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 :
492 : impl From<DatabaseError> for ApiError {
493 0 : fn from(err: DatabaseError) -> ApiError {
494 0 : match err {
495 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
496 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
497 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
498 0 : ApiError::ShuttingDown
499 : }
500 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
501 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
502 : }
503 0 : DatabaseError::Cas(reason) => ApiError::Conflict(reason),
504 : }
505 0 : }
506 : }
507 :
508 : enum InitialShardScheduleOutcome {
509 : Scheduled(TenantCreateResponseShard),
510 : NotScheduled,
511 : ShardScheduleError(ScheduleError),
512 : }
513 :
514 : pub struct Service {
515 : inner: Arc<std::sync::RwLock<ServiceState>>,
516 : config: Config,
517 : persistence: Arc<Persistence>,
518 : compute_hook: Arc<ComputeHook>,
519 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
520 :
521 : heartbeater_ps: Heartbeater<Node, PageserverState>,
522 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
523 :
524 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
525 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
526 :
527 : // Locking on a tenant granularity (covers all shards in the tenant):
528 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
529 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
530 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
531 :
532 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
533 : // that transition it to/from Active.
534 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
535 :
536 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
537 : // and reconciliation on startup.
538 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
539 :
540 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
541 : // a human user might be waiting for.
542 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
543 :
544 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
545 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
546 : ///
547 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
548 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
549 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
550 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
551 :
552 : // Process shutdown will fire this token
553 : cancel: CancellationToken,
554 :
555 : // Child token of [`Service::cancel`] used by reconcilers
556 : reconcilers_cancel: CancellationToken,
557 :
558 : // Background tasks will hold this gate
559 : gate: Gate,
560 :
561 : // Reconcilers background tasks will hold this gate
562 : reconcilers_gate: Gate,
563 :
564 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
565 : /// passes, it isn't safe to do any actions that mutate tenants.
566 : pub(crate) startup_complete: Barrier,
567 :
568 : /// HTTP client with proper CA certs.
569 : http_client: reqwest::Client,
570 :
571 : /// Handle for the step down background task if one was ever requested
572 : step_down_barrier: OnceLock<tokio::sync::watch::Receiver<Option<GlobalObservedState>>>,
573 : }
574 :
575 : impl From<ReconcileWaitError> for ApiError {
576 0 : fn from(value: ReconcileWaitError) -> Self {
577 0 : match value {
578 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
579 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
580 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
581 : }
582 0 : }
583 : }
584 :
585 : impl From<OperationError> for ApiError {
586 0 : fn from(value: OperationError) -> Self {
587 0 : match value {
588 0 : OperationError::NodeStateChanged(err)
589 0 : | OperationError::FinalizeError(err)
590 0 : | OperationError::ImpossibleConstraint(err) => {
591 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
592 : }
593 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
594 : }
595 0 : }
596 : }
597 :
598 : #[allow(clippy::large_enum_variant)]
599 : enum TenantCreateOrUpdate {
600 : Create(TenantCreateRequest),
601 : Update(Vec<ShardUpdate>),
602 : }
603 :
604 : struct ShardSplitParams {
605 : old_shard_count: ShardCount,
606 : new_shard_count: ShardCount,
607 : new_stripe_size: Option<ShardStripeSize>,
608 : targets: Vec<ShardSplitTarget>,
609 : policy: PlacementPolicy,
610 : config: TenantConfig,
611 : shard_ident: ShardIdentity,
612 : preferred_az_id: Option<AvailabilityZone>,
613 : }
614 :
615 : // When preparing for a shard split, we may either choose to proceed with the split,
616 : // or find that the work is already done and return NoOp.
617 : enum ShardSplitAction {
618 : Split(Box<ShardSplitParams>),
619 : NoOp(TenantShardSplitResponse),
620 : }
621 :
622 : // A parent shard which will be split
623 : struct ShardSplitTarget {
624 : parent_id: TenantShardId,
625 : node: Node,
626 : child_ids: Vec<TenantShardId>,
627 : }
628 :
629 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
630 : /// might not be available. We therefore use a queue of abort operations processed in the background.
631 : struct TenantShardSplitAbort {
632 : tenant_id: TenantId,
633 : /// The target values from the request that failed
634 : new_shard_count: ShardCount,
635 : new_stripe_size: Option<ShardStripeSize>,
636 : /// Until this abort op is complete, no other operations may be done on the tenant
637 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
638 : /// The reconciler gate for the duration of the split operation, and any included abort.
639 : _gate: GateGuard,
640 : }
641 :
642 : #[derive(thiserror::Error, Debug)]
643 : enum TenantShardSplitAbortError {
644 : #[error(transparent)]
645 : Database(#[from] DatabaseError),
646 : #[error(transparent)]
647 : Remote(#[from] mgmt_api::Error),
648 : #[error("Unavailable")]
649 : Unavailable,
650 : }
651 :
652 : /// Inputs for computing a target shard count for a tenant.
653 : struct ShardSplitInputs {
654 : /// Current shard count.
655 : shard_count: ShardCount,
656 : /// Total size of largest timeline summed across all shards.
657 : max_logical_size: u64,
658 : /// Size-based split threshold. Zero if size-based splits are disabled.
659 : split_threshold: u64,
660 : /// Upper bound on target shards. 0 or 1 disables splits.
661 : max_split_shards: u8,
662 : /// Initial split threshold. Zero if initial splits are disabled.
663 : initial_split_threshold: u64,
664 : /// Number of shards for initial splits. 0 or 1 disables initial splits.
665 : initial_split_shards: u8,
666 : }
667 :
668 : struct ShardUpdate {
669 : tenant_shard_id: TenantShardId,
670 : placement_policy: PlacementPolicy,
671 : tenant_config: TenantConfig,
672 :
673 : /// If this is None, generation is not updated.
674 : generation: Option<Generation>,
675 :
676 : /// If this is None, scheduling policy is not updated.
677 : scheduling_policy: Option<ShardSchedulingPolicy>,
678 : }
679 :
680 : enum StopReconciliationsReason {
681 : ShuttingDown,
682 : SteppingDown,
683 : }
684 :
685 : impl std::fmt::Display for StopReconciliationsReason {
686 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
687 0 : let s = match self {
688 0 : Self::ShuttingDown => "Shutting down",
689 0 : Self::SteppingDown => "Stepping down",
690 : };
691 0 : write!(writer, "{s}")
692 0 : }
693 : }
694 :
695 : pub(crate) enum ReconcileResultRequest {
696 : ReconcileResult(ReconcileResult),
697 : Stop,
698 : }
699 :
700 : #[derive(Clone)]
701 : pub(crate) struct MutationLocation {
702 : pub(crate) node: Node,
703 : pub(crate) generation: Generation,
704 : }
705 :
706 : #[derive(Clone)]
707 : pub(crate) struct ShardMutationLocations {
708 : pub(crate) latest: MutationLocation,
709 : pub(crate) other: Vec<MutationLocation>,
710 : }
711 :
712 : #[derive(Default, Clone)]
713 : pub(crate) struct TenantMutationLocations(pub BTreeMap<TenantShardId, ShardMutationLocations>);
714 :
715 : struct ReconcileAllResult {
716 : spawned_reconciles: usize,
717 : stuck_reconciles: usize,
718 : has_delayed_reconciles: bool,
719 : }
720 :
721 : impl ReconcileAllResult {
722 0 : fn new(
723 0 : spawned_reconciles: usize,
724 0 : stuck_reconciles: usize,
725 0 : has_delayed_reconciles: bool,
726 0 : ) -> Self {
727 0 : assert!(
728 0 : spawned_reconciles >= stuck_reconciles,
729 0 : "It is impossible to have less spawned reconciles than stuck reconciles"
730 : );
731 0 : Self {
732 0 : spawned_reconciles,
733 0 : stuck_reconciles,
734 0 : has_delayed_reconciles,
735 0 : }
736 0 : }
737 :
738 : /// We can run optimizations only if we don't have any delayed reconciles and
739 : /// all spawned reconciles are also stuck reconciles.
740 0 : fn can_run_optimizations(&self) -> bool {
741 0 : !self.has_delayed_reconciles && self.spawned_reconciles == self.stuck_reconciles
742 0 : }
743 : }
744 :
745 : enum TenantIdOrShardId {
746 : TenantId(TenantId),
747 : TenantShardId(TenantShardId),
748 : }
749 :
750 : impl TenantIdOrShardId {
751 0 : fn tenant_id(&self) -> TenantId {
752 0 : match self {
753 0 : TenantIdOrShardId::TenantId(tenant_id) => *tenant_id,
754 0 : TenantIdOrShardId::TenantShardId(tenant_shard_id) => tenant_shard_id.tenant_id,
755 : }
756 0 : }
757 :
758 0 : fn matches(&self, tenant_shard_id: &TenantShardId) -> bool {
759 0 : match self {
760 0 : TenantIdOrShardId::TenantId(tenant_id) => tenant_shard_id.tenant_id == *tenant_id,
761 0 : TenantIdOrShardId::TenantShardId(this_tenant_shard_id) => {
762 0 : this_tenant_shard_id == tenant_shard_id
763 : }
764 : }
765 0 : }
766 : }
767 :
768 : impl Service {
769 0 : pub fn get_config(&self) -> &Config {
770 0 : &self.config
771 0 : }
772 :
773 0 : pub fn get_http_client(&self) -> &reqwest::Client {
774 0 : &self.http_client
775 0 : }
776 :
777 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
778 : /// view of the world, and determine which pageservers are responsive.
779 : #[instrument(skip_all)]
780 : async fn startup_reconcile(
781 : self: &Arc<Service>,
782 : current_leader: Option<ControllerPersistence>,
783 : leader_step_down_state: Option<GlobalObservedState>,
784 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
785 : Result<(), (TenantShardId, NotifyError)>,
786 : >,
787 : ) {
788 : // Startup reconciliation does I/O to other services: whether they
789 : // are responsive or not, we should aim to finish within our deadline, because:
790 : // - If we don't, a k8s readiness hook watching /ready will kill us.
791 : // - While we're waiting for startup reconciliation, we are not fully
792 : // available for end user operations like creating/deleting tenants and timelines.
793 : //
794 : // We set multiple deadlines to break up the time available between the phases of work: this is
795 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
796 : let start_at = Instant::now();
797 : let node_scan_deadline = start_at
798 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
799 : .expect("Reconcile timeout is a modest constant");
800 :
801 : let observed = if let Some(state) = leader_step_down_state {
802 : tracing::info!(
803 : "Using observed state received from leader at {}",
804 : current_leader.as_ref().unwrap().address
805 : );
806 :
807 : state
808 : } else {
809 : self.build_global_observed_state(node_scan_deadline).await
810 : };
811 :
812 : // Accumulate a list of any tenant locations that ought to be detached
813 : let mut cleanup = Vec::new();
814 :
815 : // Send initial heartbeat requests to all nodes loaded from the database
816 : let all_nodes = {
817 : let locked = self.inner.read().unwrap();
818 : locked.nodes.clone()
819 : };
820 : let (mut nodes_online, mut sks_online) =
821 : self.initial_heartbeat_round(all_nodes.keys()).await;
822 :
823 : // List of tenants for which we will attempt to notify compute of their location at startup
824 : let mut compute_notifications = Vec::new();
825 :
826 : // Populate intent and observed states for all tenants, based on reported state on pageservers
827 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
828 : let shard_count = {
829 : let mut locked = self.inner.write().unwrap();
830 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
831 :
832 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
833 : let mut new_nodes = (**nodes).clone();
834 : for (node_id, node) in new_nodes.iter_mut() {
835 : if let Some(utilization) = nodes_online.remove(node_id) {
836 : node.set_availability(NodeAvailability::Active(utilization));
837 : scheduler.node_upsert(node);
838 : }
839 : }
840 : *nodes = Arc::new(new_nodes);
841 :
842 : let mut new_sks = (**safekeepers).clone();
843 : for (node_id, node) in new_sks.iter_mut() {
844 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
845 : node.set_availability(SafekeeperState::Available {
846 : utilization,
847 : last_seen_at,
848 : });
849 : }
850 : }
851 : *safekeepers = Arc::new(new_sks);
852 :
853 : for (tenant_shard_id, observed_state) in observed.0 {
854 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
855 : for node_id in observed_state.locations.keys() {
856 : cleanup.push((tenant_shard_id, *node_id));
857 : }
858 :
859 : continue;
860 : };
861 :
862 : tenant_shard.observed = observed_state;
863 : }
864 :
865 : // Populate each tenant's intent state
866 : let mut schedule_context = ScheduleContext::default();
867 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
868 : if tenant_shard_id.shard_number == ShardNumber(0) {
869 : // Reset scheduling context each time we advance to the next Tenant
870 : schedule_context = ScheduleContext::default();
871 : }
872 :
873 : tenant_shard.intent_from_observed(scheduler);
874 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
875 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
876 : // not enough pageservers are available. The tenant may well still be available
877 : // to clients.
878 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
879 : } else {
880 : // If we're both intending and observed to be attached at a particular node, we will
881 : // emit a compute notification for this. In the case where our observed state does not
882 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
883 : if let Some(attached_at) = tenant_shard.stably_attached() {
884 : compute_notifications.push(compute_hook::ShardUpdate {
885 : tenant_shard_id: *tenant_shard_id,
886 : node_id: attached_at,
887 : stripe_size: tenant_shard.shard.stripe_size,
888 : preferred_az: tenant_shard
889 : .preferred_az()
890 0 : .map(|az| Cow::Owned(az.clone())),
891 : });
892 : }
893 : }
894 : }
895 :
896 : tenants.len()
897 : };
898 :
899 : // Before making any obeservable changes to the cluster, persist self
900 : // as leader in database and memory.
901 : let leadership = Leadership::new(
902 : self.persistence.clone(),
903 : self.config.clone(),
904 : self.cancel.child_token(),
905 : );
906 :
907 : if let Err(e) = leadership.become_leader(current_leader).await {
908 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
909 : std::process::exit(1);
910 : }
911 :
912 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
913 : let sk_schedule_requests =
914 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
915 : Ok(v) => v,
916 : Err(e) => {
917 : tracing::warn!(
918 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
919 : );
920 : // std::process::exit(1);
921 : Vec::new()
922 : }
923 : };
924 :
925 : {
926 : let mut locked = self.inner.write().unwrap();
927 : locked.become_leader();
928 :
929 : for (sk_id, _sk) in locked.safekeepers.clone().iter() {
930 : locked.safekeeper_reconcilers.start_reconciler(*sk_id, self);
931 : }
932 :
933 : locked
934 : .safekeeper_reconcilers
935 : .schedule_request_vec(sk_schedule_requests);
936 : }
937 :
938 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
939 : // generation_pageserver in the database.
940 :
941 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
942 : // will emit compute hook notifications when they reconcile.
943 : //
944 : // Ordering: our calls to notify_attach_background synchronously establish a relative order for these notifications vs. any later
945 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
946 : // calls will be correctly ordered wrt these.
947 : //
948 : // Concurrency: we call notify_attach_background for all tenants, which will create O(N) tokio tasks, but almost all of them
949 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
950 : // unit and start doing I/O.
951 : tracing::info!(
952 : "Sending {} compute notifications",
953 : compute_notifications.len()
954 : );
955 : self.compute_hook.notify_attach_background(
956 : compute_notifications,
957 : bg_compute_notify_result_tx.clone(),
958 : &self.cancel,
959 : );
960 :
961 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
962 : // which require it: under normal circumstances this should only include tenants that were in some
963 : // transient state before we restarted, or any tenants whose compute hooks failed above.
964 : tracing::info!("Checking for shards in need of reconciliation...");
965 : let reconcile_all_result = self.reconcile_all();
966 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
967 : // normal operations may proceed.
968 :
969 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
970 : // background because it does not need to complete in order to proceed with other work.
971 : if !cleanup.is_empty() {
972 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
973 : tokio::task::spawn({
974 : let cleanup_self = self.clone();
975 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
976 : });
977 : }
978 :
979 : // Reconcile the timeline imports:
980 : // 1. Mark each tenant shard of tenants with an importing timeline as importing.
981 : // 2. Finalize the completed imports in the background. This handles the case where
982 : // the previous storage controller instance shut down whilst finalizing imports.
983 : let imports = self.persistence.list_timeline_imports().await;
984 : match imports {
985 : Ok(mut imports) => {
986 : {
987 : let mut locked = self.inner.write().unwrap();
988 : for import in &imports {
989 : locked
990 : .tenants
991 : .range_mut(TenantShardId::tenant_range(import.tenant_id))
992 0 : .for_each(|(_id, shard)| {
993 0 : shard.importing = TimelineImportState::Importing
994 0 : });
995 : }
996 : }
997 :
998 0 : imports.retain(|import| import.is_complete());
999 : tokio::task::spawn({
1000 : let finalize_imports_self = self.clone();
1001 0 : async move {
1002 0 : finalize_imports_self
1003 0 : .finalize_timeline_imports(imports)
1004 0 : .await
1005 0 : }
1006 : });
1007 : }
1008 : Err(err) => {
1009 : tracing::error!("Could not retrieve completed imports from database: {err}");
1010 : }
1011 : }
1012 :
1013 : let spawned_reconciles = reconcile_all_result.spawned_reconciles;
1014 : tracing::info!(
1015 : "Startup complete, spawned {spawned_reconciles} reconciliation tasks ({shard_count} shards total)"
1016 : );
1017 : }
1018 :
1019 0 : async fn initial_heartbeat_round<'a>(
1020 0 : &self,
1021 0 : node_ids: impl Iterator<Item = &'a NodeId>,
1022 0 : ) -> (
1023 0 : HashMap<NodeId, PageserverUtilization>,
1024 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
1025 0 : ) {
1026 0 : assert!(!self.startup_complete.is_ready());
1027 :
1028 0 : let all_nodes = {
1029 0 : let locked = self.inner.read().unwrap();
1030 0 : locked.nodes.clone()
1031 : };
1032 :
1033 0 : let mut nodes_to_heartbeat = HashMap::new();
1034 0 : for node_id in node_ids {
1035 0 : match all_nodes.get(node_id) {
1036 0 : Some(node) => {
1037 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
1038 0 : }
1039 : None => {
1040 0 : tracing::warn!("Node {node_id} was removed during start-up");
1041 : }
1042 : }
1043 : }
1044 :
1045 0 : let all_sks = {
1046 0 : let locked = self.inner.read().unwrap();
1047 0 : locked.safekeepers.clone()
1048 : };
1049 :
1050 0 : tracing::info!("Sending initial heartbeats...");
1051 0 : let (res_ps, res_sk) = tokio::join!(
1052 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
1053 0 : self.heartbeater_sk.heartbeat(all_sks)
1054 : );
1055 :
1056 0 : let mut online_nodes = HashMap::new();
1057 0 : if let Ok(deltas) = res_ps {
1058 0 : for (node_id, status) in deltas.0 {
1059 0 : match status {
1060 0 : PageserverState::Available { utilization, .. } => {
1061 0 : online_nodes.insert(node_id, utilization);
1062 0 : }
1063 0 : PageserverState::Offline => {}
1064 : PageserverState::WarmingUp { .. } => {
1065 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
1066 : }
1067 : }
1068 : }
1069 0 : }
1070 :
1071 0 : let mut online_sks = HashMap::new();
1072 0 : if let Ok(deltas) = res_sk {
1073 0 : for (node_id, status) in deltas.0 {
1074 0 : match status {
1075 : SafekeeperState::Available {
1076 0 : utilization,
1077 0 : last_seen_at,
1078 0 : } => {
1079 0 : online_sks.insert(node_id, (utilization, last_seen_at));
1080 0 : }
1081 0 : SafekeeperState::Offline => {}
1082 : }
1083 : }
1084 0 : }
1085 :
1086 0 : (online_nodes, online_sks)
1087 0 : }
1088 :
1089 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
1090 : ///
1091 : /// The result includes only nodes which responded within the deadline
1092 0 : async fn scan_node_locations(
1093 0 : &self,
1094 0 : deadline: Instant,
1095 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
1096 0 : let nodes = {
1097 0 : let locked = self.inner.read().unwrap();
1098 0 : locked.nodes.clone()
1099 : };
1100 :
1101 0 : let mut node_results = HashMap::new();
1102 :
1103 0 : let mut node_list_futs = FuturesUnordered::new();
1104 :
1105 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
1106 0 : for node in nodes.values() {
1107 0 : node_list_futs.push({
1108 0 : async move {
1109 0 : tracing::info!("Scanning shards on node {node}...");
1110 0 : let timeout = Duration::from_secs(5);
1111 0 : let response = node
1112 0 : .with_client_retries(
1113 0 : |client| async move { client.list_location_config().await },
1114 0 : &self.http_client,
1115 0 : &self.config.pageserver_jwt_token,
1116 : 1,
1117 : 5,
1118 0 : timeout,
1119 0 : &self.cancel,
1120 : )
1121 0 : .await;
1122 0 : (node.get_id(), response)
1123 0 : }
1124 : });
1125 : }
1126 :
1127 : loop {
1128 0 : let (node_id, result) = tokio::select! {
1129 0 : next = node_list_futs.next() => {
1130 0 : match next {
1131 0 : Some(result) => result,
1132 : None =>{
1133 : // We got results for all our nodes
1134 0 : break;
1135 : }
1136 :
1137 : }
1138 : },
1139 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
1140 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
1141 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
1142 0 : break;
1143 : }
1144 : };
1145 :
1146 0 : let Some(list_response) = result else {
1147 0 : tracing::info!("Shutdown during startup_reconcile");
1148 0 : break;
1149 : };
1150 :
1151 0 : match list_response {
1152 0 : Err(e) => {
1153 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
1154 : }
1155 0 : Ok(listing) => {
1156 0 : node_results.insert(node_id, listing);
1157 0 : }
1158 : }
1159 : }
1160 :
1161 0 : node_results
1162 0 : }
1163 :
1164 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
1165 0 : let node_listings = self.scan_node_locations(deadline).await;
1166 0 : let mut observed = GlobalObservedState::default();
1167 :
1168 0 : for (node_id, location_confs) in node_listings {
1169 0 : tracing::info!(
1170 0 : "Received {} shard statuses from pageserver {}",
1171 0 : location_confs.tenant_shards.len(),
1172 : node_id
1173 : );
1174 :
1175 0 : for (tid, location_conf) in location_confs.tenant_shards {
1176 0 : let entry = observed.0.entry(tid).or_default();
1177 0 : entry.locations.insert(
1178 0 : node_id,
1179 0 : ObservedStateLocation {
1180 0 : conf: location_conf,
1181 0 : },
1182 0 : );
1183 0 : }
1184 : }
1185 :
1186 0 : observed
1187 0 : }
1188 :
1189 : /// Used during [`Self::startup_reconcile`] and shard splits: detach a list of unknown-to-us
1190 : /// tenants from pageservers.
1191 : ///
1192 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1193 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1194 : /// other task trying to attach it.
1195 : #[instrument(skip_all)]
1196 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1197 : let nodes = self.inner.read().unwrap().nodes.clone();
1198 :
1199 : for (tenant_shard_id, node_id) in cleanup {
1200 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1201 : let Some(node) = nodes.get(&node_id) else {
1202 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1203 : // a location to clean up on a node that has since been removed.
1204 : tracing::info!(
1205 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1206 : );
1207 : continue;
1208 : };
1209 :
1210 : if self.cancel.is_cancelled() {
1211 : break;
1212 : }
1213 :
1214 : let client = PageserverClient::new(
1215 : node.get_id(),
1216 : self.http_client.clone(),
1217 : node.base_url(),
1218 : self.config.pageserver_jwt_token.as_deref(),
1219 : );
1220 : match client
1221 : .location_config(
1222 : tenant_shard_id,
1223 : LocationConfig {
1224 : mode: LocationConfigMode::Detached,
1225 : generation: None,
1226 : secondary_conf: None,
1227 : shard_number: tenant_shard_id.shard_number.0,
1228 : shard_count: tenant_shard_id.shard_count.literal(),
1229 : shard_stripe_size: 0,
1230 : tenant_conf: models::TenantConfig::default(),
1231 : },
1232 : None,
1233 : false,
1234 : )
1235 : .await
1236 : {
1237 : Ok(()) => {
1238 : tracing::info!(
1239 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1240 : );
1241 : }
1242 : Err(e) => {
1243 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1244 : // break anything.
1245 : tracing::error!(
1246 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1247 : );
1248 : }
1249 : }
1250 : }
1251 : }
1252 :
1253 : /// Long running background task that periodically wakes up and looks for shards that need
1254 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1255 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1256 : /// for those retries.
1257 : #[instrument(skip_all)]
1258 : async fn background_reconcile(self: &Arc<Self>) {
1259 : self.startup_complete.clone().wait().await;
1260 :
1261 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1262 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1263 : while !self.reconcilers_cancel.is_cancelled() {
1264 : tokio::select! {
1265 : _ = interval.tick() => {
1266 : let reconcile_all_result = self.reconcile_all();
1267 : if reconcile_all_result.can_run_optimizations() {
1268 : // Run optimizer only when we didn't find any other work to do
1269 : self.optimize_all().await;
1270 : }
1271 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1272 : // must be responsive when new projects begin ingesting and reach the threshold.
1273 : self.autosplit_tenants().await;
1274 : }
1275 : _ = self.reconcilers_cancel.cancelled() => return
1276 : }
1277 : }
1278 : }
1279 : /// Heartbeat all storage nodes once in a while.
1280 : #[instrument(skip_all)]
1281 : async fn spawn_heartbeat_driver(self: &Arc<Self>) {
1282 : self.startup_complete.clone().wait().await;
1283 :
1284 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1285 : while !self.cancel.is_cancelled() {
1286 : tokio::select! {
1287 : _ = interval.tick() => { }
1288 : _ = self.cancel.cancelled() => return
1289 : };
1290 :
1291 : let nodes = {
1292 : let locked = self.inner.read().unwrap();
1293 : locked.nodes.clone()
1294 : };
1295 :
1296 : let safekeepers = {
1297 : let locked = self.inner.read().unwrap();
1298 : locked.safekeepers.clone()
1299 : };
1300 :
1301 : let (res_ps, res_sk) = tokio::join!(
1302 : self.heartbeater_ps.heartbeat(nodes),
1303 : self.heartbeater_sk.heartbeat(safekeepers)
1304 : );
1305 :
1306 : if let Ok(deltas) = res_ps {
1307 : let mut to_handle = Vec::default();
1308 :
1309 : for (node_id, state) in deltas.0 {
1310 : let new_availability = match state {
1311 : PageserverState::Available { utilization, .. } => {
1312 : NodeAvailability::Active(utilization)
1313 : }
1314 : PageserverState::WarmingUp { started_at } => {
1315 : NodeAvailability::WarmingUp(started_at)
1316 : }
1317 : PageserverState::Offline => {
1318 : // The node might have been placed in the WarmingUp state
1319 : // while the heartbeat round was on-going. Hence, filter out
1320 : // offline transitions for WarmingUp nodes that are still within
1321 : // their grace period.
1322 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1323 : .get_node(node_id)
1324 : .await
1325 : .as_ref()
1326 0 : .map(|n| n.get_availability())
1327 : {
1328 : let now = Instant::now();
1329 : if now - *started_at >= self.config.max_warming_up_interval {
1330 : NodeAvailability::Offline
1331 : } else {
1332 : NodeAvailability::WarmingUp(*started_at)
1333 : }
1334 : } else {
1335 : NodeAvailability::Offline
1336 : }
1337 : }
1338 : };
1339 :
1340 : let node_lock = trace_exclusive_lock(
1341 : &self.node_op_locks,
1342 : node_id,
1343 : NodeOperations::Configure,
1344 : )
1345 : .await;
1346 :
1347 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1348 :
1349 : // This is the code path for geniune availability transitions (i.e node
1350 : // goes unavailable and/or comes back online).
1351 : let res = self
1352 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1353 : .await;
1354 :
1355 : match res {
1356 : Ok(transition) => {
1357 : // Keep hold of the lock until the availability transitions
1358 : // have been handled in
1359 : // [`Service::handle_node_availability_transitions`] in order avoid
1360 : // racing with [`Service::external_node_configure`].
1361 : to_handle.push((node_id, node_lock, transition));
1362 : }
1363 : Err(ApiError::NotFound(_)) => {
1364 : // This should be rare, but legitimate since the heartbeats are done
1365 : // on a snapshot of the nodes.
1366 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1367 : }
1368 : Err(ApiError::ShuttingDown) => {
1369 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1370 : }
1371 : Err(err) => {
1372 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1373 : // becomes unavailable again, we may get an error here.
1374 : tracing::error!(
1375 : "Failed to update node state {} after heartbeat round: {}",
1376 : node_id,
1377 : err
1378 : );
1379 : }
1380 : }
1381 : }
1382 :
1383 : // We collected all the transitions above and now we handle them.
1384 : let res = self.handle_node_availability_transitions(to_handle).await;
1385 : if let Err(errs) = res {
1386 : for (node_id, err) in errs {
1387 : match err {
1388 : ApiError::NotFound(_) => {
1389 : // This should be rare, but legitimate since the heartbeats are done
1390 : // on a snapshot of the nodes.
1391 : tracing::info!(
1392 : "Node {} was not found after heartbeat round",
1393 : node_id
1394 : );
1395 : }
1396 : err => {
1397 : tracing::error!(
1398 : "Failed to handle availability transition for {} after heartbeat round: {}",
1399 : node_id,
1400 : err
1401 : );
1402 : }
1403 : }
1404 : }
1405 : }
1406 : }
1407 : if let Ok(deltas) = res_sk {
1408 : let mut to_activate = Vec::new();
1409 : {
1410 : let mut locked = self.inner.write().unwrap();
1411 : let mut safekeepers = (*locked.safekeepers).clone();
1412 :
1413 : for (id, state) in deltas.0 {
1414 : let Some(sk) = safekeepers.get_mut(&id) else {
1415 : tracing::info!(
1416 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1417 : );
1418 : continue;
1419 : };
1420 : if sk.scheduling_policy() == SkSchedulingPolicy::Activating
1421 : && let SafekeeperState::Available { .. } = state
1422 : {
1423 : to_activate.push(id);
1424 : }
1425 : sk.set_availability(state);
1426 : }
1427 : locked.safekeepers = Arc::new(safekeepers);
1428 : }
1429 : for sk_id in to_activate {
1430 : // TODO this can race with set_scheduling_policy (can create disjoint DB <-> in-memory state)
1431 : tracing::info!("Activating safekeeper {sk_id}");
1432 : match self.persistence.activate_safekeeper(sk_id.0 as i64).await {
1433 : Ok(Some(())) => {}
1434 : Ok(None) => {
1435 : tracing::info!(
1436 : "safekeeper {sk_id} has been removed from db or has different scheduling policy than active or activating"
1437 : );
1438 : }
1439 : Err(e) => {
1440 : tracing::warn!("couldn't apply activation of {sk_id} to db: {e}");
1441 : continue;
1442 : }
1443 : }
1444 : if let Err(e) = self
1445 : .set_safekeeper_scheduling_policy_in_mem(sk_id, SkSchedulingPolicy::Active)
1446 : .await
1447 : {
1448 : tracing::info!("couldn't activate safekeeper {sk_id} in memory: {e}");
1449 : continue;
1450 : }
1451 : tracing::info!("Activation of safekeeper {sk_id} done");
1452 : }
1453 : }
1454 : }
1455 : }
1456 :
1457 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1458 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1459 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1460 : /// will indicate that reconciliation is not needed.
1461 : #[instrument(skip_all, fields(
1462 : seq=%result.sequence,
1463 : tenant_id=%result.tenant_shard_id.tenant_id,
1464 : shard_id=%result.tenant_shard_id.shard_slug(),
1465 : ))]
1466 : fn process_result(&self, result: ReconcileResult) {
1467 : let mut locked = self.inner.write().unwrap();
1468 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1469 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1470 : // A reconciliation result might race with removing a tenant: drop results for
1471 : // tenants that aren't in our map.
1472 : return;
1473 : };
1474 :
1475 : // Usually generation should only be updated via this path, so the max() isn't
1476 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1477 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1478 :
1479 : // If the reconciler signals that it failed to notify compute, set this state on
1480 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1481 : tenant.pending_compute_notification = result.pending_compute_notification;
1482 :
1483 : // Let the TenantShard know it is idle.
1484 : tenant.reconcile_complete(result.sequence);
1485 :
1486 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1487 : // make to the tenant
1488 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
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 node = nodes.get(delta.node_id())?;
1492 :
1493 0 : if node.is_available() {
1494 0 : return Some(delta);
1495 0 : }
1496 :
1497 : // In case a node became unavailable concurrently with the reconcile, observed
1498 : // locations on it are now uncertain. By convention, set them to None in order
1499 : // for them to get refreshed when the node comes back online.
1500 0 : Some(ObservedStateDelta::Upsert(Box::new((
1501 0 : node.get_id(),
1502 0 : ObservedStateLocation { conf: None },
1503 0 : ))))
1504 0 : });
1505 :
1506 : match result.result {
1507 : Ok(()) => {
1508 : tenant.apply_observed_deltas(deltas);
1509 : tenant.waiter.advance(result.sequence);
1510 : }
1511 : Err(e) => {
1512 : match e {
1513 : ReconcileError::Cancel => {
1514 : tracing::info!("Reconciler was cancelled");
1515 : }
1516 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1517 : // This might be due to the reconciler getting cancelled, or it might
1518 : // be due to the `Node` being marked offline.
1519 : tracing::info!("Reconciler cancelled during pageserver API call");
1520 : }
1521 : _ => {
1522 : tracing::warn!("Reconcile error: {}", e);
1523 : }
1524 : }
1525 :
1526 : // Ordering: populate last_error before advancing error_seq,
1527 : // so that waiters will see the correct error after waiting.
1528 : tenant.set_last_error(result.sequence, e);
1529 :
1530 : // Skip deletions on reconcile failures
1531 : let upsert_deltas =
1532 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1533 : tenant.apply_observed_deltas(upsert_deltas);
1534 : }
1535 : }
1536 :
1537 : tenant.consecutive_reconciles_count = tenant.consecutive_reconciles_count.saturating_add(1);
1538 :
1539 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1540 : if tenant.policy == PlacementPolicy::Detached {
1541 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1542 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1543 : // duration of the request.
1544 : let guard = self.tenant_op_locks.try_exclusive(
1545 : tenant.tenant_shard_id.tenant_id,
1546 : TenantOperations::DropDetached,
1547 : );
1548 : if let Some(guard) = guard {
1549 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1550 : }
1551 : }
1552 :
1553 : // Maybe some other work can proceed now that this job finished.
1554 : //
1555 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1556 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1557 : if self.reconciler_concurrency.available_permits() > 0 {
1558 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1559 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1560 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1561 : shard.delayed_reconcile = false;
1562 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1563 : }
1564 :
1565 : if self.reconciler_concurrency.available_permits() == 0 {
1566 : break;
1567 : }
1568 : }
1569 : }
1570 : }
1571 :
1572 0 : async fn process_results(
1573 0 : &self,
1574 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1575 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1576 0 : Result<(), (TenantShardId, NotifyError)>,
1577 0 : >,
1578 0 : ) {
1579 : loop {
1580 : // Wait for the next result, or for cancellation
1581 0 : tokio::select! {
1582 0 : r = result_rx.recv() => {
1583 0 : match r {
1584 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1585 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1586 : }
1587 : }
1588 0 : _ = async{
1589 0 : match bg_compute_hook_result_rx.recv().await {
1590 0 : Some(result) => {
1591 0 : if let Err((tenant_shard_id, notify_error)) = result {
1592 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1593 0 : let mut locked = self.inner.write().unwrap();
1594 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1595 0 : shard.pending_compute_notification = true;
1596 0 : }
1597 :
1598 0 : }
1599 : },
1600 : None => {
1601 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1602 0 : self.cancel.cancelled().await;
1603 : }
1604 : }
1605 0 : } => {},
1606 0 : _ = self.cancel.cancelled() => {
1607 0 : break;
1608 : }
1609 : };
1610 : }
1611 0 : }
1612 :
1613 0 : async fn process_aborts(
1614 0 : &self,
1615 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1616 0 : ) {
1617 : loop {
1618 : // Wait for the next result, or for cancellation
1619 0 : let op = tokio::select! {
1620 0 : r = abort_rx.recv() => {
1621 0 : match r {
1622 0 : Some(op) => {op},
1623 0 : None => {break;}
1624 : }
1625 : }
1626 0 : _ = self.cancel.cancelled() => {
1627 0 : break;
1628 : }
1629 : };
1630 :
1631 : // Retry until shutdown: we must keep this request object alive until it is properly
1632 : // processed, as it holds a lock guard that prevents other operations trying to do things
1633 : // to the tenant while it is in a weird part-split state.
1634 0 : while !self.reconcilers_cancel.is_cancelled() {
1635 0 : match self.abort_tenant_shard_split(&op).await {
1636 0 : Ok(_) => break,
1637 0 : Err(e) => {
1638 0 : tracing::warn!(
1639 0 : "Failed to abort shard split on {}, will retry: {e}",
1640 : op.tenant_id
1641 : );
1642 :
1643 : // If a node is unavailable, we hope that it has been properly marked Offline
1644 : // when we retry, so that the abort op will succeed. If the abort op is failing
1645 : // for some other reason, we will keep retrying forever, or until a human notices
1646 : // and does something about it (either fixing a pageserver or restarting the controller).
1647 0 : tokio::time::timeout(
1648 0 : Duration::from_secs(5),
1649 0 : self.reconcilers_cancel.cancelled(),
1650 0 : )
1651 0 : .await
1652 0 : .ok();
1653 : }
1654 : }
1655 : }
1656 : }
1657 0 : }
1658 :
1659 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1660 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1661 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1662 :
1663 0 : let leadership_cancel = CancellationToken::new();
1664 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1665 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1666 :
1667 : // Apply the migrations **after** the current leader has stepped down
1668 : // (or we've given up waiting for it), but **before** reading from the
1669 : // database. The only exception is reading the current leader before
1670 : // migrating.
1671 0 : persistence.migration_run().await?;
1672 :
1673 0 : tracing::info!("Loading nodes from database...");
1674 0 : let nodes = persistence
1675 0 : .list_nodes()
1676 0 : .await?
1677 0 : .into_iter()
1678 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1679 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1680 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1681 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1682 0 : metrics::METRICS_REGISTRY
1683 0 : .metrics_group
1684 0 : .storage_controller_pageserver_nodes
1685 0 : .set(nodes.len() as i64);
1686 0 : metrics::METRICS_REGISTRY
1687 0 : .metrics_group
1688 0 : .storage_controller_https_pageserver_nodes
1689 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
1690 :
1691 0 : tracing::info!("Loading safekeepers from database...");
1692 0 : let safekeepers = persistence
1693 0 : .list_safekeepers()
1694 0 : .await?
1695 0 : .into_iter()
1696 0 : .map(|skp| {
1697 0 : Safekeeper::from_persistence(
1698 0 : skp,
1699 0 : CancellationToken::new(),
1700 0 : config.use_https_safekeeper_api,
1701 : )
1702 0 : })
1703 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1704 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1705 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1706 0 : let count_policy = |policy| {
1707 0 : safekeepers
1708 0 : .iter()
1709 0 : .filter(|sk| sk.1.scheduling_policy() == policy)
1710 0 : .count()
1711 0 : };
1712 0 : let active_sk_count = count_policy(SkSchedulingPolicy::Active);
1713 0 : let activating_sk_count = count_policy(SkSchedulingPolicy::Activating);
1714 0 : let pause_sk_count = count_policy(SkSchedulingPolicy::Pause);
1715 0 : let decom_sk_count = count_policy(SkSchedulingPolicy::Decomissioned);
1716 0 : tracing::info!(
1717 0 : "Loaded {} safekeepers from database. Active {active_sk_count}, activating {activating_sk_count}, \
1718 0 : paused {pause_sk_count}, decomissioned {decom_sk_count}.",
1719 0 : safekeepers.len()
1720 : );
1721 0 : metrics::METRICS_REGISTRY
1722 0 : .metrics_group
1723 0 : .storage_controller_safekeeper_nodes
1724 0 : .set(safekeepers.len() as i64);
1725 0 : metrics::METRICS_REGISTRY
1726 0 : .metrics_group
1727 0 : .storage_controller_https_safekeeper_nodes
1728 0 : .set(safekeepers.values().filter(|s| s.has_https_port()).count() as i64);
1729 :
1730 0 : tracing::info!("Loading shards from database...");
1731 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1732 0 : tracing::info!(
1733 0 : "Loaded {} shards from database.",
1734 0 : tenant_shard_persistence.len()
1735 : );
1736 :
1737 : // If any shard splits were in progress, reset the database state to abort them
1738 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1739 0 : HashMap::new();
1740 0 : for tsp in &mut tenant_shard_persistence {
1741 0 : let shard = tsp.get_shard_identity()?;
1742 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1743 0 : let entry = tenant_shard_count_min_max
1744 0 : .entry(tenant_shard_id.tenant_id)
1745 0 : .or_insert_with(|| (shard.count, shard.count));
1746 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1747 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1748 : }
1749 :
1750 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1751 0 : if count_min != count_max {
1752 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1753 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1754 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1755 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1756 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1757 :
1758 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1759 : // identified this tenant has having mismatching min/max counts.
