)
8621 0 : }
8622 :
8623 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
8624 0 : fn maybe_configured_reconcile_shard(
8625 0 : &self,
8626 0 : shard: &mut TenantShard,
8627 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8628 0 : reconciler_config: ReconcilerConfig,
8629 0 : ) -> Option<ReconcilerWaiter> {
8630 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
8631 :
8632 0 : let reconcile_reason = match reconcile_needed {
8633 0 : ReconcileNeeded::No => return None,
8634 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
8635 0 : ReconcileNeeded::Yes(reason) => {
8636 : // Fall through to try and acquire units for spawning reconciler
8637 0 : reason
8638 : }
8639 : };
8640 :
8641 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
8642 0 : Ok(u) => u,
8643 : Err(_) => {
8644 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
8645 0 : "Concurrency limited: enqueued for reconcile later");
8646 0 : if !shard.delayed_reconcile {
8647 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
8648 0 : Err(TrySendError::Closed(_)) => {
8649 0 : // Weird mid-shutdown case?
8650 0 : }
8651 : Err(TrySendError::Full(_)) => {
8652 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
8653 0 : tracing::warn!(
8654 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
8655 : );
8656 : }
8657 0 : Ok(()) => {
8658 0 : shard.delayed_reconcile = true;
8659 0 : }
8660 : }
8661 0 : }
8662 :
8663 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
8664 : // number to advance. When this function is eventually called again and succeeds in getting units,
8665 : // it will spawn a reconciler that makes this waiter complete.
8666 0 : return Some(shard.future_reconcile_waiter());
8667 : }
8668 : };
8669 :
8670 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
8671 : // Gate closed: we're shutting down, drop out.
8672 0 : return None;
8673 : };
8674 :
8675 0 : shard.spawn_reconciler(
8676 0 : reconcile_reason,
8677 0 : &self.result_tx,
8678 0 : nodes,
8679 0 : &self.compute_hook,
8680 0 : reconciler_config,
8681 0 : &self.config,
8682 0 : &self.persistence,
8683 0 : units,
8684 0 : gate_guard,
8685 0 : &self.reconcilers_cancel,
8686 0 : self.http_client.clone(),
8687 : )
8688 0 : }
8689 :
8690 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
8691 : /// Additionally, reschedule tenants that require it.
8692 : ///
8693 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
8694 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
8695 : /// available. A return value of 0 indicates that everything is fully reconciled already.
8696 0 : fn reconcile_all(&self) -> ReconcileAllResult {
8697 0 : let mut locked = self.inner.write().unwrap();
8698 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8699 0 : let pageservers = nodes.clone();
8700 :
8701 : // This function is an efficient place to update lazy statistics, since we are walking
8702 : // all tenants.
8703 0 : let mut pending_reconciles = 0;
8704 0 : let mut stuck_reconciles = 0;
8705 0 : let mut az_violations = 0;
8706 :
8707 : // If we find any tenants to drop from memory, stash them to offload after
8708 : // we're done traversing the map of tenants.
8709 0 : let mut drop_detached_tenants = Vec::new();
8710 :
8711 0 : let mut spawned_reconciles = 0;
8712 0 : let mut has_delayed_reconciles = false;
8713 :
8714 0 : for shard in tenants.values_mut() {
8715 : // Accumulate scheduling statistics
8716 0 : if let (Some(attached), Some(preferred)) =
8717 0 : (shard.intent.get_attached(), shard.preferred_az())
8718 : {
8719 0 : let node_az = nodes
8720 0 : .get(attached)
8721 0 : .expect("Nodes exist if referenced")
8722 0 : .get_availability_zone_id();
8723 0 : if node_az != preferred {
8724 0 : az_violations += 1;
8725 0 : }
8726 0 : }
8727 :
8728 : // Skip checking if this shard is already enqueued for reconciliation
8729 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
8730 : // If there is something delayed, then return a nonzero count so that
8731 : // callers like reconcile_all_now do not incorrectly get the impression
8732 : // that the system is in a quiescent state.
8733 0 : has_delayed_reconciles = true;
8734 0 : pending_reconciles += 1;
8735 0 : continue;
8736 0 : }
8737 :
8738 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
8739 : // dirty, spawn another one
8740 0 : if self
8741 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
8742 0 : .is_some()
8743 : {
8744 0 : spawned_reconciles += 1;
8745 :
8746 0 : if shard.consecutive_reconciles_count >= MAX_CONSECUTIVE_RECONCILES {
8747 : // Count shards that are stuck, butwe still want to reconcile them.
8748 : // We don't want to consider them when deciding to run optimizations.
8749 0 : tracing::warn!(
8750 : tenant_id=%shard.tenant_shard_id.tenant_id,
8751 0 : shard_id=%shard.tenant_shard_id.shard_slug(),
8752 0 : "Shard reconciliation is stuck: {} consecutive launches",
8753 : shard.consecutive_reconciles_count
8754 : );
8755 0 : stuck_reconciles += 1;
8756 0 : }
8757 : } else {
8758 0 : if shard.delayed_reconcile {
8759 0 : // Shard wanted to reconcile but for some reason couldn't.
8760 0 : pending_reconciles += 1;
8761 0 : }
8762 :
8763 : // Reset the counter when we don't need to launch a reconcile.
8764 0 : shard.consecutive_reconciles_count = 0;
8765 : }
8766 : // If this tenant is detached, try dropping it from memory. This is usually done
8767 : // proactively in [`Self::process_results`], but we do it here to handle the edge
8768 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
8769 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
8770 0 : && shard.policy == PlacementPolicy::Detached
8771 : {
8772 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
8773 0 : shard.tenant_shard_id.tenant_id,
8774 0 : TenantOperations::DropDetached,
8775 0 : ) {
8776 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
8777 0 : }
8778 0 : }
8779 : }
8780 :
8781 : // Some metrics are calculated from SchedulerNode state, update these periodically
8782 0 : scheduler.update_metrics();
8783 :
8784 : // Process any deferred tenant drops
8785 0 : for (tenant_id, guard) in drop_detached_tenants {
8786 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
8787 0 : }
8788 :
8789 0 : metrics::METRICS_REGISTRY
8790 0 : .metrics_group
8791 0 : .storage_controller_schedule_az_violation
8792 0 : .set(az_violations as i64);
8793 :
8794 0 : metrics::METRICS_REGISTRY
8795 0 : .metrics_group
8796 0 : .storage_controller_pending_reconciles
8797 0 : .set(pending_reconciles as i64);
8798 :
8799 0 : metrics::METRICS_REGISTRY
8800 0 : .metrics_group
8801 0 : .storage_controller_stuck_reconciles
8802 0 : .set(stuck_reconciles as i64);
8803 :
8804 0 : ReconcileAllResult::new(spawned_reconciles, stuck_reconciles, has_delayed_reconciles)
8805 0 : }
8806 :
8807 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
8808 : /// could be scheduled somewhere better:
8809 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
8810 : /// * e.g. after a node fails then recovers, to move some work back to it
8811 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
8812 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
8813 : /// we did the split, but are probably better placed elsewhere.
8814 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
8815 : /// * e.g. after a shard split, some locations will be on the same node (where the split
8816 : /// happened), and will probably be better placed elsewhere.
8817 : ///
8818 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
8819 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
8820 : /// according to those same soft constraints.
8821 0 : async fn optimize_all(&self) -> usize {
8822 : // Limit on how many shards' optmizations each call to this function will execute. Combined
8823 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
8824 : // trickle of optimizations in the background, rather than executing a large number in parallel
8825 : // when a change occurs.
8826 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
8827 :
8828 : // Synchronous prepare: scan shards for possible scheduling optimizations
8829 0 : let candidate_work = self.optimize_all_plan();
8830 0 : let candidate_work_len = candidate_work.len();
8831 :
8832 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
8833 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
8834 :
8835 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
8836 :
8837 : // Synchronous apply: update the shards' intent states according to validated optimisations
8838 0 : let mut reconciles_spawned = 0;
8839 0 : let mut optimizations_applied = 0;
8840 0 : let mut locked = self.inner.write().unwrap();
8841 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8842 0 : for (tenant_shard_id, optimization) in validated_work {
8843 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
8844 : // Shard was dropped between planning and execution;
8845 0 : continue;
8846 : };
8847 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
8848 0 : if shard.apply_optimization(scheduler, optimization) {
8849 0 : optimizations_applied += 1;
8850 0 : if self
8851 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
8852 0 : .is_some()
8853 0 : {
8854 0 : reconciles_spawned += 1;
8855 0 : }
8856 0 : }
8857 :
8858 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
8859 0 : break;
8860 0 : }
8861 : }
8862 :
8863 0 : if was_work_filtered {
8864 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
8865 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
8866 0 : // as these validations start passing.
