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 std::borrow::Cow;
8 : use std::cmp::Ordering;
9 : use std::collections::{BTreeMap, HashMap, HashSet};
10 : use std::error::Error;
11 : use std::num::NonZeroU32;
12 : use std::ops::{Deref, DerefMut};
13 : use std::path::PathBuf;
14 : use std::str::FromStr;
15 : use std::sync::{Arc, OnceLock};
16 : use std::time::{Duration, Instant, SystemTime};
17 :
18 : use anyhow::Context;
19 : use control_plane::storage_controller::{
20 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
21 : };
22 : use diesel::result::DatabaseErrorKind;
23 : use futures::StreamExt;
24 : use futures::stream::FuturesUnordered;
25 : use http_utils::error::ApiError;
26 : use hyper::Uri;
27 : use itertools::Itertools;
28 : use pageserver_api::config::PostHogConfig;
29 : use pageserver_api::controller_api::{
30 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
31 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
32 : ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
33 : SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse, TenantCreateResponseShard,
34 : TenantDescribeResponse, TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
35 : TenantShardMigrateRequest, TenantShardMigrateResponse, TenantTimelineDescribeResponse,
36 : };
37 : use pageserver_api::models::{
38 : self, DetachBehavior, LocationConfig, LocationConfigListResponse, LocationConfigMode, LsnLease,
39 : PageserverUtilization, SecondaryProgress, ShardImportStatus, ShardParameters, TenantConfig,
40 : TenantConfigPatchRequest, TenantConfigRequest, TenantLocationConfigRequest,
41 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
42 : TenantShardSplitResponse, TenantSorting, TenantTimeTravelRequest,
43 : TimelineArchivalConfigRequest, TimelineCreateRequest, TimelineCreateResponseStorcon,
44 : TimelineInfo, TopTenantShardItem, TopTenantShardsRequest,
45 : };
46 : use pageserver_api::shard::{
47 : DEFAULT_STRIPE_SIZE, ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
48 : };
49 : use pageserver_api::upcall_api::{
50 : PutTimelineImportStatusRequest, ReAttachRequest, ReAttachResponse, ReAttachResponseTenant,
51 : TimelineImportStatusRequest, ValidateRequest, ValidateResponse, ValidateResponseTenant,
52 : };
53 : use pageserver_client::{BlockUnblock, mgmt_api};
54 : use reqwest::{Certificate, StatusCode};
55 : use safekeeper_api::models::SafekeeperUtilization;
56 : use safekeeper_reconciler::SafekeeperReconcilers;
57 : use tenant_shard_iterator::{TenantShardExclusiveIterator, create_shared_shard_iterator};
58 : use tokio::sync::TryAcquireError;
59 : use tokio::sync::mpsc::error::TrySendError;
60 : use tokio_util::sync::CancellationToken;
61 : use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
62 : use utils::completion::Barrier;
63 : use utils::env;
64 : use utils::generation::Generation;
65 : use utils::id::{NodeId, TenantId, TimelineId};
66 : use utils::lsn::Lsn;
67 : use utils::shard::ShardIndex;
68 : use utils::sync::gate::{Gate, GateGuard};
69 : use utils::{failpoint_support, pausable_failpoint};
70 :
71 : use crate::background_node_operations::{
72 : Delete, Drain, Fill, MAX_RECONCILES_PER_OPERATION, Operation, OperationError, OperationHandler,
73 : };
74 : use crate::compute_hook::{self, ComputeHook, NotifyError};
75 : use crate::heartbeater::{Heartbeater, PageserverState, SafekeeperState};
76 : use crate::id_lock_map::{
77 : IdLockMap, TracingExclusiveGuard, trace_exclusive_lock, trace_shared_lock,
78 : };
79 : use crate::leadership::Leadership;
80 : use crate::metrics;
81 : use crate::node::{AvailabilityTransition, Node};
82 : use crate::operation_utils::{self, TenantShardDrain};
83 : use crate::pageserver_client::PageserverClient;
84 : use crate::peer_client::GlobalObservedState;
85 : use crate::persistence::split_state::SplitState;
86 : use crate::persistence::{
87 : AbortShardSplitStatus, ControllerPersistence, DatabaseError, DatabaseResult,
88 : MetadataHealthPersistence, Persistence, ShardGenerationState, TenantFilter,
89 : TenantShardPersistence,
90 : };
91 : use crate::reconciler::{
92 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder, ReconcilerPriority,
93 : attached_location_conf,
94 : };
95 : use crate::safekeeper::Safekeeper;
96 : use crate::scheduler::{
97 : AttachedShardTag, MaySchedule, ScheduleContext, ScheduleError, ScheduleMode, Scheduler,
98 : };
99 : use crate::tenant_shard::{
100 : IntentState, MigrateAttachment, ObservedState, ObservedStateDelta, ObservedStateLocation,
101 : ReconcileNeeded, ReconcileResult, ReconcileWaitError, ReconcilerStatus, ReconcilerWaiter,
102 : ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
103 : };
104 : use crate::timeline_import::{
105 : FinalizingImport, ImportResult, ShardImportStatuses, TimelineImport,
106 : TimelineImportFinalizeError, TimelineImportState, UpcallClient,
107 : };
108 :
109 : const WAITER_OPERATION_POLL_TIMEOUT: Duration = Duration::from_millis(500);
110 :
111 : // For operations that should be quick, like attaching a new tenant
112 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
113 :
114 : // For operations that might be slow, like migrating a tenant with
115 : // some data in it.
116 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
117 :
118 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
119 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
120 : // input generation from future requests as authoritative.
121 : const INITIAL_GENERATION: Generation = Generation::new(0);
122 :
123 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
124 : /// up on unresponsive pageservers and proceed.
125 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
126 :
127 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
128 : /// This must be long enough to cover node restarts as well as normal operations: in future
129 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
130 :
131 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
132 : /// offline.
133 : ///
134 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
135 : /// handling of the re-attach response may take a long time and blocks heartbeats from
136 : /// being handled on the pageserver side.
137 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
138 :
139 : /// How often to send heartbeats to registered nodes?
140 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
141 :
142 : /// How long is too long for a reconciliation?
143 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
144 :
145 : #[derive(Clone, strum_macros::Display)]
146 : enum TenantOperations {
147 : Create,
148 : LocationConfig,
149 : ConfigSet,
150 : ConfigPatch,
151 : TimeTravelRemoteStorage,
152 : Delete,
153 : UpdatePolicy,
154 : ShardSplit,
155 : SecondaryDownload,
156 : TimelineCreate,
157 : TimelineDelete,
158 : AttachHook,
159 : TimelineArchivalConfig,
160 : TimelineDetachAncestor,
161 : TimelineGcBlockUnblock,
162 : DropDetached,
163 : DownloadHeatmapLayers,
164 : TimelineLsnLease,
165 : TimelineSafekeeperMigrate,
166 : }
167 :
168 : #[derive(Clone, strum_macros::Display)]
169 : enum NodeOperations {
170 : Register,
171 : Configure,
172 : Delete,
173 : DeleteTombstone,
174 : }
175 :
176 : /// The leadership status for the storage controller process.
177 : /// Allowed transitions are:
178 : /// 1. Leader -> SteppedDown
179 : /// 2. Candidate -> Leader
180 : #[derive(
181 : Eq,
182 : PartialEq,
183 : Copy,
184 : Clone,
185 : strum_macros::Display,
186 : strum_macros::EnumIter,
187 : measured::FixedCardinalityLabel,
188 : )]
189 : #[strum(serialize_all = "snake_case")]
190 : pub(crate) enum LeadershipStatus {
191 : /// This is the steady state where the storage controller can produce
192 : /// side effects in the cluster.
193 : Leader,
194 : /// We've been notified to step down by another candidate. No reconciliations
195 : /// take place in this state.
196 : SteppedDown,
197 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
198 : #[allow(unused)]
199 : Candidate,
200 : }
201 :
202 : enum ShardGenerationValidity {
203 : Valid,
204 : Mismatched {
205 : claimed: Generation,
206 : actual: Option<Generation>,
207 : },
208 : }
209 :
210 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
211 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
212 : pub const SAFEKEEPER_RECONCILER_CONCURRENCY_DEFAULT: usize = 32;
213 :
214 : // Number of consecutive reconciliations that have occurred for one shard,
215 : // after which the shard is ignored when considering to run optimizations.
216 : const MAX_CONSECUTIVE_RECONCILES: usize = 10;
217 :
218 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
219 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
220 : // than they're being pushed onto the queue.
221 : const MAX_DELAYED_RECONCILES: usize = 10000;
222 :
223 : // Top level state available to all HTTP handlers
224 : struct ServiceState {
225 : leadership_status: LeadershipStatus,
226 :
227 : tenants: BTreeMap<TenantShardId, TenantShard>,
228 :
229 : nodes: Arc<HashMap<NodeId, Node>>,
230 :
231 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
232 :
233 : safekeeper_reconcilers: SafekeeperReconcilers,
234 :
235 : scheduler: Scheduler,
236 :
237 : /// Ongoing background operation on the cluster if any is running.
238 : /// Note that only one such operation may run at any given time,
239 : /// hence the type choice.
240 : ongoing_operation: Option<OperationHandler>,
241 :
242 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
243 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
244 :
245 : /// Tracks ongoing timeline import finalization tasks
246 : imports_finalizing: BTreeMap<(TenantId, TimelineId), FinalizingImport>,
247 : }
248 :
249 : /// Transform an error from a pageserver into an error to return to callers of a storage
250 : /// controller API.
251 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
252 0 : match e {
253 0 : mgmt_api::Error::SendRequest(e) => {
254 : // Presume errors sending requests are connectivity/availability issues
255 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
256 : }
257 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
258 : // Presume errors receiving body are connectivity/availability issues
259 0 : ApiError::ResourceUnavailable(
260 0 : format!("{node} error receiving error body: {str}").into(),
261 0 : )
262 : }
263 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
264 : // Return 500 for decoding errors.
265 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
266 : }
267 0 : mgmt_api::Error::ReceiveBody(err) => {
268 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
269 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
270 0 : ApiError::ResourceUnavailable(
271 0 : format!("{node} error receiving error body: {err} {src_str}").into(),
272 0 : )
273 : }
274 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
275 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
276 : }
277 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
278 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
279 : }
280 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
281 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
282 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
283 : // internal server errors, showing that something is wrong with the pageserver or
284 : // storage controller's auth configuration.
285 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
286 : }
287 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
288 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
289 : // turn ask our clients to wait + retry
290 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
291 : }
292 0 : mgmt_api::Error::ApiError(status, msg) => {
293 : // Presume general case of pageserver API errors is that we tried to do something
294 : // that can't be done right now.
295 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
296 : }
297 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
298 0 : mgmt_api::Error::Timeout(e) => ApiError::Timeout(e.into()),
299 : }
300 0 : }
301 :
302 : impl ServiceState {
303 0 : fn new(
304 0 : nodes: HashMap<NodeId, Node>,
305 0 : safekeepers: HashMap<NodeId, Safekeeper>,
306 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
307 0 : scheduler: Scheduler,
308 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
309 0 : initial_leadership_status: LeadershipStatus,
310 0 : reconcilers_cancel: CancellationToken,
311 0 : ) -> Self {
312 0 : metrics::update_leadership_status(initial_leadership_status);
313 :
314 0 : Self {
315 0 : leadership_status: initial_leadership_status,
316 0 : tenants,
317 0 : nodes: Arc::new(nodes),
318 0 : safekeepers: Arc::new(safekeepers),
319 0 : safekeeper_reconcilers: SafekeeperReconcilers::new(reconcilers_cancel),
320 0 : scheduler,
321 0 : ongoing_operation: None,
322 0 : delayed_reconcile_rx,
323 0 : imports_finalizing: Default::default(),
324 0 : }
325 0 : }
326 :
327 0 : fn parts_mut(
328 0 : &mut self,
329 0 : ) -> (
330 0 : &mut Arc<HashMap<NodeId, Node>>,
331 0 : &mut BTreeMap<TenantShardId, TenantShard>,
332 0 : &mut Scheduler,
333 0 : ) {
334 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
335 0 : }
336 :
337 : #[allow(clippy::type_complexity)]
338 0 : fn parts_mut_sk(
339 0 : &mut self,
340 0 : ) -> (
341 0 : &mut Arc<HashMap<NodeId, Node>>,
342 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
343 0 : &mut BTreeMap<TenantShardId, TenantShard>,
344 0 : &mut Scheduler,
345 0 : ) {
346 0 : (
347 0 : &mut self.nodes,
348 0 : &mut self.safekeepers,
349 0 : &mut self.tenants,
350 0 : &mut self.scheduler,
351 0 : )
352 0 : }
353 :
354 0 : fn get_leadership_status(&self) -> LeadershipStatus {
355 0 : self.leadership_status
356 0 : }
357 :
358 0 : fn step_down(&mut self) {
359 0 : self.leadership_status = LeadershipStatus::SteppedDown;
360 0 : metrics::update_leadership_status(self.leadership_status);
361 0 : }
362 :
363 0 : fn become_leader(&mut self) {
364 0 : self.leadership_status = LeadershipStatus::Leader;
365 0 : metrics::update_leadership_status(self.leadership_status);
366 0 : }
367 : }
368 :
369 : #[derive(Clone)]
370 : pub struct Config {
371 : // All pageservers managed by one instance of this service must have
372 : // the same public key. This JWT token will be used to authenticate
373 : // this service to the pageservers it manages.
374 : pub pageserver_jwt_token: Option<String>,
375 :
376 : // All safekeepers managed by one instance of this service must have
377 : // the same public key. This JWT token will be used to authenticate
378 : // this service to the safekeepers it manages.
379 : pub safekeeper_jwt_token: Option<String>,
380 :
381 : // This JWT token will be used to authenticate this service to the control plane.
382 : pub control_plane_jwt_token: Option<String>,
383 :
384 : // This JWT token will be used to authenticate with other storage controller instances
385 : pub peer_jwt_token: Option<String>,
386 :
387 : /// Prefix for storage API endpoints of the control plane. We use this prefix to compute
388 : /// URLs that we use to send pageserver and safekeeper attachment locations.
389 : /// If this is None, the compute hook will assume it is running in a test environment
390 : /// and try to invoke neon_local instead.
391 : pub control_plane_url: Option<String>,
392 :
393 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
394 : /// considered active. Once the grace period elapses, the next heartbeat failure will
395 : /// mark the pagseserver offline.
396 : pub max_offline_interval: Duration,
397 :
398 : /// Extended grace period within which pageserver may not respond to heartbeats.
399 : /// This extended grace period kicks in after the node has been drained for restart
400 : /// and/or upon handling the re-attach request from a node.
401 : pub max_warming_up_interval: Duration,
402 :
403 : /// How many normal-priority Reconcilers may be spawned concurrently
404 : pub reconciler_concurrency: usize,
405 :
406 : /// How many high-priority Reconcilers may be spawned concurrently
407 : pub priority_reconciler_concurrency: usize,
408 :
409 : /// How many safekeeper reconciles may happen concurrently (per safekeeper)
410 : pub safekeeper_reconciler_concurrency: usize,
411 :
412 : /// How many API requests per second to allow per tenant, across all
413 : /// tenant-scoped API endpoints. Further API requests queue until ready.
414 : pub tenant_rate_limit: NonZeroU32,
415 :
416 : /// If a tenant shard's largest timeline (max_logical_size) exceeds this value, all tenant
417 : /// shards will be split in 2 until they fall below split_threshold (up to max_split_shards).
418 : ///
419 : /// This will greedily split into as many shards as necessary to fall below split_threshold, as
420 : /// powers of 2: if a tenant shard is 7 times larger than split_threshold, it will split into 8
421 : /// immediately, rather than first 2 then 4 then 8.
422 : ///
423 : /// None or 0 disables auto-splitting.
424 : ///
425 : /// TODO: consider using total logical size of all timelines instead.
426 : pub split_threshold: Option<u64>,
427 :
428 : /// The maximum number of shards a tenant can be split into during autosplits. Does not affect
429 : /// manual split requests. 0 or 1 disables autosplits, as we already have 1 shard.
430 : pub max_split_shards: u8,
431 :
432 : /// The size at which an unsharded tenant should initially split. Ingestion is significantly
433 : /// faster with multiple shards, so eagerly splitting below split_threshold will typically speed
434 : /// up initial ingestion of large tenants.
435 : ///
436 : /// This should be below split_threshold, but it is not required. If both split_threshold and
437 : /// initial_split_threshold qualify, the largest number of target shards will be used.
438 : ///
439 : /// Does not apply to already sharded tenants: changing initial_split_threshold or
440 : /// initial_split_shards is not retroactive for already-sharded tenants.
441 : ///
442 : /// None or 0 disables initial splits.
443 : pub initial_split_threshold: Option<u64>,
444 :
445 : /// The number of shards to split into when reaching initial_split_threshold. Will
446 : /// be clamped to max_split_shards.
447 : ///
448 : /// 0 or 1 disables initial splits. Has no effect if initial_split_threshold is disabled.
449 : pub initial_split_shards: u8,
450 :
451 : // TODO: make this cfg(feature = "testing")
452 : pub neon_local_repo_dir: Option<PathBuf>,
453 :
454 : // Maximum acceptable download lag for the secondary location
455 : // while draining a node. If the secondary location is lagging
456 : // by more than the configured amount, then the secondary is not
457 : // upgraded to primary.
458 : pub max_secondary_lag_bytes: Option<u64>,
459 :
460 : pub heartbeat_interval: Duration,
461 :
462 : pub address_for_peers: Option<Uri>,
463 :
464 : pub start_as_candidate: bool,
465 :
466 : pub long_reconcile_threshold: Duration,
467 :
468 : pub use_https_pageserver_api: bool,
469 :
470 : pub use_https_safekeeper_api: bool,
471 :
472 : pub ssl_ca_certs: Vec<Certificate>,
473 :
474 : pub timelines_onto_safekeepers: bool,
475 :
476 : pub use_local_compute_notifications: bool,
477 :
478 : /// Number of safekeepers to choose for a timeline when creating it.
479 : /// Safekeepers will be choosen from different availability zones.
480 : pub timeline_safekeeper_count: usize,
481 :
482 : /// PostHog integration config
483 : pub posthog_config: Option<PostHogConfig>,
484 :
485 : /// When set, actively checks and initiates heatmap downloads/uploads.
486 : pub kick_secondary_downloads: bool,
487 :
488 : /// Timeout used for HTTP client of split requests. [`Duration::MAX`] if None.
489 : pub shard_split_request_timeout: Duration,
490 :
491 : // Feature flag: Whether the storage controller should act to rectify pageserver-reported local disk loss.
492 : pub handle_ps_local_disk_loss: bool,
493 : }
494 :
495 : impl From<DatabaseError> for ApiError {
496 0 : fn from(err: DatabaseError) -> ApiError {
497 0 : match err {
498 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
499 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
500 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
501 0 : ApiError::ShuttingDown
502 : }
503 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
504 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
505 : }
506 0 : DatabaseError::Cas(reason) => ApiError::Conflict(reason),
507 : }
508 0 : }
509 : }
510 :
511 : enum InitialShardScheduleOutcome {
512 : Scheduled(TenantCreateResponseShard),
513 : NotScheduled,
514 : ShardScheduleError(ScheduleError),
515 : }
516 :
517 : pub struct Service {
518 : inner: Arc<std::sync::RwLock<ServiceState>>,
519 : config: Config,
520 : persistence: Arc<Persistence>,
521 : compute_hook: Arc<ComputeHook>,
522 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
523 :
524 : heartbeater_ps: Heartbeater<Node, PageserverState>,
525 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
526 :
527 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
528 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
529 :
530 : // Locking on a tenant granularity (covers all shards in the tenant):
531 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
532 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
533 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
534 :
535 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
536 : // that transition it to/from Active.
537 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
538 :
539 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
540 : // and reconciliation on startup.
541 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
542 :
543 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
544 : // a human user might be waiting for.
545 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
546 :
547 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
548 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
549 : ///
550 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
551 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
552 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
553 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
554 :
555 : // Process shutdown will fire this token
556 : cancel: CancellationToken,
557 :
558 : // Child token of [`Service::cancel`] used by reconcilers
559 : reconcilers_cancel: CancellationToken,
560 :
561 : // Background tasks will hold this gate
562 : gate: Gate,
563 :
564 : // Reconcilers background tasks will hold this gate
565 : reconcilers_gate: Gate,
566 :
567 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
568 : /// passes, it isn't safe to do any actions that mutate tenants.
569 : pub(crate) startup_complete: Barrier,
570 :
571 : /// HTTP client with proper CA certs.
572 : http_client: reqwest::Client,
573 :
574 : /// Handle for the step down background task if one was ever requested
575 : step_down_barrier: OnceLock<tokio::sync::watch::Receiver<Option<GlobalObservedState>>>,
576 : }
577 :
578 : impl From<ReconcileWaitError> for ApiError {
579 0 : fn from(value: ReconcileWaitError) -> Self {
580 0 : match value {
581 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
582 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
583 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
584 : }
585 0 : }
586 : }
587 :
588 : impl From<OperationError> for ApiError {
589 0 : fn from(value: OperationError) -> Self {
590 0 : match value {
591 0 : OperationError::NodeStateChanged(err)
592 0 : | OperationError::FinalizeError(err)
593 0 : | OperationError::ImpossibleConstraint(err) => {
594 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
595 : }
596 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
597 : }
598 0 : }
599 : }
600 :
601 : #[allow(clippy::large_enum_variant)]
602 : enum TenantCreateOrUpdate {
603 : Create(TenantCreateRequest),
604 : Update(Vec<ShardUpdate>),
605 : }
606 :
607 : struct ShardSplitParams {
608 : old_shard_count: ShardCount,
609 : new_shard_count: ShardCount,
610 : new_stripe_size: Option<ShardStripeSize>,
611 : targets: Vec<ShardSplitTarget>,
612 : policy: PlacementPolicy,
613 : config: TenantConfig,
614 : shard_ident: ShardIdentity,
615 : preferred_az_id: Option<AvailabilityZone>,
616 : }
617 :
618 : // When preparing for a shard split, we may either choose to proceed with the split,
619 : // or find that the work is already done and return NoOp.
620 : enum ShardSplitAction {
621 : Split(Box<ShardSplitParams>),
622 : NoOp(TenantShardSplitResponse),
623 : }
624 :
625 : // A parent shard which will be split
626 : struct ShardSplitTarget {
627 : parent_id: TenantShardId,
628 : node: Node,
629 : child_ids: Vec<TenantShardId>,
630 : }
631 :
632 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
633 : /// might not be available. We therefore use a queue of abort operations processed in the background.
634 : struct TenantShardSplitAbort {
635 : tenant_id: TenantId,
636 : /// The target values from the request that failed
637 : new_shard_count: ShardCount,
638 : new_stripe_size: Option<ShardStripeSize>,
639 : /// Until this abort op is complete, no other operations may be done on the tenant
640 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
641 : /// The reconciler gate for the duration of the split operation, and any included abort.
642 : _gate: GateGuard,
643 : }
644 :
645 : #[derive(thiserror::Error, Debug)]
646 : enum TenantShardSplitAbortError {
647 : #[error(transparent)]
648 : Database(#[from] DatabaseError),
649 : #[error(transparent)]
650 : Remote(#[from] mgmt_api::Error),
651 : #[error("Unavailable")]
652 : Unavailable,
653 : }
654 :
655 : /// Inputs for computing a target shard count for a tenant.
656 : struct ShardSplitInputs {
657 : /// Current shard count.
658 : shard_count: ShardCount,
659 : /// Total size of largest timeline summed across all shards.
660 : max_logical_size: u64,
661 : /// Size-based split threshold. Zero if size-based splits are disabled.
662 : split_threshold: u64,
663 : /// Upper bound on target shards. 0 or 1 disables splits.
664 : max_split_shards: u8,
665 : /// Initial split threshold. Zero if initial splits are disabled.
666 : initial_split_threshold: u64,
667 : /// Number of shards for initial splits. 0 or 1 disables initial splits.
668 : initial_split_shards: u8,
669 : }
670 :
671 : struct ShardUpdate {
672 : tenant_shard_id: TenantShardId,
673 : placement_policy: PlacementPolicy,
674 : tenant_config: TenantConfig,
675 :
676 : /// If this is None, generation is not updated.
677 : generation: Option<Generation>,
678 :
679 : /// If this is None, scheduling policy is not updated.
680 : scheduling_policy: Option<ShardSchedulingPolicy>,
681 : }
682 :
683 : enum StopReconciliationsReason {
684 : ShuttingDown,
685 : SteppingDown,
686 : }
687 :
688 : impl std::fmt::Display for StopReconciliationsReason {
689 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
690 0 : let s = match self {
691 0 : Self::ShuttingDown => "Shutting down",
692 0 : Self::SteppingDown => "Stepping down",
693 : };
694 0 : write!(writer, "{s}")
695 0 : }
696 : }
697 :
698 : pub(crate) enum ReconcileResultRequest {
699 : ReconcileResult(ReconcileResult),
700 : Stop,
701 : }
702 :
703 : #[derive(Clone)]
704 : pub(crate) struct MutationLocation {
705 : pub(crate) node: Node,
706 : pub(crate) generation: Generation,
707 : }
708 :
709 : #[derive(Clone)]
710 : pub(crate) struct ShardMutationLocations {
711 : pub(crate) latest: MutationLocation,
712 : pub(crate) other: Vec<MutationLocation>,
713 : }
714 :
715 : #[derive(Default, Clone)]
716 : pub(crate) struct TenantMutationLocations(pub BTreeMap<TenantShardId, ShardMutationLocations>);
717 :
718 : struct ReconcileAllResult {
719 : spawned_reconciles: usize,
720 : stuck_reconciles: usize,
721 : has_delayed_reconciles: bool,
722 : }
723 :
724 : impl ReconcileAllResult {
725 0 : fn new(
726 0 : spawned_reconciles: usize,
727 0 : stuck_reconciles: usize,
728 0 : has_delayed_reconciles: bool,
729 0 : ) -> Self {
730 0 : assert!(
731 0 : spawned_reconciles >= stuck_reconciles,
732 0 : "It is impossible to have less spawned reconciles than stuck reconciles"
733 : );
734 0 : Self {
735 0 : spawned_reconciles,
736 0 : stuck_reconciles,
737 0 : has_delayed_reconciles,
738 0 : }
739 0 : }
740 :
741 : /// We can run optimizations only if we don't have any delayed reconciles and
742 : /// all spawned reconciles are also stuck reconciles.
743 0 : fn can_run_optimizations(&self) -> bool {
744 0 : !self.has_delayed_reconciles && self.spawned_reconciles == self.stuck_reconciles
745 0 : }
746 : }
747 :
748 : enum TenantIdOrShardId {
749 : TenantId(TenantId),
750 : TenantShardId(TenantShardId),
751 : }
752 :
753 : impl TenantIdOrShardId {
754 0 : fn tenant_id(&self) -> TenantId {
755 0 : match self {
756 0 : TenantIdOrShardId::TenantId(tenant_id) => *tenant_id,
757 0 : TenantIdOrShardId::TenantShardId(tenant_shard_id) => tenant_shard_id.tenant_id,
758 : }
759 0 : }
760 :
761 0 : fn matches(&self, tenant_shard_id: &TenantShardId) -> bool {
762 0 : match self {
763 0 : TenantIdOrShardId::TenantId(tenant_id) => tenant_shard_id.tenant_id == *tenant_id,
764 0 : TenantIdOrShardId::TenantShardId(this_tenant_shard_id) => {
765 0 : this_tenant_shard_id == tenant_shard_id
766 : }
767 : }
768 0 : }
769 : }
770 :
771 : impl Service {
772 0 : pub fn get_config(&self) -> &Config {
773 0 : &self.config
774 0 : }
775 :
776 0 : pub fn get_http_client(&self) -> &reqwest::Client {
777 0 : &self.http_client
778 0 : }
779 :
780 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
781 : /// view of the world, and determine which pageservers are responsive.
782 : #[instrument(skip_all)]
783 : async fn startup_reconcile(
784 : self: &Arc<Service>,
785 : current_leader: Option<ControllerPersistence>,
786 : leader_step_down_state: Option<GlobalObservedState>,
787 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
788 : Result<(), (TenantShardId, NotifyError)>,
789 : >,
790 : ) {
791 : // Startup reconciliation does I/O to other services: whether they
792 : // are responsive or not, we should aim to finish within our deadline, because:
793 : // - If we don't, a k8s readiness hook watching /ready will kill us.
794 : // - While we're waiting for startup reconciliation, we are not fully
795 : // available for end user operations like creating/deleting tenants and timelines.
796 : //
797 : // We set multiple deadlines to break up the time available between the phases of work: this is
798 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
799 : let start_at = Instant::now();
800 : let node_scan_deadline = start_at
801 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
802 : .expect("Reconcile timeout is a modest constant");
803 :
804 : let observed = if let Some(state) = leader_step_down_state {
805 : tracing::info!(
806 : "Using observed state received from leader at {}",
807 : current_leader.as_ref().unwrap().address
808 : );
809 :
810 : state
811 : } else {
812 : self.build_global_observed_state(node_scan_deadline).await
813 : };
814 :
815 : // Accumulate a list of any tenant locations that ought to be detached
816 : let mut cleanup = Vec::new();
817 :
818 : // Send initial heartbeat requests to all nodes loaded from the database
819 : let all_nodes = {
820 : let locked = self.inner.read().unwrap();
821 : locked.nodes.clone()
822 : };
823 : let (mut nodes_online, mut sks_online) =
824 : self.initial_heartbeat_round(all_nodes.keys()).await;
825 :
826 : // List of tenants for which we will attempt to notify compute of their location at startup
827 : let mut compute_notifications = Vec::new();
828 :
829 : // Populate intent and observed states for all tenants, based on reported state on pageservers
830 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
831 : let shard_count = {
832 : let mut locked = self.inner.write().unwrap();
833 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
834 :
835 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
836 : let mut new_nodes = (**nodes).clone();
837 : for (node_id, node) in new_nodes.iter_mut() {
838 : if let Some(utilization) = nodes_online.remove(node_id) {
839 : node.set_availability(NodeAvailability::Active(utilization));
840 : scheduler.node_upsert(node);
841 : }
842 : }
843 : *nodes = Arc::new(new_nodes);
844 :
845 : let mut new_sks = (**safekeepers).clone();
846 : for (node_id, node) in new_sks.iter_mut() {
847 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
848 : node.set_availability(SafekeeperState::Available {
849 : utilization,
850 : last_seen_at,
851 : });
852 : }
853 : }
854 : *safekeepers = Arc::new(new_sks);
855 :
856 : for (tenant_shard_id, observed_state) in observed.0 {
857 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
858 : for node_id in observed_state.locations.keys() {
859 : cleanup.push((tenant_shard_id, *node_id));
860 : }
861 :
862 : continue;
863 : };
864 :
865 : tenant_shard.observed = observed_state;
866 : }
867 :
868 : // Populate each tenant's intent state
869 : let mut schedule_context = ScheduleContext::default();
870 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
871 : if tenant_shard_id.shard_number == ShardNumber(0) {
872 : // Reset scheduling context each time we advance to the next Tenant
873 : schedule_context = ScheduleContext::default();
874 : }
875 :
876 : tenant_shard.intent_from_observed(scheduler);
877 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
878 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
879 : // not enough pageservers are available. The tenant may well still be available
880 : // to clients.
881 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
882 : } else {
883 : // If we're both intending and observed to be attached at a particular node, we will
884 : // emit a compute notification for this. In the case where our observed state does not
885 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
886 : if let Some(attached_at) = tenant_shard.stably_attached() {
887 : compute_notifications.push(compute_hook::ShardUpdate {
888 : tenant_shard_id: *tenant_shard_id,
889 : node_id: attached_at,
890 : stripe_size: tenant_shard.shard.stripe_size,
891 : preferred_az: tenant_shard
892 : .preferred_az()
893 0 : .map(|az| Cow::Owned(az.clone())),
894 : });
895 : }
896 : }
897 : }
898 :
899 : tenants.len()
900 : };
901 :
902 : // Before making any obeservable changes to the cluster, persist self
903 : // as leader in database and memory.
904 : let leadership = Leadership::new(
905 : self.persistence.clone(),
906 : self.config.clone(),
907 : self.cancel.child_token(),
908 : );
909 :
910 : if let Err(e) = leadership.become_leader(current_leader).await {
911 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
912 : std::process::exit(1);
913 : }
914 :
915 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
916 : let sk_schedule_requests =
917 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
918 : Ok(v) => v,
919 : Err(e) => {
920 : tracing::warn!(
921 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
922 : );
923 : // std::process::exit(1);
924 : Vec::new()
925 : }
926 : };
927 :
928 : {
929 : let mut locked = self.inner.write().unwrap();
930 : locked.become_leader();
931 :
932 : for (sk_id, _sk) in locked.safekeepers.clone().iter() {
933 : locked.safekeeper_reconcilers.start_reconciler(*sk_id, self);
934 : }
935 :
936 : locked
937 : .safekeeper_reconcilers
938 : .schedule_request_vec(sk_schedule_requests);
939 : }
940 :
941 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
942 : // generation_pageserver in the database.
943 :
944 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
945 : // will emit compute hook notifications when they reconcile.
946 : //
947 : // Ordering: our calls to notify_attach_background synchronously establish a relative order for these notifications vs. any later
948 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
949 : // calls will be correctly ordered wrt these.
950 : //
951 : // Concurrency: we call notify_attach_background for all tenants, which will create O(N) tokio tasks, but almost all of them
952 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
953 : // unit and start doing I/O.
954 : tracing::info!(
955 : "Sending {} compute notifications",
956 : compute_notifications.len()
957 : );
958 : self.compute_hook.notify_attach_background(
959 : compute_notifications,
960 : bg_compute_notify_result_tx.clone(),
961 : &self.cancel,
962 : );
963 :
964 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
965 : // which require it: under normal circumstances this should only include tenants that were in some
966 : // transient state before we restarted, or any tenants whose compute hooks failed above.
967 : tracing::info!("Checking for shards in need of reconciliation...");
968 : let reconcile_all_result = self.reconcile_all();
969 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
970 : // normal operations may proceed.
971 :
972 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
973 : // background because it does not need to complete in order to proceed with other work.
974 : if !cleanup.is_empty() {
975 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
976 : tokio::task::spawn({
977 : let cleanup_self = self.clone();
978 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
979 : });
980 : }
981 :
982 : // Reconcile the timeline imports:
983 : // 1. Mark each tenant shard of tenants with an importing timeline as importing.
984 : // 2. Finalize the completed imports in the background. This handles the case where
985 : // the previous storage controller instance shut down whilst finalizing imports.
986 : let imports = self.persistence.list_timeline_imports().await;
987 : match imports {
988 : Ok(mut imports) => {
989 : {
990 : let mut locked = self.inner.write().unwrap();
991 : for import in &imports {
992 : locked
993 : .tenants
994 : .range_mut(TenantShardId::tenant_range(import.tenant_id))
995 0 : .for_each(|(_id, shard)| {
996 0 : shard.importing = TimelineImportState::Importing
997 0 : });
998 : }
999 : }
1000 :
1001 0 : imports.retain(|import| import.is_complete());
1002 : tokio::task::spawn({
1003 : let finalize_imports_self = self.clone();
1004 0 : async move {
1005 0 : finalize_imports_self
1006 0 : .finalize_timeline_imports(imports)
1007 0 : .await
1008 0 : }
1009 : });
1010 : }
1011 : Err(err) => {
1012 : tracing::error!("Could not retrieve completed imports from database: {err}");
1013 : }
1014 : }
1015 :
1016 : let spawned_reconciles = reconcile_all_result.spawned_reconciles;
1017 : tracing::info!(
1018 : "Startup complete, spawned {spawned_reconciles} reconciliation tasks ({shard_count} shards total)"
1019 : );
1020 : }
1021 :
1022 0 : async fn initial_heartbeat_round<'a>(
1023 0 : &self,
1024 0 : node_ids: impl Iterator<Item = &'a NodeId>,
1025 0 : ) -> (
1026 0 : HashMap<NodeId, PageserverUtilization>,
1027 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
1028 0 : ) {
1029 0 : assert!(!self.startup_complete.is_ready());
1030 :
1031 0 : let all_nodes = {
1032 0 : let locked = self.inner.read().unwrap();
1033 0 : locked.nodes.clone()
1034 : };
1035 :
1036 0 : let mut nodes_to_heartbeat = HashMap::new();
1037 0 : for node_id in node_ids {
1038 0 : match all_nodes.get(node_id) {
1039 0 : Some(node) => {
1040 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
1041 0 : }
1042 : None => {
1043 0 : tracing::warn!("Node {node_id} was removed during start-up");
1044 : }
1045 : }
1046 : }
1047 :
1048 0 : let all_sks = {
1049 0 : let locked = self.inner.read().unwrap();
1050 0 : locked.safekeepers.clone()
1051 : };
1052 :
1053 0 : tracing::info!("Sending initial heartbeats...");
1054 0 : let (res_ps, res_sk) = tokio::join!(
1055 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
1056 0 : self.heartbeater_sk.heartbeat(all_sks)
1057 : );
1058 :
1059 0 : let mut online_nodes = HashMap::new();
1060 0 : if let Ok(deltas) = res_ps {
1061 0 : for (node_id, status) in deltas.0 {
1062 0 : match status {
1063 0 : PageserverState::Available { utilization, .. } => {
1064 0 : online_nodes.insert(node_id, utilization);
1065 0 : }
1066 0 : PageserverState::Offline => {}
1067 : PageserverState::WarmingUp { .. } => {
1068 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
1069 : }
1070 : }
1071 : }
1072 0 : }
1073 :
1074 0 : let mut online_sks = HashMap::new();
1075 0 : if let Ok(deltas) = res_sk {
1076 0 : for (node_id, status) in deltas.0 {
1077 0 : match status {
1078 : SafekeeperState::Available {
1079 0 : utilization,
1080 0 : last_seen_at,
1081 0 : } => {
1082 0 : online_sks.insert(node_id, (utilization, last_seen_at));
1083 0 : }
1084 0 : SafekeeperState::Offline => {}
1085 : }
1086 : }
1087 0 : }
1088 :
1089 0 : (online_nodes, online_sks)
1090 0 : }
1091 :
1092 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
1093 : ///
1094 : /// The result includes only nodes which responded within the deadline
1095 0 : async fn scan_node_locations(
1096 0 : &self,
1097 0 : deadline: Instant,
1098 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
1099 0 : let nodes = {
1100 0 : let locked = self.inner.read().unwrap();
1101 0 : locked.nodes.clone()
1102 : };
1103 :
1104 0 : let mut node_results = HashMap::new();
1105 :
1106 0 : let mut node_list_futs = FuturesUnordered::new();
1107 :
1108 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
1109 0 : for node in nodes.values() {
1110 0 : node_list_futs.push({
1111 0 : async move {
1112 0 : tracing::info!("Scanning shards on node {node}...");
1113 0 : let timeout = Duration::from_secs(5);
1114 0 : let response = node
1115 0 : .with_client_retries(
1116 0 : |client| async move { client.list_location_config().await },
1117 0 : &self.http_client,
1118 0 : &self.config.pageserver_jwt_token,
1119 : 1,
1120 : 5,
1121 0 : timeout,
1122 0 : &self.cancel,
1123 : )
1124 0 : .await;
1125 0 : (node.get_id(), response)
1126 0 : }
1127 : });
1128 : }
1129 :
1130 : loop {
1131 0 : let (node_id, result) = tokio::select! {
1132 0 : next = node_list_futs.next() => {
1133 0 : match next {
1134 0 : Some(result) => result,
1135 : None =>{
1136 : // We got results for all our nodes
1137 0 : break;
1138 : }
1139 :
1140 : }
1141 : },
1142 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
1143 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
1144 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
1145 0 : break;
1146 : }
1147 : };
1148 :
1149 0 : let Some(list_response) = result else {
1150 0 : tracing::info!("Shutdown during startup_reconcile");
1151 0 : break;
1152 : };
1153 :
1154 0 : match list_response {
1155 0 : Err(e) => {
1156 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
1157 : }
1158 0 : Ok(listing) => {
1159 0 : node_results.insert(node_id, listing);
1160 0 : }
1161 : }
1162 : }
1163 :
1164 0 : node_results
1165 0 : }
1166 :
1167 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
1168 0 : let node_listings = self.scan_node_locations(deadline).await;
1169 0 : let mut observed = GlobalObservedState::default();
1170 :
1171 0 : for (node_id, location_confs) in node_listings {
1172 0 : tracing::info!(
1173 0 : "Received {} shard statuses from pageserver {}",
1174 0 : location_confs.tenant_shards.len(),
1175 : node_id
1176 : );
1177 :
1178 0 : for (tid, location_conf) in location_confs.tenant_shards {
1179 0 : let entry = observed.0.entry(tid).or_default();
1180 0 : entry.locations.insert(
1181 0 : node_id,
1182 0 : ObservedStateLocation {
1183 0 : conf: location_conf,
1184 0 : },
1185 0 : );
1186 0 : }
1187 : }
1188 :
1189 0 : observed
1190 0 : }
1191 :
1192 : /// Used during [`Self::startup_reconcile`] and shard splits: detach a list of unknown-to-us
1193 : /// tenants from pageservers.
1194 : ///
1195 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1196 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1197 : /// other task trying to attach it.
1198 : #[instrument(skip_all)]
1199 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1200 : let nodes = self.inner.read().unwrap().nodes.clone();
1201 :
1202 : for (tenant_shard_id, node_id) in cleanup {
1203 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1204 : let Some(node) = nodes.get(&node_id) else {
1205 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1206 : // a location to clean up on a node that has since been removed.
1207 : tracing::info!(
1208 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1209 : );
1210 : continue;
1211 : };
1212 :
1213 : if self.cancel.is_cancelled() {
1214 : break;
1215 : }
1216 :
1217 : let client = PageserverClient::new(
1218 : node.get_id(),
1219 : self.http_client.clone(),
1220 : node.base_url(),
1221 : self.config.pageserver_jwt_token.as_deref(),
1222 : );
1223 : match client
1224 : .location_config(
1225 : tenant_shard_id,
1226 : LocationConfig {
1227 : mode: LocationConfigMode::Detached,
1228 : generation: None,
1229 : secondary_conf: None,
1230 : shard_number: tenant_shard_id.shard_number.0,
1231 : shard_count: tenant_shard_id.shard_count.literal(),
1232 : shard_stripe_size: 0,
1233 : tenant_conf: models::TenantConfig::default(),
1234 : },
1235 : None,
1236 : false,
1237 : )
1238 : .await
1239 : {
1240 : Ok(()) => {
1241 : tracing::info!(
1242 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1243 : );
1244 : }
1245 : Err(e) => {
1246 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1247 : // break anything.
1248 : tracing::error!(
1249 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1250 : );
1251 : }
1252 : }
1253 : }
1254 : }
1255 :
1256 : /// Long running background task that periodically wakes up and looks for shards that need
1257 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1258 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1259 : /// for those retries.
1260 : #[instrument(skip_all)]
1261 : async fn background_reconcile(self: &Arc<Self>) {
1262 : self.startup_complete.clone().wait().await;
1263 :
1264 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1265 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1266 : while !self.reconcilers_cancel.is_cancelled() {
1267 : tokio::select! {
1268 : _ = interval.tick() => {
1269 : let reconcile_all_result = self.reconcile_all();
1270 : if reconcile_all_result.can_run_optimizations() {
1271 : // Run optimizer only when we didn't find any other work to do
1272 : self.optimize_all().await;
1273 : }
1274 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1275 : // must be responsive when new projects begin ingesting and reach the threshold.
1276 : self.autosplit_tenants().await;
1277 : }
1278 : _ = self.reconcilers_cancel.cancelled() => return
1279 : }
1280 : }
1281 : }
1282 : /// Heartbeat all storage nodes once in a while.