1760 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1761 :
1762 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1763 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1764 : // Set idle split state on those shards that we will retain.
1765 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1766 0 : if tsp_tenant_id == tenant_id
1767 0 : && tsp.get_shard_identity().unwrap().count == count_min
1768 0 : {
1769 0 : tsp.splitting = SplitState::Idle;
1770 0 : } else if tsp_tenant_id == tenant_id {
1771 : // Leave the splitting state on the child shards: this will be used next to
1772 : // drop them.
1773 0 : tracing::info!(
1774 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1775 : );
1776 0 : }
1777 0 : });
1778 :
1779 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1780 0 : tenant_shard_persistence.retain(|tsp| {
1781 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1782 0 : || tsp.splitting == SplitState::Idle
1783 0 : });
1784 0 : }
1785 : }
1786 :
1787 0 : let mut tenants = BTreeMap::new();
1788 :
1789 0 : let mut scheduler = Scheduler::new(nodes.values());
1790 :
1791 : #[cfg(feature = "testing")]
1792 : {
1793 : use pageserver_api::controller_api::AvailabilityZone;
1794 :
1795 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1796 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1797 : // after when pageservers start up and register.
1798 0 : let mut node_ids = HashSet::new();
1799 0 : for tsp in &tenant_shard_persistence {
1800 0 : if let Some(node_id) = tsp.generation_pageserver {
1801 0 : node_ids.insert(node_id);
1802 0 : }
1803 : }
1804 0 : for node_id in node_ids {
1805 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1806 0 : let node = Node::new(
1807 0 : NodeId(node_id as u64),
1808 0 : "".to_string(),
1809 : 123,
1810 0 : None,
1811 0 : "".to_string(),
1812 : 123,
1813 0 : None,
1814 0 : None,
1815 0 : AvailabilityZone("test_az".to_string()),
1816 : false,
1817 : )
1818 0 : .unwrap();
1819 :
1820 0 : scheduler.node_upsert(&node);
1821 : }
1822 : }
1823 0 : for tsp in tenant_shard_persistence {
1824 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1825 :
1826 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1827 : // it with what we can infer: the node for which a generation was most recently issued.
1828 0 : let mut intent = IntentState::new(
1829 0 : tsp.preferred_az_id
1830 0 : .as_ref()
1831 0 : .map(|az| AvailabilityZone(az.clone())),
1832 : );
1833 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1834 : {
1835 0 : if nodes.contains_key(&generation_pageserver) {
1836 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1837 0 : } else {
1838 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1839 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1840 : // on different pageservers.
1841 0 : tracing::warn!(
1842 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1843 : );
1844 : }
1845 0 : }
1846 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1847 :
1848 0 : tenants.insert(tenant_shard_id, new_tenant);
1849 : }
1850 :
1851 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1852 :
1853 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1854 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1855 0 : tokio::sync::mpsc::channel(512);
1856 :
1857 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1858 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1859 :
1860 0 : let cancel = CancellationToken::new();
1861 0 : let reconcilers_cancel = cancel.child_token();
1862 :
1863 0 : let mut http_client = reqwest::Client::builder();
1864 : // We intentionally disable the connection pool, so every request will create its own TCP connection.
1865 : // It's especially important for heartbeaters to notice more network problems.
1866 : //
1867 : // TODO: It makes sense to use this client only in heartbeaters and create a second one with
1868 : // connection pooling for everything else. But reqwest::Client may create a connection without
1869 : // ever using it (it uses hyper's Client under the hood):
1870 : // https://github.com/hyperium/hyper-util/blob/d51318df3461d40e5f5e5ca163cb3905ac960209/src/client/legacy/client.rs#L415
1871 : //
1872 : // Because of a bug in hyper0::Connection::graceful_shutdown such connections hang during
1873 : // graceful server shutdown: https://github.com/hyperium/hyper/issues/2730
1874 : //
1875 : // The bug has been fixed in hyper v1, so keep alive may be enabled only after we migrate to hyper1.
1876 0 : http_client = http_client.pool_max_idle_per_host(0);
1877 0 : for ssl_ca_cert in &config.ssl_ca_certs {
1878 0 : http_client = http_client.add_root_certificate(ssl_ca_cert.clone());
1879 0 : }
1880 0 : let http_client = http_client.build()?;
1881 :
1882 0 : let heartbeater_ps = Heartbeater::new(
1883 0 : http_client.clone(),
1884 0 : config.pageserver_jwt_token.clone(),
1885 0 : config.max_offline_interval,
1886 0 : config.max_warming_up_interval,
1887 0 : cancel.clone(),
1888 : );
1889 :
1890 0 : let heartbeater_sk = Heartbeater::new(
1891 0 : http_client.clone(),
1892 0 : config.safekeeper_jwt_token.clone(),
1893 0 : config.max_offline_interval,
1894 0 : config.max_warming_up_interval,
1895 0 : cancel.clone(),
1896 : );
1897 :
1898 0 : let initial_leadership_status = if config.start_as_candidate {
1899 0 : LeadershipStatus::Candidate
1900 : } else {
1901 0 : LeadershipStatus::Leader
1902 : };
1903 :
1904 0 : let this = Arc::new(Self {
1905 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1906 0 : nodes,
1907 0 : safekeepers,
1908 0 : tenants,
1909 0 : scheduler,
1910 0 : delayed_reconcile_rx,
1911 0 : initial_leadership_status,
1912 0 : reconcilers_cancel.clone(),
1913 : ))),
1914 0 : config: config.clone(),
1915 0 : persistence,
1916 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())?),
1917 0 : result_tx,
1918 0 : heartbeater_ps,
1919 0 : heartbeater_sk,
1920 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1921 0 : config.reconciler_concurrency,
1922 : )),
1923 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1924 0 : config.priority_reconciler_concurrency,
1925 : )),
1926 0 : delayed_reconcile_tx,
1927 0 : abort_tx,
1928 0 : startup_complete: startup_complete.clone(),
1929 0 : cancel,
1930 0 : reconcilers_cancel,
1931 0 : gate: Gate::default(),
1932 0 : reconcilers_gate: Gate::default(),
1933 0 : tenant_op_locks: Default::default(),
1934 0 : node_op_locks: Default::default(),
1935 0 : http_client,
1936 0 : step_down_barrier: Default::default(),
1937 : });
1938 :
1939 0 : let result_task_this = this.clone();
1940 0 : tokio::task::spawn(async move {
1941 : // Block shutdown until we're done (we must respect self.cancel)
1942 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1943 0 : result_task_this
1944 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1945 0 : .await
1946 0 : }
1947 0 : });
1948 :
1949 0 : tokio::task::spawn({
1950 0 : let this = this.clone();
1951 0 : async move {
1952 : // Block shutdown until we're done (we must respect self.cancel)
1953 0 : if let Ok(_gate) = this.gate.enter() {
1954 0 : this.process_aborts(abort_rx).await
1955 0 : }
1956 0 : }
1957 : });
1958 :
1959 0 : tokio::task::spawn({
1960 0 : let this = this.clone();
1961 0 : async move {
1962 0 : if let Ok(_gate) = this.gate.enter() {
1963 : loop {
1964 0 : tokio::select! {
1965 0 : _ = this.cancel.cancelled() => {
1966 0 : break;
1967 : },
1968 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1969 : };
1970 0 : this.tenant_op_locks.housekeeping();
1971 : }
1972 0 : }
1973 0 : }
1974 : });
1975 :
1976 0 : tokio::task::spawn({
1977 0 : let this = this.clone();
1978 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1979 : // is done.
1980 0 : let startup_completion = startup_completion.clone();
1981 0 : async move {
1982 : // Block shutdown until we're done (we must respect self.cancel)
1983 0 : let Ok(_gate) = this.gate.enter() else {
1984 0 : return;
1985 : };
1986 :
1987 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1988 0 : .await;
1989 :
1990 0 : drop(startup_completion);
1991 0 : }
1992 : });
1993 :
1994 0 : tokio::task::spawn({
1995 0 : let this = this.clone();
1996 0 : let startup_complete = startup_complete.clone();
1997 0 : async move {
1998 0 : startup_complete.wait().await;
1999 0 : this.background_reconcile().await;
2000 0 : }
2001 : });
2002 :
2003 0 : tokio::task::spawn({
2004 0 : let this = this.clone();
2005 0 : let startup_complete = startup_complete.clone();
2006 0 : async move {
2007 0 : startup_complete.wait().await;
2008 0 : this.spawn_heartbeat_driver().await;
2009 0 : }
2010 : });
2011 :
2012 : // Check that there is enough safekeepers configured that we can create new timelines
2013 0 : let test_sk_res_str = match this.safekeepers_for_new_timeline().await {
2014 0 : Ok(v) => format!("Ok({v:?})"),
2015 0 : Err(v) => format!("Err({v:})"),
2016 : };
2017 0 : tracing::info!(
2018 : timeline_safekeeper_count = config.timeline_safekeeper_count,
2019 : timelines_onto_safekeepers = config.timelines_onto_safekeepers,
2020 0 : "viability test result (test timeline creation on safekeepers): {test_sk_res_str}",
2021 : );
2022 :
2023 0 : Ok(this)
2024 0 : }
2025 :
2026 0 : pub(crate) async fn attach_hook(
2027 0 : &self,
2028 0 : attach_req: AttachHookRequest,
2029 0 : ) -> anyhow::Result<AttachHookResponse> {
2030 0 : let _tenant_lock = trace_exclusive_lock(
2031 0 : &self.tenant_op_locks,
2032 0 : attach_req.tenant_shard_id.tenant_id,
2033 0 : TenantOperations::AttachHook,
2034 0 : )
2035 0 : .await;
2036 :
2037 : // This is a test hook. To enable using it on tenants that were created directly with
2038 : // the pageserver API (not via this service), we will auto-create any missing tenant
2039 : // shards with default state.
2040 0 : let insert = {
2041 0 : match self
2042 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
2043 0 : .await
2044 : {
2045 0 : Ok(_) => false,
2046 0 : Err(ApiError::NotFound(_)) => true,
2047 0 : Err(e) => return Err(e.into()),
2048 : }
2049 : };
2050 :
2051 0 : if insert {
2052 0 : let config = attach_req.config.clone().unwrap_or_default();
2053 0 : let tsp = TenantShardPersistence {
2054 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
2055 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
2056 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
2057 0 : shard_stripe_size: 0,
2058 0 : generation: attach_req.generation_override.or(Some(0)),
2059 0 : generation_pageserver: None,
2060 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
2061 0 : config: serde_json::to_string(&config).unwrap(),
2062 0 : splitting: SplitState::default(),
2063 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2064 0 : .unwrap(),
2065 0 : preferred_az_id: None,
2066 0 : };
2067 :
2068 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
2069 0 : Err(e) => match e {
2070 : DatabaseError::Query(diesel::result::Error::DatabaseError(
2071 : DatabaseErrorKind::UniqueViolation,
2072 : _,
2073 : )) => {
2074 0 : tracing::info!(
2075 0 : "Raced with another request to insert tenant {}",
2076 : attach_req.tenant_shard_id
2077 : )
2078 : }
2079 0 : _ => return Err(e.into()),
2080 : },
2081 : Ok(()) => {
2082 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
2083 :
2084 0 : let mut shard = TenantShard::new(
2085 0 : attach_req.tenant_shard_id,
2086 0 : ShardIdentity::unsharded(),
2087 0 : PlacementPolicy::Attached(0),
2088 0 : None,
2089 : );
2090 0 : shard.config = config;
2091 :
2092 0 : let mut locked = self.inner.write().unwrap();
2093 0 : locked.tenants.insert(attach_req.tenant_shard_id, shard);
2094 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
2095 : }
2096 : }
2097 0 : }
2098 :
2099 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
2100 0 : let maybe_tenant_conf = {
2101 0 : let locked = self.inner.write().unwrap();
2102 0 : locked
2103 0 : .tenants
2104 0 : .get(&attach_req.tenant_shard_id)
2105 0 : .map(|t| t.config.clone())
2106 : };
2107 :
2108 0 : match maybe_tenant_conf {
2109 0 : Some(conf) => {
2110 0 : let new_generation = self
2111 0 : .persistence
2112 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
2113 0 : .await?;
2114 :
2115 : // Persist the placement policy update. This is required
2116 : // when we reattaching a detached tenant.
2117 0 : self.persistence
2118 0 : .update_tenant_shard(
2119 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
2120 0 : Some(PlacementPolicy::Attached(0)),
2121 0 : Some(conf),
2122 0 : None,
2123 0 : None,
2124 0 : )
2125 0 : .await?;
2126 0 : Some(new_generation)
2127 : }
2128 : None => {
2129 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
2130 : }
2131 : }
2132 : } else {
2133 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
2134 0 : None
2135 : };
2136 :
2137 0 : let mut locked = self.inner.write().unwrap();
2138 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2139 :
2140 0 : let tenant_shard = tenants
2141 0 : .get_mut(&attach_req.tenant_shard_id)
2142 0 : .expect("Checked for existence above");
2143 :
2144 0 : if let Some(new_generation) = new_generation {
2145 0 : tenant_shard.generation = Some(new_generation);
2146 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
2147 0 : } else {
2148 : // This is a detach notification. We must update placement policy to avoid re-attaching
2149 : // during background scheduling/reconciliation, or during storage controller restart.
2150 0 : assert!(attach_req.node_id.is_none());
2151 0 : tenant_shard.policy = PlacementPolicy::Detached;
2152 : }
2153 :
2154 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
2155 0 : tracing::info!(
2156 : tenant_id = %attach_req.tenant_shard_id,
2157 : ps_id = %attaching_pageserver,
2158 : generation = ?tenant_shard.generation,
2159 0 : "issuing",
2160 : );
2161 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
2162 0 : tracing::info!(
2163 : tenant_id = %attach_req.tenant_shard_id,
2164 : %ps_id,
2165 : generation = ?tenant_shard.generation,
2166 0 : "dropping",
2167 : );
2168 : } else {
2169 0 : tracing::info!(
2170 : tenant_id = %attach_req.tenant_shard_id,
2171 0 : "no-op: tenant already has no pageserver");
2172 : }
2173 0 : tenant_shard
2174 0 : .intent
2175 0 : .set_attached(scheduler, attach_req.node_id);
2176 :
2177 0 : tracing::info!(
2178 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}, config {:?}",
2179 : attach_req.tenant_shard_id,
2180 : tenant_shard.generation,
2181 : // TODO: this is an odd number of 0xf's
2182 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff)),
2183 : attach_req.config,
2184 : );
2185 :
2186 : // Trick the reconciler into not doing anything for this tenant: this helps
2187 : // tests that manually configure a tenant on the pagesrever, and then call this
2188 : // attach hook: they don't want background reconciliation to modify what they
2189 : // did to the pageserver.
2190 : #[cfg(feature = "testing")]
2191 : {
2192 0 : if let Some(node_id) = attach_req.node_id {
2193 0 : tenant_shard.observed.locations = HashMap::from([(
2194 0 : node_id,
2195 0 : ObservedStateLocation {
2196 0 : conf: Some(attached_location_conf(
2197 0 : tenant_shard.generation.unwrap(),
2198 0 : &tenant_shard.shard,
2199 0 : &tenant_shard.config,
2200 0 : &PlacementPolicy::Attached(0),
2201 0 : tenant_shard.intent.get_secondary().len(),
2202 0 : )),
2203 0 : },
2204 0 : )]);
2205 0 : } else {
2206 0 : tenant_shard.observed.locations.clear();
2207 0 : }
2208 : }
2209 :
2210 : Ok(AttachHookResponse {
2211 0 : generation: attach_req
2212 0 : .node_id
2213 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
2214 : })
2215 0 : }
2216 :
2217 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
2218 0 : let locked = self.inner.read().unwrap();
2219 :
2220 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
2221 :
2222 : InspectResponse {
2223 0 : attachment: tenant_shard.and_then(|s| {
2224 0 : s.intent
2225 0 : .get_attached()
2226 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
2227 0 : }),
2228 : }
2229 0 : }
2230 :
2231 : // When the availability state of a node transitions to active, we must do a full reconciliation
2232 : // of LocationConfigs on that node. This is because while a node was offline:
2233 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
2234 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
2235 : //
2236 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
2237 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
2238 : // this function.
2239 : //
2240 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
2241 : // for written for a single node rather than as a batch job for all nodes.
2242 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
2243 : async fn node_activate_reconcile(
2244 : &self,
2245 : mut node: Node,
2246 : _lock: &TracingExclusiveGuard<NodeOperations>,
2247 : ) -> Result<(), ApiError> {
2248 : // This Node is a mutable local copy: we will set it active so that we can use its
2249 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
2250 : // later.
2251 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
2252 :
2253 : let configs = match node
2254 : .with_client_retries(
2255 0 : |client| async move { client.list_location_config().await },
2256 : &self.http_client,
2257 : &self.config.pageserver_jwt_token,
2258 : 1,
2259 : 5,
2260 : SHORT_RECONCILE_TIMEOUT,
2261 : &self.cancel,
2262 : )
2263 : .await
2264 : {
2265 : None => {
2266 : // We're shutting down (the Node's cancellation token can't have fired, because
2267 : // we're the only scope that has a reference to it, and we didn't fire it).
2268 : return Err(ApiError::ShuttingDown);
2269 : }
2270 : Some(Err(e)) => {
2271 : // This node didn't succeed listing its locations: it may not proceed to active state
2272 : // as it is apparently unavailable.
2273 : return Err(ApiError::PreconditionFailed(
2274 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2275 : ));
2276 : }
2277 : Some(Ok(configs)) => configs,
2278 : };
2279 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2280 :
2281 : let mut cleanup = Vec::new();
2282 : let mut mismatched_locations = 0;
2283 : {
2284 : let mut locked = self.inner.write().unwrap();
2285 :
2286 : for (tenant_shard_id, reported) in configs.tenant_shards {
2287 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2288 : cleanup.push(tenant_shard_id);
2289 : continue;
2290 : };
2291 :
2292 : let on_record = &mut tenant_shard
2293 : .observed
2294 : .locations
2295 : .entry(node.get_id())
2296 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2297 : .conf;
2298 :
2299 : // If the location reported by the node does not match our observed state,
2300 : // then we mark it as uncertain and let the background reconciliation loop
2301 : // deal with it.
2302 : //
2303 : // Note that this also covers net new locations reported by the node.
2304 : if *on_record != reported {
2305 : mismatched_locations += 1;
2306 : *on_record = None;
2307 : }
2308 : }
2309 : }
2310 :
2311 : if mismatched_locations > 0 {
2312 : tracing::info!(
2313 : "Set observed state to None for {mismatched_locations} mismatched locations"
2314 : );
2315 : }
2316 :
2317 : for tenant_shard_id in cleanup {
2318 : tracing::info!("Detaching {tenant_shard_id}");
2319 : match node
2320 : .with_client_retries(
2321 0 : |client| async move {
2322 0 : let config = LocationConfig {
2323 0 : mode: LocationConfigMode::Detached,
2324 0 : generation: None,
2325 0 : secondary_conf: None,
2326 0 : shard_number: tenant_shard_id.shard_number.0,
2327 0 : shard_count: tenant_shard_id.shard_count.literal(),
2328 0 : shard_stripe_size: 0,
2329 0 : tenant_conf: models::TenantConfig::default(),
2330 0 : };
2331 0 : client
2332 0 : .location_config(tenant_shard_id, config, None, false)
2333 0 : .await
2334 0 : },
2335 : &self.http_client,
2336 : &self.config.pageserver_jwt_token,
2337 : 1,
2338 : 5,
2339 : SHORT_RECONCILE_TIMEOUT,
2340 : &self.cancel,
2341 : )
2342 : .await
2343 : {
2344 : None => {
2345 : // We're shutting down (the Node's cancellation token can't have fired, because
2346 : // we're the only scope that has a reference to it, and we didn't fire it).
2347 : return Err(ApiError::ShuttingDown);
2348 : }
2349 : Some(Err(e)) => {
2350 : // Do not let the node proceed to Active state if it is not responsive to requests
2351 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2352 : // detach completing: we should not let this node back into the set of nodes considered
2353 : // okay for scheduling.
2354 : return Err(ApiError::Conflict(format!(
2355 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2356 : )));
2357 : }
2358 : Some(Ok(_)) => {}
2359 : };
2360 : }
2361 :
2362 : Ok(())
2363 : }
2364 :
2365 0 : pub(crate) async fn re_attach(
2366 0 : &self,
2367 0 : reattach_req: ReAttachRequest,
2368 0 : ) -> Result<ReAttachResponse, ApiError> {
2369 0 : if let Some(register_req) = reattach_req.register {
2370 0 : self.node_register(register_req).await?;
2371 0 : }
2372 :
2373 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2374 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2375 :
2376 0 : tracing::info!(
2377 : node_id=%reattach_req.node_id,
2378 0 : "Incremented {} tenant shards' generations",
2379 0 : incremented_generations.len()
2380 : );
2381 :
2382 : // Apply the updated generation to our in-memory state, and
2383 : // gather discover secondary locations.
2384 0 : let mut locked = self.inner.write().unwrap();
2385 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2386 :
2387 0 : let mut response = ReAttachResponse {
2388 0 : tenants: Vec::new(),
2389 0 : };
2390 :
2391 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2392 : // to call location_conf API with an old generation. Wait for cancellation to complete
2393 : // before responding to this request. Requires well implemented CancellationToken logic
2394 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2395 : // request in flight over the network: TODO handle that by making location_conf API refuse
2396 : // to go backward in generations.
2397 :
2398 : // Scan through all shards, applying updates for ones where we updated generation
2399 : // and identifying shards that intend to have a secondary location on this node.
2400 0 : for (tenant_shard_id, shard) in tenants {
2401 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2402 0 : let new_gen = *new_gen;
2403 0 : response.tenants.push(ReAttachResponseTenant {
2404 0 : id: *tenant_shard_id,
2405 0 : r#gen: Some(new_gen.into().unwrap()),
2406 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2407 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2408 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2409 0 : // the stale/multi states at this point.
2410 0 : mode: LocationConfigMode::AttachedSingle,
2411 0 : stripe_size: shard.shard.stripe_size,
2412 0 : });
2413 :
2414 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2415 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2416 : // Why can we update `observed` even though we're not sure our response will be received
2417 : // by the pageserver? Because the pageserver will not proceed with startup until
2418 : // it has processed response: if it loses it, we'll see another request and increment
2419 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2420 0 : if let Some(conf) = observed.conf.as_mut() {
2421 0 : conf.generation = new_gen.into();
2422 0 : }
2423 0 : } else {
2424 0 : // This node has no observed state for the shard: perhaps it was offline
2425 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2426 0 : // will be prompted to learn the location's state before it makes changes.
2427 0 : shard
2428 0 : .observed
2429 0 : .locations
2430 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2431 0 : }
2432 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2433 0 : // Ordering: pageserver will not accept /location_config requests until it has
2434 0 : // finished processing the response from re-attach. So we can update our in-memory state
2435 0 : // now, and be confident that we are not stamping on the result of some later location config.
2436 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2437 0 : // so we might update observed state here, and then get over-written by some racing
2438 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2439 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2440 0 :
2441 0 : response.tenants.push(ReAttachResponseTenant {
2442 0 : id: *tenant_shard_id,
2443 0 : r#gen: None,
2444 0 : mode: LocationConfigMode::Secondary,
2445 0 : stripe_size: shard.shard.stripe_size,
2446 0 : });
2447 0 :
2448 0 : // We must not update observed, because we have no guarantee that our
2449 0 : // response will be received by the pageserver. This could leave it
2450 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2451 0 : }
2452 : }
2453 :
2454 : // We consider a node Active once we have composed a re-attach response, but we
2455 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2456 : // implicitly synchronizes the LocationConfigs on the node.
2457 : //
2458 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2459 : // but those requests will not be accepted by the node until it has finished processing
2460 : // the re-attach response.
2461 : //
2462 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2463 : // in [`Persistence::re_attach`].
2464 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2465 0 : let reset_scheduling = matches!(
2466 0 : node.get_scheduling(),
2467 : NodeSchedulingPolicy::PauseForRestart
2468 : | NodeSchedulingPolicy::Draining
2469 : | NodeSchedulingPolicy::Filling
2470 : | NodeSchedulingPolicy::Deleting
2471 : );
2472 :
2473 0 : let mut new_nodes = (**nodes).clone();
2474 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2475 0 : if reset_scheduling {
2476 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2477 0 : }
2478 :
2479 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2480 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2481 :
2482 0 : scheduler.node_upsert(node);
2483 0 : let new_nodes = Arc::new(new_nodes);
2484 0 : *nodes = new_nodes;
2485 : } else {
2486 0 : tracing::error!(
2487 0 : "Reattaching node {} was removed while processing the request",
2488 : reattach_req.node_id
2489 : );
2490 : }
2491 0 : }
2492 :
2493 0 : Ok(response)
2494 0 : }
2495 :
2496 0 : pub(crate) async fn validate(
2497 0 : &self,
2498 0 : validate_req: ValidateRequest,
2499 0 : ) -> Result<ValidateResponse, DatabaseError> {
2500 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2501 : // in-memory generation for a shard
2502 0 : let in_memory_result = {
2503 0 : let mut in_memory_result = Vec::new();
2504 0 : let locked = self.inner.read().unwrap();
2505 0 : for req_tenant in validate_req.tenants {
2506 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2507 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2508 0 : tracing::info!(
2509 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2510 : req_tenant.id,
2511 : req_tenant.r#gen,
2512 : tenant_shard.generation
2513 : );
2514 :
2515 0 : in_memory_result.push((
2516 0 : req_tenant.id,
2517 0 : Generation::new(req_tenant.r#gen),
2518 0 : valid,
2519 0 : ));
2520 : } else {
2521 : // This is legal: for example during a shard split the pageserver may still
2522 : // have deletions in its queue from the old pre-split shard, or after deletion
2523 : // of a tenant that was busy with compaction/gc while being deleted.
2524 0 : tracing::info!(
2525 0 : "Refusing deletion validation for missing shard {}",
2526 : req_tenant.id
2527 : );
2528 : }
2529 : }
2530 :
2531 0 : in_memory_result
2532 : };
2533 :
2534 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2535 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2536 0 : let db_generations = self
2537 0 : .persistence
2538 0 : .shard_generations(
2539 0 : in_memory_result
2540 0 : .iter()
2541 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2542 : )
2543 0 : .await?;
2544 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2545 :
2546 0 : let mut response = ValidateResponse {
2547 0 : tenants: Vec::new(),
2548 0 : };
2549 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2550 0 : let valid = if valid {
2551 0 : let db_generation = db_generations.get(&tenant_shard_id);
2552 0 : db_generation == Some(&Some(validate_generation))
2553 : } else {
2554 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2555 : // this prevents a pageserver from cleaning up an object in S3.
2556 0 : false
2557 : };
2558 :
2559 0 : response.tenants.push(ValidateResponseTenant {
2560 0 : id: tenant_shard_id,
2561 0 : valid,
2562 0 : })
2563 : }
2564 :
2565 0 : Ok(response)
2566 0 : }
2567 :
2568 0 : pub(crate) async fn tenant_create(
2569 0 : &self,
2570 0 : create_req: TenantCreateRequest,
2571 0 : ) -> Result<TenantCreateResponse, ApiError> {
2572 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2573 :
2574 : // Exclude any concurrent attempts to create/access the same tenant ID
2575 0 : let _tenant_lock = trace_exclusive_lock(
2576 0 : &self.tenant_op_locks,
2577 0 : create_req.new_tenant_id.tenant_id,
2578 0 : TenantOperations::Create,
2579 0 : )
2580 0 : .await;
2581 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2582 :
2583 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2584 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2585 : // accept compute notifications while it is in the process of creating. Reconciliation will
2586 : // be retried in the background.
2587 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2588 0 : }
2589 0 : Ok(response)
2590 0 : }
2591 :
2592 0 : pub(crate) async fn do_tenant_create(
2593 0 : &self,
2594 0 : create_req: TenantCreateRequest,
2595 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2596 0 : let placement_policy = create_req
2597 0 : .placement_policy
2598 0 : .clone()
2599 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2600 0 : .unwrap_or(PlacementPolicy::Attached(0));
2601 :
2602 : // This service expects to handle sharding itself: it is an error to try and directly create
2603 : // a particular shard here.
2604 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2605 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2606 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2607 0 : )));
2608 : } else {
2609 0 : create_req.new_tenant_id.tenant_id
2610 : };
2611 :
2612 0 : tracing::info!(
2613 0 : "Creating tenant {}, shard_count={:?}",
2614 : create_req.new_tenant_id,
2615 : create_req.shard_parameters.count,
2616 : );
2617 :
2618 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2619 0 : .map(|i| TenantShardId {
2620 0 : tenant_id,
2621 0 : shard_number: ShardNumber(i),
2622 0 : shard_count: create_req.shard_parameters.count,
2623 0 : })
2624 0 : .collect::<Vec<_>>();
2625 :
2626 : // If the caller specifies a None generation, it means "start from default". This is different
2627 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2628 : // an incompletely-onboarded tenant.
2629 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2630 0 : tracing::info!(
2631 0 : "tenant_create: secondary mode, generation is_some={}",
2632 0 : create_req.generation.is_some()
2633 : );
2634 0 : create_req.generation.map(Generation::new)
2635 : } else {
2636 0 : tracing::info!(
2637 0 : "tenant_create: not secondary mode, generation is_some={}",
2638 0 : create_req.generation.is_some()
2639 : );
2640 0 : Some(
2641 0 : create_req
2642 0 : .generation
2643 0 : .map(Generation::new)
2644 0 : .unwrap_or(INITIAL_GENERATION),
2645 0 : )
2646 : };
2647 :
2648 0 : let preferred_az_id = {
2649 0 : let locked = self.inner.read().unwrap();
2650 : // Idempotency: take the existing value if the tenant already exists
2651 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2652 0 : shard.preferred_az().cloned()
2653 : } else {
2654 0 : locked.scheduler.get_az_for_new_tenant()
2655 : }
2656 : };
2657 :
2658 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2659 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2660 : // during the creation, rather than risking leaving orphan objects in S3.
2661 0 : let persist_tenant_shards = create_ids
2662 0 : .iter()
2663 0 : .map(|tenant_shard_id| TenantShardPersistence {
2664 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2665 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2666 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2667 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2668 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2669 : // The pageserver is not known until scheduling happens: we will set this column when
2670 : // incrementing the generation the first time we attach to a pageserver.
2671 0 : generation_pageserver: None,
2672 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2673 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2674 0 : splitting: SplitState::default(),
2675 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2676 0 : .unwrap(),
2677 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2678 0 : })
2679 0 : .collect();
2680 :
2681 0 : match self
2682 0 : .persistence
2683 0 : .insert_tenant_shards(persist_tenant_shards)
2684 0 : .await
2685 : {
2686 0 : Ok(_) => {}
2687 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2688 : DatabaseErrorKind::UniqueViolation,
2689 : _,
2690 : ))) => {
2691 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2692 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2693 : // creation's shard count.
2694 0 : tracing::info!(
2695 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2696 : );
2697 : }
2698 : // Any other database error is unexpected and a bug.
2699 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2700 : };
2701 :
2702 0 : let mut schedule_context = ScheduleContext::default();
2703 0 : let mut schedule_error = None;
2704 0 : let mut response_shards = Vec::new();
2705 0 : for tenant_shard_id in create_ids {
2706 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2707 :
2708 0 : let outcome = self
2709 0 : .do_initial_shard_scheduling(
2710 0 : tenant_shard_id,
2711 0 : initial_generation,
2712 0 : create_req.shard_parameters,
2713 0 : create_req.config.clone(),
2714 0 : placement_policy.clone(),
2715 0 : preferred_az_id.as_ref(),
2716 0 : &mut schedule_context,
2717 0 : )
2718 0 : .await;
2719 :
2720 0 : match outcome {
2721 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2722 0 : InitialShardScheduleOutcome::NotScheduled => {}
2723 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2724 0 : schedule_error = Some(err);
2725 0 : }
2726 : }
2727 : }
2728 :
2729 : // If we failed to schedule shards, then they are still created in the controller,
2730 : // but we return an error to the requester to avoid a silent failure when someone
2731 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2732 : // are present in the system. We do this here rather than in the above loop, to
2733 : // avoid situations where we only create a subset of shards in the tenant.
2734 0 : if let Some(e) = schedule_error {
2735 0 : return Err(ApiError::Conflict(format!(
2736 0 : "Failed to schedule shard(s): {e}"
2737 0 : )));
2738 0 : }
2739 :
2740 0 : let waiters = {
2741 0 : let mut locked = self.inner.write().unwrap();
2742 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2743 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2744 0 : .tenant_creation_hint(true)
2745 0 : .build();
2746 0 : tenants
2747 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2748 0 : .filter_map(|(_shard_id, shard)| {
2749 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2750 0 : })
2751 0 : .collect::<Vec<_>>()
2752 : };
2753 :
2754 0 : Ok((
2755 0 : TenantCreateResponse {
2756 0 : shards: response_shards,
2757 0 : },
2758 0 : waiters,
2759 0 : ))
2760 0 : }
2761 :
2762 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2763 : /// case of a new tenant and a pre-existing one.
2764 : #[allow(clippy::too_many_arguments)]
2765 0 : async fn do_initial_shard_scheduling(
2766 0 : &self,
2767 0 : tenant_shard_id: TenantShardId,
2768 0 : initial_generation: Option<Generation>,
2769 0 : shard_params: ShardParameters,
2770 0 : config: TenantConfig,
2771 0 : placement_policy: PlacementPolicy,
2772 0 : preferred_az_id: Option<&AvailabilityZone>,
2773 0 : schedule_context: &mut ScheduleContext,
2774 0 : ) -> InitialShardScheduleOutcome {
2775 0 : let mut locked = self.inner.write().unwrap();
2776 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2777 :
2778 : use std::collections::btree_map::Entry;
2779 0 : match tenants.entry(tenant_shard_id) {
2780 0 : Entry::Occupied(mut entry) => {
2781 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2782 :
2783 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2784 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2785 0 : }
2786 :
2787 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2788 0 : let generation = entry
2789 0 : .get()
2790 0 : .generation
2791 0 : .expect("Generation is set when in attached mode");
2792 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2793 0 : shard_id: tenant_shard_id,
2794 0 : node_id: *node_id,
2795 0 : generation: generation.into().unwrap(),
2796 0 : })
2797 : } else {
2798 0 : InitialShardScheduleOutcome::NotScheduled
2799 : }
2800 : }
2801 0 : Entry::Vacant(entry) => {
2802 0 : let state = entry.insert(TenantShard::new(
2803 0 : tenant_shard_id,
2804 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2805 0 : placement_policy,
2806 0 : preferred_az_id.cloned(),
2807 : ));
2808 :
2809 0 : state.generation = initial_generation;
2810 0 : state.config = config;
2811 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2812 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2813 0 : }
2814 :
2815 : // Only include shards in result if we are attaching: the purpose
2816 : // of the response is to tell the caller where the shards are attached.
2817 0 : if let Some(node_id) = state.intent.get_attached() {
2818 0 : let generation = state
2819 0 : .generation
2820 0 : .expect("Generation is set when in attached mode");
2821 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2822 0 : shard_id: tenant_shard_id,
2823 0 : node_id: *node_id,
2824 0 : generation: generation.into().unwrap(),
2825 0 : })
2826 : } else {
2827 0 : InitialShardScheduleOutcome::NotScheduled
2828 : }
2829 : }
2830 : }
2831 0 : }
2832 :
2833 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2834 : /// wait for reconciliation to complete before responding.
2835 0 : async fn await_waiters(
2836 0 : &self,
2837 0 : waiters: Vec<ReconcilerWaiter>,
2838 0 : timeout: Duration,
2839 0 : ) -> Result<(), ReconcileWaitError> {
2840 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2841 0 : for waiter in waiters {
2842 0 : let timeout = deadline.duration_since(Instant::now());
2843 0 : waiter.wait_timeout(timeout).await?;
2844 : }
2845 :
2846 0 : Ok(())
2847 0 : }
2848 :
2849 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2850 : /// in progress
2851 0 : async fn await_waiters_remainder(
2852 0 : &self,
2853 0 : waiters: Vec<ReconcilerWaiter>,
2854 0 : timeout: Duration,
2855 0 : ) -> Vec<ReconcilerWaiter> {
2856 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2857 0 : for waiter in waiters.iter() {
2858 0 : let timeout = deadline.duration_since(Instant::now());
2859 0 : let _ = waiter.wait_timeout(timeout).await;
2860 : }
2861 :
2862 0 : waiters
2863 0 : .into_iter()
2864 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2865 0 : .collect::<Vec<_>>()
2866 0 : }
2867 :
2868 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2869 : /// and transform it into either a tenant creation of a series of shard updates.
2870 : ///
2871 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2872 : /// still be returned.
2873 0 : fn tenant_location_config_prepare(
2874 0 : &self,
2875 0 : tenant_id: TenantId,
2876 0 : req: TenantLocationConfigRequest,
2877 0 : ) -> TenantCreateOrUpdate {
2878 0 : let mut updates = Vec::new();
2879 0 : let mut locked = self.inner.write().unwrap();
2880 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2881 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2882 :
2883 : // Use location config mode as an indicator of policy.
2884 0 : let placement_policy = match req.config.mode {
2885 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2886 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2887 : LocationConfigMode::AttachedMulti
2888 : | LocationConfigMode::AttachedSingle
2889 : | LocationConfigMode::AttachedStale => {
2890 0 : if nodes.len() > 1 {
2891 0 : PlacementPolicy::Attached(1)
2892 : } else {
2893 : // Convenience for dev/test: if we just have one pageserver, import
2894 : // tenants into non-HA mode so that scheduling will succeed.
2895 0 : PlacementPolicy::Attached(0)
2896 : }
2897 : }
2898 : };
2899 :
2900 : // Ordinarily we do not update scheduling policy, but when making major changes
2901 : // like detaching or demoting to secondary-only, we need to force the scheduling
2902 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2903 0 : let scheduling_policy = match req.config.mode {
2904 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2905 : // Special case: when making major changes like detaching or demoting to secondary-only,
2906 : // we need to force the scheduling mode to Active, or nothing will happen.
2907 0 : Some(ShardSchedulingPolicy::Active)
2908 : }
2909 : LocationConfigMode::AttachedMulti
2910 : | LocationConfigMode::AttachedSingle
2911 : | LocationConfigMode::AttachedStale => {
2912 : // While attached, continue to respect whatever the existing scheduling mode is.
2913 0 : None
2914 : }
2915 : };
2916 :
2917 0 : let mut create = true;
2918 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2919 : // Saw an existing shard: this is not a creation
2920 0 : create = false;
2921 :
2922 : // Shards may have initially been created by a Secondary request, where we
2923 : // would have left generation as None.
2924 : //
2925 : // We only update generation the first time we see an attached-mode request,
2926 : // and if there is no existing generation set. The caller is responsible for
2927 : // ensuring that no non-storage-controller pageserver ever uses a higher
2928 : // generation than they passed in here.
2929 : use LocationConfigMode::*;
2930 0 : let set_generation = match req.config.mode {
2931 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2932 0 : req.config.generation.map(Generation::new)
2933 : }
2934 0 : _ => None,
2935 : };
2936 :
2937 0 : updates.push(ShardUpdate {
2938 0 : tenant_shard_id: *shard_id,
2939 0 : placement_policy: placement_policy.clone(),
2940 0 : tenant_config: req.config.tenant_conf.clone(),
2941 0 : generation: set_generation,
2942 0 : scheduling_policy,
2943 0 : });
2944 : }
2945 :
2946 0 : if create {
2947 : use LocationConfigMode::*;
2948 0 : let generation = match req.config.mode {
2949 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2950 : // If a caller provided a generation in a non-attached request, ignore it
2951 : // and leave our generation as None: this enables a subsequent update to set
2952 : // the generation when setting an attached mode for the first time.
2953 0 : _ => None,
2954 : };
2955 :
2956 0 : TenantCreateOrUpdate::Create(
2957 0 : // Synthesize a creation request
2958 0 : TenantCreateRequest {
2959 0 : new_tenant_id: tenant_shard_id,
2960 0 : generation,
2961 0 : shard_parameters: ShardParameters {
2962 0 : count: tenant_shard_id.shard_count,
2963 0 : // We only import un-sharded or single-sharded tenants, so stripe
2964 0 : // size can be made up arbitrarily here.