8867 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
8868 0 : }
8869 :
8870 0 : reconciles_spawned
8871 0 : }
8872 :
8873 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
8874 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
8875 : // this higher than the execution limit gives us a chance to execute some work even if the first
8876 : // few optimizations we find are not ready.
8877 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
8878 :
8879 0 : let mut work = Vec::new();
8880 0 : let mut locked = self.inner.write().unwrap();
8881 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
8882 :
8883 : // We are going to plan a bunch of optimisations before applying any of them, so the
8884 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
8885 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
8886 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
8887 : // to ignore the utilisation component of the score.
8888 :
8889 0 : for (_tenant_id, schedule_context, shards) in
8890 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Speculative)
8891 : {
8892 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
8893 0 : break;
8894 0 : }
8895 0 : for shard in shards {
8896 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
8897 0 : break;
8898 0 : }
8899 0 : match shard.get_scheduling_policy() {
8900 0 : ShardSchedulingPolicy::Active => {
8901 0 : // Ok to do optimization
8902 0 : }
8903 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
8904 0 : // Ok to do optimization: we are executing a graceful migration that
8905 0 : // has set preferred_node
8906 0 : }
8907 : ShardSchedulingPolicy::Essential
8908 : | ShardSchedulingPolicy::Pause
8909 : | ShardSchedulingPolicy::Stop => {
8910 : // Policy prevents optimizing this shard.
8911 0 : continue;
8912 : }
8913 : }
8914 :
8915 0 : if !matches!(shard.splitting, SplitState::Idle)
8916 0 : || matches!(shard.policy, PlacementPolicy::Detached)
8917 0 : || shard.reconciler.is_some()
8918 : {
8919 : // Do not start any optimizations while another change to the tenant is ongoing: this
8920 : // is not necessary for correctness, but simplifies operations and implicitly throttles
8921 : // optimization changes to happen in a "trickle" over time.
8922 0 : continue;
8923 0 : }
8924 :
8925 : // Fast path: we may quickly identify shards that don't have any possible optimisations
8926 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
8927 0 : if cfg!(feature = "testing") {
8928 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
8929 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
8930 : // panic in prod if we hit this, or spend cycles on it in prod.
8931 0 : assert!(
8932 0 : shard
8933 0 : .optimize_attachment(scheduler, &schedule_context)
8934 0 : .is_none()
8935 : );
8936 0 : assert!(
8937 0 : shard
8938 0 : .optimize_secondary(scheduler, &schedule_context)
8939 0 : .is_none()
8940 : );
8941 0 : }
8942 0 : continue;
8943 0 : }
8944 :
8945 0 : if let Some(optimization) =
8946 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
8947 : // its primary location based on soft constraints, cut it over.
8948 0 : shard.optimize_attachment(scheduler, &schedule_context)
8949 : {
8950 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
8951 0 : work.push((shard.tenant_shard_id, optimization));
8952 0 : break;
8953 0 : } else if let Some(optimization) =
8954 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
8955 : // better placed on another node, based on ScheduleContext, then adjust it. This
8956 : // covers cases like after a shard split, where we might have too many shards
8957 : // in the same tenant with secondary locations on the node where they originally split.
8958 0 : shard.optimize_secondary(scheduler, &schedule_context)
8959 : {
8960 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
8961 0 : work.push((shard.tenant_shard_id, optimization));
8962 0 : break;
8963 0 : }
8964 : }
8965 : }
8966 :
8967 0 : work
8968 0 : }
8969 :
8970 0 : async fn optimize_all_validate(
8971 0 : &self,
8972 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
8973 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
8974 : // Take a clone of the node map to use outside the lock in async validation phase
8975 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
8976 :
8977 0 : let mut want_secondary_status = Vec::new();
8978 :
8979 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
8980 : // check that the state of locations is acceptable to run the optimization, such as
8981 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
8982 : // in a live migration.
8983 0 : let mut validated_work = Vec::new();
8984 0 : for (tenant_shard_id, optimization) in candidate_work {
8985 0 : match optimization.action {
8986 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
8987 : old_attached_node_id: _,
8988 0 : new_attached_node_id,
8989 : }) => {
8990 0 : match validation_nodes.get(&new_attached_node_id) {
8991 0 : None => {
8992 0 : // Node was dropped between planning and validation
8993 0 : }
8994 0 : Some(node) => {
8995 0 : if !node.is_available() {
8996 0 : tracing::info!(
8997 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
8998 : );
8999 0 : } else {
9000 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
9001 0 : // remote API requests concurrently.
9002 0 : want_secondary_status.push((
9003 0 : tenant_shard_id,
9004 0 : node.clone(),
9005 0 : optimization,
9006 0 : ));
9007 0 : }
9008 : }
9009 : }
9010 : }
9011 : ScheduleOptimizationAction::ReplaceSecondary(_)
9012 : | ScheduleOptimizationAction::CreateSecondary(_)
9013 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
9014 : // No extra checks needed to manage secondaries: this does not interrupt client access
9015 0 : validated_work.push((tenant_shard_id, optimization))
9016 : }
9017 : };
9018 : }
9019 :
9020 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
9021 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
9022 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
9023 0 : let results = self
9024 0 : .tenant_for_shards_api(
9025 0 : want_secondary_status
9026 0 : .iter()
9027 0 : .map(|i| (i.0, i.1.clone()))
9028 0 : .collect(),
9029 0 : |tenant_shard_id, client| async move {
9030 0 : client.tenant_secondary_status(tenant_shard_id).await
9031 0 : },
9032 : 1,
9033 : 1,
9034 : SHORT_RECONCILE_TIMEOUT,
9035 0 : &self.cancel,
9036 : )
9037 0 : .await;
9038 :
9039 0 : for ((tenant_shard_id, node, optimization), (_, secondary_status)) in
9040 0 : want_secondary_status.into_iter().zip(results.into_iter())
9041 : {
9042 0 : match secondary_status {
9043 0 : Err(e) => {
9044 0 : tracing::info!(
9045 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
9046 : );
9047 : }
9048 0 : Ok(progress) => {
9049 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
9050 : // them in an optimization
9051 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
9052 :
9053 0 : if progress.heatmap_mtime.is_none()
9054 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
9055 0 : && progress.bytes_downloaded != progress.bytes_total
9056 0 : || progress.bytes_total - progress.bytes_downloaded
9057 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
9058 : {
9059 0 : tracing::info!(
9060 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
9061 : );
9062 :
9063 0 : if progress.heatmap_mtime.is_none() {
9064 : // No heatmap might mean the attached location has never uploaded one, or that
9065 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
9066 : // but fairly common in tests.
9067 0 : self.kick_secondary_download(tenant_shard_id).await;
9068 0 : }
9069 : } else {
9070 : // Location looks ready: proceed
9071 0 : tracing::info!(
9072 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
9073 : );
9074 0 : validated_work.push((tenant_shard_id, optimization))
9075 : }
9076 : }
9077 : }
9078 : }
9079 :
9080 0 : validated_work
9081 0 : }
9082 :
9083 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
9084 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
9085 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
9086 : /// we have this helper to move things along faster.
9087 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
9088 0 : if !self.config.kick_secondary_downloads {
9089 : // No-op if kick_secondary_downloads functionaliuty is not configured
9090 0 : return;
9091 0 : }
9092 :
9093 0 : let (attached_node, secondaries) = {
9094 0 : let locked = self.inner.read().unwrap();
9095 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
9096 0 : tracing::warn!(
9097 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
9098 : );
9099 0 : return;
9100 : };
9101 :
9102 0 : let Some(attached) = shard.intent.get_attached() else {
9103 0 : tracing::warn!(
9104 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
9105 : );
9106 0 : return;
9107 : };
9108 :
9109 0 : let secondaries = shard
9110 0 : .intent
9111 0 : .get_secondary()
9112 0 : .iter()
9113 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
9114 0 : .collect::<Vec<_>>();
9115 :
9116 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
9117 : };
9118 :
9119 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
9120 : // a 'kick' to let scheduling optimisation run more promptly.