1283 : #[instrument(skip_all)]
1284 : async fn spawn_heartbeat_driver(self: &Arc<Self>) {
1285 : self.startup_complete.clone().wait().await;
1286 :
1287 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1288 : while !self.cancel.is_cancelled() {
1289 : tokio::select! {
1290 : _ = interval.tick() => { }
1291 : _ = self.cancel.cancelled() => return
1292 : };
1293 :
1294 : let nodes = {
1295 : let locked = self.inner.read().unwrap();
1296 : locked.nodes.clone()
1297 : };
1298 :
1299 : let safekeepers = {
1300 : let locked = self.inner.read().unwrap();
1301 : locked.safekeepers.clone()
1302 : };
1303 :
1304 : let (res_ps, res_sk) = tokio::join!(
1305 : self.heartbeater_ps.heartbeat(nodes),
1306 : self.heartbeater_sk.heartbeat(safekeepers)
1307 : );
1308 :
1309 : if let Ok(deltas) = res_ps {
1310 : let mut to_handle = Vec::default();
1311 :
1312 : for (node_id, state) in deltas.0 {
1313 : let new_availability = match state {
1314 : PageserverState::Available { utilization, .. } => {
1315 : NodeAvailability::Active(utilization)
1316 : }
1317 : PageserverState::WarmingUp { started_at } => {
1318 : NodeAvailability::WarmingUp(started_at)
1319 : }
1320 : PageserverState::Offline => {
1321 : // The node might have been placed in the WarmingUp state
1322 : // while the heartbeat round was on-going. Hence, filter out
1323 : // offline transitions for WarmingUp nodes that are still within
1324 : // their grace period.
1325 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1326 : .get_node(node_id)
1327 : .await
1328 : .as_ref()
1329 0 : .map(|n| n.get_availability())
1330 : {
1331 : let now = Instant::now();
1332 : if now - *started_at >= self.config.max_warming_up_interval {
1333 : NodeAvailability::Offline
1334 : } else {
1335 : NodeAvailability::WarmingUp(*started_at)
1336 : }
1337 : } else {
1338 : NodeAvailability::Offline
1339 : }
1340 : }
1341 : };
1342 :
1343 : let node_lock = trace_exclusive_lock(
1344 : &self.node_op_locks,
1345 : node_id,
1346 : NodeOperations::Configure,
1347 : )
1348 : .await;
1349 :
1350 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1351 :
1352 : // This is the code path for geniune availability transitions (i.e node
1353 : // goes unavailable and/or comes back online).
1354 : let res = self
1355 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1356 : .await;
1357 :
1358 : match res {
1359 : Ok(transition) => {
1360 : // Keep hold of the lock until the availability transitions
1361 : // have been handled in
1362 : // [`Service::handle_node_availability_transitions`] in order avoid
1363 : // racing with [`Service::external_node_configure`].
1364 : to_handle.push((node_id, node_lock, transition));
1365 : }
1366 : Err(ApiError::NotFound(_)) => {
1367 : // This should be rare, but legitimate since the heartbeats are done
1368 : // on a snapshot of the nodes.
1369 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1370 : }
1371 : Err(ApiError::ShuttingDown) => {
1372 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1373 : }
1374 : Err(err) => {
1375 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1376 : // becomes unavailable again, we may get an error here.
1377 : tracing::error!(
1378 : "Failed to update node state {} after heartbeat round: {}",
1379 : node_id,
1380 : err
1381 : );
1382 : }
1383 : }
1384 : }
1385 :
1386 : // We collected all the transitions above and now we handle them.
1387 : let res = self.handle_node_availability_transitions(to_handle).await;
1388 : if let Err(errs) = res {
1389 : for (node_id, err) in errs {
1390 : match err {
1391 : ApiError::NotFound(_) => {
1392 : // This should be rare, but legitimate since the heartbeats are done
1393 : // on a snapshot of the nodes.
1394 : tracing::info!(
1395 : "Node {} was not found after heartbeat round",
1396 : node_id
1397 : );
1398 : }
1399 : err => {
1400 : tracing::error!(
1401 : "Failed to handle availability transition for {} after heartbeat round: {}",
1402 : node_id,
1403 : err
1404 : );
1405 : }
1406 : }
1407 : }
1408 : }
1409 : }
1410 : if let Ok(deltas) = res_sk {
1411 : let mut to_activate = Vec::new();
1412 : {
1413 : let mut locked = self.inner.write().unwrap();
1414 : let mut safekeepers = (*locked.safekeepers).clone();
1415 :
1416 : for (id, state) in deltas.0 {
1417 : let Some(sk) = safekeepers.get_mut(&id) else {
1418 : tracing::info!(
1419 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1420 : );
1421 : continue;
1422 : };
1423 : if sk.scheduling_policy() == SkSchedulingPolicy::Activating
1424 : && let SafekeeperState::Available { .. } = state
1425 : {
1426 : to_activate.push(id);
1427 : }
1428 : sk.set_availability(state);
1429 : }
1430 : locked.safekeepers = Arc::new(safekeepers);
1431 : }
1432 : for sk_id in to_activate {
1433 : // TODO this can race with set_scheduling_policy (can create disjoint DB <-> in-memory state)
1434 : tracing::info!("Activating safekeeper {sk_id}");
1435 : match self.persistence.activate_safekeeper(sk_id.0 as i64).await {
1436 : Ok(Some(())) => {}
1437 : Ok(None) => {
1438 : tracing::info!(
1439 : "safekeeper {sk_id} has been removed from db or has different scheduling policy than active or activating"
1440 : );
1441 : }
1442 : Err(e) => {
1443 : tracing::warn!("couldn't apply activation of {sk_id} to db: {e}");
1444 : continue;
1445 : }
1446 : }
1447 : if let Err(e) = self
1448 : .set_safekeeper_scheduling_policy_in_mem(sk_id, SkSchedulingPolicy::Active)
1449 : .await
1450 : {
1451 : tracing::info!("couldn't activate safekeeper {sk_id} in memory: {e}");
1452 : continue;
1453 : }
1454 : tracing::info!("Activation of safekeeper {sk_id} done");
1455 : }
1456 : }
1457 : }
1458 : }
1459 :
1460 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1461 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1462 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1463 : /// will indicate that reconciliation is not needed.
1464 : #[instrument(skip_all, fields(
1465 : seq=%result.sequence,
1466 : tenant_id=%result.tenant_shard_id.tenant_id,
1467 : shard_id=%result.tenant_shard_id.shard_slug(),
1468 : ))]
1469 : fn process_result(&self, result: ReconcileResult) {
1470 : let mut locked = self.inner.write().unwrap();
1471 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1472 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1473 : // A reconciliation result might race with removing a tenant: drop results for
1474 : // tenants that aren't in our map.
1475 : return;
1476 : };
1477 :
1478 : // Usually generation should only be updated via this path, so the max() isn't
1479 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1480 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1481 :
1482 : // If the reconciler signals that it failed to notify compute, set this state on
1483 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1484 : tenant.pending_compute_notification = result.pending_compute_notification;
1485 :
1486 : // Let the TenantShard know it is idle.
1487 : tenant.reconcile_complete(result.sequence);
1488 :
1489 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1490 : // make to the tenant
1491 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1492 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1493 : // make to the tenant
1494 0 : let node = nodes.get(delta.node_id())?;
1495 :
1496 0 : if node.is_available() {
1497 0 : return Some(delta);
1498 0 : }
1499 :
1500 : // In case a node became unavailable concurrently with the reconcile, observed
1501 : // locations on it are now uncertain. By convention, set them to None in order
1502 : // for them to get refreshed when the node comes back online.
1503 0 : Some(ObservedStateDelta::Upsert(Box::new((
1504 0 : node.get_id(),
1505 0 : ObservedStateLocation { conf: None },
1506 0 : ))))
1507 0 : });
1508 :
1509 : match result.result {
1510 : Ok(()) => {
1511 : tenant.apply_observed_deltas(deltas);
1512 : tenant.waiter.advance(result.sequence);
1513 : }
1514 : Err(e) => {
1515 : match e {
1516 : ReconcileError::Cancel => {
1517 : tracing::info!("Reconciler was cancelled");
1518 : }
1519 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1520 : // This might be due to the reconciler getting cancelled, or it might
1521 : // be due to the `Node` being marked offline.
1522 : tracing::info!("Reconciler cancelled during pageserver API call");
1523 : }
1524 : _ => {
1525 : tracing::warn!("Reconcile error: {}", e);
1526 : }
1527 : }
1528 :
1529 : // Ordering: populate last_error before advancing error_seq,
1530 : // so that waiters will see the correct error after waiting.
1531 : tenant.set_last_error(result.sequence, e);
1532 :
1533 : // Skip deletions on reconcile failures
1534 : let upsert_deltas =
1535 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1536 : tenant.apply_observed_deltas(upsert_deltas);
1537 : }
1538 : }
1539 :
1540 : tenant.consecutive_reconciles_count = tenant.consecutive_reconciles_count.saturating_add(1);
1541 :
1542 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1543 : if tenant.policy == PlacementPolicy::Detached {
1544 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1545 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1546 : // duration of the request.
1547 : let guard = self.tenant_op_locks.try_exclusive(
1548 : tenant.tenant_shard_id.tenant_id,
1549 : TenantOperations::DropDetached,
1550 : );
1551 : if let Some(guard) = guard {
1552 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1553 : }
1554 : }
1555 :
1556 : // Maybe some other work can proceed now that this job finished.
1557 : //
1558 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1559 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1560 : if self.reconciler_concurrency.available_permits() > 0 {
1561 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1562 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1563 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1564 : shard.delayed_reconcile = false;
1565 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1566 : }
1567 :
1568 : if self.reconciler_concurrency.available_permits() == 0 {
1569 : break;
1570 : }
1571 : }
1572 : }
1573 : }
1574 :
1575 0 : async fn process_results(
1576 0 : &self,
1577 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1578 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1579 0 : Result<(), (TenantShardId, NotifyError)>,
1580 0 : >,
1581 0 : ) {
1582 : loop {
1583 : // Wait for the next result, or for cancellation
1584 0 : tokio::select! {
1585 0 : r = result_rx.recv() => {
1586 0 : match r {
1587 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1588 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1589 : }
1590 : }
1591 0 : _ = async{
1592 0 : match bg_compute_hook_result_rx.recv().await {
1593 0 : Some(result) => {
1594 0 : if let Err((tenant_shard_id, notify_error)) = result {
1595 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1596 0 : let mut locked = self.inner.write().unwrap();
1597 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1598 0 : shard.pending_compute_notification = true;
1599 0 : }
1600 :
1601 0 : }
1602 : },
1603 : None => {
1604 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1605 0 : self.cancel.cancelled().await;
1606 : }
1607 : }
1608 0 : } => {},
1609 0 : _ = self.cancel.cancelled() => {
1610 0 : break;
1611 : }
1612 : };
1613 : }
1614 0 : }
1615 :
1616 0 : async fn process_aborts(
1617 0 : &self,
1618 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1619 0 : ) {
1620 : loop {
1621 : // Wait for the next result, or for cancellation
1622 0 : let op = tokio::select! {
1623 0 : r = abort_rx.recv() => {
1624 0 : match r {
1625 0 : Some(op) => {op},
1626 0 : None => {break;}
1627 : }
1628 : }
1629 0 : _ = self.cancel.cancelled() => {
1630 0 : break;
1631 : }
1632 : };
1633 :
1634 : // Retry until shutdown: we must keep this request object alive until it is properly
1635 : // processed, as it holds a lock guard that prevents other operations trying to do things
1636 : // to the tenant while it is in a weird part-split state.
1637 0 : while !self.reconcilers_cancel.is_cancelled() {
1638 0 : match self.abort_tenant_shard_split(&op).await {
1639 0 : Ok(_) => break,
1640 0 : Err(e) => {
1641 0 : tracing::warn!(
1642 0 : "Failed to abort shard split on {}, will retry: {e}",
1643 : op.tenant_id
1644 : );
1645 :
1646 : // If a node is unavailable, we hope that it has been properly marked Offline
1647 : // when we retry, so that the abort op will succeed. If the abort op is failing
1648 : // for some other reason, we will keep retrying forever, or until a human notices
1649 : // and does something about it (either fixing a pageserver or restarting the controller).
1650 0 : tokio::time::timeout(
1651 0 : Duration::from_secs(5),
1652 0 : self.reconcilers_cancel.cancelled(),
1653 0 : )
1654 0 : .await
1655 0 : .ok();
1656 : }
1657 : }
1658 : }
1659 : }
1660 0 : }
1661 :
1662 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1663 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1664 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1665 :
1666 0 : let leadership_cancel = CancellationToken::new();
1667 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1668 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1669 :
1670 : // Apply the migrations **after** the current leader has stepped down
1671 : // (or we've given up waiting for it), but **before** reading from the
1672 : // database. The only exception is reading the current leader before
1673 : // migrating.
1674 0 : persistence.migration_run().await?;
1675 :
1676 0 : tracing::info!("Loading nodes from database...");
1677 0 : let nodes = persistence
1678 0 : .list_nodes()
1679 0 : .await?
1680 0 : .into_iter()
1681 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1682 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1683 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1684 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1685 0 : metrics::METRICS_REGISTRY
1686 0 : .metrics_group
1687 0 : .storage_controller_pageserver_nodes
1688 0 : .set(nodes.len() as i64);
1689 0 : metrics::METRICS_REGISTRY
1690 0 : .metrics_group
1691 0 : .storage_controller_https_pageserver_nodes
1692 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
1693 :
1694 0 : tracing::info!("Loading safekeepers from database...");
1695 0 : let safekeepers = persistence
1696 0 : .list_safekeepers()
1697 0 : .await?
1698 0 : .into_iter()
1699 0 : .map(|skp| {
1700 0 : Safekeeper::from_persistence(
1701 0 : skp,
1702 0 : CancellationToken::new(),
1703 0 : config.use_https_safekeeper_api,
1704 : )
1705 0 : })
1706 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1707 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1708 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1709 0 : let count_policy = |policy| {
1710 0 : safekeepers
1711 0 : .iter()
1712 0 : .filter(|sk| sk.1.scheduling_policy() == policy)
1713 0 : .count()
1714 0 : };
1715 0 : let active_sk_count = count_policy(SkSchedulingPolicy::Active);
1716 0 : let activating_sk_count = count_policy(SkSchedulingPolicy::Activating);
1717 0 : let pause_sk_count = count_policy(SkSchedulingPolicy::Pause);
1718 0 : let decom_sk_count = count_policy(SkSchedulingPolicy::Decomissioned);
1719 0 : tracing::info!(
1720 0 : "Loaded {} safekeepers from database. Active {active_sk_count}, activating {activating_sk_count}, \
1721 0 : paused {pause_sk_count}, decomissioned {decom_sk_count}.",
1722 0 : safekeepers.len()
1723 : );
1724 0 : metrics::METRICS_REGISTRY
1725 0 : .metrics_group
1726 0 : .storage_controller_safekeeper_nodes
1727 0 : .set(safekeepers.len() as i64);
1728 0 : metrics::METRICS_REGISTRY
1729 0 : .metrics_group
1730 0 : .storage_controller_https_safekeeper_nodes
1731 0 : .set(safekeepers.values().filter(|s| s.has_https_port()).count() as i64);
1732 :
1733 0 : tracing::info!("Loading shards from database...");
1734 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1735 0 : tracing::info!(
1736 0 : "Loaded {} shards from database.",
1737 0 : tenant_shard_persistence.len()
1738 : );
1739 :
1740 : // If any shard splits were in progress, reset the database state to abort them
1741 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1742 0 : HashMap::new();
1743 0 : for tsp in &mut tenant_shard_persistence {
1744 0 : let shard = tsp.get_shard_identity()?;
1745 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1746 0 : let entry = tenant_shard_count_min_max
1747 0 : .entry(tenant_shard_id.tenant_id)
1748 0 : .or_insert_with(|| (shard.count, shard.count));
1749 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1750 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1751 : }
1752 :
1753 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1754 0 : if count_min != count_max {
1755 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1756 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1757 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1758 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1759 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1760 :
1761 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1762 : // identified this tenant has having mismatching min/max counts.
1763 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1764 :
1765 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1766 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1767 : // Set idle split state on those shards that we will retain.
1768 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1769 0 : if tsp_tenant_id == tenant_id
1770 0 : && tsp.get_shard_identity().unwrap().count == count_min
1771 0 : {
1772 0 : tsp.splitting = SplitState::Idle;
1773 0 : } else if tsp_tenant_id == tenant_id {
1774 : // Leave the splitting state on the child shards: this will be used next to
1775 : // drop them.
1776 0 : tracing::info!(
1777 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1778 : );
1779 0 : }
1780 0 : });
1781 :
1782 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1783 0 : tenant_shard_persistence.retain(|tsp| {
1784 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1785 0 : || tsp.splitting == SplitState::Idle
1786 0 : });
1787 0 : }
1788 : }
1789 :
1790 0 : let mut tenants = BTreeMap::new();
1791 :
1792 0 : let mut scheduler = Scheduler::new(nodes.values());
1793 :
1794 : #[cfg(feature = "testing")]
1795 : {
1796 : use pageserver_api::controller_api::AvailabilityZone;
1797 :
1798 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1799 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1800 : // after when pageservers start up and register.
1801 0 : let mut node_ids = HashSet::new();
1802 0 : for tsp in &tenant_shard_persistence {
1803 0 : if let Some(node_id) = tsp.generation_pageserver {
1804 0 : node_ids.insert(node_id);
1805 0 : }
1806 : }
1807 0 : for node_id in node_ids {
1808 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1809 0 : let node = Node::new(
1810 0 : NodeId(node_id as u64),
1811 0 : "".to_string(),
1812 : 123,
1813 0 : None,
1814 0 : "".to_string(),
1815 : 123,
1816 0 : None,
1817 0 : None,
1818 0 : AvailabilityZone("test_az".to_string()),
1819 : false,
1820 : )
1821 0 : .unwrap();
1822 :
1823 0 : scheduler.node_upsert(&node);
1824 : }
1825 : }
1826 0 : for tsp in tenant_shard_persistence {
1827 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1828 :
1829 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1830 : // it with what we can infer: the node for which a generation was most recently issued.
1831 0 : let mut intent = IntentState::new(
1832 0 : tsp.preferred_az_id
1833 0 : .as_ref()
1834 0 : .map(|az| AvailabilityZone(az.clone())),
1835 : );
1836 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1837 : {
1838 0 : if nodes.contains_key(&generation_pageserver) {
1839 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1840 0 : } else {
1841 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1842 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1843 : // on different pageservers.
1844 0 : tracing::warn!(
1845 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1846 : );
1847 : }
1848 0 : }
1849 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1850 :
1851 0 : tenants.insert(tenant_shard_id, new_tenant);
1852 : }
1853 :
1854 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1855 :
1856 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1857 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1858 0 : tokio::sync::mpsc::channel(512);
1859 :
1860 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1861 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1862 :
1863 0 : let cancel = CancellationToken::new();
1864 0 : let reconcilers_cancel = cancel.child_token();
1865 :
1866 0 : let mut http_client = reqwest::Client::builder();
1867 : // We intentionally disable the connection pool, so every request will create its own TCP connection.
1868 : // It's especially important for heartbeaters to notice more network problems.
1869 : //
1870 : // TODO: It makes sense to use this client only in heartbeaters and create a second one with
1871 : // connection pooling for everything else. But reqwest::Client may create a connection without
1872 : // ever using it (it uses hyper's Client under the hood):
1873 : // https://github.com/hyperium/hyper-util/blob/d51318df3461d40e5f5e5ca163cb3905ac960209/src/client/legacy/client.rs#L415
1874 : //
1875 : // Because of a bug in hyper0::Connection::graceful_shutdown such connections hang during
1876 : // graceful server shutdown: https://github.com/hyperium/hyper/issues/2730
1877 : //
1878 : // The bug has been fixed in hyper v1, so keep alive may be enabled only after we migrate to hyper1.
1879 0 : http_client = http_client.pool_max_idle_per_host(0);
1880 0 : for ssl_ca_cert in &config.ssl_ca_certs {
1881 0 : http_client = http_client.add_root_certificate(ssl_ca_cert.clone());
1882 0 : }
1883 0 : let http_client = http_client.build()?;
1884 :
1885 0 : let heartbeater_ps = Heartbeater::new(
1886 0 : http_client.clone(),
1887 0 : config.pageserver_jwt_token.clone(),
1888 0 : config.max_offline_interval,
1889 0 : config.max_warming_up_interval,
1890 0 : cancel.clone(),
1891 : );
1892 :
1893 0 : let heartbeater_sk = Heartbeater::new(
1894 0 : http_client.clone(),
1895 0 : config.safekeeper_jwt_token.clone(),
1896 0 : config.max_offline_interval,
1897 0 : config.max_warming_up_interval,
1898 0 : cancel.clone(),
1899 : );
1900 :
1901 0 : let initial_leadership_status = if config.start_as_candidate {
1902 0 : LeadershipStatus::Candidate
1903 : } else {
1904 0 : LeadershipStatus::Leader
1905 : };
1906 :
1907 0 : let this = Arc::new(Self {
1908 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1909 0 : nodes,
1910 0 : safekeepers,
1911 0 : tenants,
1912 0 : scheduler,
1913 0 : delayed_reconcile_rx,
1914 0 : initial_leadership_status,
1915 0 : reconcilers_cancel.clone(),
1916 : ))),
1917 0 : config: config.clone(),
1918 0 : persistence,
1919 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())?),
1920 0 : result_tx,
1921 0 : heartbeater_ps,
1922 0 : heartbeater_sk,
1923 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1924 0 : config.reconciler_concurrency,
1925 : )),
1926 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1927 0 : config.priority_reconciler_concurrency,
1928 : )),
1929 0 : delayed_reconcile_tx,
1930 0 : abort_tx,
1931 0 : startup_complete: startup_complete.clone(),
1932 0 : cancel,
1933 0 : reconcilers_cancel,
1934 0 : gate: Gate::default(),
1935 0 : reconcilers_gate: Gate::default(),
1936 0 : tenant_op_locks: Default::default(),
1937 0 : node_op_locks: Default::default(),
1938 0 : http_client,
1939 0 : step_down_barrier: Default::default(),
1940 : });
1941 :
1942 0 : let result_task_this = this.clone();
1943 0 : tokio::task::spawn(async move {
1944 : // Block shutdown until we're done (we must respect self.cancel)
1945 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1946 0 : result_task_this
1947 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1948 0 : .await
1949 0 : }
1950 0 : });
1951 :
1952 0 : tokio::task::spawn({
1953 0 : let this = this.clone();
1954 0 : async move {
1955 : // Block shutdown until we're done (we must respect self.cancel)
1956 0 : if let Ok(_gate) = this.gate.enter() {
1957 0 : this.process_aborts(abort_rx).await
1958 0 : }
1959 0 : }
1960 : });
1961 :
1962 0 : tokio::task::spawn({
1963 0 : let this = this.clone();
1964 0 : async move {
1965 0 : if let Ok(_gate) = this.gate.enter() {
1966 : loop {
1967 0 : tokio::select! {
1968 0 : _ = this.cancel.cancelled() => {
1969 0 : break;
1970 : },
1971 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1972 : };
1973 0 : this.tenant_op_locks.housekeeping();
1974 : }
1975 0 : }
1976 0 : }
1977 : });
1978 :
1979 0 : tokio::task::spawn({
1980 0 : let this = this.clone();
1981 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1982 : // is done.
1983 0 : let startup_completion = startup_completion.clone();
1984 0 : async move {
1985 : // Block shutdown until we're done (we must respect self.cancel)
1986 0 : let Ok(_gate) = this.gate.enter() else {
1987 0 : return;
1988 : };
1989 :
1990 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1991 0 : .await;
1992 :
1993 0 : drop(startup_completion);
1994 0 : }
1995 : });
1996 :
1997 0 : tokio::task::spawn({
1998 0 : let this = this.clone();
1999 0 : let startup_complete = startup_complete.clone();
2000 0 : async move {
2001 0 : startup_complete.wait().await;
2002 0 : this.background_reconcile().await;
2003 0 : }
2004 : });
2005 :
2006 0 : tokio::task::spawn({
2007 0 : let this = this.clone();
2008 0 : let startup_complete = startup_complete.clone();
2009 0 : async move {
2010 0 : startup_complete.wait().await;
2011 0 : this.spawn_heartbeat_driver().await;
2012 0 : }
2013 : });
2014 :
2015 : // Check that there is enough safekeepers configured that we can create new timelines
2016 0 : let test_sk_res_str = match this.safekeepers_for_new_timeline().await {
2017 0 : Ok(v) => format!("Ok({v:?})"),
2018 0 : Err(v) => format!("Err({v:})"),
2019 : };
2020 0 : tracing::info!(
2021 : timeline_safekeeper_count = config.timeline_safekeeper_count,
2022 : timelines_onto_safekeepers = config.timelines_onto_safekeepers,
2023 0 : "viability test result (test timeline creation on safekeepers): {test_sk_res_str}",
2024 : );
2025 :
2026 0 : Ok(this)
2027 0 : }
2028 :
2029 0 : pub(crate) async fn attach_hook(
2030 0 : &self,
2031 0 : attach_req: AttachHookRequest,
2032 0 : ) -> anyhow::Result<AttachHookResponse> {
2033 0 : let _tenant_lock = trace_exclusive_lock(
2034 0 : &self.tenant_op_locks,
2035 0 : attach_req.tenant_shard_id.tenant_id,
2036 0 : TenantOperations::AttachHook,
2037 0 : )
2038 0 : .await;
2039 :
2040 : // This is a test hook. To enable using it on tenants that were created directly with
2041 : // the pageserver API (not via this service), we will auto-create any missing tenant
2042 : // shards with default state.
2043 0 : let insert = {
2044 0 : match self
2045 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
2046 0 : .await
2047 : {
2048 0 : Ok(_) => false,
2049 0 : Err(ApiError::NotFound(_)) => true,
2050 0 : Err(e) => return Err(e.into()),
2051 : }
2052 : };
2053 :
2054 0 : if insert {
2055 0 : let config = attach_req.config.clone().unwrap_or_default();
2056 0 : let tsp = TenantShardPersistence {
2057 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
2058 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
2059 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
2060 0 : shard_stripe_size: 0,
2061 0 : generation: attach_req.generation_override.or(Some(0)),
2062 0 : generation_pageserver: None,
2063 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
2064 0 : config: serde_json::to_string(&config).unwrap(),
2065 0 : splitting: SplitState::default(),
2066 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2067 0 : .unwrap(),
2068 0 : preferred_az_id: None,
2069 0 : };
2070 :
2071 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
2072 0 : Err(e) => match e {
2073 : DatabaseError::Query(diesel::result::Error::DatabaseError(
2074 : DatabaseErrorKind::UniqueViolation,
2075 : _,
2076 : )) => {
2077 0 : tracing::info!(
2078 0 : "Raced with another request to insert tenant {}",
2079 : attach_req.tenant_shard_id
2080 : )
2081 : }
2082 0 : _ => return Err(e.into()),
2083 : },
2084 : Ok(()) => {
2085 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
2086 :
2087 0 : let mut shard = TenantShard::new(
2088 0 : attach_req.tenant_shard_id,
2089 0 : ShardIdentity::unsharded(),
2090 0 : PlacementPolicy::Attached(0),
2091 0 : None,
2092 : );
2093 0 : shard.config = config;
2094 :
2095 0 : let mut locked = self.inner.write().unwrap();
2096 0 : locked.tenants.insert(attach_req.tenant_shard_id, shard);
2097 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
2098 : }
2099 : }
2100 0 : }
2101 :
2102 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
2103 0 : let maybe_tenant_conf = {
2104 0 : let locked = self.inner.write().unwrap();
2105 0 : locked
2106 0 : .tenants
2107 0 : .get(&attach_req.tenant_shard_id)
2108 0 : .map(|t| t.config.clone())
2109 : };
2110 :
2111 0 : match maybe_tenant_conf {
2112 0 : Some(conf) => {
2113 0 : let new_generation = self
2114 0 : .persistence
2115 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
2116 0 : .await?;
2117 :
2118 : // Persist the placement policy update. This is required
2119 : // when we reattaching a detached tenant.
2120 0 : self.persistence
2121 0 : .update_tenant_shard(
2122 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
2123 0 : Some(PlacementPolicy::Attached(0)),
2124 0 : Some(conf),
2125 0 : None,
2126 0 : None,
2127 0 : )
2128 0 : .await?;
2129 0 : Some(new_generation)
2130 : }
2131 : None => {
2132 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
2133 : }
2134 : }
2135 : } else {
2136 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
2137 0 : None
2138 : };
2139 :
2140 0 : let mut locked = self.inner.write().unwrap();
2141 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2142 :
2143 0 : let tenant_shard = tenants
2144 0 : .get_mut(&attach_req.tenant_shard_id)
2145 0 : .expect("Checked for existence above");
2146 :
2147 0 : if let Some(new_generation) = new_generation {
2148 0 : tenant_shard.generation = Some(new_generation);
2149 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
2150 0 : } else {
2151 : // This is a detach notification. We must update placement policy to avoid re-attaching
2152 : // during background scheduling/reconciliation, or during storage controller restart.
2153 0 : assert!(attach_req.node_id.is_none());
2154 0 : tenant_shard.policy = PlacementPolicy::Detached;
2155 : }
2156 :
2157 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
2158 0 : tracing::info!(
2159 : tenant_id = %attach_req.tenant_shard_id,
2160 : ps_id = %attaching_pageserver,
2161 : generation = ?tenant_shard.generation,
2162 0 : "issuing",
2163 : );
2164 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
2165 0 : tracing::info!(
2166 : tenant_id = %attach_req.tenant_shard_id,
2167 : %ps_id,
2168 : generation = ?tenant_shard.generation,
2169 0 : "dropping",
2170 : );
2171 : } else {
2172 0 : tracing::info!(
2173 : tenant_id = %attach_req.tenant_shard_id,
2174 0 : "no-op: tenant already has no pageserver");
2175 : }
2176 0 : tenant_shard
2177 0 : .intent
2178 0 : .set_attached(scheduler, attach_req.node_id);
2179 :
2180 0 : tracing::info!(
2181 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}, config {:?}",
2182 : attach_req.tenant_shard_id,
2183 : tenant_shard.generation,
2184 : // TODO: this is an odd number of 0xf's
2185 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff)),
2186 : attach_req.config,
2187 : );
2188 :
2189 : // Trick the reconciler into not doing anything for this tenant: this helps
2190 : // tests that manually configure a tenant on the pagesrever, and then call this
2191 : // attach hook: they don't want background reconciliation to modify what they
2192 : // did to the pageserver.
2193 : #[cfg(feature = "testing")]
2194 : {
2195 0 : if let Some(node_id) = attach_req.node_id {
2196 0 : tenant_shard.observed.locations = HashMap::from([(
2197 0 : node_id,
2198 0 : ObservedStateLocation {
2199 0 : conf: Some(attached_location_conf(
2200 0 : tenant_shard.generation.unwrap(),
2201 0 : &tenant_shard.shard,
2202 0 : &tenant_shard.config,
2203 0 : &PlacementPolicy::Attached(0),
2204 0 : tenant_shard.intent.get_secondary().len(),
2205 0 : )),
2206 0 : },
2207 0 : )]);
2208 0 : } else {
2209 0 : tenant_shard.observed.locations.clear();
2210 0 : }
2211 : }
2212 :
2213 : Ok(AttachHookResponse {
2214 0 : generation: attach_req
2215 0 : .node_id
2216 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
2217 : })
2218 0 : }
2219 :
2220 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
2221 0 : let locked = self.inner.read().unwrap();
2222 :
2223 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
2224 :
2225 : InspectResponse {
2226 0 : attachment: tenant_shard.and_then(|s| {
2227 0 : s.intent
2228 0 : .get_attached()
2229 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
2230 0 : }),
2231 : }
2232 0 : }
2233 :
2234 : // When the availability state of a node transitions to active, we must do a full reconciliation
2235 : // of LocationConfigs on that node. This is because while a node was offline:
2236 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
2237 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
2238 : //
2239 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
2240 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
2241 : // this function.
2242 : //
2243 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
2244 : // for written for a single node rather than as a batch job for all nodes.
2245 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
2246 : async fn node_activate_reconcile(
2247 : &self,
2248 : mut node: Node,
2249 : _lock: &TracingExclusiveGuard<NodeOperations>,
2250 : ) -> Result<(), ApiError> {
2251 : // This Node is a mutable local copy: we will set it active so that we can use its
2252 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
2253 : // later.
2254 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
2255 :
2256 : let configs = match node
2257 : .with_client_retries(
2258 0 : |client| async move { client.list_location_config().await },
2259 : &self.http_client,
2260 : &self.config.pageserver_jwt_token,
2261 : 1,
2262 : 5,
2263 : SHORT_RECONCILE_TIMEOUT,
2264 : &self.cancel,
2265 : )
2266 : .await
2267 : {
2268 : None => {
2269 : // We're shutting down (the Node's cancellation token can't have fired, because
2270 : // we're the only scope that has a reference to it, and we didn't fire it).
2271 : return Err(ApiError::ShuttingDown);
2272 : }
2273 : Some(Err(e)) => {
2274 : // This node didn't succeed listing its locations: it may not proceed to active state
2275 : // as it is apparently unavailable.
2276 : return Err(ApiError::PreconditionFailed(
2277 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2278 : ));
2279 : }
2280 : Some(Ok(configs)) => configs,
2281 : };
2282 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2283 :
2284 : let mut cleanup = Vec::new();
2285 : let mut mismatched_locations = 0;
2286 : {
2287 : let mut locked = self.inner.write().unwrap();
2288 :
2289 : for (tenant_shard_id, reported) in configs.tenant_shards {
2290 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2291 : cleanup.push(tenant_shard_id);
2292 : continue;
2293 : };
2294 :
2295 : let on_record = &mut tenant_shard
2296 : .observed
2297 : .locations
2298 : .entry(node.get_id())
2299 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2300 : .conf;
2301 :
2302 : // If the location reported by the node does not match our observed state,
2303 : // then we mark it as uncertain and let the background reconciliation loop
2304 : // deal with it.
2305 : //
2306 : // Note that this also covers net new locations reported by the node.
2307 : if *on_record != reported {
2308 : mismatched_locations += 1;
2309 : *on_record = None;
2310 : }
2311 : }
2312 : }
2313 :
2314 : if mismatched_locations > 0 {
2315 : tracing::info!(
2316 : "Set observed state to None for {mismatched_locations} mismatched locations"
2317 : );
2318 : }
2319 :
2320 : for tenant_shard_id in cleanup {
2321 : tracing::info!("Detaching {tenant_shard_id}");
2322 : match node
2323 : .with_client_retries(
2324 0 : |client| async move {
2325 0 : let config = LocationConfig {
2326 0 : mode: LocationConfigMode::Detached,
2327 0 : generation: None,
2328 0 : secondary_conf: None,
2329 0 : shard_number: tenant_shard_id.shard_number.0,
2330 0 : shard_count: tenant_shard_id.shard_count.literal(),
2331 0 : shard_stripe_size: 0,
2332 0 : tenant_conf: models::TenantConfig::default(),
2333 0 : };
2334 0 : client
2335 0 : .location_config(tenant_shard_id, config, None, false)
2336 0 : .await
2337 0 : },
2338 : &self.http_client,
2339 : &self.config.pageserver_jwt_token,
2340 : 1,
2341 : 5,
2342 : SHORT_RECONCILE_TIMEOUT,
2343 : &self.cancel,
2344 : )
2345 : .await
2346 : {
2347 : None => {
2348 : // We're shutting down (the Node's cancellation token can't have fired, because
2349 : // we're the only scope that has a reference to it, and we didn't fire it).
2350 : return Err(ApiError::ShuttingDown);
2351 : }
2352 : Some(Err(e)) => {
2353 : // Do not let the node proceed to Active state if it is not responsive to requests
2354 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2355 : // detach completing: we should not let this node back into the set of nodes considered
2356 : // okay for scheduling.
2357 : return Err(ApiError::Conflict(format!(
2358 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2359 : )));
2360 : }
2361 : Some(Ok(_)) => {}
2362 : };
2363 : }
2364 :
2365 : Ok(())
2366 : }
2367 :
2368 0 : pub(crate) async fn re_attach(
2369 0 : &self,
2370 0 : reattach_req: ReAttachRequest,
2371 0 : ) -> Result<ReAttachResponse, ApiError> {
2372 0 : if let Some(register_req) = reattach_req.register {
2373 0 : self.node_register(register_req).await?;
2374 0 : }
2375 :
2376 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2377 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2378 :
2379 0 : tracing::info!(
2380 : node_id=%reattach_req.node_id,
2381 0 : "Incremented {} tenant shards' generations",
2382 0 : incremented_generations.len()
2383 : );
2384 :
2385 : // Apply the updated generation to our in-memory state, and
2386 : // gather discover secondary locations.
2387 0 : let mut locked = self.inner.write().unwrap();
2388 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2389 :
2390 0 : let mut response = ReAttachResponse {
2391 0 : tenants: Vec::new(),
2392 0 : };
2393 :
2394 : // [Hadron] If the pageserver reports in the reattach message that it has an empty disk, it's possible that it just
2395 : // recovered from a local disk failure. The response of the reattach request will contain a list of tenants but it
2396 : // will not be honored by the pageserver in this case (disk failure). We should make sure we clear any observed
2397 : // locations of tenants attached to the node so that the reconciler will discover the discrpancy and reconfigure the
2398 : // missing tenants on the node properly.
2399 0 : if self.config.handle_ps_local_disk_loss && reattach_req.empty_local_disk.unwrap_or(false) {
2400 0 : tracing::info!(
2401 0 : "Pageserver {node_id} reports empty local disk, clearing observed locations referencing the pageserver for all tenants",
2402 : node_id = reattach_req.node_id
2403 : );
2404 0 : let mut num_tenant_shards_affected = 0;
2405 0 : for (tenant_shard_id, shard) in tenants.iter_mut() {
2406 0 : if shard
2407 0 : .observed
2408 0 : .locations
2409 0 : .remove(&reattach_req.node_id)
2410 0 : .is_some()
2411 : {
2412 0 : tracing::info!("Cleared observed location for tenant shard {tenant_shard_id}");
2413 0 : num_tenant_shards_affected += 1;
2414 0 : }
2415 : }
2416 0 : tracing::info!(
2417 0 : "Cleared observed locations for {num_tenant_shards_affected} tenant shards"
2418 : );
2419 0 : }
2420 :
2421 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2422 : // to call location_conf API with an old generation. Wait for cancellation to complete
2423 : // before responding to this request. Requires well implemented CancellationToken logic
2424 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2425 : // request in flight over the network: TODO handle that by making location_conf API refuse
2426 : // to go backward in generations.
2427 :
2428 : // Scan through all shards, applying updates for ones where we updated generation
2429 : // and identifying shards that intend to have a secondary location on this node.
2430 0 : for (tenant_shard_id, shard) in tenants {
2431 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2432 0 : let new_gen = *new_gen;
2433 0 : response.tenants.push(ReAttachResponseTenant {
2434 0 : id: *tenant_shard_id,
2435 0 : r#gen: Some(new_gen.into().unwrap()),
2436 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2437 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2438 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2439 0 : // the stale/multi states at this point.
2440 0 : mode: LocationConfigMode::AttachedSingle,
2441 0 : stripe_size: shard.shard.stripe_size,
2442 0 : });
2443 :
2444 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2445 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2446 : // Why can we update `observed` even though we're not sure our response will be received
2447 : // by the pageserver? Because the pageserver will not proceed with startup until
2448 : // it has processed response: if it loses it, we'll see another request and increment
2449 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2450 0 : if let Some(conf) = observed.conf.as_mut() {
2451 0 : conf.generation = new_gen.into();
2452 0 : }
2453 0 : } else {
2454 0 : // This node has no observed state for the shard: perhaps it was offline
2455 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2456 0 : // will be prompted to learn the location's state before it makes changes.
2457 0 : shard
2458 0 : .observed
2459 0 : .locations
2460 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2461 0 : }
2462 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2463 0 : // Ordering: pageserver will not accept /location_config requests until it has
2464 0 : // finished processing the response from re-attach. So we can update our in-memory state
2465 0 : // now, and be confident that we are not stamping on the result of some later location config.
2466 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2467 0 : // so we might update observed state here, and then get over-written by some racing
2468 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2469 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2470 0 :
2471 0 : response.tenants.push(ReAttachResponseTenant {
2472 0 : id: *tenant_shard_id,
2473 0 : r#gen: None,
2474 0 : mode: LocationConfigMode::Secondary,
2475 0 : stripe_size: shard.shard.stripe_size,
2476 0 : });
2477 0 :
2478 0 : // We must not update observed, because we have no guarantee that our
2479 0 : // response will be received by the pageserver. This could leave it
2480 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2481 0 : }
2482 : }
2483 :
2484 : // We consider a node Active once we have composed a re-attach response, but we
2485 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2486 : // implicitly synchronizes the LocationConfigs on the node.
2487 : //
2488 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2489 : // but those requests will not be accepted by the node until it has finished processing
2490 : // the re-attach response.
2491 : //
2492 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2493 : // in [`Persistence::re_attach`].
2494 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2495 0 : let reset_scheduling = matches!(
2496 0 : node.get_scheduling(),
2497 : NodeSchedulingPolicy::PauseForRestart
2498 : | NodeSchedulingPolicy::Draining
2499 : | NodeSchedulingPolicy::Filling
2500 : | NodeSchedulingPolicy::Deleting
2501 : );
2502 :
2503 0 : let mut new_nodes = (**nodes).clone();
2504 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2505 0 : if reset_scheduling {
2506 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2507 0 : }
2508 :
2509 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2510 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2511 :
2512 0 : scheduler.node_upsert(node);
2513 0 : let new_nodes = Arc::new(new_nodes);
2514 0 : *nodes = new_nodes;
2515 : } else {
2516 0 : tracing::error!(
2517 0 : "Reattaching node {} was removed while processing the request",
2518 : reattach_req.node_id
2519 : );
2520 : }
2521 0 : }
2522 :
2523 0 : Ok(response)
2524 0 : }
2525 :
2526 0 : pub(crate) async fn validate(
2527 0 : &self,
2528 0 : validate_req: ValidateRequest,
2529 0 : ) -> Result<ValidateResponse, DatabaseError> {
2530 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2531 : // in-memory generation for a shard
2532 0 : let in_memory_result = {
2533 0 : let mut in_memory_result = Vec::new();
2534 0 : let locked = self.inner.read().unwrap();
2535 0 : for req_tenant in validate_req.tenants {
2536 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2537 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2538 0 : tracing::info!(
2539 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2540 : req_tenant.id,
2541 : req_tenant.r#gen,
2542 : tenant_shard.generation
2543 : );
2544 :
2545 0 : in_memory_result.push((
2546 0 : req_tenant.id,
2547 0 : Generation::new(req_tenant.r#gen),
2548 0 : valid,
2549 0 : ));
2550 : } else {
2551 : // This is legal: for example during a shard split the pageserver may still
2552 : // have deletions in its queue from the old pre-split shard, or after deletion
2553 : // of a tenant that was busy with compaction/gc while being deleted.
2554 0 : tracing::info!(
2555 0 : "Refusing deletion validation for missing shard {}",
2556 : req_tenant.id
2557 : );
2558 : }
2559 : }
2560 :
2561 0 : in_memory_result
2562 : };
2563 :
2564 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2565 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2566 0 : let db_generations = self
2567 0 : .persistence
2568 0 : .shard_generations(
2569 0 : in_memory_result
2570 0 : .iter()
2571 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2572 : )
2573 0 : .await?;
2574 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2575 :
2576 0 : let mut response = ValidateResponse {
2577 0 : tenants: Vec::new(),
2578 0 : };
2579 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2580 0 : let valid = if valid {
2581 0 : let db_generation = db_generations.get(&tenant_shard_id);
2582 0 : db_generation == Some(&Some(validate_generation))
2583 : } else {
2584 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2585 : // this prevents a pageserver from cleaning up an object in S3.
2586 0 : false
2587 : };
2588 :
2589 0 : response.tenants.push(ValidateResponseTenant {
2590 0 : id: tenant_shard_id,
2591 0 : valid,
2592 0 : })
2593 : }
2594 :
2595 0 : Ok(response)
2596 0 : }
2597 :
2598 0 : pub(crate) async fn tenant_create(
2599 0 : &self,
2600 0 : create_req: TenantCreateRequest,
2601 0 : ) -> Result<TenantCreateResponse, ApiError> {
2602 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2603 :
2604 : // Exclude any concurrent attempts to create/access the same tenant ID
2605 0 : let _tenant_lock = trace_exclusive_lock(
2606 0 : &self.tenant_op_locks,
2607 0 : create_req.new_tenant_id.tenant_id,
2608 0 : TenantOperations::Create,
2609 0 : )
2610 0 : .await;
2611 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2612 :
2613 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2614 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2615 : // accept compute notifications while it is in the process of creating. Reconciliation will
2616 : // be retried in the background.