2965 0 : stripe_size: DEFAULT_STRIPE_SIZE,
2966 0 : },
2967 0 : placement_policy: Some(placement_policy),
2968 0 : config: req.config.tenant_conf,
2969 0 : },
2970 0 : )
2971 : } else {
2972 0 : assert!(!updates.is_empty());
2973 0 : TenantCreateOrUpdate::Update(updates)
2974 : }
2975 0 : }
2976 :
2977 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2978 : /// the tenant is present in memory. If not, load it from the database. If it is found
2979 : /// in neither location, return a NotFound error.
2980 : ///
2981 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2982 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2983 0 : async fn maybe_load_tenant(
2984 0 : &self,
2985 0 : tenant_id: TenantId,
2986 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2987 0 : ) -> Result<(), ApiError> {
2988 : // Check if the tenant is present in memory, and select an AZ to use when loading
2989 : // if we will load it.
2990 0 : let load_in_az = {
2991 0 : let locked = self.inner.read().unwrap();
2992 0 : let existing = locked
2993 0 : .tenants
2994 0 : .range(TenantShardId::tenant_range(tenant_id))
2995 0 : .next();
2996 :
2997 : // If the tenant is not present in memory, we expect to load it from database,
2998 : // so let's figure out what AZ to load it into while we have self.inner locked.
2999 0 : if existing.is_none() {
3000 0 : locked
3001 0 : .scheduler
3002 0 : .get_az_for_new_tenant()
3003 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
3004 0 : "No AZ with nodes found to load tenant"
3005 0 : )))?
3006 : } else {
3007 : // We already have this tenant in memory
3008 0 : return Ok(());
3009 : }
3010 : };
3011 :
3012 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
3013 0 : if tenant_shards.is_empty() {
3014 0 : return Err(ApiError::NotFound(
3015 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3016 0 : ));
3017 0 : }
3018 :
3019 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
3020 0 : self.persistence
3021 0 : .set_tenant_shard_preferred_azs(
3022 0 : tenant_shards
3023 0 : .iter()
3024 0 : .map(|t| {
3025 0 : (
3026 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
3027 0 : Some(load_in_az.clone()),
3028 0 : )
3029 0 : })
3030 0 : .collect(),
3031 : )
3032 0 : .await?;
3033 :
3034 0 : let mut locked = self.inner.write().unwrap();
3035 0 : tracing::info!(
3036 0 : "Loaded {} shards for tenant {}",
3037 0 : tenant_shards.len(),
3038 : tenant_id
3039 : );
3040 :
3041 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
3042 0 : let intent = IntentState::new(Some(load_in_az.clone()));
3043 0 : let shard =
3044 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
3045 :
3046 : // Sanity check: when loading on-demand, we should always be loaded something Detached
3047 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
3048 0 : if shard.policy != PlacementPolicy::Detached {
3049 0 : tracing::error!(
3050 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
3051 : shard.tenant_shard_id,
3052 : shard.policy
3053 : );
3054 0 : }
3055 :
3056 0 : (shard.tenant_shard_id, shard)
3057 0 : }));
3058 :
3059 0 : Ok(())
3060 0 : }
3061 :
3062 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
3063 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
3064 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
3065 : ///
3066 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
3067 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
3068 0 : fn maybe_drop_tenant(
3069 0 : &self,
3070 0 : tenant_id: TenantId,
3071 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
3072 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
3073 0 : ) {
3074 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
3075 0 : if tenant_shards.all(|(_id, shard)| {
3076 0 : shard.policy == PlacementPolicy::Detached
3077 0 : && shard.reconciler.is_none()
3078 0 : && shard.observed.is_empty()
3079 0 : }) {
3080 0 : let keys = locked
3081 0 : .tenants
3082 0 : .range(TenantShardId::tenant_range(tenant_id))
3083 0 : .map(|(id, _)| id)
3084 0 : .copied()
3085 0 : .collect::<Vec<_>>();
3086 0 : for key in keys {
3087 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
3088 0 : locked.tenants.remove(&key);
3089 : }
3090 0 : }
3091 0 : }
3092 :
3093 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
3094 : /// directly with pageservers into this service.
3095 : ///
3096 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
3097 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
3098 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
3099 : /// tenant's source of generation numbers.
3100 : ///
3101 : /// The mode in this request coarse-grained control of tenants:
3102 : /// - Call with mode Attached* to upsert the tenant.
3103 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
3104 : /// to set an existing tenant to PolicyMode::Secondary
3105 : /// - Call with mode Detached to switch to PolicyMode::Detached
3106 0 : pub(crate) async fn tenant_location_config(
3107 0 : &self,
3108 0 : tenant_shard_id: TenantShardId,
3109 0 : req: TenantLocationConfigRequest,
3110 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
3111 : // We require an exclusive lock, because we are updating both persistent and in-memory state
3112 0 : let _tenant_lock = trace_exclusive_lock(
3113 0 : &self.tenant_op_locks,
3114 0 : tenant_shard_id.tenant_id,
3115 0 : TenantOperations::LocationConfig,
3116 0 : )
3117 0 : .await;
3118 :
3119 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
3120 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3121 0 : "This API is for importing single-sharded or unsharded tenants"
3122 0 : )));
3123 : } else {
3124 0 : tenant_shard_id.tenant_id
3125 : };
3126 :
3127 : // In case we are waking up a Detached tenant
3128 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
3129 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
3130 0 : // This is a creation or an update
3131 0 : }
3132 0 : Err(e) => {
3133 0 : return Err(e);
3134 : }
3135 : };
3136 :
3137 : // First check if this is a creation or an update
3138 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
3139 :
3140 0 : let mut result = TenantLocationConfigResponse {
3141 0 : shards: Vec::new(),
3142 0 : stripe_size: None,
3143 0 : };
3144 0 : let waiters = match create_or_update {
3145 0 : TenantCreateOrUpdate::Create(create_req) => {
3146 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
3147 0 : result.shards = create_resp
3148 0 : .shards
3149 0 : .into_iter()
3150 0 : .map(|s| TenantShardLocation {
3151 0 : node_id: s.node_id,
3152 0 : shard_id: s.shard_id,
3153 0 : })
3154 0 : .collect();
3155 0 : waiters
3156 : }
3157 0 : TenantCreateOrUpdate::Update(updates) => {
3158 : // Persist updates
3159 : // Ordering: write to the database before applying changes in-memory, so that
3160 : // we will not appear time-travel backwards on a restart.
3161 :
3162 0 : let mut schedule_context = ScheduleContext::default();
3163 : for ShardUpdate {
3164 0 : tenant_shard_id,
3165 0 : placement_policy,
3166 0 : tenant_config,
3167 0 : generation,
3168 0 : scheduling_policy,
3169 0 : } in &updates
3170 : {
3171 0 : self.persistence
3172 0 : .update_tenant_shard(
3173 0 : TenantFilter::Shard(*tenant_shard_id),
3174 0 : Some(placement_policy.clone()),
3175 0 : Some(tenant_config.clone()),
3176 0 : *generation,
3177 0 : *scheduling_policy,
3178 0 : )
3179 0 : .await?;
3180 : }
3181 :
3182 : // Apply updates in-memory
3183 0 : let mut waiters = Vec::new();
3184 : {
3185 0 : let mut locked = self.inner.write().unwrap();
3186 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3187 :
3188 : for ShardUpdate {
3189 0 : tenant_shard_id,
3190 0 : placement_policy,
3191 0 : tenant_config,
3192 0 : generation: update_generation,
3193 0 : scheduling_policy,
3194 0 : } in updates
3195 : {
3196 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
3197 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
3198 0 : continue;
3199 : };
3200 :
3201 : // Update stripe size
3202 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
3203 0 : result.stripe_size = Some(shard.shard.stripe_size);
3204 0 : }
3205 :
3206 0 : shard.policy = placement_policy;
3207 0 : shard.config = tenant_config;
3208 0 : if let Some(generation) = update_generation {
3209 0 : shard.generation = Some(generation);
3210 0 : }
3211 :
3212 0 : if let Some(scheduling_policy) = scheduling_policy {
3213 0 : shard.set_scheduling_policy(scheduling_policy);
3214 0 : }
3215 :
3216 0 : shard.schedule(scheduler, &mut schedule_context)?;
3217 :
3218 0 : let maybe_waiter =
3219 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3220 0 : if let Some(waiter) = maybe_waiter {
3221 0 : waiters.push(waiter);
3222 0 : }
3223 :
3224 0 : if let Some(node_id) = shard.intent.get_attached() {
3225 0 : result.shards.push(TenantShardLocation {
3226 0 : shard_id: tenant_shard_id,
3227 0 : node_id: *node_id,
3228 0 : })
3229 0 : }
3230 : }
3231 : }
3232 0 : waiters
3233 : }
3234 : };
3235 :
3236 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3237 : // Do not treat a reconcile error as fatal: we have already applied any requested
3238 : // Intent changes, and the reconcile can fail for external reasons like unavailable
3239 : // compute notification API. In these cases, it is important that we do not
3240 : // cause the cloud control plane to retry forever on this API.
3241 0 : tracing::warn!(
3242 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
3243 : );
3244 0 : }
3245 :
3246 : // Logging the full result is useful because it lets us cross-check what the cloud control
3247 : // plane's tenant_shards table should contain.
3248 0 : tracing::info!("Complete, returning {result:?}");
3249 :
3250 0 : Ok(result)
3251 0 : }
3252 :
3253 0 : pub(crate) async fn tenant_config_patch(
3254 0 : &self,
3255 0 : req: TenantConfigPatchRequest,
3256 0 : ) -> Result<(), ApiError> {
3257 0 : let _tenant_lock = trace_exclusive_lock(
3258 0 : &self.tenant_op_locks,
3259 0 : req.tenant_id,
3260 0 : TenantOperations::ConfigPatch,
3261 0 : )
3262 0 : .await;
3263 :
3264 0 : let tenant_id = req.tenant_id;
3265 0 : let patch = req.config;
3266 :
3267 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3268 :
3269 0 : let base = {
3270 0 : let locked = self.inner.read().unwrap();
3271 0 : let shards = locked
3272 0 : .tenants
3273 0 : .range(TenantShardId::tenant_range(req.tenant_id));
3274 :
3275 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
3276 :
3277 0 : let first = match configs.peek() {
3278 0 : Some(first) => (*first).clone(),
3279 : None => {
3280 0 : return Err(ApiError::NotFound(
3281 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3282 0 : ));
3283 : }
3284 : };
3285 :
3286 0 : if !configs.all_equal() {
3287 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3288 : // This can't happen because we atomically update the database records
3289 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3290 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3291 0 : "Tenant configs for {} are mismatched",
3292 0 : req.tenant_id
3293 0 : )));
3294 0 : }
3295 :
3296 0 : first
3297 : };
3298 :
3299 0 : let updated_config = base
3300 0 : .apply_patch(patch)
3301 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3302 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3303 0 : .await
3304 0 : }
3305 :
3306 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3307 : // We require an exclusive lock, because we are updating persistent and in-memory state
3308 0 : let _tenant_lock = trace_exclusive_lock(
3309 0 : &self.tenant_op_locks,
3310 0 : req.tenant_id,
3311 0 : TenantOperations::ConfigSet,
3312 0 : )
3313 0 : .await;
3314 :
3315 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3316 :
3317 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3318 0 : .await
3319 0 : }
3320 :
3321 0 : async fn set_tenant_config_and_reconcile(
3322 0 : &self,
3323 0 : tenant_id: TenantId,
3324 0 : config: TenantConfig,
3325 0 : ) -> Result<(), ApiError> {
3326 0 : self.persistence
3327 0 : .update_tenant_shard(
3328 0 : TenantFilter::Tenant(tenant_id),
3329 0 : None,
3330 0 : Some(config.clone()),
3331 0 : None,
3332 0 : None,
3333 0 : )
3334 0 : .await?;
3335 :
3336 0 : let waiters = {
3337 0 : let mut waiters = Vec::new();
3338 0 : let mut locked = self.inner.write().unwrap();
3339 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3340 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3341 0 : shard.config = config.clone();
3342 0 : if let Some(waiter) =
3343 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3344 0 : {
3345 0 : waiters.push(waiter);
3346 0 : }
3347 : }
3348 0 : waiters
3349 : };
3350 :
3351 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3352 : // Treat this as success because we have stored the configuration. If e.g.
3353 : // a node was unavailable at this time, it should not stop us accepting a
3354 : // configuration change.
3355 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3356 0 : }
3357 :
3358 0 : Ok(())
3359 0 : }
3360 :
3361 0 : pub(crate) fn tenant_config_get(
3362 0 : &self,
3363 0 : tenant_id: TenantId,
3364 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3365 0 : let config = {
3366 0 : let locked = self.inner.read().unwrap();
3367 :
3368 0 : match locked
3369 0 : .tenants
3370 0 : .range(TenantShardId::tenant_range(tenant_id))
3371 0 : .next()
3372 : {
3373 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3374 : None => {
3375 0 : return Err(ApiError::NotFound(
3376 0 : anyhow::anyhow!("Tenant not found").into(),
3377 0 : ));
3378 : }
3379 : }
3380 : };
3381 :
3382 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3383 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3384 : // and `effective_config` in the response is meaningless, but we retain that syntax
3385 : // in order to remain compatible with the pageserver API.
3386 :
3387 0 : let response = HashMap::from([
3388 : (
3389 : "tenant_specific_overrides",
3390 0 : serde_json::to_value(&config)
3391 0 : .context("serializing tenant specific overrides")
3392 0 : .map_err(ApiError::InternalServerError)?,
3393 : ),
3394 : (
3395 0 : "effective_config",
3396 0 : serde_json::to_value(&config)
3397 0 : .context("serializing effective config")
3398 0 : .map_err(ApiError::InternalServerError)?,
3399 : ),
3400 : ]);
3401 :
3402 0 : Ok(response)
3403 0 : }
3404 :
3405 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3406 0 : &self,
3407 0 : time_travel_req: &TenantTimeTravelRequest,
3408 0 : tenant_id: TenantId,
3409 0 : timestamp: Cow<'_, str>,
3410 0 : done_if_after: Cow<'_, str>,
3411 0 : ) -> Result<(), ApiError> {
3412 0 : let _tenant_lock = trace_exclusive_lock(
3413 0 : &self.tenant_op_locks,
3414 0 : tenant_id,
3415 0 : TenantOperations::TimeTravelRemoteStorage,
3416 0 : )
3417 0 : .await;
3418 :
3419 0 : let node = {
3420 0 : let mut locked = self.inner.write().unwrap();
3421 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3422 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3423 : // but only at the start of the process, so it's really just to prevent operator
3424 : // mistakes.
3425 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3426 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3427 : {
3428 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3429 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3430 0 : )));
3431 0 : }
3432 0 : let maybe_attached = shard
3433 0 : .observed
3434 0 : .locations
3435 0 : .iter()
3436 0 : .filter_map(|(node_id, observed_location)| {
3437 0 : observed_location
3438 0 : .conf
3439 0 : .as_ref()
3440 0 : .map(|loc| (node_id, observed_location, loc.mode))
3441 0 : })
3442 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3443 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3444 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3445 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3446 0 : )));
3447 0 : }
3448 : }
3449 0 : let scheduler = &mut locked.scheduler;
3450 : // Right now we only perform the operation on a single node without parallelization
3451 : // TODO fan out the operation to multiple nodes for better performance
3452 0 : let node_id = scheduler.any_available_node()?;
3453 0 : let node = locked
3454 0 : .nodes
3455 0 : .get(&node_id)
3456 0 : .expect("Pageservers may not be deleted while lock is active");
3457 0 : node.clone()
3458 : };
3459 :
3460 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3461 0 : let mut counts = time_travel_req
3462 0 : .shard_counts
3463 0 : .iter()
3464 0 : .copied()
3465 0 : .collect::<HashSet<_>>()
3466 0 : .into_iter()
3467 0 : .collect::<Vec<_>>();
3468 0 : counts.sort_unstable();
3469 :
3470 0 : for count in counts {
3471 0 : let shard_ids = (0..count.count())
3472 0 : .map(|i| TenantShardId {
3473 0 : tenant_id,
3474 0 : shard_number: ShardNumber(i),
3475 0 : shard_count: count,
3476 0 : })
3477 0 : .collect::<Vec<_>>();
3478 0 : for tenant_shard_id in shard_ids {
3479 0 : let client = PageserverClient::new(
3480 0 : node.get_id(),
3481 0 : self.http_client.clone(),
3482 0 : node.base_url(),
3483 0 : self.config.pageserver_jwt_token.as_deref(),
3484 : );
3485 :
3486 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3487 :
3488 0 : client
3489 0 : .tenant_time_travel_remote_storage(
3490 0 : tenant_shard_id,
3491 0 : ×tamp,
3492 0 : &done_if_after,
3493 0 : )
3494 0 : .await
3495 0 : .map_err(|e| {
3496 0 : ApiError::InternalServerError(anyhow::anyhow!(
3497 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3498 0 : node
3499 0 : ))
3500 0 : })?;
3501 : }
3502 : }
3503 0 : Ok(())
3504 0 : }
3505 :
3506 0 : pub(crate) async fn tenant_secondary_download(
3507 0 : &self,
3508 0 : tenant_id: TenantId,
3509 0 : wait: Option<Duration>,
3510 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3511 0 : let _tenant_lock = trace_shared_lock(
3512 0 : &self.tenant_op_locks,
3513 0 : tenant_id,
3514 0 : TenantOperations::SecondaryDownload,
3515 0 : )
3516 0 : .await;
3517 :
3518 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3519 0 : let targets = {
3520 0 : let locked = self.inner.read().unwrap();
3521 0 : let mut targets = Vec::new();
3522 :
3523 0 : for (tenant_shard_id, shard) in
3524 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3525 : {
3526 0 : for node_id in shard.intent.get_secondary() {
3527 0 : let node = locked
3528 0 : .nodes
3529 0 : .get(node_id)
3530 0 : .expect("Pageservers may not be deleted while referenced");
3531 0 :
3532 0 : targets.push((*tenant_shard_id, node.clone()));
3533 0 : }
3534 : }
3535 0 : targets
3536 : };
3537 :
3538 : // Issue concurrent requests to all shards' locations
3539 0 : let mut futs = FuturesUnordered::new();
3540 0 : for (tenant_shard_id, node) in targets {
3541 0 : let client = PageserverClient::new(
3542 0 : node.get_id(),
3543 0 : self.http_client.clone(),
3544 0 : node.base_url(),
3545 0 : self.config.pageserver_jwt_token.as_deref(),
3546 : );
3547 0 : futs.push(async move {
3548 0 : let result = client
3549 0 : .tenant_secondary_download(tenant_shard_id, wait)
3550 0 : .await;
3551 0 : (result, node, tenant_shard_id)
3552 0 : })
3553 : }
3554 :
3555 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3556 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3557 : // well as more general cases like 503s, 500s, or timeouts.
3558 0 : let mut aggregate_progress = SecondaryProgress::default();
3559 0 : let mut aggregate_status: Option<StatusCode> = None;
3560 0 : let mut error: Option<mgmt_api::Error> = None;
3561 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3562 0 : match result {
3563 0 : Err(e) => {
3564 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3565 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3566 : // than they had hoped for.
3567 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3568 0 : error = Some(e)
3569 : }
3570 0 : Ok((status_code, progress)) => {
3571 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3572 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3573 0 : aggregate_progress.layers_total += progress.layers_total;
3574 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3575 0 : aggregate_progress.bytes_total += progress.bytes_total;
3576 0 : aggregate_progress.heatmap_mtime =
3577 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3578 0 : aggregate_status = match aggregate_status {
3579 0 : None => Some(status_code),
3580 0 : Some(StatusCode::OK) => Some(status_code),
3581 0 : Some(cur) => {
3582 : // Other status codes (e.g. 202) -- do not overwrite.
3583 0 : Some(cur)
3584 : }
3585 : };
3586 : }
3587 : }
3588 : }
3589 :
3590 : // If any of the shards return 202, indicate our result as 202.
3591 0 : match aggregate_status {
3592 : None => {
3593 0 : match error {
3594 0 : Some(e) => {
3595 : // No successes, and an error: surface it
3596 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3597 : }
3598 : None => {
3599 : // No shards found
3600 0 : Err(ApiError::NotFound(
3601 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3602 0 : ))
3603 : }
3604 : }
3605 : }
3606 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3607 : }
3608 0 : }
3609 :
3610 0 : pub(crate) async fn tenant_delete(
3611 0 : self: &Arc<Self>,
3612 0 : tenant_id: TenantId,
3613 0 : ) -> Result<StatusCode, ApiError> {
3614 0 : let _tenant_lock =
3615 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3616 :
3617 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3618 :
3619 : // Detach all shards. This also deletes local pageserver shard data.
3620 0 : let (detach_waiters, node) = {
3621 0 : let mut detach_waiters = Vec::new();
3622 0 : let mut locked = self.inner.write().unwrap();
3623 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3624 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3625 : // Update the tenant's intent to remove all attachments
3626 0 : shard.policy = PlacementPolicy::Detached;
3627 0 : shard
3628 0 : .schedule(scheduler, &mut ScheduleContext::default())
3629 0 : .expect("De-scheduling is infallible");
3630 0 : debug_assert!(shard.intent.get_attached().is_none());
3631 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3632 :
3633 0 : if let Some(waiter) =
3634 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3635 0 : {
3636 0 : detach_waiters.push(waiter);
3637 0 : }
3638 : }
3639 :
3640 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3641 : // was attached, just has to be able to see the S3 content)
3642 0 : let node_id = scheduler.any_available_node()?;
3643 0 : let node = nodes
3644 0 : .get(&node_id)
3645 0 : .expect("Pageservers may not be deleted while lock is active");
3646 0 : (detach_waiters, node.clone())
3647 : };
3648 :
3649 : // This reconcile wait can fail in a few ways:
3650 : // A there is a very long queue for the reconciler semaphore
3651 : // B some pageserver is failing to handle a detach promptly
3652 : // C some pageserver goes offline right at the moment we send it a request.
3653 : //
3654 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3655 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3656 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3657 : // deleting the underlying data).
3658 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3659 0 : .await?;
3660 :
3661 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3662 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3663 : // the tenant ID prefix, including all shards (even possibly stale ones).
3664 0 : match node
3665 0 : .with_client_retries(
3666 0 : |client| async move {
3667 0 : client
3668 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3669 0 : .await
3670 0 : },
3671 0 : &self.http_client,
3672 0 : &self.config.pageserver_jwt_token,
3673 : 1,
3674 : 3,
3675 : RECONCILE_TIMEOUT,
3676 0 : &self.cancel,
3677 : )
3678 0 : .await
3679 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3680 : {
3681 0 : Ok(_) => {}
3682 : Err(mgmt_api::Error::Cancelled) => {
3683 0 : return Err(ApiError::ShuttingDown);
3684 : }
3685 0 : Err(e) => {
3686 : // This is unexpected: remote deletion should be infallible, unless the object store
3687 : // at large is unavailable.
3688 0 : tracing::error!("Error deleting via node {node}: {e}");
3689 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3690 : }
3691 : }
3692 :
3693 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3694 : // our in-memory state and database state.
3695 :
3696 : // Ordering: we delete persistent state first: if we then
3697 : // crash, we will drop the in-memory state.
3698 :
3699 : // Drop persistent state.
3700 0 : self.persistence.delete_tenant(tenant_id).await?;
3701 :
3702 : // Drop in-memory state
3703 : {
3704 0 : let mut locked = self.inner.write().unwrap();
3705 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3706 :
3707 : // Dereference Scheduler from shards before dropping them
3708 0 : for (_tenant_shard_id, shard) in
3709 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3710 0 : {
3711 0 : shard.intent.clear(scheduler);
3712 0 : }
3713 :
3714 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3715 0 : tracing::info!(
3716 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3717 0 : locked.tenants.len()
3718 : );
3719 : };
3720 :
3721 : // Delete the tenant from safekeepers (if needed)
3722 0 : self.tenant_delete_safekeepers(tenant_id)
3723 0 : .instrument(tracing::info_span!("tenant_delete_safekeepers", %tenant_id))
3724 0 : .await?;
3725 :
3726 : // Success is represented as 404, to imitate the existing pageserver deletion API
3727 0 : Ok(StatusCode::NOT_FOUND)
3728 0 : }
3729 :
3730 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3731 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3732 : /// the tenant's policies (configuration) within the storage controller
3733 0 : pub(crate) async fn tenant_update_policy(
3734 0 : &self,
3735 0 : tenant_id: TenantId,
3736 0 : req: TenantPolicyRequest,
3737 0 : ) -> Result<(), ApiError> {
3738 : // We require an exclusive lock, because we are updating persistent and in-memory state
3739 0 : let _tenant_lock = trace_exclusive_lock(
3740 0 : &self.tenant_op_locks,
3741 0 : tenant_id,
3742 0 : TenantOperations::UpdatePolicy,
3743 0 : )
3744 0 : .await;
3745 :
3746 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3747 :
3748 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3749 :
3750 : let TenantPolicyRequest {
3751 0 : placement,
3752 0 : mut scheduling,
3753 0 : } = req;
3754 :
3755 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3756 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3757 : // this to take effect.
3758 0 : if scheduling.is_none() {
3759 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3760 0 : }
3761 0 : }
3762 :
3763 0 : self.persistence
3764 0 : .update_tenant_shard(
3765 0 : TenantFilter::Tenant(tenant_id),
3766 0 : placement.clone(),
3767 0 : None,
3768 0 : None,
3769 0 : scheduling,
3770 0 : )
3771 0 : .await?;
3772 :
3773 0 : let mut schedule_context = ScheduleContext::default();
3774 0 : let mut locked = self.inner.write().unwrap();
3775 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3776 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3777 0 : if let Some(placement) = &placement {
3778 0 : shard.policy = placement.clone();
3779 :
3780 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3781 0 : "Updated placement policy to {placement:?}");
3782 0 : }
3783 :
3784 0 : if let Some(scheduling) = &scheduling {
3785 0 : shard.set_scheduling_policy(*scheduling);
3786 :
3787 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3788 0 : "Updated scheduling policy to {scheduling:?}");
3789 0 : }
3790 :
3791 : // In case scheduling is being switched back on, try it now.
3792 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3793 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3794 : }
3795 :
3796 0 : Ok(())
3797 0 : }
3798 :
3799 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3800 0 : &self,
3801 0 : tenant_id: TenantId,
3802 0 : mut create_req: TimelineCreateRequest,
3803 0 : ) -> Result<TimelineInfo, ApiError> {
3804 0 : tracing::info!(
3805 0 : "Creating timeline {}/{}",
3806 : tenant_id,
3807 : create_req.new_timeline_id,
3808 : );
3809 :
3810 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3811 0 : if targets.0.is_empty() {
3812 0 : return Err(ApiError::NotFound(
3813 0 : anyhow::anyhow!("Tenant not found").into(),
3814 0 : ));
3815 0 : };
3816 :
3817 0 : let (shard_zero_tid, shard_zero_locations) =
3818 0 : targets.0.pop_first().expect("Must have at least one shard");
3819 0 : assert!(shard_zero_tid.is_shard_zero());
3820 :
3821 0 : async fn create_one(
3822 0 : tenant_shard_id: TenantShardId,
3823 0 : locations: ShardMutationLocations,
3824 0 : http_client: reqwest::Client,
3825 0 : jwt: Option<String>,
3826 0 : mut create_req: TimelineCreateRequest,
3827 0 : ) -> Result<TimelineInfo, ApiError> {
3828 0 : let latest = locations.latest.node;
3829 :
3830 0 : tracing::info!(
3831 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3832 : tenant_shard_id,
3833 : create_req.new_timeline_id,
3834 : locations.latest.generation
3835 : );
3836 :
3837 0 : let client =
3838 0 : PageserverClient::new(latest.get_id(), http_client.clone(), latest.base_url(), jwt.as_deref());
3839 :
3840 0 : let timeline_info = client
3841 0 : .timeline_create(tenant_shard_id, &create_req)
3842 0 : .await
3843 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3844 :
3845 : // If we are going to create the timeline on some stale locations for shard 0, then ask them to re-use
3846 : // the initdb generated by the latest location, rather than generating their own. This avoids racing uploads
3847 : // of initdb to S3 which might not be binary-identical if different pageservers have different postgres binaries.
3848 0 : if tenant_shard_id.is_shard_zero() {
3849 0 : if let models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } = &mut create_req.mode {
3850 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id);
3851 0 : }
3852 0 : }
3853 :
3854 : // We propagate timeline creations to all attached locations such that a compute
3855 : // for the new timeline is able to start regardless of the current state of the
3856 : // tenant shard reconciliation.
3857 0 : for location in locations.other {
3858 0 : tracing::info!(
3859 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3860 : tenant_shard_id,
3861 : create_req.new_timeline_id,
3862 : location.node,
3863 : location.generation
3864 : );
3865 :
3866 0 : let client = PageserverClient::new(
3867 0 : location.node.get_id(),
3868 0 : http_client.clone(),
3869 0 : location.node.base_url(),
3870 0 : jwt.as_deref(),
3871 : );
3872 :
3873 0 : let res = client
3874 0 : .timeline_create(tenant_shard_id, &create_req)
3875 0 : .await;
3876 :
3877 0 : if let Err(e) = res {
3878 0 : match e {
3879 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3880 0 : // Tenant might have been detached from the stale location,
3881 0 : // so ignore 404s.
3882 0 : },
3883 : _ => {
3884 0 : return Err(passthrough_api_error(&location.node, e));
3885 : }
3886 : }
3887 0 : }
3888 : }
3889 :
3890 0 : Ok(timeline_info)
3891 0 : }
3892 :
3893 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3894 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3895 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3896 : //
3897 : // This also enables non-zero shards to use the initdb that shard 0 generated and uploaded to S3, rather than
3898 : // independently generating their own initdb. This guarantees that shards cannot end up with different initial
3899 : // states if e.g. they have different postgres binary versions.
3900 0 : let timeline_info = create_one(
3901 0 : shard_zero_tid,
3902 0 : shard_zero_locations,
3903 0 : self.http_client.clone(),
3904 0 : self.config.pageserver_jwt_token.clone(),
3905 0 : create_req.clone(),
3906 0 : )
3907 0 : .await?;
3908 :
3909 : // Update the create request for shards >= 0
3910 0 : match &mut create_req.mode {
3911 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3912 0 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3913 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3914 0 : },
3915 0 : models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } => {
3916 : // For shards >= 0, do not run initdb: use the one that shard 0 uploaded to S3
3917 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id)
3918 : }
3919 0 : _ => {}
3920 : }
3921 :
3922 : // Create timeline on remaining shards with number >0
3923 0 : if !targets.0.is_empty() {
3924 : // If we had multiple shards, issue requests for the remainder now.
3925 0 : let jwt = &self.config.pageserver_jwt_token;
3926 0 : self.tenant_for_shards(
3927 0 : targets
3928 0 : .0
3929 0 : .iter()
3930 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3931 0 : .collect(),
3932 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3933 0 : let create_req = create_req.clone();
3934 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3935 0 : Box::pin(create_one(
3936 0 : tenant_shard_id,
3937 0 : mutation_locations,
3938 0 : self.http_client.clone(),
3939 0 : jwt.clone(),
3940 0 : create_req,
3941 0 : ))
3942 0 : },
3943 : )
3944 0 : .await?;
3945 0 : }
3946 :
3947 0 : Ok(timeline_info)
3948 0 : })
3949 0 : .await?
3950 0 : }
3951 :
3952 0 : pub(crate) async fn tenant_timeline_create(
3953 0 : self: &Arc<Self>,
3954 0 : tenant_id: TenantId,
3955 0 : create_req: TimelineCreateRequest,
3956 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
3957 0 : let safekeepers = self.config.timelines_onto_safekeepers;
3958 0 : let timeline_id = create_req.new_timeline_id;
3959 :
3960 0 : tracing::info!(
3961 0 : mode=%create_req.mode_tag(),
3962 : %safekeepers,
3963 0 : "Creating timeline {}/{}",
3964 : tenant_id,
3965 : timeline_id,
3966 : );
3967 :
3968 0 : let _tenant_lock = trace_shared_lock(
3969 0 : &self.tenant_op_locks,
3970 0 : tenant_id,
3971 0 : TenantOperations::TimelineCreate,
3972 0 : )
3973 0 : .await;
3974 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3975 0 : let is_import = create_req.is_import();
3976 0 : let read_only = matches!(
3977 0 : create_req.mode,
3978 : models::TimelineCreateRequestMode::Branch {
3979 : read_only: true,
3980 : ..
3981 : }
3982 : );
3983 :
3984 0 : if is_import {
3985 : // Ensure that there is no split on-going.
3986 : // [`Self::tenant_shard_split`] holds the exclusive tenant lock
3987 : // for the duration of the split, but here we handle the case
3988 : // where we restarted and the split is being aborted.
3989 0 : let locked = self.inner.read().unwrap();
3990 0 : let splitting = locked
3991 0 : .tenants
3992 0 : .range(TenantShardId::tenant_range(tenant_id))
3993 0 : .any(|(_id, shard)| shard.splitting != SplitState::Idle);
3994 :
3995 0 : if splitting {
3996 0 : return Err(ApiError::Conflict("Tenant is splitting shard".to_string()));
3997 0 : }
3998 0 : }
3999 :
4000 0 : let timeline_info = self
4001 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
4002 0 : .await?;
4003 :
4004 0 : let selected_safekeepers = if is_import {
4005 0 : let shards = {
4006 0 : let locked = self.inner.read().unwrap();
4007 0 : locked
4008 0 : .tenants
4009 0 : .range(TenantShardId::tenant_range(tenant_id))
4010 0 : .map(|(ts_id, _)| ts_id.to_index())
4011 0 : .collect::<Vec<_>>()
4012 : };
4013 :
4014 0 : if !shards
4015 0 : .iter()
4016 0 : .map(|shard_index| shard_index.shard_count)
4017 0 : .all_equal()
4018 : {
4019 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4020 0 : "Inconsistent shard count"
4021 0 : )));
4022 0 : }
4023 :
4024 0 : let import = TimelineImport {
4025 0 : tenant_id,
4026 0 : timeline_id,
4027 0 : shard_statuses: ShardImportStatuses::new(shards),
4028 0 : };
4029 :
4030 0 : let inserted = self
4031 0 : .persistence
4032 0 : .insert_timeline_import(import.to_persistent())
4033 0 : .await
4034 0 : .context("timeline import insert")
4035 0 : .map_err(ApiError::InternalServerError)?;
4036 :
4037 : // Set the importing flag on the tenant shards
4038 0 : self.inner
4039 0 : .write()
4040 0 : .unwrap()
4041 0 : .tenants
4042 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4043 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Importing);
4044 :
4045 0 : match inserted {
4046 : true => {
4047 0 : tracing::info!(%tenant_id, %timeline_id, "Inserted timeline import");
4048 : }
4049 : false => {
4050 0 : tracing::info!(%tenant_id, %timeline_id, "Timeline import entry already present");
4051 : }
4052 : }
4053 :
4054 0 : None
4055 0 : } else if safekeepers || read_only {
4056 : // Note that for imported timelines, we do not create the timeline on the safekeepers
4057 : // straight away. Instead, we do it once the import finalized such that we know what
4058 : // start LSN to provide for the safekeepers. This is done in
4059 : // [`Self::finalize_timeline_import`].
4060 0 : let res = self
4061 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
4062 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
4063 0 : .await?;
4064 0 : Some(res)
4065 : } else {
4066 0 : None
4067 : };
4068 :
4069 0 : Ok(TimelineCreateResponseStorcon {
4070 0 : timeline_info,
4071 0 : safekeepers: selected_safekeepers,
4072 0 : })
4073 0 : }
4074 :
4075 : #[instrument(skip_all, fields(
4076 : tenant_id=%req.tenant_shard_id.tenant_id,
4077 : shard_id=%req.tenant_shard_id.shard_slug(),
4078 : timeline_id=%req.timeline_id,
4079 : ))]
4080 : pub(crate) async fn handle_timeline_shard_import_progress(
4081 : self: &Arc<Self>,
4082 : req: TimelineImportStatusRequest,
4083 : ) -> Result<ShardImportStatus, ApiError> {
4084 : let validity = self
4085 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4086 : .await?;
4087 : match validity {
4088 : ShardGenerationValidity::Valid => {
4089 : // fallthrough
4090 : }
4091 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4092 : tracing::info!(
4093 : claimed=?claimed.into(),
4094 0 : actual=?actual.and_then(|g| g.into()),
4095 : "Rejecting import progress fetch from stale generation"
4096 : );
4097 :
4098 : return Err(ApiError::BadRequest(anyhow::anyhow!("Invalid generation")));
4099 : }
4100 : }
4101 :
4102 : let maybe_import = self
4103 : .persistence
4104 : .get_timeline_import(req.tenant_shard_id.tenant_id, req.timeline_id)
4105 : .await?;
4106 :
4107 0 : let import = maybe_import.ok_or_else(|| {
4108 0 : ApiError::NotFound(
4109 0 : format!(
4110 0 : "import for {}/{} not found",
4111 0 : req.tenant_shard_id.tenant_id, req.timeline_id
4112 0 : )
4113 0 : .into(),
4114 0 : )
4115 0 : })?;
4116 :
4117 : import
4118 : .shard_statuses
4119 : .0
4120 : .get(&req.tenant_shard_id.to_index())
4121 : .cloned()
4122 0 : .ok_or_else(|| {
4123 0 : ApiError::NotFound(
4124 0 : format!("shard {} not found", req.tenant_shard_id.shard_slug()).into(),
4125 0 : )
4126 0 : })
4127 : }
4128 :
4129 : #[instrument(skip_all, fields(
4130 : tenant_id=%req.tenant_shard_id.tenant_id,
4131 : shard_id=%req.tenant_shard_id.shard_slug(),
4132 : timeline_id=%req.timeline_id,
4133 : ))]
4134 : pub(crate) async fn handle_timeline_shard_import_progress_upcall(
4135 : self: &Arc<Self>,
4136 : req: PutTimelineImportStatusRequest,
4137 : ) -> Result<(), ApiError> {
4138 : let validity = self
4139 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4140 : .await?;
4141 : match validity {
4142 : ShardGenerationValidity::Valid => {
4143 : // fallthrough
4144 : }
4145 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4146 : tracing::info!(
4147 : claimed=?claimed.into(),
4148 0 : actual=?actual.and_then(|g| g.into()),
4149 : "Rejecting import progress update from stale generation"
4150 : );
4151 :
4152 : return Err(ApiError::PreconditionFailed("Invalid generation".into()));
4153 : }
4154 : }
4155 :
4156 : let res = self
4157 : .persistence
4158 : .update_timeline_import(req.tenant_shard_id, req.timeline_id, req.status)
4159 : .await;
4160 : let timeline_import = match res {
4161 : Ok(Ok(Some(timeline_import))) => timeline_import,
4162 : Ok(Ok(None)) => {
4163 : // Idempotency: we've already seen and handled this update.
4164 : return Ok(());
4165 : }
4166 : Ok(Err(logical_err)) => {
4167 : return Err(logical_err.into());
4168 : }
4169 : Err(db_err) => {
4170 : return Err(db_err.into());
4171 : }
4172 : };
4173 :
4174 : tracing::info!(
4175 : tenant_id=%req.tenant_shard_id.tenant_id,
4176 : timeline_id=%req.timeline_id,
4177 : shard_id=%req.tenant_shard_id.shard_slug(),
4178 : "Updated timeline import status to: {timeline_import:?}");
4179 :
4180 : if timeline_import.is_complete() {
4181 : tokio::task::spawn({
4182 : let this = self.clone();
4183 0 : async move { this.finalize_timeline_import(timeline_import).await }
4184 : });
4185 : }
4186 :
4187 : Ok(())
4188 : }
4189 :
4190 : /// Check that a provided generation for some tenant shard is the most recent one.
4191 : ///
4192 : /// Validate with the in-mem state first, and, if that passes, validate with the
4193 : /// database state which is authoritative.
4194 0 : async fn validate_shard_generation(
4195 0 : self: &Arc<Self>,
4196 0 : tenant_shard_id: TenantShardId,
4197 0 : generation: Generation,
4198 0 : ) -> Result<ShardGenerationValidity, ApiError> {
4199 : {
4200 0 : let locked = self.inner.read().unwrap();
4201 0 : let tenant_shard =
4202 0 : locked
4203 0 : .tenants
4204 0 : .get(&tenant_shard_id)
4205 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4206 0 : "{} shard not found",
4207 0 : tenant_shard_id
4208 0 : )))?;
4209 :
4210 0 : if tenant_shard.generation != Some(generation) {
4211 0 : return Ok(ShardGenerationValidity::Mismatched {
4212 0 : claimed: generation,
4213 0 : actual: tenant_shard.generation,
4214 0 : });
4215 0 : }
4216 : }
4217 :
4218 0 : let mut db_generations = self
4219 0 : .persistence
4220 0 : .shard_generations(std::iter::once(&tenant_shard_id))
4221 0 : .await?;
4222 0 : let (_tid, db_generation) =
4223 0 : db_generations
4224 0 : .pop()
4225 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4226 0 : "{} shard not found",
4227 0 : tenant_shard_id
4228 0 : )))?;
4229 :
4230 0 : if db_generation != Some(generation) {
4231 0 : return Ok(ShardGenerationValidity::Mismatched {
4232 0 : claimed: generation,
4233 0 : actual: db_generation,
4234 0 : });
4235 0 : }
4236 :
4237 0 : Ok(ShardGenerationValidity::Valid)
4238 0 : }
4239 :
4240 : /// Finalize the import of a timeline
4241 : ///
4242 : /// This method should be called once all shards have reported that the import is complete.