9121 0 : match attached_node
9122 0 : .with_client_retries(
9123 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
9124 0 : &self.http_client,
9125 0 : &self.config.pageserver_jwt_token,
9126 : 3,
9127 : 10,
9128 : SHORT_RECONCILE_TIMEOUT,
9129 0 : &self.cancel,
9130 : )
9131 0 : .await
9132 : {
9133 0 : Some(Err(e)) => {
9134 0 : tracing::info!(
9135 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
9136 : );
9137 : }
9138 : None => {
9139 0 : tracing::info!(
9140 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
9141 : );
9142 : }
9143 : Some(Ok(_)) => {
9144 0 : tracing::info!(
9145 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
9146 : );
9147 : }
9148 : }
9149 :
9150 0 : for secondary_node in secondaries {
9151 0 : match secondary_node
9152 0 : .with_client_retries(
9153 0 : |client| async move {
9154 0 : client
9155 0 : .tenant_secondary_download(
9156 0 : tenant_shard_id,
9157 0 : Some(Duration::from_secs(1)),
9158 0 : )
9159 0 : .await
9160 0 : },
9161 0 : &self.http_client,
9162 0 : &self.config.pageserver_jwt_token,
9163 : 3,
9164 : 10,
9165 : SHORT_RECONCILE_TIMEOUT,
9166 0 : &self.cancel,
9167 : )
9168 0 : .await
9169 : {
9170 0 : Some(Err(e)) => {
9171 0 : tracing::info!(
9172 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
9173 : );
9174 : }
9175 : None => {
9176 0 : tracing::info!(
9177 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
9178 : );
9179 : }
9180 0 : Some(Ok(progress)) => {
9181 0 : tracing::info!(
9182 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
9183 : );
9184 : }
9185 : }
9186 : }
9187 0 : }
9188 :
9189 : /// Asynchronously split a tenant that's eligible for automatic splits. At most one tenant will
9190 : /// be split per call.
9191 : ///
9192 : /// Two sets of criteria are used: initial splits and size-based splits (in that order).
9193 : /// Initial splits are used to eagerly split unsharded tenants that may be performing initial
9194 : /// ingestion, since sharded tenants have significantly better ingestion throughput. Size-based
9195 : /// splits are used to bound the maximum shard size and balance out load.
9196 : ///
9197 : /// Splits are based on max_logical_size, i.e. the logical size of the largest timeline in a
9198 : /// tenant. We use this instead of the total logical size because branches will duplicate
9199 : /// logical size without actually using more storage. We could also use visible physical size,
9200 : /// but this might overestimate tenants that frequently churn branches.
9201 : ///
9202 : /// Initial splits (initial_split_threshold):
9203 : /// * Applies to tenants with 1 shard.
9204 : /// * The largest timeline (max_logical_size) exceeds initial_split_threshold.
9205 : /// * Splits into initial_split_shards.
9206 : ///
9207 : /// Size-based splits (split_threshold):
9208 : /// * Applies to all tenants.
9209 : /// * The largest timeline (max_logical_size) divided by shard count exceeds split_threshold.
9210 : /// * Splits such that max_logical_size / shard_count <= split_threshold, in powers of 2.
9211 : ///
9212 : /// Tenant shards are ordered by descending max_logical_size, first initial split candidates
9213 : /// then size-based split candidates. The first matching candidate is split.
9214 : ///
9215 : /// The shard count is clamped to max_split_shards. If a candidate is eligible for both initial
9216 : /// and size-based splits, the largest shard count will be used.
9217 : ///
9218 : /// An unsharded tenant will get DEFAULT_STRIPE_SIZE, regardless of what its ShardIdentity says.
9219 : /// A sharded tenant will retain its stripe size, as splits do not allow changing it.
9220 : ///
9221 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
9222 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
9223 : /// block all other splits.
9224 0 : async fn autosplit_tenants(self: &Arc<Self>) {
9225 : // If max_split_shards is set to 0 or 1, we can't split.
9226 0 : let max_split_shards = self.config.max_split_shards;
9227 0 : if max_split_shards <= 1 {
9228 0 : return;
9229 0 : }
9230 :
9231 : // If initial_split_shards is set to 0 or 1, disable initial splits.
9232 0 : let mut initial_split_threshold = self.config.initial_split_threshold.unwrap_or(0);
9233 0 : let initial_split_shards = self.config.initial_split_shards;
9234 0 : if initial_split_shards <= 1 {
9235 0 : initial_split_threshold = 0;
9236 0 : }
9237 :
9238 : // If no split_threshold nor initial_split_threshold, disable autosplits.
9239 0 : let split_threshold = self.config.split_threshold.unwrap_or(0);
9240 0 : if split_threshold == 0 && initial_split_threshold == 0 {
9241 0 : return;
9242 0 : }
9243 :
9244 : // Fetch split candidates in prioritized order.
9245 : //
9246 : // If initial splits are enabled, fetch eligible tenants first. We prioritize initial splits
9247 : // over size-based splits, since these are often performing initial ingestion and rely on
9248 : // splits to improve ingest throughput.
9249 0 : let mut candidates = Vec::new();
9250 :
9251 0 : if initial_split_threshold > 0 {
9252 : // Initial splits: fetch tenants with 1 shard where the logical size of the largest
9253 : // timeline exceeds the initial split threshold.
9254 0 : let initial_candidates = self
9255 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9256 0 : order_by: TenantSorting::MaxLogicalSize,
9257 0 : limit: 10,
9258 0 : where_shards_lt: Some(ShardCount(2)),
9259 0 : where_gt: Some(initial_split_threshold),
9260 0 : })
9261 0 : .await;
9262 0 : candidates.extend(initial_candidates);
9263 0 : }
9264 :
9265 0 : if split_threshold > 0 {
9266 : // Size-based splits: fetch tenants where the logical size of the largest timeline
9267 : // divided by shard count exceeds the split threshold.
9268 : //
9269 : // max_logical_size is only tracked on shard 0, and contains the total logical size
9270 : // across all shards. We have to order and filter by MaxLogicalSizePerShard, i.e.
9271 : // max_logical_size / shard_count, such that we only receive tenants that are actually
9272 : // eligible for splits. But we still use max_logical_size for later split calculations.
9273 0 : let size_candidates = self
9274 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9275 0 : order_by: TenantSorting::MaxLogicalSizePerShard,
9276 0 : limit: 10,
9277 0 : where_shards_lt: Some(ShardCount(max_split_shards)),
9278 0 : where_gt: Some(split_threshold),
9279 0 : })
9280 0 : .await;
9281 : #[cfg(feature = "testing")]
9282 0 : assert!(
9283 0 : size_candidates.iter().all(|c| c.id.is_shard_zero()),
9284 0 : "MaxLogicalSizePerShard returned non-zero shard: {size_candidates:?}",
9285 : );
9286 0 : candidates.extend(size_candidates);
9287 0 : }
9288 :
9289 : // Filter out tenants in a prohibiting scheduling modes
9290 : // and tenants with an ongoing import.
9291 : //
9292 : // Note that the import check here is oportunistic. An import might start
9293 : // after the check before we actually update [`TenantShard::splitting`].
9294 : // [`Self::tenant_shard_split`] checks the database whilst holding the exclusive
9295 : // tenant lock. Imports might take a long time, so the check here allows us
9296 : // to split something else instead of trying the same shard over and over.
9297 : {
9298 0 : let state = self.inner.read().unwrap();
9299 0 : candidates.retain(|i| {
9300 0 : let shard = state.tenants.get(&i.id);
9301 0 : match shard {
9302 0 : Some(t) => {
9303 0 : t.get_scheduling_policy() == ShardSchedulingPolicy::Active
9304 0 : && t.importing == TimelineImportState::Idle
9305 : }
9306 0 : None => false,
9307 : }
9308 0 : });
9309 : }
9310 :
9311 : // Pick the first candidate to split. This will generally always be the first one in
9312 : // candidates, but we defensively skip candidates that end up not actually splitting.
9313 0 : let Some((candidate, new_shard_count)) = candidates
9314 0 : .into_iter()
9315 0 : .filter_map(|candidate| {
9316 0 : let new_shard_count = Self::compute_split_shards(ShardSplitInputs {
9317 0 : shard_count: candidate.id.shard_count,
9318 0 : max_logical_size: candidate.max_logical_size,
9319 0 : split_threshold,
9320 0 : max_split_shards,
9321 0 : initial_split_threshold,
9322 0 : initial_split_shards,
9323 0 : });
9324 0 : new_shard_count.map(|shards| (candidate, shards.count()))
9325 0 : })
9326 0 : .next()
9327 : else {
9328 0 : debug!("no split-eligible tenants found");
9329 0 : return;
9330 : };
9331 :
9332 : // Retain the stripe size of sharded tenants, as splits don't allow changing it. Otherwise,
9333 : // use DEFAULT_STRIPE_SIZE for unsharded tenants -- their stripe size doesn't really matter,
9334 : // and if we change the default stripe size we want to use the new default rather than an
9335 : // old, persisted stripe size.