2617 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2618 0 : }
2619 0 : Ok(response)
2620 0 : }
2621 :
2622 0 : pub(crate) async fn do_tenant_create(
2623 0 : &self,
2624 0 : create_req: TenantCreateRequest,
2625 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2626 0 : let placement_policy = create_req
2627 0 : .placement_policy
2628 0 : .clone()
2629 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2630 0 : .unwrap_or(PlacementPolicy::Attached(0));
2631 :
2632 : // This service expects to handle sharding itself: it is an error to try and directly create
2633 : // a particular shard here.
2634 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2635 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2636 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2637 0 : )));
2638 : } else {
2639 0 : create_req.new_tenant_id.tenant_id
2640 : };
2641 :
2642 0 : tracing::info!(
2643 0 : "Creating tenant {}, shard_count={:?}",
2644 : create_req.new_tenant_id,
2645 : create_req.shard_parameters.count,
2646 : );
2647 :
2648 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2649 0 : .map(|i| TenantShardId {
2650 0 : tenant_id,
2651 0 : shard_number: ShardNumber(i),
2652 0 : shard_count: create_req.shard_parameters.count,
2653 0 : })
2654 0 : .collect::<Vec<_>>();
2655 :
2656 : // If the caller specifies a None generation, it means "start from default". This is different
2657 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2658 : // an incompletely-onboarded tenant.
2659 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2660 0 : tracing::info!(
2661 0 : "tenant_create: secondary mode, generation is_some={}",
2662 0 : create_req.generation.is_some()
2663 : );
2664 0 : create_req.generation.map(Generation::new)
2665 : } else {
2666 0 : tracing::info!(
2667 0 : "tenant_create: not secondary mode, generation is_some={}",
2668 0 : create_req.generation.is_some()
2669 : );
2670 0 : Some(
2671 0 : create_req
2672 0 : .generation
2673 0 : .map(Generation::new)
2674 0 : .unwrap_or(INITIAL_GENERATION),
2675 0 : )
2676 : };
2677 :
2678 0 : let preferred_az_id = {
2679 0 : let locked = self.inner.read().unwrap();
2680 : // Idempotency: take the existing value if the tenant already exists
2681 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2682 0 : shard.preferred_az().cloned()
2683 : } else {
2684 0 : locked.scheduler.get_az_for_new_tenant()
2685 : }
2686 : };
2687 :
2688 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2689 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2690 : // during the creation, rather than risking leaving orphan objects in S3.
2691 0 : let persist_tenant_shards = create_ids
2692 0 : .iter()
2693 0 : .map(|tenant_shard_id| TenantShardPersistence {
2694 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2695 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2696 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2697 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2698 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2699 : // The pageserver is not known until scheduling happens: we will set this column when
2700 : // incrementing the generation the first time we attach to a pageserver.
2701 0 : generation_pageserver: None,
2702 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2703 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2704 0 : splitting: SplitState::default(),
2705 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2706 0 : .unwrap(),
2707 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2708 0 : })
2709 0 : .collect();
2710 :
2711 0 : match self
2712 0 : .persistence
2713 0 : .insert_tenant_shards(persist_tenant_shards)
2714 0 : .await
2715 : {
2716 0 : Ok(_) => {}
2717 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2718 : DatabaseErrorKind::UniqueViolation,
2719 : _,
2720 : ))) => {
2721 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2722 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2723 : // creation's shard count.
2724 0 : tracing::info!(
2725 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2726 : );
2727 : }
2728 : // Any other database error is unexpected and a bug.
2729 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2730 : };
2731 :
2732 0 : let mut schedule_context = ScheduleContext::default();
2733 0 : let mut schedule_error = None;
2734 0 : let mut response_shards = Vec::new();
2735 0 : for tenant_shard_id in create_ids {
2736 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2737 :
2738 0 : let outcome = self
2739 0 : .do_initial_shard_scheduling(
2740 0 : tenant_shard_id,
2741 0 : initial_generation,
2742 0 : create_req.shard_parameters,
2743 0 : create_req.config.clone(),
2744 0 : placement_policy.clone(),
2745 0 : preferred_az_id.as_ref(),
2746 0 : &mut schedule_context,
2747 0 : )
2748 0 : .await;
2749 :
2750 0 : match outcome {
2751 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2752 0 : InitialShardScheduleOutcome::NotScheduled => {}
2753 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2754 0 : schedule_error = Some(err);
2755 0 : }
2756 : }
2757 : }
2758 :
2759 : // If we failed to schedule shards, then they are still created in the controller,
2760 : // but we return an error to the requester to avoid a silent failure when someone
2761 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2762 : // are present in the system. We do this here rather than in the above loop, to
2763 : // avoid situations where we only create a subset of shards in the tenant.
2764 0 : if let Some(e) = schedule_error {
2765 0 : return Err(ApiError::Conflict(format!(
2766 0 : "Failed to schedule shard(s): {e}"
2767 0 : )));
2768 0 : }
2769 :
2770 0 : let waiters = {
2771 0 : let mut locked = self.inner.write().unwrap();
2772 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2773 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2774 0 : .tenant_creation_hint(true)
2775 0 : .build();
2776 0 : tenants
2777 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2778 0 : .filter_map(|(_shard_id, shard)| {
2779 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2780 0 : })
2781 0 : .collect::<Vec<_>>()
2782 : };
2783 :
2784 0 : Ok((
2785 0 : TenantCreateResponse {
2786 0 : shards: response_shards,
2787 0 : },
2788 0 : waiters,
2789 0 : ))
2790 0 : }
2791 :
2792 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2793 : /// case of a new tenant and a pre-existing one.
2794 : #[allow(clippy::too_many_arguments)]
2795 0 : async fn do_initial_shard_scheduling(
2796 0 : &self,
2797 0 : tenant_shard_id: TenantShardId,
2798 0 : initial_generation: Option<Generation>,
2799 0 : shard_params: ShardParameters,
2800 0 : config: TenantConfig,
2801 0 : placement_policy: PlacementPolicy,
2802 0 : preferred_az_id: Option<&AvailabilityZone>,
2803 0 : schedule_context: &mut ScheduleContext,
2804 0 : ) -> InitialShardScheduleOutcome {
2805 0 : let mut locked = self.inner.write().unwrap();
2806 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2807 :
2808 : use std::collections::btree_map::Entry;
2809 0 : match tenants.entry(tenant_shard_id) {
2810 0 : Entry::Occupied(mut entry) => {
2811 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2812 :
2813 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2814 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2815 0 : }
2816 :
2817 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2818 0 : let generation = entry
2819 0 : .get()
2820 0 : .generation
2821 0 : .expect("Generation is set when in attached mode");
2822 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2823 0 : shard_id: tenant_shard_id,
2824 0 : node_id: *node_id,
2825 0 : generation: generation.into().unwrap(),
2826 0 : })
2827 : } else {
2828 0 : InitialShardScheduleOutcome::NotScheduled
2829 : }
2830 : }
2831 0 : Entry::Vacant(entry) => {
2832 0 : let state = entry.insert(TenantShard::new(
2833 0 : tenant_shard_id,
2834 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2835 0 : placement_policy,
2836 0 : preferred_az_id.cloned(),
2837 : ));
2838 :
2839 0 : state.generation = initial_generation;
2840 0 : state.config = config;
2841 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2842 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2843 0 : }
2844 :
2845 : // Only include shards in result if we are attaching: the purpose
2846 : // of the response is to tell the caller where the shards are attached.
2847 0 : if let Some(node_id) = state.intent.get_attached() {
2848 0 : let generation = state
2849 0 : .generation
2850 0 : .expect("Generation is set when in attached mode");
2851 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2852 0 : shard_id: tenant_shard_id,
2853 0 : node_id: *node_id,
2854 0 : generation: generation.into().unwrap(),
2855 0 : })
2856 : } else {
2857 0 : InitialShardScheduleOutcome::NotScheduled
2858 : }
2859 : }
2860 : }
2861 0 : }
2862 :
2863 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2864 : /// wait for reconciliation to complete before responding.
2865 0 : async fn await_waiters(
2866 0 : &self,
2867 0 : waiters: Vec<ReconcilerWaiter>,
2868 0 : timeout: Duration,
2869 0 : ) -> Result<(), ReconcileWaitError> {
2870 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2871 0 : for waiter in waiters {
2872 0 : let timeout = deadline.duration_since(Instant::now());
2873 0 : waiter.wait_timeout(timeout).await?;
2874 : }
2875 :
2876 0 : Ok(())
2877 0 : }
2878 :
2879 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2880 : /// in progress
2881 0 : async fn await_waiters_remainder(
2882 0 : &self,
2883 0 : waiters: Vec<ReconcilerWaiter>,
2884 0 : timeout: Duration,
2885 0 : ) -> Vec<ReconcilerWaiter> {
2886 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2887 0 : for waiter in waiters.iter() {
2888 0 : let timeout = deadline.duration_since(Instant::now());
2889 0 : let _ = waiter.wait_timeout(timeout).await;
2890 : }
2891 :
2892 0 : waiters
2893 0 : .into_iter()
2894 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2895 0 : .collect::<Vec<_>>()
2896 0 : }
2897 :
2898 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2899 : /// and transform it into either a tenant creation of a series of shard updates.
2900 : ///
2901 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2902 : /// still be returned.
2903 0 : fn tenant_location_config_prepare(
2904 0 : &self,
2905 0 : tenant_id: TenantId,
2906 0 : req: TenantLocationConfigRequest,
2907 0 : ) -> TenantCreateOrUpdate {
2908 0 : let mut updates = Vec::new();
2909 0 : let mut locked = self.inner.write().unwrap();
2910 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2911 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2912 :
2913 : // Use location config mode as an indicator of policy.
2914 0 : let placement_policy = match req.config.mode {
2915 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2916 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2917 : LocationConfigMode::AttachedMulti
2918 : | LocationConfigMode::AttachedSingle
2919 : | LocationConfigMode::AttachedStale => {
2920 0 : if nodes.len() > 1 {
2921 0 : PlacementPolicy::Attached(1)
2922 : } else {
2923 : // Convenience for dev/test: if we just have one pageserver, import
2924 : // tenants into non-HA mode so that scheduling will succeed.
2925 0 : PlacementPolicy::Attached(0)
2926 : }
2927 : }
2928 : };
2929 :
2930 : // Ordinarily we do not update scheduling policy, but when making major changes
2931 : // like detaching or demoting to secondary-only, we need to force the scheduling
2932 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2933 0 : let scheduling_policy = match req.config.mode {
2934 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2935 : // Special case: when making major changes like detaching or demoting to secondary-only,
2936 : // we need to force the scheduling mode to Active, or nothing will happen.
2937 0 : Some(ShardSchedulingPolicy::Active)
2938 : }
2939 : LocationConfigMode::AttachedMulti
2940 : | LocationConfigMode::AttachedSingle
2941 : | LocationConfigMode::AttachedStale => {
2942 : // While attached, continue to respect whatever the existing scheduling mode is.
2943 0 : None
2944 : }
2945 : };
2946 :
2947 0 : let mut create = true;
2948 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2949 : // Saw an existing shard: this is not a creation
2950 0 : create = false;
2951 :
2952 : // Shards may have initially been created by a Secondary request, where we
2953 : // would have left generation as None.
2954 : //
2955 : // We only update generation the first time we see an attached-mode request,
2956 : // and if there is no existing generation set. The caller is responsible for
2957 : // ensuring that no non-storage-controller pageserver ever uses a higher
2958 : // generation than they passed in here.
2959 : use LocationConfigMode::*;
2960 0 : let set_generation = match req.config.mode {
2961 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2962 0 : req.config.generation.map(Generation::new)
2963 : }
2964 0 : _ => None,
2965 : };
2966 :
2967 0 : updates.push(ShardUpdate {
2968 0 : tenant_shard_id: *shard_id,
2969 0 : placement_policy: placement_policy.clone(),
2970 0 : tenant_config: req.config.tenant_conf.clone(),
2971 0 : generation: set_generation,
2972 0 : scheduling_policy,
2973 0 : });
2974 : }
2975 :
2976 0 : if create {
2977 : use LocationConfigMode::*;
2978 0 : let generation = match req.config.mode {
2979 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2980 : // If a caller provided a generation in a non-attached request, ignore it
2981 : // and leave our generation as None: this enables a subsequent update to set
2982 : // the generation when setting an attached mode for the first time.
2983 0 : _ => None,
2984 : };
2985 :
2986 0 : TenantCreateOrUpdate::Create(
2987 0 : // Synthesize a creation request
2988 0 : TenantCreateRequest {
2989 0 : new_tenant_id: tenant_shard_id,
2990 0 : generation,
2991 0 : shard_parameters: ShardParameters {
2992 0 : count: tenant_shard_id.shard_count,
2993 0 : // We only import un-sharded or single-sharded tenants, so stripe
2994 0 : // size can be made up arbitrarily here.
2995 0 : stripe_size: DEFAULT_STRIPE_SIZE,
2996 0 : },
2997 0 : placement_policy: Some(placement_policy),
2998 0 : config: req.config.tenant_conf,
2999 0 : },
3000 0 : )
3001 : } else {
3002 0 : assert!(!updates.is_empty());
3003 0 : TenantCreateOrUpdate::Update(updates)
3004 : }
3005 0 : }
3006 :
3007 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
3008 : /// the tenant is present in memory. If not, load it from the database. If it is found
3009 : /// in neither location, return a NotFound error.
3010 : ///
3011 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
3012 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
3013 0 : async fn maybe_load_tenant(
3014 0 : &self,
3015 0 : tenant_id: TenantId,
3016 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
3017 0 : ) -> Result<(), ApiError> {
3018 : // Check if the tenant is present in memory, and select an AZ to use when loading
3019 : // if we will load it.
3020 0 : let load_in_az = {
3021 0 : let locked = self.inner.read().unwrap();
3022 0 : let existing = locked
3023 0 : .tenants
3024 0 : .range(TenantShardId::tenant_range(tenant_id))
3025 0 : .next();
3026 :
3027 : // If the tenant is not present in memory, we expect to load it from database,
3028 : // so let's figure out what AZ to load it into while we have self.inner locked.
3029 0 : if existing.is_none() {
3030 0 : locked
3031 0 : .scheduler
3032 0 : .get_az_for_new_tenant()
3033 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
3034 0 : "No AZ with nodes found to load tenant"
3035 0 : )))?
3036 : } else {
3037 : // We already have this tenant in memory
3038 0 : return Ok(());
3039 : }
3040 : };
3041 :
3042 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
3043 0 : if tenant_shards.is_empty() {
3044 0 : return Err(ApiError::NotFound(
3045 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3046 0 : ));
3047 0 : }
3048 :
3049 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
3050 0 : self.persistence
3051 0 : .set_tenant_shard_preferred_azs(
3052 0 : tenant_shards
3053 0 : .iter()
3054 0 : .map(|t| {
3055 0 : (
3056 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
3057 0 : Some(load_in_az.clone()),
3058 0 : )
3059 0 : })
3060 0 : .collect(),
3061 : )
3062 0 : .await?;
3063 :
3064 0 : let mut locked = self.inner.write().unwrap();
3065 0 : tracing::info!(
3066 0 : "Loaded {} shards for tenant {}",
3067 0 : tenant_shards.len(),
3068 : tenant_id
3069 : );
3070 :
3071 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
3072 0 : let intent = IntentState::new(Some(load_in_az.clone()));
3073 0 : let shard =
3074 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
3075 :
3076 : // Sanity check: when loading on-demand, we should always be loaded something Detached
3077 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
3078 0 : if shard.policy != PlacementPolicy::Detached {
3079 0 : tracing::error!(
3080 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
3081 : shard.tenant_shard_id,
3082 : shard.policy
3083 : );
3084 0 : }
3085 :
3086 0 : (shard.tenant_shard_id, shard)
3087 0 : }));
3088 :
3089 0 : Ok(())
3090 0 : }
3091 :
3092 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
3093 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
3094 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
3095 : ///
3096 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
3097 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
3098 0 : fn maybe_drop_tenant(
3099 0 : &self,
3100 0 : tenant_id: TenantId,
3101 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
3102 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
3103 0 : ) {
3104 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
3105 0 : if tenant_shards.all(|(_id, shard)| {
3106 0 : shard.policy == PlacementPolicy::Detached
3107 0 : && shard.reconciler.is_none()
3108 0 : && shard.observed.is_empty()
3109 0 : }) {
3110 0 : let keys = locked
3111 0 : .tenants
3112 0 : .range(TenantShardId::tenant_range(tenant_id))
3113 0 : .map(|(id, _)| id)
3114 0 : .copied()
3115 0 : .collect::<Vec<_>>();
3116 0 : for key in keys {
3117 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
3118 0 : locked.tenants.remove(&key);
3119 : }
3120 0 : }
3121 0 : }
3122 :
3123 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
3124 : /// directly with pageservers into this service.
3125 : ///
3126 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
3127 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
3128 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
3129 : /// tenant's source of generation numbers.
3130 : ///
3131 : /// The mode in this request coarse-grained control of tenants:
3132 : /// - Call with mode Attached* to upsert the tenant.
3133 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
3134 : /// to set an existing tenant to PolicyMode::Secondary
3135 : /// - Call with mode Detached to switch to PolicyMode::Detached
3136 0 : pub(crate) async fn tenant_location_config(
3137 0 : &self,
3138 0 : tenant_shard_id: TenantShardId,
3139 0 : req: TenantLocationConfigRequest,
3140 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
3141 : // We require an exclusive lock, because we are updating both persistent and in-memory state
3142 0 : let _tenant_lock = trace_exclusive_lock(
3143 0 : &self.tenant_op_locks,
3144 0 : tenant_shard_id.tenant_id,
3145 0 : TenantOperations::LocationConfig,
3146 0 : )
3147 0 : .await;
3148 :
3149 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
3150 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3151 0 : "This API is for importing single-sharded or unsharded tenants"
3152 0 : )));
3153 : } else {
3154 0 : tenant_shard_id.tenant_id
3155 : };
3156 :
3157 : // In case we are waking up a Detached tenant
3158 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
3159 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
3160 0 : // This is a creation or an update
3161 0 : }
3162 0 : Err(e) => {
3163 0 : return Err(e);
3164 : }
3165 : };
3166 :
3167 : // First check if this is a creation or an update
3168 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
3169 :
3170 0 : let mut result = TenantLocationConfigResponse {
3171 0 : shards: Vec::new(),
3172 0 : stripe_size: None,
3173 0 : };
3174 0 : let waiters = match create_or_update {
3175 0 : TenantCreateOrUpdate::Create(create_req) => {
3176 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
3177 0 : result.shards = create_resp
3178 0 : .shards
3179 0 : .into_iter()
3180 0 : .map(|s| TenantShardLocation {
3181 0 : node_id: s.node_id,
3182 0 : shard_id: s.shard_id,
3183 0 : })
3184 0 : .collect();
3185 0 : waiters
3186 : }
3187 0 : TenantCreateOrUpdate::Update(updates) => {
3188 : // Persist updates
3189 : // Ordering: write to the database before applying changes in-memory, so that
3190 : // we will not appear time-travel backwards on a restart.
3191 :
3192 0 : let mut schedule_context = ScheduleContext::default();
3193 : for ShardUpdate {
3194 0 : tenant_shard_id,
3195 0 : placement_policy,
3196 0 : tenant_config,
3197 0 : generation,
3198 0 : scheduling_policy,
3199 0 : } in &updates
3200 : {
3201 0 : self.persistence
3202 0 : .update_tenant_shard(
3203 0 : TenantFilter::Shard(*tenant_shard_id),
3204 0 : Some(placement_policy.clone()),
3205 0 : Some(tenant_config.clone()),
3206 0 : *generation,
3207 0 : *scheduling_policy,
3208 0 : )
3209 0 : .await?;
3210 : }
3211 :
3212 : // Apply updates in-memory
3213 0 : let mut waiters = Vec::new();
3214 : {
3215 0 : let mut locked = self.inner.write().unwrap();
3216 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3217 :
3218 : for ShardUpdate {
3219 0 : tenant_shard_id,
3220 0 : placement_policy,
3221 0 : tenant_config,
3222 0 : generation: update_generation,
3223 0 : scheduling_policy,
3224 0 : } in updates
3225 : {
3226 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
3227 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
3228 0 : continue;
3229 : };
3230 :
3231 : // Update stripe size
3232 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
3233 0 : result.stripe_size = Some(shard.shard.stripe_size);
3234 0 : }
3235 :
3236 0 : shard.policy = placement_policy;
3237 0 : shard.config = tenant_config;
3238 0 : if let Some(generation) = update_generation {
3239 0 : shard.generation = Some(generation);
3240 0 : }
3241 :
3242 0 : if let Some(scheduling_policy) = scheduling_policy {
3243 0 : shard.set_scheduling_policy(scheduling_policy);
3244 0 : }
3245 :
3246 0 : shard.schedule(scheduler, &mut schedule_context)?;
3247 :
3248 0 : let maybe_waiter =
3249 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3250 0 : if let Some(waiter) = maybe_waiter {
3251 0 : waiters.push(waiter);
3252 0 : }
3253 :
3254 0 : if let Some(node_id) = shard.intent.get_attached() {
3255 0 : result.shards.push(TenantShardLocation {
3256 0 : shard_id: tenant_shard_id,
3257 0 : node_id: *node_id,
3258 0 : })
3259 0 : }
3260 : }
3261 : }
3262 0 : waiters
3263 : }
3264 : };
3265 :
3266 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3267 : // Do not treat a reconcile error as fatal: we have already applied any requested
3268 : // Intent changes, and the reconcile can fail for external reasons like unavailable
3269 : // compute notification API. In these cases, it is important that we do not
3270 : // cause the cloud control plane to retry forever on this API.
3271 0 : tracing::warn!(
3272 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
3273 : );
3274 0 : }
3275 :
3276 : // Logging the full result is useful because it lets us cross-check what the cloud control
3277 : // plane's tenant_shards table should contain.
3278 0 : tracing::info!("Complete, returning {result:?}");
3279 :
3280 0 : Ok(result)
3281 0 : }
3282 :
3283 0 : pub(crate) async fn tenant_config_patch(
3284 0 : &self,
3285 0 : req: TenantConfigPatchRequest,
3286 0 : ) -> Result<(), ApiError> {
3287 0 : let _tenant_lock = trace_exclusive_lock(
3288 0 : &self.tenant_op_locks,
3289 0 : req.tenant_id,
3290 0 : TenantOperations::ConfigPatch,
3291 0 : )
3292 0 : .await;
3293 :
3294 0 : let tenant_id = req.tenant_id;
3295 0 : let patch = req.config;
3296 :
3297 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3298 :
3299 0 : let base = {
3300 0 : let locked = self.inner.read().unwrap();
3301 0 : let shards = locked
3302 0 : .tenants
3303 0 : .range(TenantShardId::tenant_range(req.tenant_id));
3304 :
3305 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
3306 :
3307 0 : let first = match configs.peek() {
3308 0 : Some(first) => (*first).clone(),
3309 : None => {
3310 0 : return Err(ApiError::NotFound(
3311 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3312 0 : ));
3313 : }
3314 : };
3315 :
3316 0 : if !configs.all_equal() {
3317 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3318 : // This can't happen because we atomically update the database records
3319 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3320 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3321 0 : "Tenant configs for {} are mismatched",
3322 0 : req.tenant_id
3323 0 : )));
3324 0 : }
3325 :
3326 0 : first
3327 : };
3328 :
3329 0 : let updated_config = base
3330 0 : .apply_patch(patch)
3331 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3332 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3333 0 : .await
3334 0 : }
3335 :
3336 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3337 : // We require an exclusive lock, because we are updating persistent and in-memory state
3338 0 : let _tenant_lock = trace_exclusive_lock(
3339 0 : &self.tenant_op_locks,
3340 0 : req.tenant_id,
3341 0 : TenantOperations::ConfigSet,
3342 0 : )
3343 0 : .await;
3344 :
3345 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3346 :
3347 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3348 0 : .await
3349 0 : }
3350 :
3351 0 : async fn set_tenant_config_and_reconcile(
3352 0 : &self,
3353 0 : tenant_id: TenantId,
3354 0 : config: TenantConfig,
3355 0 : ) -> Result<(), ApiError> {
3356 0 : self.persistence
3357 0 : .update_tenant_shard(
3358 0 : TenantFilter::Tenant(tenant_id),
3359 0 : None,
3360 0 : Some(config.clone()),
3361 0 : None,
3362 0 : None,
3363 0 : )
3364 0 : .await?;
3365 :
3366 0 : let waiters = {
3367 0 : let mut waiters = Vec::new();
3368 0 : let mut locked = self.inner.write().unwrap();
3369 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3370 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3371 0 : shard.config = config.clone();
3372 0 : if let Some(waiter) =
3373 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3374 0 : {
3375 0 : waiters.push(waiter);
3376 0 : }
3377 : }
3378 0 : waiters
3379 : };
3380 :
3381 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3382 : // Treat this as success because we have stored the configuration. If e.g.
3383 : // a node was unavailable at this time, it should not stop us accepting a
3384 : // configuration change.
3385 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3386 0 : }
3387 :
3388 0 : Ok(())
3389 0 : }
3390 :
3391 0 : pub(crate) fn tenant_config_get(
3392 0 : &self,
3393 0 : tenant_id: TenantId,
3394 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3395 0 : let config = {
3396 0 : let locked = self.inner.read().unwrap();
3397 :
3398 0 : match locked
3399 0 : .tenants
3400 0 : .range(TenantShardId::tenant_range(tenant_id))
3401 0 : .next()
3402 : {
3403 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3404 : None => {
3405 0 : return Err(ApiError::NotFound(
3406 0 : anyhow::anyhow!("Tenant not found").into(),
3407 0 : ));
3408 : }
3409 : }
3410 : };
3411 :
3412 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3413 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3414 : // and `effective_config` in the response is meaningless, but we retain that syntax
3415 : // in order to remain compatible with the pageserver API.
3416 :
3417 0 : let response = HashMap::from([
3418 : (
3419 : "tenant_specific_overrides",
3420 0 : serde_json::to_value(&config)
3421 0 : .context("serializing tenant specific overrides")
3422 0 : .map_err(ApiError::InternalServerError)?,
3423 : ),
3424 : (
3425 0 : "effective_config",
3426 0 : serde_json::to_value(&config)
3427 0 : .context("serializing effective config")
3428 0 : .map_err(ApiError::InternalServerError)?,
3429 : ),
3430 : ]);
3431 :
3432 0 : Ok(response)
3433 0 : }
3434 :
3435 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3436 0 : &self,
3437 0 : time_travel_req: &TenantTimeTravelRequest,
3438 0 : tenant_id: TenantId,
3439 0 : timestamp: Cow<'_, str>,
3440 0 : done_if_after: Cow<'_, str>,
3441 0 : ) -> Result<(), ApiError> {
3442 0 : let _tenant_lock = trace_exclusive_lock(
3443 0 : &self.tenant_op_locks,
3444 0 : tenant_id,
3445 0 : TenantOperations::TimeTravelRemoteStorage,
3446 0 : )
3447 0 : .await;
3448 :
3449 0 : let node = {
3450 0 : let mut locked = self.inner.write().unwrap();
3451 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3452 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3453 : // but only at the start of the process, so it's really just to prevent operator
3454 : // mistakes.
3455 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3456 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3457 : {
3458 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3459 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3460 0 : )));
3461 0 : }
3462 0 : let maybe_attached = shard
3463 0 : .observed
3464 0 : .locations
3465 0 : .iter()
3466 0 : .filter_map(|(node_id, observed_location)| {
3467 0 : observed_location
3468 0 : .conf
3469 0 : .as_ref()
3470 0 : .map(|loc| (node_id, observed_location, loc.mode))
3471 0 : })
3472 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3473 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3474 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3475 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3476 0 : )));
3477 0 : }
3478 : }
3479 0 : let scheduler = &mut locked.scheduler;
3480 : // Right now we only perform the operation on a single node without parallelization
3481 : // TODO fan out the operation to multiple nodes for better performance
3482 0 : let node_id = scheduler.any_available_node()?;
3483 0 : let node = locked
3484 0 : .nodes
3485 0 : .get(&node_id)
3486 0 : .expect("Pageservers may not be deleted while lock is active");
3487 0 : node.clone()
3488 : };
3489 :
3490 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3491 0 : let mut counts = time_travel_req
3492 0 : .shard_counts
3493 0 : .iter()
3494 0 : .copied()
3495 0 : .collect::<HashSet<_>>()
3496 0 : .into_iter()
3497 0 : .collect::<Vec<_>>();
3498 0 : counts.sort_unstable();
3499 :
3500 0 : for count in counts {
3501 0 : let shard_ids = (0..count.count())
3502 0 : .map(|i| TenantShardId {
3503 0 : tenant_id,
3504 0 : shard_number: ShardNumber(i),
3505 0 : shard_count: count,
3506 0 : })
3507 0 : .collect::<Vec<_>>();
3508 0 : for tenant_shard_id in shard_ids {
3509 0 : let client = PageserverClient::new(
3510 0 : node.get_id(),
3511 0 : self.http_client.clone(),
3512 0 : node.base_url(),
3513 0 : self.config.pageserver_jwt_token.as_deref(),
3514 : );
3515 :
3516 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3517 :
3518 0 : client
3519 0 : .tenant_time_travel_remote_storage(
3520 0 : tenant_shard_id,
3521 0 : ×tamp,
3522 0 : &done_if_after,
3523 0 : )
3524 0 : .await
3525 0 : .map_err(|e| {
3526 0 : ApiError::InternalServerError(anyhow::anyhow!(
3527 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3528 0 : node
3529 0 : ))
3530 0 : })?;
3531 : }
3532 : }
3533 0 : Ok(())
3534 0 : }
3535 :
3536 0 : pub(crate) async fn tenant_secondary_download(
3537 0 : &self,
3538 0 : tenant_id: TenantId,
3539 0 : wait: Option<Duration>,
3540 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3541 0 : let _tenant_lock = trace_shared_lock(
3542 0 : &self.tenant_op_locks,
3543 0 : tenant_id,
3544 0 : TenantOperations::SecondaryDownload,
3545 0 : )
3546 0 : .await;
3547 :
3548 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3549 0 : let targets = {
3550 0 : let locked = self.inner.read().unwrap();
3551 0 : let mut targets = Vec::new();
3552 :
3553 0 : for (tenant_shard_id, shard) in
3554 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3555 : {
3556 0 : for node_id in shard.intent.get_secondary() {
3557 0 : let node = locked
3558 0 : .nodes
3559 0 : .get(node_id)
3560 0 : .expect("Pageservers may not be deleted while referenced");
3561 0 :
3562 0 : targets.push((*tenant_shard_id, node.clone()));
3563 0 : }
3564 : }
3565 0 : targets
3566 : };
3567 :
3568 : // Issue concurrent requests to all shards' locations
3569 0 : let mut futs = FuturesUnordered::new();
3570 0 : for (tenant_shard_id, node) in targets {
3571 0 : let client = PageserverClient::new(
3572 0 : node.get_id(),
3573 0 : self.http_client.clone(),
3574 0 : node.base_url(),
3575 0 : self.config.pageserver_jwt_token.as_deref(),
3576 : );
3577 0 : futs.push(async move {
3578 0 : let result = client
3579 0 : .tenant_secondary_download(tenant_shard_id, wait)
3580 0 : .await;
3581 0 : (result, node, tenant_shard_id)
3582 0 : })
3583 : }
3584 :
3585 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3586 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3587 : // well as more general cases like 503s, 500s, or timeouts.
3588 0 : let mut aggregate_progress = SecondaryProgress::default();
3589 0 : let mut aggregate_status: Option<StatusCode> = None;
3590 0 : let mut error: Option<mgmt_api::Error> = None;
3591 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3592 0 : match result {
3593 0 : Err(e) => {
3594 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3595 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3596 : // than they had hoped for.
3597 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3598 0 : error = Some(e)
3599 : }
3600 0 : Ok((status_code, progress)) => {
3601 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3602 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3603 0 : aggregate_progress.layers_total += progress.layers_total;
3604 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3605 0 : aggregate_progress.bytes_total += progress.bytes_total;
3606 0 : aggregate_progress.heatmap_mtime =
3607 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3608 0 : aggregate_status = match aggregate_status {
3609 0 : None => Some(status_code),
3610 0 : Some(StatusCode::OK) => Some(status_code),
3611 0 : Some(cur) => {
3612 : // Other status codes (e.g. 202) -- do not overwrite.
3613 0 : Some(cur)
3614 : }
3615 : };
3616 : }
3617 : }
3618 : }
3619 :
3620 : // If any of the shards return 202, indicate our result as 202.
3621 0 : match aggregate_status {
3622 : None => {
3623 0 : match error {
3624 0 : Some(e) => {
3625 : // No successes, and an error: surface it
3626 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3627 : }
3628 : None => {
3629 : // No shards found
3630 0 : Err(ApiError::NotFound(
3631 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3632 0 : ))
3633 : }
3634 : }
3635 : }
3636 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3637 : }
3638 0 : }
3639 :
3640 0 : pub(crate) async fn tenant_delete(
3641 0 : self: &Arc<Self>,
3642 0 : tenant_id: TenantId,
3643 0 : ) -> Result<StatusCode, ApiError> {
3644 0 : let _tenant_lock =
3645 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3646 :
3647 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3648 :
3649 : // Detach all shards. This also deletes local pageserver shard data.
3650 0 : let (detach_waiters, node) = {
3651 0 : let mut detach_waiters = Vec::new();
3652 0 : let mut locked = self.inner.write().unwrap();
3653 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3654 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3655 : // Update the tenant's intent to remove all attachments
3656 0 : shard.policy = PlacementPolicy::Detached;
3657 0 : shard
3658 0 : .schedule(scheduler, &mut ScheduleContext::default())
3659 0 : .expect("De-scheduling is infallible");
3660 0 : debug_assert!(shard.intent.get_attached().is_none());
3661 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3662 :
3663 0 : if let Some(waiter) =
3664 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3665 0 : {
3666 0 : detach_waiters.push(waiter);
3667 0 : }
3668 : }
3669 :
3670 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3671 : // was attached, just has to be able to see the S3 content)
3672 0 : let node_id = scheduler.any_available_node()?;
3673 0 : let node = nodes
3674 0 : .get(&node_id)
3675 0 : .expect("Pageservers may not be deleted while lock is active");
3676 0 : (detach_waiters, node.clone())
3677 : };
3678 :
3679 : // This reconcile wait can fail in a few ways:
3680 : // A there is a very long queue for the reconciler semaphore
3681 : // B some pageserver is failing to handle a detach promptly
3682 : // C some pageserver goes offline right at the moment we send it a request.
3683 : //
3684 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3685 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3686 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3687 : // deleting the underlying data).
3688 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3689 0 : .await?;
3690 :
3691 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3692 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3693 : // the tenant ID prefix, including all shards (even possibly stale ones).
3694 0 : match node
3695 0 : .with_client_retries(
3696 0 : |client| async move {
3697 0 : client
3698 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3699 0 : .await
3700 0 : },
3701 0 : &self.http_client,
3702 0 : &self.config.pageserver_jwt_token,
3703 : 1,
3704 : 3,
3705 : RECONCILE_TIMEOUT,
3706 0 : &self.cancel,
3707 : )
3708 0 : .await
3709 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3710 : {
3711 0 : Ok(_) => {}
3712 : Err(mgmt_api::Error::Cancelled) => {
3713 0 : return Err(ApiError::ShuttingDown);
3714 : }
3715 0 : Err(e) => {
3716 : // This is unexpected: remote deletion should be infallible, unless the object store
3717 : // at large is unavailable.
3718 0 : tracing::error!("Error deleting via node {node}: {e}");
3719 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3720 : }
3721 : }
3722 :
3723 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3724 : // our in-memory state and database state.
3725 :
3726 : // Ordering: we delete persistent state first: if we then
3727 : // crash, we will drop the in-memory state.
3728 :
3729 : // Drop persistent state.
3730 0 : self.persistence.delete_tenant(tenant_id).await?;
3731 :
3732 : // Drop in-memory state
3733 : {
3734 0 : let mut locked = self.inner.write().unwrap();
3735 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3736 :
3737 : // Dereference Scheduler from shards before dropping them
3738 0 : for (_tenant_shard_id, shard) in
3739 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3740 0 : {
3741 0 : shard.intent.clear(scheduler);
3742 0 : }
3743 :
3744 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3745 0 : tracing::info!(
3746 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3747 0 : locked.tenants.len()
3748 : );
3749 : };
3750 :
3751 : // Delete the tenant from safekeepers (if needed)
3752 0 : self.tenant_delete_safekeepers(tenant_id)
3753 0 : .instrument(tracing::info_span!("tenant_delete_safekeepers", %tenant_id))
3754 0 : .await?;
3755 :
3756 : // Success is represented as 404, to imitate the existing pageserver deletion API
3757 0 : Ok(StatusCode::NOT_FOUND)
3758 0 : }
3759 :
3760 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3761 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3762 : /// the tenant's policies (configuration) within the storage controller
3763 0 : pub(crate) async fn tenant_update_policy(
3764 0 : &self,
3765 0 : tenant_id: TenantId,
3766 0 : req: TenantPolicyRequest,
3767 0 : ) -> Result<(), ApiError> {
3768 : // We require an exclusive lock, because we are updating persistent and in-memory state
3769 0 : let _tenant_lock = trace_exclusive_lock(
3770 0 : &self.tenant_op_locks,
3771 0 : tenant_id,
3772 0 : TenantOperations::UpdatePolicy,
3773 0 : )
3774 0 : .await;
3775 :
3776 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3777 :
3778 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3779 :
3780 : let TenantPolicyRequest {
3781 0 : placement,
3782 0 : mut scheduling,
3783 0 : } = req;
3784 :
3785 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3786 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3787 : // this to take effect.
3788 0 : if scheduling.is_none() {
3789 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3790 0 : }
3791 0 : }
3792 :
3793 0 : self.persistence
3794 0 : .update_tenant_shard(
3795 0 : TenantFilter::Tenant(tenant_id),
3796 0 : placement.clone(),
3797 0 : None,
3798 0 : None,
3799 0 : scheduling,
3800 0 : )
3801 0 : .await?;
3802 :
3803 0 : let mut schedule_context = ScheduleContext::default();
3804 0 : let mut locked = self.inner.write().unwrap();
3805 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3806 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3807 0 : if let Some(placement) = &placement {
3808 0 : shard.policy = placement.clone();
3809 :
3810 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3811 0 : "Updated placement policy to {placement:?}");
3812 0 : }
3813 :
3814 0 : if let Some(scheduling) = &scheduling {
3815 0 : shard.set_scheduling_policy(*scheduling);
3816 :
3817 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3818 0 : "Updated scheduling policy to {scheduling:?}");
3819 0 : }
3820 :
3821 : // In case scheduling is being switched back on, try it now.
3822 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3823 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3824 : }
3825 :
3826 0 : Ok(())
3827 0 : }
3828 :
3829 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3830 0 : &self,
3831 0 : tenant_id: TenantId,
3832 0 : mut create_req: TimelineCreateRequest,
3833 0 : ) -> Result<TimelineInfo, ApiError> {
3834 0 : tracing::info!(
3835 0 : "Creating timeline {}/{}",
3836 : tenant_id,
3837 : create_req.new_timeline_id,
3838 : );
3839 :
3840 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3841 0 : if targets.0.is_empty() {
3842 0 : return Err(ApiError::NotFound(
3843 0 : anyhow::anyhow!("Tenant not found").into(),
3844 0 : ));
3845 0 : };
3846 :
3847 0 : let (shard_zero_tid, shard_zero_locations) =
3848 0 : targets.0.pop_first().expect("Must have at least one shard");
3849 0 : assert!(shard_zero_tid.is_shard_zero());
3850 :
3851 0 : async fn create_one(
3852 0 : tenant_shard_id: TenantShardId,
3853 0 : locations: ShardMutationLocations,
3854 0 : http_client: reqwest::Client,
3855 0 : jwt: Option<String>,
3856 0 : mut create_req: TimelineCreateRequest,
3857 0 : ) -> Result<TimelineInfo, ApiError> {
3858 0 : let latest = locations.latest.node;
3859 :
3860 0 : tracing::info!(
3861 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3862 : tenant_shard_id,
3863 : create_req.new_timeline_id,
3864 : locations.latest.generation
3865 : );
3866 :
3867 0 : let client =
3868 0 : PageserverClient::new(latest.get_id(), http_client.clone(), latest.base_url(), jwt.as_deref());
3869 :
3870 0 : let timeline_info = client
3871 0 : .timeline_create(tenant_shard_id, &create_req)
3872 0 : .await
3873 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3874 :
3875 : // If we are going to create the timeline on some stale locations for shard 0, then ask them to re-use
3876 : // the initdb generated by the latest location, rather than generating their own. This avoids racing uploads
3877 : // of initdb to S3 which might not be binary-identical if different pageservers have different postgres binaries.
3878 0 : if tenant_shard_id.is_shard_zero() {
3879 0 : if let models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } = &mut create_req.mode {
3880 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id);
3881 0 : }
3882 0 : }
3883 :
3884 : // We propagate timeline creations to all attached locations such that a compute
3885 : // for the new timeline is able to start regardless of the current state of the
3886 : // tenant shard reconciliation.
3887 0 : for location in locations.other {
3888 0 : tracing::info!(
3889 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3890 : tenant_shard_id,
3891 : create_req.new_timeline_id,
3892 : location.node,
3893 : location.generation
3894 : );
3895 :
3896 0 : let client = PageserverClient::new(
3897 0 : location.node.get_id(),
3898 0 : http_client.clone(),
3899 0 : location.node.base_url(),
3900 0 : jwt.as_deref(),
3901 : );
3902 :
3903 0 : let res = client
3904 0 : .timeline_create(tenant_shard_id, &create_req)
3905 0 : .await;
3906 :
3907 0 : if let Err(e) = res {
3908 0 : match e {
3909 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3910 0 : // Tenant might have been detached from the stale location,
3911 0 : // so ignore 404s.
3912 0 : },
3913 : _ => {
3914 0 : return Err(passthrough_api_error(&location.node, e));
3915 : }
3916 : }
3917 0 : }
3918 : }
3919 :
3920 0 : Ok(timeline_info)
3921 0 : }
3922 :
3923 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3924 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3925 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3926 : //
3927 : // This also enables non-zero shards to use the initdb that shard 0 generated and uploaded to S3, rather than
3928 : // independently generating their own initdb. This guarantees that shards cannot end up with different initial
3929 : // states if e.g. they have different postgres binary versions.
3930 0 : let timeline_info = create_one(
3931 0 : shard_zero_tid,
3932 0 : shard_zero_locations,
3933 0 : self.http_client.clone(),
3934 0 : self.config.pageserver_jwt_token.clone(),
3935 0 : create_req.clone(),
3936 0 : )
3937 0 : .await?;
3938 :
3939 : // Update the create request for shards >= 0
3940 0 : match &mut create_req.mode {
3941 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3942 0 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3943 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3944 0 : },
3945 0 : models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } => {
3946 : // For shards >= 0, do not run initdb: use the one that shard 0 uploaded to S3
3947 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id)
3948 : }
3949 0 : _ => {}
3950 : }
3951 :
3952 : // Create timeline on remaining shards with number >0
3953 0 : if !targets.0.is_empty() {
3954 : // If we had multiple shards, issue requests for the remainder now.
3955 0 : let jwt = &self.config.pageserver_jwt_token;
3956 0 : self.tenant_for_shards(
3957 0 : targets
3958 0 : .0
3959 0 : .iter()
3960 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3961 0 : .collect(),
3962 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3963 0 : let create_req = create_req.clone();
3964 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3965 0 : Box::pin(create_one(
3966 0 : tenant_shard_id,
3967 0 : mutation_locations,
3968 0 : self.http_client.clone(),
3969 0 : jwt.clone(),
3970 0 : create_req,
3971 0 : ))
3972 0 : },
3973 : )
3974 0 : .await?;
3975 0 : }
3976 :
3977 0 : Ok(timeline_info)
3978 0 : })
3979 0 : .await?
3980 0 : }
3981 :
3982 0 : pub(crate) async fn tenant_timeline_create(
3983 0 : self: &Arc<Self>,
3984 0 : tenant_id: TenantId,
3985 0 : create_req: TimelineCreateRequest,
3986 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
3987 0 : let safekeepers = self.config.timelines_onto_safekeepers;
3988 0 : let timeline_id = create_req.new_timeline_id;
3989 :
3990 0 : tracing::info!(
3991 0 : mode=%create_req.mode_tag(),
3992 : %safekeepers,
3993 0 : "Creating timeline {}/{}",
3994 : tenant_id,
3995 : timeline_id,
3996 : );
3997 :
3998 0 : let _tenant_lock = trace_shared_lock(
3999 0 : &self.tenant_op_locks,
4000 0 : tenant_id,
4001 0 : TenantOperations::TimelineCreate,
4002 0 : )
4003 0 : .await;
4004 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
4005 0 : let is_import = create_req.is_import();
4006 0 : let read_only = matches!(
4007 0 : create_req.mode,
4008 : models::TimelineCreateRequestMode::Branch {
4009 : read_only: true,
4010 : ..