4243 : /// Firstly, it polls the post import timeline activation endpoint exposed by the pageserver.
4244 : /// Once the timeline is active on all shards, the timeline also gets created on the
4245 : /// safekeepers. Finally, notify cplane of the import completion (whether failed or
4246 : /// successful), and remove the import from the database and in-memory.
4247 : ///
4248 : /// If this method gets pre-empted by shut down, it will be called again at start-up (on-going
4249 : /// imports are stored in the database).
4250 : ///
4251 : /// # Cancel-Safety
4252 : /// Not cancel safe.
4253 : /// If the caller stops polling, the import will not be removed from
4254 : /// [`ServiceState::imports_finalizing`].
4255 : #[instrument(skip_all, fields(
4256 : tenant_id=%import.tenant_id,
4257 : timeline_id=%import.timeline_id,
4258 : ))]
4259 :
4260 : async fn finalize_timeline_import(
4261 : self: &Arc<Self>,
4262 : import: TimelineImport,
4263 : ) -> Result<(), TimelineImportFinalizeError> {
4264 : let tenant_timeline = (import.tenant_id, import.timeline_id);
4265 :
4266 : let (_finalize_import_guard, cancel) = {
4267 : let mut locked = self.inner.write().unwrap();
4268 : let gate = Gate::default();
4269 : let cancel = CancellationToken::default();
4270 :
4271 : let guard = gate.enter().unwrap();
4272 :
4273 : locked.imports_finalizing.insert(
4274 : tenant_timeline,
4275 : FinalizingImport {
4276 : gate,
4277 : cancel: cancel.clone(),
4278 : },
4279 : );
4280 :
4281 : (guard, cancel)
4282 : };
4283 :
4284 : let res = tokio::select! {
4285 : res = self.finalize_timeline_import_impl(import) => {
4286 : res
4287 : },
4288 : _ = cancel.cancelled() => {
4289 : Err(TimelineImportFinalizeError::Cancelled)
4290 : }
4291 : };
4292 :
4293 : let mut locked = self.inner.write().unwrap();
4294 : locked.imports_finalizing.remove(&tenant_timeline);
4295 :
4296 : res
4297 : }
4298 :
4299 0 : async fn finalize_timeline_import_impl(
4300 0 : self: &Arc<Self>,
4301 0 : import: TimelineImport,
4302 0 : ) -> Result<(), TimelineImportFinalizeError> {
4303 0 : tracing::info!("Finalizing timeline import");
4304 :
4305 0 : pausable_failpoint!("timeline-import-pre-cplane-notification");
4306 :
4307 0 : let tenant_id = import.tenant_id;
4308 0 : let timeline_id = import.timeline_id;
4309 :
4310 0 : let import_error = import.completion_error();
4311 0 : match import_error {
4312 0 : Some(err) => {
4313 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Err(err))
4314 0 : .await?;
4315 0 : tracing::warn!("Timeline import completed with shard errors");
4316 0 : Ok(())
4317 : }
4318 0 : None => match self.activate_timeline_post_import(&import).await {
4319 0 : Ok(timeline_info) => {
4320 0 : tracing::info!("Post import timeline activation complete");
4321 :
4322 0 : if self.config.timelines_onto_safekeepers {
4323 : // Now that we know the start LSN of this timeline, create it on the
4324 : // safekeepers.
4325 0 : self.tenant_timeline_create_safekeepers_until_success(
4326 0 : import.tenant_id,
4327 0 : timeline_info,
4328 0 : )
4329 0 : .await?;
4330 0 : }
4331 :
4332 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Ok(()))
4333 0 : .await?;
4334 :
4335 0 : tracing::info!("Timeline import completed successfully");
4336 0 : Ok(())
4337 : }
4338 : Err(TimelineImportFinalizeError::ShuttingDown) => {
4339 : // We got pre-empted by shut down and will resume after the restart.
4340 0 : Err(TimelineImportFinalizeError::ShuttingDown)
4341 : }
4342 0 : Err(err) => {
4343 : // Any finalize error apart from shut down is permanent and requires us to notify
4344 : // cplane such that it can clean up.
4345 0 : tracing::error!("Import finalize failed with permanent error: {err}");
4346 0 : self.notify_cplane_and_delete_import(
4347 0 : tenant_id,
4348 0 : timeline_id,
4349 0 : Err(err.to_string()),
4350 0 : )
4351 0 : .await?;
4352 0 : Err(err)
4353 : }
4354 : },
4355 : }
4356 0 : }
4357 :
4358 0 : async fn notify_cplane_and_delete_import(
4359 0 : self: &Arc<Self>,
4360 0 : tenant_id: TenantId,
4361 0 : timeline_id: TimelineId,
4362 0 : import_result: ImportResult,
4363 0 : ) -> Result<(), TimelineImportFinalizeError> {
4364 0 : let import_failed = import_result.is_err();
4365 0 : tracing::info!(%import_failed, "Notifying cplane of import completion");
4366 :
4367 0 : let client = UpcallClient::new(self.get_config(), self.cancel.child_token());
4368 0 : client
4369 0 : .notify_import_complete(tenant_id, timeline_id, import_result)
4370 0 : .await
4371 0 : .map_err(|_err| TimelineImportFinalizeError::ShuttingDown)?;
4372 :
4373 0 : if let Err(err) = self
4374 0 : .persistence
4375 0 : .delete_timeline_import(tenant_id, timeline_id)
4376 0 : .await
4377 : {
4378 0 : tracing::warn!("Failed to delete timeline import entry from database: {err}");
4379 0 : }
4380 :
4381 0 : self.inner
4382 0 : .write()
4383 0 : .unwrap()
4384 0 : .tenants
4385 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4386 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Idle);
4387 :
4388 0 : Ok(())
4389 0 : }
4390 :
4391 : /// Activate an imported timeline on all shards once the import is complete.
4392 : /// Returns the [`TimelineInfo`] reported by shard zero.
4393 0 : async fn activate_timeline_post_import(
4394 0 : self: &Arc<Self>,
4395 0 : import: &TimelineImport,
4396 0 : ) -> Result<TimelineInfo, TimelineImportFinalizeError> {
4397 : const TIMELINE_ACTIVATE_TIMEOUT: Duration = Duration::from_millis(128);
4398 :
4399 0 : let mut shards_to_activate: HashSet<ShardIndex> =
4400 0 : import.shard_statuses.0.keys().cloned().collect();
4401 0 : let mut shard_zero_timeline_info = None;
4402 :
4403 0 : while !shards_to_activate.is_empty() {
4404 0 : if self.cancel.is_cancelled() {
4405 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4406 0 : }
4407 :
4408 0 : let targets = {
4409 0 : let locked = self.inner.read().unwrap();
4410 0 : let mut targets = Vec::new();
4411 :
4412 0 : for (tenant_shard_id, shard) in locked
4413 0 : .tenants
4414 0 : .range(TenantShardId::tenant_range(import.tenant_id))
4415 : {
4416 0 : if !import
4417 0 : .shard_statuses
4418 0 : .0
4419 0 : .contains_key(&tenant_shard_id.to_index())
4420 : {
4421 0 : return Err(TimelineImportFinalizeError::MismatchedShards(
4422 0 : tenant_shard_id.to_index(),
4423 0 : ));
4424 0 : }
4425 :
4426 0 : if let Some(node_id) = shard.intent.get_attached() {
4427 0 : let node = locked
4428 0 : .nodes
4429 0 : .get(node_id)
4430 0 : .expect("Pageservers may not be deleted while referenced");
4431 0 : targets.push((*tenant_shard_id, node.clone()));
4432 0 : }
4433 : }
4434 :
4435 0 : targets
4436 : };
4437 :
4438 0 : let targeted_tenant_shards: Vec<_> = targets.iter().map(|(tid, _node)| *tid).collect();
4439 :
4440 0 : let results = self
4441 0 : .tenant_for_shards_api(
4442 0 : targets,
4443 0 : |tenant_shard_id, client| async move {
4444 0 : client
4445 0 : .activate_post_import(
4446 0 : tenant_shard_id,
4447 0 : import.timeline_id,
4448 0 : TIMELINE_ACTIVATE_TIMEOUT,
4449 0 : )
4450 0 : .await
4451 0 : },
4452 : 1,
4453 : 1,
4454 : SHORT_RECONCILE_TIMEOUT,
4455 0 : &self.cancel,
4456 : )
4457 0 : .await;
4458 :
4459 0 : let mut failed = 0;
4460 0 : for (tid, (_, result)) in targeted_tenant_shards.iter().zip(results.into_iter()) {
4461 0 : match result {
4462 0 : Ok(ok) => {
4463 0 : if tid.is_shard_zero() {
4464 0 : shard_zero_timeline_info = Some(ok);
4465 0 : }
4466 :
4467 0 : shards_to_activate.remove(&tid.to_index());
4468 : }
4469 0 : Err(_err) => {
4470 0 : failed += 1;
4471 0 : }
4472 : }
4473 : }
4474 :
4475 0 : if failed > 0 {
4476 0 : tracing::info!(
4477 0 : "Failed to activate timeline on {failed} shards post import. Will retry"
4478 : );
4479 0 : }
4480 :
4481 0 : tokio::select! {
4482 0 : _ = tokio::time::sleep(Duration::from_millis(250)) => {},
4483 0 : _ = self.cancel.cancelled() => {
4484 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4485 : }
4486 : }
4487 : }
4488 :
4489 0 : Ok(shard_zero_timeline_info.expect("All shards replied"))
4490 0 : }
4491 :
4492 0 : async fn finalize_timeline_imports(self: &Arc<Self>, imports: Vec<TimelineImport>) {
4493 0 : futures::future::join_all(
4494 0 : imports
4495 0 : .into_iter()
4496 0 : .map(|import| self.finalize_timeline_import(import)),
4497 : )
4498 0 : .await;
4499 0 : }
4500 :
4501 : /// Delete a timeline import if it exists
4502 : ///
4503 : /// Firstly, delete the entry from the database. Any updates
4504 : /// from pageservers after the update will fail with a 404, so the
4505 : /// import cannot progress into finalizing state if it's not there already.
4506 : /// Secondly, cancel the finalization if one is in progress.
4507 0 : pub(crate) async fn maybe_delete_timeline_import(
4508 0 : self: &Arc<Self>,
4509 0 : tenant_id: TenantId,
4510 0 : timeline_id: TimelineId,
4511 0 : ) -> Result<(), DatabaseError> {
4512 0 : let tenant_has_ongoing_import = {
4513 0 : let locked = self.inner.read().unwrap();
4514 0 : locked
4515 0 : .tenants
4516 0 : .range(TenantShardId::tenant_range(tenant_id))
4517 0 : .any(|(_tid, shard)| shard.importing == TimelineImportState::Importing)
4518 : };
4519 :
4520 0 : if !tenant_has_ongoing_import {
4521 0 : return Ok(());
4522 0 : }
4523 :
4524 0 : self.persistence
4525 0 : .delete_timeline_import(tenant_id, timeline_id)
4526 0 : .await?;
4527 :
4528 0 : let maybe_finalizing = {
4529 0 : let mut locked = self.inner.write().unwrap();
4530 0 : locked.imports_finalizing.remove(&(tenant_id, timeline_id))
4531 : };
4532 :
4533 0 : if let Some(finalizing) = maybe_finalizing {
4534 0 : finalizing.cancel.cancel();
4535 0 : finalizing.gate.close().await;
4536 0 : }
4537 :
4538 0 : Ok(())
4539 0 : }
4540 :
4541 0 : pub(crate) async fn tenant_timeline_archival_config(
4542 0 : &self,
4543 0 : tenant_id: TenantId,
4544 0 : timeline_id: TimelineId,
4545 0 : req: TimelineArchivalConfigRequest,
4546 0 : ) -> Result<(), ApiError> {
4547 0 : tracing::info!(
4548 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
4549 : req.state
4550 : );
4551 :
4552 0 : let _tenant_lock = trace_shared_lock(
4553 0 : &self.tenant_op_locks,
4554 0 : tenant_id,
4555 0 : TenantOperations::TimelineArchivalConfig,
4556 0 : )
4557 0 : .await;
4558 :
4559 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4560 0 : if targets.0.is_empty() {
4561 0 : return Err(ApiError::NotFound(
4562 0 : anyhow::anyhow!("Tenant not found").into(),
4563 0 : ));
4564 0 : }
4565 0 : async fn config_one(
4566 0 : tenant_shard_id: TenantShardId,
4567 0 : timeline_id: TimelineId,
4568 0 : node: Node,
4569 0 : http_client: reqwest::Client,
4570 0 : jwt: Option<String>,
4571 0 : req: TimelineArchivalConfigRequest,
4572 0 : ) -> Result<(), ApiError> {
4573 0 : tracing::info!(
4574 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4575 : );
4576 :
4577 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4578 :
4579 0 : client
4580 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
4581 0 : .await
4582 0 : .map_err(|e| match e {
4583 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
4584 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
4585 : }
4586 0 : _ => passthrough_api_error(&node, e),
4587 0 : })
4588 0 : }
4589 :
4590 : // no shard needs to go first/last; the operation should be idempotent
4591 : // TODO: it would be great to ensure that all shards return the same error
4592 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4593 0 : let results = self
4594 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4595 0 : futures::FutureExt::boxed(config_one(
4596 0 : tenant_shard_id,
4597 0 : timeline_id,
4598 0 : node,
4599 0 : self.http_client.clone(),
4600 0 : self.config.pageserver_jwt_token.clone(),
4601 0 : req.clone(),
4602 0 : ))
4603 0 : })
4604 0 : .await?;
4605 0 : assert!(!results.is_empty(), "must have at least one result");
4606 :
4607 0 : Ok(())
4608 0 : }).await?
4609 0 : }
4610 :
4611 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
4612 0 : &self,
4613 0 : tenant_id: TenantId,
4614 0 : timeline_id: TimelineId,
4615 0 : behavior: Option<DetachBehavior>,
4616 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
4617 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
4618 :
4619 0 : let _tenant_lock = trace_shared_lock(
4620 0 : &self.tenant_op_locks,
4621 0 : tenant_id,
4622 0 : TenantOperations::TimelineDetachAncestor,
4623 0 : )
4624 0 : .await;
4625 :
4626 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4627 0 : if targets.0.is_empty() {
4628 0 : return Err(ApiError::NotFound(
4629 0 : anyhow::anyhow!("Tenant not found").into(),
4630 0 : ));
4631 0 : }
4632 :
4633 0 : async fn detach_one(
4634 0 : tenant_shard_id: TenantShardId,
4635 0 : timeline_id: TimelineId,
4636 0 : node: Node,
4637 0 : http_client: reqwest::Client,
4638 0 : jwt: Option<String>,
4639 0 : behavior: Option<DetachBehavior>,
4640 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
4641 0 : tracing::info!(
4642 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4643 : );
4644 :
4645 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4646 :
4647 0 : client
4648 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id, behavior)
4649 0 : .await
4650 0 : .map_err(|e| {
4651 : use mgmt_api::Error;
4652 :
4653 0 : match e {
4654 : // no ancestor (ever)
4655 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
4656 0 : "{node}: {}",
4657 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
4658 0 : )),
4659 : // too many ancestors
4660 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
4661 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
4662 : }
4663 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
4664 : // avoid turning these into conflicts to remain compatible with
4665 : // pageservers, 500 errors are sadly retryable with timeline ancestor
4666 : // detach
4667 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
4668 : }
4669 : // rest can be mapped as usual
4670 0 : other => passthrough_api_error(&node, other),
4671 : }
4672 0 : })
4673 0 : .map(|res| (tenant_shard_id.shard_number, res))
4674 0 : }
4675 :
4676 : // no shard needs to go first/last; the operation should be idempotent
4677 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4678 0 : let mut results = self
4679 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4680 0 : futures::FutureExt::boxed(detach_one(
4681 0 : tenant_shard_id,
4682 0 : timeline_id,
4683 0 : node,
4684 0 : self.http_client.clone(),
4685 0 : self.config.pageserver_jwt_token.clone(),
4686 0 : behavior,
4687 0 : ))
4688 0 : })
4689 0 : .await?;
4690 :
4691 0 : let any = results.pop().expect("we must have at least one response");
4692 :
4693 0 : let mismatching = results
4694 0 : .iter()
4695 0 : .filter(|(_, res)| res != &any.1)
4696 0 : .collect::<Vec<_>>();
4697 0 : if !mismatching.is_empty() {
4698 : // this can be hit by races which should not happen because operation lock on cplane
4699 0 : let matching = results.len() - mismatching.len();
4700 0 : tracing::error!(
4701 : matching,
4702 : compared_against=?any,
4703 : ?mismatching,
4704 0 : "shards returned different results"
4705 : );
4706 :
4707 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
4708 0 : }
4709 :
4710 0 : Ok(any.1)
4711 0 : }).await?
4712 0 : }
4713 :
4714 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
4715 0 : &self,
4716 0 : tenant_id: TenantId,
4717 0 : timeline_id: TimelineId,
4718 0 : dir: BlockUnblock,
4719 0 : ) -> Result<(), ApiError> {
4720 0 : let _tenant_lock = trace_shared_lock(
4721 0 : &self.tenant_op_locks,
4722 0 : tenant_id,
4723 0 : TenantOperations::TimelineGcBlockUnblock,
4724 0 : )
4725 0 : .await;
4726 :
4727 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4728 0 : if targets.0.is_empty() {
4729 0 : return Err(ApiError::NotFound(
4730 0 : anyhow::anyhow!("Tenant not found").into(),
4731 0 : ));
4732 0 : }
4733 :
4734 0 : async fn do_one(
4735 0 : tenant_shard_id: TenantShardId,
4736 0 : timeline_id: TimelineId,
4737 0 : node: Node,
4738 0 : http_client: reqwest::Client,
4739 0 : jwt: Option<String>,
4740 0 : dir: BlockUnblock,
4741 0 : ) -> Result<(), ApiError> {
4742 0 : let client = PageserverClient::new(
4743 0 : node.get_id(),
4744 0 : http_client,
4745 0 : node.base_url(),
4746 0 : jwt.as_deref(),
4747 : );
4748 :
4749 0 : client
4750 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
4751 0 : .await
4752 0 : .map_err(|e| passthrough_api_error(&node, e))
4753 0 : }
4754 :
4755 : // no shard needs to go first/last; the operation should be idempotent
4756 0 : let locations = targets
4757 0 : .0
4758 0 : .iter()
4759 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
4760 0 : .collect();
4761 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
4762 0 : futures::FutureExt::boxed(do_one(
4763 0 : tenant_shard_id,
4764 0 : timeline_id,
4765 0 : node,
4766 0 : self.http_client.clone(),
4767 0 : self.config.pageserver_jwt_token.clone(),
4768 0 : dir,
4769 0 : ))
4770 0 : })
4771 0 : .await
4772 0 : })
4773 0 : .await??;
4774 0 : Ok(())
4775 0 : }
4776 :
4777 0 : pub(crate) fn is_tenant_not_found_error(body: &str, tenant_id: TenantId) -> bool {
4778 0 : body.contains(&format!("tenant {tenant_id}"))
4779 0 : }
4780 :
4781 0 : fn process_result_and_passthrough_errors<T>(
4782 0 : &self,
4783 0 : tenant_id: TenantId,
4784 0 : results: Vec<(Node, Result<T, mgmt_api::Error>)>,
4785 0 : ) -> Result<Vec<(Node, T)>, ApiError> {
4786 0 : let mut processed_results: Vec<(Node, T)> = Vec::with_capacity(results.len());
4787 0 : for (node, res) in results {
4788 0 : match res {
4789 0 : Ok(res) => processed_results.push((node, res)),
4790 0 : Err(mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, body))
4791 0 : if Self::is_tenant_not_found_error(&body, tenant_id) =>
4792 : {
4793 : // If there's a tenant not found, we are still in the process of attaching the tenant.
4794 : // Return 503 so that the client can retry.
4795 0 : return Err(ApiError::ResourceUnavailable(
4796 0 : format!(
4797 0 : "Timeline is not attached to the pageserver {} yet, please retry",
4798 0 : node.get_id()
4799 0 : )
4800 0 : .into(),
4801 0 : ));
4802 : }
4803 0 : Err(e) => return Err(passthrough_api_error(&node, e)),
4804 : }
4805 : }
4806 0 : Ok(processed_results)
4807 0 : }
4808 :
4809 0 : pub(crate) async fn tenant_timeline_lsn_lease(
4810 0 : &self,
4811 0 : tenant_id: TenantId,
4812 0 : timeline_id: TimelineId,
4813 0 : lsn: Lsn,
4814 0 : ) -> Result<LsnLease, ApiError> {
4815 0 : let _tenant_lock = trace_shared_lock(
4816 0 : &self.tenant_op_locks,
4817 0 : tenant_id,
4818 0 : TenantOperations::TimelineLsnLease,
4819 0 : )
4820 0 : .await;
4821 :
4822 0 : self.tenant_remote_mutation(tenant_id, |locations| async move {
4823 0 : if locations.0.is_empty() {
4824 0 : return Err(ApiError::NotFound(
4825 0 : anyhow::anyhow!("Tenant not found").into(),
4826 0 : ));
4827 0 : }
4828 :
4829 0 : let results = self
4830 0 : .tenant_for_shards_api(
4831 0 : locations
4832 0 : .0
4833 0 : .iter()
4834 0 : .map(|(tenant_shard_id, ShardMutationLocations { latest, .. })| {
4835 0 : (*tenant_shard_id, latest.node.clone())
4836 0 : })
4837 0 : .collect(),
4838 0 : |tenant_shard_id, client| async move {
4839 0 : client
4840 0 : .timeline_lease_lsn(tenant_shard_id, timeline_id, lsn)
4841 0 : .await
4842 0 : },
4843 : 1,
4844 : 1,
4845 : SHORT_RECONCILE_TIMEOUT,
4846 0 : &self.cancel,
4847 : )
4848 0 : .await;
4849 :
4850 0 : let leases = self.process_result_and_passthrough_errors(tenant_id, results)?;
4851 0 : let mut valid_until = None;
4852 0 : for (_, lease) in leases {
4853 0 : if let Some(ref mut valid_until) = valid_until {
4854 0 : *valid_until = std::cmp::min(*valid_until, lease.valid_until);
4855 0 : } else {
4856 0 : valid_until = Some(lease.valid_until);
4857 0 : }
4858 : }
4859 0 : Ok(LsnLease {
4860 0 : valid_until: valid_until.unwrap_or_else(SystemTime::now),
4861 0 : })
4862 0 : })
4863 0 : .await?
4864 0 : }
4865 :
4866 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4867 0 : &self,
4868 0 : tenant_shard_id: TenantShardId,
4869 0 : timeline_id: TimelineId,
4870 0 : concurrency: Option<usize>,
4871 0 : recurse: bool,
4872 0 : ) -> Result<(), ApiError> {
4873 0 : let _tenant_lock = trace_shared_lock(
4874 0 : &self.tenant_op_locks,
4875 0 : tenant_shard_id.tenant_id,
4876 0 : TenantOperations::DownloadHeatmapLayers,
4877 0 : )
4878 0 : .await;
4879 :
4880 0 : let targets = {
4881 0 : let locked = self.inner.read().unwrap();
4882 0 : let mut targets = Vec::new();
4883 :
4884 : // If the request got an unsharded tenant id, then apply
4885 : // the operation to all shards. Otherwise, apply it to a specific shard.
4886 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4887 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4888 : } else {
4889 0 : tenant_shard_id.range()
4890 : };
4891 :
4892 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4893 0 : if let Some(node_id) = shard.intent.get_attached() {
4894 0 : let node = locked
4895 0 : .nodes
4896 0 : .get(node_id)
4897 0 : .expect("Pageservers may not be deleted while referenced");
4898 0 :
4899 0 : targets.push((*tenant_shard_id, node.clone()));
4900 0 : }
4901 : }
4902 0 : targets
4903 : };
4904 :
4905 0 : self.tenant_for_shards_api(
4906 0 : targets,
4907 0 : |tenant_shard_id, client| async move {
4908 0 : client
4909 0 : .timeline_download_heatmap_layers(
4910 0 : tenant_shard_id,
4911 0 : timeline_id,
4912 0 : concurrency,
4913 0 : recurse,
4914 0 : )
4915 0 : .await
4916 0 : },
4917 : 1,
4918 : 1,
4919 : SHORT_RECONCILE_TIMEOUT,
4920 0 : &self.cancel,
4921 : )
4922 0 : .await;
4923 :
4924 0 : Ok(())
4925 0 : }
4926 :
4927 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4928 : ///
4929 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`
4930 : /// and returned element at index `i` is the result for `req_fn(op(locations[i])`.
4931 0 : async fn tenant_for_shards<F, R>(
4932 0 : &self,
4933 0 : locations: Vec<(TenantShardId, Node)>,
4934 0 : mut req_fn: F,
4935 0 : ) -> Result<Vec<R>, ApiError>
4936 0 : where
4937 0 : F: FnMut(
4938 0 : TenantShardId,
4939 0 : Node,
4940 0 : )
4941 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4942 0 : {
4943 0 : let mut futs = FuturesUnordered::new();
4944 0 : let mut results = Vec::with_capacity(locations.len());
4945 :
4946 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4947 0 : let fut = req_fn(tenant_shard_id, node);
4948 0 : futs.push(async move { (idx, fut.await) });
4949 : }
4950 :
4951 0 : while let Some((idx, r)) = futs.next().await {
4952 0 : results.push((idx, r?));
4953 : }
4954 :
4955 0 : results.sort_by_key(|(idx, _)| *idx);
4956 0 : Ok(results.into_iter().map(|(_, r)| r).collect())
4957 0 : }
4958 :
4959 : /// Concurrently invoke a pageserver API call on many shards at once.
4960 : ///
4961 : /// The returned Vec has the same length as the `locations` Vec,
4962 : /// and returned element at index `i` is the result for `op(locations[i])`.
4963 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
4964 0 : &self,
4965 0 : locations: Vec<(TenantShardId, Node)>,
4966 0 : op: O,
4967 0 : warn_threshold: u32,
4968 0 : max_retries: u32,
4969 0 : timeout: Duration,
4970 0 : cancel: &CancellationToken,
4971 0 : ) -> Vec<(Node, mgmt_api::Result<T>)>
4972 0 : where
4973 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
4974 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
4975 0 : {
4976 0 : let mut futs = FuturesUnordered::new();
4977 0 : let mut results = Vec::with_capacity(locations.len());
4978 :
4979 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4980 0 : futs.push(async move {
4981 0 : let r = node
4982 0 : .with_client_retries(
4983 0 : |client| op(tenant_shard_id, client),
4984 0 : &self.http_client,
4985 0 : &self.config.pageserver_jwt_token,
4986 0 : warn_threshold,
4987 0 : max_retries,
4988 0 : timeout,
4989 0 : cancel,
4990 : )
4991 0 : .await;
4992 0 : (idx, node, r)
4993 0 : });
4994 : }
4995 :
4996 0 : while let Some((idx, node, r)) = futs.next().await {
4997 0 : results.push((idx, node, r.unwrap_or(Err(mgmt_api::Error::Cancelled))));
4998 0 : }
4999 :
5000 0 : results.sort_by_key(|(idx, _, _)| *idx);
5001 0 : results.into_iter().map(|(_, node, r)| (node, r)).collect()
5002 0 : }
5003 :
5004 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
5005 : /// when creating and deleting timelines:
5006 : /// - Makes sure shards are attached somewhere if they weren't already
5007 : /// - Looks up the shards and the nodes where they were most recently attached
5008 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
5009 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
5010 0 : pub(crate) async fn tenant_remote_mutation<R, O, F>(
5011 0 : &self,
5012 0 : tenant_id: TenantId,
5013 0 : op: O,
5014 0 : ) -> Result<R, ApiError>
5015 0 : where
5016 0 : O: FnOnce(TenantMutationLocations) -> F,
5017 0 : F: std::future::Future<Output = R>,
5018 0 : {
5019 0 : self.tenant_remote_mutation_inner(TenantIdOrShardId::TenantId(tenant_id), op)
5020 0 : .await
5021 0 : }
5022 :
5023 0 : pub(crate) async fn tenant_shard_remote_mutation<R, O, F>(
5024 0 : &self,
5025 0 : tenant_shard_id: TenantShardId,
5026 0 : op: O,
5027 0 : ) -> Result<R, ApiError>
5028 0 : where
5029 0 : O: FnOnce(TenantMutationLocations) -> F,
5030 0 : F: std::future::Future<Output = R>,
5031 0 : {
5032 0 : self.tenant_remote_mutation_inner(TenantIdOrShardId::TenantShardId(tenant_shard_id), op)
5033 0 : .await
5034 0 : }
5035 :
5036 0 : async fn tenant_remote_mutation_inner<R, O, F>(
5037 0 : &self,
5038 0 : tenant_id_or_shard_id: TenantIdOrShardId,
5039 0 : op: O,
5040 0 : ) -> Result<R, ApiError>
5041 0 : where
5042 0 : O: FnOnce(TenantMutationLocations) -> F,
5043 0 : F: std::future::Future<Output = R>,
5044 0 : {
5045 0 : let mutation_locations = {
5046 0 : let mut locations = TenantMutationLocations::default();
5047 :
5048 : // Load the currently attached pageservers for the latest generation of each shard. This can
5049 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
5050 : // will still be the latest when we're done: we will check generations again at the end of
5051 : // this function to handle that.
5052 0 : let generations = self
5053 0 : .persistence
5054 0 : .tenant_generations(tenant_id_or_shard_id.tenant_id())
5055 0 : .await?
5056 0 : .into_iter()
5057 0 : .filter(|i| tenant_id_or_shard_id.matches(&i.tenant_shard_id))
5058 0 : .collect::<Vec<_>>();
5059 :
5060 0 : if generations
5061 0 : .iter()
5062 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
5063 : {
5064 0 : let shard_generations = generations
5065 0 : .into_iter()
5066 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
5067 0 : .collect::<HashMap<_, _>>();
5068 :
5069 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
5070 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
5071 0 : let locked = self.inner.read().unwrap();
5072 0 : let tenant_shards = locked
5073 0 : .tenants
5074 0 : .range(TenantShardId::tenant_range(
5075 0 : tenant_id_or_shard_id.tenant_id(),
5076 : ))
5077 0 : .filter(|(shard_id, _)| tenant_id_or_shard_id.matches(shard_id))
5078 0 : .collect::<Vec<_>>();
5079 0 : for (shard_id, shard) in tenant_shards {
5080 0 : match shard.policy {
5081 : PlacementPolicy::Attached(_) => {
5082 : // This shard is meant to be attached: the caller is not wrong to try and
5083 : // use this function, but we can't service the request right now.
5084 0 : let Some(generation) = shard_generations.get(shard_id) else {
5085 : // This can only happen if there is a split brain controller modifying the database. This should
5086 : // never happen when testing, and if it happens in production we can only log the issue.
5087 0 : debug_assert!(false);
5088 0 : tracing::error!(
5089 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
5090 : );
5091 0 : continue;
5092 : };
5093 0 : let (generation, generation_pageserver) = generation;
5094 0 : if let Some(generation) = generation {
5095 0 : if generation_pageserver.is_none() {
5096 : // This is legitimate only in a very narrow window where the shard was only just configured into
5097 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
5098 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
5099 0 : tracing::warn!(
5100 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
5101 : );
5102 0 : }
5103 : } else {
5104 : // This should never happen: a shard with no generation is only permitted when it was created in some state
5105 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
5106 0 : debug_assert!(false);
5107 0 : tracing::error!(
5108 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
5109 : );
5110 0 : continue;
5111 : }
5112 : }
5113 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
5114 0 : return Err(ApiError::Conflict(format!(
5115 0 : "Shard {shard_id} tenant has policy {:?}",
5116 0 : shard.policy
5117 0 : )));
5118 : }
5119 : }
5120 : }
5121 :
5122 0 : return Err(ApiError::ResourceUnavailable(
5123 0 : "One or more shards in tenant is not yet attached".into(),
5124 0 : ));
5125 0 : }
5126 :
5127 0 : let locked = self.inner.read().unwrap();
5128 : for ShardGenerationState {
5129 0 : tenant_shard_id,
5130 0 : generation,
5131 0 : generation_pageserver,
5132 0 : } in generations
5133 : {
5134 0 : let node_id = generation_pageserver.expect("We checked for None above");
5135 0 : let node = locked
5136 0 : .nodes
5137 0 : .get(&node_id)
5138 0 : .ok_or(ApiError::Conflict(format!(
5139 0 : "Raced with removal of node {node_id}"
5140 0 : )))?;
5141 0 : let generation = generation.expect("Checked above");
5142 :
5143 0 : let tenant = locked.tenants.get(&tenant_shard_id);
5144 :
5145 : // TODO(vlad): Abstract the logic that finds stale attached locations
5146 : // from observed state into a [`Service`] method.
5147 0 : let other_locations = match tenant {
5148 0 : Some(tenant) => {
5149 0 : let mut other = tenant.attached_locations();
5150 0 : let latest_location_index =
5151 0 : other.iter().position(|&l| l == (node.get_id(), generation));
5152 0 : if let Some(idx) = latest_location_index {
5153 0 : other.remove(idx);
5154 0 : }
5155 :
5156 0 : other
5157 : }
5158 0 : None => Vec::default(),
5159 : };
5160 :
5161 0 : let location = ShardMutationLocations {
5162 0 : latest: MutationLocation {
5163 0 : node: node.clone(),
5164 0 : generation,
5165 0 : },
5166 0 : other: other_locations
5167 0 : .into_iter()
5168 0 : .filter_map(|(node_id, generation)| {
5169 0 : let node = locked.nodes.get(&node_id)?;
5170 :
5171 0 : Some(MutationLocation {
5172 0 : node: node.clone(),
5173 0 : generation,
5174 0 : })
5175 0 : })
5176 0 : .collect(),
5177 : };
5178 0 : locations.0.insert(tenant_shard_id, location);
5179 : }
5180 :
5181 0 : locations
5182 : };
5183 :
5184 0 : let result = op(mutation_locations.clone()).await;
5185 :
5186 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
5187 : // our remote operation executed on the latest generation and is therefore persistent.
5188 : {
5189 0 : let latest_generations = self
5190 0 : .persistence
5191 0 : .tenant_generations(tenant_id_or_shard_id.tenant_id())
5192 0 : .await?
5193 0 : .into_iter()
5194 0 : .filter(|i| tenant_id_or_shard_id.matches(&i.tenant_shard_id))
5195 0 : .collect::<Vec<_>>();
5196 :
5197 0 : if latest_generations
5198 0 : .into_iter()
5199 0 : .map(
5200 : |ShardGenerationState {
5201 : tenant_shard_id,
5202 : generation,
5203 : generation_pageserver: _,
5204 0 : }| (tenant_shard_id, generation),
5205 : )
5206 0 : .collect::<Vec<_>>()
5207 0 : != mutation_locations
5208 0 : .0
5209 0 : .into_iter()
5210 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
5211 0 : .collect::<Vec<_>>()
5212 : {
5213 : // We raced with something that incremented the generation, and therefore cannot be
5214 : // confident that our actions are persistent (they might have hit an old generation).
5215 : //
5216 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
5217 0 : return Err(ApiError::ResourceUnavailable(
5218 0 : "Tenant attachment changed, please retry".into(),
5219 0 : ));
5220 0 : }
5221 : }
5222 :
5223 0 : Ok(result)
5224 0 : }
5225 :
5226 0 : pub(crate) async fn tenant_timeline_delete(
5227 0 : self: &Arc<Self>,
5228 0 : tenant_id: TenantId,
5229 0 : timeline_id: TimelineId,
5230 0 : ) -> Result<StatusCode, ApiError> {
5231 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
5232 0 : let _tenant_lock = trace_shared_lock(
5233 0 : &self.tenant_op_locks,
5234 0 : tenant_id,
5235 0 : TenantOperations::TimelineDelete,
5236 0 : )
5237 0 : .await;
5238 :
5239 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
5240 0 : if targets.0.is_empty() {
5241 0 : return Err(ApiError::NotFound(
5242 0 : anyhow::anyhow!("Tenant not found").into(),
5243 0 : ));
5244 0 : }
5245 :
5246 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
5247 0 : assert!(shard_zero_tid.is_shard_zero());
5248 :
5249 0 : async fn delete_one(
5250 0 : tenant_shard_id: TenantShardId,
5251 0 : timeline_id: TimelineId,
5252 0 : node: Node,
5253 0 : http_client: reqwest::Client,
5254 0 : jwt: Option<String>,
5255 0 : ) -> Result<StatusCode, ApiError> {
5256 0 : tracing::info!(
5257 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
5258 : );
5259 :
5260 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
5261 0 : let res = client
5262 0 : .timeline_delete(tenant_shard_id, timeline_id)
5263 0 : .await;
5264 :
5265 0 : match res {
5266 0 : Ok(ok) => Ok(ok),
5267 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
5268 0 : Err(mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg)) if msg.contains("Requested tenant is missing") => {
5269 0 : Err(ApiError::ResourceUnavailable("Tenant migration in progress".into()))
5270 : },
5271 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
5272 0 : Err(e) => {
5273 0 : Err(
5274 0 : ApiError::InternalServerError(anyhow::anyhow!(
5275 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
5276 0 : ))
5277 0 : )
5278 : }
5279 : }
5280 0 : }
5281 :
5282 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
5283 0 : let statuses = self
5284 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
5285 0 : Box::pin(delete_one(
5286 0 : tenant_shard_id,
5287 0 : timeline_id,
5288 0 : node,
5289 0 : self.http_client.clone(),
5290 0 : self.config.pageserver_jwt_token.clone(),
5291 0 : ))
5292 0 : })
5293 0 : .await?;
5294 :
5295 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
5296 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
5297 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
5298 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
5299 0 : return Ok(StatusCode::CONFLICT);
5300 0 : }
5301 :
5302 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
5303 0 : return Ok(StatusCode::ACCEPTED);
5304 0 : }
5305 :
5306 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
5307 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
5308 0 : let shard_zero_status = delete_one(
5309 0 : shard_zero_tid,
5310 0 : timeline_id,
5311 0 : shard_zero_locations.latest.node,
5312 0 : self.http_client.clone(),
5313 0 : self.config.pageserver_jwt_token.clone(),
5314 0 : )
5315 0 : .await?;
5316 0 : Ok(shard_zero_status)
5317 0 : }).await?;
5318 :
5319 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
5320 0 : .await?;
5321 :
5322 0 : status_code
5323 0 : }
5324 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
5325 : ///
5326 : /// Returns the node, tenant shard id, and whether it is consistent with the observed state.
5327 0 : pub(crate) async fn tenant_shard0_node(
5328 0 : &self,
5329 0 : tenant_id: TenantId,
5330 0 : ) -> Result<(Node, TenantShardId), ApiError> {
5331 0 : let tenant_shard_id = {
5332 0 : let locked = self.inner.read().unwrap();
5333 0 : let Some((tenant_shard_id, _shard)) = locked
5334 0 : .tenants
5335 0 : .range(TenantShardId::tenant_range(tenant_id))
5336 0 : .next()
5337 : else {
5338 0 : return Err(ApiError::NotFound(
5339 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
5340 0 : ));
5341 : };
5342 :
5343 0 : *tenant_shard_id
5344 : };
5345 :
5346 0 : self.tenant_shard_node(tenant_shard_id)
5347 0 : .await
5348 0 : .map(|node| (node, tenant_shard_id))
5349 0 : }
5350 :
5351 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
5352 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
5353 : ///
5354 : /// Returns the intent node and whether it is consistent with the observed state.
5355 0 : pub(crate) async fn tenant_shard_node(
5356 0 : &self,
5357 0 : tenant_shard_id: TenantShardId,
5358 0 : ) -> Result<Node, ApiError> {
5359 : // Look up in-memory state and maybe use the node from there.