9336 0 : let new_stripe_size = match candidate.id.shard_count.count() {
9337 0 : 0 => panic!("invalid shard count 0"),
9338 0 : 1 => Some(DEFAULT_STRIPE_SIZE),
9339 0 : 2.. => None,
9340 : };
9341 :
9342 : // We spawn a task to run this, so it's exactly like some external API client requesting
9343 : // it. We don't want to block the background reconcile loop on this.
9344 0 : let old_shard_count = candidate.id.shard_count.count();
9345 0 : info!(
9346 0 : "auto-splitting tenant {old_shard_count} → {new_shard_count} shards, \
9347 0 : current size {candidate:?} (split_threshold={split_threshold} \
9348 0 : initial_split_threshold={initial_split_threshold})"
9349 : );
9350 :
9351 0 : let this = self.clone();
9352 0 : tokio::spawn(
9353 0 : async move {
9354 0 : match this
9355 0 : .tenant_shard_split(
9356 0 : candidate.id.tenant_id,
9357 0 : TenantShardSplitRequest {
9358 0 : new_shard_count,
9359 0 : new_stripe_size,
9360 0 : },
9361 0 : )
9362 0 : .await
9363 : {
9364 : Ok(_) => {
9365 0 : info!("successful auto-split {old_shard_count} → {new_shard_count} shards")
9366 : }
9367 0 : Err(err) => error!("auto-split failed: {err}"),
9368 : }
9369 0 : }
9370 0 : .instrument(info_span!("auto_split", tenant_id=%candidate.id.tenant_id)),
9371 : );
9372 0 : }
9373 :
9374 : /// Returns the number of shards to split a tenant into, or None if the tenant shouldn't split,
9375 : /// based on the total logical size of the largest timeline summed across all shards. Uses the
9376 : /// larger of size-based and initial splits, clamped to max_split_shards.
9377 : ///
9378 : /// NB: the thresholds are exclusive, since TopTenantShardsRequest uses where_gt.
9379 25 : fn compute_split_shards(inputs: ShardSplitInputs) -> Option<ShardCount> {
9380 : let ShardSplitInputs {
9381 25 : shard_count,
9382 25 : max_logical_size,
9383 25 : split_threshold,
9384 25 : max_split_shards,
9385 25 : initial_split_threshold,
9386 25 : initial_split_shards,
9387 25 : } = inputs;
9388 :
9389 25 : let mut new_shard_count: u8 = shard_count.count();
9390 :
9391 : // Size-based splits. Ensures max_logical_size / new_shard_count <= split_threshold, using
9392 : // power-of-two shard counts.
9393 : //
9394 : // If the current shard count is not a power of two, and does not exceed split_threshold,
9395 : // then we leave it alone rather than forcing a power-of-two split.
9396 25 : if split_threshold > 0
9397 18 : && max_logical_size.div_ceil(split_threshold) > shard_count.count() as u64
9398 12 : {
9399 12 : new_shard_count = max_logical_size
9400 12 : .div_ceil(split_threshold)
9401 12 : .checked_next_power_of_two()
9402 12 : .unwrap_or(u8::MAX as u64)
9403 12 : .try_into()
9404 12 : .unwrap_or(u8::MAX);
9405 13 : }
9406 :
9407 : // Initial splits. Use the larger of size-based and initial split shard counts. This only
9408 : // applies to unsharded tenants, i.e. changes to initial_split_threshold or
9409 : // initial_split_shards are not retroactive for sharded tenants.
9410 25 : if initial_split_threshold > 0
9411 14 : && shard_count.count() <= 1
9412 11 : && max_logical_size > initial_split_threshold
9413 8 : {
9414 8 : new_shard_count = new_shard_count.max(initial_split_shards);
9415 17 : }
9416 :
9417 : // Clamp to max shards.
9418 25 : new_shard_count = new_shard_count.min(max_split_shards);
9419 :
9420 : // Don't split if we're not increasing the shard count.
9421 25 : if new_shard_count <= shard_count.count() {
9422 10 : return None;
9423 15 : }
9424 :
9425 15 : Some(ShardCount(new_shard_count))
9426 25 : }
9427 :
9428 : /// Fetches the top tenant shards from every available node, in descending order of
9429 : /// max logical size. Offline nodes are skipped, and any errors from available nodes
9430 : /// will be logged and ignored.
9431 0 : async fn get_top_tenant_shards(
9432 0 : &self,
9433 0 : request: &TopTenantShardsRequest,
9434 0 : ) -> Vec<TopTenantShardItem> {
9435 0 : let nodes = self
9436 0 : .inner
9437 0 : .read()
9438 0 : .unwrap()
9439 0 : .nodes
9440 0 : .values()
9441 0 : .filter(|node| node.is_available())
9442 0 : .cloned()
9443 0 : .collect_vec();
9444 :
9445 0 : let mut futures = FuturesUnordered::new();
9446 0 : for node in nodes {
9447 0 : futures.push(async move {
9448 0 : node.with_client_retries(
9449 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
9450 0 : &self.http_client,
9451 0 : &self.config.pageserver_jwt_token,
9452 : 3,
9453 : 3,
9454 0 : Duration::from_secs(5),
9455 0 : &self.cancel,
9456 : )
9457 0 : .await
9458 0 : });
9459 : }
9460 :
9461 0 : let mut top = Vec::new();
9462 0 : while let Some(output) = futures.next().await {
9463 0 : match output {
9464 0 : Some(Ok(response)) => top.extend(response.shards),
9465 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
9466 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
9467 0 : None => {} // node is shutting down
9468 : }
9469 : }
9470 :
9471 0 : top.sort_by_key(|i| i.max_logical_size);
9472 0 : top.reverse();
9473 0 : top
9474 0 : }
9475 :
9476 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
9477 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
9478 : /// put the system into a quiescent state where future background reconciliations won't do anything.
9479 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
9480 0 : let reconcile_all_result = self.reconcile_all();
9481 0 : let mut spawned_reconciles = reconcile_all_result.spawned_reconciles;
9482 0 : if reconcile_all_result.can_run_optimizations() {
9483 : // Only optimize when we are otherwise idle
9484 0 : let optimization_reconciles = self.optimize_all().await;
9485 0 : spawned_reconciles += optimization_reconciles;
9486 0 : }
9487 :
9488 0 : let waiters = {
9489 0 : let mut waiters = Vec::new();
9490 0 : let locked = self.inner.read().unwrap();
9491 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
9492 0 : if let Some(waiter) = shard.get_waiter() {
9493 0 : waiters.push(waiter);
9494 0 : }
9495 : }
9496 0 : waiters
9497 : };
9498 :
9499 0 : let waiter_count = waiters.len();
9500 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
9501 0 : Ok(()) => {}
9502 0 : Err(e) => {
9503 0 : if let ReconcileWaitError::Failed(_, reconcile_error) = &e {
9504 0 : match **reconcile_error {
9505 : ReconcileError::Cancel
9506 0 : | ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
9507 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
9508 0 : // because some other change superceded it. We will return a nonzero number,
9509 0 : // so the caller knows they might have to call again to quiesce the system.
9510 0 : }
9511 : _ => {
9512 0 : return Err(e);
9513 : }
9514 : }
9515 : } else {
9516 0 : return Err(e);
9517 : }
9518 : }
9519 : };
9520 :
9521 0 : tracing::info!(
9522 0 : "{} reconciles in reconcile_all, {} waiters",
9523 : spawned_reconciles,
9524 : waiter_count
9525 : );
9526 :
9527 0 : Ok(std::cmp::max(waiter_count, spawned_reconciles))
9528 0 : }
9529 :
9530 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
9531 : // Cancel all on-going reconciles and wait for them to exit the gate.
9532 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
9533 0 : self.reconcilers_cancel.cancel();
9534 0 : self.reconcilers_gate.close().await;
9535 :
9536 : // Signal the background loop in [`Service::process_results`] to exit once
9537 : // it has proccessed the results from all the reconciles we cancelled earlier.
9538 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
9539 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
9540 0 : self.result_tx.closed().await;
9541 0 : }
9542 :
9543 0 : pub async fn shutdown(&self) {
9544 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
9545 0 : .await;
9546 :
9547 : // Background tasks hold gate guards: this notifies them of the cancellation and
9548 : // waits for them all to complete.
9549 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
9550 0 : self.cancel.cancel();
9551 0 : self.gate.close().await;
9552 0 : }
9553 :
9554 : /// Spot check the download lag for a secondary location of a shard.
9555 : /// Should be used as a heuristic, since it's not always precise: the
9556 : /// secondary might have not downloaded the new heat map yet and, hence,
9557 : /// is not aware of the lag.