4011 : }
4012 : );
4013 :
4014 0 : if is_import {
4015 : // Ensure that there is no split on-going.
4016 : // [`Self::tenant_shard_split`] holds the exclusive tenant lock
4017 : // for the duration of the split, but here we handle the case
4018 : // where we restarted and the split is being aborted.
4019 0 : let locked = self.inner.read().unwrap();
4020 0 : let splitting = locked
4021 0 : .tenants
4022 0 : .range(TenantShardId::tenant_range(tenant_id))
4023 0 : .any(|(_id, shard)| shard.splitting != SplitState::Idle);
4024 :
4025 0 : if splitting {
4026 0 : return Err(ApiError::Conflict("Tenant is splitting shard".to_string()));
4027 0 : }
4028 0 : }
4029 :
4030 0 : let timeline_info = self
4031 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
4032 0 : .await?;
4033 :
4034 0 : let selected_safekeepers = if is_import {
4035 0 : let shards = {
4036 0 : let locked = self.inner.read().unwrap();
4037 0 : locked
4038 0 : .tenants
4039 0 : .range(TenantShardId::tenant_range(tenant_id))
4040 0 : .map(|(ts_id, _)| ts_id.to_index())
4041 0 : .collect::<Vec<_>>()
4042 : };
4043 :
4044 0 : if !shards
4045 0 : .iter()
4046 0 : .map(|shard_index| shard_index.shard_count)
4047 0 : .all_equal()
4048 : {
4049 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4050 0 : "Inconsistent shard count"
4051 0 : )));
4052 0 : }
4053 :
4054 0 : let import = TimelineImport {
4055 0 : tenant_id,
4056 0 : timeline_id,
4057 0 : shard_statuses: ShardImportStatuses::new(shards),
4058 0 : };
4059 :
4060 0 : let inserted = self
4061 0 : .persistence
4062 0 : .insert_timeline_import(import.to_persistent())
4063 0 : .await
4064 0 : .context("timeline import insert")
4065 0 : .map_err(ApiError::InternalServerError)?;
4066 :
4067 : // Set the importing flag on the tenant shards
4068 0 : self.inner
4069 0 : .write()
4070 0 : .unwrap()
4071 0 : .tenants
4072 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4073 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Importing);
4074 :
4075 0 : match inserted {
4076 : true => {
4077 0 : tracing::info!(%tenant_id, %timeline_id, "Inserted timeline import");
4078 : }
4079 : false => {
4080 0 : tracing::info!(%tenant_id, %timeline_id, "Timeline import entry already present");
4081 : }
4082 : }
4083 :
4084 0 : None
4085 0 : } else if safekeepers || read_only {
4086 : // Note that for imported timelines, we do not create the timeline on the safekeepers
4087 : // straight away. Instead, we do it once the import finalized such that we know what
4088 : // start LSN to provide for the safekeepers. This is done in
4089 : // [`Self::finalize_timeline_import`].
4090 0 : let res = self
4091 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
4092 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
4093 0 : .await?;
4094 0 : Some(res)
4095 : } else {
4096 0 : None
4097 : };
4098 :
4099 0 : Ok(TimelineCreateResponseStorcon {
4100 0 : timeline_info,
4101 0 : safekeepers: selected_safekeepers,
4102 0 : })
4103 0 : }
4104 :
4105 : #[instrument(skip_all, fields(
4106 : tenant_id=%req.tenant_shard_id.tenant_id,
4107 : shard_id=%req.tenant_shard_id.shard_slug(),
4108 : timeline_id=%req.timeline_id,
4109 : ))]
4110 : pub(crate) async fn handle_timeline_shard_import_progress(
4111 : self: &Arc<Self>,
4112 : req: TimelineImportStatusRequest,
4113 : ) -> Result<ShardImportStatus, ApiError> {
4114 : let validity = self
4115 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4116 : .await?;
4117 : match validity {
4118 : ShardGenerationValidity::Valid => {
4119 : // fallthrough
4120 : }
4121 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4122 : tracing::info!(
4123 : claimed=?claimed.into(),
4124 0 : actual=?actual.and_then(|g| g.into()),
4125 : "Rejecting import progress fetch from stale generation"
4126 : );
4127 :
4128 : return Err(ApiError::BadRequest(anyhow::anyhow!("Invalid generation")));
4129 : }
4130 : }
4131 :
4132 : let maybe_import = self
4133 : .persistence
4134 : .get_timeline_import(req.tenant_shard_id.tenant_id, req.timeline_id)
4135 : .await?;
4136 :
4137 0 : let import = maybe_import.ok_or_else(|| {
4138 0 : ApiError::NotFound(
4139 0 : format!(
4140 0 : "import for {}/{} not found",
4141 0 : req.tenant_shard_id.tenant_id, req.timeline_id
4142 0 : )
4143 0 : .into(),
4144 0 : )
4145 0 : })?;
4146 :
4147 : import
4148 : .shard_statuses
4149 : .0
4150 : .get(&req.tenant_shard_id.to_index())
4151 : .cloned()
4152 0 : .ok_or_else(|| {
4153 0 : ApiError::NotFound(
4154 0 : format!("shard {} not found", req.tenant_shard_id.shard_slug()).into(),
4155 0 : )
4156 0 : })
4157 : }
4158 :
4159 : #[instrument(skip_all, fields(
4160 : tenant_id=%req.tenant_shard_id.tenant_id,
4161 : shard_id=%req.tenant_shard_id.shard_slug(),
4162 : timeline_id=%req.timeline_id,
4163 : ))]
4164 : pub(crate) async fn handle_timeline_shard_import_progress_upcall(
4165 : self: &Arc<Self>,
4166 : req: PutTimelineImportStatusRequest,
4167 : ) -> Result<(), ApiError> {
4168 : let validity = self
4169 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4170 : .await?;
4171 : match validity {
4172 : ShardGenerationValidity::Valid => {
4173 : // fallthrough
4174 : }
4175 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4176 : tracing::info!(
4177 : claimed=?claimed.into(),
4178 0 : actual=?actual.and_then(|g| g.into()),
4179 : "Rejecting import progress update from stale generation"
4180 : );
4181 :
4182 : return Err(ApiError::PreconditionFailed("Invalid generation".into()));
4183 : }
4184 : }
4185 :
4186 : let res = self
4187 : .persistence
4188 : .update_timeline_import(req.tenant_shard_id, req.timeline_id, req.status)
4189 : .await;
4190 : let timeline_import = match res {
4191 : Ok(Ok(Some(timeline_import))) => timeline_import,
4192 : Ok(Ok(None)) => {
4193 : // Idempotency: we've already seen and handled this update.
4194 : return Ok(());
4195 : }
4196 : Ok(Err(logical_err)) => {
4197 : return Err(logical_err.into());
4198 : }
4199 : Err(db_err) => {
4200 : return Err(db_err.into());
4201 : }
4202 : };
4203 :
4204 : tracing::info!(
4205 : tenant_id=%req.tenant_shard_id.tenant_id,
4206 : timeline_id=%req.timeline_id,
4207 : shard_id=%req.tenant_shard_id.shard_slug(),
4208 : "Updated timeline import status to: {timeline_import:?}");
4209 :
4210 : if timeline_import.is_complete() {
4211 : tokio::task::spawn({
4212 : let this = self.clone();
4213 0 : async move { this.finalize_timeline_import(timeline_import).await }
4214 : });
4215 : }
4216 :
4217 : Ok(())
4218 : }
4219 :
4220 : /// Check that a provided generation for some tenant shard is the most recent one.
4221 : ///
4222 : /// Validate with the in-mem state first, and, if that passes, validate with the
4223 : /// database state which is authoritative.
4224 0 : async fn validate_shard_generation(
4225 0 : self: &Arc<Self>,
4226 0 : tenant_shard_id: TenantShardId,
4227 0 : generation: Generation,
4228 0 : ) -> Result<ShardGenerationValidity, ApiError> {
4229 : {
4230 0 : let locked = self.inner.read().unwrap();
4231 0 : let tenant_shard =
4232 0 : locked
4233 0 : .tenants
4234 0 : .get(&tenant_shard_id)
4235 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4236 0 : "{} shard not found",
4237 0 : tenant_shard_id
4238 0 : )))?;
4239 :
4240 0 : if tenant_shard.generation != Some(generation) {
4241 0 : return Ok(ShardGenerationValidity::Mismatched {
4242 0 : claimed: generation,
4243 0 : actual: tenant_shard.generation,
4244 0 : });
4245 0 : }
4246 : }
4247 :
4248 0 : let mut db_generations = self
4249 0 : .persistence
4250 0 : .shard_generations(std::iter::once(&tenant_shard_id))
4251 0 : .await?;
4252 0 : let (_tid, db_generation) =
4253 0 : db_generations
4254 0 : .pop()
4255 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4256 0 : "{} shard not found",
4257 0 : tenant_shard_id
4258 0 : )))?;
4259 :
4260 0 : if db_generation != Some(generation) {
4261 0 : return Ok(ShardGenerationValidity::Mismatched {
4262 0 : claimed: generation,
4263 0 : actual: db_generation,
4264 0 : });
4265 0 : }
4266 :
4267 0 : Ok(ShardGenerationValidity::Valid)
4268 0 : }
4269 :
4270 : /// Finalize the import of a timeline
4271 : ///
4272 : /// This method should be called once all shards have reported that the import is complete.
4273 : /// Firstly, it polls the post import timeline activation endpoint exposed by the pageserver.
4274 : /// Once the timeline is active on all shards, the timeline also gets created on the
4275 : /// safekeepers. Finally, notify cplane of the import completion (whether failed or
4276 : /// successful), and remove the import from the database and in-memory.
4277 : ///
4278 : /// If this method gets pre-empted by shut down, it will be called again at start-up (on-going
4279 : /// imports are stored in the database).
4280 : ///
4281 : /// # Cancel-Safety
4282 : /// Not cancel safe.
4283 : /// If the caller stops polling, the import will not be removed from
4284 : /// [`ServiceState::imports_finalizing`].
4285 : #[instrument(skip_all, fields(
4286 : tenant_id=%import.tenant_id,
4287 : timeline_id=%import.timeline_id,
4288 : ))]
4289 :
4290 : async fn finalize_timeline_import(
4291 : self: &Arc<Self>,
4292 : import: TimelineImport,
4293 : ) -> Result<(), TimelineImportFinalizeError> {
4294 : let tenant_timeline = (import.tenant_id, import.timeline_id);
4295 :
4296 : let (_finalize_import_guard, cancel) = {
4297 : let mut locked = self.inner.write().unwrap();
4298 : let gate = Gate::default();
4299 : let cancel = CancellationToken::default();
4300 :
4301 : let guard = gate.enter().unwrap();
4302 :
4303 : locked.imports_finalizing.insert(
4304 : tenant_timeline,
4305 : FinalizingImport {
4306 : gate,
4307 : cancel: cancel.clone(),
4308 : },
4309 : );
4310 :
4311 : (guard, cancel)
4312 : };
4313 :
4314 : let res = tokio::select! {
4315 : res = self.finalize_timeline_import_impl(import) => {
4316 : res
4317 : },
4318 : _ = cancel.cancelled() => {
4319 : Err(TimelineImportFinalizeError::Cancelled)
4320 : }
4321 : };
4322 :
4323 : let mut locked = self.inner.write().unwrap();
4324 : locked.imports_finalizing.remove(&tenant_timeline);
4325 :
4326 : res
4327 : }
4328 :
4329 0 : async fn finalize_timeline_import_impl(
4330 0 : self: &Arc<Self>,
4331 0 : import: TimelineImport,
4332 0 : ) -> Result<(), TimelineImportFinalizeError> {
4333 0 : tracing::info!("Finalizing timeline import");
4334 :
4335 0 : pausable_failpoint!("timeline-import-pre-cplane-notification");
4336 :
4337 0 : let tenant_id = import.tenant_id;
4338 0 : let timeline_id = import.timeline_id;
4339 :
4340 0 : let import_error = import.completion_error();
4341 0 : match import_error {
4342 0 : Some(err) => {
4343 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Err(err))
4344 0 : .await?;
4345 0 : tracing::warn!("Timeline import completed with shard errors");
4346 0 : Ok(())
4347 : }
4348 0 : None => match self.activate_timeline_post_import(&import).await {
4349 0 : Ok(timeline_info) => {
4350 0 : tracing::info!("Post import timeline activation complete");
4351 :
4352 0 : if self.config.timelines_onto_safekeepers {
4353 : // Now that we know the start LSN of this timeline, create it on the
4354 : // safekeepers.
4355 0 : self.tenant_timeline_create_safekeepers_until_success(
4356 0 : import.tenant_id,
4357 0 : timeline_info,
4358 0 : )
4359 0 : .await?;
4360 0 : }
4361 :
4362 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Ok(()))
4363 0 : .await?;
4364 :
4365 0 : tracing::info!("Timeline import completed successfully");
4366 0 : Ok(())
4367 : }
4368 : Err(TimelineImportFinalizeError::ShuttingDown) => {
4369 : // We got pre-empted by shut down and will resume after the restart.
4370 0 : Err(TimelineImportFinalizeError::ShuttingDown)
4371 : }
4372 0 : Err(err) => {
4373 : // Any finalize error apart from shut down is permanent and requires us to notify
4374 : // cplane such that it can clean up.
4375 0 : tracing::error!("Import finalize failed with permanent error: {err}");
4376 0 : self.notify_cplane_and_delete_import(
4377 0 : tenant_id,
4378 0 : timeline_id,
4379 0 : Err(err.to_string()),
4380 0 : )
4381 0 : .await?;
4382 0 : Err(err)
4383 : }
4384 : },
4385 : }
4386 0 : }
4387 :
4388 0 : async fn notify_cplane_and_delete_import(
4389 0 : self: &Arc<Self>,
4390 0 : tenant_id: TenantId,
4391 0 : timeline_id: TimelineId,
4392 0 : import_result: ImportResult,
4393 0 : ) -> Result<(), TimelineImportFinalizeError> {
4394 0 : let import_failed = import_result.is_err();
4395 0 : tracing::info!(%import_failed, "Notifying cplane of import completion");
4396 :
4397 0 : let client = UpcallClient::new(self.get_config(), self.cancel.child_token());
4398 0 : client
4399 0 : .notify_import_complete(tenant_id, timeline_id, import_result)
4400 0 : .await
4401 0 : .map_err(|_err| TimelineImportFinalizeError::ShuttingDown)?;
4402 :
4403 0 : if let Err(err) = self
4404 0 : .persistence
4405 0 : .delete_timeline_import(tenant_id, timeline_id)
4406 0 : .await
4407 : {
4408 0 : tracing::warn!("Failed to delete timeline import entry from database: {err}");
4409 0 : }
4410 :
4411 0 : self.inner
4412 0 : .write()
4413 0 : .unwrap()
4414 0 : .tenants
4415 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4416 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Idle);
4417 :
4418 0 : Ok(())
4419 0 : }
4420 :
4421 : /// Activate an imported timeline on all shards once the import is complete.
4422 : /// Returns the [`TimelineInfo`] reported by shard zero.
4423 0 : async fn activate_timeline_post_import(
4424 0 : self: &Arc<Self>,
4425 0 : import: &TimelineImport,
4426 0 : ) -> Result<TimelineInfo, TimelineImportFinalizeError> {
4427 : const TIMELINE_ACTIVATE_TIMEOUT: Duration = Duration::from_millis(128);
4428 :
4429 0 : let mut shards_to_activate: HashSet<ShardIndex> =
4430 0 : import.shard_statuses.0.keys().cloned().collect();
4431 0 : let mut shard_zero_timeline_info = None;
4432 :
4433 0 : while !shards_to_activate.is_empty() {
4434 0 : if self.cancel.is_cancelled() {
4435 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4436 0 : }
4437 :
4438 0 : let targets = {
4439 0 : let locked = self.inner.read().unwrap();
4440 0 : let mut targets = Vec::new();
4441 :
4442 0 : for (tenant_shard_id, shard) in locked
4443 0 : .tenants
4444 0 : .range(TenantShardId::tenant_range(import.tenant_id))
4445 : {
4446 0 : if !import
4447 0 : .shard_statuses
4448 0 : .0
4449 0 : .contains_key(&tenant_shard_id.to_index())
4450 : {
4451 0 : return Err(TimelineImportFinalizeError::MismatchedShards(
4452 0 : tenant_shard_id.to_index(),
4453 0 : ));
4454 0 : }
4455 :
4456 0 : if let Some(node_id) = shard.intent.get_attached() {
4457 0 : let node = locked
4458 0 : .nodes
4459 0 : .get(node_id)
4460 0 : .expect("Pageservers may not be deleted while referenced");
4461 0 : targets.push((*tenant_shard_id, node.clone()));
4462 0 : }
4463 : }
4464 :
4465 0 : targets
4466 : };
4467 :
4468 0 : let targeted_tenant_shards: Vec<_> = targets.iter().map(|(tid, _node)| *tid).collect();
4469 :
4470 0 : let results = self
4471 0 : .tenant_for_shards_api(
4472 0 : targets,
4473 0 : |tenant_shard_id, client| async move {
4474 0 : client
4475 0 : .activate_post_import(
4476 0 : tenant_shard_id,
4477 0 : import.timeline_id,
4478 0 : TIMELINE_ACTIVATE_TIMEOUT,
4479 0 : )
4480 0 : .await
4481 0 : },
4482 : 1,
4483 : 1,
4484 : SHORT_RECONCILE_TIMEOUT,
4485 0 : &self.cancel,
4486 : )
4487 0 : .await;
4488 :
4489 0 : let mut failed = 0;
4490 0 : for (tid, (_, result)) in targeted_tenant_shards.iter().zip(results.into_iter()) {
4491 0 : match result {
4492 0 : Ok(ok) => {
4493 0 : if tid.is_shard_zero() {
4494 0 : shard_zero_timeline_info = Some(ok);
4495 0 : }
4496 :
4497 0 : shards_to_activate.remove(&tid.to_index());
4498 : }
4499 0 : Err(_err) => {
4500 0 : failed += 1;
4501 0 : }
4502 : }
4503 : }
4504 :
4505 0 : if failed > 0 {
4506 0 : tracing::info!(
4507 0 : "Failed to activate timeline on {failed} shards post import. Will retry"
4508 : );
4509 0 : }
4510 :
4511 0 : tokio::select! {
4512 0 : _ = tokio::time::sleep(Duration::from_millis(250)) => {},
4513 0 : _ = self.cancel.cancelled() => {
4514 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4515 : }
4516 : }
4517 : }
4518 :
4519 0 : Ok(shard_zero_timeline_info.expect("All shards replied"))
4520 0 : }
4521 :
4522 0 : async fn finalize_timeline_imports(self: &Arc<Self>, imports: Vec<TimelineImport>) {
4523 0 : futures::future::join_all(
4524 0 : imports
4525 0 : .into_iter()
4526 0 : .map(|import| self.finalize_timeline_import(import)),
4527 : )
4528 0 : .await;
4529 0 : }
4530 :
4531 : /// Delete a timeline import if it exists
4532 : ///
4533 : /// Firstly, delete the entry from the database. Any updates
4534 : /// from pageservers after the update will fail with a 404, so the
4535 : /// import cannot progress into finalizing state if it's not there already.
4536 : /// Secondly, cancel the finalization if one is in progress.
4537 0 : pub(crate) async fn maybe_delete_timeline_import(
4538 0 : self: &Arc<Self>,
4539 0 : tenant_id: TenantId,
4540 0 : timeline_id: TimelineId,
4541 0 : ) -> Result<(), DatabaseError> {
4542 0 : let tenant_has_ongoing_import = {
4543 0 : let locked = self.inner.read().unwrap();
4544 0 : locked
4545 0 : .tenants
4546 0 : .range(TenantShardId::tenant_range(tenant_id))
4547 0 : .any(|(_tid, shard)| shard.importing == TimelineImportState::Importing)
4548 : };
4549 :
4550 0 : if !tenant_has_ongoing_import {
4551 0 : return Ok(());
4552 0 : }
4553 :
4554 0 : self.persistence
4555 0 : .delete_timeline_import(tenant_id, timeline_id)
4556 0 : .await?;
4557 :
4558 0 : let maybe_finalizing = {
4559 0 : let mut locked = self.inner.write().unwrap();
4560 0 : locked.imports_finalizing.remove(&(tenant_id, timeline_id))
4561 : };
4562 :
4563 0 : if let Some(finalizing) = maybe_finalizing {
4564 0 : finalizing.cancel.cancel();
4565 0 : finalizing.gate.close().await;
4566 0 : }
4567 :
4568 0 : Ok(())
4569 0 : }
4570 :
4571 0 : pub(crate) async fn tenant_timeline_archival_config(
4572 0 : &self,
4573 0 : tenant_id: TenantId,
4574 0 : timeline_id: TimelineId,
4575 0 : req: TimelineArchivalConfigRequest,
4576 0 : ) -> Result<(), ApiError> {
4577 0 : tracing::info!(
4578 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
4579 : req.state
4580 : );
4581 :
4582 0 : let _tenant_lock = trace_shared_lock(
4583 0 : &self.tenant_op_locks,
4584 0 : tenant_id,
4585 0 : TenantOperations::TimelineArchivalConfig,
4586 0 : )
4587 0 : .await;
4588 :
4589 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4590 0 : if targets.0.is_empty() {
4591 0 : return Err(ApiError::NotFound(
4592 0 : anyhow::anyhow!("Tenant not found").into(),
4593 0 : ));
4594 0 : }
4595 0 : async fn config_one(
4596 0 : tenant_shard_id: TenantShardId,
4597 0 : timeline_id: TimelineId,
4598 0 : node: Node,
4599 0 : http_client: reqwest::Client,
4600 0 : jwt: Option<String>,
4601 0 : req: TimelineArchivalConfigRequest,
4602 0 : ) -> Result<(), ApiError> {
4603 0 : tracing::info!(
4604 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4605 : );
4606 :
4607 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4608 :
4609 0 : client
4610 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
4611 0 : .await
4612 0 : .map_err(|e| match e {
4613 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
4614 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
4615 : }
4616 0 : _ => passthrough_api_error(&node, e),
4617 0 : })
4618 0 : }
4619 :
4620 : // no shard needs to go first/last; the operation should be idempotent
4621 : // TODO: it would be great to ensure that all shards return the same error
4622 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4623 0 : let results = self
4624 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4625 0 : futures::FutureExt::boxed(config_one(
4626 0 : tenant_shard_id,
4627 0 : timeline_id,
4628 0 : node,
4629 0 : self.http_client.clone(),
4630 0 : self.config.pageserver_jwt_token.clone(),
4631 0 : req.clone(),
4632 0 : ))
4633 0 : })
4634 0 : .await?;
4635 0 : assert!(!results.is_empty(), "must have at least one result");
4636 :
4637 0 : Ok(())
4638 0 : }).await?
4639 0 : }
4640 :
4641 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
4642 0 : &self,
4643 0 : tenant_id: TenantId,
4644 0 : timeline_id: TimelineId,
4645 0 : behavior: Option<DetachBehavior>,
4646 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
4647 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
4648 :
4649 0 : let _tenant_lock = trace_shared_lock(
4650 0 : &self.tenant_op_locks,
4651 0 : tenant_id,
4652 0 : TenantOperations::TimelineDetachAncestor,
4653 0 : )
4654 0 : .await;
4655 :
4656 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4657 0 : if targets.0.is_empty() {
4658 0 : return Err(ApiError::NotFound(
4659 0 : anyhow::anyhow!("Tenant not found").into(),
4660 0 : ));
4661 0 : }
4662 :
4663 0 : async fn detach_one(
4664 0 : tenant_shard_id: TenantShardId,
4665 0 : timeline_id: TimelineId,
4666 0 : node: Node,
4667 0 : http_client: reqwest::Client,
4668 0 : jwt: Option<String>,
4669 0 : behavior: Option<DetachBehavior>,
4670 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
4671 0 : tracing::info!(
4672 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4673 : );
4674 :
4675 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4676 :
4677 0 : client
4678 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id, behavior)
4679 0 : .await
4680 0 : .map_err(|e| {
4681 : use mgmt_api::Error;
4682 :
4683 0 : match e {
4684 : // no ancestor (ever)
4685 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
4686 0 : "{node}: {}",
4687 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
4688 0 : )),
4689 : // too many ancestors
4690 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
4691 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
4692 : }
4693 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
4694 : // avoid turning these into conflicts to remain compatible with
4695 : // pageservers, 500 errors are sadly retryable with timeline ancestor
4696 : // detach
4697 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
4698 : }
4699 : // rest can be mapped as usual
4700 0 : other => passthrough_api_error(&node, other),
4701 : }
4702 0 : })
4703 0 : .map(|res| (tenant_shard_id.shard_number, res))
4704 0 : }
4705 :
4706 : // no shard needs to go first/last; the operation should be idempotent
4707 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4708 0 : let mut results = self
4709 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4710 0 : futures::FutureExt::boxed(detach_one(
4711 0 : tenant_shard_id,
4712 0 : timeline_id,
4713 0 : node,
4714 0 : self.http_client.clone(),
4715 0 : self.config.pageserver_jwt_token.clone(),
4716 0 : behavior,
4717 0 : ))
4718 0 : })
4719 0 : .await?;
4720 :
4721 0 : let any = results.pop().expect("we must have at least one response");
4722 :
4723 0 : let mismatching = results
4724 0 : .iter()
4725 0 : .filter(|(_, res)| res != &any.1)
4726 0 : .collect::<Vec<_>>();
4727 0 : if !mismatching.is_empty() {
4728 : // this can be hit by races which should not happen because operation lock on cplane
4729 0 : let matching = results.len() - mismatching.len();
4730 0 : tracing::error!(
4731 : matching,
4732 : compared_against=?any,
4733 : ?mismatching,
4734 0 : "shards returned different results"
4735 : );
4736 :
4737 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
4738 0 : }
4739 :
4740 0 : Ok(any.1)
4741 0 : }).await?
4742 0 : }
4743 :
4744 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
4745 0 : &self,
4746 0 : tenant_id: TenantId,
4747 0 : timeline_id: TimelineId,
4748 0 : dir: BlockUnblock,
4749 0 : ) -> Result<(), ApiError> {
4750 0 : let _tenant_lock = trace_shared_lock(
4751 0 : &self.tenant_op_locks,
4752 0 : tenant_id,
4753 0 : TenantOperations::TimelineGcBlockUnblock,
4754 0 : )
4755 0 : .await;
4756 :
4757 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4758 0 : if targets.0.is_empty() {
4759 0 : return Err(ApiError::NotFound(
4760 0 : anyhow::anyhow!("Tenant not found").into(),
4761 0 : ));
4762 0 : }
4763 :
4764 0 : async fn do_one(
4765 0 : tenant_shard_id: TenantShardId,
4766 0 : timeline_id: TimelineId,
4767 0 : node: Node,
4768 0 : http_client: reqwest::Client,
4769 0 : jwt: Option<String>,
4770 0 : dir: BlockUnblock,
4771 0 : ) -> Result<(), ApiError> {
4772 0 : let client = PageserverClient::new(
4773 0 : node.get_id(),
4774 0 : http_client,
4775 0 : node.base_url(),
4776 0 : jwt.as_deref(),
4777 : );
4778 :
4779 0 : client
4780 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
4781 0 : .await
4782 0 : .map_err(|e| passthrough_api_error(&node, e))
4783 0 : }
4784 :
4785 : // no shard needs to go first/last; the operation should be idempotent
4786 0 : let locations = targets
4787 0 : .0
4788 0 : .iter()
4789 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
4790 0 : .collect();
4791 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
4792 0 : futures::FutureExt::boxed(do_one(
4793 0 : tenant_shard_id,
4794 0 : timeline_id,
4795 0 : node,
4796 0 : self.http_client.clone(),
4797 0 : self.config.pageserver_jwt_token.clone(),
4798 0 : dir,
4799 0 : ))
4800 0 : })
4801 0 : .await
4802 0 : })
4803 0 : .await??;
4804 0 : Ok(())
4805 0 : }
4806 :
4807 0 : pub(crate) fn is_tenant_not_found_error(body: &str, tenant_id: TenantId) -> bool {
4808 0 : body.contains(&format!("tenant {tenant_id}"))
4809 0 : }
4810 :
4811 0 : fn process_result_and_passthrough_errors<T>(
4812 0 : &self,
4813 0 : tenant_id: TenantId,
4814 0 : results: Vec<(Node, Result<T, mgmt_api::Error>)>,
4815 0 : ) -> Result<Vec<(Node, T)>, ApiError> {
4816 0 : let mut processed_results: Vec<(Node, T)> = Vec::with_capacity(results.len());
4817 0 : for (node, res) in results {
4818 0 : match res {
4819 0 : Ok(res) => processed_results.push((node, res)),
4820 0 : Err(mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, body))
4821 0 : if Self::is_tenant_not_found_error(&body, tenant_id) =>
4822 : {
4823 : // If there's a tenant not found, we are still in the process of attaching the tenant.
4824 : // Return 503 so that the client can retry.
4825 0 : return Err(ApiError::ResourceUnavailable(
4826 0 : format!(
4827 0 : "Timeline is not attached to the pageserver {} yet, please retry",
4828 0 : node.get_id()
4829 0 : )
4830 0 : .into(),
4831 0 : ));
4832 : }
4833 0 : Err(e) => return Err(passthrough_api_error(&node, e)),
4834 : }
4835 : }
4836 0 : Ok(processed_results)
4837 0 : }
4838 :
4839 0 : pub(crate) async fn tenant_timeline_lsn_lease(
4840 0 : &self,
4841 0 : tenant_id: TenantId,
4842 0 : timeline_id: TimelineId,
4843 0 : lsn: Lsn,
4844 0 : ) -> Result<LsnLease, ApiError> {
4845 0 : let _tenant_lock = trace_shared_lock(
4846 0 : &self.tenant_op_locks,
4847 0 : tenant_id,
4848 0 : TenantOperations::TimelineLsnLease,
4849 0 : )
4850 0 : .await;
4851 :
4852 0 : self.tenant_remote_mutation(tenant_id, |locations| async move {
4853 0 : if locations.0.is_empty() {
4854 0 : return Err(ApiError::NotFound(
4855 0 : anyhow::anyhow!("Tenant not found").into(),
4856 0 : ));
4857 0 : }
4858 :
4859 0 : let results = self
4860 0 : .tenant_for_shards_api(
4861 0 : locations
4862 0 : .0
4863 0 : .iter()
4864 0 : .map(|(tenant_shard_id, ShardMutationLocations { latest, .. })| {
4865 0 : (*tenant_shard_id, latest.node.clone())
4866 0 : })
4867 0 : .collect(),
4868 0 : |tenant_shard_id, client| async move {
4869 0 : client
4870 0 : .timeline_lease_lsn(tenant_shard_id, timeline_id, lsn)
4871 0 : .await
4872 0 : },
4873 : 1,
4874 : 1,
4875 : SHORT_RECONCILE_TIMEOUT,
4876 0 : &self.cancel,
4877 : )
4878 0 : .await;
4879 :
4880 0 : let leases = self.process_result_and_passthrough_errors(tenant_id, results)?;
4881 0 : let mut valid_until = None;
4882 0 : for (_, lease) in leases {
4883 0 : if let Some(ref mut valid_until) = valid_until {
4884 0 : *valid_until = std::cmp::min(*valid_until, lease.valid_until);
4885 0 : } else {
4886 0 : valid_until = Some(lease.valid_until);
4887 0 : }
4888 : }
4889 0 : Ok(LsnLease {
4890 0 : valid_until: valid_until.unwrap_or_else(SystemTime::now),
4891 0 : })
4892 0 : })
4893 0 : .await?
4894 0 : }
4895 :
4896 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4897 0 : &self,
4898 0 : tenant_shard_id: TenantShardId,
4899 0 : timeline_id: TimelineId,
4900 0 : concurrency: Option<usize>,
4901 0 : recurse: bool,
4902 0 : ) -> Result<(), ApiError> {
4903 0 : let _tenant_lock = trace_shared_lock(
4904 0 : &self.tenant_op_locks,
4905 0 : tenant_shard_id.tenant_id,
4906 0 : TenantOperations::DownloadHeatmapLayers,
4907 0 : )
4908 0 : .await;
4909 :
4910 0 : let targets = {
4911 0 : let locked = self.inner.read().unwrap();
4912 0 : let mut targets = Vec::new();
4913 :
4914 : // If the request got an unsharded tenant id, then apply
4915 : // the operation to all shards. Otherwise, apply it to a specific shard.
4916 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4917 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4918 : } else {
4919 0 : tenant_shard_id.range()
4920 : };
4921 :
4922 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4923 0 : if let Some(node_id) = shard.intent.get_attached() {
4924 0 : let node = locked
4925 0 : .nodes
4926 0 : .get(node_id)
4927 0 : .expect("Pageservers may not be deleted while referenced");
4928 0 :
4929 0 : targets.push((*tenant_shard_id, node.clone()));
4930 0 : }
4931 : }
4932 0 : targets
4933 : };
4934 :
4935 0 : self.tenant_for_shards_api(
4936 0 : targets,
4937 0 : |tenant_shard_id, client| async move {
4938 0 : client
4939 0 : .timeline_download_heatmap_layers(
4940 0 : tenant_shard_id,
4941 0 : timeline_id,
4942 0 : concurrency,
4943 0 : recurse,
4944 0 : )
4945 0 : .await
4946 0 : },
4947 : 1,
4948 : 1,
4949 : SHORT_RECONCILE_TIMEOUT,
4950 0 : &self.cancel,
4951 : )
4952 0 : .await;
4953 :
4954 0 : Ok(())
4955 0 : }
4956 :
4957 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4958 : ///
4959 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`
4960 : /// and returned element at index `i` is the result for `req_fn(op(locations[i])`.
4961 0 : async fn tenant_for_shards<F, R>(
4962 0 : &self,
4963 0 : locations: Vec<(TenantShardId, Node)>,
4964 0 : mut req_fn: F,
4965 0 : ) -> Result<Vec<R>, ApiError>
4966 0 : where
4967 0 : F: FnMut(
4968 0 : TenantShardId,
4969 0 : Node,
4970 0 : )
4971 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4972 0 : {
4973 0 : let mut futs = FuturesUnordered::new();
4974 0 : let mut results = Vec::with_capacity(locations.len());
4975 :
4976 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4977 0 : let fut = req_fn(tenant_shard_id, node);
4978 0 : futs.push(async move { (idx, fut.await) });
4979 : }
4980 :
4981 0 : while let Some((idx, r)) = futs.next().await {
4982 0 : results.push((idx, r?));
4983 : }
4984 :
4985 0 : results.sort_by_key(|(idx, _)| *idx);
4986 0 : Ok(results.into_iter().map(|(_, r)| r).collect())
4987 0 : }
4988 :
4989 : /// Concurrently invoke a pageserver API call on many shards at once.
4990 : ///
4991 : /// The returned Vec has the same length as the `locations` Vec,
4992 : /// and returned element at index `i` is the result for `op(locations[i])`.
4993 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
4994 0 : &self,
4995 0 : locations: Vec<(TenantShardId, Node)>,
4996 0 : op: O,
4997 0 : warn_threshold: u32,
4998 0 : max_retries: u32,
4999 0 : timeout: Duration,
5000 0 : cancel: &CancellationToken,
5001 0 : ) -> Vec<(Node, mgmt_api::Result<T>)>
5002 0 : where
5003 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
5004 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
5005 0 : {
5006 0 : let mut futs = FuturesUnordered::new();
5007 0 : let mut results = Vec::with_capacity(locations.len());
5008 :
5009 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
5010 0 : futs.push(async move {
5011 0 : let r = node
5012 0 : .with_client_retries(
5013 0 : |client| op(tenant_shard_id, client),
5014 0 : &self.http_client,
5015 0 : &self.config.pageserver_jwt_token,
5016 0 : warn_threshold,
5017 0 : max_retries,
5018 0 : timeout,
5019 0 : cancel,
5020 : )
5021 0 : .await;
5022 0 : (idx, node, r)
5023 0 : });
5024 : }
5025 :
5026 0 : while let Some((idx, node, r)) = futs.next().await {
5027 0 : results.push((idx, node, r.unwrap_or(Err(mgmt_api::Error::Cancelled))));
5028 0 : }
5029 :
5030 0 : results.sort_by_key(|(idx, _, _)| *idx);
5031 0 : results.into_iter().map(|(_, node, r)| (node, r)).collect()
5032 0 : }
5033 :
5034 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
5035 : /// when creating and deleting timelines:
5036 : /// - Makes sure shards are attached somewhere if they weren't already
5037 : /// - Looks up the shards and the nodes where they were most recently attached
5038 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
5039 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
5040 0 : pub(crate) async fn tenant_remote_mutation<R, O, F>(
5041 0 : &self,
5042 0 : tenant_id: TenantId,
5043 0 : op: O,
5044 0 : ) -> Result<R, ApiError>
5045 0 : where
5046 0 : O: FnOnce(TenantMutationLocations) -> F,
5047 0 : F: std::future::Future<Output = R>,
5048 0 : {
5049 0 : self.tenant_remote_mutation_inner(TenantIdOrShardId::TenantId(tenant_id), op)
5050 0 : .await
5051 0 : }
5052 :
5053 0 : pub(crate) async fn tenant_shard_remote_mutation<R, O, F>(
5054 0 : &self,
5055 0 : tenant_shard_id: TenantShardId,
5056 0 : op: O,
5057 0 : ) -> Result<R, ApiError>
5058 0 : where
5059 0 : O: FnOnce(TenantMutationLocations) -> F,
5060 0 : F: std::future::Future<Output = R>,
5061 0 : {
5062 0 : self.tenant_remote_mutation_inner(TenantIdOrShardId::TenantShardId(tenant_shard_id), op)
5063 0 : .await
5064 0 : }
5065 :
5066 0 : async fn tenant_remote_mutation_inner<R, O, F>(
5067 0 : &self,
5068 0 : tenant_id_or_shard_id: TenantIdOrShardId,
5069 0 : op: O,
5070 0 : ) -> Result<R, ApiError>
5071 0 : where
5072 0 : O: FnOnce(TenantMutationLocations) -> F,
5073 0 : F: std::future::Future<Output = R>,
5074 0 : {
5075 0 : let mutation_locations = {
5076 0 : let mut locations = TenantMutationLocations::default();
5077 :
5078 : // Load the currently attached pageservers for the latest generation of each shard. This can
5079 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
5080 : // will still be the latest when we're done: we will check generations again at the end of
5081 : // this function to handle that.
5082 0 : let generations = self
5083 0 : .persistence
5084 0 : .tenant_generations(tenant_id_or_shard_id.tenant_id())
5085 0 : .await?
5086 0 : .into_iter()
5087 0 : .filter(|i| tenant_id_or_shard_id.matches(&i.tenant_shard_id))
5088 0 : .collect::<Vec<_>>();
5089 :
5090 0 : if generations
5091 0 : .iter()
5092 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
5093 : {
5094 0 : let shard_generations = generations
5095 0 : .into_iter()
5096 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
5097 0 : .collect::<HashMap<_, _>>();
5098 :
5099 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
5100 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
5101 0 : let locked = self.inner.read().unwrap();
5102 0 : let tenant_shards = locked
5103 0 : .tenants
5104 0 : .range(TenantShardId::tenant_range(
5105 0 : tenant_id_or_shard_id.tenant_id(),
5106 : ))
5107 0 : .filter(|(shard_id, _)| tenant_id_or_shard_id.matches(shard_id))
5108 0 : .collect::<Vec<_>>();
5109 0 : for (shard_id, shard) in tenant_shards {
5110 0 : match shard.policy {
5111 : PlacementPolicy::Attached(_) => {
5112 : // This shard is meant to be attached: the caller is not wrong to try and
5113 : // use this function, but we can't service the request right now.
5114 0 : let Some(generation) = shard_generations.get(shard_id) else {
5115 : // This can only happen if there is a split brain controller modifying the database. This should
5116 : // never happen when testing, and if it happens in production we can only log the issue.
5117 0 : debug_assert!(false);
5118 0 : tracing::error!(
5119 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
5120 : );
5121 0 : continue;
5122 : };
5123 0 : let (generation, generation_pageserver) = generation;
5124 0 : if let Some(generation) = generation {
5125 0 : if generation_pageserver.is_none() {
5126 : // This is legitimate only in a very narrow window where the shard was only just configured into
5127 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
5128 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
5129 0 : tracing::warn!(
5130 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
5131 : );
5132 0 : }
5133 : } else {
5134 : // This should never happen: a shard with no generation is only permitted when it was created in some state
5135 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
5136 0 : debug_assert!(false);
5137 0 : tracing::error!(
5138 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
5139 : );
5140 0 : continue;
5141 : }
5142 : }
5143 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
5144 0 : return Err(ApiError::Conflict(format!(
5145 0 : "Shard {shard_id} tenant has policy {:?}",
5146 0 : shard.policy
5147 0 : )));
5148 : }
5149 : }
5150 : }
5151 :
5152 0 : return Err(ApiError::ResourceUnavailable(
5153 0 : "One or more shards in tenant is not yet attached".into(),
5154 0 : ));
5155 0 : }
5156 :
5157 0 : let locked = self.inner.read().unwrap();
5158 : for ShardGenerationState {
5159 0 : tenant_shard_id,
5160 0 : generation,
5161 0 : generation_pageserver,
5162 0 : } in generations
5163 : {
5164 0 : let node_id = generation_pageserver.expect("We checked for None above");
5165 0 : let node = locked
5166 0 : .nodes
5167 0 : .get(&node_id)
5168 0 : .ok_or(ApiError::Conflict(format!(
5169 0 : "Raced with removal of node {node_id}"
5170 0 : )))?;
5171 0 : let generation = generation.expect("Checked above");
5172 :
5173 0 : let tenant = locked.tenants.get(&tenant_shard_id);
5174 :
5175 : // TODO(vlad): Abstract the logic that finds stale attached locations
5176 : // from observed state into a [`Service`] method.
5177 0 : let other_locations = match tenant {
5178 0 : Some(tenant) => {
5179 0 : let mut other = tenant.attached_locations();
5180 0 : let latest_location_index =
5181 0 : other.iter().position(|&l| l == (node.get_id(), generation));
5182 0 : if let Some(idx) = latest_location_index {
5183 0 : other.remove(idx);
5184 0 : }
5185 :
5186 0 : other
5187 : }
5188 0 : None => Vec::default(),
5189 : };
5190 :
5191 0 : let location = ShardMutationLocations {
5192 0 : latest: MutationLocation {
5193 0 : node: node.clone(),
5194 0 : generation,
5195 0 : },
5196 0 : other: other_locations
5197 0 : .into_iter()
5198 0 : .filter_map(|(node_id, generation)| {
5199 0 : let node = locked.nodes.get(&node_id)?;
5200 :
5201 0 : Some(MutationLocation {
5202 0 : node: node.clone(),
5203 0 : generation,
5204 0 : })
5205 0 : })
5206 0 : .collect(),
5207 : };
5208 0 : locations.0.insert(tenant_shard_id, location);
5209 : }
5210 :
5211 0 : locations
5212 : };
5213 :
5214 0 : let result = op(mutation_locations.clone()).await;
5215 :
5216 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
5217 : // our remote operation executed on the latest generation and is therefore persistent.
5218 : {
5219 0 : let latest_generations = self
5220 0 : .persistence
5221 0 : .tenant_generations(tenant_id_or_shard_id.tenant_id())
5222 0 : .await?
5223 0 : .into_iter()
5224 0 : .filter(|i| tenant_id_or_shard_id.matches(&i.tenant_shard_id))
5225 0 : .collect::<Vec<_>>();
5226 :
5227 0 : if latest_generations
5228 0 : .into_iter()
5229 0 : .map(
5230 : |ShardGenerationState {
5231 : tenant_shard_id,
5232 : generation,
5233 : generation_pageserver: _,
5234 0 : }| (tenant_shard_id, generation),
5235 : )
5236 0 : .collect::<Vec<_>>()
5237 0 : != mutation_locations
5238 0 : .0
5239 0 : .into_iter()
5240 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
5241 0 : .collect::<Vec<_>>()
5242 : {
5243 : // We raced with something that incremented the generation, and therefore cannot be
5244 : // confident that our actions are persistent (they might have hit an old generation).