5360 : {
5361 0 : let locked = self.inner.read().unwrap();
5362 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5363 0 : return Err(ApiError::NotFound(
5364 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
5365 0 : ));
5366 : };
5367 :
5368 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
5369 0 : tracing::warn!(
5370 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
5371 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
5372 : shard.policy
5373 : );
5374 0 : return Err(ApiError::Conflict(
5375 0 : "Cannot call timeline API on non-attached tenant".to_string(),
5376 0 : ));
5377 : };
5378 :
5379 0 : if shard.reconciler.is_none() {
5380 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
5381 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
5382 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
5383 : // This should never happen
5384 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5385 0 : "Shard refers to nonexistent node"
5386 0 : )));
5387 : };
5388 0 : return Ok(node.clone());
5389 0 : }
5390 : };
5391 :
5392 : // Look up the latest attached pageserver location from the database
5393 : // generation state: this will reflect the progress of any ongoing migration.
5394 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
5395 : // the case where they call through to the pageserver and get a 404.
5396 0 : let db_result = self
5397 0 : .persistence
5398 0 : .tenant_generations(tenant_shard_id.tenant_id)
5399 0 : .await?;
5400 : let Some(ShardGenerationState {
5401 : tenant_shard_id: _,
5402 : generation: _,
5403 0 : generation_pageserver: Some(node_id),
5404 0 : }) = db_result
5405 0 : .into_iter()
5406 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
5407 : else {
5408 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
5409 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
5410 : // or a conflict response (case where tenant was detached in background)
5411 0 : return Err(ApiError::ResourceUnavailable(
5412 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
5413 0 : ));
5414 : };
5415 0 : let locked = self.inner.read().unwrap();
5416 0 : let Some(node) = locked.nodes.get(&node_id) else {
5417 : // This should never happen
5418 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5419 0 : "Shard refers to nonexistent node"
5420 0 : )));
5421 : };
5422 : // As a reconciliation is in flight, we do not have the observed state yet, and therefore we assume it is always inconsistent.
5423 0 : Ok(node.clone())
5424 0 : }
5425 :
5426 0 : pub(crate) fn tenant_locate(
5427 0 : &self,
5428 0 : tenant_id: TenantId,
5429 0 : ) -> Result<TenantLocateResponse, ApiError> {
5430 0 : let locked = self.inner.read().unwrap();
5431 0 : tracing::info!("Locating shards for tenant {tenant_id}");
5432 :
5433 0 : let mut result = Vec::new();
5434 0 : let mut shard_params: Option<ShardParameters> = None;
5435 :
5436 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5437 : {
5438 0 : let node_id =
5439 0 : shard
5440 0 : .intent
5441 0 : .get_attached()
5442 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
5443 0 : "Cannot locate a tenant that is not attached"
5444 0 : )))?;
5445 :
5446 0 : let node = locked
5447 0 : .nodes
5448 0 : .get(&node_id)
5449 0 : .expect("Pageservers may not be deleted while referenced");
5450 :
5451 0 : result.push(node.shard_location(*tenant_shard_id));
5452 :
5453 0 : match &shard_params {
5454 0 : None => {
5455 0 : shard_params = Some(ShardParameters {
5456 0 : stripe_size: shard.shard.stripe_size,
5457 0 : count: shard.shard.count,
5458 0 : });
5459 0 : }
5460 0 : Some(params) => {
5461 0 : if params.stripe_size != shard.shard.stripe_size {
5462 : // This should never happen. We enforce at runtime because it's simpler than
5463 : // adding an extra per-tenant data structure to store the things that should be the same
5464 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5465 0 : "Inconsistent shard stripe size parameters!"
5466 0 : )));
5467 0 : }
5468 : }
5469 : }
5470 : }
5471 :
5472 0 : if result.is_empty() {
5473 0 : return Err(ApiError::NotFound(
5474 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
5475 0 : ));
5476 0 : }
5477 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
5478 0 : tracing::info!(
5479 0 : "Located tenant {} with params {:?} on shards {}",
5480 : tenant_id,
5481 : shard_params,
5482 0 : result
5483 0 : .iter()
5484 0 : .map(|s| format!("{s:?}"))
5485 0 : .collect::<Vec<_>>()
5486 0 : .join(",")
5487 : );
5488 :
5489 0 : Ok(TenantLocateResponse {
5490 0 : shards: result,
5491 0 : shard_params,
5492 0 : })
5493 0 : }
5494 :
5495 : /// Returns None if the input iterator of shards does not include a shard with number=0
5496 0 : fn tenant_describe_impl<'a>(
5497 0 : &self,
5498 0 : shards: impl Iterator<Item = &'a TenantShard>,
5499 0 : ) -> Option<TenantDescribeResponse> {
5500 0 : let mut shard_zero = None;
5501 0 : let mut describe_shards = Vec::new();
5502 :
5503 0 : for shard in shards {
5504 0 : if shard.tenant_shard_id.is_shard_zero() {
5505 0 : shard_zero = Some(shard);
5506 0 : }
5507 :
5508 0 : describe_shards.push(TenantDescribeResponseShard {
5509 0 : tenant_shard_id: shard.tenant_shard_id,
5510 0 : node_attached: *shard.intent.get_attached(),
5511 0 : node_secondary: shard.intent.get_secondary().to_vec(),
5512 0 : last_error: shard
5513 0 : .last_error
5514 0 : .lock()
5515 0 : .unwrap()
5516 0 : .as_ref()
5517 0 : .map(|e| format!("{e}"))
5518 0 : .unwrap_or("".to_string())
5519 0 : .clone(),
5520 0 : is_reconciling: shard.reconciler.is_some(),
5521 0 : is_pending_compute_notification: shard.pending_compute_notification,
5522 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
5523 0 : is_importing: shard.importing == TimelineImportState::Importing,
5524 0 : scheduling_policy: shard.get_scheduling_policy(),
5525 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
5526 : })
5527 : }
5528 :
5529 0 : let shard_zero = shard_zero?;
5530 :
5531 0 : Some(TenantDescribeResponse {
5532 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
5533 0 : shards: describe_shards,
5534 0 : stripe_size: shard_zero.shard.stripe_size,
5535 0 : policy: shard_zero.policy.clone(),
5536 0 : config: shard_zero.config.clone(),
5537 0 : })
5538 0 : }
5539 :
5540 0 : pub(crate) fn tenant_describe(
5541 0 : &self,
5542 0 : tenant_id: TenantId,
5543 0 : ) -> Result<TenantDescribeResponse, ApiError> {
5544 0 : let locked = self.inner.read().unwrap();
5545 :
5546 0 : self.tenant_describe_impl(
5547 0 : locked
5548 0 : .tenants
5549 0 : .range(TenantShardId::tenant_range(tenant_id))
5550 0 : .map(|(_k, v)| v),
5551 : )
5552 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
5553 0 : }
5554 :
5555 : /* BEGIN_HADRON */
5556 0 : pub(crate) async fn tenant_timeline_describe(
5557 0 : &self,
5558 0 : tenant_id: TenantId,
5559 0 : timeline_id: TimelineId,
5560 0 : ) -> Result<TenantTimelineDescribeResponse, ApiError> {
5561 0 : self.tenant_remote_mutation(tenant_id, |locations| async move {
5562 0 : if locations.0.is_empty() {
5563 0 : return Err(ApiError::NotFound(
5564 0 : anyhow::anyhow!("Tenant not found").into(),
5565 0 : ));
5566 0 : };
5567 :
5568 0 : let locations: Vec<(TenantShardId, Node)> = locations
5569 0 : .0
5570 0 : .iter()
5571 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
5572 0 : .collect();
5573 0 : let mut futs = FuturesUnordered::new();
5574 :
5575 0 : for (shard_id, node) in locations {
5576 0 : futs.push({
5577 0 : async move {
5578 0 : let result = node
5579 0 : .with_client_retries(
5580 0 : |client| async move {
5581 0 : client
5582 0 : .tenant_timeline_describe(&shard_id, &timeline_id)
5583 0 : .await
5584 0 : },
5585 0 : &self.http_client,
5586 0 : &self.config.pageserver_jwt_token,
5587 : 3,
5588 : 3,
5589 0 : Duration::from_secs(30),
5590 0 : &self.cancel,
5591 : )
5592 0 : .await;
5593 0 : (result, shard_id, node.get_id())
5594 0 : }
5595 : });
5596 : }
5597 :
5598 0 : let mut results: Vec<TimelineInfo> = Vec::new();
5599 0 : while let Some((result, tenant_shard_id, node_id)) = futs.next().await {
5600 0 : match result {
5601 0 : Some(Ok(timeline_info)) => results.push(timeline_info),
5602 0 : Some(Err(e)) => {
5603 0 : tracing::warn!(
5604 0 : "Failed to describe tenant {} timeline {} for pageserver {}: {e}",
5605 : tenant_shard_id,
5606 : timeline_id,
5607 : node_id,
5608 : );
5609 0 : return Err(ApiError::ResourceUnavailable(format!("{e}").into()));
5610 : }
5611 0 : None => return Err(ApiError::Cancelled),
5612 : }
5613 : }
5614 0 : let mut image_consistent_lsn: Option<Lsn> = Some(Lsn::MAX);
5615 0 : for timeline_info in &results {
5616 0 : if let Some(tline_image_consistent_lsn) = timeline_info.image_consistent_lsn {
5617 0 : image_consistent_lsn = Some(std::cmp::min(
5618 0 : image_consistent_lsn.unwrap(),
5619 0 : tline_image_consistent_lsn,
5620 0 : ));
5621 0 : } else {
5622 0 : tracing::warn!(
5623 0 : "Timeline {} on shard {} does not have image consistent lsn",
5624 : timeline_info.timeline_id,
5625 : timeline_info.tenant_id
5626 : );
5627 0 : image_consistent_lsn = None;
5628 0 : break;
5629 : }
5630 : }
5631 :
5632 0 : Ok(TenantTimelineDescribeResponse {
5633 0 : shards: results,
5634 0 : image_consistent_lsn,
5635 0 : })
5636 0 : })
5637 0 : .await?
5638 0 : }
5639 : /* END_HADRON */
5640 :
5641 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
5642 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
5643 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
5644 : /// in our external API.
5645 0 : pub(crate) fn tenant_list(
5646 0 : &self,
5647 0 : limit: Option<usize>,
5648 0 : start_after: Option<TenantId>,
5649 0 : ) -> Vec<TenantDescribeResponse> {
5650 0 : let locked = self.inner.read().unwrap();
5651 :
5652 : // Apply start_from parameter
5653 0 : let shard_range = match start_after {
5654 0 : None => locked.tenants.range(..),
5655 0 : Some(tenant_id) => locked.tenants.range(
5656 0 : TenantShardId {
5657 0 : tenant_id,
5658 0 : shard_number: ShardNumber(u8::MAX),
5659 0 : shard_count: ShardCount(u8::MAX),
5660 0 : }..,
5661 : ),
5662 : };
5663 :
5664 0 : let mut result = Vec::new();
5665 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
5666 0 : result.push(
5667 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
5668 0 : .expect("Groups are always non-empty"),
5669 : );
5670 :
5671 : // Enforce `limit` parameter
5672 0 : if let Some(limit) = limit {
5673 0 : if result.len() >= limit {
5674 0 : break;
5675 0 : }
5676 0 : }
5677 : }
5678 :
5679 0 : result
5680 0 : }
5681 :
5682 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
5683 : async fn abort_tenant_shard_split(
5684 : &self,
5685 : op: &TenantShardSplitAbort,
5686 : ) -> Result<(), TenantShardSplitAbortError> {
5687 : // Cleaning up a split:
5688 : // - Parent shards are not destroyed during a split, just detached.
5689 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
5690 : // just the children attached, or both.
5691 : //
5692 : // Therefore our work to do is to:
5693 : // 1. Clean up storage controller's internal state to just refer to parents, no children
5694 : // 2. Call out to pageservers to ensure that children are detached
5695 : // 3. Call out to pageservers to ensure that parents are attached.
5696 : //
5697 : // Crash safety:
5698 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
5699 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
5700 : // and detach them.
5701 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
5702 : // from our database, then we will re-enter this cleanup routine on startup.
5703 :
5704 : let TenantShardSplitAbort {
5705 : tenant_id,
5706 : new_shard_count,
5707 : new_stripe_size,
5708 : ..
5709 : } = op;
5710 :
5711 : // First abort persistent state, if any exists.
5712 : match self
5713 : .persistence
5714 : .abort_shard_split(*tenant_id, *new_shard_count)
5715 : .await?
5716 : {
5717 : AbortShardSplitStatus::Aborted => {
5718 : // Proceed to roll back any child shards created on pageservers
5719 : }
5720 : AbortShardSplitStatus::Complete => {
5721 : // The split completed (we might hit that path if e.g. our database transaction
5722 : // to write the completion landed in the database, but we dropped connection
5723 : // before seeing the result).
5724 : //
5725 : // We must update in-memory state to reflect the successful split.
5726 : self.tenant_shard_split_commit_inmem(
5727 : *tenant_id,
5728 : *new_shard_count,
5729 : *new_stripe_size,
5730 : );
5731 : return Ok(());
5732 : }
5733 : }
5734 :
5735 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
5736 : let detach_locations: Vec<(Node, TenantShardId)> = {
5737 : let mut detach_locations = Vec::new();
5738 : let mut locked = self.inner.write().unwrap();
5739 : let (nodes, tenants, scheduler) = locked.parts_mut();
5740 :
5741 : for (tenant_shard_id, shard) in
5742 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
5743 : {
5744 : if shard.shard.count == op.new_shard_count {
5745 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
5746 : // is infallible, so if we got an error we shouldn't have got that far.
5747 : tracing::warn!(
5748 : "During split abort, child shard {tenant_shard_id} found in-memory"
5749 : );
5750 : continue;
5751 : }
5752 :
5753 : // Add the children of this shard to this list of things to detach
5754 : if let Some(node_id) = shard.intent.get_attached() {
5755 : for child_id in tenant_shard_id.split(*new_shard_count) {
5756 : detach_locations.push((
5757 : nodes
5758 : .get(node_id)
5759 : .expect("Intent references nonexistent node")
5760 : .clone(),
5761 : child_id,
5762 : ));
5763 : }
5764 : } else {
5765 : tracing::warn!(
5766 : "During split abort, shard {tenant_shard_id} has no attached location"
5767 : );
5768 : }
5769 :
5770 : tracing::info!("Restoring parent shard {tenant_shard_id}");
5771 :
5772 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
5773 : // if the original attachment location is offline.
5774 : if let Some(node_id) = shard.intent.get_attached() {
5775 : if !nodes.get(node_id).unwrap().is_available() {
5776 : tracing::info!(
5777 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
5778 : );
5779 : shard.intent.demote_attached(scheduler, *node_id);
5780 : }
5781 : }
5782 : for node_id in shard.intent.get_secondary().clone() {
5783 : if !nodes.get(&node_id).unwrap().is_available() {
5784 : tracing::info!(
5785 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
5786 : );
5787 : shard.intent.remove_secondary(scheduler, node_id);
5788 : }
5789 : }
5790 :
5791 : shard.splitting = SplitState::Idle;
5792 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
5793 : // If this shard can't be scheduled now (perhaps due to offline nodes or
5794 : // capacity issues), that must not prevent us rolling back a split. In this
5795 : // case it should be eventually scheduled in the background.
5796 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
5797 : }
5798 :
5799 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
5800 : }
5801 :
5802 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
5803 : tenants
5804 0 : .retain(|id, s| !(id.tenant_id == *tenant_id && s.shard.count == *new_shard_count));
5805 :
5806 : detach_locations
5807 : };
5808 :
5809 : for (node, child_id) in detach_locations {
5810 : if !node.is_available() {
5811 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
5812 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
5813 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
5814 : // them from the node.
5815 : tracing::warn!(
5816 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
5817 : );
5818 : continue;
5819 : }
5820 :
5821 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
5822 : // a 503 and retry, up to our limit.
5823 : tracing::info!("Detaching {child_id} on {node}...");
5824 : match node
5825 : .with_client_retries(
5826 0 : |client| async move {
5827 0 : let config = LocationConfig {
5828 0 : mode: LocationConfigMode::Detached,
5829 0 : generation: None,
5830 0 : secondary_conf: None,
5831 0 : shard_number: child_id.shard_number.0,
5832 0 : shard_count: child_id.shard_count.literal(),
5833 0 : // Stripe size and tenant config don't matter when detaching
5834 0 : shard_stripe_size: 0,
5835 0 : tenant_conf: TenantConfig::default(),
5836 0 : };
5837 :
5838 0 : client.location_config(child_id, config, None, false).await
5839 0 : },
5840 : &self.http_client,
5841 : &self.config.pageserver_jwt_token,
5842 : 1,
5843 : 10,
5844 : Duration::from_secs(5),
5845 : &self.reconcilers_cancel,
5846 : )
5847 : .await
5848 : {
5849 : Some(Ok(_)) => {}
5850 : Some(Err(e)) => {
5851 : // We failed to communicate with the remote node. This is problematic: we may be
5852 : // leaving it with a rogue child shard.
5853 : tracing::warn!(
5854 : "Failed to detach child {child_id} from node {node} during abort"
5855 : );
5856 : return Err(e.into());
5857 : }
5858 : None => {
5859 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
5860 : // clean up on restart. The node going offline requires a retry.
5861 : return Err(TenantShardSplitAbortError::Unavailable);
5862 : }
5863 : };
5864 : }
5865 :
5866 : tracing::info!("Successfully aborted split");
5867 : Ok(())
5868 : }
5869 :
5870 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
5871 : /// of the tenant map to reflect the child shards that exist after the split.
5872 0 : fn tenant_shard_split_commit_inmem(
5873 0 : &self,
5874 0 : tenant_id: TenantId,
5875 0 : new_shard_count: ShardCount,
5876 0 : new_stripe_size: Option<ShardStripeSize>,
5877 0 : ) -> (
5878 0 : TenantShardSplitResponse,
5879 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
5880 0 : Vec<ReconcilerWaiter>,
5881 0 : ) {
5882 0 : let mut response = TenantShardSplitResponse {
5883 0 : new_shards: Vec::new(),
5884 0 : };
5885 0 : let mut child_locations = Vec::new();
5886 0 : let mut waiters = Vec::new();
5887 :
5888 : {
5889 0 : let mut locked = self.inner.write().unwrap();
5890 :
5891 0 : let parent_ids = locked
5892 0 : .tenants
5893 0 : .range(TenantShardId::tenant_range(tenant_id))
5894 0 : .map(|(shard_id, _)| *shard_id)
5895 0 : .collect::<Vec<_>>();
5896 :
5897 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5898 0 : for parent_id in parent_ids {
5899 0 : let child_ids = parent_id.split(new_shard_count);
5900 :
5901 : let (
5902 0 : pageserver,
5903 0 : generation,
5904 0 : policy,
5905 0 : parent_ident,
5906 0 : config,
5907 0 : preferred_az,
5908 0 : secondary_count,
5909 : ) = {
5910 0 : let mut old_state = tenants
5911 0 : .remove(&parent_id)
5912 0 : .expect("It was present, we just split it");
5913 :
5914 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
5915 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
5916 : // nothing else can clear this.
5917 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
5918 :
5919 0 : let old_attached = old_state.intent.get_attached().unwrap();
5920 0 : old_state.intent.clear(scheduler);
5921 0 : let generation = old_state.generation.expect("Shard must have been attached");
5922 0 : (
5923 0 : old_attached,
5924 0 : generation,
5925 0 : old_state.policy.clone(),
5926 0 : old_state.shard,
5927 0 : old_state.config.clone(),
5928 0 : old_state.preferred_az().cloned(),
5929 0 : old_state.intent.get_secondary().len(),
5930 0 : )
5931 : };
5932 :
5933 0 : let mut schedule_context = ScheduleContext::default();
5934 0 : for child in child_ids {
5935 0 : let mut child_shard = parent_ident;
5936 0 : child_shard.number = child.shard_number;
5937 0 : child_shard.count = child.shard_count;
5938 0 : if let Some(stripe_size) = new_stripe_size {
5939 0 : child_shard.stripe_size = stripe_size;
5940 0 : }
5941 :
5942 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5943 0 : child_observed.insert(
5944 0 : pageserver,
5945 0 : ObservedStateLocation {
5946 0 : conf: Some(attached_location_conf(
5947 0 : generation,
5948 0 : &child_shard,
5949 0 : &config,
5950 0 : &policy,
5951 0 : secondary_count,
5952 0 : )),
5953 0 : },
5954 : );
5955 :
5956 0 : let mut child_state =
5957 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
5958 0 : child_state.intent =
5959 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
5960 0 : child_state.observed = ObservedState {
5961 0 : locations: child_observed,
5962 0 : };
5963 0 : child_state.generation = Some(generation);
5964 0 : child_state.config = config.clone();
5965 :
5966 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
5967 : // as at this point in the split process we have succeeded and this part is infallible:
5968 : // we will never need to do any special recovery from this state.
5969 :
5970 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
5971 :
5972 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
5973 : // This is not fatal, because we've implicitly already got an attached
5974 : // location for the child shard. Failure here just means we couldn't
5975 : // find a secondary (e.g. because cluster is overloaded).
5976 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
5977 0 : }
5978 : // In the background, attach secondary locations for the new shards
5979 0 : if let Some(waiter) = self.maybe_reconcile_shard(
5980 0 : &mut child_state,
5981 0 : nodes,
5982 0 : ReconcilerPriority::High,
5983 0 : ) {
5984 0 : waiters.push(waiter);
5985 0 : }
5986 :
5987 0 : tenants.insert(child, child_state);
5988 0 : response.new_shards.push(child);
5989 : }
5990 : }
5991 0 : (response, child_locations, waiters)
5992 : }
5993 0 : }
5994 :
5995 0 : async fn tenant_shard_split_start_secondaries(
5996 0 : &self,
5997 0 : tenant_id: TenantId,
5998 0 : waiters: Vec<ReconcilerWaiter>,
5999 0 : ) {
6000 : // Wait for initial reconcile of child shards, this creates the secondary locations
6001 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
6002 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
6003 : // their secondaries couldn't be attached.
6004 0 : tracing::warn!("Failed to reconcile after split: {e}");
6005 0 : return;
6006 0 : }
6007 :
6008 : // Take the state lock to discover the attached & secondary intents for all shards
6009 0 : let (attached, secondary) = {
6010 0 : let locked = self.inner.read().unwrap();
6011 0 : let mut attached = Vec::new();
6012 0 : let mut secondary = Vec::new();
6013 :
6014 0 : for (tenant_shard_id, shard) in
6015 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
6016 : {
6017 0 : let Some(node_id) = shard.intent.get_attached() else {
6018 : // Unexpected. Race with a PlacementPolicy change?
6019 0 : tracing::warn!(
6020 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
6021 : );
6022 0 : continue;
6023 : };
6024 :
6025 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
6026 : // No secondary location. Nothing for us to do.
6027 0 : continue;
6028 : };
6029 :
6030 0 : let attached_node = locked
6031 0 : .nodes
6032 0 : .get(node_id)
6033 0 : .expect("Pageservers may not be deleted while referenced");
6034 :
6035 0 : let secondary_node = locked
6036 0 : .nodes
6037 0 : .get(secondary_node_id)
6038 0 : .expect("Pageservers may not be deleted while referenced");
6039 :
6040 0 : attached.push((*tenant_shard_id, attached_node.clone()));
6041 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
6042 : }
6043 0 : (attached, secondary)
6044 : };
6045 :
6046 0 : if secondary.is_empty() {
6047 : // No secondary locations; nothing for us to do
6048 0 : return;
6049 0 : }
6050 :
6051 0 : for (_, result) in self
6052 0 : .tenant_for_shards_api(
6053 0 : attached,
6054 0 : |tenant_shard_id, client| async move {
6055 0 : client.tenant_heatmap_upload(tenant_shard_id).await
6056 0 : },
6057 : 1,
6058 : 1,
6059 : SHORT_RECONCILE_TIMEOUT,
6060 0 : &self.cancel,
6061 : )
6062 0 : .await
6063 : {
6064 0 : if let Err(e) = result {
6065 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
6066 0 : return;
6067 0 : }
6068 : }
6069 :
6070 0 : for (_, result) in self
6071 0 : .tenant_for_shards_api(
6072 0 : secondary,
6073 0 : |tenant_shard_id, client| async move {
6074 0 : client
6075 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
6076 0 : .await
6077 0 : },
6078 : 1,
6079 : 1,
6080 : SHORT_RECONCILE_TIMEOUT,
6081 0 : &self.cancel,
6082 : )
6083 0 : .await
6084 : {
6085 0 : if let Err(e) = result {
6086 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
6087 0 : return;
6088 0 : }
6089 : }
6090 0 : }
6091 :
6092 0 : pub(crate) async fn tenant_shard_split(
6093 0 : &self,
6094 0 : tenant_id: TenantId,
6095 0 : split_req: TenantShardSplitRequest,
6096 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
6097 : // TODO: return 503 if we get stuck waiting for this lock
6098 : // (issue https://github.com/neondatabase/neon/issues/7108)
6099 0 : let _tenant_lock = trace_exclusive_lock(
6100 0 : &self.tenant_op_locks,
6101 0 : tenant_id,
6102 0 : TenantOperations::ShardSplit,
6103 0 : )
6104 0 : .await;
6105 :
6106 0 : let _gate = self
6107 0 : .reconcilers_gate
6108 0 : .enter()
6109 0 : .map_err(|_| ApiError::ShuttingDown)?;
6110 :
6111 : // Timeline imports on the pageserver side can't handle shard-splits.
6112 : // If the tenant is importing a timeline, dont't shard split it.
6113 0 : match self
6114 0 : .persistence
6115 0 : .is_tenant_importing_timeline(tenant_id)
6116 0 : .await
6117 : {
6118 0 : Ok(importing) => {
6119 0 : if importing {
6120 0 : return Err(ApiError::Conflict(
6121 0 : "Cannot shard split during timeline import".to_string(),
6122 0 : ));
6123 0 : }
6124 : }
6125 0 : Err(err) => {
6126 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6127 0 : "Failed to check for running imports: {err}"
6128 0 : )));
6129 : }
6130 : }
6131 :
6132 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
6133 0 : let new_stripe_size = split_req.new_stripe_size;
6134 :
6135 : // Validate the request and construct parameters. This phase is fallible, but does not require
6136 : // rollback on errors, as it does no I/O and mutates no state.
6137 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
6138 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
6139 0 : ShardSplitAction::Split(params) => params,
6140 : };
6141 :
6142 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
6143 : // we must roll back.
6144 0 : let r = self
6145 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
6146 0 : .await;
6147 :
6148 0 : let (response, waiters) = match r {
6149 0 : Ok(r) => r,
6150 0 : Err(e) => {
6151 : // Split might be part-done, we must do work to abort it.
6152 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
6153 0 : self.abort_tx
6154 0 : .send(TenantShardSplitAbort {
6155 0 : tenant_id,
6156 0 : new_shard_count,
6157 0 : new_stripe_size,
6158 0 : _tenant_lock,
6159 0 : _gate,
6160 0 : })
6161 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
6162 0 : .ok();
6163 0 : return Err(e);
6164 : }
6165 : };
6166 :
6167 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
6168 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
6169 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
6170 : // in [`Self::optimize_all`]
6171 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
6172 0 : .await;
6173 0 : Ok(response)
6174 0 : }
6175 :
6176 0 : fn prepare_tenant_shard_split(
6177 0 : &self,
6178 0 : tenant_id: TenantId,
6179 0 : split_req: TenantShardSplitRequest,
6180 0 : ) -> Result<ShardSplitAction, ApiError> {
6181 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
6182 0 : anyhow::anyhow!("failpoint")
6183 0 : )));
6184 :
6185 0 : let mut policy = None;
6186 0 : let mut config = None;
6187 0 : let mut shard_ident = None;
6188 0 : let mut preferred_az_id = None;
6189 : // Validate input, and calculate which shards we will create
6190 0 : let (old_shard_count, targets) =
6191 : {
6192 0 : let locked = self.inner.read().unwrap();
6193 :
6194 0 : let pageservers = locked.nodes.clone();
6195 :
6196 0 : let mut targets = Vec::new();
6197 :
6198 : // In case this is a retry, count how many already-split shards we found
6199 0 : let mut children_found = Vec::new();
6200 0 : let mut old_shard_count = None;
6201 :
6202 0 : for (tenant_shard_id, shard) in
6203 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
6204 : {
6205 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
6206 : Ordering::Equal => {
6207 : // Already split this
6208 0 : children_found.push(*tenant_shard_id);
6209 0 : continue;
6210 : }
6211 : Ordering::Greater => {
6212 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6213 0 : "Requested count {} but already have shards at count {}",
6214 0 : split_req.new_shard_count,
6215 0 : shard.shard.count.count()
6216 0 : )));
6217 : }
6218 0 : Ordering::Less => {
6219 0 : // Fall through: this shard has lower count than requested,
6220 0 : // is a candidate for splitting.
6221 0 : }
6222 : }
6223 :
6224 0 : match old_shard_count {
6225 0 : None => old_shard_count = Some(shard.shard.count),
6226 0 : Some(old_shard_count) => {
6227 0 : if old_shard_count != shard.shard.count {
6228 : // We may hit this case if a caller asked for two splits to
6229 : // different sizes, before the first one is complete.
6230 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
6231 : // of shard_count=1 and shard_count=2 shards in the map.
6232 0 : return Err(ApiError::Conflict(
6233 0 : "Cannot split, currently mid-split".to_string(),
6234 0 : ));
6235 0 : }
6236 : }
6237 : }
6238 0 : if policy.is_none() {
6239 0 : policy = Some(shard.policy.clone());
6240 0 : }
6241 0 : if shard_ident.is_none() {
6242 0 : shard_ident = Some(shard.shard);
6243 0 : }
6244 0 : if config.is_none() {
6245 0 : config = Some(shard.config.clone());
6246 0 : }
6247 0 : if preferred_az_id.is_none() {
6248 0 : preferred_az_id = shard.preferred_az().cloned();
6249 0 : }
6250 :
6251 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
6252 0 : tracing::info!(
6253 0 : "Tenant shard {} already has shard count {}",
6254 : tenant_shard_id,
6255 : split_req.new_shard_count
6256 : );
6257 0 : continue;
6258 0 : }
6259 :
6260 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
6261 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
6262 0 : ))?;
6263 :
6264 0 : let node = pageservers
6265 0 : .get(&node_id)
6266 0 : .expect("Pageservers may not be deleted while referenced");
6267 :
6268 0 : targets.push(ShardSplitTarget {
6269 0 : parent_id: *tenant_shard_id,
6270 0 : node: node.clone(),
6271 0 : child_ids: tenant_shard_id
6272 0 : .split(ShardCount::new(split_req.new_shard_count)),
6273 0 : });
6274 : }
6275 :
6276 0 : if targets.is_empty() {
6277 0 : if children_found.len() == split_req.new_shard_count as usize {
6278 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
6279 0 : new_shards: children_found,
6280 0 : }));
6281 : } else {
6282 : // No shards found to split, and no existing children found: the
6283 : // tenant doesn't exist at all.
6284 0 : return Err(ApiError::NotFound(
6285 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
6286 0 : ));
6287 : }
6288 0 : }
6289 :
6290 0 : (old_shard_count, targets)
6291 : };
6292 :
6293 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
6294 0 : let old_shard_count = old_shard_count.unwrap();
6295 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
6296 : // This ShardIdentity will be used as the template for all children, so this implicitly
6297 : // applies the new stripe size to the children.
6298 0 : let mut shard_ident = shard_ident.unwrap();
6299 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
6300 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6301 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
6302 0 : shard_ident.stripe_size
6303 0 : )));
6304 0 : }
6305 :
6306 0 : shard_ident.stripe_size = new_stripe_size;
6307 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
6308 0 : shard_ident
6309 : } else {
6310 0 : shard_ident.unwrap()
6311 : };
6312 0 : let policy = policy.unwrap();
6313 0 : let config = config.unwrap();
6314 :
6315 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
6316 0 : old_shard_count,
6317 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
6318 0 : new_stripe_size: split_req.new_stripe_size,
6319 0 : targets,
6320 0 : policy,
6321 0 : config,
6322 0 : shard_ident,
6323 0 : preferred_az_id,
6324 0 : })))
6325 0 : }
6326 :
6327 0 : async fn do_tenant_shard_split(
6328 0 : &self,
6329 0 : tenant_id: TenantId,
6330 0 : params: Box<ShardSplitParams>,
6331 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
6332 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
6333 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
6334 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
6335 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
6336 : // (https://github.com/neondatabase/neon/issues/6676)
6337 :
6338 : let ShardSplitParams {
6339 0 : old_shard_count,
6340 0 : new_shard_count,
6341 0 : new_stripe_size,
6342 0 : mut targets,
6343 0 : policy,
6344 0 : config,
6345 0 : shard_ident,
6346 0 : preferred_az_id,
6347 0 : } = *params;
6348 :
6349 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
6350 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
6351 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
6352 : // at the time of split.
6353 0 : let waiters = {
6354 0 : let mut locked = self.inner.write().unwrap();
6355 0 : let mut waiters = Vec::new();
6356 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6357 0 : for target in &mut targets {
6358 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
6359 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6360 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6361 0 : "Shard {} not found",
6362 0 : target.parent_id
6363 0 : )));
6364 : };
6365 :
6366 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
6367 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6368 0 : return Err(ApiError::Conflict(format!(
6369 0 : "Shard {} unexpectedly rescheduled during split",
6370 0 : target.parent_id
6371 0 : )));
6372 0 : }
6373 :
6374 : // Irrespective of PlacementPolicy, clear secondary locations from intent
6375 0 : shard.intent.clear_secondary(scheduler);
6376 :
6377 : // Run Reconciler to execute detach fo secondary locations.
6378 0 : if let Some(waiter) =
6379 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6380 0 : {
6381 0 : waiters.push(waiter);
6382 0 : }
6383 : }
6384 0 : waiters
6385 : };
6386 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
6387 :
6388 : // Before creating any new child shards in memory or on the pageservers, persist them: this
6389 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
6390 : // acts as the protection against two concurrent attempts to split: one of them will get a database
6391 : // error trying to insert the child shards.
6392 0 : let mut child_tsps = Vec::new();
6393 0 : for target in &targets {
6394 0 : let mut this_child_tsps = Vec::new();
6395 0 : for child in &target.child_ids {
6396 0 : let mut child_shard = shard_ident;
6397 0 : child_shard.number = child.shard_number;
6398 0 : child_shard.count = child.shard_count;
6399 :
6400 0 : tracing::info!(
6401 0 : "Create child shard persistence with stripe size {}",
6402 : shard_ident.stripe_size.0
6403 : );
6404 :
6405 0 : this_child_tsps.push(TenantShardPersistence {
6406 0 : tenant_id: child.tenant_id.to_string(),
6407 0 : shard_number: child.shard_number.0 as i32,
6408 0 : shard_count: child.shard_count.literal() as i32,
6409 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
6410 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
6411 : // populate the correct generation as part of its transaction, to protect us
6412 : // against racing with changes in the state of the parent.
6413 0 : generation: None,
6414 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
6415 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
6416 0 : config: serde_json::to_string(&config).unwrap(),
6417 0 : splitting: SplitState::Splitting,
6418 :
6419 : // Scheduling policies and preferred AZ do not carry through to children
6420 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
6421 0 : .unwrap(),
6422 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
6423 : });
6424 : }
6425 :
6426 0 : child_tsps.push((target.parent_id, this_child_tsps));
6427 : }
6428 :
6429 0 : if let Err(e) = self
6430 0 : .persistence
6431 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
6432 0 : .await
6433 : {
6434 0 : match e {
6435 : DatabaseError::Query(diesel::result::Error::DatabaseError(
6436 : DatabaseErrorKind::UniqueViolation,
6437 : _,
6438 : )) => {
6439 : // Inserting a child shard violated a unique constraint: we raced with another call to
6440 : // this function
6441 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
6442 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
6443 : }
6444 0 : _ => return Err(ApiError::InternalServerError(e.into())),
6445 : }
6446 0 : }
6447 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
6448 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6449 : ));
6450 :
6451 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
6452 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
6453 : // is not set in memory, then it was not persisted.
6454 : {
6455 0 : let mut locked = self.inner.write().unwrap();
6456 0 : for target in &targets {
6457 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
6458 0 : parent_shard.splitting = SplitState::Splitting;
6459 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
6460 0 : // split operation.
6461 0 : parent_shard
6462 0 : .observed
6463 0 : .locations
6464 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
6465 0 : }
6466 : }
6467 : }
6468 :
6469 : // TODO: issue split calls concurrently (this only matters once we're splitting
6470 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
6471 :
6472 : // HADRON: set a timeout for splitting individual shards on page servers.
6473 : // Currently we do not perform any retry because it's not clear if page server can handle
6474 : // partially split shards correctly.
6475 0 : let shard_split_timeout =
6476 0 : if let Some(env::DeploymentMode::Local) = env::get_deployment_mode() {
6477 0 : Duration::from_secs(30)
6478 : } else {
6479 0 : self.config.shard_split_request_timeout
6480 : };
6481 0 : let mut http_client_builder = reqwest::ClientBuilder::new()
6482 0 : .pool_max_idle_per_host(0)
6483 0 : .timeout(shard_split_timeout);
6484 :
6485 0 : for ssl_ca_cert in &self.config.ssl_ca_certs {
6486 0 : http_client_builder = http_client_builder.add_root_certificate(ssl_ca_cert.clone());
6487 0 : }
6488 0 : let http_client = http_client_builder
6489 0 : .build()
6490 0 : .expect("Failed to construct HTTP client");
6491 0 : for target in &targets {
6492 : let ShardSplitTarget {
6493 0 : parent_id,
6494 0 : node,
6495 0 : child_ids,
6496 0 : } = target;
6497 :
6498 0 : let client = PageserverClient::new(
6499 0 : node.get_id(),
6500 0 : http_client.clone(),
6501 0 : node.base_url(),
6502 0 : self.config.pageserver_jwt_token.as_deref(),
6503 : );
6504 :
6505 0 : let response = client
6506 0 : .tenant_shard_split(
6507 0 : *parent_id,
6508 0 : TenantShardSplitRequest {
6509 0 : new_shard_count: new_shard_count.literal(),
6510 0 : new_stripe_size,
6511 0 : },
6512 0 : )
6513 0 : .await
6514 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {parent_id}: {e}")))?;
6515 :
6516 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
6517 0 : "failpoint".to_string()
6518 0 : )));
6519 :
6520 0 : failpoint_support::sleep_millis_async!(
6521 : "shard-split-post-remote-sleep",
6522 0 : &self.reconcilers_cancel
6523 : );
6524 :
6525 0 : tracing::info!(
6526 0 : "Split {} into {}",
6527 : parent_id,
6528 0 : response
6529 0 : .new_shards
6530 0 : .iter()
6531 0 : .map(|s| format!("{s:?}"))
6532 0 : .collect::<Vec<_>>()
6533 0 : .join(",")
6534 : );
6535 :
6536 0 : if &response.new_shards != child_ids {
6537 : // This should never happen: the pageserver should agree with us on how shard splits work.
6538 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6539 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
6540 0 : parent_id,
6541 0 : response.new_shards,
6542 0 : child_ids
6543 0 : )));
6544 0 : }
6545 : }
6546 :
6547 0 : fail::fail_point!("shard-split-pre-complete", |_| Err(ApiError::Conflict(
6548 0 : "failpoint".to_string()
6549 0 : )));
6550 :
6551 0 : pausable_failpoint!("shard-split-pre-complete-pause");
6552 :
6553 : // TODO: if the pageserver restarted concurrently with our split API call,
6554 : // the actual generation of the child shard might differ from the generation
6555 : // we expect it to have. In order for our in-database generation to end up
6556 : // correct, we should carry the child generation back in the response and apply it here
6557 : // in complete_shard_split (and apply the correct generation in memory)
6558 : // (or, we can carry generation in the request and reject the request if
6559 : // it doesn't match, but that requires more retry logic on this side)
6560 :
6561 0 : self.persistence
6562 0 : .complete_shard_split(tenant_id, old_shard_count, new_shard_count)
6563 0 : .await?;
6564 :
6565 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
6566 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6567 : ));
6568 :
6569 : // Replace all the shards we just split with their children: this phase is infallible.
6570 0 : let (response, child_locations, waiters) =
6571 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
6572 :
6573 : // Notify all page servers to detach and clean up the old shards because they will no longer
6574 : // be needed. This is best-effort: if it fails, it will be cleaned up on a subsequent
6575 : // Pageserver re-attach/startup.