9558 : ///
9559 : /// Returns:
9560 : /// * Ok(None) if the lag could not be determined from the status,
9561 : /// * Ok(Some(_)) if the lag could be determind
9562 : /// * Err on failures to query the pageserver.
9563 0 : async fn secondary_lag(
9564 0 : &self,
9565 0 : secondary: &NodeId,
9566 0 : tenant_shard_id: TenantShardId,
9567 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
9568 0 : let nodes = self.inner.read().unwrap().nodes.clone();
9569 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
9570 0 : StatusCode::NOT_FOUND,
9571 0 : format!("Node with id {secondary} not found"),
9572 0 : ))?;
9573 :
9574 0 : match node
9575 0 : .with_client_retries(
9576 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
9577 0 : &self.http_client,
9578 0 : &self.config.pageserver_jwt_token,
9579 : 1,
9580 : 3,
9581 0 : Duration::from_millis(250),
9582 0 : &self.cancel,
9583 : )
9584 0 : .await
9585 : {
9586 0 : Some(Ok(status)) => match status.heatmap_mtime {
9587 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
9588 0 : None => Ok(None),
9589 : },
9590 0 : Some(Err(e)) => Err(e),
9591 0 : None => Err(mgmt_api::Error::Cancelled),
9592 : }
9593 0 : }
9594 :
9595 : /// Drain a node by moving the shards attached to it as primaries.
9596 : /// This is a long running operation and it should run as a separate Tokio task.
9597 0 : pub(crate) async fn drain_node(
9598 0 : self: &Arc<Self>,
9599 0 : node_id: NodeId,
9600 0 : cancel: CancellationToken,
9601 0 : ) -> Result<(), OperationError> {
9602 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
9603 0 : let max_secondary_lag_bytes = self
9604 0 : .config
9605 0 : .max_secondary_lag_bytes
9606 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
9607 :
9608 : // By default, live migrations are generous about the wait time for getting
9609 : // the secondary location up to speed. When draining, give up earlier in order
9610 : // to not stall the operation when a cold secondary is encountered.
9611 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9612 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9613 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9614 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9615 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9616 0 : .build();
9617 :
9618 0 : let mut waiters = Vec::new();
9619 :
9620 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
9621 :
9622 0 : while !tid_iter.finished() {
9623 0 : if cancel.is_cancelled() {
9624 0 : match self
9625 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9626 0 : .await
9627 : {
9628 0 : Ok(()) => return Err(OperationError::Cancelled),
9629 0 : Err(err) => {
9630 0 : return Err(OperationError::FinalizeError(
9631 0 : format!(
9632 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9633 0 : )
9634 0 : .into(),
9635 0 : ));
9636 : }
9637 : }
9638 0 : }
9639 :
9640 0 : operation_utils::validate_node_state(
9641 0 : &node_id,
9642 0 : self.inner.read().unwrap().nodes.clone(),
9643 0 : NodeSchedulingPolicy::Draining,
9644 0 : )?;
9645 :
9646 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9647 0 : let tid = match tid_iter.next() {
9648 0 : Some(tid) => tid,
9649 : None => {
9650 0 : break;
9651 : }
9652 : };
9653 :
9654 0 : let tid_drain = TenantShardDrain {
9655 0 : drained_node: node_id,
9656 0 : tenant_shard_id: tid,
9657 0 : };
9658 :
9659 0 : let drain_action = {
9660 0 : let locked = self.inner.read().unwrap();
9661 0 : tid_drain.tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
9662 : };
9663 :
9664 0 : let dest_node_id = match drain_action {
9665 0 : TenantShardDrainAction::RescheduleToSecondary(dest_node_id) => dest_node_id,
9666 0 : TenantShardDrainAction::Reconcile(intent_node_id) => intent_node_id,
9667 : TenantShardDrainAction::Skip => {
9668 0 : continue;
9669 : }
9670 : };
9671 :
9672 0 : match self.secondary_lag(&dest_node_id, tid).await {
9673 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
9674 0 : // The secondary is reasonably up to date.
9675 0 : // Migrate to it
9676 0 : }
9677 0 : Ok(Some(lag)) => {
9678 0 : tracing::info!(
9679 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9680 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
9681 : );
9682 0 : continue;
9683 : }
9684 : Ok(None) => {
9685 0 : tracing::info!(
9686 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9687 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
9688 : );
9689 0 : continue;
9690 : }
9691 0 : Err(err) => {
9692 0 : tracing::warn!(
9693 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9694 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
9695 : );
9696 0 : continue;
9697 : }
9698 : }
9699 :
9700 : {
9701 0 : let mut locked = self.inner.write().unwrap();
9702 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9703 :
9704 0 : let tenant_shard = match drain_action {
9705 0 : TenantShardDrainAction::RescheduleToSecondary(dest_node_id) => tid_drain
9706 0 : .reschedule_to_secondary(dest_node_id, tenants, scheduler, nodes)?,
9707 0 : TenantShardDrainAction::Reconcile(_) => tenants.get_mut(&tid),
9708 : // Note: Unreachable, handled above.
9709 0 : TenantShardDrainAction::Skip => None,
9710 : };
9711 :
9712 0 : if let Some(tenant_shard) = tenant_shard {
9713 0 : let waiter = self.maybe_configured_reconcile_shard(
9714 0 : tenant_shard,
9715 0 : nodes,
9716 0 : reconciler_config,
9717 0 : );
9718 0 : if let Some(some) = waiter {
9719 0 : waiters.push(some);
9720 0 : }
9721 0 : }
9722 : }
9723 : }
9724 :
9725 0 : waiters = self
9726 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9727 0 : .await;
9728 :
9729 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
9730 : }
9731 :
9732 0 : while !waiters.is_empty() {
9733 0 : if cancel.is_cancelled() {
9734 0 : match self
9735 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9736 0 : .await
9737 : {
9738 0 : Ok(()) => return Err(OperationError::Cancelled),
9739 0 : Err(err) => {
9740 0 : return Err(OperationError::FinalizeError(
9741 0 : format!(
9742 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9743 0 : )
9744 0 : .into(),
9745 0 : ));
9746 : }
9747 : }
9748 0 : }
9749 :
9750 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
9751 :
9752 0 : waiters = self
9753 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9754 0 : .await;
9755 : }
9756 :
9757 : // At this point we have done the best we could to drain shards from this node.
9758 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
9759 : // to complete the drain.
9760 0 : if let Err(err) = self
9761 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
9762 0 : .await
9763 : {
9764 : // This is not fatal. Anything that is polling the node scheduling policy to detect
9765 : // the end of the drain operations will hang, but all such places should enforce an
9766 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
9767 : // by the counterpart fill operation.
9768 0 : return Err(OperationError::FinalizeError(
9769 0 : format!(
9770 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
9771 0 : )
9772 0 : .into(),
9773 0 : ));
9774 0 : }
9775 :
9776 0 : Ok(())
9777 0 : }
9778 :
9779 : /// Create a node fill plan (pick secondaries to promote), based on:
9780 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
9781 : /// outside their home AZ, should be migrated back here.
9782 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
9783 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
9784 : /// those shards have a home AZ that doesn't match the node we're filling.
9785 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
9786 0 : let mut locked = self.inner.write().unwrap();
9787 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
9788 :
9789 0 : let node_az = nodes
9790 0 : .get(&node_id)
9791 0 : .expect("Node must exist")
9792 0 : .get_availability_zone_id()
9793 0 : .clone();
9794 :
9795 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
9796 0 : let mut plan = Vec::new();
9797 :
9798 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
9799 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
9800 :
9801 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
9802 : // conceivably come up in real life if deploying a single-AZ region intentionally.
9803 0 : let respect_azs = nodes
9804 0 : .values()
9805 0 : .map(|n| n.get_availability_zone_id())
9806 0 : .unique()
9807 0 : .count()
9808 : > 1;
9809 :
9810 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
9811 : // requires it.
9812 0 : for (tsid, tenant_shard) in tenants {
9813 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9814 : // Shard doesn't have a secondary on this node, ignore it.
9815 0 : continue;
9816 0 : }
9817 :
9818 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
9819 : // shards which belong on this node, not to promote shards whose scheduling preference
9820 : // would be on their currently attached node. So will avoid promoting shards whose
9821 : // home AZ doesn't match the AZ of the node we're filling.
9822 0 : match tenant_shard.preferred_az() {
9823 0 : _ if !respect_azs => {
9824 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9825 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9826 0 : }
9827 : }
9828 : None => {
9829 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
9830 : // will only do so if our target shard count requires it.