5245 : //
5246 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
5247 0 : return Err(ApiError::ResourceUnavailable(
5248 0 : "Tenant attachment changed, please retry".into(),
5249 0 : ));
5250 0 : }
5251 : }
5252 :
5253 0 : Ok(result)
5254 0 : }
5255 :
5256 0 : pub(crate) async fn tenant_timeline_delete(
5257 0 : self: &Arc<Self>,
5258 0 : tenant_id: TenantId,
5259 0 : timeline_id: TimelineId,
5260 0 : ) -> Result<StatusCode, ApiError> {
5261 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
5262 0 : let _tenant_lock = trace_shared_lock(
5263 0 : &self.tenant_op_locks,
5264 0 : tenant_id,
5265 0 : TenantOperations::TimelineDelete,
5266 0 : )
5267 0 : .await;
5268 :
5269 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
5270 0 : if targets.0.is_empty() {
5271 0 : return Err(ApiError::NotFound(
5272 0 : anyhow::anyhow!("Tenant not found").into(),
5273 0 : ));
5274 0 : }
5275 :
5276 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
5277 0 : assert!(shard_zero_tid.is_shard_zero());
5278 :
5279 0 : async fn delete_one(
5280 0 : tenant_shard_id: TenantShardId,
5281 0 : timeline_id: TimelineId,
5282 0 : node: Node,
5283 0 : http_client: reqwest::Client,
5284 0 : jwt: Option<String>,
5285 0 : ) -> Result<StatusCode, ApiError> {
5286 0 : tracing::info!(
5287 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
5288 : );
5289 :
5290 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
5291 0 : let res = client
5292 0 : .timeline_delete(tenant_shard_id, timeline_id)
5293 0 : .await;
5294 :
5295 0 : match res {
5296 0 : Ok(ok) => Ok(ok),
5297 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
5298 0 : Err(mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg)) if msg.contains("Requested tenant is missing") => {
5299 0 : Err(ApiError::ResourceUnavailable("Tenant migration in progress".into()))
5300 : },
5301 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
5302 0 : Err(e) => {
5303 0 : Err(
5304 0 : ApiError::InternalServerError(anyhow::anyhow!(
5305 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
5306 0 : ))
5307 0 : )
5308 : }
5309 : }
5310 0 : }
5311 :
5312 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
5313 0 : let statuses = self
5314 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
5315 0 : Box::pin(delete_one(
5316 0 : tenant_shard_id,
5317 0 : timeline_id,
5318 0 : node,
5319 0 : self.http_client.clone(),
5320 0 : self.config.pageserver_jwt_token.clone(),
5321 0 : ))
5322 0 : })
5323 0 : .await?;
5324 :
5325 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
5326 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
5327 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
5328 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
5329 0 : return Ok(StatusCode::CONFLICT);
5330 0 : }
5331 :
5332 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
5333 0 : return Ok(StatusCode::ACCEPTED);
5334 0 : }
5335 :
5336 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
5337 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
5338 0 : let shard_zero_status = delete_one(
5339 0 : shard_zero_tid,
5340 0 : timeline_id,
5341 0 : shard_zero_locations.latest.node,
5342 0 : self.http_client.clone(),
5343 0 : self.config.pageserver_jwt_token.clone(),
5344 0 : )
5345 0 : .await?;
5346 0 : Ok(shard_zero_status)
5347 0 : }).await?;
5348 :
5349 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
5350 0 : .await?;
5351 :
5352 0 : status_code
5353 0 : }
5354 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
5355 : ///
5356 : /// Returns the node, tenant shard id, and whether it is consistent with the observed state.
5357 0 : pub(crate) async fn tenant_shard0_node(
5358 0 : &self,
5359 0 : tenant_id: TenantId,
5360 0 : ) -> Result<(Node, TenantShardId), ApiError> {
5361 0 : let tenant_shard_id = {
5362 0 : let locked = self.inner.read().unwrap();
5363 0 : let Some((tenant_shard_id, _shard)) = locked
5364 0 : .tenants
5365 0 : .range(TenantShardId::tenant_range(tenant_id))
5366 0 : .next()
5367 : else {
5368 0 : return Err(ApiError::NotFound(
5369 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
5370 0 : ));
5371 : };
5372 :
5373 0 : *tenant_shard_id
5374 : };
5375 :
5376 0 : self.tenant_shard_node(tenant_shard_id)
5377 0 : .await
5378 0 : .map(|node| (node, tenant_shard_id))
5379 0 : }
5380 :
5381 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
5382 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
5383 : ///
5384 : /// Returns the intent node and whether it is consistent with the observed state.
5385 0 : pub(crate) async fn tenant_shard_node(
5386 0 : &self,
5387 0 : tenant_shard_id: TenantShardId,
5388 0 : ) -> Result<Node, ApiError> {
5389 : // Look up in-memory state and maybe use the node from there.
5390 : {
5391 0 : let locked = self.inner.read().unwrap();
5392 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5393 0 : return Err(ApiError::NotFound(
5394 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
5395 0 : ));
5396 : };
5397 :
5398 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
5399 0 : tracing::warn!(
5400 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
5401 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
5402 : shard.policy
5403 : );
5404 0 : return Err(ApiError::Conflict(
5405 0 : "Cannot call timeline API on non-attached tenant".to_string(),
5406 0 : ));
5407 : };
5408 :
5409 0 : if shard.reconciler.is_none() {
5410 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
5411 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
5412 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
5413 : // This should never happen
5414 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5415 0 : "Shard refers to nonexistent node"
5416 0 : )));
5417 : };
5418 0 : return Ok(node.clone());
5419 0 : }
5420 : };
5421 :
5422 : // Look up the latest attached pageserver location from the database
5423 : // generation state: this will reflect the progress of any ongoing migration.
5424 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
5425 : // the case where they call through to the pageserver and get a 404.
5426 0 : let db_result = self
5427 0 : .persistence
5428 0 : .tenant_generations(tenant_shard_id.tenant_id)
5429 0 : .await?;
5430 : let Some(ShardGenerationState {
5431 : tenant_shard_id: _,
5432 : generation: _,
5433 0 : generation_pageserver: Some(node_id),
5434 0 : }) = db_result
5435 0 : .into_iter()
5436 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
5437 : else {
5438 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
5439 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
5440 : // or a conflict response (case where tenant was detached in background)
5441 0 : return Err(ApiError::ResourceUnavailable(
5442 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
5443 0 : ));
5444 : };
5445 0 : let locked = self.inner.read().unwrap();
5446 0 : let Some(node) = locked.nodes.get(&node_id) else {
5447 : // This should never happen
5448 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5449 0 : "Shard refers to nonexistent node"
5450 0 : )));
5451 : };
5452 : // As a reconciliation is in flight, we do not have the observed state yet, and therefore we assume it is always inconsistent.
5453 0 : Ok(node.clone())
5454 0 : }
5455 :
5456 0 : pub(crate) fn tenant_locate(
5457 0 : &self,
5458 0 : tenant_id: TenantId,
5459 0 : ) -> Result<TenantLocateResponse, ApiError> {
5460 0 : let locked = self.inner.read().unwrap();
5461 0 : tracing::info!("Locating shards for tenant {tenant_id}");
5462 :
5463 0 : let mut result = Vec::new();
5464 0 : let mut shard_params: Option<ShardParameters> = None;
5465 :
5466 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5467 : {
5468 0 : let node_id =
5469 0 : shard
5470 0 : .intent
5471 0 : .get_attached()
5472 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
5473 0 : "Cannot locate a tenant that is not attached"
5474 0 : )))?;
5475 :
5476 0 : let node = locked
5477 0 : .nodes
5478 0 : .get(&node_id)
5479 0 : .expect("Pageservers may not be deleted while referenced");
5480 :
5481 0 : result.push(node.shard_location(*tenant_shard_id));
5482 :
5483 0 : match &shard_params {
5484 0 : None => {
5485 0 : shard_params = Some(ShardParameters {
5486 0 : stripe_size: shard.shard.stripe_size,
5487 0 : count: shard.shard.count,
5488 0 : });
5489 0 : }
5490 0 : Some(params) => {
5491 0 : if params.stripe_size != shard.shard.stripe_size {
5492 : // This should never happen. We enforce at runtime because it's simpler than
5493 : // adding an extra per-tenant data structure to store the things that should be the same
5494 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5495 0 : "Inconsistent shard stripe size parameters!"
5496 0 : )));
5497 0 : }
5498 : }
5499 : }
5500 : }
5501 :
5502 0 : if result.is_empty() {
5503 0 : return Err(ApiError::NotFound(
5504 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
5505 0 : ));
5506 0 : }
5507 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
5508 0 : tracing::info!(
5509 0 : "Located tenant {} with params {:?} on shards {}",
5510 : tenant_id,
5511 : shard_params,
5512 0 : result
5513 0 : .iter()
5514 0 : .map(|s| format!("{s:?}"))
5515 0 : .collect::<Vec<_>>()
5516 0 : .join(",")
5517 : );
5518 :
5519 0 : Ok(TenantLocateResponse {
5520 0 : shards: result,
5521 0 : shard_params,
5522 0 : })
5523 0 : }
5524 :
5525 : /// Returns None if the input iterator of shards does not include a shard with number=0
5526 0 : fn tenant_describe_impl<'a>(
5527 0 : &self,
5528 0 : shards: impl Iterator<Item = &'a TenantShard>,
5529 0 : ) -> Option<TenantDescribeResponse> {
5530 0 : let mut shard_zero = None;
5531 0 : let mut describe_shards = Vec::new();
5532 :
5533 0 : for shard in shards {
5534 0 : if shard.tenant_shard_id.is_shard_zero() {
5535 0 : shard_zero = Some(shard);
5536 0 : }
5537 :
5538 0 : describe_shards.push(TenantDescribeResponseShard {
5539 0 : tenant_shard_id: shard.tenant_shard_id,
5540 0 : node_attached: *shard.intent.get_attached(),
5541 0 : node_secondary: shard.intent.get_secondary().to_vec(),
5542 0 : last_error: shard
5543 0 : .last_error
5544 0 : .lock()
5545 0 : .unwrap()
5546 0 : .as_ref()
5547 0 : .map(|e| format!("{e}"))
5548 0 : .unwrap_or("".to_string())
5549 0 : .clone(),
5550 0 : is_reconciling: shard.reconciler.is_some(),
5551 0 : is_pending_compute_notification: shard.pending_compute_notification,
5552 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
5553 0 : is_importing: shard.importing == TimelineImportState::Importing,
5554 0 : scheduling_policy: shard.get_scheduling_policy(),
5555 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
5556 : })
5557 : }
5558 :
5559 0 : let shard_zero = shard_zero?;
5560 :
5561 0 : Some(TenantDescribeResponse {
5562 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
5563 0 : shards: describe_shards,
5564 0 : stripe_size: shard_zero.shard.stripe_size,
5565 0 : policy: shard_zero.policy.clone(),
5566 0 : config: shard_zero.config.clone(),
5567 0 : })
5568 0 : }
5569 :
5570 0 : pub(crate) fn tenant_describe(
5571 0 : &self,
5572 0 : tenant_id: TenantId,
5573 0 : ) -> Result<TenantDescribeResponse, ApiError> {
5574 0 : let locked = self.inner.read().unwrap();
5575 :
5576 0 : self.tenant_describe_impl(
5577 0 : locked
5578 0 : .tenants
5579 0 : .range(TenantShardId::tenant_range(tenant_id))
5580 0 : .map(|(_k, v)| v),
5581 : )
5582 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
5583 0 : }
5584 :
5585 : /* BEGIN_HADRON */
5586 0 : pub(crate) async fn tenant_timeline_describe(
5587 0 : &self,
5588 0 : tenant_id: TenantId,
5589 0 : timeline_id: TimelineId,
5590 0 : ) -> Result<TenantTimelineDescribeResponse, ApiError> {
5591 0 : self.tenant_remote_mutation(tenant_id, |locations| async move {
5592 0 : if locations.0.is_empty() {
5593 0 : return Err(ApiError::NotFound(
5594 0 : anyhow::anyhow!("Tenant not found").into(),
5595 0 : ));
5596 0 : };
5597 :
5598 0 : let locations: Vec<(TenantShardId, Node)> = locations
5599 0 : .0
5600 0 : .iter()
5601 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
5602 0 : .collect();
5603 0 : let mut futs = FuturesUnordered::new();
5604 :
5605 0 : for (shard_id, node) in locations {
5606 0 : futs.push({
5607 0 : async move {
5608 0 : let result = node
5609 0 : .with_client_retries(
5610 0 : |client| async move {
5611 0 : client
5612 0 : .tenant_timeline_describe(&shard_id, &timeline_id)
5613 0 : .await
5614 0 : },
5615 0 : &self.http_client,
5616 0 : &self.config.pageserver_jwt_token,
5617 : 3,
5618 : 3,
5619 0 : Duration::from_secs(30),
5620 0 : &self.cancel,
5621 : )
5622 0 : .await;
5623 0 : (result, shard_id, node.get_id())
5624 0 : }
5625 : });
5626 : }
5627 :
5628 0 : let mut results: Vec<TimelineInfo> = Vec::new();
5629 0 : while let Some((result, tenant_shard_id, node_id)) = futs.next().await {
5630 0 : match result {
5631 0 : Some(Ok(timeline_info)) => results.push(timeline_info),
5632 0 : Some(Err(e)) => {
5633 0 : tracing::warn!(
5634 0 : "Failed to describe tenant {} timeline {} for pageserver {}: {e}",
5635 : tenant_shard_id,
5636 : timeline_id,
5637 : node_id,
5638 : );
5639 0 : return Err(ApiError::ResourceUnavailable(format!("{e}").into()));
5640 : }
5641 0 : None => return Err(ApiError::Cancelled),
5642 : }
5643 : }
5644 0 : let mut image_consistent_lsn: Option<Lsn> = Some(Lsn::MAX);
5645 0 : for timeline_info in &results {
5646 0 : if let Some(tline_image_consistent_lsn) = timeline_info.image_consistent_lsn {
5647 0 : image_consistent_lsn = Some(std::cmp::min(
5648 0 : image_consistent_lsn.unwrap(),
5649 0 : tline_image_consistent_lsn,
5650 0 : ));
5651 0 : } else {
5652 0 : tracing::warn!(
5653 0 : "Timeline {} on shard {} does not have image consistent lsn",
5654 : timeline_info.timeline_id,
5655 : timeline_info.tenant_id
5656 : );
5657 0 : image_consistent_lsn = None;
5658 0 : break;
5659 : }
5660 : }
5661 :
5662 0 : Ok(TenantTimelineDescribeResponse {
5663 0 : shards: results,
5664 0 : image_consistent_lsn,
5665 0 : })
5666 0 : })
5667 0 : .await?
5668 0 : }
5669 : /* END_HADRON */
5670 :
5671 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
5672 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
5673 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
5674 : /// in our external API.
5675 0 : pub(crate) fn tenant_list(
5676 0 : &self,
5677 0 : limit: Option<usize>,
5678 0 : start_after: Option<TenantId>,
5679 0 : ) -> Vec<TenantDescribeResponse> {
5680 0 : let locked = self.inner.read().unwrap();
5681 :
5682 : // Apply start_from parameter
5683 0 : let shard_range = match start_after {
5684 0 : None => locked.tenants.range(..),
5685 0 : Some(tenant_id) => locked.tenants.range(
5686 0 : TenantShardId {
5687 0 : tenant_id,
5688 0 : shard_number: ShardNumber(u8::MAX),
5689 0 : shard_count: ShardCount(u8::MAX),
5690 0 : }..,
5691 : ),
5692 : };
5693 :
5694 0 : let mut result = Vec::new();
5695 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
5696 0 : result.push(
5697 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
5698 0 : .expect("Groups are always non-empty"),
5699 : );
5700 :
5701 : // Enforce `limit` parameter
5702 0 : if let Some(limit) = limit {
5703 0 : if result.len() >= limit {
5704 0 : break;
5705 0 : }
5706 0 : }
5707 : }
5708 :
5709 0 : result
5710 0 : }
5711 :
5712 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
5713 : async fn abort_tenant_shard_split(
5714 : &self,
5715 : op: &TenantShardSplitAbort,
5716 : ) -> Result<(), TenantShardSplitAbortError> {
5717 : // Cleaning up a split:
5718 : // - Parent shards are not destroyed during a split, just detached.
5719 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
5720 : // just the children attached, or both.
5721 : //
5722 : // Therefore our work to do is to:
5723 : // 1. Clean up storage controller's internal state to just refer to parents, no children
5724 : // 2. Call out to pageservers to ensure that children are detached
5725 : // 3. Call out to pageservers to ensure that parents are attached.
5726 : //
5727 : // Crash safety:
5728 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
5729 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
5730 : // and detach them.
5731 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
5732 : // from our database, then we will re-enter this cleanup routine on startup.
5733 :
5734 : let TenantShardSplitAbort {
5735 : tenant_id,
5736 : new_shard_count,
5737 : new_stripe_size,
5738 : ..
5739 : } = op;
5740 :
5741 : // First abort persistent state, if any exists.
5742 : match self
5743 : .persistence
5744 : .abort_shard_split(*tenant_id, *new_shard_count)
5745 : .await?
5746 : {
5747 : AbortShardSplitStatus::Aborted => {
5748 : // Proceed to roll back any child shards created on pageservers
5749 : }
5750 : AbortShardSplitStatus::Complete => {
5751 : // The split completed (we might hit that path if e.g. our database transaction
5752 : // to write the completion landed in the database, but we dropped connection
5753 : // before seeing the result).
5754 : //
5755 : // We must update in-memory state to reflect the successful split.
5756 : self.tenant_shard_split_commit_inmem(
5757 : *tenant_id,
5758 : *new_shard_count,
5759 : *new_stripe_size,
5760 : );
5761 : return Ok(());
5762 : }
5763 : }
5764 :
5765 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
5766 : let detach_locations: Vec<(Node, TenantShardId)> = {
5767 : let mut detach_locations = Vec::new();
5768 : let mut locked = self.inner.write().unwrap();
5769 : let (nodes, tenants, scheduler) = locked.parts_mut();
5770 :
5771 : for (tenant_shard_id, shard) in
5772 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
5773 : {
5774 : if shard.shard.count == op.new_shard_count {
5775 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
5776 : // is infallible, so if we got an error we shouldn't have got that far.
5777 : tracing::warn!(
5778 : "During split abort, child shard {tenant_shard_id} found in-memory"
5779 : );
5780 : continue;
5781 : }
5782 :
5783 : // Add the children of this shard to this list of things to detach
5784 : if let Some(node_id) = shard.intent.get_attached() {
5785 : for child_id in tenant_shard_id.split(*new_shard_count) {
5786 : detach_locations.push((
5787 : nodes
5788 : .get(node_id)
5789 : .expect("Intent references nonexistent node")
5790 : .clone(),
5791 : child_id,
5792 : ));
5793 : }
5794 : } else {
5795 : tracing::warn!(
5796 : "During split abort, shard {tenant_shard_id} has no attached location"
5797 : );
5798 : }
5799 :
5800 : tracing::info!("Restoring parent shard {tenant_shard_id}");
5801 :
5802 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
5803 : // if the original attachment location is offline.
5804 : if let Some(node_id) = shard.intent.get_attached() {
5805 : if !nodes.get(node_id).unwrap().is_available() {
5806 : tracing::info!(
5807 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
5808 : );
5809 : shard.intent.demote_attached(scheduler, *node_id);
5810 : }
5811 : }
5812 : for node_id in shard.intent.get_secondary().clone() {
5813 : if !nodes.get(&node_id).unwrap().is_available() {
5814 : tracing::info!(
5815 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
5816 : );
5817 : shard.intent.remove_secondary(scheduler, node_id);
5818 : }
5819 : }
5820 :
5821 : shard.splitting = SplitState::Idle;
5822 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
5823 : // If this shard can't be scheduled now (perhaps due to offline nodes or
5824 : // capacity issues), that must not prevent us rolling back a split. In this
5825 : // case it should be eventually scheduled in the background.
5826 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
5827 : }
5828 :
5829 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
5830 : }
5831 :
5832 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
5833 : tenants
5834 0 : .retain(|id, s| !(id.tenant_id == *tenant_id && s.shard.count == *new_shard_count));
5835 :
5836 : detach_locations
5837 : };
5838 :
5839 : for (node, child_id) in detach_locations {
5840 : if !node.is_available() {
5841 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
5842 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
5843 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
5844 : // them from the node.
5845 : tracing::warn!(
5846 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
5847 : );
5848 : continue;
5849 : }
5850 :
5851 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
5852 : // a 503 and retry, up to our limit.
5853 : tracing::info!("Detaching {child_id} on {node}...");
5854 : match node
5855 : .with_client_retries(
5856 0 : |client| async move {
5857 0 : let config = LocationConfig {
5858 0 : mode: LocationConfigMode::Detached,
5859 0 : generation: None,
5860 0 : secondary_conf: None,
5861 0 : shard_number: child_id.shard_number.0,
5862 0 : shard_count: child_id.shard_count.literal(),
5863 0 : // Stripe size and tenant config don't matter when detaching
5864 0 : shard_stripe_size: 0,
5865 0 : tenant_conf: TenantConfig::default(),
5866 0 : };
5867 :
5868 0 : client.location_config(child_id, config, None, false).await
5869 0 : },
5870 : &self.http_client,
5871 : &self.config.pageserver_jwt_token,
5872 : 1,
5873 : 10,
5874 : Duration::from_secs(5),
5875 : &self.reconcilers_cancel,
5876 : )
5877 : .await
5878 : {
5879 : Some(Ok(_)) => {}
5880 : Some(Err(e)) => {
5881 : // We failed to communicate with the remote node. This is problematic: we may be
5882 : // leaving it with a rogue child shard.
5883 : tracing::warn!(
5884 : "Failed to detach child {child_id} from node {node} during abort"
5885 : );
5886 : return Err(e.into());
5887 : }
5888 : None => {
5889 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
5890 : // clean up on restart. The node going offline requires a retry.
5891 : return Err(TenantShardSplitAbortError::Unavailable);
5892 : }
5893 : };
5894 : }
5895 :
5896 : tracing::info!("Successfully aborted split");
5897 : Ok(())
5898 : }
5899 :
5900 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
5901 : /// of the tenant map to reflect the child shards that exist after the split.
5902 0 : fn tenant_shard_split_commit_inmem(
5903 0 : &self,
5904 0 : tenant_id: TenantId,
5905 0 : new_shard_count: ShardCount,
5906 0 : new_stripe_size: Option<ShardStripeSize>,
5907 0 : ) -> (
5908 0 : TenantShardSplitResponse,
5909 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
5910 0 : Vec<ReconcilerWaiter>,
5911 0 : ) {
5912 0 : let mut response = TenantShardSplitResponse {
5913 0 : new_shards: Vec::new(),
5914 0 : };
5915 0 : let mut child_locations = Vec::new();
5916 0 : let mut waiters = Vec::new();
5917 :
5918 : {
5919 0 : let mut locked = self.inner.write().unwrap();
5920 :
5921 0 : let parent_ids = locked
5922 0 : .tenants
5923 0 : .range(TenantShardId::tenant_range(tenant_id))
5924 0 : .map(|(shard_id, _)| *shard_id)
5925 0 : .collect::<Vec<_>>();
5926 :
5927 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5928 0 : for parent_id in parent_ids {
5929 0 : let child_ids = parent_id.split(new_shard_count);
5930 :
5931 : let (
5932 0 : pageserver,
5933 0 : generation,
5934 0 : policy,
5935 0 : parent_ident,
5936 0 : config,
5937 0 : preferred_az,
5938 0 : secondary_count,
5939 : ) = {
5940 0 : let mut old_state = tenants
5941 0 : .remove(&parent_id)
5942 0 : .expect("It was present, we just split it");
5943 :
5944 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
5945 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
5946 : // nothing else can clear this.
5947 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
5948 :
5949 0 : let old_attached = old_state.intent.get_attached().unwrap();
5950 0 : old_state.intent.clear(scheduler);
5951 0 : let generation = old_state.generation.expect("Shard must have been attached");
5952 0 : (
5953 0 : old_attached,
5954 0 : generation,
5955 0 : old_state.policy.clone(),
5956 0 : old_state.shard,
5957 0 : old_state.config.clone(),
5958 0 : old_state.preferred_az().cloned(),
5959 0 : old_state.intent.get_secondary().len(),
5960 0 : )
5961 : };
5962 :
5963 0 : let mut schedule_context = ScheduleContext::default();
5964 0 : for child in child_ids {
5965 0 : let mut child_shard = parent_ident;
5966 0 : child_shard.number = child.shard_number;
5967 0 : child_shard.count = child.shard_count;
5968 0 : if let Some(stripe_size) = new_stripe_size {
5969 0 : child_shard.stripe_size = stripe_size;
5970 0 : }
5971 :
5972 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5973 0 : child_observed.insert(
5974 0 : pageserver,
5975 0 : ObservedStateLocation {
5976 0 : conf: Some(attached_location_conf(
5977 0 : generation,
5978 0 : &child_shard,
5979 0 : &config,
5980 0 : &policy,
5981 0 : secondary_count,
5982 0 : )),
5983 0 : },
5984 : );
5985 :
5986 0 : let mut child_state =
5987 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
5988 0 : child_state.intent =
5989 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
5990 0 : child_state.observed = ObservedState {
5991 0 : locations: child_observed,
5992 0 : };
5993 0 : child_state.generation = Some(generation);
5994 0 : child_state.config = config.clone();
5995 :
5996 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
5997 : // as at this point in the split process we have succeeded and this part is infallible:
5998 : // we will never need to do any special recovery from this state.
5999 :
6000 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
6001 :
6002 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
6003 : // This is not fatal, because we've implicitly already got an attached
6004 : // location for the child shard. Failure here just means we couldn't
6005 : // find a secondary (e.g. because cluster is overloaded).
6006 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
6007 0 : }
6008 : // In the background, attach secondary locations for the new shards
6009 0 : if let Some(waiter) = self.maybe_reconcile_shard(
6010 0 : &mut child_state,
6011 0 : nodes,
6012 0 : ReconcilerPriority::High,
6013 0 : ) {
6014 0 : waiters.push(waiter);
6015 0 : }
6016 :
6017 0 : tenants.insert(child, child_state);
6018 0 : response.new_shards.push(child);
6019 : }
6020 : }
6021 0 : (response, child_locations, waiters)
6022 : }
6023 0 : }
6024 :
6025 0 : async fn tenant_shard_split_start_secondaries(
6026 0 : &self,
6027 0 : tenant_id: TenantId,
6028 0 : waiters: Vec<ReconcilerWaiter>,
6029 0 : ) {
6030 : // Wait for initial reconcile of child shards, this creates the secondary locations
6031 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
6032 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
6033 : // their secondaries couldn't be attached.
6034 0 : tracing::warn!("Failed to reconcile after split: {e}");
6035 0 : return;
6036 0 : }
6037 :
6038 : // Take the state lock to discover the attached & secondary intents for all shards
6039 0 : let (attached, secondary) = {
6040 0 : let locked = self.inner.read().unwrap();
6041 0 : let mut attached = Vec::new();
6042 0 : let mut secondary = Vec::new();
6043 :
6044 0 : for (tenant_shard_id, shard) in
6045 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
6046 : {
6047 0 : let Some(node_id) = shard.intent.get_attached() else {
6048 : // Unexpected. Race with a PlacementPolicy change?
6049 0 : tracing::warn!(
6050 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
6051 : );
6052 0 : continue;
6053 : };
6054 :
6055 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
6056 : // No secondary location. Nothing for us to do.
6057 0 : continue;
6058 : };
6059 :
6060 0 : let attached_node = locked
6061 0 : .nodes
6062 0 : .get(node_id)
6063 0 : .expect("Pageservers may not be deleted while referenced");
6064 :
6065 0 : let secondary_node = locked
6066 0 : .nodes
6067 0 : .get(secondary_node_id)
6068 0 : .expect("Pageservers may not be deleted while referenced");
6069 :
6070 0 : attached.push((*tenant_shard_id, attached_node.clone()));
6071 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
6072 : }
6073 0 : (attached, secondary)
6074 : };
6075 :
6076 0 : if secondary.is_empty() {
6077 : // No secondary locations; nothing for us to do
6078 0 : return;
6079 0 : }
6080 :
6081 0 : for (_, result) in self
6082 0 : .tenant_for_shards_api(
6083 0 : attached,
6084 0 : |tenant_shard_id, client| async move {
6085 0 : client.tenant_heatmap_upload(tenant_shard_id).await
6086 0 : },
6087 : 1,
6088 : 1,
6089 : SHORT_RECONCILE_TIMEOUT,
6090 0 : &self.cancel,
6091 : )
6092 0 : .await
6093 : {
6094 0 : if let Err(e) = result {
6095 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
6096 0 : return;
6097 0 : }
6098 : }
6099 :
6100 0 : for (_, result) in self
6101 0 : .tenant_for_shards_api(
6102 0 : secondary,
6103 0 : |tenant_shard_id, client| async move {
6104 0 : client
6105 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
6106 0 : .await
6107 0 : },
6108 : 1,
6109 : 1,
6110 : SHORT_RECONCILE_TIMEOUT,
6111 0 : &self.cancel,
6112 : )
6113 0 : .await
6114 : {
6115 0 : if let Err(e) = result {
6116 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
6117 0 : return;
6118 0 : }
6119 : }
6120 0 : }
6121 :
6122 0 : pub(crate) async fn tenant_shard_split(
6123 0 : &self,
6124 0 : tenant_id: TenantId,
6125 0 : split_req: TenantShardSplitRequest,
6126 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
6127 : // TODO: return 503 if we get stuck waiting for this lock
6128 : // (issue https://github.com/neondatabase/neon/issues/7108)
6129 0 : let _tenant_lock = trace_exclusive_lock(
6130 0 : &self.tenant_op_locks,
6131 0 : tenant_id,
6132 0 : TenantOperations::ShardSplit,
6133 0 : )
6134 0 : .await;
6135 :
6136 0 : let _gate = self
6137 0 : .reconcilers_gate
6138 0 : .enter()
6139 0 : .map_err(|_| ApiError::ShuttingDown)?;
6140 :
6141 : // Timeline imports on the pageserver side can't handle shard-splits.
6142 : // If the tenant is importing a timeline, dont't shard split it.
6143 0 : match self
6144 0 : .persistence
6145 0 : .is_tenant_importing_timeline(tenant_id)
6146 0 : .await
6147 : {
6148 0 : Ok(importing) => {
6149 0 : if importing {
6150 0 : return Err(ApiError::Conflict(
6151 0 : "Cannot shard split during timeline import".to_string(),
6152 0 : ));
6153 0 : }
6154 : }
6155 0 : Err(err) => {
6156 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6157 0 : "Failed to check for running imports: {err}"
6158 0 : )));
6159 : }
6160 : }
6161 :
6162 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
6163 0 : let new_stripe_size = split_req.new_stripe_size;
6164 :
6165 : // Validate the request and construct parameters. This phase is fallible, but does not require
6166 : // rollback on errors, as it does no I/O and mutates no state.
6167 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
6168 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
6169 0 : ShardSplitAction::Split(params) => params,
6170 : };
6171 :
6172 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
6173 : // we must roll back.
6174 0 : let r = self
6175 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
6176 0 : .await;
6177 :
6178 0 : let (response, waiters) = match r {
6179 0 : Ok(r) => r,
6180 0 : Err(e) => {
6181 : // Split might be part-done, we must do work to abort it.
6182 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
6183 0 : self.abort_tx
6184 0 : .send(TenantShardSplitAbort {
6185 0 : tenant_id,
6186 0 : new_shard_count,
6187 0 : new_stripe_size,
6188 0 : _tenant_lock,
6189 0 : _gate,
6190 0 : })
6191 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
6192 0 : .ok();
6193 0 : return Err(e);
6194 : }
6195 : };
6196 :
6197 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
6198 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
6199 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
6200 : // in [`Self::optimize_all`]
6201 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
6202 0 : .await;
6203 0 : Ok(response)
6204 0 : }
6205 :
6206 0 : fn prepare_tenant_shard_split(
6207 0 : &self,
6208 0 : tenant_id: TenantId,
6209 0 : split_req: TenantShardSplitRequest,
6210 0 : ) -> Result<ShardSplitAction, ApiError> {
6211 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
6212 0 : anyhow::anyhow!("failpoint")
6213 0 : )));
6214 :
6215 0 : let mut policy = None;
6216 0 : let mut config = None;
6217 0 : let mut shard_ident = None;
6218 0 : let mut preferred_az_id = None;
6219 : // Validate input, and calculate which shards we will create
6220 0 : let (old_shard_count, targets) =
6221 : {
6222 0 : let locked = self.inner.read().unwrap();
6223 :
6224 0 : let pageservers = locked.nodes.clone();
6225 :
6226 0 : let mut targets = Vec::new();
6227 :
6228 : // In case this is a retry, count how many already-split shards we found
6229 0 : let mut children_found = Vec::new();
6230 0 : let mut old_shard_count = None;
6231 :
6232 0 : for (tenant_shard_id, shard) in
6233 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
6234 : {
6235 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
6236 : Ordering::Equal => {
6237 : // Already split this
6238 0 : children_found.push(*tenant_shard_id);
6239 0 : continue;
6240 : }
6241 : Ordering::Greater => {
6242 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6243 0 : "Requested count {} but already have shards at count {}",
6244 0 : split_req.new_shard_count,
6245 0 : shard.shard.count.count()
6246 0 : )));
6247 : }
6248 0 : Ordering::Less => {
6249 0 : // Fall through: this shard has lower count than requested,
6250 0 : // is a candidate for splitting.
6251 0 : }
6252 : }
6253 :
6254 0 : match old_shard_count {
6255 0 : None => old_shard_count = Some(shard.shard.count),
6256 0 : Some(old_shard_count) => {
6257 0 : if old_shard_count != shard.shard.count {
6258 : // We may hit this case if a caller asked for two splits to
6259 : // different sizes, before the first one is complete.
6260 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
6261 : // of shard_count=1 and shard_count=2 shards in the map.
6262 0 : return Err(ApiError::Conflict(
6263 0 : "Cannot split, currently mid-split".to_string(),
6264 0 : ));
6265 0 : }
6266 : }
6267 : }
6268 0 : if policy.is_none() {
6269 0 : policy = Some(shard.policy.clone());
6270 0 : }
6271 0 : if shard_ident.is_none() {
6272 0 : shard_ident = Some(shard.shard);
6273 0 : }
6274 0 : if config.is_none() {
6275 0 : config = Some(shard.config.clone());
6276 0 : }
6277 0 : if preferred_az_id.is_none() {
6278 0 : preferred_az_id = shard.preferred_az().cloned();
6279 0 : }
6280 :
6281 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
6282 0 : tracing::info!(
6283 0 : "Tenant shard {} already has shard count {}",
6284 : tenant_shard_id,
6285 : split_req.new_shard_count
6286 : );
6287 0 : continue;
6288 0 : }
6289 :
6290 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
6291 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
6292 0 : ))?;
6293 :
6294 0 : let node = pageservers
6295 0 : .get(&node_id)
6296 0 : .expect("Pageservers may not be deleted while referenced");
6297 :
6298 0 : targets.push(ShardSplitTarget {
6299 0 : parent_id: *tenant_shard_id,
6300 0 : node: node.clone(),
6301 0 : child_ids: tenant_shard_id
6302 0 : .split(ShardCount::new(split_req.new_shard_count)),
6303 0 : });
6304 : }
6305 :
6306 0 : if targets.is_empty() {
6307 0 : if children_found.len() == split_req.new_shard_count as usize {
6308 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
6309 0 : new_shards: children_found,
6310 0 : }));
6311 : } else {
6312 : // No shards found to split, and no existing children found: the
6313 : // tenant doesn't exist at all.
6314 0 : return Err(ApiError::NotFound(
6315 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
6316 0 : ));
6317 : }
6318 0 : }
6319 :
6320 0 : (old_shard_count, targets)
6321 : };
6322 :
6323 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
6324 0 : let old_shard_count = old_shard_count.unwrap();
6325 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
6326 : // This ShardIdentity will be used as the template for all children, so this implicitly
6327 : // applies the new stripe size to the children.
6328 0 : let mut shard_ident = shard_ident.unwrap();
6329 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
6330 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6331 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
6332 0 : shard_ident.stripe_size
6333 0 : )));
6334 0 : }
6335 :
6336 0 : shard_ident.stripe_size = new_stripe_size;
6337 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
6338 0 : shard_ident
6339 : } else {
6340 0 : shard_ident.unwrap()
6341 : };
6342 0 : let policy = policy.unwrap();
6343 0 : let config = config.unwrap();
6344 :
6345 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
6346 0 : old_shard_count,
6347 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
6348 0 : new_stripe_size: split_req.new_stripe_size,
6349 0 : targets,
6350 0 : policy,
6351 0 : config,
6352 0 : shard_ident,
6353 0 : preferred_az_id,
6354 0 : })))
6355 0 : }
6356 :
6357 0 : async fn do_tenant_shard_split(
6358 0 : &self,
6359 0 : tenant_id: TenantId,
6360 0 : params: Box<ShardSplitParams>,
6361 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
6362 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
6363 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
6364 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
6365 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
6366 : // (https://github.com/neondatabase/neon/issues/6676)
6367 :
6368 : let ShardSplitParams {
6369 0 : old_shard_count,
6370 0 : new_shard_count,
6371 0 : new_stripe_size,
6372 0 : mut targets,
6373 0 : policy,
6374 0 : config,
6375 0 : shard_ident,
6376 0 : preferred_az_id,
6377 0 : } = *params;
6378 :
6379 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
6380 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
6381 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
6382 : // at the time of split.
6383 0 : let waiters = {
6384 0 : let mut locked = self.inner.write().unwrap();
6385 0 : let mut waiters = Vec::new();
6386 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6387 0 : for target in &mut targets {
6388 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
6389 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6390 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6391 0 : "Shard {} not found",
6392 0 : target.parent_id
6393 0 : )));
6394 : };
6395 :
6396 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
6397 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6398 0 : return Err(ApiError::Conflict(format!(
6399 0 : "Shard {} unexpectedly rescheduled during split",
6400 0 : target.parent_id
6401 0 : )));
6402 0 : }
6403 :
6404 : // Irrespective of PlacementPolicy, clear secondary locations from intent
6405 0 : shard.intent.clear_secondary(scheduler);
6406 :
6407 : // Run Reconciler to execute detach fo secondary locations.
6408 0 : if let Some(waiter) =
6409 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6410 0 : {
6411 0 : waiters.push(waiter);
6412 0 : }
6413 : }
6414 0 : waiters
6415 : };
6416 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
6417 :
6418 : // Before creating any new child shards in memory or on the pageservers, persist them: this
6419 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
6420 : // acts as the protection against two concurrent attempts to split: one of them will get a database
6421 : // error trying to insert the child shards.
6422 0 : let mut child_tsps = Vec::new();
6423 0 : for target in &targets {
6424 0 : let mut this_child_tsps = Vec::new();
6425 0 : for child in &target.child_ids {
6426 0 : let mut child_shard = shard_ident;
6427 0 : child_shard.number = child.shard_number;
6428 0 : child_shard.count = child.shard_count;
6429 :
6430 0 : tracing::info!(
6431 0 : "Create child shard persistence with stripe size {}",
6432 : shard_ident.stripe_size.0
6433 : );
6434 :
6435 0 : this_child_tsps.push(TenantShardPersistence {
6436 0 : tenant_id: child.tenant_id.to_string(),
6437 0 : shard_number: child.shard_number.0 as i32,
6438 0 : shard_count: child.shard_count.literal() as i32,
6439 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
6440 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
6441 : // populate the correct generation as part of its transaction, to protect us
6442 : // against racing with changes in the state of the parent.
6443 0 : generation: None,
6444 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
6445 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
6446 0 : config: serde_json::to_string(&config).unwrap(),
6447 0 : splitting: SplitState::Splitting,
6448 :
6449 : // Scheduling policies and preferred AZ do not carry through to children
6450 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
6451 0 : .unwrap(),
6452 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
6453 : });
6454 : }
6455 :
6456 0 : child_tsps.push((target.parent_id, this_child_tsps));
6457 : }
6458 :
6459 0 : if let Err(e) = self
6460 0 : .persistence
6461 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
6462 0 : .await
6463 : {
6464 0 : match e {
6465 : DatabaseError::Query(diesel::result::Error::DatabaseError(
6466 : DatabaseErrorKind::UniqueViolation,
6467 : _,
6468 : )) => {
6469 : // Inserting a child shard violated a unique constraint: we raced with another call to
6470 : // this function
6471 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
6472 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
6473 : }
6474 0 : _ => return Err(ApiError::InternalServerError(e.into())),
6475 : }
6476 0 : }
6477 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
6478 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6479 : ));
6480 :
6481 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
6482 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
6483 : // is not set in memory, then it was not persisted.
6484 : {
6485 0 : let mut locked = self.inner.write().unwrap();
6486 0 : for target in &targets {
6487 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
6488 0 : parent_shard.splitting = SplitState::Splitting;
6489 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
6490 0 : // split operation.
6491 0 : parent_shard
6492 0 : .observed
6493 0 : .locations
6494 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
6495 0 : }
6496 : }
6497 : }
6498 :
6499 : // TODO: issue split calls concurrently (this only matters once we're splitting
6500 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
6501 :
6502 : // HADRON: set a timeout for splitting individual shards on page servers.
6503 : // Currently we do not perform any retry because it's not clear if page server can handle
6504 : // partially split shards correctly.
6505 0 : let shard_split_timeout =
6506 0 : if let Some(env::DeploymentMode::Local) = env::get_deployment_mode() {
6507 0 : Duration::from_secs(30)
6508 : } else {
6509 0 : self.config.shard_split_request_timeout
6510 : };
6511 0 : let mut http_client_builder = reqwest::ClientBuilder::new()
6512 0 : .pool_max_idle_per_host(0)
6513 0 : .timeout(shard_split_timeout);
6514 :
6515 0 : for ssl_ca_cert in &self.config.ssl_ca_certs {
6516 0 : http_client_builder = http_client_builder.add_root_certificate(ssl_ca_cert.clone());
6517 0 : }
6518 0 : let http_client = http_client_builder
6519 0 : .build()
6520 0 : .expect("Failed to construct HTTP client");
6521 0 : for target in &targets {
6522 : let ShardSplitTarget {
6523 0 : parent_id,
6524 0 : node,
6525 0 : child_ids,
6526 0 : } = target;
6527 :
6528 0 : let client = PageserverClient::new(
6529 0 : node.get_id(),
6530 0 : http_client.clone(),
6531 0 : node.base_url(),
6532 0 : self.config.pageserver_jwt_token.as_deref(),
6533 : );
6534 :
6535 0 : let response = client
6536 0 : .tenant_shard_split(
6537 0 : *parent_id,
6538 0 : TenantShardSplitRequest {
6539 0 : new_shard_count: new_shard_count.literal(),
6540 0 : new_stripe_size,
6541 0 : },
6542 0 : )
6543 0 : .await
6544 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {parent_id}: {e}")))?;
6545 :
6546 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
6547 0 : "failpoint".to_string()
6548 0 : )));
6549 :
6550 0 : failpoint_support::sleep_millis_async!(
6551 : "shard-split-post-remote-sleep",
6552 0 : &self.reconcilers_cancel
6553 : );
6554 :
6555 0 : tracing::info!(
6556 0 : "Split {} into {}",
6557 : parent_id,
6558 0 : response
6559 0 : .new_shards
6560 0 : .iter()
6561 0 : .map(|s| format!("{s:?}"))
6562 0 : .collect::<Vec<_>>()
6563 0 : .join(",")
6564 : );
6565 :
6566 0 : if &response.new_shards != child_ids {
6567 : // This should never happen: the pageserver should agree with us on how shard splits work.
6568 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6569 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
6570 0 : parent_id,
6571 0 : response.new_shards,
6572 0 : child_ids
6573 0 : )));
6574 0 : }
6575 : }
6576 :
6577 0 : fail::fail_point!("shard-split-pre-complete", |_| Err(ApiError::Conflict(
6578 0 : "failpoint".to_string()
6579 0 : )));
6580 :
6581 0 : pausable_failpoint!("shard-split-pre-complete-pause");
6582 :
6583 : // TODO: if the pageserver restarted concurrently with our split API call,
6584 : // the actual generation of the child shard might differ from the generation
6585 : // we expect it to have. In order for our in-database generation to end up
6586 : // correct, we should carry the child generation back in the response and apply it here
6587 : // in complete_shard_split (and apply the correct generation in memory)
6588 : // (or, we can carry generation in the request and reject the request if
6589 : // it doesn't match, but that requires more retry logic on this side)
6590 :
6591 0 : self.persistence
6592 0 : .complete_shard_split(tenant_id, old_shard_count, new_shard_count)
6593 0 : .await?;
6594 :
6595 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
6596 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6597 : ));
6598 :
6599 : // Replace all the shards we just split with their children: this phase is infallible.