6576 0 : let shards_to_cleanup = targets
6577 0 : .iter()
6578 0 : .map(|target| (target.parent_id, target.node.get_id()))
6579 0 : .collect();
6580 0 : self.cleanup_locations(shards_to_cleanup).await;
6581 :
6582 : // Send compute notifications for all the new shards
6583 0 : let mut failed_notifications = Vec::new();
6584 0 : for (child_id, child_ps, stripe_size) in child_locations {
6585 0 : if let Err(e) = self
6586 0 : .compute_hook
6587 0 : .notify_attach(
6588 0 : compute_hook::ShardUpdate {
6589 0 : tenant_shard_id: child_id,
6590 0 : node_id: child_ps,
6591 0 : stripe_size,
6592 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
6593 0 : },
6594 0 : &self.reconcilers_cancel,
6595 0 : )
6596 0 : .await
6597 : {
6598 0 : tracing::warn!(
6599 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
6600 : child_id,
6601 : child_ps
6602 : );
6603 0 : failed_notifications.push(child_id);
6604 0 : }
6605 : }
6606 :
6607 : // If we failed any compute notifications, make a note to retry later.
6608 0 : if !failed_notifications.is_empty() {
6609 0 : let mut locked = self.inner.write().unwrap();
6610 0 : for failed in failed_notifications {
6611 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
6612 0 : shard.pending_compute_notification = true;
6613 0 : }
6614 : }
6615 0 : }
6616 :
6617 0 : Ok((response, waiters))
6618 0 : }
6619 :
6620 : /// A graceful migration: update the preferred node and let optimisation handle the migration
6621 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
6622 : ///
6623 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
6624 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
6625 0 : fn tenant_shard_migrate_with_prewarm(
6626 0 : &self,
6627 0 : migrate_req: &TenantShardMigrateRequest,
6628 0 : shard: &mut TenantShard,
6629 0 : scheduler: &mut Scheduler,
6630 0 : schedule_context: ScheduleContext,
6631 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
6632 0 : shard.set_preferred_node(Some(migrate_req.node_id));
6633 :
6634 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
6635 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
6636 : // e.g. by checking secondary is warm enough.
6637 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
6638 0 : }
6639 :
6640 : /// Immediate migration: directly update the intent state and kick off a reconciler
6641 0 : fn tenant_shard_migrate_immediate(
6642 0 : &self,
6643 0 : migrate_req: &TenantShardMigrateRequest,
6644 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6645 0 : shard: &mut TenantShard,
6646 0 : scheduler: &mut Scheduler,
6647 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
6648 : // Non-graceful migration: update the intent state immediately
6649 0 : let old_attached = *shard.intent.get_attached();
6650 0 : match shard.policy {
6651 0 : PlacementPolicy::Attached(n) => {
6652 : // If our new attached node was a secondary, it no longer should be.
6653 0 : shard
6654 0 : .intent
6655 0 : .remove_secondary(scheduler, migrate_req.node_id);
6656 :
6657 0 : shard
6658 0 : .intent
6659 0 : .set_attached(scheduler, Some(migrate_req.node_id));
6660 :
6661 : // If we were already attached to something, demote that to a secondary
6662 0 : if let Some(old_attached) = old_attached {
6663 0 : if n > 0 {
6664 : // Remove other secondaries to make room for the location we'll demote
6665 0 : while shard.intent.get_secondary().len() >= n {
6666 0 : shard.intent.pop_secondary(scheduler);
6667 0 : }
6668 :
6669 0 : shard.intent.push_secondary(scheduler, old_attached);
6670 0 : }
6671 0 : }
6672 : }
6673 0 : PlacementPolicy::Secondary => {
6674 0 : shard.intent.clear(scheduler);
6675 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6676 0 : }
6677 : PlacementPolicy::Detached => {
6678 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6679 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
6680 0 : )));
6681 : }
6682 : }
6683 :
6684 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
6685 0 : shard.sequence = shard.sequence.next();
6686 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
6687 0 : Ok(self.maybe_configured_reconcile_shard(
6688 0 : shard,
6689 0 : nodes,
6690 0 : (&migrate_req.migration_config).into(),
6691 0 : ))
6692 0 : }
6693 :
6694 0 : pub(crate) async fn tenant_shard_migrate(
6695 0 : &self,
6696 0 : tenant_shard_id: TenantShardId,
6697 0 : migrate_req: TenantShardMigrateRequest,
6698 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6699 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
6700 : // get a different outcome to handle
6701 : enum MigrationOutcome {
6702 : Optimization(Option<ScheduleOptimization>),
6703 : Reconcile(Option<ReconcilerWaiter>),
6704 : }
6705 :
6706 0 : let outcome = {
6707 0 : let mut locked = self.inner.write().unwrap();
6708 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6709 :
6710 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6711 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6712 0 : "Node {} not found",
6713 0 : migrate_req.node_id
6714 0 : )));
6715 : };
6716 :
6717 : // Migration to unavavailable node requires force flag
6718 0 : if !node.is_available() {
6719 0 : if migrate_req.migration_config.override_scheduler {
6720 : // Warn but proceed: the caller may intend to manually adjust the placement of
6721 : // a shard even if the node is down, e.g. if intervening during an incident.
6722 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
6723 : } else {
6724 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
6725 0 : return Err(ApiError::PreconditionFailed(
6726 0 : format!("Node {node} is unavailable").into_boxed_str(),
6727 0 : ));
6728 : }
6729 0 : }
6730 :
6731 : // Calculate the ScheduleContext for this tenant
6732 0 : let mut schedule_context = ScheduleContext::default();
6733 0 : for (_shard_id, shard) in
6734 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
6735 0 : {
6736 0 : schedule_context.avoid(&shard.intent.all_pageservers());
6737 0 : }
6738 :
6739 : // Look up the specific shard we will migrate
6740 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6741 0 : return Err(ApiError::NotFound(
6742 0 : anyhow::anyhow!("Tenant shard not found").into(),
6743 0 : ));
6744 : };
6745 :
6746 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
6747 : // be migrated back by the optimiser.
6748 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
6749 0 : scheduler,
6750 0 : &schedule_context,
6751 0 : migrate_req.node_id,
6752 0 : &[],
6753 0 : ) {
6754 0 : if !migrate_req.migration_config.override_scheduler {
6755 0 : return Err(ApiError::PreconditionFailed(
6756 0 : "Migration to a worse-scoring node".into(),
6757 0 : ));
6758 : } else {
6759 0 : tracing::info!(
6760 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
6761 : migrate_req.node_id
6762 : );
6763 : }
6764 0 : }
6765 :
6766 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
6767 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
6768 0 : return Err(ApiError::PreconditionFailed(
6769 0 : format!(
6770 0 : "Migration expected to originate from {} but shard is on {:?}",
6771 0 : origin_node_id,
6772 0 : shard.intent.get_attached()
6773 0 : )
6774 0 : .into(),
6775 0 : ));
6776 0 : }
6777 0 : }
6778 :
6779 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6780 : // No-op case: we will still proceed to wait for reconciliation in case it is
6781 : // incomplete from an earlier update to the intent.
6782 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
6783 :
6784 : // An instruction to migrate to the currently attached node should
6785 : // cancel any pending graceful migration
6786 0 : shard.set_preferred_node(None);
6787 :
6788 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
6789 0 : shard,
6790 0 : nodes,
6791 0 : (&migrate_req.migration_config).into(),
6792 0 : ))
6793 0 : } else if migrate_req.migration_config.prewarm {
6794 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
6795 0 : &migrate_req,
6796 0 : shard,
6797 0 : scheduler,
6798 0 : schedule_context,
6799 0 : )?)
6800 : } else {
6801 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
6802 0 : &migrate_req,
6803 0 : nodes,
6804 0 : shard,
6805 0 : scheduler,
6806 0 : )?)
6807 : }
6808 : };
6809 :
6810 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
6811 0 : let waiter = match outcome {
6812 0 : MigrationOutcome::Optimization(Some(optimization)) => {
6813 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
6814 : // we might as well do it more promptly as this is a direct external request.
6815 0 : let mut validated = self
6816 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
6817 0 : .await;
6818 0 : if let Some((_shard_id, optimization)) = validated.pop() {
6819 0 : let mut locked = self.inner.write().unwrap();
6820 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6821 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6822 : // Rare but possible: tenant is removed between generating optimisation and validating it.
6823 0 : return Err(ApiError::NotFound(
6824 0 : anyhow::anyhow!("Tenant shard not found").into(),
6825 0 : ));
6826 : };
6827 :
6828 0 : if !shard.apply_optimization(scheduler, optimization) {
6829 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
6830 : // and therefore not applied.
6831 0 : tracing::warn!(
6832 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
6833 : );
6834 0 : }
6835 0 : self.maybe_configured_reconcile_shard(
6836 0 : shard,
6837 0 : nodes,
6838 0 : (&migrate_req.migration_config).into(),
6839 : )
6840 : } else {
6841 0 : None
6842 : }
6843 : }
6844 0 : MigrationOutcome::Optimization(None) => None,
6845 0 : MigrationOutcome::Reconcile(waiter) => waiter,
6846 : };
6847 :
6848 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
6849 : // locations, this has a good chance of timing out.
6850 0 : if let Some(waiter) = waiter {
6851 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6852 : } else {
6853 0 : tracing::info!("Migration is a no-op");
6854 : }
6855 :
6856 0 : Ok(TenantShardMigrateResponse {})
6857 0 : }
6858 :
6859 0 : pub(crate) async fn tenant_shard_migrate_secondary(
6860 0 : &self,
6861 0 : tenant_shard_id: TenantShardId,
6862 0 : migrate_req: TenantShardMigrateRequest,
6863 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6864 0 : let waiter = {
6865 0 : let mut locked = self.inner.write().unwrap();
6866 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6867 :
6868 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6869 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6870 0 : "Node {} not found",
6871 0 : migrate_req.node_id
6872 0 : )));
6873 : };
6874 :
6875 0 : if !node.is_available() {
6876 : // Warn but proceed: the caller may intend to manually adjust the placement of
6877 : // a shard even if the node is down, e.g. if intervening during an incident.
6878 0 : tracing::warn!("Migrating to unavailable node {node}");
6879 0 : }
6880 :
6881 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6882 0 : return Err(ApiError::NotFound(
6883 0 : anyhow::anyhow!("Tenant shard not found").into(),
6884 0 : ));
6885 : };
6886 :
6887 0 : if shard.intent.get_secondary().len() == 1
6888 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
6889 : {
6890 0 : tracing::info!(
6891 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
6892 : shard.intent
6893 : );
6894 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6895 0 : tracing::info!(
6896 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
6897 : );
6898 : } else {
6899 0 : let old_secondaries = shard.intent.get_secondary().clone();
6900 0 : for secondary in old_secondaries {
6901 0 : shard.intent.remove_secondary(scheduler, secondary);
6902 0 : }
6903 :
6904 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6905 0 : shard.sequence = shard.sequence.next();
6906 0 : tracing::info!(
6907 0 : "Migrating secondary to {node}: new intent {:?}",
6908 : shard.intent
6909 : );
6910 : }
6911 :
6912 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6913 : };
6914 :
6915 0 : if let Some(waiter) = waiter {
6916 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6917 : } else {
6918 0 : tracing::info!("Migration is a no-op");
6919 : }
6920 :
6921 0 : Ok(TenantShardMigrateResponse {})
6922 0 : }
6923 :
6924 : /// 'cancel' in this context means cancel any ongoing reconcile
6925 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
6926 0 : &self,
6927 0 : tenant_shard_id: TenantShardId,
6928 0 : ) -> Result<(), ApiError> {
6929 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
6930 0 : let waiter = {
6931 0 : let locked = self.inner.write().unwrap();
6932 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6933 0 : return Err(ApiError::NotFound(
6934 0 : anyhow::anyhow!("Tenant shard not found").into(),
6935 0 : ));
6936 : };
6937 :
6938 0 : let waiter = shard.get_waiter();
6939 0 : match waiter {
6940 : None => {
6941 0 : tracing::info!("Shard does not have an ongoing Reconciler");
6942 0 : return Ok(());
6943 : }
6944 0 : Some(waiter) => {
6945 0 : tracing::info!("Cancelling Reconciler");
6946 0 : shard.cancel_reconciler();
6947 0 : waiter
6948 : }
6949 : }
6950 : };
6951 :
6952 : // Cancellation should be prompt. If this fails we have still done our job of firing the
6953 : // cancellation token, but by returning an ApiError we will indicate to the caller that
6954 : // the Reconciler is misbehaving and not respecting the cancellation token
6955 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
6956 0 : .await?;
6957 :
6958 0 : Ok(())
6959 0 : }
6960 :
6961 : /// This is for debug/support only: we simply drop all state for a tenant, without
6962 : /// detaching or deleting it on pageservers.
6963 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
6964 0 : self.persistence.delete_tenant(tenant_id).await?;
6965 :
6966 0 : let mut locked = self.inner.write().unwrap();
6967 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6968 0 : let mut shards = Vec::new();
6969 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
6970 0 : shards.push(*tenant_shard_id);
6971 0 : }
6972 :
6973 0 : for shard_id in shards {
6974 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
6975 0 : shard.intent.clear(scheduler);
6976 0 : }
6977 : }
6978 :
6979 0 : Ok(())
6980 0 : }
6981 :
6982 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
6983 : /// tenant with a very high generation number so that it will see the existing data.
6984 : /// It does not create timelines on safekeepers, because they might already exist on some
6985 : /// safekeeper set. So, the timelines are not storcon-managed after the import.
6986 0 : pub(crate) async fn tenant_import(
6987 0 : &self,
6988 0 : tenant_id: TenantId,
6989 0 : ) -> Result<TenantCreateResponse, ApiError> {
6990 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
6991 0 : let maybe_node = {
6992 0 : self.inner
6993 0 : .read()
6994 0 : .unwrap()
6995 0 : .nodes
6996 0 : .values()
6997 0 : .find(|n| n.is_available())
6998 0 : .cloned()
6999 : };
7000 0 : let Some(node) = maybe_node else {
7001 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
7002 : };
7003 :
7004 0 : let client = PageserverClient::new(
7005 0 : node.get_id(),
7006 0 : self.http_client.clone(),
7007 0 : node.base_url(),
7008 0 : self.config.pageserver_jwt_token.as_deref(),
7009 : );
7010 :
7011 0 : let scan_result = client
7012 0 : .tenant_scan_remote_storage(tenant_id)
7013 0 : .await
7014 0 : .map_err(|e| passthrough_api_error(&node, e))?;
7015 :
7016 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
7017 0 : let Some(shard_count) = scan_result
7018 0 : .shards
7019 0 : .iter()
7020 0 : .map(|s| s.tenant_shard_id.shard_count)
7021 0 : .max()
7022 : else {
7023 0 : return Err(ApiError::NotFound(
7024 0 : anyhow::anyhow!("No shards found").into(),
7025 0 : ));
7026 : };
7027 :
7028 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
7029 : // to
7030 0 : let generation = scan_result
7031 0 : .shards
7032 0 : .iter()
7033 0 : .map(|s| s.generation)
7034 0 : .max()
7035 0 : .expect("We already validated >0 shards");
7036 :
7037 : // Find the tenant's stripe size. This wasn't always persisted in the tenant manifest, so
7038 : // fall back to the original default stripe size of 32768 (256 MB) if it's not specified.
7039 : const ORIGINAL_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(32768);
7040 0 : let stripe_size = scan_result
7041 0 : .shards
7042 0 : .iter()
7043 0 : .find(|s| s.tenant_shard_id.shard_count == shard_count && s.generation == generation)
7044 0 : .expect("we validated >0 shards above")
7045 : .stripe_size
7046 0 : .unwrap_or_else(|| {
7047 0 : if shard_count.count() > 1 {
7048 0 : warn!("unknown stripe size, assuming {ORIGINAL_STRIPE_SIZE}");
7049 0 : }
7050 0 : ORIGINAL_STRIPE_SIZE
7051 0 : });
7052 :
7053 0 : let (response, waiters) = self
7054 0 : .do_tenant_create(TenantCreateRequest {
7055 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
7056 0 : generation,
7057 0 :
7058 0 : shard_parameters: ShardParameters {
7059 0 : count: shard_count,
7060 0 : stripe_size,
7061 0 : },
7062 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
7063 0 : config: TenantConfig::default(),
7064 0 : })
7065 0 : .await?;
7066 :
7067 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
7068 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
7069 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
7070 : // reconcile, as reconciliation includes notifying compute.
7071 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
7072 0 : }
7073 :
7074 0 : Ok(response)
7075 0 : }
7076 :
7077 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
7078 : /// we don't have to make TenantShard clonable in the return path.
7079 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7080 0 : let serialized = {
7081 0 : let locked = self.inner.read().unwrap();
7082 0 : let result = locked.tenants.values().collect::<Vec<_>>();
7083 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
7084 : };
7085 :
7086 0 : hyper::Response::builder()
7087 0 : .status(hyper::StatusCode::OK)
7088 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7089 0 : .body(hyper::Body::from(serialized))
7090 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7091 0 : }
7092 :
7093 : /// Check the consistency of in-memory state vs. persistent state, and check that the
7094 : /// scheduler's statistics are up to date.
7095 : ///
7096 : /// These consistency checks expect an **idle** system. If changes are going on while
7097 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
7098 : /// checks, but not suitable for running continuously in the background in the field.
7099 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
7100 0 : let (mut expect_nodes, mut expect_shards) = {
7101 0 : let locked = self.inner.read().unwrap();
7102 :
7103 0 : locked
7104 0 : .scheduler
7105 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
7106 0 : .context("Scheduler checks")
7107 0 : .map_err(ApiError::InternalServerError)?;
7108 :
7109 0 : let expect_nodes = locked
7110 0 : .nodes
7111 0 : .values()
7112 0 : .map(|n| n.to_persistent())
7113 0 : .collect::<Vec<_>>();
7114 :
7115 0 : let expect_shards = locked
7116 0 : .tenants
7117 0 : .values()
7118 0 : .map(|t| t.to_persistent())
7119 0 : .collect::<Vec<_>>();
7120 :
7121 : // This method can only validate the state of an idle system: if a reconcile is in
7122 : // progress, fail out early to avoid giving false errors on state that won't match
7123 : // between database and memory under a ReconcileResult is processed.
7124 0 : for t in locked.tenants.values() {
7125 0 : if t.reconciler.is_some() {
7126 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7127 0 : "Shard {} reconciliation in progress",
7128 0 : t.tenant_shard_id
7129 0 : )));
7130 0 : }
7131 : }
7132 :
7133 0 : (expect_nodes, expect_shards)
7134 : };
7135 :
7136 0 : let mut nodes = self.persistence.list_nodes().await?;
7137 0 : expect_nodes.sort_by_key(|n| n.node_id);
7138 0 : nodes.sort_by_key(|n| n.node_id);
7139 :
7140 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
7141 0 : let node_result = if nodes != expect_nodes {
7142 0 : tracing::error!("Consistency check failed on nodes.");
7143 0 : tracing::error!(
7144 0 : "Nodes in memory: {}",
7145 0 : serde_json::to_string(&expect_nodes)
7146 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7147 : );
7148 0 : tracing::error!(
7149 0 : "Nodes in database: {}",
7150 0 : serde_json::to_string(&nodes)
7151 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7152 : );
7153 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
7154 0 : "Node consistency failure"
7155 0 : )))
7156 : } else {
7157 0 : Ok(())
7158 : };
7159 :
7160 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
7161 0 : persistent_shards
7162 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7163 :
7164 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7165 :
7166 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
7167 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
7168 : // fields are added, before doing a comparison.
7169 0 : for tsp in &mut persistent_shards {
7170 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
7171 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
7172 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
7173 : }
7174 :
7175 0 : if persistent_shards != expect_shards {
7176 0 : tracing::error!("Consistency check failed on shards.");
7177 :
7178 0 : tracing::error!(
7179 0 : "Shards in memory: {}",
7180 0 : serde_json::to_string(&expect_shards)
7181 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7182 : );
7183 0 : tracing::error!(
7184 0 : "Shards in database: {}",
7185 0 : serde_json::to_string(&persistent_shards)
7186 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7187 : );
7188 :
7189 : // The total dump log lines above are useful in testing but in the field grafana will
7190 : // usually just drop them because they're so large. So we also do some explicit logging
7191 : // of just the diffs.
7192 0 : let persistent_shards = persistent_shards
7193 0 : .into_iter()
7194 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7195 0 : .collect::<HashMap<_, _>>();
7196 0 : let expect_shards = expect_shards
7197 0 : .into_iter()
7198 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7199 0 : .collect::<HashMap<_, _>>();
7200 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
7201 0 : match expect_shards.get(tenant_shard_id) {
7202 : None => {
7203 0 : tracing::error!(
7204 0 : "Shard {} found in database but not in memory",
7205 : tenant_shard_id
7206 : );
7207 : }
7208 0 : Some(expect_tsp) => {
7209 0 : if expect_tsp != persistent_tsp {
7210 0 : tracing::error!(
7211 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
7212 : tenant_shard_id,
7213 0 : serde_json::to_string(expect_tsp).unwrap(),
7214 0 : serde_json::to_string(&persistent_tsp).unwrap()
7215 : );
7216 0 : }
7217 : }
7218 : }
7219 : }
7220 :
7221 : // Having already logged any differences, log any shards that simply aren't present in the database
7222 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
7223 0 : if !persistent_shards.contains_key(tenant_shard_id) {
7224 0 : tracing::error!(
7225 0 : "Shard {} found in memory but not in database: {}",
7226 : tenant_shard_id,
7227 0 : serde_json::to_string(memory_tsp)
7228 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7229 : );
7230 0 : }
7231 : }
7232 :
7233 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7234 0 : "Shard consistency failure"
7235 0 : )));
7236 0 : }
7237 :
7238 0 : node_result
7239 0 : }
7240 :
7241 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
7242 : /// we don't have to make TenantShard clonable in the return path.
7243 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7244 0 : let serialized = {
7245 0 : let locked = self.inner.read().unwrap();
7246 0 : serde_json::to_string(&locked.scheduler)
7247 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7248 : };
7249 :
7250 0 : hyper::Response::builder()
7251 0 : .status(hyper::StatusCode::OK)
7252 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7253 0 : .body(hyper::Body::from(serialized))
7254 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7255 0 : }
7256 :
7257 : /// This is for debug/support only: we simply drop all state for a tenant, without
7258 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
7259 : /// tenants that were on this node.
7260 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
7261 0 : self.persistence.set_tombstone(node_id).await?;
7262 :
7263 0 : let mut locked = self.inner.write().unwrap();
7264 :
7265 0 : for shard in locked.tenants.values_mut() {
7266 0 : shard.deref_node(node_id);
7267 0 : shard.observed.locations.remove(&node_id);
7268 0 : }
7269 :
7270 0 : let mut nodes = (*locked.nodes).clone();
7271 0 : nodes.remove(&node_id);
7272 0 : locked.nodes = Arc::new(nodes);
7273 0 : metrics::METRICS_REGISTRY
7274 0 : .metrics_group
7275 0 : .storage_controller_pageserver_nodes
7276 0 : .set(locked.nodes.len() as i64);
7277 0 : metrics::METRICS_REGISTRY
7278 0 : .metrics_group
7279 0 : .storage_controller_https_pageserver_nodes
7280 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7281 :
7282 0 : locked.scheduler.node_remove(node_id);
7283 :
7284 0 : Ok(())
7285 0 : }
7286 :
7287 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
7288 : /// that we don't leave any bad state behind in the storage controller, but unclean
7289 : /// in the sense that we are not carefully draining the node.
7290 0 : pub(crate) async fn node_delete_old(&self, node_id: NodeId) -> Result<(), ApiError> {
7291 0 : let _node_lock =
7292 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
7293 :
7294 : // 1. Atomically update in-memory state:
7295 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
7296 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
7297 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
7298 : // re-insert references to this node into the ObservedState of shards
7299 : // - drop the node from the scheduler
7300 : {
7301 0 : let mut locked = self.inner.write().unwrap();
7302 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7303 :
7304 : {
7305 0 : let mut nodes_mut = (*nodes).deref().clone();
7306 0 : match nodes_mut.get_mut(&node_id) {
7307 0 : Some(node) => {
7308 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
7309 0 : // if we crash it would not be desirable to leave the node paused after a restart.
7310 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
7311 0 : }
7312 : None => {
7313 0 : tracing::info!(
7314 0 : "Node not found: presuming this is a retry and returning success"
7315 : );
7316 0 : return Ok(());
7317 : }
7318 : }
7319 :
7320 0 : *nodes = Arc::new(nodes_mut);
7321 : }
7322 :
7323 0 : for (_tenant_id, mut schedule_context, shards) in
7324 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7325 : {
7326 0 : for shard in shards {
7327 0 : if shard.deref_node(node_id) {
7328 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
7329 : // TODO: implement force flag to remove a node even if we can't reschedule
7330 : // a tenant
7331 0 : tracing::error!(
7332 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7333 : shard.tenant_shard_id
7334 : );
7335 0 : return Err(e.into());
7336 : } else {
7337 0 : tracing::info!(
7338 0 : "Rescheduled shard {} away from node during deletion",
7339 : shard.tenant_shard_id
7340 : )
7341 : }
7342 :
7343 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
7344 0 : }
7345 :
7346 : // Here we remove an existing observed location for the node we're removing, and it will
7347 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7348 : // process_result.
7349 : //
7350 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
7351 : // means any reconciles we spawned will know about the node we're deleting, enabling them
7352 : // to do live migrations if it's still online.
7353 0 : shard.observed.locations.remove(&node_id);
7354 : }
7355 : }
7356 :
7357 0 : scheduler.node_remove(node_id);
7358 :
7359 : {
7360 0 : let mut nodes_mut = (**nodes).clone();
7361 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7362 0 : // Ensure that any reconciler holding an Arc<> to this node will
7363 0 : // drop out when trying to RPC to it (setting Offline state sets the
7364 0 : // cancellation token on the Node object).
7365 0 : removed_node.set_availability(NodeAvailability::Offline);
7366 0 : }
7367 0 : *nodes = Arc::new(nodes_mut);
7368 0 : metrics::METRICS_REGISTRY
7369 0 : .metrics_group
7370 0 : .storage_controller_pageserver_nodes
7371 0 : .set(nodes.len() as i64);
7372 0 : metrics::METRICS_REGISTRY
7373 0 : .metrics_group
7374 0 : .storage_controller_https_pageserver_nodes
7375 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7376 : }
7377 : }
7378 :
7379 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
7380 : // the removed node, as this column means "The pageserver to which this generation was issued", and
7381 : // their generations won't get updated until the reconcilers moving them away from this node complete.
7382 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
7383 : // that exists.
7384 :
7385 : // 2. Actually delete the node from in-memory state and set tombstone to the database
7386 : // for preventing the node to register again.
7387 0 : tracing::info!("Deleting node from database");
7388 0 : self.persistence.set_tombstone(node_id).await?;
7389 :
7390 0 : Ok(())
7391 0 : }
7392 :
7393 0 : pub(crate) async fn delete_node(
7394 0 : self: &Arc<Self>,
7395 0 : node_id: NodeId,
7396 0 : policy_on_start: NodeSchedulingPolicy,
7397 0 : force: bool,
7398 0 : cancel: CancellationToken,
7399 0 : ) -> Result<(), OperationError> {
7400 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal).build();
7401 :
7402 0 : let mut waiters: Vec<ReconcilerWaiter> = Vec::new();
7403 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
7404 :
7405 0 : let reset_node_policy_on_cancel = || async {
7406 0 : match self
7407 0 : .node_configure(node_id, None, Some(policy_on_start))
7408 0 : .await
7409 : {
7410 0 : Ok(()) => OperationError::Cancelled,
7411 0 : Err(err) => {
7412 0 : OperationError::FinalizeError(
7413 0 : format!(
7414 0 : "Failed to finalise delete cancel of {} by setting scheduling policy to {}: {}",
7415 0 : node_id, String::from(policy_on_start), err
7416 0 : )
7417 0 : .into(),
7418 0 : )
7419 : }
7420 : }
7421 0 : };
7422 :
7423 0 : while !tid_iter.finished() {
7424 0 : if cancel.is_cancelled() {
7425 0 : return Err(reset_node_policy_on_cancel().await);
7426 0 : }
7427 :
7428 0 : operation_utils::validate_node_state(
7429 0 : &node_id,
7430 0 : self.inner.read().unwrap().nodes.clone(),
7431 0 : NodeSchedulingPolicy::Deleting,
7432 0 : )?;
7433 :
7434 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7435 0 : let tid = match tid_iter.next() {
7436 0 : Some(tid) => tid,
7437 : None => {
7438 0 : break;
7439 : }
7440 : };
7441 :
7442 0 : let mut locked = self.inner.write().unwrap();
7443 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7444 :
7445 : // Calculate a schedule context here to avoid borrow checker issues.
7446 0 : let mut schedule_context = ScheduleContext::default();
7447 0 : for (_, shard) in tenants.range(TenantShardId::tenant_range(tid.tenant_id)) {
7448 0 : schedule_context.avoid(&shard.intent.all_pageservers());
7449 0 : }
7450 :
7451 0 : let tenant_shard = match tenants.get_mut(&tid) {
7452 0 : Some(tenant_shard) => tenant_shard,
7453 : None => {
7454 : // Tenant shard was deleted by another operation. Skip it.
7455 0 : continue;
7456 : }
7457 : };
7458 :
7459 0 : match tenant_shard.get_scheduling_policy() {
7460 0 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {
7461 0 : // A migration during delete is classed as 'essential' because it is required to
7462 0 : // uphold our availability goals for the tenant: this shard is elegible for migration.
7463 0 : }
7464 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
7465 : // If we have been asked to avoid rescheduling this shard, then do not migrate it during a deletion
7466 0 : tracing::warn!(
7467 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7468 0 : "Skip migration during deletion because shard scheduling policy {:?} disallows it",
7469 0 : tenant_shard.get_scheduling_policy(),
7470 : );
7471 0 : continue;
7472 : }
7473 : }
7474 :
7475 0 : if tenant_shard.deref_node(node_id) {
7476 0 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
7477 0 : tracing::error!(
7478 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7479 : tenant_shard.tenant_shard_id
7480 : );
7481 0 : return Err(OperationError::ImpossibleConstraint(e.to_string().into()));
7482 : } else {
7483 0 : tracing::info!(
7484 0 : "Rescheduled shard {} away from node during deletion",
7485 : tenant_shard.tenant_shard_id
7486 : )
7487 : }
7488 :
7489 0 : let waiter = self.maybe_configured_reconcile_shard(
7490 0 : tenant_shard,
7491 0 : nodes,
7492 0 : reconciler_config,
7493 0 : );
7494 :
7495 0 : if force {
7496 0 : // Here we remove an existing observed location for the node we're removing, and it will
7497 0 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7498 0 : // process_result.
7499 0 : //
7500 0 : // Note that we update the shard's observed state _after_ calling maybe_configured_reconcile_shard:
7501 0 : // that means any reconciles we spawned will know about the node we're deleting,
7502 0 : // enabling them to do live migrations if it's still online.
7503 0 : tenant_shard.observed.locations.remove(&node_id);
7504 0 : } else if let Some(waiter) = waiter {
7505 0 : waiters.push(waiter);
7506 0 : }
7507 0 : }
7508 : }
7509 :
7510 0 : waiters = self
7511 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
7512 0 : .await;
7513 :
7514 0 : failpoint_support::sleep_millis_async!("sleepy-delete-loop", &cancel);
7515 : }
7516 :
7517 0 : while !waiters.is_empty() {
7518 0 : if cancel.is_cancelled() {
7519 0 : return Err(reset_node_policy_on_cancel().await);
7520 0 : }
7521 :
7522 0 : tracing::info!("Awaiting {} pending delete reconciliations", waiters.len());
7523 :
7524 0 : waiters = self
7525 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7526 0 : .await;
7527 : }
7528 :
7529 0 : let pf = pausable_failpoint!("delete-node-after-reconciles-spawned", &cancel);
7530 0 : if pf.is_err() {
7531 : // An error from pausable_failpoint indicates the cancel token was triggered.
7532 0 : return Err(reset_node_policy_on_cancel().await);
7533 0 : }
7534 :
7535 0 : self.persistence
7536 0 : .set_tombstone(node_id)
7537 0 : .await
7538 0 : .map_err(|e| OperationError::FinalizeError(e.to_string().into()))?;
7539 :
7540 : {
7541 0 : let mut locked = self.inner.write().unwrap();
7542 0 : let (nodes, _, scheduler) = locked.parts_mut();
7543 :
7544 0 : scheduler.node_remove(node_id);
7545 :
7546 0 : let mut nodes_mut = (**nodes).clone();
7547 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7548 0 : // Ensure that any reconciler holding an Arc<> to this node will
7549 0 : // drop out when trying to RPC to it (setting Offline state sets the
7550 0 : // cancellation token on the Node object).
7551 0 : removed_node.set_availability(NodeAvailability::Offline);
7552 0 : }
7553 0 : *nodes = Arc::new(nodes_mut);
7554 :
7555 0 : metrics::METRICS_REGISTRY
7556 0 : .metrics_group
7557 0 : .storage_controller_pageserver_nodes
7558 0 : .set(nodes.len() as i64);
7559 0 : metrics::METRICS_REGISTRY
7560 0 : .metrics_group
7561 0 : .storage_controller_https_pageserver_nodes
7562 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7563 : }
7564 :
7565 0 : Ok(())
7566 0 : }
7567 :
7568 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
7569 0 : let nodes = {
7570 0 : self.inner
7571 0 : .read()
7572 0 : .unwrap()
7573 0 : .nodes
7574 0 : .values()
7575 0 : .cloned()
7576 0 : .collect::<Vec<_>>()
7577 : };
7578 :
7579 0 : Ok(nodes)
7580 0 : }
7581 :
7582 0 : pub(crate) async fn tombstone_list(&self) -> Result<Vec<Node>, ApiError> {
7583 0 : self.persistence
7584 0 : .list_tombstones()
7585 0 : .await?
7586 0 : .into_iter()
7587 0 : .map(|np| Node::from_persistent(np, false))
7588 0 : .collect::<Result<Vec<_>, _>>()
7589 0 : .map_err(ApiError::InternalServerError)
7590 0 : }
7591 :
7592 0 : pub(crate) async fn tombstone_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
7593 0 : let _node_lock = trace_exclusive_lock(
7594 0 : &self.node_op_locks,
7595 0 : node_id,
7596 0 : NodeOperations::DeleteTombstone,
7597 0 : )
7598 0 : .await;
7599 :
7600 0 : if matches!(self.get_node(node_id).await, Err(ApiError::NotFound(_))) {
7601 0 : self.persistence.delete_node(node_id).await?;
7602 0 : Ok(())
7603 : } else {
7604 0 : Err(ApiError::Conflict(format!(
7605 0 : "Node {node_id} is in use, consider using tombstone API first"
7606 0 : )))
7607 : }
7608 0 : }
7609 :
7610 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
7611 0 : self.inner
7612 0 : .read()
7613 0 : .unwrap()
7614 0 : .nodes
7615 0 : .get(&node_id)
7616 0 : .cloned()
7617 0 : .ok_or(ApiError::NotFound(
7618 0 : format!("Node {node_id} not registered").into(),
7619 0 : ))
7620 0 : }
7621 :
7622 0 : pub(crate) async fn get_node_shards(
7623 0 : &self,
7624 0 : node_id: NodeId,
7625 0 : ) -> Result<NodeShardResponse, ApiError> {
7626 0 : let locked = self.inner.read().unwrap();
7627 0 : let mut shards = Vec::new();
7628 0 : for (tid, tenant) in locked.tenants.iter() {
7629 0 : let is_intended_secondary = match (
7630 0 : tenant.intent.get_attached() == &Some(node_id),
7631 0 : tenant.intent.get_secondary().contains(&node_id),
7632 0 : ) {
7633 : (true, true) => {
7634 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7635 0 : "{} attached as primary+secondary on the same node",
7636 0 : tid
7637 0 : )));
7638 : }
7639 0 : (true, false) => Some(false),
7640 0 : (false, true) => Some(true),
7641 0 : (false, false) => None,
7642 : };
7643 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
7644 0 : tenant.observed.locations.get(&node_id)
7645 : {
7646 0 : Some(conf.secondary_conf.is_some())
7647 : } else {
7648 0 : None
7649 : };
7650 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
7651 0 : shards.push(NodeShard {
7652 0 : tenant_shard_id: *tid,
7653 0 : is_intended_secondary,
7654 0 : is_observed_secondary,
7655 0 : });
7656 0 : }
7657 : }
7658 0 : Ok(NodeShardResponse { node_id, shards })
7659 0 : }
7660 :
7661 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
7662 0 : self.persistence.get_leader().await
7663 0 : }
7664 :
7665 0 : pub(crate) async fn node_register(
7666 0 : &self,
7667 0 : register_req: NodeRegisterRequest,
7668 0 : ) -> Result<(), ApiError> {
7669 0 : let _node_lock = trace_exclusive_lock(
7670 0 : &self.node_op_locks,
7671 0 : register_req.node_id,
7672 0 : NodeOperations::Register,
7673 0 : )
7674 0 : .await;
7675 :
7676 : #[derive(PartialEq)]
7677 : enum RegistrationStatus {
7678 : UpToDate,
7679 : NeedUpdate,
7680 : Mismatched,
7681 : New,
7682 : }
7683 :
7684 0 : let registration_status = {
7685 0 : let locked = self.inner.read().unwrap();
7686 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
7687 0 : if node.registration_match(®ister_req) {
7688 0 : if node.need_update(®ister_req) {
7689 0 : RegistrationStatus::NeedUpdate
7690 : } else {
7691 0 : RegistrationStatus::UpToDate
7692 : }
7693 : } else {
7694 0 : RegistrationStatus::Mismatched
7695 : }
7696 : } else {
7697 0 : RegistrationStatus::New
7698 : }
7699 : };
7700 :
7701 0 : match registration_status {
7702 : RegistrationStatus::UpToDate => {
7703 0 : tracing::info!(
7704 0 : "Node {} re-registered with matching address and is up to date",
7705 : register_req.node_id
7706 : );
7707 :
7708 0 : return Ok(());
7709 : }
7710 : RegistrationStatus::Mismatched => {
7711 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
7712 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
7713 : // a fixed address through the lifetime of a node.
7714 0 : tracing::warn!(
7715 0 : "Node {} tried to register with different address",
7716 : register_req.node_id
7717 : );
7718 0 : return Err(ApiError::Conflict(
7719 0 : "Node is already registered with different address".to_string(),
7720 0 : ));
7721 : }
7722 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
7723 0 : // fallthrough
7724 0 : }
7725 : }
7726 :
7727 : // We do not require that a node is actually online when registered (it will start life
7728 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
7729 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
7730 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
7731 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
7732 0 : if tokio::net::lookup_host(format!(
7733 0 : "{}:{}",
7734 : register_req.listen_http_addr, register_req.listen_http_port
7735 : ))
7736 0 : .await
7737 0 : .is_err()
7738 : {
7739 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
7740 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
7741 : // we return a soft 503 error, to encourage callers to retry past transient issues.
7742 0 : return Err(ApiError::ResourceUnavailable(
7743 0 : format!(
7744 0 : "Node {} tried to register with unknown DNS name '{}'",
7745 0 : register_req.node_id, register_req.listen_http_addr
7746 0 : )
7747 0 : .into(),
7748 0 : ));
7749 0 : }
7750 :
7751 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
7752 0 : return Err(ApiError::PreconditionFailed(
7753 0 : format!(
7754 0 : "Node {} has no https port, but use_https is enabled",
7755 0 : register_req.node_id
7756 0 : )
7757 0 : .into(),
7758 0 : ));
7759 0 : }
7760 :
7761 0 : if register_req.listen_grpc_addr.is_some() != register_req.listen_grpc_port.is_some() {
7762 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
7763 0 : "must specify both gRPC address and port"
7764 0 : )));
7765 0 : }
7766 :
7767 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
7768 : // This ensures that before we use it for anything or expose it via any external
7769 : // API, it is guaranteed to be available after a restart.
7770 0 : let new_node = Node::new(
7771 0 : register_req.node_id,
7772 0 : register_req.listen_http_addr,
7773 0 : register_req.listen_http_port,
7774 0 : register_req.listen_https_port,
7775 0 : register_req.listen_pg_addr,
7776 0 : register_req.listen_pg_port,
7777 0 : register_req.listen_grpc_addr,
7778 0 : register_req.listen_grpc_port,
7779 0 : register_req.availability_zone_id.clone(),
7780 0 : self.config.use_https_pageserver_api,
7781 : );
7782 0 : let new_node = match new_node {
7783 0 : Ok(new_node) => new_node,
7784 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
7785 : };
7786 :
7787 0 : match registration_status {
7788 : RegistrationStatus::New => {
7789 0 : self.persistence.insert_node(&new_node).await.map_err(|e| {
7790 0 : if matches!(
7791 0 : e,
7792 : crate::persistence::DatabaseError::Query(
7793 : diesel::result::Error::DatabaseError(
7794 : diesel::result::DatabaseErrorKind::UniqueViolation,
7795 : _,
7796 : )
7797 : )
7798 : ) {
7799 : // The node can be deleted by tombstone API, and not show up in the list of nodes.
7800 : // If you see this error, check tombstones first.