9831 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9832 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9833 0 : }
9834 : }
9835 0 : Some(az) if az == &node_az => {
9836 : // This shard's home AZ is equal to the node we're filling: it should
9837 : // be moved back to this node as part of filling, unless its currently
9838 : // attached location is also in its home AZ.
9839 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9840 0 : if nodes
9841 0 : .get(primary)
9842 0 : .expect("referenced node must exist")
9843 0 : .get_availability_zone_id()
9844 0 : != tenant_shard
9845 0 : .preferred_az()
9846 0 : .expect("tenant must have an AZ preference")
9847 : {
9848 0 : plan.push(*tsid)
9849 0 : }
9850 : } else {
9851 0 : plan.push(*tsid)
9852 : }
9853 : }
9854 0 : Some(_) => {
9855 0 : // This shard's home AZ is somewhere other than the node we're filling,
9856 0 : // it may not be moved back to this node as part of filling. Ignore it
9857 0 : }
9858 : }
9859 : }
9860 :
9861 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
9862 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
9863 :
9864 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
9865 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
9866 :
9867 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
9868 :
9869 0 : for (node_id, attached) in nodes_by_load {
9870 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
9871 0 : if !available {
9872 0 : continue;
9873 0 : }
9874 :
9875 0 : if plan.len() >= fill_requirement
9876 0 : || free_tids_by_node.is_empty()
9877 0 : || attached <= expected_attached
9878 : {
9879 0 : break;
9880 0 : }
9881 :
9882 0 : let can_take = attached - expected_attached;
9883 0 : let needed = fill_requirement - plan.len();
9884 0 : let mut take = std::cmp::min(can_take, needed);
9885 :
9886 0 : let mut remove_node = false;
9887 0 : while take > 0 {
9888 0 : match free_tids_by_node.get_mut(&node_id) {
9889 0 : Some(tids) => match tids.pop() {
9890 0 : Some(tid) => {
9891 0 : let max_promote_for_tenant = std::cmp::max(
9892 0 : tid.shard_count.count() as usize / locked.nodes.len(),
9893 : 1,
9894 : );
9895 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
9896 0 : if *promoted < max_promote_for_tenant {
9897 0 : plan.push(tid);
9898 0 : *promoted += 1;
9899 0 : take -= 1;
9900 0 : }
9901 : }
9902 : None => {
9903 0 : remove_node = true;
9904 0 : break;
9905 : }
9906 : },
9907 : None => {
9908 0 : break;
9909 : }
9910 : }
9911 : }
9912 :
9913 0 : if remove_node {
9914 0 : free_tids_by_node.remove(&node_id);
9915 0 : }
9916 : }
9917 :
9918 0 : plan
9919 0 : }
9920 :
9921 : /// Fill a node by promoting its secondaries until the cluster is balanced
9922 : /// with regards to attached shard counts. Note that this operation only
9923 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
9924 : /// This is a long running operation and it should run as a separate Tokio task.
9925 0 : pub(crate) async fn fill_node(
9926 0 : &self,
9927 0 : node_id: NodeId,
9928 0 : cancel: CancellationToken,
9929 0 : ) -> Result<(), OperationError> {
9930 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9931 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9932 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9933 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9934 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9935 0 : .build();
9936 :
9937 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
9938 0 : let mut waiters = Vec::new();
9939 :
9940 : // Execute the plan we've composed above. Before aplying each move from the plan,
9941 : // we validate to ensure that it has not gone stale in the meantime.
9942 0 : while !tids_to_promote.is_empty() {
9943 0 : if cancel.is_cancelled() {
9944 0 : match self
9945 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9946 0 : .await
9947 : {
9948 0 : Ok(()) => return Err(OperationError::Cancelled),
9949 0 : Err(err) => {
9950 0 : return Err(OperationError::FinalizeError(
9951 0 : format!(
9952 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9953 0 : )
9954 0 : .into(),
9955 0 : ));
9956 : }
9957 : }
9958 0 : }
9959 :
9960 : {
9961 0 : let mut locked = self.inner.write().unwrap();
9962 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9963 :
9964 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
9965 0 : format!("node {node_id} was removed").into(),
9966 0 : ))?;
9967 :
9968 0 : let current_policy = node.get_scheduling();
9969 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
9970 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
9971 : // about it
9972 0 : return Err(OperationError::NodeStateChanged(
9973 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
9974 0 : ));
9975 0 : }
9976 :
9977 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9978 0 : if let Some(tid) = tids_to_promote.pop() {
9979 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
9980 : // If the node being filled is not a secondary anymore,
9981 : // skip the promotion.
9982 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9983 0 : continue;
9984 0 : }
9985 :
9986 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
9987 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
9988 0 : Err(e) => {
9989 0 : tracing::warn!(
9990 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9991 0 : "Scheduling error when filling pageserver {} : {e}", node_id
9992 : );
9993 : }
9994 : Ok(()) => {
9995 0 : tracing::info!(
9996 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9997 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
9998 : node_id,
9999 : previously_attached_to,
10000 : node_id
10001 : );
10002 :
10003 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
10004 0 : tenant_shard,
10005 0 : nodes,
10006 0 : reconciler_config,
10007 0 : ) {
10008 0 : waiters.push(waiter);
10009 0 : }
10010 : }
10011 : }
10012 0 : }
10013 : } else {
10014 0 : break;
10015 : }
10016 : }
10017 : }
10018 :
10019 0 : waiters = self
10020 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
10021 0 : .await;
10022 : }
10023 :
10024 0 : while !waiters.is_empty() {
10025 0 : if cancel.is_cancelled() {
10026 0 : match self
10027 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
10028 0 : .await
10029 : {
10030 0 : Ok(()) => return Err(OperationError::Cancelled),
10031 0 : Err(err) => {
10032 0 : return Err(OperationError::FinalizeError(
10033 0 : format!(
10034 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
10035 0 : )
10036 0 : .into(),
10037 0 : ));
10038 : }
10039 : }
10040 0 : }
10041 :
10042 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
10043 :
10044 0 : waiters = self
10045 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
10046 0 : .await;
10047 : }
10048 :
10049 0 : if let Err(err) = self
10050 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
10051 0 : .await
10052 : {
10053 : // This isn't a huge issue since the filling process starts upon request. However, it
10054 : // will prevent the next drain from starting. The only case in which this can fail
10055 : // is database unavailability. Such a case will require manual intervention.
10056 0 : return Err(OperationError::FinalizeError(
10057 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
10058 0 : .into(),
10059 0 : ));
10060 0 : }
10061 :
10062 0 : Ok(())
10063 0 : }
10064 :
10065 : /// Updates scrubber metadata health check results.
10066 0 : pub(crate) async fn metadata_health_update(
10067 0 : &self,
10068 0 : update_req: MetadataHealthUpdateRequest,
10069 0 : ) -> Result<(), ApiError> {
10070 0 : let now = chrono::offset::Utc::now();
10071 0 : let (healthy_records, unhealthy_records) = {
10072 0 : let locked = self.inner.read().unwrap();
10073 0 : let healthy_records = update_req
10074 0 : .healthy_tenant_shards
10075 0 : .into_iter()
10076 : // Retain only health records associated with tenant shards managed by storage controller.
10077 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
10078 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
10079 0 : .collect();
10080 0 : let unhealthy_records = update_req
10081 0 : .unhealthy_tenant_shards
10082 0 : .into_iter()
10083 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
10084 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
10085 0 : .collect();
10086 :
10087 0 : (healthy_records, unhealthy_records)
10088 : };
10089 :
10090 0 : self.persistence
10091 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
10092 0 : .await?;
10093 0 : Ok(())
10094 0 : }
10095 :
10096 : /// Lists the tenant shards that has unhealthy metadata status.
10097 0 : pub(crate) async fn metadata_health_list_unhealthy(
10098 0 : &self,
10099 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
10100 0 : let result = self
10101 0 : .persistence
10102 0 : .list_unhealthy_metadata_health_records()
10103 0 : .await?
10104 0 : .iter()
10105 0 : .map(|p| p.get_tenant_shard_id().unwrap())
10106 0 : .collect();
10107 :
10108 0 : Ok(result)
10109 0 : }
10110 :
10111 : /// Lists the tenant shards that have not been scrubbed for some duration.
10112 0 : pub(crate) async fn metadata_health_list_outdated(
10113 0 : &self,
10114 0 : not_scrubbed_for: Duration,
10115 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
10116 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
10117 0 : let result = self
10118 0 : .persistence
10119 0 : .list_outdated_metadata_health_records(earlier)
10120 0 : .await?