6600 0 : let (response, child_locations, waiters) =
6601 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
6602 :
6603 : // Notify all page servers to detach and clean up the old shards because they will no longer
6604 : // be needed. This is best-effort: if it fails, it will be cleaned up on a subsequent
6605 : // Pageserver re-attach/startup.
6606 0 : let shards_to_cleanup = targets
6607 0 : .iter()
6608 0 : .map(|target| (target.parent_id, target.node.get_id()))
6609 0 : .collect();
6610 0 : self.cleanup_locations(shards_to_cleanup).await;
6611 :
6612 : // Send compute notifications for all the new shards
6613 0 : let mut failed_notifications = Vec::new();
6614 0 : for (child_id, child_ps, stripe_size) in child_locations {
6615 0 : if let Err(e) = self
6616 0 : .compute_hook
6617 0 : .notify_attach(
6618 0 : compute_hook::ShardUpdate {
6619 0 : tenant_shard_id: child_id,
6620 0 : node_id: child_ps,
6621 0 : stripe_size,
6622 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
6623 0 : },
6624 0 : &self.reconcilers_cancel,
6625 0 : )
6626 0 : .await
6627 : {
6628 0 : tracing::warn!(
6629 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
6630 : child_id,
6631 : child_ps
6632 : );
6633 0 : failed_notifications.push(child_id);
6634 0 : }
6635 : }
6636 :
6637 : // If we failed any compute notifications, make a note to retry later.
6638 0 : if !failed_notifications.is_empty() {
6639 0 : let mut locked = self.inner.write().unwrap();
6640 0 : for failed in failed_notifications {
6641 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
6642 0 : shard.pending_compute_notification = true;
6643 0 : }
6644 : }
6645 0 : }
6646 :
6647 0 : Ok((response, waiters))
6648 0 : }
6649 :
6650 : /// A graceful migration: update the preferred node and let optimisation handle the migration
6651 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
6652 : ///
6653 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
6654 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
6655 0 : fn tenant_shard_migrate_with_prewarm(
6656 0 : &self,
6657 0 : migrate_req: &TenantShardMigrateRequest,
6658 0 : shard: &mut TenantShard,
6659 0 : scheduler: &mut Scheduler,
6660 0 : schedule_context: ScheduleContext,
6661 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
6662 0 : shard.set_preferred_node(Some(migrate_req.node_id));
6663 :
6664 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
6665 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
6666 : // e.g. by checking secondary is warm enough.
6667 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
6668 0 : }
6669 :
6670 : /// Immediate migration: directly update the intent state and kick off a reconciler
6671 0 : fn tenant_shard_migrate_immediate(
6672 0 : &self,
6673 0 : migrate_req: &TenantShardMigrateRequest,
6674 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6675 0 : shard: &mut TenantShard,
6676 0 : scheduler: &mut Scheduler,
6677 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
6678 : // Non-graceful migration: update the intent state immediately
6679 0 : let old_attached = *shard.intent.get_attached();
6680 0 : match shard.policy {
6681 0 : PlacementPolicy::Attached(n) => {
6682 : // If our new attached node was a secondary, it no longer should be.
6683 0 : shard
6684 0 : .intent
6685 0 : .remove_secondary(scheduler, migrate_req.node_id);
6686 :
6687 0 : shard
6688 0 : .intent
6689 0 : .set_attached(scheduler, Some(migrate_req.node_id));
6690 :
6691 : // If we were already attached to something, demote that to a secondary
6692 0 : if let Some(old_attached) = old_attached {
6693 0 : if n > 0 {
6694 : // Remove other secondaries to make room for the location we'll demote
6695 0 : while shard.intent.get_secondary().len() >= n {
6696 0 : shard.intent.pop_secondary(scheduler);
6697 0 : }
6698 :
6699 0 : shard.intent.push_secondary(scheduler, old_attached);
6700 0 : }
6701 0 : }
6702 : }
6703 0 : PlacementPolicy::Secondary => {
6704 0 : shard.intent.clear(scheduler);
6705 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6706 0 : }
6707 : PlacementPolicy::Detached => {
6708 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6709 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
6710 0 : )));
6711 : }
6712 : }
6713 :
6714 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
6715 0 : shard.sequence = shard.sequence.next();
6716 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
6717 0 : Ok(self.maybe_configured_reconcile_shard(
6718 0 : shard,
6719 0 : nodes,
6720 0 : (&migrate_req.migration_config).into(),
6721 0 : ))
6722 0 : }
6723 :
6724 0 : pub(crate) async fn tenant_shard_migrate(
6725 0 : &self,
6726 0 : tenant_shard_id: TenantShardId,
6727 0 : migrate_req: TenantShardMigrateRequest,
6728 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6729 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
6730 : // get a different outcome to handle
6731 : enum MigrationOutcome {
6732 : Optimization(Option<ScheduleOptimization>),
6733 : Reconcile(Option<ReconcilerWaiter>),
6734 : }
6735 :
6736 0 : let outcome = {
6737 0 : let mut locked = self.inner.write().unwrap();
6738 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6739 :
6740 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6741 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6742 0 : "Node {} not found",
6743 0 : migrate_req.node_id
6744 0 : )));
6745 : };
6746 :
6747 : // Migration to unavavailable node requires force flag
6748 0 : if !node.is_available() {
6749 0 : if migrate_req.migration_config.override_scheduler {
6750 : // Warn but proceed: the caller may intend to manually adjust the placement of
6751 : // a shard even if the node is down, e.g. if intervening during an incident.
6752 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
6753 : } else {
6754 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
6755 0 : return Err(ApiError::PreconditionFailed(
6756 0 : format!("Node {node} is unavailable").into_boxed_str(),
6757 0 : ));
6758 : }
6759 0 : }
6760 :
6761 : // Calculate the ScheduleContext for this tenant
6762 0 : let mut schedule_context = ScheduleContext::default();
6763 0 : for (_shard_id, shard) in
6764 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
6765 0 : {
6766 0 : schedule_context.avoid(&shard.intent.all_pageservers());
6767 0 : }
6768 :
6769 : // Look up the specific shard we will migrate
6770 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6771 0 : return Err(ApiError::NotFound(
6772 0 : anyhow::anyhow!("Tenant shard not found").into(),
6773 0 : ));
6774 : };
6775 :
6776 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
6777 : // be migrated back by the optimiser.
6778 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
6779 0 : scheduler,
6780 0 : &schedule_context,
6781 0 : migrate_req.node_id,
6782 0 : &[],
6783 0 : ) {
6784 0 : if !migrate_req.migration_config.override_scheduler {
6785 0 : return Err(ApiError::PreconditionFailed(
6786 0 : "Migration to a worse-scoring node".into(),
6787 0 : ));
6788 : } else {
6789 0 : tracing::info!(
6790 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
6791 : migrate_req.node_id
6792 : );
6793 : }
6794 0 : }
6795 :
6796 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
6797 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
6798 0 : return Err(ApiError::PreconditionFailed(
6799 0 : format!(
6800 0 : "Migration expected to originate from {} but shard is on {:?}",
6801 0 : origin_node_id,
6802 0 : shard.intent.get_attached()
6803 0 : )
6804 0 : .into(),
6805 0 : ));
6806 0 : }
6807 0 : }
6808 :
6809 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6810 : // No-op case: we will still proceed to wait for reconciliation in case it is
6811 : // incomplete from an earlier update to the intent.
6812 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
6813 :
6814 : // An instruction to migrate to the currently attached node should
6815 : // cancel any pending graceful migration
6816 0 : shard.set_preferred_node(None);
6817 :
6818 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
6819 0 : shard,
6820 0 : nodes,
6821 0 : (&migrate_req.migration_config).into(),
6822 0 : ))
6823 0 : } else if migrate_req.migration_config.prewarm {
6824 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
6825 0 : &migrate_req,
6826 0 : shard,
6827 0 : scheduler,
6828 0 : schedule_context,
6829 0 : )?)
6830 : } else {
6831 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
6832 0 : &migrate_req,
6833 0 : nodes,
6834 0 : shard,
6835 0 : scheduler,
6836 0 : )?)
6837 : }
6838 : };
6839 :
6840 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
6841 0 : let waiter = match outcome {
6842 0 : MigrationOutcome::Optimization(Some(optimization)) => {
6843 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
6844 : // we might as well do it more promptly as this is a direct external request.
6845 0 : let mut validated = self
6846 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
6847 0 : .await;
6848 0 : if let Some((_shard_id, optimization)) = validated.pop() {
6849 0 : let mut locked = self.inner.write().unwrap();
6850 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6851 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6852 : // Rare but possible: tenant is removed between generating optimisation and validating it.
6853 0 : return Err(ApiError::NotFound(
6854 0 : anyhow::anyhow!("Tenant shard not found").into(),
6855 0 : ));
6856 : };
6857 :
6858 0 : if !shard.apply_optimization(scheduler, optimization) {
6859 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
6860 : // and therefore not applied.
6861 0 : tracing::warn!(
6862 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
6863 : );
6864 0 : }
6865 0 : self.maybe_configured_reconcile_shard(
6866 0 : shard,
6867 0 : nodes,
6868 0 : (&migrate_req.migration_config).into(),
6869 : )
6870 : } else {
6871 0 : None
6872 : }
6873 : }
6874 0 : MigrationOutcome::Optimization(None) => None,
6875 0 : MigrationOutcome::Reconcile(waiter) => waiter,
6876 : };
6877 :
6878 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
6879 : // locations, this has a good chance of timing out.
6880 0 : if let Some(waiter) = waiter {
6881 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6882 : } else {
6883 0 : tracing::info!("Migration is a no-op");
6884 : }
6885 :
6886 0 : Ok(TenantShardMigrateResponse {})
6887 0 : }
6888 :
6889 0 : pub(crate) async fn tenant_shard_migrate_secondary(
6890 0 : &self,
6891 0 : tenant_shard_id: TenantShardId,
6892 0 : migrate_req: TenantShardMigrateRequest,
6893 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6894 0 : let waiter = {
6895 0 : let mut locked = self.inner.write().unwrap();
6896 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6897 :
6898 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6899 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6900 0 : "Node {} not found",
6901 0 : migrate_req.node_id
6902 0 : )));
6903 : };
6904 :
6905 0 : if !node.is_available() {
6906 : // Warn but proceed: the caller may intend to manually adjust the placement of
6907 : // a shard even if the node is down, e.g. if intervening during an incident.
6908 0 : tracing::warn!("Migrating to unavailable node {node}");
6909 0 : }
6910 :
6911 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6912 0 : return Err(ApiError::NotFound(
6913 0 : anyhow::anyhow!("Tenant shard not found").into(),
6914 0 : ));
6915 : };
6916 :
6917 0 : if shard.intent.get_secondary().len() == 1
6918 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
6919 : {
6920 0 : tracing::info!(
6921 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
6922 : shard.intent
6923 : );
6924 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6925 0 : tracing::info!(
6926 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
6927 : );
6928 : } else {
6929 0 : let old_secondaries = shard.intent.get_secondary().clone();
6930 0 : for secondary in old_secondaries {
6931 0 : shard.intent.remove_secondary(scheduler, secondary);
6932 0 : }
6933 :
6934 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6935 0 : shard.sequence = shard.sequence.next();
6936 0 : tracing::info!(
6937 0 : "Migrating secondary to {node}: new intent {:?}",
6938 : shard.intent
6939 : );
6940 : }
6941 :
6942 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6943 : };
6944 :
6945 0 : if let Some(waiter) = waiter {
6946 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6947 : } else {
6948 0 : tracing::info!("Migration is a no-op");
6949 : }
6950 :
6951 0 : Ok(TenantShardMigrateResponse {})
6952 0 : }
6953 :
6954 : /// 'cancel' in this context means cancel any ongoing reconcile
6955 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
6956 0 : &self,
6957 0 : tenant_shard_id: TenantShardId,
6958 0 : ) -> Result<(), ApiError> {
6959 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
6960 0 : let waiter = {
6961 0 : let locked = self.inner.write().unwrap();
6962 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6963 0 : return Err(ApiError::NotFound(
6964 0 : anyhow::anyhow!("Tenant shard not found").into(),
6965 0 : ));
6966 : };
6967 :
6968 0 : let waiter = shard.get_waiter();
6969 0 : match waiter {
6970 : None => {
6971 0 : tracing::info!("Shard does not have an ongoing Reconciler");
6972 0 : return Ok(());
6973 : }
6974 0 : Some(waiter) => {
6975 0 : tracing::info!("Cancelling Reconciler");
6976 0 : shard.cancel_reconciler();
6977 0 : waiter
6978 : }
6979 : }
6980 : };
6981 :
6982 : // Cancellation should be prompt. If this fails we have still done our job of firing the
6983 : // cancellation token, but by returning an ApiError we will indicate to the caller that
6984 : // the Reconciler is misbehaving and not respecting the cancellation token
6985 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
6986 0 : .await?;
6987 :
6988 0 : Ok(())
6989 0 : }
6990 :
6991 : /// This is for debug/support only: we simply drop all state for a tenant, without
6992 : /// detaching or deleting it on pageservers.
6993 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
6994 0 : self.persistence.delete_tenant(tenant_id).await?;
6995 :
6996 0 : let mut locked = self.inner.write().unwrap();
6997 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6998 0 : let mut shards = Vec::new();
6999 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
7000 0 : shards.push(*tenant_shard_id);
7001 0 : }
7002 :
7003 0 : for shard_id in shards {
7004 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
7005 0 : shard.intent.clear(scheduler);
7006 0 : }
7007 : }
7008 :
7009 0 : Ok(())
7010 0 : }
7011 :
7012 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
7013 : /// tenant with a very high generation number so that it will see the existing data.
7014 : /// It does not create timelines on safekeepers, because they might already exist on some
7015 : /// safekeeper set. So, the timelines are not storcon-managed after the import.
7016 0 : pub(crate) async fn tenant_import(
7017 0 : &self,
7018 0 : tenant_id: TenantId,
7019 0 : ) -> Result<TenantCreateResponse, ApiError> {
7020 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
7021 0 : let maybe_node = {
7022 0 : self.inner
7023 0 : .read()
7024 0 : .unwrap()
7025 0 : .nodes
7026 0 : .values()
7027 0 : .find(|n| n.is_available())
7028 0 : .cloned()
7029 : };
7030 0 : let Some(node) = maybe_node else {
7031 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
7032 : };
7033 :
7034 0 : let client = PageserverClient::new(
7035 0 : node.get_id(),
7036 0 : self.http_client.clone(),
7037 0 : node.base_url(),
7038 0 : self.config.pageserver_jwt_token.as_deref(),
7039 : );
7040 :
7041 0 : let scan_result = client
7042 0 : .tenant_scan_remote_storage(tenant_id)
7043 0 : .await
7044 0 : .map_err(|e| passthrough_api_error(&node, e))?;
7045 :
7046 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
7047 0 : let Some(shard_count) = scan_result
7048 0 : .shards
7049 0 : .iter()
7050 0 : .map(|s| s.tenant_shard_id.shard_count)
7051 0 : .max()
7052 : else {
7053 0 : return Err(ApiError::NotFound(
7054 0 : anyhow::anyhow!("No shards found").into(),
7055 0 : ));
7056 : };
7057 :
7058 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
7059 : // to
7060 0 : let generation = scan_result
7061 0 : .shards
7062 0 : .iter()
7063 0 : .map(|s| s.generation)
7064 0 : .max()
7065 0 : .expect("We already validated >0 shards");
7066 :
7067 : // Find the tenant's stripe size. This wasn't always persisted in the tenant manifest, so
7068 : // fall back to the original default stripe size of 32768 (256 MB) if it's not specified.
7069 : const ORIGINAL_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(32768);
7070 0 : let stripe_size = scan_result
7071 0 : .shards
7072 0 : .iter()
7073 0 : .find(|s| s.tenant_shard_id.shard_count == shard_count && s.generation == generation)
7074 0 : .expect("we validated >0 shards above")
7075 : .stripe_size
7076 0 : .unwrap_or_else(|| {
7077 0 : if shard_count.count() > 1 {
7078 0 : warn!("unknown stripe size, assuming {ORIGINAL_STRIPE_SIZE}");
7079 0 : }
7080 0 : ORIGINAL_STRIPE_SIZE
7081 0 : });
7082 :
7083 0 : let (response, waiters) = self
7084 0 : .do_tenant_create(TenantCreateRequest {
7085 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
7086 0 : generation,
7087 0 :
7088 0 : shard_parameters: ShardParameters {
7089 0 : count: shard_count,
7090 0 : stripe_size,
7091 0 : },
7092 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
7093 0 : config: TenantConfig::default(),
7094 0 : })
7095 0 : .await?;
7096 :
7097 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
7098 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
7099 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
7100 : // reconcile, as reconciliation includes notifying compute.
7101 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
7102 0 : }
7103 :
7104 0 : Ok(response)
7105 0 : }
7106 :
7107 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
7108 : /// we don't have to make TenantShard clonable in the return path.
7109 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7110 0 : let serialized = {
7111 0 : let locked = self.inner.read().unwrap();
7112 0 : let result = locked.tenants.values().collect::<Vec<_>>();
7113 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
7114 : };
7115 :
7116 0 : hyper::Response::builder()
7117 0 : .status(hyper::StatusCode::OK)
7118 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7119 0 : .body(hyper::Body::from(serialized))
7120 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7121 0 : }
7122 :
7123 : /// Check the consistency of in-memory state vs. persistent state, and check that the
7124 : /// scheduler's statistics are up to date.
7125 : ///
7126 : /// These consistency checks expect an **idle** system. If changes are going on while
7127 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
7128 : /// checks, but not suitable for running continuously in the background in the field.
7129 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
7130 0 : let (mut expect_nodes, mut expect_shards) = {
7131 0 : let locked = self.inner.read().unwrap();
7132 :
7133 0 : locked
7134 0 : .scheduler
7135 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
7136 0 : .context("Scheduler checks")
7137 0 : .map_err(ApiError::InternalServerError)?;
7138 :
7139 0 : let expect_nodes = locked
7140 0 : .nodes
7141 0 : .values()
7142 0 : .map(|n| n.to_persistent())
7143 0 : .collect::<Vec<_>>();
7144 :
7145 0 : let expect_shards = locked
7146 0 : .tenants
7147 0 : .values()
7148 0 : .map(|t| t.to_persistent())
7149 0 : .collect::<Vec<_>>();
7150 :
7151 : // This method can only validate the state of an idle system: if a reconcile is in
7152 : // progress, fail out early to avoid giving false errors on state that won't match
7153 : // between database and memory under a ReconcileResult is processed.
7154 0 : for t in locked.tenants.values() {
7155 0 : if t.reconciler.is_some() {
7156 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7157 0 : "Shard {} reconciliation in progress",
7158 0 : t.tenant_shard_id
7159 0 : )));
7160 0 : }
7161 : }
7162 :
7163 0 : (expect_nodes, expect_shards)
7164 : };
7165 :
7166 0 : let mut nodes = self.persistence.list_nodes().await?;
7167 0 : expect_nodes.sort_by_key(|n| n.node_id);
7168 0 : nodes.sort_by_key(|n| n.node_id);
7169 :
7170 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
7171 0 : let node_result = if nodes != expect_nodes {
7172 0 : tracing::error!("Consistency check failed on nodes.");
7173 0 : tracing::error!(
7174 0 : "Nodes in memory: {}",
7175 0 : serde_json::to_string(&expect_nodes)
7176 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7177 : );
7178 0 : tracing::error!(
7179 0 : "Nodes in database: {}",
7180 0 : serde_json::to_string(&nodes)
7181 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7182 : );
7183 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
7184 0 : "Node consistency failure"
7185 0 : )))
7186 : } else {
7187 0 : Ok(())
7188 : };
7189 :
7190 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
7191 0 : persistent_shards
7192 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7193 :
7194 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7195 :
7196 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
7197 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
7198 : // fields are added, before doing a comparison.
7199 0 : for tsp in &mut persistent_shards {
7200 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
7201 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
7202 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
7203 : }
7204 :
7205 0 : if persistent_shards != expect_shards {
7206 0 : tracing::error!("Consistency check failed on shards.");
7207 :
7208 0 : tracing::error!(
7209 0 : "Shards in memory: {}",
7210 0 : serde_json::to_string(&expect_shards)
7211 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7212 : );
7213 0 : tracing::error!(
7214 0 : "Shards in database: {}",
7215 0 : serde_json::to_string(&persistent_shards)
7216 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7217 : );
7218 :
7219 : // The total dump log lines above are useful in testing but in the field grafana will
7220 : // usually just drop them because they're so large. So we also do some explicit logging
7221 : // of just the diffs.
7222 0 : let persistent_shards = persistent_shards
7223 0 : .into_iter()
7224 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7225 0 : .collect::<HashMap<_, _>>();
7226 0 : let expect_shards = expect_shards
7227 0 : .into_iter()
7228 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7229 0 : .collect::<HashMap<_, _>>();
7230 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
7231 0 : match expect_shards.get(tenant_shard_id) {
7232 : None => {
7233 0 : tracing::error!(
7234 0 : "Shard {} found in database but not in memory",
7235 : tenant_shard_id
7236 : );
7237 : }
7238 0 : Some(expect_tsp) => {
7239 0 : if expect_tsp != persistent_tsp {
7240 0 : tracing::error!(
7241 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
7242 : tenant_shard_id,
7243 0 : serde_json::to_string(expect_tsp).unwrap(),
7244 0 : serde_json::to_string(&persistent_tsp).unwrap()
7245 : );
7246 0 : }
7247 : }
7248 : }
7249 : }
7250 :
7251 : // Having already logged any differences, log any shards that simply aren't present in the database
7252 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
7253 0 : if !persistent_shards.contains_key(tenant_shard_id) {
7254 0 : tracing::error!(
7255 0 : "Shard {} found in memory but not in database: {}",
7256 : tenant_shard_id,
7257 0 : serde_json::to_string(memory_tsp)
7258 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7259 : );
7260 0 : }
7261 : }
7262 :
7263 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7264 0 : "Shard consistency failure"
7265 0 : )));
7266 0 : }
7267 :
7268 0 : node_result
7269 0 : }
7270 :
7271 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
7272 : /// we don't have to make TenantShard clonable in the return path.
7273 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7274 0 : let serialized = {
7275 0 : let locked = self.inner.read().unwrap();
7276 0 : serde_json::to_string(&locked.scheduler)
7277 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7278 : };
7279 :
7280 0 : hyper::Response::builder()
7281 0 : .status(hyper::StatusCode::OK)
7282 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7283 0 : .body(hyper::Body::from(serialized))
7284 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7285 0 : }
7286 :
7287 : /// This is for debug/support only: we simply drop all state for a tenant, without
7288 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
7289 : /// tenants that were on this node.
7290 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
7291 0 : self.persistence.set_tombstone(node_id).await?;
7292 :
7293 0 : let mut locked = self.inner.write().unwrap();
7294 :
7295 0 : for shard in locked.tenants.values_mut() {
7296 0 : shard.deref_node(node_id);
7297 0 : shard.observed.locations.remove(&node_id);
7298 0 : }
7299 :
7300 0 : let mut nodes = (*locked.nodes).clone();
7301 0 : nodes.remove(&node_id);
7302 0 : locked.nodes = Arc::new(nodes);
7303 0 : metrics::METRICS_REGISTRY
7304 0 : .metrics_group
7305 0 : .storage_controller_pageserver_nodes
7306 0 : .set(locked.nodes.len() as i64);
7307 0 : metrics::METRICS_REGISTRY
7308 0 : .metrics_group
7309 0 : .storage_controller_https_pageserver_nodes
7310 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7311 :
7312 0 : locked.scheduler.node_remove(node_id);
7313 :
7314 0 : Ok(())
7315 0 : }
7316 :
7317 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
7318 : /// that we don't leave any bad state behind in the storage controller, but unclean
7319 : /// in the sense that we are not carefully draining the node.
7320 0 : pub(crate) async fn node_delete_old(&self, node_id: NodeId) -> Result<(), ApiError> {
7321 0 : let _node_lock =
7322 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
7323 :
7324 : // 1. Atomically update in-memory state:
7325 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
7326 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
7327 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
7328 : // re-insert references to this node into the ObservedState of shards
7329 : // - drop the node from the scheduler
7330 : {
7331 0 : let mut locked = self.inner.write().unwrap();
7332 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7333 :
7334 : {
7335 0 : let mut nodes_mut = (*nodes).deref().clone();
7336 0 : match nodes_mut.get_mut(&node_id) {
7337 0 : Some(node) => {
7338 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
7339 0 : // if we crash it would not be desirable to leave the node paused after a restart.
7340 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
7341 0 : }
7342 : None => {
7343 0 : tracing::info!(
7344 0 : "Node not found: presuming this is a retry and returning success"
7345 : );
7346 0 : return Ok(());
7347 : }
7348 : }
7349 :
7350 0 : *nodes = Arc::new(nodes_mut);
7351 : }
7352 :
7353 0 : for (_tenant_id, mut schedule_context, shards) in
7354 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7355 : {
7356 0 : for shard in shards {
7357 0 : if shard.deref_node(node_id) {
7358 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
7359 : // TODO: implement force flag to remove a node even if we can't reschedule
7360 : // a tenant
7361 0 : tracing::error!(
7362 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7363 : shard.tenant_shard_id
7364 : );
7365 0 : return Err(e.into());
7366 : } else {
7367 0 : tracing::info!(
7368 0 : "Rescheduled shard {} away from node during deletion",
7369 : shard.tenant_shard_id
7370 : )
7371 : }
7372 :
7373 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
7374 0 : }
7375 :
7376 : // Here we remove an existing observed location for the node we're removing, and it will
7377 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7378 : // process_result.
7379 : //
7380 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
7381 : // means any reconciles we spawned will know about the node we're deleting, enabling them
7382 : // to do live migrations if it's still online.
7383 0 : shard.observed.locations.remove(&node_id);
7384 : }
7385 : }
7386 :
7387 0 : scheduler.node_remove(node_id);
7388 :
7389 : {
7390 0 : let mut nodes_mut = (**nodes).clone();
7391 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7392 0 : // Ensure that any reconciler holding an Arc<> to this node will
7393 0 : // drop out when trying to RPC to it (setting Offline state sets the
7394 0 : // cancellation token on the Node object).
7395 0 : removed_node.set_availability(NodeAvailability::Offline);
7396 0 : }
7397 0 : *nodes = Arc::new(nodes_mut);
7398 0 : metrics::METRICS_REGISTRY
7399 0 : .metrics_group
7400 0 : .storage_controller_pageserver_nodes
7401 0 : .set(nodes.len() as i64);
7402 0 : metrics::METRICS_REGISTRY
7403 0 : .metrics_group
7404 0 : .storage_controller_https_pageserver_nodes
7405 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7406 : }
7407 : }
7408 :
7409 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
7410 : // the removed node, as this column means "The pageserver to which this generation was issued", and
7411 : // their generations won't get updated until the reconcilers moving them away from this node complete.
7412 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
7413 : // that exists.
7414 :
7415 : // 2. Actually delete the node from in-memory state and set tombstone to the database
7416 : // for preventing the node to register again.
7417 0 : tracing::info!("Deleting node from database");
7418 0 : self.persistence.set_tombstone(node_id).await?;
7419 :
7420 0 : Ok(())
7421 0 : }
7422 :
7423 0 : pub(crate) async fn delete_node(
7424 0 : self: &Arc<Self>,
7425 0 : node_id: NodeId,
7426 0 : policy_on_start: NodeSchedulingPolicy,
7427 0 : force: bool,
7428 0 : cancel: CancellationToken,
7429 0 : ) -> Result<(), OperationError> {
7430 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal).build();
7431 :
7432 0 : let mut waiters: Vec<ReconcilerWaiter> = Vec::new();
7433 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
7434 :
7435 0 : let reset_node_policy_on_cancel = || async {
7436 0 : match self
7437 0 : .node_configure(node_id, None, Some(policy_on_start))
7438 0 : .await
7439 : {
7440 0 : Ok(()) => OperationError::Cancelled,
7441 0 : Err(err) => {
7442 0 : OperationError::FinalizeError(
7443 0 : format!(
7444 0 : "Failed to finalise delete cancel of {} by setting scheduling policy to {}: {}",
7445 0 : node_id, String::from(policy_on_start), err
7446 0 : )
7447 0 : .into(),
7448 0 : )
7449 : }
7450 : }
7451 0 : };
7452 :
7453 0 : while !tid_iter.finished() {
7454 0 : if cancel.is_cancelled() {
7455 0 : return Err(reset_node_policy_on_cancel().await);
7456 0 : }
7457 :
7458 0 : operation_utils::validate_node_state(
7459 0 : &node_id,
7460 0 : self.inner.read().unwrap().nodes.clone(),
7461 0 : NodeSchedulingPolicy::Deleting,
7462 0 : )?;
7463 :
7464 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7465 0 : let tid = match tid_iter.next() {
7466 0 : Some(tid) => tid,
7467 : None => {
7468 0 : break;
7469 : }
7470 : };
7471 :
7472 0 : let mut locked = self.inner.write().unwrap();
7473 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7474 :
7475 : // Calculate a schedule context here to avoid borrow checker issues.
7476 0 : let mut schedule_context = ScheduleContext::default();
7477 0 : for (_, shard) in tenants.range(TenantShardId::tenant_range(tid.tenant_id)) {
7478 0 : schedule_context.avoid(&shard.intent.all_pageservers());
7479 0 : }
7480 :
7481 0 : let tenant_shard = match tenants.get_mut(&tid) {
7482 0 : Some(tenant_shard) => tenant_shard,
7483 : None => {
7484 : // Tenant shard was deleted by another operation. Skip it.
7485 0 : continue;
7486 : }
7487 : };
7488 :
7489 0 : match tenant_shard.get_scheduling_policy() {
7490 0 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {
7491 0 : // A migration during delete is classed as 'essential' because it is required to
7492 0 : // uphold our availability goals for the tenant: this shard is elegible for migration.
7493 0 : }
7494 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
7495 : // If we have been asked to avoid rescheduling this shard, then do not migrate it during a deletion
7496 0 : tracing::warn!(
7497 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7498 0 : "Skip migration during deletion because shard scheduling policy {:?} disallows it",
7499 0 : tenant_shard.get_scheduling_policy(),
7500 : );
7501 0 : continue;
7502 : }
7503 : }
7504 :
7505 0 : if tenant_shard.deref_node(node_id) {
7506 0 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
7507 0 : tracing::error!(
7508 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7509 : tenant_shard.tenant_shard_id
7510 : );
7511 0 : return Err(OperationError::ImpossibleConstraint(e.to_string().into()));
7512 : } else {
7513 0 : tracing::info!(
7514 0 : "Rescheduled shard {} away from node during deletion",
7515 : tenant_shard.tenant_shard_id
7516 : )
7517 : }
7518 :
7519 0 : let waiter = self.maybe_configured_reconcile_shard(
7520 0 : tenant_shard,
7521 0 : nodes,
7522 0 : reconciler_config,
7523 0 : );
7524 :
7525 0 : if force {
7526 0 : // Here we remove an existing observed location for the node we're removing, and it will
7527 0 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7528 0 : // process_result.
7529 0 : //
7530 0 : // Note that we update the shard's observed state _after_ calling maybe_configured_reconcile_shard:
7531 0 : // that means any reconciles we spawned will know about the node we're deleting,
7532 0 : // enabling them to do live migrations if it's still online.
7533 0 : tenant_shard.observed.locations.remove(&node_id);
7534 0 : } else if let Some(waiter) = waiter {
7535 0 : waiters.push(waiter);
7536 0 : }
7537 0 : }
7538 : }
7539 :
7540 0 : waiters = self
7541 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
7542 0 : .await;
7543 :
7544 0 : failpoint_support::sleep_millis_async!("sleepy-delete-loop", &cancel);
7545 : }
7546 :
7547 0 : while !waiters.is_empty() {
7548 0 : if cancel.is_cancelled() {
7549 0 : return Err(reset_node_policy_on_cancel().await);
7550 0 : }
7551 :
7552 0 : tracing::info!("Awaiting {} pending delete reconciliations", waiters.len());
7553 :
7554 0 : waiters = self
7555 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7556 0 : .await;
7557 : }
7558 :
7559 0 : let pf = pausable_failpoint!("delete-node-after-reconciles-spawned", &cancel);
7560 0 : if pf.is_err() {
7561 : // An error from pausable_failpoint indicates the cancel token was triggered.
7562 0 : return Err(reset_node_policy_on_cancel().await);
7563 0 : }
7564 :
7565 0 : self.persistence
7566 0 : .set_tombstone(node_id)
7567 0 : .await
7568 0 : .map_err(|e| OperationError::FinalizeError(e.to_string().into()))?;
7569 :
7570 : {
7571 0 : let mut locked = self.inner.write().unwrap();
7572 0 : let (nodes, _, scheduler) = locked.parts_mut();
7573 :
7574 0 : scheduler.node_remove(node_id);
7575 :
7576 0 : let mut nodes_mut = (**nodes).clone();
7577 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7578 0 : // Ensure that any reconciler holding an Arc<> to this node will
7579 0 : // drop out when trying to RPC to it (setting Offline state sets the
7580 0 : // cancellation token on the Node object).
7581 0 : removed_node.set_availability(NodeAvailability::Offline);
7582 0 : }
7583 0 : *nodes = Arc::new(nodes_mut);
7584 :
7585 0 : metrics::METRICS_REGISTRY
7586 0 : .metrics_group
7587 0 : .storage_controller_pageserver_nodes
7588 0 : .set(nodes.len() as i64);
7589 0 : metrics::METRICS_REGISTRY
7590 0 : .metrics_group
7591 0 : .storage_controller_https_pageserver_nodes
7592 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7593 : }
7594 :
7595 0 : Ok(())
7596 0 : }
7597 :
7598 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
7599 0 : let nodes = {
7600 0 : self.inner
7601 0 : .read()
7602 0 : .unwrap()
7603 0 : .nodes
7604 0 : .values()
7605 0 : .cloned()
7606 0 : .collect::<Vec<_>>()
7607 : };
7608 :
7609 0 : Ok(nodes)
7610 0 : }
7611 :
7612 0 : pub(crate) async fn tombstone_list(&self) -> Result<Vec<Node>, ApiError> {
7613 0 : self.persistence
7614 0 : .list_tombstones()
7615 0 : .await?
7616 0 : .into_iter()
7617 0 : .map(|np| Node::from_persistent(np, false))
7618 0 : .collect::<Result<Vec<_>, _>>()
7619 0 : .map_err(ApiError::InternalServerError)
7620 0 : }
7621 :
7622 0 : pub(crate) async fn tombstone_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
7623 0 : let _node_lock = trace_exclusive_lock(
7624 0 : &self.node_op_locks,
7625 0 : node_id,
7626 0 : NodeOperations::DeleteTombstone,
7627 0 : )
7628 0 : .await;
7629 :
7630 0 : if matches!(self.get_node(node_id).await, Err(ApiError::NotFound(_))) {
7631 0 : self.persistence.delete_node(node_id).await?;
7632 0 : Ok(())
7633 : } else {
7634 0 : Err(ApiError::Conflict(format!(
7635 0 : "Node {node_id} is in use, consider using tombstone API first"
7636 0 : )))
7637 : }
7638 0 : }
7639 :
7640 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
7641 0 : self.inner
7642 0 : .read()
7643 0 : .unwrap()
7644 0 : .nodes
7645 0 : .get(&node_id)
7646 0 : .cloned()
7647 0 : .ok_or(ApiError::NotFound(
7648 0 : format!("Node {node_id} not registered").into(),
7649 0 : ))
7650 0 : }
7651 :
7652 0 : pub(crate) async fn get_node_shards(
7653 0 : &self,
7654 0 : node_id: NodeId,
7655 0 : ) -> Result<NodeShardResponse, ApiError> {
7656 0 : let locked = self.inner.read().unwrap();
7657 0 : let mut shards = Vec::new();
7658 0 : for (tid, tenant) in locked.tenants.iter() {
7659 0 : let is_intended_secondary = match (
7660 0 : tenant.intent.get_attached() == &Some(node_id),
7661 0 : tenant.intent.get_secondary().contains(&node_id),
7662 0 : ) {
7663 : (true, true) => {
7664 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7665 0 : "{} attached as primary+secondary on the same node",
7666 0 : tid
7667 0 : )));
7668 : }
7669 0 : (true, false) => Some(false),
7670 0 : (false, true) => Some(true),
7671 0 : (false, false) => None,
7672 : };
7673 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
7674 0 : tenant.observed.locations.get(&node_id)
7675 : {
7676 0 : Some(conf.secondary_conf.is_some())
7677 : } else {
7678 0 : None
7679 : };
7680 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
7681 0 : shards.push(NodeShard {
7682 0 : tenant_shard_id: *tid,
7683 0 : is_intended_secondary,
7684 0 : is_observed_secondary,
7685 0 : });
7686 0 : }
7687 : }
7688 0 : Ok(NodeShardResponse { node_id, shards })
7689 0 : }
7690 :
7691 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
7692 0 : self.persistence.get_leader().await
7693 0 : }
7694 :
7695 0 : pub(crate) async fn node_register(
7696 0 : &self,
7697 0 : register_req: NodeRegisterRequest,
7698 0 : ) -> Result<(), ApiError> {
7699 0 : let _node_lock = trace_exclusive_lock(
7700 0 : &self.node_op_locks,
7701 0 : register_req.node_id,
7702 0 : NodeOperations::Register,
7703 0 : )
7704 0 : .await;
7705 :
7706 : #[derive(PartialEq)]
7707 : enum RegistrationStatus {
7708 : UpToDate,
7709 : NeedUpdate,
7710 : Mismatched,
7711 : New,
7712 : }
7713 :
7714 0 : let registration_status = {
7715 0 : let locked = self.inner.read().unwrap();
7716 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
7717 0 : if node.registration_match(®ister_req) {
7718 0 : if node.need_update(®ister_req) {
7719 0 : RegistrationStatus::NeedUpdate
7720 : } else {
7721 0 : RegistrationStatus::UpToDate
7722 : }
7723 : } else {
7724 0 : RegistrationStatus::Mismatched
7725 : }
7726 : } else {
7727 0 : RegistrationStatus::New
7728 : }
7729 : };
7730 :
7731 0 : match registration_status {
7732 : RegistrationStatus::UpToDate => {
7733 0 : tracing::info!(
7734 0 : "Node {} re-registered with matching address and is up to date",
7735 : register_req.node_id
7736 : );
7737 :
7738 0 : return Ok(());
7739 : }
7740 : RegistrationStatus::Mismatched => {
7741 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
7742 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
7743 : // a fixed address through the lifetime of a node.
7744 0 : tracing::warn!(
7745 0 : "Node {} tried to register with different address",
7746 : register_req.node_id
7747 : );
7748 0 : return Err(ApiError::Conflict(
7749 0 : "Node is already registered with different address".to_string(),
7750 0 : ));
7751 : }
7752 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
7753 0 : // fallthrough
7754 0 : }
7755 : }
7756 :
7757 : // We do not require that a node is actually online when registered (it will start life
7758 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
7759 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
7760 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
7761 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
7762 0 : if tokio::net::lookup_host(format!(
7763 0 : "{}:{}",
7764 : register_req.listen_http_addr, register_req.listen_http_port
7765 : ))
7766 0 : .await
7767 0 : .is_err()
7768 : {
7769 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
7770 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
7771 : // we return a soft 503 error, to encourage callers to retry past transient issues.
7772 0 : return Err(ApiError::ResourceUnavailable(
7773 0 : format!(
7774 0 : "Node {} tried to register with unknown DNS name '{}'",
7775 0 : register_req.node_id, register_req.listen_http_addr
7776 0 : )
7777 0 : .into(),
7778 0 : ));
7779 0 : }
7780 :
7781 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
7782 0 : return Err(ApiError::PreconditionFailed(
7783 0 : format!(
7784 0 : "Node {} has no https port, but use_https is enabled",
7785 0 : register_req.node_id
7786 0 : )
7787 0 : .into(),
7788 0 : ));
7789 0 : }
7790 :
7791 0 : if register_req.listen_grpc_addr.is_some() != register_req.listen_grpc_port.is_some() {
7792 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
7793 0 : "must specify both gRPC address and port"
7794 0 : )));
7795 0 : }
7796 :
7797 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
7798 : // This ensures that before we use it for anything or expose it via any external
7799 : // API, it is guaranteed to be available after a restart.
7800 0 : let new_node = Node::new(
7801 0 : register_req.node_id,
7802 0 : register_req.listen_http_addr,
7803 0 : register_req.listen_http_port,
7804 0 : register_req.listen_https_port,
7805 0 : register_req.listen_pg_addr,
7806 0 : register_req.listen_pg_port,
7807 0 : register_req.listen_grpc_addr,
7808 0 : register_req.listen_grpc_port,
7809 0 : register_req.availability_zone_id.clone(),
7810 0 : self.config.use_https_pageserver_api,
7811 : );
7812 0 : let new_node = match new_node {
7813 0 : Ok(new_node) => new_node,
7814 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
7815 : };
7816 :
7817 0 : match registration_status {
7818 : RegistrationStatus::New => {
7819 0 : self.persistence.insert_node(&new_node).await.map_err(|e| {
7820 0 : if matches!(
7821 0 : e,
7822 : crate::persistence::DatabaseError::Query(
7823 : diesel::result::Error::DatabaseError(
7824 : diesel::result::DatabaseErrorKind::UniqueViolation,
7825 : _,
7826 : )
7827 : )
7828 : ) {
7829 : // The node can be deleted by tombstone API, and not show up in the list of nodes.
7830 : // If you see this error, check tombstones first.
7831 0 : ApiError::Conflict(format!("Node {} is already exists", new_node.get_id()))
7832 : } else {
7833 0 : ApiError::from(e)
7834 : }
7835 0 : })?;
7836 : }
7837 : RegistrationStatus::NeedUpdate => {
7838 0 : self.persistence
7839 0 : .update_node_on_registration(
7840 0 : register_req.node_id,
7841 0 : register_req.listen_https_port,
7842 0 : )
7843 0 : .await?
7844 : }
7845 0 : _ => unreachable!("Other statuses have been processed earlier"),
7846 : }
7847 :
7848 0 : let mut locked = self.inner.write().unwrap();
7849 0 : let mut new_nodes = (*locked.nodes).clone();
7850 :
7851 0 : locked.scheduler.node_upsert(&new_node);
7852 0 : new_nodes.insert(register_req.node_id, new_node);
7853 :
7854 0 : locked.nodes = Arc::new(new_nodes);
7855 :
7856 0 : metrics::METRICS_REGISTRY
7857 0 : .metrics_group
7858 0 : .storage_controller_pageserver_nodes
7859 0 : .set(locked.nodes.len() as i64);
7860 0 : metrics::METRICS_REGISTRY
7861 0 : .metrics_group
7862 0 : .storage_controller_https_pageserver_nodes
7863 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7864 :
7865 0 : match registration_status {
7866 : RegistrationStatus::New => {
7867 0 : tracing::info!(
7868 0 : "Registered pageserver {} ({}), now have {} pageservers",
7869 : register_req.node_id,
7870 : register_req.availability_zone_id,
7871 0 : locked.nodes.len()
7872 : );
7873 : }
7874 : RegistrationStatus::NeedUpdate => {
7875 0 : tracing::info!(
7876 0 : "Re-registered and updated node {} ({})",
7877 : register_req.node_id,
7878 : register_req.availability_zone_id,
7879 : );
7880 : }
7881 0 : _ => unreachable!("Other statuses have been processed earlier"),
7882 : }
7883 0 : Ok(())
7884 0 : }
7885 :
7886 : /// Configure in-memory and persistent state of a node as requested
7887 : ///
7888 : /// Note that this function does not trigger any immediate side effects in response
7889 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
7890 0 : async fn node_state_configure(
7891 0 : &self,
7892 0 : node_id: NodeId,
7893 0 : availability: Option<NodeAvailability>,
7894 0 : scheduling: Option<NodeSchedulingPolicy>,
7895 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
7896 0 : ) -> Result<AvailabilityTransition, ApiError> {
7897 0 : if let Some(scheduling) = scheduling {
7898 : // Scheduling is a persistent part of Node: we must write updates to the database before
7899 : // applying them in memory
7900 0 : self.persistence
7901 0 : .update_node_scheduling_policy(node_id, scheduling)
7902 0 : .await?;
7903 0 : }
7904 :
7905 : // If we're activating a node, then before setting it active we must reconcile any shard locations
7906 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
7907 : // by calling [`Self::node_activate_reconcile`]
7908 : //
7909 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
7910 : // nothing else can mutate its availability while we run.