7801 0 : ApiError::Conflict(format!("Node {} is already exists", new_node.get_id()))
7802 : } else {
7803 0 : ApiError::from(e)
7804 : }
7805 0 : })?;
7806 : }
7807 : RegistrationStatus::NeedUpdate => {
7808 0 : self.persistence
7809 0 : .update_node_on_registration(
7810 0 : register_req.node_id,
7811 0 : register_req.listen_https_port,
7812 0 : )
7813 0 : .await?
7814 : }
7815 0 : _ => unreachable!("Other statuses have been processed earlier"),
7816 : }
7817 :
7818 0 : let mut locked = self.inner.write().unwrap();
7819 0 : let mut new_nodes = (*locked.nodes).clone();
7820 :
7821 0 : locked.scheduler.node_upsert(&new_node);
7822 0 : new_nodes.insert(register_req.node_id, new_node);
7823 :
7824 0 : locked.nodes = Arc::new(new_nodes);
7825 :
7826 0 : metrics::METRICS_REGISTRY
7827 0 : .metrics_group
7828 0 : .storage_controller_pageserver_nodes
7829 0 : .set(locked.nodes.len() as i64);
7830 0 : metrics::METRICS_REGISTRY
7831 0 : .metrics_group
7832 0 : .storage_controller_https_pageserver_nodes
7833 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7834 :
7835 0 : match registration_status {
7836 : RegistrationStatus::New => {
7837 0 : tracing::info!(
7838 0 : "Registered pageserver {} ({}), now have {} pageservers",
7839 : register_req.node_id,
7840 : register_req.availability_zone_id,
7841 0 : locked.nodes.len()
7842 : );
7843 : }
7844 : RegistrationStatus::NeedUpdate => {
7845 0 : tracing::info!(
7846 0 : "Re-registered and updated node {} ({})",
7847 : register_req.node_id,
7848 : register_req.availability_zone_id,
7849 : );
7850 : }
7851 0 : _ => unreachable!("Other statuses have been processed earlier"),
7852 : }
7853 0 : Ok(())
7854 0 : }
7855 :
7856 : /// Configure in-memory and persistent state of a node as requested
7857 : ///
7858 : /// Note that this function does not trigger any immediate side effects in response
7859 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
7860 0 : async fn node_state_configure(
7861 0 : &self,
7862 0 : node_id: NodeId,
7863 0 : availability: Option<NodeAvailability>,
7864 0 : scheduling: Option<NodeSchedulingPolicy>,
7865 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
7866 0 : ) -> Result<AvailabilityTransition, ApiError> {
7867 0 : if let Some(scheduling) = scheduling {
7868 : // Scheduling is a persistent part of Node: we must write updates to the database before
7869 : // applying them in memory
7870 0 : self.persistence
7871 0 : .update_node_scheduling_policy(node_id, scheduling)
7872 0 : .await?;
7873 0 : }
7874 :
7875 : // If we're activating a node, then before setting it active we must reconcile any shard locations
7876 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
7877 : // by calling [`Self::node_activate_reconcile`]
7878 : //
7879 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
7880 : // nothing else can mutate its availability while we run.
7881 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
7882 0 : let (activate_node, availability_transition) = {
7883 0 : let locked = self.inner.read().unwrap();
7884 0 : let Some(node) = locked.nodes.get(&node_id) else {
7885 0 : return Err(ApiError::NotFound(
7886 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7887 0 : ));
7888 : };
7889 :
7890 0 : (
7891 0 : node.clone(),
7892 0 : node.get_availability_transition(input_availability),
7893 0 : )
7894 : };
7895 :
7896 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
7897 0 : self.node_activate_reconcile(activate_node, node_lock)
7898 0 : .await?;
7899 0 : }
7900 0 : availability_transition
7901 : } else {
7902 0 : AvailabilityTransition::Unchanged
7903 : };
7904 :
7905 : // Apply changes from the request to our in-memory state for the Node
7906 0 : let mut locked = self.inner.write().unwrap();
7907 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
7908 :
7909 0 : let mut new_nodes = (**nodes).clone();
7910 :
7911 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
7912 0 : return Err(ApiError::NotFound(
7913 0 : anyhow::anyhow!("Node not registered").into(),
7914 0 : ));
7915 : };
7916 :
7917 0 : if let Some(availability) = availability {
7918 0 : node.set_availability(availability);
7919 0 : }
7920 :
7921 0 : if let Some(scheduling) = scheduling {
7922 0 : node.set_scheduling(scheduling);
7923 0 : }
7924 :
7925 : // Update the scheduler, in case the elegibility of the node for new shards has changed
7926 0 : scheduler.node_upsert(node);
7927 :
7928 0 : let new_nodes = Arc::new(new_nodes);
7929 0 : locked.nodes = new_nodes;
7930 :
7931 0 : Ok(availability_transition)
7932 0 : }
7933 :
7934 : /// Handle availability transition of one node
7935 : ///
7936 : /// Note that you should first call [`Self::node_state_configure`] to update
7937 : /// the in-memory state referencing that node. If you need to handle more than one transition
7938 : /// consider using [`Self::handle_node_availability_transitions`].
7939 0 : async fn handle_node_availability_transition(
7940 0 : &self,
7941 0 : node_id: NodeId,
7942 0 : transition: AvailabilityTransition,
7943 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
7944 0 : ) -> Result<(), ApiError> {
7945 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
7946 0 : match transition {
7947 : AvailabilityTransition::ToOffline => {
7948 0 : tracing::info!("Node {} transition to offline", node_id);
7949 :
7950 0 : let mut locked = self.inner.write().unwrap();
7951 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7952 :
7953 0 : let mut tenants_affected: usize = 0;
7954 :
7955 0 : for (_tenant_id, mut schedule_context, shards) in
7956 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7957 : {
7958 0 : for tenant_shard in shards {
7959 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
7960 0 : if let Some(observed_loc) =
7961 0 : tenant_shard.observed.locations.get_mut(&node_id)
7962 0 : {
7963 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
7964 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
7965 0 : observed_loc.conf = None;
7966 0 : }
7967 :
7968 0 : if nodes.len() == 1 {
7969 : // Special case for single-node cluster: there is no point trying to reschedule
7970 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
7971 : // failures to schedule them.
7972 0 : continue;
7973 0 : }
7974 :
7975 0 : if !nodes
7976 0 : .values()
7977 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
7978 : {
7979 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
7980 : // trying to reschedule since there's nowhere else to go. Without this
7981 : // branch we incorrectly detach tenants in response to node unavailability.
7982 0 : continue;
7983 0 : }
7984 :
7985 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
7986 0 : tenant_shard.sequence = tenant_shard.sequence.next();
7987 :
7988 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
7989 0 : Err(e) => {
7990 : // It is possible that some tenants will become unschedulable when too many pageservers
7991 : // go offline: in this case there isn't much we can do other than make the issue observable.
7992 : // TODO: give TenantShard a scheduling error attribute to be queried later.
7993 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
7994 : }
7995 : Ok(()) => {
7996 0 : if self
7997 0 : .maybe_reconcile_shard(
7998 0 : tenant_shard,
7999 0 : nodes,
8000 0 : ReconcilerPriority::Normal,
8001 0 : )
8002 0 : .is_some()
8003 0 : {
8004 0 : tenants_affected += 1;
8005 0 : };
8006 : }
8007 : }
8008 0 : }
8009 : }
8010 : }
8011 0 : tracing::info!(
8012 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
8013 : tenants_affected,
8014 : node_id
8015 : )
8016 : }
8017 : AvailabilityTransition::ToActive => {
8018 0 : tracing::info!("Node {} transition to active", node_id);
8019 :
8020 0 : let mut locked = self.inner.write().unwrap();
8021 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
8022 :
8023 : // When a node comes back online, we must reconcile any tenant that has a None observed
8024 : // location on the node.
8025 0 : for tenant_shard in tenants.values_mut() {
8026 : // If a reconciliation is already in progress, rely on the previous scheduling
8027 : // decision and skip triggering a new reconciliation.
8028 0 : if tenant_shard.reconciler.is_some() {
8029 0 : continue;
8030 0 : }
8031 :
8032 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
8033 0 : if observed_loc.conf.is_none() {
8034 0 : self.maybe_reconcile_shard(
8035 0 : tenant_shard,
8036 0 : nodes,
8037 0 : ReconcilerPriority::Normal,
8038 0 : );
8039 0 : }
8040 0 : }
8041 : }
8042 :
8043 : // TODO: in the background, we should balance work back onto this pageserver
8044 : }
8045 : // No action required for the intermediate unavailable state.
8046 : // When we transition into active or offline from the unavailable state,
8047 : // the correct handling above will kick in.
8048 : AvailabilityTransition::ToWarmingUpFromActive => {
8049 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
8050 : }
8051 : AvailabilityTransition::ToWarmingUpFromOffline => {
8052 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
8053 : }
8054 : AvailabilityTransition::Unchanged => {
8055 0 : tracing::debug!("Node {} no availability change during config", node_id);
8056 : }
8057 : }
8058 :
8059 0 : Ok(())
8060 0 : }
8061 :
8062 : /// Handle availability transition for multiple nodes
8063 : ///
8064 : /// Note that you should first call [`Self::node_state_configure`] for
8065 : /// all nodes being handled here for the handling to use fresh in-memory state.
8066 0 : async fn handle_node_availability_transitions(
8067 0 : &self,
8068 0 : transitions: Vec<(
8069 0 : NodeId,
8070 0 : TracingExclusiveGuard<NodeOperations>,
8071 0 : AvailabilityTransition,
8072 0 : )>,
8073 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
8074 0 : let mut errors = Vec::default();
8075 0 : for (node_id, node_lock, transition) in transitions {
8076 0 : let res = self
8077 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
8078 0 : .await;
8079 0 : if let Err(err) = res {
8080 0 : errors.push((node_id, err));
8081 0 : }
8082 : }
8083 :
8084 0 : if errors.is_empty() {
8085 0 : Ok(())
8086 : } else {
8087 0 : Err(errors)
8088 : }
8089 0 : }
8090 :
8091 0 : pub(crate) async fn node_configure(
8092 0 : &self,
8093 0 : node_id: NodeId,
8094 0 : availability: Option<NodeAvailability>,
8095 0 : scheduling: Option<NodeSchedulingPolicy>,
8096 0 : ) -> Result<(), ApiError> {
8097 0 : let node_lock =
8098 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
8099 :
8100 0 : let transition = self
8101 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
8102 0 : .await?;
8103 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
8104 0 : .await
8105 0 : }
8106 :
8107 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
8108 : /// operation for HTTP api.
8109 0 : pub(crate) async fn external_node_configure(
8110 0 : &self,
8111 0 : node_id: NodeId,
8112 0 : availability: Option<NodeAvailability>,
8113 0 : scheduling: Option<NodeSchedulingPolicy>,
8114 0 : ) -> Result<(), ApiError> {
8115 : {
8116 0 : let locked = self.inner.read().unwrap();
8117 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
8118 0 : return Err(ApiError::PreconditionFailed(
8119 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
8120 0 : ));
8121 0 : }
8122 : }
8123 :
8124 0 : self.node_configure(node_id, availability, scheduling).await
8125 0 : }
8126 :
8127 0 : pub(crate) async fn start_node_delete(
8128 0 : self: &Arc<Self>,
8129 0 : node_id: NodeId,
8130 0 : force: bool,
8131 0 : ) -> Result<(), ApiError> {
8132 0 : let (ongoing_op, node_policy, schedulable_nodes_count) = {
8133 0 : let locked = self.inner.read().unwrap();
8134 0 : let nodes = &locked.nodes;
8135 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8136 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8137 0 : ))?;
8138 0 : let schedulable_nodes_count = nodes
8139 0 : .iter()
8140 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8141 0 : .count();
8142 :
8143 : (
8144 0 : locked
8145 0 : .ongoing_operation
8146 0 : .as_ref()
8147 0 : .map(|ongoing| ongoing.operation),
8148 0 : node.get_scheduling(),
8149 0 : schedulable_nodes_count,
8150 : )
8151 : };
8152 :
8153 0 : if let Some(ongoing) = ongoing_op {
8154 0 : return Err(ApiError::PreconditionFailed(
8155 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8156 0 : ));
8157 0 : }
8158 :
8159 0 : if schedulable_nodes_count == 0 {
8160 0 : return Err(ApiError::PreconditionFailed(
8161 0 : "No other schedulable nodes to move shards".into(),
8162 0 : ));
8163 0 : }
8164 :
8165 0 : match node_policy {
8166 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
8167 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Deleting))
8168 0 : .await?;
8169 :
8170 0 : let cancel = self.cancel.child_token();
8171 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8172 0 : let policy_on_start = node_policy;
8173 :
8174 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8175 0 : operation: Operation::Delete(Delete { node_id }),
8176 0 : cancel: cancel.clone(),
8177 0 : });
8178 :
8179 0 : let span = tracing::info_span!(parent: None, "delete_node", %node_id);
8180 :
8181 0 : tokio::task::spawn(
8182 : {
8183 0 : let service = self.clone();
8184 0 : let cancel = cancel.clone();
8185 0 : async move {
8186 0 : let _gate_guard = gate_guard;
8187 :
8188 0 : scopeguard::defer! {
8189 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8190 :
8191 : if let Some(Operation::Delete(removed_delete)) = prev.map(|h| h.operation) {
8192 : assert_eq!(removed_delete.node_id, node_id, "We always take the same operation");
8193 : } else {
8194 : panic!("We always remove the same operation")
8195 : }
8196 : }
8197 :
8198 0 : tracing::info!("Delete background operation starting");
8199 0 : let res = service
8200 0 : .delete_node(node_id, policy_on_start, force, cancel)
8201 0 : .await;
8202 0 : match res {
8203 : Ok(()) => {
8204 0 : tracing::info!(
8205 0 : "Delete background operation completed successfully"
8206 : );
8207 : }
8208 : Err(OperationError::Cancelled) => {
8209 0 : tracing::info!("Delete background operation was cancelled");
8210 : }
8211 0 : Err(err) => {
8212 0 : tracing::error!(
8213 0 : "Delete background operation encountered: {err}"
8214 : )
8215 : }
8216 : }
8217 0 : }
8218 : }
8219 0 : .instrument(span),
8220 : );
8221 : }
8222 : NodeSchedulingPolicy::Deleting => {
8223 0 : return Err(ApiError::Conflict(format!(
8224 0 : "Node {node_id} has delete in progress"
8225 0 : )));
8226 : }
8227 0 : policy => {
8228 0 : return Err(ApiError::PreconditionFailed(
8229 0 : format!("Node {node_id} cannot be deleted due to {policy:?} policy").into(),
8230 0 : ));
8231 : }
8232 : }
8233 :
8234 0 : Ok(())
8235 0 : }
8236 :
8237 0 : pub(crate) async fn cancel_node_delete(
8238 0 : self: &Arc<Self>,
8239 0 : node_id: NodeId,
8240 0 : ) -> Result<(), ApiError> {
8241 : {
8242 0 : let locked = self.inner.read().unwrap();
8243 0 : let nodes = &locked.nodes;
8244 0 : nodes.get(&node_id).ok_or(ApiError::NotFound(
8245 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8246 0 : ))?;
8247 : }
8248 :
8249 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8250 0 : if let Operation::Delete(delete) = op_handler.operation {
8251 0 : if delete.node_id == node_id {
8252 0 : tracing::info!("Cancelling background delete operation for node {node_id}");
8253 0 : op_handler.cancel.cancel();
8254 0 : return Ok(());
8255 0 : }
8256 0 : }
8257 0 : }
8258 :
8259 0 : Err(ApiError::PreconditionFailed(
8260 0 : format!("Node {node_id} has no delete in progress").into(),
8261 0 : ))
8262 0 : }
8263 :
8264 0 : pub(crate) async fn start_node_drain(
8265 0 : self: &Arc<Self>,
8266 0 : node_id: NodeId,
8267 0 : ) -> Result<(), ApiError> {
8268 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
8269 0 : let locked = self.inner.read().unwrap();
8270 0 : let nodes = &locked.nodes;
8271 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8272 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8273 0 : ))?;
8274 0 : let schedulable_nodes_count = nodes
8275 0 : .iter()
8276 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8277 0 : .count();
8278 :
8279 : (
8280 0 : locked
8281 0 : .ongoing_operation
8282 0 : .as_ref()
8283 0 : .map(|ongoing| ongoing.operation),
8284 0 : node.is_available(),
8285 0 : node.get_scheduling(),
8286 0 : schedulable_nodes_count,
8287 : )
8288 : };
8289 :
8290 0 : if let Some(ongoing) = ongoing_op {
8291 0 : return Err(ApiError::PreconditionFailed(
8292 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8293 0 : ));
8294 0 : }
8295 :
8296 0 : if !node_available {
8297 0 : return Err(ApiError::ResourceUnavailable(
8298 0 : format!("Node {node_id} is currently unavailable").into(),
8299 0 : ));
8300 0 : }
8301 :
8302 0 : if schedulable_nodes_count == 0 {
8303 0 : return Err(ApiError::PreconditionFailed(
8304 0 : "No other schedulable nodes to drain to".into(),
8305 0 : ));
8306 0 : }
8307 :
8308 0 : match node_policy {
8309 : NodeSchedulingPolicy::Active => {
8310 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
8311 0 : .await?;
8312 :
8313 0 : let cancel = self.cancel.child_token();
8314 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8315 :
8316 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8317 0 : operation: Operation::Drain(Drain { node_id }),
8318 0 : cancel: cancel.clone(),
8319 0 : });
8320 :
8321 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
8322 :
8323 0 : tokio::task::spawn({
8324 0 : let service = self.clone();
8325 0 : let cancel = cancel.clone();
8326 0 : async move {
8327 0 : let _gate_guard = gate_guard;
8328 :
8329 0 : scopeguard::defer! {
8330 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8331 :
8332 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
8333 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
8334 : } else {
8335 : panic!("We always remove the same operation")
8336 : }
8337 : }
8338 :
8339 0 : tracing::info!("Drain background operation starting");
8340 0 : let res = service.drain_node(node_id, cancel).await;
8341 0 : match res {
8342 : Ok(()) => {
8343 0 : tracing::info!("Drain background operation completed successfully");
8344 : }
8345 : Err(OperationError::Cancelled) => {
8346 0 : tracing::info!("Drain background operation was cancelled");
8347 : }
8348 0 : Err(err) => {
8349 0 : tracing::error!("Drain background operation encountered: {err}")
8350 : }
8351 : }
8352 0 : }
8353 0 : }.instrument(span));
8354 : }
8355 : NodeSchedulingPolicy::Draining => {
8356 0 : return Err(ApiError::Conflict(format!(
8357 0 : "Node {node_id} has drain in progress"
8358 0 : )));
8359 : }
8360 0 : policy => {
8361 0 : return Err(ApiError::PreconditionFailed(
8362 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
8363 0 : ));
8364 : }
8365 : }
8366 :
8367 0 : Ok(())
8368 0 : }
8369 :
8370 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
8371 0 : let node_available = {
8372 0 : let locked = self.inner.read().unwrap();
8373 0 : let nodes = &locked.nodes;
8374 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8375 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8376 0 : ))?;
8377 :
8378 0 : node.is_available()
8379 : };
8380 :
8381 0 : if !node_available {
8382 0 : return Err(ApiError::ResourceUnavailable(
8383 0 : format!("Node {node_id} is currently unavailable").into(),
8384 0 : ));
8385 0 : }
8386 :
8387 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8388 0 : if let Operation::Drain(drain) = op_handler.operation {
8389 0 : if drain.node_id == node_id {
8390 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8391 0 : op_handler.cancel.cancel();
8392 0 : return Ok(());
8393 0 : }
8394 0 : }
8395 0 : }
8396 :
8397 0 : Err(ApiError::PreconditionFailed(
8398 0 : format!("Node {node_id} has no drain in progress").into(),
8399 0 : ))
8400 0 : }
8401 :
8402 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
8403 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
8404 0 : let locked = self.inner.read().unwrap();
8405 0 : let nodes = &locked.nodes;
8406 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8407 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8408 0 : ))?;
8409 :
8410 : (
8411 0 : locked
8412 0 : .ongoing_operation
8413 0 : .as_ref()
8414 0 : .map(|ongoing| ongoing.operation),
8415 0 : node.is_available(),
8416 0 : node.get_scheduling(),
8417 0 : nodes.len(),
8418 : )
8419 : };
8420 :
8421 0 : if let Some(ongoing) = ongoing_op {
8422 0 : return Err(ApiError::PreconditionFailed(
8423 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8424 0 : ));
8425 0 : }
8426 :
8427 0 : if !node_available {
8428 0 : return Err(ApiError::ResourceUnavailable(
8429 0 : format!("Node {node_id} is currently unavailable").into(),
8430 0 : ));
8431 0 : }
8432 :
8433 0 : if total_nodes_count <= 1 {
8434 0 : return Err(ApiError::PreconditionFailed(
8435 0 : "No other nodes to fill from".into(),
8436 0 : ));
8437 0 : }
8438 :
8439 0 : match node_policy {
8440 : NodeSchedulingPolicy::Active => {
8441 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
8442 0 : .await?;
8443 :
8444 0 : let cancel = self.cancel.child_token();
8445 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8446 :
8447 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8448 0 : operation: Operation::Fill(Fill { node_id }),
8449 0 : cancel: cancel.clone(),
8450 0 : });
8451 :
8452 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
8453 :
8454 0 : tokio::task::spawn({
8455 0 : let service = self.clone();
8456 0 : let cancel = cancel.clone();
8457 0 : async move {
8458 0 : let _gate_guard = gate_guard;
8459 :
8460 0 : scopeguard::defer! {
8461 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8462 :
8463 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
8464 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
8465 : } else {
8466 : panic!("We always remove the same operation")
8467 : }
8468 : }
8469 :
8470 0 : tracing::info!("Fill background operation starting");
8471 0 : let res = service.fill_node(node_id, cancel).await;
8472 0 : match res {
8473 : Ok(()) => {
8474 0 : tracing::info!("Fill background operation completed successfully");
8475 : }
8476 : Err(OperationError::Cancelled) => {
8477 0 : tracing::info!("Fill background operation was cancelled");
8478 : }
8479 0 : Err(err) => {
8480 0 : tracing::error!("Fill background operation encountered: {err}")
8481 : }
8482 : }
8483 0 : }
8484 0 : }.instrument(span));
8485 : }
8486 : NodeSchedulingPolicy::Filling => {
8487 0 : return Err(ApiError::Conflict(format!(
8488 0 : "Node {node_id} has fill in progress"
8489 0 : )));
8490 : }
8491 0 : policy => {
8492 0 : return Err(ApiError::PreconditionFailed(
8493 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
8494 0 : ));
8495 : }
8496 : }
8497 :
8498 0 : Ok(())
8499 0 : }
8500 :
8501 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
8502 0 : let node_available = {
8503 0 : let locked = self.inner.read().unwrap();
8504 0 : let nodes = &locked.nodes;
8505 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8506 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8507 0 : ))?;
8508 :
8509 0 : node.is_available()
8510 : };
8511 :
8512 0 : if !node_available {
8513 0 : return Err(ApiError::ResourceUnavailable(
8514 0 : format!("Node {node_id} is currently unavailable").into(),
8515 0 : ));
8516 0 : }
8517 :
8518 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8519 0 : if let Operation::Fill(fill) = op_handler.operation {
8520 0 : if fill.node_id == node_id {
8521 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8522 0 : op_handler.cancel.cancel();
8523 0 : return Ok(());
8524 0 : }
8525 0 : }
8526 0 : }
8527 :
8528 0 : Err(ApiError::PreconditionFailed(
8529 0 : format!("Node {node_id} has no fill in progress").into(),
8530 0 : ))
8531 0 : }
8532 :
8533 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
8534 : /// configuration
8535 0 : fn maybe_reconcile_shard(
8536 0 : &self,
8537 0 : shard: &mut TenantShard,
8538 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8539 0 : priority: ReconcilerPriority,
8540 0 : ) -> Option<ReconcilerWaiter> {
8541 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
8542 0 : }
8543 :
8544 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
8545 0 : fn get_reconciler_units(
8546 0 : &self,
8547 0 : priority: ReconcilerPriority,
8548 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
8549 0 : let units = match priority {
8550 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
8551 : ReconcilerPriority::High => {
8552 0 : match self
8553 0 : .priority_reconciler_concurrency
8554 0 : .clone()
8555 0 : .try_acquire_owned()
8556 : {
8557 0 : Ok(u) => Ok(u),
8558 : Err(TryAcquireError::NoPermits) => {
8559 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
8560 : // the normal priority semaphore.
8561 0 : self.reconciler_concurrency.clone().try_acquire_owned()
8562 : }
8563 0 : Err(e) => Err(e),
8564 : }
8565 : }
8566 : };
8567 :
8568 0 : units.map(ReconcileUnits::new)
8569 0 : }
8570 :
8571 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
8572 0 : fn maybe_configured_reconcile_shard(
8573 0 : &self,
8574 0 : shard: &mut TenantShard,
8575 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8576 0 : reconciler_config: ReconcilerConfig,
8577 0 : ) -> Option<ReconcilerWaiter> {
8578 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
8579 :
8580 0 : let reconcile_reason = match reconcile_needed {
8581 0 : ReconcileNeeded::No => return None,
8582 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
8583 0 : ReconcileNeeded::Yes(reason) => {
8584 : // Fall through to try and acquire units for spawning reconciler
8585 0 : reason
8586 : }
8587 : };
8588 :
8589 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
8590 0 : Ok(u) => u,
8591 : Err(_) => {
8592 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
8593 0 : "Concurrency limited: enqueued for reconcile later");
8594 0 : if !shard.delayed_reconcile {
8595 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
8596 0 : Err(TrySendError::Closed(_)) => {
8597 0 : // Weird mid-shutdown case?
8598 0 : }
8599 : Err(TrySendError::Full(_)) => {
8600 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
8601 0 : tracing::warn!(
8602 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
8603 : );
8604 : }
8605 0 : Ok(()) => {
8606 0 : shard.delayed_reconcile = true;
8607 0 : }
8608 : }
8609 0 : }
8610 :
8611 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
8612 : // number to advance. When this function is eventually called again and succeeds in getting units,
8613 : // it will spawn a reconciler that makes this waiter complete.
8614 0 : return Some(shard.future_reconcile_waiter());
8615 : }
8616 : };
8617 :
8618 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
8619 : // Gate closed: we're shutting down, drop out.
8620 0 : return None;
8621 : };
8622 :
8623 0 : shard.spawn_reconciler(
8624 0 : reconcile_reason,
8625 0 : &self.result_tx,
8626 0 : nodes,
8627 0 : &self.compute_hook,
8628 0 : reconciler_config,
8629 0 : &self.config,
8630 0 : &self.persistence,
8631 0 : units,
8632 0 : gate_guard,
8633 0 : &self.reconcilers_cancel,
8634 0 : self.http_client.clone(),
8635 : )
8636 0 : }
8637 :
8638 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
8639 : /// Additionally, reschedule tenants that require it.
8640 : ///
8641 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
8642 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
8643 : /// available. A return value of 0 indicates that everything is fully reconciled already.
8644 0 : fn reconcile_all(&self) -> ReconcileAllResult {
8645 0 : let mut locked = self.inner.write().unwrap();
8646 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8647 0 : let pageservers = nodes.clone();
8648 :
8649 : // This function is an efficient place to update lazy statistics, since we are walking
8650 : // all tenants.
8651 0 : let mut pending_reconciles = 0;
8652 0 : let mut stuck_reconciles = 0;
8653 0 : let mut az_violations = 0;
8654 :
8655 : // If we find any tenants to drop from memory, stash them to offload after
8656 : // we're done traversing the map of tenants.
8657 0 : let mut drop_detached_tenants = Vec::new();
8658 :
8659 0 : let mut spawned_reconciles = 0;
8660 0 : let mut has_delayed_reconciles = false;
8661 :
8662 0 : for shard in tenants.values_mut() {
8663 : // Accumulate scheduling statistics
8664 0 : if let (Some(attached), Some(preferred)) =
8665 0 : (shard.intent.get_attached(), shard.preferred_az())
8666 : {
8667 0 : let node_az = nodes
8668 0 : .get(attached)
8669 0 : .expect("Nodes exist if referenced")
8670 0 : .get_availability_zone_id();
8671 0 : if node_az != preferred {
8672 0 : az_violations += 1;
8673 0 : }
8674 0 : }
8675 :
8676 : // Skip checking if this shard is already enqueued for reconciliation
8677 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
8678 : // If there is something delayed, then return a nonzero count so that
8679 : // callers like reconcile_all_now do not incorrectly get the impression
8680 : // that the system is in a quiescent state.
8681 0 : has_delayed_reconciles = true;
8682 0 : pending_reconciles += 1;
8683 0 : continue;
8684 0 : }
8685 :
8686 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
8687 : // dirty, spawn another one
8688 0 : if self
8689 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
8690 0 : .is_some()
8691 : {
8692 0 : spawned_reconciles += 1;
8693 :
8694 0 : if shard.consecutive_reconciles_count >= MAX_CONSECUTIVE_RECONCILES {
8695 : // Count shards that are stuck, butwe still want to reconcile them.
8696 : // We don't want to consider them when deciding to run optimizations.
8697 0 : tracing::warn!(
8698 : tenant_id=%shard.tenant_shard_id.tenant_id,
8699 0 : shard_id=%shard.tenant_shard_id.shard_slug(),
8700 0 : "Shard reconciliation is stuck: {} consecutive launches",
8701 : shard.consecutive_reconciles_count
8702 : );
8703 0 : stuck_reconciles += 1;
8704 0 : }
8705 : } else {
8706 0 : if shard.delayed_reconcile {
8707 0 : // Shard wanted to reconcile but for some reason couldn't.
8708 0 : pending_reconciles += 1;
8709 0 : }
8710 :
8711 : // Reset the counter when we don't need to launch a reconcile.
8712 0 : shard.consecutive_reconciles_count = 0;
8713 : }
8714 : // If this tenant is detached, try dropping it from memory. This is usually done
8715 : // proactively in [`Self::process_results`], but we do it here to handle the edge
8716 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
8717 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
8718 0 : && shard.policy == PlacementPolicy::Detached
8719 : {
8720 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
8721 0 : shard.tenant_shard_id.tenant_id,
8722 0 : TenantOperations::DropDetached,
8723 0 : ) {
8724 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
8725 0 : }
8726 0 : }
8727 : }
8728 :
8729 : // Some metrics are calculated from SchedulerNode state, update these periodically
8730 0 : scheduler.update_metrics();
8731 :
8732 : // Process any deferred tenant drops
8733 0 : for (tenant_id, guard) in drop_detached_tenants {
8734 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
8735 0 : }
8736 :
8737 0 : metrics::METRICS_REGISTRY
8738 0 : .metrics_group
8739 0 : .storage_controller_schedule_az_violation
8740 0 : .set(az_violations as i64);
8741 :
8742 0 : metrics::METRICS_REGISTRY
8743 0 : .metrics_group
8744 0 : .storage_controller_pending_reconciles
8745 0 : .set(pending_reconciles as i64);
8746 :
8747 0 : metrics::METRICS_REGISTRY
8748 0 : .metrics_group
8749 0 : .storage_controller_stuck_reconciles
8750 0 : .set(stuck_reconciles as i64);
8751 :
8752 0 : ReconcileAllResult::new(spawned_reconciles, stuck_reconciles, has_delayed_reconciles)
8753 0 : }
8754 :
8755 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
8756 : /// could be scheduled somewhere better:
8757 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
8758 : /// * e.g. after a node fails then recovers, to move some work back to it
8759 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
8760 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
8761 : /// we did the split, but are probably better placed elsewhere.
8762 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
8763 : /// * e.g. after a shard split, some locations will be on the same node (where the split
8764 : /// happened), and will probably be better placed elsewhere.
8765 : ///
8766 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
8767 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
8768 : /// according to those same soft constraints.
8769 0 : async fn optimize_all(&self) -> usize {
8770 : // Limit on how many shards' optmizations each call to this function will execute. Combined
8771 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
8772 : // trickle of optimizations in the background, rather than executing a large number in parallel
8773 : // when a change occurs.
8774 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
8775 :
8776 : // Synchronous prepare: scan shards for possible scheduling optimizations
8777 0 : let candidate_work = self.optimize_all_plan();
8778 0 : let candidate_work_len = candidate_work.len();
8779 :
8780 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
8781 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
8782 :
8783 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
8784 :
8785 : // Synchronous apply: update the shards' intent states according to validated optimisations
8786 0 : let mut reconciles_spawned = 0;
8787 0 : let mut optimizations_applied = 0;
8788 0 : let mut locked = self.inner.write().unwrap();
8789 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8790 0 : for (tenant_shard_id, optimization) in validated_work {
8791 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
8792 : // Shard was dropped between planning and execution;
8793 0 : continue;
8794 : };
8795 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
8796 0 : if shard.apply_optimization(scheduler, optimization) {
8797 0 : optimizations_applied += 1;
8798 0 : if self
8799 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
8800 0 : .is_some()
8801 0 : {
8802 0 : reconciles_spawned += 1;
8803 0 : }
8804 0 : }
8805 :
8806 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
8807 0 : break;
8808 0 : }
8809 : }
8810 :
8811 0 : if was_work_filtered {
8812 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
8813 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
8814 0 : // as these validations start passing.
8815 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
8816 0 : }
8817 :
8818 0 : reconciles_spawned
8819 0 : }
8820 :
8821 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
8822 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
8823 : // this higher than the execution limit gives us a chance to execute some work even if the first
8824 : // few optimizations we find are not ready.
8825 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
8826 :
8827 0 : let mut work = Vec::new();
8828 0 : let mut locked = self.inner.write().unwrap();
8829 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
8830 :
8831 : // We are going to plan a bunch of optimisations before applying any of them, so the
8832 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
8833 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
8834 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
8835 : // to ignore the utilisation component of the score.
8836 :
8837 0 : for (_tenant_id, schedule_context, shards) in
8838 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Speculative)
8839 : {
8840 0 : for shard in shards {
8841 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
8842 0 : break;
8843 0 : }
8844 0 : match shard.get_scheduling_policy() {
8845 0 : ShardSchedulingPolicy::Active => {
8846 0 : // Ok to do optimization
8847 0 : }
8848 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
8849 0 : // Ok to do optimization: we are executing a graceful migration that
8850 0 : // has set preferred_node
8851 0 : }
8852 : ShardSchedulingPolicy::Essential
8853 : | ShardSchedulingPolicy::Pause
8854 : | ShardSchedulingPolicy::Stop => {
8855 : // Policy prevents optimizing this shard.
8856 0 : continue;
8857 : }
8858 : }
8859 :
8860 0 : if !matches!(shard.splitting, SplitState::Idle)
8861 0 : || matches!(shard.policy, PlacementPolicy::Detached)
8862 0 : || shard.reconciler.is_some()
8863 : {
8864 : // Do not start any optimizations while another change to the tenant is ongoing: this
8865 : // is not necessary for correctness, but simplifies operations and implicitly throttles
8866 : // optimization changes to happen in a "trickle" over time.
8867 0 : continue;
8868 0 : }
8869 :
8870 : // Fast path: we may quickly identify shards that don't have any possible optimisations
8871 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
8872 0 : if cfg!(feature = "testing") {
8873 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
8874 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
8875 : // panic in prod if we hit this, or spend cycles on it in prod.
8876 0 : assert!(
8877 0 : shard
8878 0 : .optimize_attachment(scheduler, &schedule_context)
8879 0 : .is_none()
8880 : );
8881 0 : assert!(
8882 0 : shard
8883 0 : .optimize_secondary(scheduler, &schedule_context)
8884 0 : .is_none()
8885 : );
8886 0 : }
8887 0 : continue;
8888 0 : }
8889 :
8890 0 : if let Some(optimization) =
8891 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
8892 : // its primary location based on soft constraints, cut it over.
8893 0 : shard.optimize_attachment(scheduler, &schedule_context)
8894 : {
8895 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
8896 0 : work.push((shard.tenant_shard_id, optimization));
8897 0 : break;
8898 0 : } else if let Some(optimization) =
8899 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
8900 : // better placed on another node, based on ScheduleContext, then adjust it. This
8901 : // covers cases like after a shard split, where we might have too many shards
8902 : // in the same tenant with secondary locations on the node where they originally split.
8903 0 : shard.optimize_secondary(scheduler, &schedule_context)
8904 : {
8905 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
8906 0 : work.push((shard.tenant_shard_id, optimization));
8907 0 : break;
8908 0 : }
8909 : }
8910 : }
8911 :
8912 0 : work
8913 0 : }
8914 :
8915 0 : async fn optimize_all_validate(
8916 0 : &self,
8917 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
8918 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
8919 : // Take a clone of the node map to use outside the lock in async validation phase
8920 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
8921 :
8922 0 : let mut want_secondary_status = Vec::new();
8923 :
8924 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
8925 : // check that the state of locations is acceptable to run the optimization, such as
8926 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
8927 : // in a live migration.
8928 0 : let mut validated_work = Vec::new();
8929 0 : for (tenant_shard_id, optimization) in candidate_work {
8930 0 : match optimization.action {
8931 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
8932 : old_attached_node_id: _,
8933 0 : new_attached_node_id,
8934 : }) => {
8935 0 : match validation_nodes.get(&new_attached_node_id) {
8936 0 : None => {
8937 0 : // Node was dropped between planning and validation
8938 0 : }
8939 0 : Some(node) => {
8940 0 : if !node.is_available() {
8941 0 : tracing::info!(
8942 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
8943 : );
8944 0 : } else {
8945 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
8946 0 : // remote API requests concurrently.
8947 0 : want_secondary_status.push((
8948 0 : tenant_shard_id,
8949 0 : node.clone(),
8950 0 : optimization,
8951 0 : ));
8952 0 : }
8953 : }
8954 : }
8955 : }
8956 : ScheduleOptimizationAction::ReplaceSecondary(_)
8957 : | ScheduleOptimizationAction::CreateSecondary(_)
8958 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
8959 : // No extra checks needed to manage secondaries: this does not interrupt client access
8960 0 : validated_work.push((tenant_shard_id, optimization))
8961 : }
8962 : };
8963 : }
8964 :
8965 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
8966 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
8967 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
8968 0 : let results = self
8969 0 : .tenant_for_shards_api(
8970 0 : want_secondary_status
8971 0 : .iter()
8972 0 : .map(|i| (i.0, i.1.clone()))
8973 0 : .collect(),
8974 0 : |tenant_shard_id, client| async move {
8975 0 : client.tenant_secondary_status(tenant_shard_id).await
8976 0 : },
8977 : 1,
8978 : 1,
8979 : SHORT_RECONCILE_TIMEOUT,
8980 0 : &self.cancel,
8981 : )
8982 0 : .await;
8983 :
8984 0 : for ((tenant_shard_id, node, optimization), (_, secondary_status)) in
8985 0 : want_secondary_status.into_iter().zip(results.into_iter())
8986 : {
8987 0 : match secondary_status {
8988 0 : Err(e) => {
8989 0 : tracing::info!(
8990 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
8991 : );
8992 : }
8993 0 : Ok(progress) => {
8994 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
8995 : // them in an optimization
8996 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
8997 :
8998 0 : if progress.heatmap_mtime.is_none()
8999 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
9000 0 : && progress.bytes_downloaded != progress.bytes_total
9001 0 : || progress.bytes_total - progress.bytes_downloaded
9002 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
9003 : {
9004 0 : tracing::info!(
9005 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
9006 : );
9007 :
9008 0 : if progress.heatmap_mtime.is_none() {
9009 : // No heatmap might mean the attached location has never uploaded one, or that
9010 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
9011 : // but fairly common in tests.
9012 0 : self.kick_secondary_download(tenant_shard_id).await;
9013 0 : }
9014 : } else {
9015 : // Location looks ready: proceed
9016 0 : tracing::info!(
9017 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
9018 : );
9019 0 : validated_work.push((tenant_shard_id, optimization))
9020 : }
9021 : }
9022 : }
9023 : }
9024 :
9025 0 : validated_work
9026 0 : }
9027 :
9028 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
9029 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
9030 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
9031 : /// we have this helper to move things along faster.
9032 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
9033 0 : if !self.config.kick_secondary_downloads {
9034 : // No-op if kick_secondary_downloads functionaliuty is not configured
9035 0 : return;
9036 0 : }
9037 :
9038 0 : let (attached_node, secondaries) = {
9039 0 : let locked = self.inner.read().unwrap();
9040 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
9041 0 : tracing::warn!(
9042 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
9043 : );
9044 0 : return;
9045 : };
9046 :
9047 0 : let Some(attached) = shard.intent.get_attached() else {
9048 0 : tracing::warn!(
9049 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
9050 : );
9051 0 : return;
9052 : };
9053 :
9054 0 : let secondaries = shard
9055 0 : .intent
9056 0 : .get_secondary()
9057 0 : .iter()
9058 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
9059 0 : .collect::<Vec<_>>();
9060 :
9061 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
9062 : };
9063 :
9064 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
9065 : // a 'kick' to let scheduling optimisation run more promptly.