10121 0 : .into_iter()
10122 0 : .map(|record| record.into())
10123 0 : .collect();
10124 0 : Ok(result)
10125 0 : }
10126 :
10127 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
10128 0 : self.inner.read().unwrap().get_leadership_status()
10129 0 : }
10130 :
10131 : /// Handler for step down requests
10132 : ///
10133 : /// Step down runs in separate task since once it's called it should
10134 : /// be driven to completion. Subsequent requests will wait on the same
10135 : /// step down task.
10136 0 : pub(crate) async fn step_down(self: &Arc<Self>) -> GlobalObservedState {
10137 0 : let handle = self.step_down_barrier.get_or_init(|| {
10138 0 : let step_down_self = self.clone();
10139 0 : let (tx, rx) = tokio::sync::watch::channel::<Option<GlobalObservedState>>(None);
10140 0 : tokio::spawn(async move {
10141 0 : let state = step_down_self.step_down_task().await;
10142 0 : tx.send(Some(state))
10143 0 : .expect("Task Arc<Service> keeps receiver alive");
10144 0 : });
10145 :
10146 0 : rx
10147 0 : });
10148 :
10149 0 : handle
10150 0 : .clone()
10151 0 : .wait_for(|observed_state| observed_state.is_some())
10152 0 : .await
10153 0 : .expect("Task Arc<Service> keeps sender alive")
10154 0 : .deref()
10155 0 : .clone()
10156 0 : .expect("Checked above")
10157 0 : }
10158 :
10159 0 : async fn step_down_task(&self) -> GlobalObservedState {
10160 0 : tracing::info!("Received step down request from peer");
10161 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
10162 :
10163 0 : self.inner.write().unwrap().step_down();
10164 :
10165 0 : let stop_reconciliations =
10166 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown);
10167 0 : let mut stop_reconciliations = std::pin::pin!(stop_reconciliations);
10168 :
10169 0 : let started_at = Instant::now();
10170 :
10171 : // Wait for reconciliations to stop and warn if that's taking a long time
10172 : loop {
10173 0 : tokio::select! {
10174 0 : _ = &mut stop_reconciliations => {
10175 0 : tracing::info!("Reconciliations stopped, proceeding with step down");
10176 0 : break;
10177 : }
10178 0 : _ = tokio::time::sleep(Duration::from_secs(10)) => {
10179 0 : tracing::warn!(
10180 0 : elapsed_sec=%started_at.elapsed().as_secs(),
10181 0 : "Stopping reconciliations during step down is taking too long"
10182 : );
10183 : }
10184 : }
10185 : }
10186 :
10187 0 : let mut global_observed = GlobalObservedState::default();
10188 0 : let locked = self.inner.read().unwrap();
10189 0 : for (tid, tenant_shard) in locked.tenants.iter() {
10190 0 : global_observed
10191 0 : .0
10192 0 : .insert(*tid, tenant_shard.observed.clone());
10193 0 : }
10194 :
10195 0 : global_observed
10196 0 : }
10197 :
10198 0 : pub(crate) async fn update_shards_preferred_azs(
10199 0 : &self,
10200 0 : req: ShardsPreferredAzsRequest,
10201 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
10202 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
10203 0 : let updated = self
10204 0 : .persistence
10205 0 : .set_tenant_shard_preferred_azs(preferred_azs)
10206 0 : .await
10207 0 : .map_err(|err| {
10208 0 : ApiError::InternalServerError(anyhow::anyhow!(
10209 0 : "Failed to persist preferred AZs: {err}"
10210 0 : ))
10211 0 : })?;
10212 :
10213 0 : let mut updated_in_mem_and_db = Vec::default();
10214 :
10215 0 : let mut locked = self.inner.write().unwrap();
10216 0 : let state = locked.deref_mut();
10217 0 : for (tid, az_id) in updated {
10218 0 : let shard = state.tenants.get_mut(&tid);
10219 0 : if let Some(shard) = shard {
10220 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
10221 0 : updated_in_mem_and_db.push(tid);
10222 0 : }
10223 : }
10224 :
10225 0 : Ok(ShardsPreferredAzsResponse {
10226 0 : updated: updated_in_mem_and_db,
10227 0 : })
10228 0 : }
10229 : }
10230 :
10231 : #[cfg(test)]
10232 : mod tests {
10233 : use super::*;
10234 :
10235 : /// Tests Service::compute_split_shards. For readability, this specifies sizes in GBs rather
10236 : /// than bytes. Note that max_logical_size is the total logical size of the largest timeline
10237 : /// summed across all shards.
10238 : #[test]
10239 1 : fn compute_split_shards() {
10240 : // Size-based split: two shards have a 500 GB timeline, which need to split into 8 shards
10241 : // that are <= 64 GB,
10242 1 : assert_eq!(
10243 1 : Service::compute_split_shards(ShardSplitInputs {
10244 1 : shard_count: ShardCount(2),
10245 1 : max_logical_size: 500,
10246 1 : split_threshold: 64,
10247 1 : max_split_shards: 16,
10248 1 : initial_split_threshold: 0,
10249 1 : initial_split_shards: 0,
10250 1 : }),
10251 : Some(ShardCount(8))
10252 : );
10253 :
10254 : // Size-based split: noop at or below threshold, fires above.
10255 1 : assert_eq!(
10256 1 : Service::compute_split_shards(ShardSplitInputs {
10257 1 : shard_count: ShardCount(2),
10258 1 : max_logical_size: 127,
10259 1 : split_threshold: 64,
10260 1 : max_split_shards: 16,
10261 1 : initial_split_threshold: 0,
10262 1 : initial_split_shards: 0,
10263 1 : }),
10264 : None,
10265 : );
10266 1 : assert_eq!(
10267 1 : Service::compute_split_shards(ShardSplitInputs {
10268 1 : shard_count: ShardCount(2),
10269 1 : max_logical_size: 128,
10270 1 : split_threshold: 64,
10271 1 : max_split_shards: 16,
10272 1 : initial_split_threshold: 0,
10273 1 : initial_split_shards: 0,
10274 1 : }),
10275 : None,
10276 : );
10277 1 : assert_eq!(
10278 1 : Service::compute_split_shards(ShardSplitInputs {
10279 1 : shard_count: ShardCount(2),
10280 1 : max_logical_size: 129,
10281 1 : split_threshold: 64,
10282 1 : max_split_shards: 16,
10283 1 : initial_split_threshold: 0,
10284 1 : initial_split_shards: 0,
10285 1 : }),
10286 : Some(ShardCount(4)),
10287 : );
10288 :
10289 : // Size-based split: clamped to max_split_shards.
10290 1 : assert_eq!(
10291 1 : Service::compute_split_shards(ShardSplitInputs {
10292 1 : shard_count: ShardCount(2),
10293 1 : max_logical_size: 10000,
10294 1 : split_threshold: 64,
10295 1 : max_split_shards: 16,
10296 1 : initial_split_threshold: 0,
10297 1 : initial_split_shards: 0,
10298 1 : }),
10299 : Some(ShardCount(16))
10300 : );
10301 :
10302 : // Size-based split: tenant already at or beyond max_split_shards is not split.
10303 1 : assert_eq!(
10304 1 : Service::compute_split_shards(ShardSplitInputs {
10305 1 : shard_count: ShardCount(16),
10306 1 : max_logical_size: 10000,
10307 1 : split_threshold: 64,
10308 1 : max_split_shards: 16,
10309 1 : initial_split_threshold: 0,
10310 1 : initial_split_shards: 0,
10311 1 : }),
10312 : None
10313 : );
10314 :
10315 1 : assert_eq!(
10316 1 : Service::compute_split_shards(ShardSplitInputs {
10317 1 : shard_count: ShardCount(32),
10318 1 : max_logical_size: 10000,
10319 1 : split_threshold: 64,
10320 1 : max_split_shards: 16,
10321 1 : initial_split_threshold: 0,
10322 1 : initial_split_shards: 0,
10323 1 : }),
10324 : None
10325 : );
10326 :
10327 : // Size-based split: a non-power-of-2 shard count is normalized to power-of-2 if it
10328 : // exceeds split_threshold (i.e. a 3-shard tenant splits into 8, not 6).
10329 1 : assert_eq!(
10330 1 : Service::compute_split_shards(ShardSplitInputs {
10331 1 : shard_count: ShardCount(3),
10332 1 : max_logical_size: 320,
10333 1 : split_threshold: 64,
10334 1 : max_split_shards: 16,
10335 1 : initial_split_threshold: 0,
10336 1 : initial_split_shards: 0,
10337 1 : }),
10338 : Some(ShardCount(8))
10339 : );
10340 :
10341 : // Size-based split: a non-power-of-2 shard count is not normalized to power-of-2 if the
10342 : // existing shards are below or at split_threshold, but splits into 4 if it exceeds it.