7911 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
7912 0 : let (activate_node, availability_transition) = {
7913 0 : let locked = self.inner.read().unwrap();
7914 0 : let Some(node) = locked.nodes.get(&node_id) else {
7915 0 : return Err(ApiError::NotFound(
7916 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7917 0 : ));
7918 : };
7919 :
7920 0 : (
7921 0 : node.clone(),
7922 0 : node.get_availability_transition(input_availability),
7923 0 : )
7924 : };
7925 :
7926 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
7927 0 : self.node_activate_reconcile(activate_node, node_lock)
7928 0 : .await?;
7929 0 : }
7930 0 : availability_transition
7931 : } else {
7932 0 : AvailabilityTransition::Unchanged
7933 : };
7934 :
7935 : // Apply changes from the request to our in-memory state for the Node
7936 0 : let mut locked = self.inner.write().unwrap();
7937 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
7938 :
7939 0 : let mut new_nodes = (**nodes).clone();
7940 :
7941 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
7942 0 : return Err(ApiError::NotFound(
7943 0 : anyhow::anyhow!("Node not registered").into(),
7944 0 : ));
7945 : };
7946 :
7947 0 : if let Some(availability) = availability {
7948 0 : node.set_availability(availability);
7949 0 : }
7950 :
7951 0 : if let Some(scheduling) = scheduling {
7952 0 : node.set_scheduling(scheduling);
7953 0 : }
7954 :
7955 : // Update the scheduler, in case the elegibility of the node for new shards has changed
7956 0 : scheduler.node_upsert(node);
7957 :
7958 0 : let new_nodes = Arc::new(new_nodes);
7959 0 : locked.nodes = new_nodes;
7960 :
7961 0 : Ok(availability_transition)
7962 0 : }
7963 :
7964 : /// Handle availability transition of one node
7965 : ///
7966 : /// Note that you should first call [`Self::node_state_configure`] to update
7967 : /// the in-memory state referencing that node. If you need to handle more than one transition
7968 : /// consider using [`Self::handle_node_availability_transitions`].
7969 0 : async fn handle_node_availability_transition(
7970 0 : &self,
7971 0 : node_id: NodeId,
7972 0 : transition: AvailabilityTransition,
7973 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
7974 0 : ) -> Result<(), ApiError> {
7975 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
7976 0 : match transition {
7977 : AvailabilityTransition::ToOffline => {
7978 0 : tracing::info!("Node {} transition to offline", node_id);
7979 :
7980 0 : let mut locked = self.inner.write().unwrap();
7981 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7982 :
7983 0 : let mut tenants_affected: usize = 0;
7984 :
7985 0 : for (_tenant_id, mut schedule_context, shards) in
7986 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7987 : {
7988 0 : for tenant_shard in shards {
7989 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
7990 0 : if let Some(observed_loc) =
7991 0 : tenant_shard.observed.locations.get_mut(&node_id)
7992 0 : {
7993 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
7994 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
7995 0 : observed_loc.conf = None;
7996 0 : }
7997 :
7998 0 : if nodes.len() == 1 {
7999 : // Special case for single-node cluster: there is no point trying to reschedule
8000 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
8001 : // failures to schedule them.
8002 0 : continue;
8003 0 : }
8004 :
8005 0 : if !nodes
8006 0 : .values()
8007 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8008 : {
8009 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
8010 : // trying to reschedule since there's nowhere else to go. Without this
8011 : // branch we incorrectly detach tenants in response to node unavailability.
8012 0 : continue;
8013 0 : }
8014 :
8015 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
8016 0 : tenant_shard.sequence = tenant_shard.sequence.next();
8017 :
8018 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
8019 0 : Err(e) => {
8020 : // It is possible that some tenants will become unschedulable when too many pageservers
8021 : // go offline: in this case there isn't much we can do other than make the issue observable.
8022 : // TODO: give TenantShard a scheduling error attribute to be queried later.
8023 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
8024 : }
8025 : Ok(()) => {
8026 0 : if self
8027 0 : .maybe_reconcile_shard(
8028 0 : tenant_shard,
8029 0 : nodes,
8030 0 : ReconcilerPriority::Normal,
8031 0 : )
8032 0 : .is_some()
8033 0 : {
8034 0 : tenants_affected += 1;
8035 0 : };
8036 : }
8037 : }
8038 0 : }
8039 : }
8040 : }
8041 0 : tracing::info!(
8042 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
8043 : tenants_affected,
8044 : node_id
8045 : )
8046 : }
8047 : AvailabilityTransition::ToActive => {
8048 0 : tracing::info!("Node {} transition to active", node_id);
8049 :
8050 0 : let mut locked = self.inner.write().unwrap();
8051 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
8052 :
8053 : // When a node comes back online, we must reconcile any tenant that has a None observed
8054 : // location on the node.
8055 0 : for tenant_shard in tenants.values_mut() {
8056 : // If a reconciliation is already in progress, rely on the previous scheduling
8057 : // decision and skip triggering a new reconciliation.
8058 0 : if tenant_shard.reconciler.is_some() {
8059 0 : continue;
8060 0 : }
8061 :
8062 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
8063 0 : if observed_loc.conf.is_none() {
8064 0 : self.maybe_reconcile_shard(
8065 0 : tenant_shard,
8066 0 : nodes,
8067 0 : ReconcilerPriority::Normal,
8068 0 : );
8069 0 : }
8070 0 : }
8071 : }
8072 :
8073 : // TODO: in the background, we should balance work back onto this pageserver
8074 : }
8075 : // No action required for the intermediate unavailable state.
8076 : // When we transition into active or offline from the unavailable state,
8077 : // the correct handling above will kick in.
8078 : AvailabilityTransition::ToWarmingUpFromActive => {
8079 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
8080 : }
8081 : AvailabilityTransition::ToWarmingUpFromOffline => {
8082 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
8083 : }
8084 : AvailabilityTransition::Unchanged => {
8085 0 : tracing::debug!("Node {} no availability change during config", node_id);
8086 : }
8087 : }
8088 :
8089 0 : Ok(())
8090 0 : }
8091 :
8092 : /// Handle availability transition for multiple nodes
8093 : ///
8094 : /// Note that you should first call [`Self::node_state_configure`] for
8095 : /// all nodes being handled here for the handling to use fresh in-memory state.
8096 0 : async fn handle_node_availability_transitions(
8097 0 : &self,
8098 0 : transitions: Vec<(
8099 0 : NodeId,
8100 0 : TracingExclusiveGuard<NodeOperations>,
8101 0 : AvailabilityTransition,
8102 0 : )>,
8103 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
8104 0 : let mut errors = Vec::default();
8105 0 : for (node_id, node_lock, transition) in transitions {
8106 0 : let res = self
8107 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
8108 0 : .await;
8109 0 : if let Err(err) = res {
8110 0 : errors.push((node_id, err));
8111 0 : }
8112 : }
8113 :
8114 0 : if errors.is_empty() {
8115 0 : Ok(())
8116 : } else {
8117 0 : Err(errors)
8118 : }
8119 0 : }
8120 :
8121 0 : pub(crate) async fn node_configure(
8122 0 : &self,
8123 0 : node_id: NodeId,
8124 0 : availability: Option<NodeAvailability>,
8125 0 : scheduling: Option<NodeSchedulingPolicy>,
8126 0 : ) -> Result<(), ApiError> {
8127 0 : let node_lock =
8128 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
8129 :
8130 0 : let transition = self
8131 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
8132 0 : .await?;
8133 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
8134 0 : .await
8135 0 : }
8136 :
8137 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
8138 : /// operation for HTTP api.
8139 0 : pub(crate) async fn external_node_configure(
8140 0 : &self,
8141 0 : node_id: NodeId,
8142 0 : availability: Option<NodeAvailability>,
8143 0 : scheduling: Option<NodeSchedulingPolicy>,
8144 0 : ) -> Result<(), ApiError> {
8145 : {
8146 0 : let locked = self.inner.read().unwrap();
8147 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
8148 0 : return Err(ApiError::PreconditionFailed(
8149 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
8150 0 : ));
8151 0 : }
8152 : }
8153 :
8154 0 : self.node_configure(node_id, availability, scheduling).await
8155 0 : }
8156 :
8157 0 : pub(crate) async fn start_node_delete(
8158 0 : self: &Arc<Self>,
8159 0 : node_id: NodeId,
8160 0 : force: bool,
8161 0 : ) -> Result<(), ApiError> {
8162 0 : let (ongoing_op, node_policy, schedulable_nodes_count) = {
8163 0 : let locked = self.inner.read().unwrap();
8164 0 : let nodes = &locked.nodes;
8165 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8166 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8167 0 : ))?;
8168 0 : let schedulable_nodes_count = nodes
8169 0 : .iter()
8170 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8171 0 : .count();
8172 :
8173 : (
8174 0 : locked
8175 0 : .ongoing_operation
8176 0 : .as_ref()
8177 0 : .map(|ongoing| ongoing.operation),
8178 0 : node.get_scheduling(),
8179 0 : schedulable_nodes_count,
8180 : )
8181 : };
8182 :
8183 0 : if let Some(ongoing) = ongoing_op {
8184 0 : return Err(ApiError::PreconditionFailed(
8185 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8186 0 : ));
8187 0 : }
8188 :
8189 0 : if schedulable_nodes_count == 0 {
8190 0 : return Err(ApiError::PreconditionFailed(
8191 0 : "No other schedulable nodes to move shards".into(),
8192 0 : ));
8193 0 : }
8194 :
8195 0 : match node_policy {
8196 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
8197 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Deleting))
8198 0 : .await?;
8199 :
8200 0 : let cancel = self.cancel.child_token();
8201 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8202 0 : let policy_on_start = node_policy;
8203 :
8204 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8205 0 : operation: Operation::Delete(Delete { node_id }),
8206 0 : cancel: cancel.clone(),
8207 0 : });
8208 :
8209 0 : let span = tracing::info_span!(parent: None, "delete_node", %node_id);
8210 :
8211 0 : tokio::task::spawn(
8212 : {
8213 0 : let service = self.clone();
8214 0 : let cancel = cancel.clone();
8215 0 : async move {
8216 0 : let _gate_guard = gate_guard;
8217 :
8218 0 : scopeguard::defer! {
8219 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8220 :
8221 : if let Some(Operation::Delete(removed_delete)) = prev.map(|h| h.operation) {
8222 : assert_eq!(removed_delete.node_id, node_id, "We always take the same operation");
8223 : } else {
8224 : panic!("We always remove the same operation")
8225 : }
8226 : }
8227 :
8228 0 : tracing::info!("Delete background operation starting");
8229 0 : let res = service
8230 0 : .delete_node(node_id, policy_on_start, force, cancel)
8231 0 : .await;
8232 0 : match res {
8233 : Ok(()) => {
8234 0 : tracing::info!(
8235 0 : "Delete background operation completed successfully"
8236 : );
8237 : }
8238 : Err(OperationError::Cancelled) => {
8239 0 : tracing::info!("Delete background operation was cancelled");
8240 : }
8241 0 : Err(err) => {
8242 0 : tracing::error!(
8243 0 : "Delete background operation encountered: {err}"
8244 : )
8245 : }
8246 : }
8247 0 : }
8248 : }
8249 0 : .instrument(span),
8250 : );
8251 : }
8252 : NodeSchedulingPolicy::Deleting => {
8253 0 : return Err(ApiError::Conflict(format!(
8254 0 : "Node {node_id} has delete in progress"
8255 0 : )));
8256 : }
8257 0 : policy => {
8258 0 : return Err(ApiError::PreconditionFailed(
8259 0 : format!("Node {node_id} cannot be deleted due to {policy:?} policy").into(),
8260 0 : ));
8261 : }
8262 : }
8263 :
8264 0 : Ok(())
8265 0 : }
8266 :
8267 0 : pub(crate) async fn cancel_node_delete(
8268 0 : self: &Arc<Self>,
8269 0 : node_id: NodeId,
8270 0 : ) -> Result<(), ApiError> {
8271 : {
8272 0 : let locked = self.inner.read().unwrap();
8273 0 : let nodes = &locked.nodes;
8274 0 : nodes.get(&node_id).ok_or(ApiError::NotFound(
8275 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8276 0 : ))?;
8277 : }
8278 :
8279 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8280 0 : if let Operation::Delete(delete) = op_handler.operation {
8281 0 : if delete.node_id == node_id {
8282 0 : tracing::info!("Cancelling background delete operation for node {node_id}");
8283 0 : op_handler.cancel.cancel();
8284 0 : return Ok(());
8285 0 : }
8286 0 : }
8287 0 : }
8288 :
8289 0 : Err(ApiError::PreconditionFailed(
8290 0 : format!("Node {node_id} has no delete in progress").into(),
8291 0 : ))
8292 0 : }
8293 :
8294 0 : pub(crate) async fn start_node_drain(
8295 0 : self: &Arc<Self>,
8296 0 : node_id: NodeId,
8297 0 : ) -> Result<(), ApiError> {
8298 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
8299 0 : let locked = self.inner.read().unwrap();
8300 0 : let nodes = &locked.nodes;
8301 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8302 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8303 0 : ))?;
8304 0 : let schedulable_nodes_count = nodes
8305 0 : .iter()
8306 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8307 0 : .count();
8308 :
8309 : (
8310 0 : locked
8311 0 : .ongoing_operation
8312 0 : .as_ref()
8313 0 : .map(|ongoing| ongoing.operation),
8314 0 : node.is_available(),
8315 0 : node.get_scheduling(),
8316 0 : schedulable_nodes_count,
8317 : )
8318 : };
8319 :
8320 0 : if let Some(ongoing) = ongoing_op {
8321 0 : return Err(ApiError::PreconditionFailed(
8322 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8323 0 : ));
8324 0 : }
8325 :
8326 0 : if !node_available {
8327 0 : return Err(ApiError::ResourceUnavailable(
8328 0 : format!("Node {node_id} is currently unavailable").into(),
8329 0 : ));
8330 0 : }
8331 :
8332 0 : if schedulable_nodes_count == 0 {
8333 0 : return Err(ApiError::PreconditionFailed(
8334 0 : "No other schedulable nodes to drain to".into(),
8335 0 : ));
8336 0 : }
8337 :
8338 0 : match node_policy {
8339 : NodeSchedulingPolicy::Active => {
8340 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
8341 0 : .await?;
8342 :
8343 0 : let cancel = self.cancel.child_token();
8344 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8345 :
8346 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8347 0 : operation: Operation::Drain(Drain { node_id }),
8348 0 : cancel: cancel.clone(),
8349 0 : });
8350 :
8351 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
8352 :
8353 0 : tokio::task::spawn({
8354 0 : let service = self.clone();
8355 0 : let cancel = cancel.clone();
8356 0 : async move {
8357 0 : let _gate_guard = gate_guard;
8358 :
8359 0 : scopeguard::defer! {
8360 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8361 :
8362 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
8363 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
8364 : } else {
8365 : panic!("We always remove the same operation")
8366 : }
8367 : }
8368 :
8369 0 : tracing::info!("Drain background operation starting");
8370 0 : let res = service.drain_node(node_id, cancel).await;
8371 0 : match res {
8372 : Ok(()) => {
8373 0 : tracing::info!("Drain background operation completed successfully");
8374 : }
8375 : Err(OperationError::Cancelled) => {
8376 0 : tracing::info!("Drain background operation was cancelled");
8377 : }
8378 0 : Err(err) => {
8379 0 : tracing::error!("Drain background operation encountered: {err}")
8380 : }
8381 : }
8382 0 : }
8383 0 : }.instrument(span));
8384 : }
8385 : NodeSchedulingPolicy::Draining => {
8386 0 : return Err(ApiError::Conflict(format!(
8387 0 : "Node {node_id} has drain in progress"
8388 0 : )));
8389 : }
8390 0 : policy => {
8391 0 : return Err(ApiError::PreconditionFailed(
8392 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
8393 0 : ));
8394 : }
8395 : }
8396 :
8397 0 : Ok(())
8398 0 : }
8399 :
8400 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
8401 0 : let node_available = {
8402 0 : let locked = self.inner.read().unwrap();
8403 0 : let nodes = &locked.nodes;
8404 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8405 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8406 0 : ))?;
8407 :
8408 0 : node.is_available()
8409 : };
8410 :
8411 0 : if !node_available {
8412 0 : return Err(ApiError::ResourceUnavailable(
8413 0 : format!("Node {node_id} is currently unavailable").into(),
8414 0 : ));
8415 0 : }
8416 :
8417 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8418 0 : if let Operation::Drain(drain) = op_handler.operation {
8419 0 : if drain.node_id == node_id {
8420 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8421 0 : op_handler.cancel.cancel();
8422 0 : return Ok(());
8423 0 : }
8424 0 : }
8425 0 : }
8426 :
8427 0 : Err(ApiError::PreconditionFailed(
8428 0 : format!("Node {node_id} has no drain in progress").into(),
8429 0 : ))
8430 0 : }
8431 :
8432 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
8433 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
8434 0 : let locked = self.inner.read().unwrap();
8435 0 : let nodes = &locked.nodes;
8436 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8437 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8438 0 : ))?;
8439 :
8440 : (
8441 0 : locked
8442 0 : .ongoing_operation
8443 0 : .as_ref()
8444 0 : .map(|ongoing| ongoing.operation),
8445 0 : node.is_available(),
8446 0 : node.get_scheduling(),
8447 0 : nodes.len(),
8448 : )
8449 : };
8450 :
8451 0 : if let Some(ongoing) = ongoing_op {
8452 0 : return Err(ApiError::PreconditionFailed(
8453 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8454 0 : ));
8455 0 : }
8456 :
8457 0 : if !node_available {
8458 0 : return Err(ApiError::ResourceUnavailable(
8459 0 : format!("Node {node_id} is currently unavailable").into(),
8460 0 : ));
8461 0 : }
8462 :
8463 0 : if total_nodes_count <= 1 {
8464 0 : return Err(ApiError::PreconditionFailed(
8465 0 : "No other nodes to fill from".into(),
8466 0 : ));
8467 0 : }
8468 :
8469 0 : match node_policy {
8470 : NodeSchedulingPolicy::Active => {
8471 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
8472 0 : .await?;
8473 :
8474 0 : let cancel = self.cancel.child_token();
8475 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8476 :
8477 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8478 0 : operation: Operation::Fill(Fill { node_id }),
8479 0 : cancel: cancel.clone(),
8480 0 : });
8481 :
8482 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
8483 :
8484 0 : tokio::task::spawn({
8485 0 : let service = self.clone();
8486 0 : let cancel = cancel.clone();
8487 0 : async move {
8488 0 : let _gate_guard = gate_guard;
8489 :
8490 0 : scopeguard::defer! {
8491 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8492 :
8493 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
8494 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
8495 : } else {
8496 : panic!("We always remove the same operation")
8497 : }
8498 : }
8499 :
8500 0 : tracing::info!("Fill background operation starting");
8501 0 : let res = service.fill_node(node_id, cancel).await;
8502 0 : match res {
8503 : Ok(()) => {
8504 0 : tracing::info!("Fill background operation completed successfully");
8505 : }
8506 : Err(OperationError::Cancelled) => {
8507 0 : tracing::info!("Fill background operation was cancelled");
8508 : }
8509 0 : Err(err) => {
8510 0 : tracing::error!("Fill background operation encountered: {err}")
8511 : }
8512 : }
8513 0 : }
8514 0 : }.instrument(span));
8515 : }
8516 : NodeSchedulingPolicy::Filling => {
8517 0 : return Err(ApiError::Conflict(format!(
8518 0 : "Node {node_id} has fill in progress"
8519 0 : )));
8520 : }
8521 0 : policy => {
8522 0 : return Err(ApiError::PreconditionFailed(
8523 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
8524 0 : ));
8525 : }
8526 : }
8527 :
8528 0 : Ok(())
8529 0 : }
8530 :
8531 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
8532 0 : let node_available = {
8533 0 : let locked = self.inner.read().unwrap();
8534 0 : let nodes = &locked.nodes;
8535 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8536 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8537 0 : ))?;
8538 :
8539 0 : node.is_available()
8540 : };
8541 :
8542 0 : if !node_available {
8543 0 : return Err(ApiError::ResourceUnavailable(
8544 0 : format!("Node {node_id} is currently unavailable").into(),
8545 0 : ));
8546 0 : }
8547 :
8548 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8549 0 : if let Operation::Fill(fill) = op_handler.operation {
8550 0 : if fill.node_id == node_id {
8551 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8552 0 : op_handler.cancel.cancel();
8553 0 : return Ok(());
8554 0 : }
8555 0 : }
8556 0 : }
8557 :
8558 0 : Err(ApiError::PreconditionFailed(
8559 0 : format!("Node {node_id} has no fill in progress").into(),
8560 0 : ))
8561 0 : }
8562 :
8563 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
8564 : /// configuration
8565 0 : fn maybe_reconcile_shard(
8566 0 : &self,
8567 0 : shard: &mut TenantShard,
8568 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8569 0 : priority: ReconcilerPriority,
8570 0 : ) -> Option<ReconcilerWaiter> {
8571 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
8572 0 : }
8573 :
8574 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
8575 0 : fn get_reconciler_units(
8576 0 : &self,
8577 0 : priority: ReconcilerPriority,
8578 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
8579 0 : let units = match priority {
8580 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
8581 : ReconcilerPriority::High => {
8582 0 : match self
8583 0 : .priority_reconciler_concurrency
8584 0 : .clone()
8585 0 : .try_acquire_owned()
8586 : {
8587 0 : Ok(u) => Ok(u),
8588 : Err(TryAcquireError::NoPermits) => {
8589 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
8590 : // the normal priority semaphore.
8591 0 : self.reconciler_concurrency.clone().try_acquire_owned()
8592 : }
8593 0 : Err(e) => Err(e),
8594 : }
8595 : }
8596 : };
8597 :
8598 0 : units.map(ReconcileUnits::new)
8599 0 : }
8600 :
8601 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
8602 0 : fn maybe_configured_reconcile_shard(
8603 0 : &self,
8604 0 : shard: &mut TenantShard,
8605 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8606 0 : reconciler_config: ReconcilerConfig,
8607 0 : ) -> Option<ReconcilerWaiter> {
8608 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
8609 :
8610 0 : let reconcile_reason = match reconcile_needed {
8611 0 : ReconcileNeeded::No => return None,
8612 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
8613 0 : ReconcileNeeded::Yes(reason) => {
8614 : // Fall through to try and acquire units for spawning reconciler
8615 0 : reason
8616 : }
8617 : };
8618 :
8619 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
8620 0 : Ok(u) => u,
8621 : Err(_) => {
8622 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
8623 0 : "Concurrency limited: enqueued for reconcile later");
8624 0 : if !shard.delayed_reconcile {
8625 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
8626 0 : Err(TrySendError::Closed(_)) => {
8627 0 : // Weird mid-shutdown case?
8628 0 : }
8629 : Err(TrySendError::Full(_)) => {
8630 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
8631 0 : tracing::warn!(
8632 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
8633 : );
8634 : }
8635 0 : Ok(()) => {
8636 0 : shard.delayed_reconcile = true;
8637 0 : }
8638 : }
8639 0 : }
8640 :
8641 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
8642 : // number to advance. When this function is eventually called again and succeeds in getting units,
8643 : // it will spawn a reconciler that makes this waiter complete.
8644 0 : return Some(shard.future_reconcile_waiter());
8645 : }
8646 : };
8647 :
8648 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
8649 : // Gate closed: we're shutting down, drop out.
8650 0 : return None;
8651 : };
8652 :
8653 0 : shard.spawn_reconciler(
8654 0 : reconcile_reason,
8655 0 : &self.result_tx,
8656 0 : nodes,
8657 0 : &self.compute_hook,
8658 0 : reconciler_config,
8659 0 : &self.config,
8660 0 : &self.persistence,
8661 0 : units,
8662 0 : gate_guard,
8663 0 : &self.reconcilers_cancel,
8664 0 : self.http_client.clone(),
8665 : )
8666 0 : }
8667 :
8668 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
8669 : /// Additionally, reschedule tenants that require it.
8670 : ///
8671 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
8672 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
8673 : /// available. A return value of 0 indicates that everything is fully reconciled already.
8674 0 : fn reconcile_all(&self) -> ReconcileAllResult {
8675 0 : let mut locked = self.inner.write().unwrap();
8676 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8677 0 : let pageservers = nodes.clone();
8678 :
8679 : // This function is an efficient place to update lazy statistics, since we are walking
8680 : // all tenants.
8681 0 : let mut pending_reconciles = 0;
8682 0 : let mut stuck_reconciles = 0;
8683 0 : let mut az_violations = 0;
8684 :
8685 : // If we find any tenants to drop from memory, stash them to offload after
8686 : // we're done traversing the map of tenants.
8687 0 : let mut drop_detached_tenants = Vec::new();
8688 :
8689 0 : let mut spawned_reconciles = 0;
8690 0 : let mut has_delayed_reconciles = false;
8691 :
8692 0 : for shard in tenants.values_mut() {
8693 : // Accumulate scheduling statistics
8694 0 : if let (Some(attached), Some(preferred)) =
8695 0 : (shard.intent.get_attached(), shard.preferred_az())
8696 : {
8697 0 : let node_az = nodes
8698 0 : .get(attached)
8699 0 : .expect("Nodes exist if referenced")
8700 0 : .get_availability_zone_id();
8701 0 : if node_az != preferred {
8702 0 : az_violations += 1;
8703 0 : }
8704 0 : }
8705 :
8706 : // Skip checking if this shard is already enqueued for reconciliation
8707 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
8708 : // If there is something delayed, then return a nonzero count so that
8709 : // callers like reconcile_all_now do not incorrectly get the impression
8710 : // that the system is in a quiescent state.
8711 0 : has_delayed_reconciles = true;
8712 0 : pending_reconciles += 1;
8713 0 : continue;
8714 0 : }
8715 :
8716 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
8717 : // dirty, spawn another one
8718 0 : if self
8719 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
8720 0 : .is_some()
8721 : {
8722 0 : spawned_reconciles += 1;
8723 :
8724 0 : if shard.consecutive_reconciles_count >= MAX_CONSECUTIVE_RECONCILES {
8725 : // Count shards that are stuck, butwe still want to reconcile them.
8726 : // We don't want to consider them when deciding to run optimizations.
8727 0 : tracing::warn!(
8728 : tenant_id=%shard.tenant_shard_id.tenant_id,
8729 0 : shard_id=%shard.tenant_shard_id.shard_slug(),
8730 0 : "Shard reconciliation is stuck: {} consecutive launches",
8731 : shard.consecutive_reconciles_count
8732 : );
8733 0 : stuck_reconciles += 1;
8734 0 : }
8735 : } else {
8736 0 : if shard.delayed_reconcile {
8737 0 : // Shard wanted to reconcile but for some reason couldn't.
8738 0 : pending_reconciles += 1;
8739 0 : }
8740 :
8741 : // Reset the counter when we don't need to launch a reconcile.
8742 0 : shard.consecutive_reconciles_count = 0;
8743 : }
8744 : // If this tenant is detached, try dropping it from memory. This is usually done
8745 : // proactively in [`Self::process_results`], but we do it here to handle the edge
8746 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
8747 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
8748 0 : && shard.policy == PlacementPolicy::Detached
8749 : {
8750 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
8751 0 : shard.tenant_shard_id.tenant_id,
8752 0 : TenantOperations::DropDetached,
8753 0 : ) {
8754 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
8755 0 : }
8756 0 : }
8757 : }
8758 :
8759 : // Some metrics are calculated from SchedulerNode state, update these periodically
8760 0 : scheduler.update_metrics();
8761 :
8762 : // Process any deferred tenant drops
8763 0 : for (tenant_id, guard) in drop_detached_tenants {
8764 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
8765 0 : }
8766 :
8767 0 : metrics::METRICS_REGISTRY
8768 0 : .metrics_group
8769 0 : .storage_controller_schedule_az_violation
8770 0 : .set(az_violations as i64);
8771 :
8772 0 : metrics::METRICS_REGISTRY
8773 0 : .metrics_group
8774 0 : .storage_controller_pending_reconciles
8775 0 : .set(pending_reconciles as i64);
8776 :
8777 0 : metrics::METRICS_REGISTRY
8778 0 : .metrics_group
8779 0 : .storage_controller_stuck_reconciles
8780 0 : .set(stuck_reconciles as i64);
8781 :
8782 0 : ReconcileAllResult::new(spawned_reconciles, stuck_reconciles, has_delayed_reconciles)
8783 0 : }
8784 :
8785 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
8786 : /// could be scheduled somewhere better:
8787 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
8788 : /// * e.g. after a node fails then recovers, to move some work back to it
8789 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
8790 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
8791 : /// we did the split, but are probably better placed elsewhere.
8792 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
8793 : /// * e.g. after a shard split, some locations will be on the same node (where the split
8794 : /// happened), and will probably be better placed elsewhere.
8795 : ///
8796 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
8797 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
8798 : /// according to those same soft constraints.
8799 0 : async fn optimize_all(&self) -> usize {
8800 : // Limit on how many shards' optmizations each call to this function will execute. Combined
8801 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
8802 : // trickle of optimizations in the background, rather than executing a large number in parallel
8803 : // when a change occurs.
8804 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
8805 :
8806 : // Synchronous prepare: scan shards for possible scheduling optimizations
8807 0 : let candidate_work = self.optimize_all_plan();
8808 0 : let candidate_work_len = candidate_work.len();
8809 :
8810 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
8811 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
8812 :
8813 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
8814 :
8815 : // Synchronous apply: update the shards' intent states according to validated optimisations
8816 0 : let mut reconciles_spawned = 0;
8817 0 : let mut optimizations_applied = 0;
8818 0 : let mut locked = self.inner.write().unwrap();
8819 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8820 0 : for (tenant_shard_id, optimization) in validated_work {
8821 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
8822 : // Shard was dropped between planning and execution;
8823 0 : continue;
8824 : };
8825 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
8826 0 : if shard.apply_optimization(scheduler, optimization) {
8827 0 : optimizations_applied += 1;
8828 0 : if self
8829 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
8830 0 : .is_some()
8831 0 : {
8832 0 : reconciles_spawned += 1;
8833 0 : }
8834 0 : }
8835 :
8836 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
8837 0 : break;
8838 0 : }
8839 : }
8840 :
8841 0 : if was_work_filtered {
8842 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
8843 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
8844 0 : // as these validations start passing.
8845 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
8846 0 : }
8847 :
8848 0 : reconciles_spawned
8849 0 : }
8850 :
8851 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
8852 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
8853 : // this higher than the execution limit gives us a chance to execute some work even if the first
8854 : // few optimizations we find are not ready.
8855 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
8856 :
8857 0 : let mut work = Vec::new();
8858 0 : let mut locked = self.inner.write().unwrap();
8859 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
8860 :
8861 : // We are going to plan a bunch of optimisations before applying any of them, so the
8862 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
8863 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
8864 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
8865 : // to ignore the utilisation component of the score.
8866 :
8867 0 : for (_tenant_id, schedule_context, shards) in
8868 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Speculative)
8869 : {
8870 0 : for shard in shards {
8871 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
8872 0 : break;
8873 0 : }
8874 0 : match shard.get_scheduling_policy() {
8875 0 : ShardSchedulingPolicy::Active => {
8876 0 : // Ok to do optimization
8877 0 : }
8878 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
8879 0 : // Ok to do optimization: we are executing a graceful migration that
8880 0 : // has set preferred_node
8881 0 : }
8882 : ShardSchedulingPolicy::Essential
8883 : | ShardSchedulingPolicy::Pause
8884 : | ShardSchedulingPolicy::Stop => {
8885 : // Policy prevents optimizing this shard.
8886 0 : continue;
8887 : }
8888 : }
8889 :
8890 0 : if !matches!(shard.splitting, SplitState::Idle)
8891 0 : || matches!(shard.policy, PlacementPolicy::Detached)
8892 0 : || shard.reconciler.is_some()
8893 : {
8894 : // Do not start any optimizations while another change to the tenant is ongoing: this
8895 : // is not necessary for correctness, but simplifies operations and implicitly throttles
8896 : // optimization changes to happen in a "trickle" over time.
8897 0 : continue;
8898 0 : }
8899 :
8900 : // Fast path: we may quickly identify shards that don't have any possible optimisations
8901 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
8902 0 : if cfg!(feature = "testing") {
8903 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
8904 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
8905 : // panic in prod if we hit this, or spend cycles on it in prod.
8906 0 : assert!(
8907 0 : shard
8908 0 : .optimize_attachment(scheduler, &schedule_context)
8909 0 : .is_none()
8910 : );
8911 0 : assert!(
8912 0 : shard
8913 0 : .optimize_secondary(scheduler, &schedule_context)
8914 0 : .is_none()
8915 : );
8916 0 : }
8917 0 : continue;
8918 0 : }
8919 :
8920 0 : if let Some(optimization) =
8921 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
8922 : // its primary location based on soft constraints, cut it over.
8923 0 : shard.optimize_attachment(scheduler, &schedule_context)
8924 : {
8925 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
8926 0 : work.push((shard.tenant_shard_id, optimization));
8927 0 : break;
8928 0 : } else if let Some(optimization) =
8929 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
8930 : // better placed on another node, based on ScheduleContext, then adjust it. This
8931 : // covers cases like after a shard split, where we might have too many shards
8932 : // in the same tenant with secondary locations on the node where they originally split.
8933 0 : shard.optimize_secondary(scheduler, &schedule_context)
8934 : {
8935 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
8936 0 : work.push((shard.tenant_shard_id, optimization));
8937 0 : break;
8938 0 : }
8939 : }
8940 : }
8941 :
8942 0 : work
8943 0 : }
8944 :
8945 0 : async fn optimize_all_validate(
8946 0 : &self,
8947 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
8948 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
8949 : // Take a clone of the node map to use outside the lock in async validation phase
8950 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
8951 :
8952 0 : let mut want_secondary_status = Vec::new();
8953 :
8954 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
8955 : // check that the state of locations is acceptable to run the optimization, such as
8956 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
8957 : // in a live migration.
8958 0 : let mut validated_work = Vec::new();
8959 0 : for (tenant_shard_id, optimization) in candidate_work {
8960 0 : match optimization.action {
8961 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
8962 : old_attached_node_id: _,
8963 0 : new_attached_node_id,
8964 : }) => {
8965 0 : match validation_nodes.get(&new_attached_node_id) {
8966 0 : None => {
8967 0 : // Node was dropped between planning and validation
8968 0 : }
8969 0 : Some(node) => {
8970 0 : if !node.is_available() {
8971 0 : tracing::info!(
8972 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
8973 : );
8974 0 : } else {
8975 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
8976 0 : // remote API requests concurrently.
8977 0 : want_secondary_status.push((
8978 0 : tenant_shard_id,
8979 0 : node.clone(),
8980 0 : optimization,
8981 0 : ));
8982 0 : }
8983 : }
8984 : }
8985 : }
8986 : ScheduleOptimizationAction::ReplaceSecondary(_)
8987 : | ScheduleOptimizationAction::CreateSecondary(_)
8988 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
8989 : // No extra checks needed to manage secondaries: this does not interrupt client access
8990 0 : validated_work.push((tenant_shard_id, optimization))
8991 : }
8992 : };
8993 : }
8994 :
8995 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
8996 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
8997 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
8998 0 : let results = self
8999 0 : .tenant_for_shards_api(
9000 0 : want_secondary_status
9001 0 : .iter()
9002 0 : .map(|i| (i.0, i.1.clone()))
9003 0 : .collect(),
9004 0 : |tenant_shard_id, client| async move {
9005 0 : client.tenant_secondary_status(tenant_shard_id).await
9006 0 : },
9007 : 1,
9008 : 1,
9009 : SHORT_RECONCILE_TIMEOUT,
9010 0 : &self.cancel,
9011 : )
9012 0 : .await;
9013 :
9014 0 : for ((tenant_shard_id, node, optimization), (_, secondary_status)) in
9015 0 : want_secondary_status.into_iter().zip(results.into_iter())
9016 : {
9017 0 : match secondary_status {
9018 0 : Err(e) => {
9019 0 : tracing::info!(
9020 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
9021 : );
9022 : }
9023 0 : Ok(progress) => {
9024 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
9025 : // them in an optimization
9026 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
9027 :
9028 0 : if progress.heatmap_mtime.is_none()
9029 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
9030 0 : && progress.bytes_downloaded != progress.bytes_total
9031 0 : || progress.bytes_total - progress.bytes_downloaded
9032 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
9033 : {
9034 0 : tracing::info!(
9035 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
9036 : );
9037 :
9038 0 : if progress.heatmap_mtime.is_none() {
9039 : // No heatmap might mean the attached location has never uploaded one, or that
9040 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
9041 : // but fairly common in tests.
9042 0 : self.kick_secondary_download(tenant_shard_id).await;
9043 0 : }
9044 : } else {
9045 : // Location looks ready: proceed
9046 0 : tracing::info!(
9047 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
9048 : );
9049 0 : validated_work.push((tenant_shard_id, optimization))
9050 : }
9051 : }
9052 : }
9053 : }
9054 :
9055 0 : validated_work
9056 0 : }
9057 :
9058 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
9059 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
9060 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
9061 : /// we have this helper to move things along faster.
9062 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
9063 0 : if !self.config.kick_secondary_downloads {
9064 : // No-op if kick_secondary_downloads functionaliuty is not configured
9065 0 : return;
9066 0 : }
9067 :
9068 0 : let (attached_node, secondaries) = {
9069 0 : let locked = self.inner.read().unwrap();
9070 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
9071 0 : tracing::warn!(
9072 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
9073 : );
9074 0 : return;
9075 : };
9076 :
9077 0 : let Some(attached) = shard.intent.get_attached() else {
9078 0 : tracing::warn!(
9079 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
9080 : );
9081 0 : return;
9082 : };
9083 :
9084 0 : let secondaries = shard
9085 0 : .intent
9086 0 : .get_secondary()
9087 0 : .iter()
9088 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
9089 0 : .collect::<Vec<_>>();
9090 :
9091 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
9092 : };
9093 :
9094 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
9095 : // a 'kick' to let scheduling optimisation run more promptly.
9096 0 : match attached_node
9097 0 : .with_client_retries(
9098 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
9099 0 : &self.http_client,
9100 0 : &self.config.pageserver_jwt_token,
9101 : 3,
9102 : 10,
9103 : SHORT_RECONCILE_TIMEOUT,
9104 0 : &self.cancel,
9105 : )
9106 0 : .await
9107 : {
9108 0 : Some(Err(e)) => {
9109 0 : tracing::info!(
9110 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
9111 : );
9112 : }
9113 : None => {
9114 0 : tracing::info!(
9115 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
9116 : );
9117 : }
9118 : Some(Ok(_)) => {
9119 0 : tracing::info!(
9120 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
9121 : );
9122 : }
9123 : }
9124 :
9125 0 : for secondary_node in secondaries {
9126 0 : match secondary_node
9127 0 : .with_client_retries(
9128 0 : |client| async move {
9129 0 : client
9130 0 : .tenant_secondary_download(
9131 0 : tenant_shard_id,
9132 0 : Some(Duration::from_secs(1)),
9133 0 : )
9134 0 : .await
9135 0 : },
9136 0 : &self.http_client,
9137 0 : &self.config.pageserver_jwt_token,
9138 : 3,
9139 : 10,
9140 : SHORT_RECONCILE_TIMEOUT,
9141 0 : &self.cancel,
9142 : )
9143 0 : .await
9144 : {
9145 0 : Some(Err(e)) => {
9146 0 : tracing::info!(
9147 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
9148 : );
9149 : }
9150 : None => {
9151 0 : tracing::info!(
9152 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
9153 : );
9154 : }
9155 0 : Some(Ok(progress)) => {
9156 0 : tracing::info!(
9157 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
9158 : );
9159 : }
9160 : }
9161 : }
9162 0 : }
9163 :
9164 : /// Asynchronously split a tenant that's eligible for automatic splits. At most one tenant will
9165 : /// be split per call.
9166 : ///
9167 : /// Two sets of criteria are used: initial splits and size-based splits (in that order).
9168 : /// Initial splits are used to eagerly split unsharded tenants that may be performing initial
9169 : /// ingestion, since sharded tenants have significantly better ingestion throughput. Size-based
9170 : /// splits are used to bound the maximum shard size and balance out load.
9171 : ///
9172 : /// Splits are based on max_logical_size, i.e. the logical size of the largest timeline in a
9173 : /// tenant. We use this instead of the total logical size because branches will duplicate
9174 : /// logical size without actually using more storage. We could also use visible physical size,
9175 : /// but this might overestimate tenants that frequently churn branches.
9176 : ///
9177 : /// Initial splits (initial_split_threshold):
9178 : /// * Applies to tenants with 1 shard.
9179 : /// * The largest timeline (max_logical_size) exceeds initial_split_threshold.
9180 : /// * Splits into initial_split_shards.
9181 : ///
9182 : /// Size-based splits (split_threshold):
9183 : /// * Applies to all tenants.
9184 : /// * The largest timeline (max_logical_size) divided by shard count exceeds split_threshold.
9185 : /// * Splits such that max_logical_size / shard_count <= split_threshold, in powers of 2.
9186 : ///
9187 : /// Tenant shards are ordered by descending max_logical_size, first initial split candidates
9188 : /// then size-based split candidates. The first matching candidate is split.
9189 : ///
9190 : /// The shard count is clamped to max_split_shards. If a candidate is eligible for both initial
9191 : /// and size-based splits, the largest shard count will be used.
9192 : ///
9193 : /// An unsharded tenant will get DEFAULT_STRIPE_SIZE, regardless of what its ShardIdentity says.
9194 : /// A sharded tenant will retain its stripe size, as splits do not allow changing it.
9195 : ///
9196 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
9197 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
9198 : /// block all other splits.
9199 0 : async fn autosplit_tenants(self: &Arc<Self>) {
9200 : // If max_split_shards is set to 0 or 1, we can't split.
9201 0 : let max_split_shards = self.config.max_split_shards;
9202 0 : if max_split_shards <= 1 {
9203 0 : return;
9204 0 : }
9205 :
9206 : // If initial_split_shards is set to 0 or 1, disable initial splits.
9207 0 : let mut initial_split_threshold = self.config.initial_split_threshold.unwrap_or(0);
9208 0 : let initial_split_shards = self.config.initial_split_shards;
9209 0 : if initial_split_shards <= 1 {
9210 0 : initial_split_threshold = 0;
9211 0 : }
9212 :
9213 : // If no split_threshold nor initial_split_threshold, disable autosplits.
9214 0 : let split_threshold = self.config.split_threshold.unwrap_or(0);
9215 0 : if split_threshold == 0 && initial_split_threshold == 0 {
9216 0 : return;
9217 0 : }
9218 :
9219 : // Fetch split candidates in prioritized order.
9220 : //
9221 : // If initial splits are enabled, fetch eligible tenants first. We prioritize initial splits
9222 : // over size-based splits, since these are often performing initial ingestion and rely on
9223 : // splits to improve ingest throughput.
9224 0 : let mut candidates = Vec::new();
9225 :
9226 0 : if initial_split_threshold > 0 {
9227 : // Initial splits: fetch tenants with 1 shard where the logical size of the largest
9228 : // timeline exceeds the initial split threshold.
9229 0 : let initial_candidates = self
9230 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9231 0 : order_by: TenantSorting::MaxLogicalSize,
9232 0 : limit: 10,
9233 0 : where_shards_lt: Some(ShardCount(2)),
9234 0 : where_gt: Some(initial_split_threshold),
9235 0 : })
9236 0 : .await;
9237 0 : candidates.extend(initial_candidates);
9238 0 : }
9239 :
9240 0 : if split_threshold > 0 {
9241 : // Size-based splits: fetch tenants where the logical size of the largest timeline
9242 : // divided by shard count exceeds the split threshold.
9243 : //
9244 : // max_logical_size is only tracked on shard 0, and contains the total logical size
9245 : // across all shards. We have to order and filter by MaxLogicalSizePerShard, i.e.