10343 1 : assert_eq!(
10344 1 : Service::compute_split_shards(ShardSplitInputs {
10345 1 : shard_count: ShardCount(3),
10346 1 : max_logical_size: 191,
10347 1 : split_threshold: 64,
10348 1 : max_split_shards: 16,
10349 1 : initial_split_threshold: 0,
10350 1 : initial_split_shards: 0,
10351 1 : }),
10352 : None
10353 : );
10354 1 : assert_eq!(
10355 1 : Service::compute_split_shards(ShardSplitInputs {
10356 1 : shard_count: ShardCount(3),
10357 1 : max_logical_size: 192,
10358 1 : split_threshold: 64,
10359 1 : max_split_shards: 16,
10360 1 : initial_split_threshold: 0,
10361 1 : initial_split_shards: 0,
10362 1 : }),
10363 : None
10364 : );
10365 1 : assert_eq!(
10366 1 : Service::compute_split_shards(ShardSplitInputs {
10367 1 : shard_count: ShardCount(3),
10368 1 : max_logical_size: 193,
10369 1 : split_threshold: 64,
10370 1 : max_split_shards: 16,
10371 1 : initial_split_threshold: 0,
10372 1 : initial_split_shards: 0,
10373 1 : }),
10374 : Some(ShardCount(4))
10375 : );
10376 :
10377 : // Initial split: tenant has a 10 GB timeline, split into 4 shards.
10378 1 : assert_eq!(
10379 1 : Service::compute_split_shards(ShardSplitInputs {
10380 1 : shard_count: ShardCount(1),
10381 1 : max_logical_size: 10,
10382 1 : split_threshold: 0,
10383 1 : max_split_shards: 16,
10384 1 : initial_split_threshold: 8,
10385 1 : initial_split_shards: 4,
10386 1 : }),
10387 : Some(ShardCount(4))
10388 : );
10389 :
10390 : // Initial split: 0 ShardCount is equivalent to 1.
10391 1 : assert_eq!(
10392 1 : Service::compute_split_shards(ShardSplitInputs {
10393 1 : shard_count: ShardCount(0),
10394 1 : max_logical_size: 10,
10395 1 : split_threshold: 0,
10396 1 : max_split_shards: 16,
10397 1 : initial_split_threshold: 8,
10398 1 : initial_split_shards: 4,
10399 1 : }),
10400 : Some(ShardCount(4))
10401 : );
10402 :
10403 : // Initial split: at or below threshold is noop.
10404 1 : assert_eq!(
10405 1 : Service::compute_split_shards(ShardSplitInputs {
10406 1 : shard_count: ShardCount(1),
10407 1 : max_logical_size: 7,
10408 1 : split_threshold: 0,
10409 1 : max_split_shards: 16,
10410 1 : initial_split_threshold: 8,
10411 1 : initial_split_shards: 4,
10412 1 : }),
10413 : None,
10414 : );
10415 1 : assert_eq!(
10416 1 : Service::compute_split_shards(ShardSplitInputs {
10417 1 : shard_count: ShardCount(1),
10418 1 : max_logical_size: 8,
10419 1 : split_threshold: 0,
10420 1 : max_split_shards: 16,
10421 1 : initial_split_threshold: 8,
10422 1 : initial_split_shards: 4,
10423 1 : }),
10424 : None,
10425 : );
10426 1 : assert_eq!(
10427 1 : Service::compute_split_shards(ShardSplitInputs {
10428 1 : shard_count: ShardCount(1),
10429 1 : max_logical_size: 9,
10430 1 : split_threshold: 0,
10431 1 : max_split_shards: 16,
10432 1 : initial_split_threshold: 8,
10433 1 : initial_split_shards: 4,
10434 1 : }),
10435 : Some(ShardCount(4))
10436 : );
10437 :
10438 : // Initial split: already sharded tenant is not affected, even if above threshold and below
10439 : // shard count.
10440 1 : assert_eq!(
10441 1 : Service::compute_split_shards(ShardSplitInputs {
10442 1 : shard_count: ShardCount(2),
10443 1 : max_logical_size: 20,
10444 1 : split_threshold: 0,
10445 1 : max_split_shards: 16,
10446 1 : initial_split_threshold: 8,
10447 1 : initial_split_shards: 4,
10448 1 : }),
10449 : None,
10450 : );
10451 :
10452 : // Initial split: clamped to max_shards.
10453 1 : assert_eq!(
10454 1 : Service::compute_split_shards(ShardSplitInputs {
10455 1 : shard_count: ShardCount(1),
10456 1 : max_logical_size: 10,
10457 1 : split_threshold: 0,
10458 1 : max_split_shards: 3,
10459 1 : initial_split_threshold: 8,
10460 1 : initial_split_shards: 4,
10461 1 : }),
10462 : Some(ShardCount(3)),
10463 : );
10464 :
10465 : // Initial+size split: tenant eligible for both will use the larger shard count.
10466 1 : assert_eq!(
10467 1 : Service::compute_split_shards(ShardSplitInputs {
10468 1 : shard_count: ShardCount(1),
10469 1 : max_logical_size: 10,
10470 1 : split_threshold: 64,
10471 1 : max_split_shards: 16,
10472 1 : initial_split_threshold: 8,
10473 1 : initial_split_shards: 4,
10474 1 : }),
10475 : Some(ShardCount(4)),
10476 : );
10477 1 : assert_eq!(
10478 1 : Service::compute_split_shards(ShardSplitInputs {
10479 1 : shard_count: ShardCount(1),
10480 1 : max_logical_size: 500,
10481 1 : split_threshold: 64,
10482 1 : max_split_shards: 16,
10483 1 : initial_split_threshold: 8,
10484 1 : initial_split_shards: 4,
10485 1 : }),
10486 : Some(ShardCount(8)),
10487 : );
10488 :
10489 : // Initial+size split: sharded tenant is only eligible for size-based split.
10490 1 : assert_eq!(
10491 1 : Service::compute_split_shards(ShardSplitInputs {
10492 1 : shard_count: ShardCount(2),
10493 1 : max_logical_size: 200,
10494 1 : split_threshold: 64,
10495 1 : max_split_shards: 16,
10496 1 : initial_split_threshold: 8,
10497 1 : initial_split_shards: 8,
10498 1 : }),
10499 : Some(ShardCount(4)),
10500 : );
10501 :
10502 : // Initial+size split: uses the larger shard count even with initial_split_threshold above
10503 : // split_threshold.
10504 1 : assert_eq!(
10505 1 : Service::compute_split_shards(ShardSplitInputs {
10506 1 : shard_count: ShardCount(1),
10507 1 : max_logical_size: 10,
10508 1 : split_threshold: 4,
10509 1 : max_split_shards: 16,
10510 1 : initial_split_threshold: 8,
10511 1 : initial_split_shards: 8,
10512 1 : }),
10513 : Some(ShardCount(8)),
10514 : );
10515 :
10516 : // Test backwards compatibility with production settings when initial/size-based splits were
10517 : // rolled out: a single split into 8 shards at 64 GB. Any already sharded tenants with <8
10518 : // shards will split according to split_threshold.
10519 1 : assert_eq!(
10520 1 : Service::compute_split_shards(ShardSplitInputs {
10521 1 : shard_count: ShardCount(1),
10522 1 : max_logical_size: 65,
10523 1 : split_threshold: 64,
10524 1 : max_split_shards: 8,
10525 1 : initial_split_threshold: 64,
10526 1 : initial_split_shards: 8,
10527 1 : }),
10528 : Some(ShardCount(8)),
10529 : );
10530 :
10531 1 : assert_eq!(
10532 1 : Service::compute_split_shards(ShardSplitInputs {
10533 1 : shard_count: ShardCount(1),
10534 1 : max_logical_size: 64,
10535 1 : split_threshold: 64,
10536 1 : max_split_shards: 8,
10537 1 : initial_split_threshold: 64,
10538 1 : initial_split_shards: 8,
10539 1 : }),
10540 : None,
10541 : );
10542 :
10543 1 : assert_eq!(
10544 1 : Service::compute_split_shards(ShardSplitInputs {
10545 1 : shard_count: ShardCount(2),
10546 1 : max_logical_size: 129,
10547 1 : split_threshold: 64,
10548 1 : max_split_shards: 8,
10549 1 : initial_split_threshold: 64,
10550 1 : initial_split_shards: 8,
10551 1 : }),
10552 : Some(ShardCount(4)),
10553 : );
10554 1 : }
10555 : }
|