9246 : // max_logical_size / shard_count, such that we only receive tenants that are actually
9247 : // eligible for splits. But we still use max_logical_size for later split calculations.
9248 0 : let size_candidates = self
9249 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9250 0 : order_by: TenantSorting::MaxLogicalSizePerShard,
9251 0 : limit: 10,
9252 0 : where_shards_lt: Some(ShardCount(max_split_shards)),
9253 0 : where_gt: Some(split_threshold),
9254 0 : })
9255 0 : .await;
9256 : #[cfg(feature = "testing")]
9257 0 : assert!(
9258 0 : size_candidates.iter().all(|c| c.id.is_shard_zero()),
9259 0 : "MaxLogicalSizePerShard returned non-zero shard: {size_candidates:?}",
9260 : );
9261 0 : candidates.extend(size_candidates);
9262 0 : }
9263 :
9264 : // Filter out tenants in a prohibiting scheduling modes
9265 : // and tenants with an ongoing import.
9266 : //
9267 : // Note that the import check here is oportunistic. An import might start
9268 : // after the check before we actually update [`TenantShard::splitting`].
9269 : // [`Self::tenant_shard_split`] checks the database whilst holding the exclusive
9270 : // tenant lock. Imports might take a long time, so the check here allows us
9271 : // to split something else instead of trying the same shard over and over.
9272 : {
9273 0 : let state = self.inner.read().unwrap();
9274 0 : candidates.retain(|i| {
9275 0 : let shard = state.tenants.get(&i.id);
9276 0 : match shard {
9277 0 : Some(t) => {
9278 0 : t.get_scheduling_policy() == ShardSchedulingPolicy::Active
9279 0 : && t.importing == TimelineImportState::Idle
9280 : }
9281 0 : None => false,
9282 : }
9283 0 : });
9284 : }
9285 :
9286 : // Pick the first candidate to split. This will generally always be the first one in
9287 : // candidates, but we defensively skip candidates that end up not actually splitting.
9288 0 : let Some((candidate, new_shard_count)) = candidates
9289 0 : .into_iter()
9290 0 : .filter_map(|candidate| {
9291 0 : let new_shard_count = Self::compute_split_shards(ShardSplitInputs {
9292 0 : shard_count: candidate.id.shard_count,
9293 0 : max_logical_size: candidate.max_logical_size,
9294 0 : split_threshold,
9295 0 : max_split_shards,
9296 0 : initial_split_threshold,
9297 0 : initial_split_shards,
9298 0 : });
9299 0 : new_shard_count.map(|shards| (candidate, shards.count()))
9300 0 : })
9301 0 : .next()
9302 : else {
9303 0 : debug!("no split-eligible tenants found");
9304 0 : return;
9305 : };
9306 :
9307 : // Retain the stripe size of sharded tenants, as splits don't allow changing it. Otherwise,
9308 : // use DEFAULT_STRIPE_SIZE for unsharded tenants -- their stripe size doesn't really matter,
9309 : // and if we change the default stripe size we want to use the new default rather than an
9310 : // old, persisted stripe size.
9311 0 : let new_stripe_size = match candidate.id.shard_count.count() {
9312 0 : 0 => panic!("invalid shard count 0"),
9313 0 : 1 => Some(DEFAULT_STRIPE_SIZE),
9314 0 : 2.. => None,
9315 : };
9316 :
9317 : // We spawn a task to run this, so it's exactly like some external API client requesting
9318 : // it. We don't want to block the background reconcile loop on this.
9319 0 : let old_shard_count = candidate.id.shard_count.count();
9320 0 : info!(
9321 0 : "auto-splitting tenant {old_shard_count} → {new_shard_count} shards, \
9322 0 : current size {candidate:?} (split_threshold={split_threshold} \
9323 0 : initial_split_threshold={initial_split_threshold})"
9324 : );
9325 :
9326 0 : let this = self.clone();
9327 0 : tokio::spawn(
9328 0 : async move {
9329 0 : match this
9330 0 : .tenant_shard_split(
9331 0 : candidate.id.tenant_id,
9332 0 : TenantShardSplitRequest {
9333 0 : new_shard_count,
9334 0 : new_stripe_size,
9335 0 : },
9336 0 : )
9337 0 : .await
9338 : {
9339 : Ok(_) => {
9340 0 : info!("successful auto-split {old_shard_count} → {new_shard_count} shards")
9341 : }
9342 0 : Err(err) => error!("auto-split failed: {err}"),
9343 : }
9344 0 : }
9345 0 : .instrument(info_span!("auto_split", tenant_id=%candidate.id.tenant_id)),
9346 : );
9347 0 : }
9348 :
9349 : /// Returns the number of shards to split a tenant into, or None if the tenant shouldn't split,
9350 : /// based on the total logical size of the largest timeline summed across all shards. Uses the
9351 : /// larger of size-based and initial splits, clamped to max_split_shards.
9352 : ///
9353 : /// NB: the thresholds are exclusive, since TopTenantShardsRequest uses where_gt.
9354 25 : fn compute_split_shards(inputs: ShardSplitInputs) -> Option<ShardCount> {
9355 : let ShardSplitInputs {
9356 25 : shard_count,
9357 25 : max_logical_size,
9358 25 : split_threshold,
9359 25 : max_split_shards,
9360 25 : initial_split_threshold,
9361 25 : initial_split_shards,
9362 25 : } = inputs;
9363 :
9364 25 : let mut new_shard_count: u8 = shard_count.count();
9365 :
9366 : // Size-based splits. Ensures max_logical_size / new_shard_count <= split_threshold, using
9367 : // power-of-two shard counts.
9368 : //
9369 : // If the current shard count is not a power of two, and does not exceed split_threshold,
9370 : // then we leave it alone rather than forcing a power-of-two split.
9371 25 : if split_threshold > 0
9372 18 : && max_logical_size.div_ceil(split_threshold) > shard_count.count() as u64
9373 12 : {
9374 12 : new_shard_count = max_logical_size
9375 12 : .div_ceil(split_threshold)
9376 12 : .checked_next_power_of_two()
9377 12 : .unwrap_or(u8::MAX as u64)
9378 12 : .try_into()
9379 12 : .unwrap_or(u8::MAX);
9380 13 : }
9381 :
9382 : // Initial splits. Use the larger of size-based and initial split shard counts. This only
9383 : // applies to unsharded tenants, i.e. changes to initial_split_threshold or
9384 : // initial_split_shards are not retroactive for sharded tenants.
9385 25 : if initial_split_threshold > 0
9386 14 : && shard_count.count() <= 1
9387 11 : && max_logical_size > initial_split_threshold
9388 8 : {
9389 8 : new_shard_count = new_shard_count.max(initial_split_shards);
9390 17 : }
9391 :
9392 : // Clamp to max shards.
9393 25 : new_shard_count = new_shard_count.min(max_split_shards);
9394 :
9395 : // Don't split if we're not increasing the shard count.
9396 25 : if new_shard_count <= shard_count.count() {
9397 10 : return None;
9398 15 : }
9399 :
9400 15 : Some(ShardCount(new_shard_count))
9401 25 : }
9402 :
9403 : /// Fetches the top tenant shards from every available node, in descending order of
9404 : /// max logical size. Offline nodes are skipped, and any errors from available nodes
9405 : /// will be logged and ignored.
9406 0 : async fn get_top_tenant_shards(
9407 0 : &self,
9408 0 : request: &TopTenantShardsRequest,
9409 0 : ) -> Vec<TopTenantShardItem> {
9410 0 : let nodes = self
9411 0 : .inner
9412 0 : .read()
9413 0 : .unwrap()
9414 0 : .nodes
9415 0 : .values()
9416 0 : .filter(|node| node.is_available())
9417 0 : .cloned()
9418 0 : .collect_vec();
9419 :
9420 0 : let mut futures = FuturesUnordered::new();
9421 0 : for node in nodes {
9422 0 : futures.push(async move {
9423 0 : node.with_client_retries(
9424 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
9425 0 : &self.http_client,
9426 0 : &self.config.pageserver_jwt_token,
9427 : 3,
9428 : 3,
9429 0 : Duration::from_secs(5),
9430 0 : &self.cancel,
9431 : )
9432 0 : .await
9433 0 : });
9434 : }
9435 :
9436 0 : let mut top = Vec::new();
9437 0 : while let Some(output) = futures.next().await {
9438 0 : match output {
9439 0 : Some(Ok(response)) => top.extend(response.shards),
9440 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
9441 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
9442 0 : None => {} // node is shutting down
9443 : }
9444 : }
9445 :
9446 0 : top.sort_by_key(|i| i.max_logical_size);
9447 0 : top.reverse();
9448 0 : top
9449 0 : }
9450 :
9451 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
9452 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
9453 : /// put the system into a quiescent state where future background reconciliations won't do anything.
9454 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
9455 0 : let reconcile_all_result = self.reconcile_all();
9456 0 : let mut spawned_reconciles = reconcile_all_result.spawned_reconciles;
9457 0 : if reconcile_all_result.can_run_optimizations() {
9458 : // Only optimize when we are otherwise idle
9459 0 : let optimization_reconciles = self.optimize_all().await;
9460 0 : spawned_reconciles += optimization_reconciles;
9461 0 : }
9462 :
9463 0 : let waiters = {
9464 0 : let mut waiters = Vec::new();
9465 0 : let locked = self.inner.read().unwrap();
9466 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
9467 0 : if let Some(waiter) = shard.get_waiter() {
9468 0 : waiters.push(waiter);
9469 0 : }
9470 : }
9471 0 : waiters
9472 : };
9473 :
9474 0 : let waiter_count = waiters.len();
9475 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
9476 0 : Ok(()) => {}
9477 0 : Err(e) => {
9478 0 : if let ReconcileWaitError::Failed(_, reconcile_error) = &e {
9479 0 : match **reconcile_error {
9480 : ReconcileError::Cancel
9481 0 : | ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
9482 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
9483 0 : // because some other change superceded it. We will return a nonzero number,
9484 0 : // so the caller knows they might have to call again to quiesce the system.
9485 0 : }
9486 : _ => {
9487 0 : return Err(e);
9488 : }
9489 : }
9490 : } else {
9491 0 : return Err(e);
9492 : }
9493 : }
9494 : };
9495 :
9496 0 : tracing::info!(
9497 0 : "{} reconciles in reconcile_all, {} waiters",
9498 : spawned_reconciles,
9499 : waiter_count
9500 : );
9501 :
9502 0 : Ok(std::cmp::max(waiter_count, spawned_reconciles))
9503 0 : }
9504 :
9505 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
9506 : // Cancel all on-going reconciles and wait for them to exit the gate.
9507 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
9508 0 : self.reconcilers_cancel.cancel();
9509 0 : self.reconcilers_gate.close().await;
9510 :
9511 : // Signal the background loop in [`Service::process_results`] to exit once
9512 : // it has proccessed the results from all the reconciles we cancelled earlier.
9513 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
9514 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
9515 0 : self.result_tx.closed().await;
9516 0 : }
9517 :
9518 0 : pub async fn shutdown(&self) {
9519 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
9520 0 : .await;
9521 :
9522 : // Background tasks hold gate guards: this notifies them of the cancellation and
9523 : // waits for them all to complete.
9524 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
9525 0 : self.cancel.cancel();
9526 0 : self.gate.close().await;
9527 0 : }
9528 :
9529 : /// Spot check the download lag for a secondary location of a shard.
9530 : /// Should be used as a heuristic, since it's not always precise: the
9531 : /// secondary might have not downloaded the new heat map yet and, hence,
9532 : /// is not aware of the lag.
9533 : ///
9534 : /// Returns:
9535 : /// * Ok(None) if the lag could not be determined from the status,
9536 : /// * Ok(Some(_)) if the lag could be determind
9537 : /// * Err on failures to query the pageserver.
9538 0 : async fn secondary_lag(
9539 0 : &self,
9540 0 : secondary: &NodeId,
9541 0 : tenant_shard_id: TenantShardId,
9542 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
9543 0 : let nodes = self.inner.read().unwrap().nodes.clone();
9544 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
9545 0 : StatusCode::NOT_FOUND,
9546 0 : format!("Node with id {secondary} not found"),
9547 0 : ))?;
9548 :
9549 0 : match node
9550 0 : .with_client_retries(
9551 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
9552 0 : &self.http_client,
9553 0 : &self.config.pageserver_jwt_token,
9554 : 1,
9555 : 3,
9556 0 : Duration::from_millis(250),
9557 0 : &self.cancel,
9558 : )
9559 0 : .await
9560 : {
9561 0 : Some(Ok(status)) => match status.heatmap_mtime {
9562 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
9563 0 : None => Ok(None),
9564 : },
9565 0 : Some(Err(e)) => Err(e),
9566 0 : None => Err(mgmt_api::Error::Cancelled),
9567 : }
9568 0 : }
9569 :
9570 : /// Drain a node by moving the shards attached to it as primaries.
9571 : /// This is a long running operation and it should run as a separate Tokio task.
9572 0 : pub(crate) async fn drain_node(
9573 0 : self: &Arc<Self>,
9574 0 : node_id: NodeId,
9575 0 : cancel: CancellationToken,
9576 0 : ) -> Result<(), OperationError> {
9577 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
9578 0 : let max_secondary_lag_bytes = self
9579 0 : .config
9580 0 : .max_secondary_lag_bytes
9581 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
9582 :
9583 : // By default, live migrations are generous about the wait time for getting
9584 : // the secondary location up to speed. When draining, give up earlier in order
9585 : // to not stall the operation when a cold secondary is encountered.
9586 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9587 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9588 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9589 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9590 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9591 0 : .build();
9592 :
9593 0 : let mut waiters = Vec::new();
9594 :
9595 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
9596 :
9597 0 : while !tid_iter.finished() {
9598 0 : if cancel.is_cancelled() {
9599 0 : match self
9600 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9601 0 : .await
9602 : {
9603 0 : Ok(()) => return Err(OperationError::Cancelled),
9604 0 : Err(err) => {
9605 0 : return Err(OperationError::FinalizeError(
9606 0 : format!(
9607 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9608 0 : )
9609 0 : .into(),
9610 0 : ));
9611 : }
9612 : }
9613 0 : }
9614 :
9615 0 : operation_utils::validate_node_state(
9616 0 : &node_id,
9617 0 : self.inner.read().unwrap().nodes.clone(),
9618 0 : NodeSchedulingPolicy::Draining,
9619 0 : )?;
9620 :
9621 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9622 0 : let tid = match tid_iter.next() {
9623 0 : Some(tid) => tid,
9624 : None => {
9625 0 : break;
9626 : }
9627 : };
9628 :
9629 0 : let tid_drain = TenantShardDrain {
9630 0 : drained_node: node_id,
9631 0 : tenant_shard_id: tid,
9632 0 : };
9633 :
9634 0 : let dest_node_id = {
9635 0 : let locked = self.inner.read().unwrap();
9636 :
9637 0 : match tid_drain
9638 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
9639 : {
9640 0 : Some(node_id) => node_id,
9641 : None => {
9642 0 : continue;
9643 : }
9644 : }
9645 : };
9646 :
9647 0 : match self.secondary_lag(&dest_node_id, tid).await {
9648 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
9649 0 : // The secondary is reasonably up to date.
9650 0 : // Migrate to it
9651 0 : }
9652 0 : Ok(Some(lag)) => {
9653 0 : tracing::info!(
9654 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9655 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
9656 : );
9657 0 : continue;
9658 : }
9659 : Ok(None) => {
9660 0 : tracing::info!(
9661 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9662 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
9663 : );
9664 0 : continue;
9665 : }
9666 0 : Err(err) => {
9667 0 : tracing::warn!(
9668 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9669 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
9670 : );
9671 0 : continue;
9672 : }
9673 : }
9674 :
9675 : {
9676 0 : let mut locked = self.inner.write().unwrap();
9677 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9678 0 : let rescheduled = tid_drain.reschedule_to_secondary(
9679 0 : dest_node_id,
9680 0 : tenants,
9681 0 : scheduler,
9682 0 : nodes,
9683 0 : )?;
9684 :
9685 0 : if let Some(tenant_shard) = rescheduled {
9686 0 : let waiter = self.maybe_configured_reconcile_shard(
9687 0 : tenant_shard,
9688 0 : nodes,
9689 0 : reconciler_config,
9690 0 : );
9691 0 : if let Some(some) = waiter {
9692 0 : waiters.push(some);
9693 0 : }
9694 0 : }
9695 : }
9696 : }
9697 :
9698 0 : waiters = self
9699 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9700 0 : .await;
9701 :
9702 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
9703 : }
9704 :
9705 0 : while !waiters.is_empty() {
9706 0 : if cancel.is_cancelled() {
9707 0 : match self
9708 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9709 0 : .await
9710 : {
9711 0 : Ok(()) => return Err(OperationError::Cancelled),
9712 0 : Err(err) => {
9713 0 : return Err(OperationError::FinalizeError(
9714 0 : format!(
9715 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9716 0 : )
9717 0 : .into(),
9718 0 : ));
9719 : }
9720 : }
9721 0 : }
9722 :
9723 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
9724 :
9725 0 : waiters = self
9726 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9727 0 : .await;
9728 : }
9729 :
9730 : // At this point we have done the best we could to drain shards from this node.
9731 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
9732 : // to complete the drain.
9733 0 : if let Err(err) = self
9734 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
9735 0 : .await
9736 : {
9737 : // This is not fatal. Anything that is polling the node scheduling policy to detect
9738 : // the end of the drain operations will hang, but all such places should enforce an
9739 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
9740 : // by the counterpart fill operation.
9741 0 : return Err(OperationError::FinalizeError(
9742 0 : format!(
9743 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
9744 0 : )
9745 0 : .into(),
9746 0 : ));
9747 0 : }
9748 :
9749 0 : Ok(())
9750 0 : }
9751 :
9752 : /// Create a node fill plan (pick secondaries to promote), based on:
9753 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
9754 : /// outside their home AZ, should be migrated back here.
9755 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
9756 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
9757 : /// those shards have a home AZ that doesn't match the node we're filling.
9758 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
9759 0 : let mut locked = self.inner.write().unwrap();
9760 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
9761 :
9762 0 : let node_az = nodes
9763 0 : .get(&node_id)
9764 0 : .expect("Node must exist")
9765 0 : .get_availability_zone_id()
9766 0 : .clone();
9767 :
9768 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
9769 0 : let mut plan = Vec::new();
9770 :
9771 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
9772 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
9773 :
9774 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
9775 : // conceivably come up in real life if deploying a single-AZ region intentionally.
9776 0 : let respect_azs = nodes
9777 0 : .values()
9778 0 : .map(|n| n.get_availability_zone_id())
9779 0 : .unique()
9780 0 : .count()
9781 : > 1;
9782 :
9783 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
9784 : // requires it.
9785 0 : for (tsid, tenant_shard) in tenants {
9786 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9787 : // Shard doesn't have a secondary on this node, ignore it.
9788 0 : continue;
9789 0 : }
9790 :
9791 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
9792 : // shards which belong on this node, not to promote shards whose scheduling preference
9793 : // would be on their currently attached node. So will avoid promoting shards whose
9794 : // home AZ doesn't match the AZ of the node we're filling.
9795 0 : match tenant_shard.preferred_az() {
9796 0 : _ if !respect_azs => {
9797 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9798 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9799 0 : }
9800 : }
9801 : None => {
9802 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
9803 : // will only do so if our target shard count requires it.
9804 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9805 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9806 0 : }
9807 : }
9808 0 : Some(az) if az == &node_az => {
9809 : // This shard's home AZ is equal to the node we're filling: it should
9810 : // be moved back to this node as part of filling, unless its currently
9811 : // attached location is also in its home AZ.
9812 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9813 0 : if nodes
9814 0 : .get(primary)
9815 0 : .expect("referenced node must exist")
9816 0 : .get_availability_zone_id()
9817 0 : != tenant_shard
9818 0 : .preferred_az()
9819 0 : .expect("tenant must have an AZ preference")
9820 : {
9821 0 : plan.push(*tsid)
9822 0 : }
9823 : } else {
9824 0 : plan.push(*tsid)
9825 : }
9826 : }
9827 0 : Some(_) => {
9828 0 : // This shard's home AZ is somewhere other than the node we're filling,
9829 0 : // it may not be moved back to this node as part of filling. Ignore it
9830 0 : }
9831 : }
9832 : }
9833 :
9834 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
9835 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
9836 :
9837 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
9838 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
9839 :
9840 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
9841 :
9842 0 : for (node_id, attached) in nodes_by_load {
9843 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
9844 0 : if !available {
9845 0 : continue;
9846 0 : }
9847 :
9848 0 : if plan.len() >= fill_requirement
9849 0 : || free_tids_by_node.is_empty()
9850 0 : || attached <= expected_attached
9851 : {
9852 0 : break;
9853 0 : }
9854 :
9855 0 : let can_take = attached - expected_attached;
9856 0 : let needed = fill_requirement - plan.len();
9857 0 : let mut take = std::cmp::min(can_take, needed);
9858 :
9859 0 : let mut remove_node = false;
9860 0 : while take > 0 {
9861 0 : match free_tids_by_node.get_mut(&node_id) {
9862 0 : Some(tids) => match tids.pop() {
9863 0 : Some(tid) => {
9864 0 : let max_promote_for_tenant = std::cmp::max(
9865 0 : tid.shard_count.count() as usize / locked.nodes.len(),
9866 : 1,
9867 : );
9868 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
9869 0 : if *promoted < max_promote_for_tenant {
9870 0 : plan.push(tid);
9871 0 : *promoted += 1;
9872 0 : take -= 1;
9873 0 : }
9874 : }
9875 : None => {
9876 0 : remove_node = true;
9877 0 : break;
9878 : }
9879 : },
9880 : None => {
9881 0 : break;
9882 : }
9883 : }
9884 : }
9885 :
9886 0 : if remove_node {
9887 0 : free_tids_by_node.remove(&node_id);
9888 0 : }
9889 : }
9890 :
9891 0 : plan
9892 0 : }
9893 :
9894 : /// Fill a node by promoting its secondaries until the cluster is balanced
9895 : /// with regards to attached shard counts. Note that this operation only
9896 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
9897 : /// This is a long running operation and it should run as a separate Tokio task.
9898 0 : pub(crate) async fn fill_node(
9899 0 : &self,
9900 0 : node_id: NodeId,
9901 0 : cancel: CancellationToken,
9902 0 : ) -> Result<(), OperationError> {
9903 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9904 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9905 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9906 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9907 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9908 0 : .build();
9909 :
9910 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
9911 0 : let mut waiters = Vec::new();
9912 :
9913 : // Execute the plan we've composed above. Before aplying each move from the plan,
9914 : // we validate to ensure that it has not gone stale in the meantime.
9915 0 : while !tids_to_promote.is_empty() {
9916 0 : if cancel.is_cancelled() {
9917 0 : match self
9918 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9919 0 : .await
9920 : {
9921 0 : Ok(()) => return Err(OperationError::Cancelled),
9922 0 : Err(err) => {
9923 0 : return Err(OperationError::FinalizeError(
9924 0 : format!(
9925 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9926 0 : )
9927 0 : .into(),
9928 0 : ));
9929 : }
9930 : }
9931 0 : }
9932 :
9933 : {
9934 0 : let mut locked = self.inner.write().unwrap();
9935 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9936 :
9937 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
9938 0 : format!("node {node_id} was removed").into(),
9939 0 : ))?;
9940 :
9941 0 : let current_policy = node.get_scheduling();
9942 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
9943 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
9944 : // about it
9945 0 : return Err(OperationError::NodeStateChanged(
9946 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
9947 0 : ));
9948 0 : }
9949 :
9950 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9951 0 : if let Some(tid) = tids_to_promote.pop() {
9952 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
9953 : // If the node being filled is not a secondary anymore,
9954 : // skip the promotion.
9955 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9956 0 : continue;
9957 0 : }
9958 :
9959 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
9960 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
9961 0 : Err(e) => {
9962 0 : tracing::warn!(
9963 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9964 0 : "Scheduling error when filling pageserver {} : {e}", node_id
9965 : );
9966 : }
9967 : Ok(()) => {
9968 0 : tracing::info!(
9969 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9970 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
9971 : node_id,
9972 : previously_attached_to,
9973 : node_id
9974 : );
9975 :
9976 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
9977 0 : tenant_shard,
9978 0 : nodes,
9979 0 : reconciler_config,
9980 0 : ) {
9981 0 : waiters.push(waiter);
9982 0 : }
9983 : }
9984 : }
9985 0 : }
9986 : } else {
9987 0 : break;
9988 : }
9989 : }
9990 : }
9991 :
9992 0 : waiters = self
9993 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9994 0 : .await;
9995 : }
9996 :
9997 0 : while !waiters.is_empty() {
9998 0 : if cancel.is_cancelled() {
9999 0 : match self
10000 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
10001 0 : .await
10002 : {
10003 0 : Ok(()) => return Err(OperationError::Cancelled),
10004 0 : Err(err) => {
10005 0 : return Err(OperationError::FinalizeError(
10006 0 : format!(
10007 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
10008 0 : )
10009 0 : .into(),
10010 0 : ));
10011 : }
10012 : }
10013 0 : }
10014 :
10015 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
10016 :
10017 0 : waiters = self
10018 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
10019 0 : .await;
10020 : }
10021 :
10022 0 : if let Err(err) = self
10023 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
10024 0 : .await
10025 : {
10026 : // This isn't a huge issue since the filling process starts upon request. However, it
10027 : // will prevent the next drain from starting. The only case in which this can fail
10028 : // is database unavailability. Such a case will require manual intervention.
10029 0 : return Err(OperationError::FinalizeError(
10030 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
10031 0 : .into(),
10032 0 : ));
10033 0 : }
10034 :
10035 0 : Ok(())
10036 0 : }
10037 :
10038 : /// Updates scrubber metadata health check results.
10039 0 : pub(crate) async fn metadata_health_update(
10040 0 : &self,
10041 0 : update_req: MetadataHealthUpdateRequest,
10042 0 : ) -> Result<(), ApiError> {
10043 0 : let now = chrono::offset::Utc::now();
10044 0 : let (healthy_records, unhealthy_records) = {
10045 0 : let locked = self.inner.read().unwrap();
10046 0 : let healthy_records = update_req
10047 0 : .healthy_tenant_shards
10048 0 : .into_iter()
10049 : // Retain only health records associated with tenant shards managed by storage controller.
10050 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
10051 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
10052 0 : .collect();
10053 0 : let unhealthy_records = update_req
10054 0 : .unhealthy_tenant_shards
10055 0 : .into_iter()
10056 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
10057 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
10058 0 : .collect();
10059 :
10060 0 : (healthy_records, unhealthy_records)
10061 : };
10062 :
10063 0 : self.persistence
10064 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
10065 0 : .await?;
10066 0 : Ok(())
10067 0 : }
10068 :
10069 : /// Lists the tenant shards that has unhealthy metadata status.
10070 0 : pub(crate) async fn metadata_health_list_unhealthy(
10071 0 : &self,
10072 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
10073 0 : let result = self
10074 0 : .persistence
10075 0 : .list_unhealthy_metadata_health_records()
10076 0 : .await?
10077 0 : .iter()
10078 0 : .map(|p| p.get_tenant_shard_id().unwrap())
10079 0 : .collect();
10080 :
10081 0 : Ok(result)
10082 0 : }
10083 :
10084 : /// Lists the tenant shards that have not been scrubbed for some duration.
10085 0 : pub(crate) async fn metadata_health_list_outdated(
10086 0 : &self,
10087 0 : not_scrubbed_for: Duration,
10088 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
10089 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
10090 0 : let result = self
10091 0 : .persistence
10092 0 : .list_outdated_metadata_health_records(earlier)
10093 0 : .await?
10094 0 : .into_iter()
10095 0 : .map(|record| record.into())
10096 0 : .collect();
10097 0 : Ok(result)
10098 0 : }
10099 :
10100 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
10101 0 : self.inner.read().unwrap().get_leadership_status()
10102 0 : }
10103 :
10104 : /// Handler for step down requests
10105 : ///
10106 : /// Step down runs in separate task since once it's called it should
10107 : /// be driven to completion. Subsequent requests will wait on the same
10108 : /// step down task.
10109 0 : pub(crate) async fn step_down(self: &Arc<Self>) -> GlobalObservedState {
10110 0 : let handle = self.step_down_barrier.get_or_init(|| {
10111 0 : let step_down_self = self.clone();
10112 0 : let (tx, rx) = tokio::sync::watch::channel::<Option<GlobalObservedState>>(None);
10113 0 : tokio::spawn(async move {
10114 0 : let state = step_down_self.step_down_task().await;
10115 0 : tx.send(Some(state))
10116 0 : .expect("Task Arc<Service> keeps receiver alive");
10117 0 : });
10118 :
10119 0 : rx
10120 0 : });
10121 :
10122 0 : handle
10123 0 : .clone()
10124 0 : .wait_for(|observed_state| observed_state.is_some())
10125 0 : .await
10126 0 : .expect("Task Arc<Service> keeps sender alive")
10127 0 : .deref()
10128 0 : .clone()
10129 0 : .expect("Checked above")
10130 0 : }
10131 :
10132 0 : async fn step_down_task(&self) -> GlobalObservedState {
10133 0 : tracing::info!("Received step down request from peer");
10134 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
10135 :
10136 0 : self.inner.write().unwrap().step_down();
10137 :
10138 0 : let stop_reconciliations =
10139 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown);
10140 0 : let mut stop_reconciliations = std::pin::pin!(stop_reconciliations);
10141 :
10142 0 : let started_at = Instant::now();
10143 :
10144 : // Wait for reconciliations to stop and warn if that's taking a long time
10145 : loop {
10146 0 : tokio::select! {
10147 0 : _ = &mut stop_reconciliations => {
10148 0 : tracing::info!("Reconciliations stopped, proceeding with step down");
10149 0 : break;
10150 : }
10151 0 : _ = tokio::time::sleep(Duration::from_secs(10)) => {
10152 0 : tracing::warn!(
10153 0 : elapsed_sec=%started_at.elapsed().as_secs(),
10154 0 : "Stopping reconciliations during step down is taking too long"
10155 : );
10156 : }
10157 : }
10158 : }
10159 :
10160 0 : let mut global_observed = GlobalObservedState::default();
10161 0 : let locked = self.inner.read().unwrap();
10162 0 : for (tid, tenant_shard) in locked.tenants.iter() {
10163 0 : global_observed
10164 0 : .0
10165 0 : .insert(*tid, tenant_shard.observed.clone());
10166 0 : }
10167 :
10168 0 : global_observed
10169 0 : }
10170 :
10171 0 : pub(crate) async fn update_shards_preferred_azs(
10172 0 : &self,
10173 0 : req: ShardsPreferredAzsRequest,
10174 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
10175 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
10176 0 : let updated = self
10177 0 : .persistence
10178 0 : .set_tenant_shard_preferred_azs(preferred_azs)
10179 0 : .await
10180 0 : .map_err(|err| {
10181 0 : ApiError::InternalServerError(anyhow::anyhow!(
10182 0 : "Failed to persist preferred AZs: {err}"
10183 0 : ))
10184 0 : })?;
10185 :
10186 0 : let mut updated_in_mem_and_db = Vec::default();
10187 :
10188 0 : let mut locked = self.inner.write().unwrap();
10189 0 : let state = locked.deref_mut();
10190 0 : for (tid, az_id) in updated {
10191 0 : let shard = state.tenants.get_mut(&tid);
10192 0 : if let Some(shard) = shard {
10193 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
10194 0 : updated_in_mem_and_db.push(tid);
10195 0 : }
10196 : }
10197 :
10198 0 : Ok(ShardsPreferredAzsResponse {
10199 0 : updated: updated_in_mem_and_db,
10200 0 : })
10201 0 : }
10202 : }
10203 :
10204 : #[cfg(test)]
10205 : mod tests {
10206 : use super::*;
10207 :
10208 : /// Tests Service::compute_split_shards. For readability, this specifies sizes in GBs rather
10209 : /// than bytes. Note that max_logical_size is the total logical size of the largest timeline
10210 : /// summed across all shards.
10211 : #[test]
10212 1 : fn compute_split_shards() {
10213 : // Size-based split: two shards have a 500 GB timeline, which need to split into 8 shards
10214 : // that are <= 64 GB,
10215 1 : assert_eq!(
10216 1 : Service::compute_split_shards(ShardSplitInputs {
10217 1 : shard_count: ShardCount(2),
10218 1 : max_logical_size: 500,
10219 1 : split_threshold: 64,
10220 1 : max_split_shards: 16,
10221 1 : initial_split_threshold: 0,
10222 1 : initial_split_shards: 0,
10223 1 : }),
10224 : Some(ShardCount(8))
10225 : );
10226 :
10227 : // Size-based split: noop at or below threshold, fires above.
10228 1 : assert_eq!(
10229 1 : Service::compute_split_shards(ShardSplitInputs {
10230 1 : shard_count: ShardCount(2),
10231 1 : max_logical_size: 127,
10232 1 : split_threshold: 64,
10233 1 : max_split_shards: 16,
10234 1 : initial_split_threshold: 0,
10235 1 : initial_split_shards: 0,
10236 1 : }),
10237 : None,
10238 : );
10239 1 : assert_eq!(
10240 1 : Service::compute_split_shards(ShardSplitInputs {
10241 1 : shard_count: ShardCount(2),
10242 1 : max_logical_size: 128,
10243 1 : split_threshold: 64,
10244 1 : max_split_shards: 16,
10245 1 : initial_split_threshold: 0,
10246 1 : initial_split_shards: 0,
10247 1 : }),
10248 : None,
10249 : );
10250 1 : assert_eq!(
10251 1 : Service::compute_split_shards(ShardSplitInputs {
10252 1 : shard_count: ShardCount(2),
10253 1 : max_logical_size: 129,
10254 1 : split_threshold: 64,
10255 1 : max_split_shards: 16,
10256 1 : initial_split_threshold: 0,
10257 1 : initial_split_shards: 0,
10258 1 : }),
10259 : Some(ShardCount(4)),
10260 : );
10261 :
10262 : // Size-based split: clamped to max_split_shards.
10263 1 : assert_eq!(
10264 1 : Service::compute_split_shards(ShardSplitInputs {
10265 1 : shard_count: ShardCount(2),
10266 1 : max_logical_size: 10000,
10267 1 : split_threshold: 64,
10268 1 : max_split_shards: 16,
10269 1 : initial_split_threshold: 0,
10270 1 : initial_split_shards: 0,
10271 1 : }),
10272 : Some(ShardCount(16))
10273 : );
10274 :
10275 : // Size-based split: tenant already at or beyond max_split_shards is not split.
10276 1 : assert_eq!(
10277 1 : Service::compute_split_shards(ShardSplitInputs {
10278 1 : shard_count: ShardCount(16),
10279 1 : max_logical_size: 10000,
10280 1 : split_threshold: 64,
10281 1 : max_split_shards: 16,
10282 1 : initial_split_threshold: 0,
10283 1 : initial_split_shards: 0,
10284 1 : }),
10285 : None
10286 : );
10287 :
10288 1 : assert_eq!(
10289 1 : Service::compute_split_shards(ShardSplitInputs {
10290 1 : shard_count: ShardCount(32),
10291 1 : max_logical_size: 10000,
10292 1 : split_threshold: 64,
10293 1 : max_split_shards: 16,
10294 1 : initial_split_threshold: 0,
10295 1 : initial_split_shards: 0,
10296 1 : }),
10297 : None
10298 : );
10299 :
10300 : // Size-based split: a non-power-of-2 shard count is normalized to power-of-2 if it
10301 : // exceeds split_threshold (i.e. a 3-shard tenant splits into 8, not 6).
10302 1 : assert_eq!(
10303 1 : Service::compute_split_shards(ShardSplitInputs {
10304 1 : shard_count: ShardCount(3),
10305 1 : max_logical_size: 320,
10306 1 : split_threshold: 64,
10307 1 : max_split_shards: 16,
10308 1 : initial_split_threshold: 0,
10309 1 : initial_split_shards: 0,
10310 1 : }),
10311 : Some(ShardCount(8))
10312 : );
10313 :
10314 : // Size-based split: a non-power-of-2 shard count is not normalized to power-of-2 if the
10315 : // existing shards are below or at split_threshold, but splits into 4 if it exceeds it.
10316 1 : assert_eq!(
10317 1 : Service::compute_split_shards(ShardSplitInputs {
10318 1 : shard_count: ShardCount(3),
10319 1 : max_logical_size: 191,
10320 1 : split_threshold: 64,
10321 1 : max_split_shards: 16,
10322 1 : initial_split_threshold: 0,
10323 1 : initial_split_shards: 0,
10324 1 : }),
10325 : None
10326 : );
10327 1 : assert_eq!(
10328 1 : Service::compute_split_shards(ShardSplitInputs {
10329 1 : shard_count: ShardCount(3),
10330 1 : max_logical_size: 192,
10331 1 : split_threshold: 64,
10332 1 : max_split_shards: 16,
10333 1 : initial_split_threshold: 0,
10334 1 : initial_split_shards: 0,
10335 1 : }),
10336 : None
10337 : );
10338 1 : assert_eq!(
10339 1 : Service::compute_split_shards(ShardSplitInputs {
10340 1 : shard_count: ShardCount(3),
10341 1 : max_logical_size: 193,
10342 1 : split_threshold: 64,
10343 1 : max_split_shards: 16,
10344 1 : initial_split_threshold: 0,
10345 1 : initial_split_shards: 0,
10346 1 : }),
10347 : Some(ShardCount(4))
10348 : );
10349 :
10350 : // Initial split: tenant has a 10 GB timeline, split into 4 shards.
10351 1 : assert_eq!(
10352 1 : Service::compute_split_shards(ShardSplitInputs {
10353 1 : shard_count: ShardCount(1),
10354 1 : max_logical_size: 10,
10355 1 : split_threshold: 0,
10356 1 : max_split_shards: 16,
10357 1 : initial_split_threshold: 8,
10358 1 : initial_split_shards: 4,
10359 1 : }),
10360 : Some(ShardCount(4))
10361 : );
10362 :
10363 : // Initial split: 0 ShardCount is equivalent to 1.
10364 1 : assert_eq!(
10365 1 : Service::compute_split_shards(ShardSplitInputs {
10366 1 : shard_count: ShardCount(0),
10367 1 : max_logical_size: 10,
10368 1 : split_threshold: 0,
10369 1 : max_split_shards: 16,
10370 1 : initial_split_threshold: 8,
10371 1 : initial_split_shards: 4,
10372 1 : }),
10373 : Some(ShardCount(4))
10374 : );
10375 :
10376 : // Initial split: at or below threshold is noop.
10377 1 : assert_eq!(
10378 1 : Service::compute_split_shards(ShardSplitInputs {
10379 1 : shard_count: ShardCount(1),
10380 1 : max_logical_size: 7,
10381 1 : split_threshold: 0,
10382 1 : max_split_shards: 16,
10383 1 : initial_split_threshold: 8,
10384 1 : initial_split_shards: 4,
10385 1 : }),
10386 : None,
10387 : );
10388 1 : assert_eq!(
10389 1 : Service::compute_split_shards(ShardSplitInputs {
10390 1 : shard_count: ShardCount(1),
10391 1 : max_logical_size: 8,
10392 1 : split_threshold: 0,
10393 1 : max_split_shards: 16,
10394 1 : initial_split_threshold: 8,
10395 1 : initial_split_shards: 4,
10396 1 : }),
10397 : None,
10398 : );
10399 1 : assert_eq!(
10400 1 : Service::compute_split_shards(ShardSplitInputs {
10401 1 : shard_count: ShardCount(1),
10402 1 : max_logical_size: 9,
10403 1 : split_threshold: 0,
10404 1 : max_split_shards: 16,
10405 1 : initial_split_threshold: 8,
10406 1 : initial_split_shards: 4,
10407 1 : }),
10408 : Some(ShardCount(4))
10409 : );
10410 :
10411 : // Initial split: already sharded tenant is not affected, even if above threshold and below
10412 : // shard count.
10413 1 : assert_eq!(
10414 1 : Service::compute_split_shards(ShardSplitInputs {
10415 1 : shard_count: ShardCount(2),
10416 1 : max_logical_size: 20,
10417 1 : split_threshold: 0,
10418 1 : max_split_shards: 16,
10419 1 : initial_split_threshold: 8,
10420 1 : initial_split_shards: 4,
10421 1 : }),
10422 : None,
10423 : );
10424 :
10425 : // Initial split: clamped to max_shards.
10426 1 : assert_eq!(
10427 1 : Service::compute_split_shards(ShardSplitInputs {
10428 1 : shard_count: ShardCount(1),
10429 1 : max_logical_size: 10,
10430 1 : split_threshold: 0,
10431 1 : max_split_shards: 3,
10432 1 : initial_split_threshold: 8,
10433 1 : initial_split_shards: 4,
10434 1 : }),
10435 : Some(ShardCount(3)),
10436 : );
10437 :
10438 : // Initial+size split: tenant eligible for both will use the larger shard count.
10439 1 : assert_eq!(
10440 1 : Service::compute_split_shards(ShardSplitInputs {
10441 1 : shard_count: ShardCount(1),
10442 1 : max_logical_size: 10,
10443 1 : split_threshold: 64,
10444 1 : max_split_shards: 16,
10445 1 : initial_split_threshold: 8,
10446 1 : initial_split_shards: 4,
10447 1 : }),
10448 : Some(ShardCount(4)),
10449 : );
10450 1 : assert_eq!(
10451 1 : Service::compute_split_shards(ShardSplitInputs {
10452 1 : shard_count: ShardCount(1),
10453 1 : max_logical_size: 500,
10454 1 : split_threshold: 64,
10455 1 : max_split_shards: 16,
10456 1 : initial_split_threshold: 8,
10457 1 : initial_split_shards: 4,
10458 1 : }),
10459 : Some(ShardCount(8)),
10460 : );
10461 :
10462 : // Initial+size split: sharded tenant is only eligible for size-based split.
10463 1 : assert_eq!(
10464 1 : Service::compute_split_shards(ShardSplitInputs {
10465 1 : shard_count: ShardCount(2),
10466 1 : max_logical_size: 200,
10467 1 : split_threshold: 64,
10468 1 : max_split_shards: 16,
10469 1 : initial_split_threshold: 8,
10470 1 : initial_split_shards: 8,
10471 1 : }),
10472 : Some(ShardCount(4)),
10473 : );
10474 :
10475 : // Initial+size split: uses the larger shard count even with initial_split_threshold above
10476 : // split_threshold.
10477 1 : assert_eq!(
10478 1 : Service::compute_split_shards(ShardSplitInputs {
10479 1 : shard_count: ShardCount(1),
10480 1 : max_logical_size: 10,
10481 1 : split_threshold: 4,
10482 1 : max_split_shards: 16,
10483 1 : initial_split_threshold: 8,
10484 1 : initial_split_shards: 8,
10485 1 : }),
10486 : Some(ShardCount(8)),
10487 : );
10488 :
10489 : // Test backwards compatibility with production settings when initial/size-based splits were
10490 : // rolled out: a single split into 8 shards at 64 GB. Any already sharded tenants with <8
10491 : // shards will split according to split_threshold.
10492 1 : assert_eq!(
10493 1 : Service::compute_split_shards(ShardSplitInputs {
10494 1 : shard_count: ShardCount(1),
10495 1 : max_logical_size: 65,
10496 1 : split_threshold: 64,
10497 1 : max_split_shards: 8,
10498 1 : initial_split_threshold: 64,
10499 1 : initial_split_shards: 8,
10500 1 : }),
10501 : Some(ShardCount(8)),
10502 : );
10503 :
10504 1 : assert_eq!(
10505 1 : Service::compute_split_shards(ShardSplitInputs {
10506 1 : shard_count: ShardCount(1),
10507 1 : max_logical_size: 64,
10508 1 : split_threshold: 64,
10509 1 : max_split_shards: 8,
10510 1 : initial_split_threshold: 64,
10511 1 : initial_split_shards: 8,
10512 1 : }),
10513 : None,
10514 : );
10515 :
10516 1 : assert_eq!(
10517 1 : Service::compute_split_shards(ShardSplitInputs {
10518 1 : shard_count: ShardCount(2),
10519 1 : max_logical_size: 129,
10520 1 : split_threshold: 64,
10521 1 : max_split_shards: 8,
10522 1 : initial_split_threshold: 64,
10523 1 : initial_split_shards: 8,
10524 1 : }),
10525 : Some(ShardCount(4)),
10526 : );
10527 1 : }
10528 : }
|