9066 0 : match attached_node
9067 0 : .with_client_retries(
9068 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
9069 0 : &self.http_client,
9070 0 : &self.config.pageserver_jwt_token,
9071 : 3,
9072 : 10,
9073 : SHORT_RECONCILE_TIMEOUT,
9074 0 : &self.cancel,
9075 : )
9076 0 : .await
9077 : {
9078 0 : Some(Err(e)) => {
9079 0 : tracing::info!(
9080 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
9081 : );
9082 : }
9083 : None => {
9084 0 : tracing::info!(
9085 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
9086 : );
9087 : }
9088 : Some(Ok(_)) => {
9089 0 : tracing::info!(
9090 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
9091 : );
9092 : }
9093 : }
9094 :
9095 0 : for secondary_node in secondaries {
9096 0 : match secondary_node
9097 0 : .with_client_retries(
9098 0 : |client| async move {
9099 0 : client
9100 0 : .tenant_secondary_download(
9101 0 : tenant_shard_id,
9102 0 : Some(Duration::from_secs(1)),
9103 0 : )
9104 0 : .await
9105 0 : },
9106 0 : &self.http_client,
9107 0 : &self.config.pageserver_jwt_token,
9108 : 3,
9109 : 10,
9110 : SHORT_RECONCILE_TIMEOUT,
9111 0 : &self.cancel,
9112 : )
9113 0 : .await
9114 : {
9115 0 : Some(Err(e)) => {
9116 0 : tracing::info!(
9117 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
9118 : );
9119 : }
9120 : None => {
9121 0 : tracing::info!(
9122 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
9123 : );
9124 : }
9125 0 : Some(Ok(progress)) => {
9126 0 : tracing::info!(
9127 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
9128 : );
9129 : }
9130 : }
9131 : }
9132 0 : }
9133 :
9134 : /// Asynchronously split a tenant that's eligible for automatic splits. At most one tenant will
9135 : /// be split per call.
9136 : ///
9137 : /// Two sets of criteria are used: initial splits and size-based splits (in that order).
9138 : /// Initial splits are used to eagerly split unsharded tenants that may be performing initial
9139 : /// ingestion, since sharded tenants have significantly better ingestion throughput. Size-based
9140 : /// splits are used to bound the maximum shard size and balance out load.
9141 : ///
9142 : /// Splits are based on max_logical_size, i.e. the logical size of the largest timeline in a
9143 : /// tenant. We use this instead of the total logical size because branches will duplicate
9144 : /// logical size without actually using more storage. We could also use visible physical size,
9145 : /// but this might overestimate tenants that frequently churn branches.
9146 : ///
9147 : /// Initial splits (initial_split_threshold):
9148 : /// * Applies to tenants with 1 shard.
9149 : /// * The largest timeline (max_logical_size) exceeds initial_split_threshold.
9150 : /// * Splits into initial_split_shards.
9151 : ///
9152 : /// Size-based splits (split_threshold):
9153 : /// * Applies to all tenants.
9154 : /// * The largest timeline (max_logical_size) divided by shard count exceeds split_threshold.
9155 : /// * Splits such that max_logical_size / shard_count <= split_threshold, in powers of 2.
9156 : ///
9157 : /// Tenant shards are ordered by descending max_logical_size, first initial split candidates
9158 : /// then size-based split candidates. The first matching candidate is split.
9159 : ///
9160 : /// The shard count is clamped to max_split_shards. If a candidate is eligible for both initial
9161 : /// and size-based splits, the largest shard count will be used.
9162 : ///
9163 : /// An unsharded tenant will get DEFAULT_STRIPE_SIZE, regardless of what its ShardIdentity says.
9164 : /// A sharded tenant will retain its stripe size, as splits do not allow changing it.
9165 : ///
9166 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
9167 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
9168 : /// block all other splits.
9169 0 : async fn autosplit_tenants(self: &Arc<Self>) {
9170 : // If max_split_shards is set to 0 or 1, we can't split.
9171 0 : let max_split_shards = self.config.max_split_shards;
9172 0 : if max_split_shards <= 1 {
9173 0 : return;
9174 0 : }
9175 :
9176 : // If initial_split_shards is set to 0 or 1, disable initial splits.
9177 0 : let mut initial_split_threshold = self.config.initial_split_threshold.unwrap_or(0);
9178 0 : let initial_split_shards = self.config.initial_split_shards;
9179 0 : if initial_split_shards <= 1 {
9180 0 : initial_split_threshold = 0;
9181 0 : }
9182 :
9183 : // If no split_threshold nor initial_split_threshold, disable autosplits.
9184 0 : let split_threshold = self.config.split_threshold.unwrap_or(0);
9185 0 : if split_threshold == 0 && initial_split_threshold == 0 {
9186 0 : return;
9187 0 : }
9188 :
9189 : // Fetch split candidates in prioritized order.
9190 : //
9191 : // If initial splits are enabled, fetch eligible tenants first. We prioritize initial splits
9192 : // over size-based splits, since these are often performing initial ingestion and rely on
9193 : // splits to improve ingest throughput.
9194 0 : let mut candidates = Vec::new();
9195 :
9196 0 : if initial_split_threshold > 0 {
9197 : // Initial splits: fetch tenants with 1 shard where the logical size of the largest
9198 : // timeline exceeds the initial split threshold.
9199 0 : let initial_candidates = self
9200 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9201 0 : order_by: TenantSorting::MaxLogicalSize,
9202 0 : limit: 10,
9203 0 : where_shards_lt: Some(ShardCount(2)),
9204 0 : where_gt: Some(initial_split_threshold),
9205 0 : })
9206 0 : .await;
9207 0 : candidates.extend(initial_candidates);
9208 0 : }
9209 :
9210 0 : if split_threshold > 0 {
9211 : // Size-based splits: fetch tenants where the logical size of the largest timeline
9212 : // divided by shard count exceeds the split threshold.
9213 : //
9214 : // max_logical_size is only tracked on shard 0, and contains the total logical size
9215 : // across all shards. We have to order and filter by MaxLogicalSizePerShard, i.e.
9216 : // max_logical_size / shard_count, such that we only receive tenants that are actually
9217 : // eligible for splits. But we still use max_logical_size for later split calculations.
9218 0 : let size_candidates = self
9219 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9220 0 : order_by: TenantSorting::MaxLogicalSizePerShard,
9221 0 : limit: 10,
9222 0 : where_shards_lt: Some(ShardCount(max_split_shards)),
9223 0 : where_gt: Some(split_threshold),
9224 0 : })
9225 0 : .await;
9226 : #[cfg(feature = "testing")]
9227 0 : assert!(
9228 0 : size_candidates.iter().all(|c| c.id.is_shard_zero()),
9229 0 : "MaxLogicalSizePerShard returned non-zero shard: {size_candidates:?}",
9230 : );
9231 0 : candidates.extend(size_candidates);
9232 0 : }
9233 :
9234 : // Filter out tenants in a prohibiting scheduling modes
9235 : // and tenants with an ongoing import.
9236 : //
9237 : // Note that the import check here is oportunistic. An import might start
9238 : // after the check before we actually update [`TenantShard::splitting`].
9239 : // [`Self::tenant_shard_split`] checks the database whilst holding the exclusive
9240 : // tenant lock. Imports might take a long time, so the check here allows us
9241 : // to split something else instead of trying the same shard over and over.
9242 : {
9243 0 : let state = self.inner.read().unwrap();
9244 0 : candidates.retain(|i| {
9245 0 : let shard = state.tenants.get(&i.id);
9246 0 : match shard {
9247 0 : Some(t) => {
9248 0 : t.get_scheduling_policy() == ShardSchedulingPolicy::Active
9249 0 : && t.importing == TimelineImportState::Idle
9250 : }
9251 0 : None => false,
9252 : }
9253 0 : });
9254 : }
9255 :
9256 : // Pick the first candidate to split. This will generally always be the first one in
9257 : // candidates, but we defensively skip candidates that end up not actually splitting.
9258 0 : let Some((candidate, new_shard_count)) = candidates
9259 0 : .into_iter()
9260 0 : .filter_map(|candidate| {
9261 0 : let new_shard_count = Self::compute_split_shards(ShardSplitInputs {
9262 0 : shard_count: candidate.id.shard_count,
9263 0 : max_logical_size: candidate.max_logical_size,
9264 0 : split_threshold,
9265 0 : max_split_shards,
9266 0 : initial_split_threshold,
9267 0 : initial_split_shards,
9268 0 : });
9269 0 : new_shard_count.map(|shards| (candidate, shards.count()))
9270 0 : })
9271 0 : .next()
9272 : else {
9273 0 : debug!("no split-eligible tenants found");
9274 0 : return;
9275 : };
9276 :
9277 : // Retain the stripe size of sharded tenants, as splits don't allow changing it. Otherwise,
9278 : // use DEFAULT_STRIPE_SIZE for unsharded tenants -- their stripe size doesn't really matter,
9279 : // and if we change the default stripe size we want to use the new default rather than an
9280 : // old, persisted stripe size.
9281 0 : let new_stripe_size = match candidate.id.shard_count.count() {
9282 0 : 0 => panic!("invalid shard count 0"),
9283 0 : 1 => Some(DEFAULT_STRIPE_SIZE),
9284 0 : 2.. => None,
9285 : };
9286 :
9287 : // We spawn a task to run this, so it's exactly like some external API client requesting
9288 : // it. We don't want to block the background reconcile loop on this.
9289 0 : let old_shard_count = candidate.id.shard_count.count();
9290 0 : info!(
9291 0 : "auto-splitting tenant {old_shard_count} → {new_shard_count} shards, \
9292 0 : current size {candidate:?} (split_threshold={split_threshold} \
9293 0 : initial_split_threshold={initial_split_threshold})"
9294 : );
9295 :
9296 0 : let this = self.clone();
9297 0 : tokio::spawn(
9298 0 : async move {
9299 0 : match this
9300 0 : .tenant_shard_split(
9301 0 : candidate.id.tenant_id,
9302 0 : TenantShardSplitRequest {
9303 0 : new_shard_count,
9304 0 : new_stripe_size,
9305 0 : },
9306 0 : )
9307 0 : .await
9308 : {
9309 : Ok(_) => {
9310 0 : info!("successful auto-split {old_shard_count} → {new_shard_count} shards")
9311 : }
9312 0 : Err(err) => error!("auto-split failed: {err}"),
9313 : }
9314 0 : }
9315 0 : .instrument(info_span!("auto_split", tenant_id=%candidate.id.tenant_id)),
9316 : );
9317 0 : }
9318 :
9319 : /// Returns the number of shards to split a tenant into, or None if the tenant shouldn't split,
9320 : /// based on the total logical size of the largest timeline summed across all shards. Uses the
9321 : /// larger of size-based and initial splits, clamped to max_split_shards.
9322 : ///
9323 : /// NB: the thresholds are exclusive, since TopTenantShardsRequest uses where_gt.
9324 25 : fn compute_split_shards(inputs: ShardSplitInputs) -> Option<ShardCount> {
9325 : let ShardSplitInputs {
9326 25 : shard_count,
9327 25 : max_logical_size,
9328 25 : split_threshold,
9329 25 : max_split_shards,
9330 25 : initial_split_threshold,
9331 25 : initial_split_shards,
9332 25 : } = inputs;
9333 :
9334 25 : let mut new_shard_count: u8 = shard_count.count();
9335 :
9336 : // Size-based splits. Ensures max_logical_size / new_shard_count <= split_threshold, using
9337 : // power-of-two shard counts.
9338 : //
9339 : // If the current shard count is not a power of two, and does not exceed split_threshold,
9340 : // then we leave it alone rather than forcing a power-of-two split.
9341 25 : if split_threshold > 0
9342 18 : && max_logical_size.div_ceil(split_threshold) > shard_count.count() as u64
9343 12 : {
9344 12 : new_shard_count = max_logical_size
9345 12 : .div_ceil(split_threshold)
9346 12 : .checked_next_power_of_two()
9347 12 : .unwrap_or(u8::MAX as u64)
9348 12 : .try_into()
9349 12 : .unwrap_or(u8::MAX);
9350 13 : }
9351 :
9352 : // Initial splits. Use the larger of size-based and initial split shard counts. This only
9353 : // applies to unsharded tenants, i.e. changes to initial_split_threshold or
9354 : // initial_split_shards are not retroactive for sharded tenants.
9355 25 : if initial_split_threshold > 0
9356 14 : && shard_count.count() <= 1
9357 11 : && max_logical_size > initial_split_threshold
9358 8 : {
9359 8 : new_shard_count = new_shard_count.max(initial_split_shards);
9360 17 : }
9361 :
9362 : // Clamp to max shards.
9363 25 : new_shard_count = new_shard_count.min(max_split_shards);
9364 :
9365 : // Don't split if we're not increasing the shard count.
9366 25 : if new_shard_count <= shard_count.count() {
9367 10 : return None;
9368 15 : }
9369 :
9370 15 : Some(ShardCount(new_shard_count))
9371 25 : }
9372 :
9373 : /// Fetches the top tenant shards from every available node, in descending order of
9374 : /// max logical size. Offline nodes are skipped, and any errors from available nodes
9375 : /// will be logged and ignored.
9376 0 : async fn get_top_tenant_shards(
9377 0 : &self,
9378 0 : request: &TopTenantShardsRequest,
9379 0 : ) -> Vec<TopTenantShardItem> {
9380 0 : let nodes = self
9381 0 : .inner
9382 0 : .read()
9383 0 : .unwrap()
9384 0 : .nodes
9385 0 : .values()
9386 0 : .filter(|node| node.is_available())
9387 0 : .cloned()
9388 0 : .collect_vec();
9389 :
9390 0 : let mut futures = FuturesUnordered::new();
9391 0 : for node in nodes {
9392 0 : futures.push(async move {
9393 0 : node.with_client_retries(
9394 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
9395 0 : &self.http_client,
9396 0 : &self.config.pageserver_jwt_token,
9397 : 3,
9398 : 3,
9399 0 : Duration::from_secs(5),
9400 0 : &self.cancel,
9401 : )
9402 0 : .await
9403 0 : });
9404 : }
9405 :
9406 0 : let mut top = Vec::new();
9407 0 : while let Some(output) = futures.next().await {
9408 0 : match output {
9409 0 : Some(Ok(response)) => top.extend(response.shards),
9410 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
9411 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
9412 0 : None => {} // node is shutting down
9413 : }
9414 : }
9415 :
9416 0 : top.sort_by_key(|i| i.max_logical_size);
9417 0 : top.reverse();
9418 0 : top
9419 0 : }
9420 :
9421 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
9422 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
9423 : /// put the system into a quiescent state where future background reconciliations won't do anything.
9424 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
9425 0 : let reconcile_all_result = self.reconcile_all();
9426 0 : let mut spawned_reconciles = reconcile_all_result.spawned_reconciles;
9427 0 : if reconcile_all_result.can_run_optimizations() {
9428 : // Only optimize when we are otherwise idle
9429 0 : let optimization_reconciles = self.optimize_all().await;
9430 0 : spawned_reconciles += optimization_reconciles;
9431 0 : }
9432 :
9433 0 : let waiters = {
9434 0 : let mut waiters = Vec::new();
9435 0 : let locked = self.inner.read().unwrap();
9436 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
9437 0 : if let Some(waiter) = shard.get_waiter() {
9438 0 : waiters.push(waiter);
9439 0 : }
9440 : }
9441 0 : waiters
9442 : };
9443 :
9444 0 : let waiter_count = waiters.len();
9445 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
9446 0 : Ok(()) => {}
9447 0 : Err(e) => {
9448 0 : if let ReconcileWaitError::Failed(_, reconcile_error) = &e {
9449 0 : match **reconcile_error {
9450 : ReconcileError::Cancel
9451 0 : | ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
9452 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
9453 0 : // because some other change superceded it. We will return a nonzero number,
9454 0 : // so the caller knows they might have to call again to quiesce the system.
9455 0 : }
9456 : _ => {
9457 0 : return Err(e);
9458 : }
9459 : }
9460 : } else {
9461 0 : return Err(e);
9462 : }
9463 : }
9464 : };
9465 :
9466 0 : tracing::info!(
9467 0 : "{} reconciles in reconcile_all, {} waiters",
9468 : spawned_reconciles,
9469 : waiter_count
9470 : );
9471 :
9472 0 : Ok(std::cmp::max(waiter_count, spawned_reconciles))
9473 0 : }
9474 :
9475 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
9476 : // Cancel all on-going reconciles and wait for them to exit the gate.
9477 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
9478 0 : self.reconcilers_cancel.cancel();
9479 0 : self.reconcilers_gate.close().await;
9480 :
9481 : // Signal the background loop in [`Service::process_results`] to exit once
9482 : // it has proccessed the results from all the reconciles we cancelled earlier.
9483 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
9484 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
9485 0 : self.result_tx.closed().await;
9486 0 : }
9487 :
9488 0 : pub async fn shutdown(&self) {
9489 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
9490 0 : .await;
9491 :
9492 : // Background tasks hold gate guards: this notifies them of the cancellation and
9493 : // waits for them all to complete.
9494 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
9495 0 : self.cancel.cancel();
9496 0 : self.gate.close().await;
9497 0 : }
9498 :
9499 : /// Spot check the download lag for a secondary location of a shard.
9500 : /// Should be used as a heuristic, since it's not always precise: the
9501 : /// secondary might have not downloaded the new heat map yet and, hence,
9502 : /// is not aware of the lag.
9503 : ///
9504 : /// Returns:
9505 : /// * Ok(None) if the lag could not be determined from the status,
9506 : /// * Ok(Some(_)) if the lag could be determind
9507 : /// * Err on failures to query the pageserver.
9508 0 : async fn secondary_lag(
9509 0 : &self,
9510 0 : secondary: &NodeId,
9511 0 : tenant_shard_id: TenantShardId,
9512 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
9513 0 : let nodes = self.inner.read().unwrap().nodes.clone();
9514 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
9515 0 : StatusCode::NOT_FOUND,
9516 0 : format!("Node with id {secondary} not found"),
9517 0 : ))?;
9518 :
9519 0 : match node
9520 0 : .with_client_retries(
9521 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
9522 0 : &self.http_client,
9523 0 : &self.config.pageserver_jwt_token,
9524 : 1,
9525 : 3,
9526 0 : Duration::from_millis(250),
9527 0 : &self.cancel,
9528 : )
9529 0 : .await
9530 : {
9531 0 : Some(Ok(status)) => match status.heatmap_mtime {
9532 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
9533 0 : None => Ok(None),
9534 : },
9535 0 : Some(Err(e)) => Err(e),
9536 0 : None => Err(mgmt_api::Error::Cancelled),
9537 : }
9538 0 : }
9539 :
9540 : /// Drain a node by moving the shards attached to it as primaries.
9541 : /// This is a long running operation and it should run as a separate Tokio task.
9542 0 : pub(crate) async fn drain_node(
9543 0 : self: &Arc<Self>,
9544 0 : node_id: NodeId,
9545 0 : cancel: CancellationToken,
9546 0 : ) -> Result<(), OperationError> {
9547 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
9548 0 : let max_secondary_lag_bytes = self
9549 0 : .config
9550 0 : .max_secondary_lag_bytes
9551 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
9552 :
9553 : // By default, live migrations are generous about the wait time for getting
9554 : // the secondary location up to speed. When draining, give up earlier in order
9555 : // to not stall the operation when a cold secondary is encountered.
9556 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9557 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9558 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9559 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9560 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9561 0 : .build();
9562 :
9563 0 : let mut waiters = Vec::new();
9564 :
9565 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
9566 :
9567 0 : while !tid_iter.finished() {
9568 0 : if cancel.is_cancelled() {
9569 0 : match self
9570 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9571 0 : .await
9572 : {
9573 0 : Ok(()) => return Err(OperationError::Cancelled),
9574 0 : Err(err) => {
9575 0 : return Err(OperationError::FinalizeError(
9576 0 : format!(
9577 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9578 0 : )
9579 0 : .into(),
9580 0 : ));
9581 : }
9582 : }
9583 0 : }
9584 :
9585 0 : operation_utils::validate_node_state(
9586 0 : &node_id,
9587 0 : self.inner.read().unwrap().nodes.clone(),
9588 0 : NodeSchedulingPolicy::Draining,
9589 0 : )?;
9590 :
9591 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9592 0 : let tid = match tid_iter.next() {
9593 0 : Some(tid) => tid,
9594 : None => {
9595 0 : break;
9596 : }
9597 : };
9598 :
9599 0 : let tid_drain = TenantShardDrain {
9600 0 : drained_node: node_id,
9601 0 : tenant_shard_id: tid,
9602 0 : };
9603 :
9604 0 : let dest_node_id = {
9605 0 : let locked = self.inner.read().unwrap();
9606 :
9607 0 : match tid_drain
9608 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
9609 : {
9610 0 : Some(node_id) => node_id,
9611 : None => {
9612 0 : continue;
9613 : }
9614 : }
9615 : };
9616 :
9617 0 : match self.secondary_lag(&dest_node_id, tid).await {
9618 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
9619 0 : // The secondary is reasonably up to date.
9620 0 : // Migrate to it
9621 0 : }
9622 0 : Ok(Some(lag)) => {
9623 0 : tracing::info!(
9624 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9625 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
9626 : );
9627 0 : continue;
9628 : }
9629 : Ok(None) => {
9630 0 : tracing::info!(
9631 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9632 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
9633 : );
9634 0 : continue;
9635 : }
9636 0 : Err(err) => {
9637 0 : tracing::warn!(
9638 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9639 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
9640 : );
9641 0 : continue;
9642 : }
9643 : }
9644 :
9645 : {
9646 0 : let mut locked = self.inner.write().unwrap();
9647 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9648 0 : let rescheduled = tid_drain.reschedule_to_secondary(
9649 0 : dest_node_id,
9650 0 : tenants,
9651 0 : scheduler,
9652 0 : nodes,
9653 0 : )?;
9654 :
9655 0 : if let Some(tenant_shard) = rescheduled {
9656 0 : let waiter = self.maybe_configured_reconcile_shard(
9657 0 : tenant_shard,
9658 0 : nodes,
9659 0 : reconciler_config,
9660 0 : );
9661 0 : if let Some(some) = waiter {
9662 0 : waiters.push(some);
9663 0 : }
9664 0 : }
9665 : }
9666 : }
9667 :
9668 0 : waiters = self
9669 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9670 0 : .await;
9671 :
9672 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
9673 : }
9674 :
9675 0 : while !waiters.is_empty() {
9676 0 : if cancel.is_cancelled() {
9677 0 : match self
9678 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9679 0 : .await
9680 : {
9681 0 : Ok(()) => return Err(OperationError::Cancelled),
9682 0 : Err(err) => {
9683 0 : return Err(OperationError::FinalizeError(
9684 0 : format!(
9685 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9686 0 : )
9687 0 : .into(),
9688 0 : ));
9689 : }
9690 : }
9691 0 : }
9692 :
9693 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
9694 :
9695 0 : waiters = self
9696 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9697 0 : .await;
9698 : }
9699 :
9700 : // At this point we have done the best we could to drain shards from this node.
9701 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
9702 : // to complete the drain.
9703 0 : if let Err(err) = self
9704 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
9705 0 : .await
9706 : {
9707 : // This is not fatal. Anything that is polling the node scheduling policy to detect
9708 : // the end of the drain operations will hang, but all such places should enforce an
9709 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
9710 : // by the counterpart fill operation.
9711 0 : return Err(OperationError::FinalizeError(
9712 0 : format!(
9713 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
9714 0 : )
9715 0 : .into(),
9716 0 : ));
9717 0 : }
9718 :
9719 0 : Ok(())
9720 0 : }
9721 :
9722 : /// Create a node fill plan (pick secondaries to promote), based on:
9723 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
9724 : /// outside their home AZ, should be migrated back here.
9725 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
9726 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
9727 : /// those shards have a home AZ that doesn't match the node we're filling.
9728 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
9729 0 : let mut locked = self.inner.write().unwrap();
9730 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
9731 :
9732 0 : let node_az = nodes
9733 0 : .get(&node_id)
9734 0 : .expect("Node must exist")
9735 0 : .get_availability_zone_id()
9736 0 : .clone();
9737 :
9738 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
9739 0 : let mut plan = Vec::new();
9740 :
9741 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
9742 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
9743 :
9744 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
9745 : // conceivably come up in real life if deploying a single-AZ region intentionally.
9746 0 : let respect_azs = nodes
9747 0 : .values()
9748 0 : .map(|n| n.get_availability_zone_id())
9749 0 : .unique()
9750 0 : .count()
9751 : > 1;
9752 :
9753 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
9754 : // requires it.
9755 0 : for (tsid, tenant_shard) in tenants {
9756 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9757 : // Shard doesn't have a secondary on this node, ignore it.
9758 0 : continue;
9759 0 : }
9760 :
9761 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
9762 : // shards which belong on this node, not to promote shards whose scheduling preference
9763 : // would be on their currently attached node. So will avoid promoting shards whose
9764 : // home AZ doesn't match the AZ of the node we're filling.
9765 0 : match tenant_shard.preferred_az() {
9766 0 : _ if !respect_azs => {
9767 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9768 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9769 0 : }
9770 : }
9771 : None => {
9772 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
9773 : // will only do so if our target shard count requires it.
9774 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9775 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9776 0 : }
9777 : }
9778 0 : Some(az) if az == &node_az => {
9779 : // This shard's home AZ is equal to the node we're filling: it should
9780 : // be moved back to this node as part of filling, unless its currently
9781 : // attached location is also in its home AZ.
9782 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9783 0 : if nodes
9784 0 : .get(primary)
9785 0 : .expect("referenced node must exist")
9786 0 : .get_availability_zone_id()
9787 0 : != tenant_shard
9788 0 : .preferred_az()
9789 0 : .expect("tenant must have an AZ preference")
9790 : {
9791 0 : plan.push(*tsid)
9792 0 : }
9793 : } else {
9794 0 : plan.push(*tsid)
9795 : }
9796 : }
9797 0 : Some(_) => {
9798 0 : // This shard's home AZ is somewhere other than the node we're filling,
9799 0 : // it may not be moved back to this node as part of filling. Ignore it
9800 0 : }
9801 : }
9802 : }
9803 :
9804 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
9805 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
9806 :
9807 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
9808 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
9809 :
9810 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
9811 :
9812 0 : for (node_id, attached) in nodes_by_load {
9813 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
9814 0 : if !available {
9815 0 : continue;
9816 0 : }
9817 :
9818 0 : if plan.len() >= fill_requirement
9819 0 : || free_tids_by_node.is_empty()
9820 0 : || attached <= expected_attached
9821 : {
9822 0 : break;
9823 0 : }
9824 :
9825 0 : let can_take = attached - expected_attached;
9826 0 : let needed = fill_requirement - plan.len();
9827 0 : let mut take = std::cmp::min(can_take, needed);
9828 :
9829 0 : let mut remove_node = false;
9830 0 : while take > 0 {
9831 0 : match free_tids_by_node.get_mut(&node_id) {
9832 0 : Some(tids) => match tids.pop() {
9833 0 : Some(tid) => {
9834 0 : let max_promote_for_tenant = std::cmp::max(
9835 0 : tid.shard_count.count() as usize / locked.nodes.len(),
9836 : 1,
9837 : );
9838 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
9839 0 : if *promoted < max_promote_for_tenant {
9840 0 : plan.push(tid);
9841 0 : *promoted += 1;
9842 0 : take -= 1;
9843 0 : }
9844 : }
9845 : None => {
9846 0 : remove_node = true;
9847 0 : break;
9848 : }
9849 : },
9850 : None => {
9851 0 : break;
9852 : }
9853 : }
9854 : }
9855 :
9856 0 : if remove_node {
9857 0 : free_tids_by_node.remove(&node_id);
9858 0 : }
9859 : }
9860 :
9861 0 : plan
9862 0 : }
9863 :
9864 : /// Fill a node by promoting its secondaries until the cluster is balanced
9865 : /// with regards to attached shard counts. Note that this operation only
9866 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
9867 : /// This is a long running operation and it should run as a separate Tokio task.
9868 0 : pub(crate) async fn fill_node(
9869 0 : &self,
9870 0 : node_id: NodeId,
9871 0 : cancel: CancellationToken,
9872 0 : ) -> Result<(), OperationError> {
9873 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9874 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9875 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9876 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9877 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9878 0 : .build();
9879 :
9880 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
9881 0 : let mut waiters = Vec::new();
9882 :
9883 : // Execute the plan we've composed above. Before aplying each move from the plan,
9884 : // we validate to ensure that it has not gone stale in the meantime.
9885 0 : while !tids_to_promote.is_empty() {
9886 0 : if cancel.is_cancelled() {
9887 0 : match self
9888 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9889 0 : .await
9890 : {
9891 0 : Ok(()) => return Err(OperationError::Cancelled),
9892 0 : Err(err) => {
9893 0 : return Err(OperationError::FinalizeError(
9894 0 : format!(
9895 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9896 0 : )
9897 0 : .into(),
9898 0 : ));
9899 : }
9900 : }
9901 0 : }
9902 :
9903 : {
9904 0 : let mut locked = self.inner.write().unwrap();
9905 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9906 :
9907 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
9908 0 : format!("node {node_id} was removed").into(),
9909 0 : ))?;
9910 :
9911 0 : let current_policy = node.get_scheduling();
9912 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
9913 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
9914 : // about it
9915 0 : return Err(OperationError::NodeStateChanged(
9916 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
9917 0 : ));
9918 0 : }
9919 :
9920 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9921 0 : if let Some(tid) = tids_to_promote.pop() {
9922 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
9923 : // If the node being filled is not a secondary anymore,
9924 : // skip the promotion.
9925 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9926 0 : continue;
9927 0 : }
9928 :
9929 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
9930 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
9931 0 : Err(e) => {
9932 0 : tracing::warn!(
9933 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9934 0 : "Scheduling error when filling pageserver {} : {e}", node_id
9935 : );
9936 : }
9937 : Ok(()) => {
9938 0 : tracing::info!(
9939 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9940 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
9941 : node_id,
9942 : previously_attached_to,
9943 : node_id
9944 : );
9945 :
9946 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
9947 0 : tenant_shard,
9948 0 : nodes,
9949 0 : reconciler_config,
9950 0 : ) {
9951 0 : waiters.push(waiter);
9952 0 : }
9953 : }
9954 : }
9955 0 : }
9956 : } else {
9957 0 : break;
9958 : }
9959 : }
9960 : }
9961 :
9962 0 : waiters = self
9963 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9964 0 : .await;
9965 : }
9966 :
9967 0 : while !waiters.is_empty() {
9968 0 : if cancel.is_cancelled() {
9969 0 : match self
9970 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9971 0 : .await
9972 : {
9973 0 : Ok(()) => return Err(OperationError::Cancelled),
9974 0 : Err(err) => {
9975 0 : return Err(OperationError::FinalizeError(
9976 0 : format!(
9977 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9978 0 : )
9979 0 : .into(),
9980 0 : ));
9981 : }
9982 : }
9983 0 : }
9984 :
9985 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
9986 :
9987 0 : waiters = self
9988 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9989 0 : .await;
9990 : }
9991 :
9992 0 : if let Err(err) = self
9993 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9994 0 : .await
9995 : {
9996 : // This isn't a huge issue since the filling process starts upon request. However, it
9997 : // will prevent the next drain from starting. The only case in which this can fail
9998 : // is database unavailability. Such a case will require manual intervention.
9999 0 : return Err(OperationError::FinalizeError(
10000 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
10001 0 : .into(),
10002 0 : ));
10003 0 : }
10004 :
10005 0 : Ok(())
10006 0 : }
10007 :
10008 : /// Updates scrubber metadata health check results.
10009 0 : pub(crate) async fn metadata_health_update(
10010 0 : &self,
10011 0 : update_req: MetadataHealthUpdateRequest,
10012 0 : ) -> Result<(), ApiError> {
10013 0 : let now = chrono::offset::Utc::now();
10014 0 : let (healthy_records, unhealthy_records) = {
10015 0 : let locked = self.inner.read().unwrap();
10016 0 : let healthy_records = update_req
10017 0 : .healthy_tenant_shards
10018 0 : .into_iter()
10019 : // Retain only health records associated with tenant shards managed by storage controller.
10020 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
10021 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
10022 0 : .collect();
10023 0 : let unhealthy_records = update_req
10024 0 : .unhealthy_tenant_shards
10025 0 : .into_iter()
10026 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
10027 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
10028 0 : .collect();
10029 :
10030 0 : (healthy_records, unhealthy_records)
10031 : };
10032 :
10033 0 : self.persistence
10034 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
10035 0 : .await?;
10036 0 : Ok(())
10037 0 : }
10038 :
10039 : /// Lists the tenant shards that has unhealthy metadata status.
10040 0 : pub(crate) async fn metadata_health_list_unhealthy(
10041 0 : &self,
10042 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
10043 0 : let result = self
10044 0 : .persistence
10045 0 : .list_unhealthy_metadata_health_records()
10046 0 : .await?
10047 0 : .iter()
10048 0 : .map(|p| p.get_tenant_shard_id().unwrap())
10049 0 : .collect();
10050 :
10051 0 : Ok(result)
10052 0 : }
10053 :
10054 : /// Lists the tenant shards that have not been scrubbed for some duration.
10055 0 : pub(crate) async fn metadata_health_list_outdated(
10056 0 : &self,
10057 0 : not_scrubbed_for: Duration,
10058 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
10059 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
10060 0 : let result = self
10061 0 : .persistence
10062 0 : .list_outdated_metadata_health_records(earlier)
10063 0 : .await?
10064 0 : .into_iter()
10065 0 : .map(|record| record.into())
10066 0 : .collect();
10067 0 : Ok(result)
10068 0 : }
10069 :
10070 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
10071 0 : self.inner.read().unwrap().get_leadership_status()
10072 0 : }
10073 :
10074 : /// Handler for step down requests
10075 : ///
10076 : /// Step down runs in separate task since once it's called it should
10077 : /// be driven to completion. Subsequent requests will wait on the same
10078 : /// step down task.
10079 0 : pub(crate) async fn step_down(self: &Arc<Self>) -> GlobalObservedState {
10080 0 : let handle = self.step_down_barrier.get_or_init(|| {
10081 0 : let step_down_self = self.clone();
10082 0 : let (tx, rx) = tokio::sync::watch::channel::<Option<GlobalObservedState>>(None);
10083 0 : tokio::spawn(async move {
10084 0 : let state = step_down_self.step_down_task().await;
10085 0 : tx.send(Some(state))
10086 0 : .expect("Task Arc<Service> keeps receiver alive");
10087 0 : });
10088 :
10089 0 : rx
10090 0 : });
10091 :
10092 0 : handle
10093 0 : .clone()
10094 0 : .wait_for(|observed_state| observed_state.is_some())
10095 0 : .await
10096 0 : .expect("Task Arc<Service> keeps sender alive")
10097 0 : .deref()
10098 0 : .clone()
10099 0 : .expect("Checked above")
10100 0 : }
10101 :
10102 0 : async fn step_down_task(&self) -> GlobalObservedState {
10103 0 : tracing::info!("Received step down request from peer");
10104 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
10105 :
10106 0 : self.inner.write().unwrap().step_down();
10107 :
10108 0 : let stop_reconciliations =
10109 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown);
10110 0 : let mut stop_reconciliations = std::pin::pin!(stop_reconciliations);
10111 :
10112 0 : let started_at = Instant::now();
10113 :
10114 : // Wait for reconciliations to stop and warn if that's taking a long time
10115 : loop {
10116 0 : tokio::select! {
10117 0 : _ = &mut stop_reconciliations => {
10118 0 : tracing::info!("Reconciliations stopped, proceeding with step down");
10119 0 : break;
10120 : }
10121 0 : _ = tokio::time::sleep(Duration::from_secs(10)) => {
10122 0 : tracing::warn!(
10123 0 : elapsed_sec=%started_at.elapsed().as_secs(),
10124 0 : "Stopping reconciliations during step down is taking too long"
10125 : );
10126 : }
10127 : }
10128 : }
10129 :
10130 0 : let mut global_observed = GlobalObservedState::default();
10131 0 : let locked = self.inner.read().unwrap();
10132 0 : for (tid, tenant_shard) in locked.tenants.iter() {
10133 0 : global_observed
10134 0 : .0
10135 0 : .insert(*tid, tenant_shard.observed.clone());
10136 0 : }
10137 :
10138 0 : global_observed
10139 0 : }
10140 :
10141 0 : pub(crate) async fn update_shards_preferred_azs(
10142 0 : &self,
10143 0 : req: ShardsPreferredAzsRequest,
10144 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
10145 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
10146 0 : let updated = self
10147 0 : .persistence
10148 0 : .set_tenant_shard_preferred_azs(preferred_azs)
10149 0 : .await
10150 0 : .map_err(|err| {
10151 0 : ApiError::InternalServerError(anyhow::anyhow!(
10152 0 : "Failed to persist preferred AZs: {err}"
10153 0 : ))
10154 0 : })?;
10155 :
10156 0 : let mut updated_in_mem_and_db = Vec::default();
10157 :
10158 0 : let mut locked = self.inner.write().unwrap();
10159 0 : let state = locked.deref_mut();
10160 0 : for (tid, az_id) in updated {
10161 0 : let shard = state.tenants.get_mut(&tid);
10162 0 : if let Some(shard) = shard {
10163 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
10164 0 : updated_in_mem_and_db.push(tid);
10165 0 : }
10166 : }
10167 :
10168 0 : Ok(ShardsPreferredAzsResponse {
10169 0 : updated: updated_in_mem_and_db,
10170 0 : })
10171 0 : }
10172 : }
10173 :
10174 : #[cfg(test)]
10175 : mod tests {
10176 : use super::*;
10177 :
10178 : /// Tests Service::compute_split_shards. For readability, this specifies sizes in GBs rather
10179 : /// than bytes. Note that max_logical_size is the total logical size of the largest timeline
10180 : /// summed across all shards.
10181 : #[test]
10182 1 : fn compute_split_shards() {
10183 : // Size-based split: two shards have a 500 GB timeline, which need to split into 8 shards
10184 : // that are <= 64 GB,
10185 1 : assert_eq!(
10186 1 : Service::compute_split_shards(ShardSplitInputs {
10187 1 : shard_count: ShardCount(2),
10188 1 : max_logical_size: 500,
10189 1 : split_threshold: 64,
10190 1 : max_split_shards: 16,
10191 1 : initial_split_threshold: 0,
10192 1 : initial_split_shards: 0,
10193 1 : }),
10194 : Some(ShardCount(8))
10195 : );
10196 :
10197 : // Size-based split: noop at or below threshold, fires above.
10198 1 : assert_eq!(
10199 1 : Service::compute_split_shards(ShardSplitInputs {
10200 1 : shard_count: ShardCount(2),
10201 1 : max_logical_size: 127,
10202 1 : split_threshold: 64,
10203 1 : max_split_shards: 16,
10204 1 : initial_split_threshold: 0,
10205 1 : initial_split_shards: 0,
10206 1 : }),
10207 : None,
10208 : );
10209 1 : assert_eq!(
10210 1 : Service::compute_split_shards(ShardSplitInputs {
10211 1 : shard_count: ShardCount(2),
10212 1 : max_logical_size: 128,
10213 1 : split_threshold: 64,
10214 1 : max_split_shards: 16,
10215 1 : initial_split_threshold: 0,
10216 1 : initial_split_shards: 0,
10217 1 : }),
10218 : None,
10219 : );
10220 1 : assert_eq!(
10221 1 : Service::compute_split_shards(ShardSplitInputs {
10222 1 : shard_count: ShardCount(2),
10223 1 : max_logical_size: 129,
10224 1 : split_threshold: 64,
10225 1 : max_split_shards: 16,
10226 1 : initial_split_threshold: 0,
10227 1 : initial_split_shards: 0,
10228 1 : }),
10229 : Some(ShardCount(4)),
10230 : );
10231 :
10232 : // Size-based split: clamped to max_split_shards.
10233 1 : assert_eq!(
10234 1 : Service::compute_split_shards(ShardSplitInputs {
10235 1 : shard_count: ShardCount(2),
10236 1 : max_logical_size: 10000,
10237 1 : split_threshold: 64,
10238 1 : max_split_shards: 16,
10239 1 : initial_split_threshold: 0,
10240 1 : initial_split_shards: 0,
10241 1 : }),
10242 : Some(ShardCount(16))
10243 : );
10244 :
10245 : // Size-based split: tenant already at or beyond max_split_shards is not split.
10246 1 : assert_eq!(
10247 1 : Service::compute_split_shards(ShardSplitInputs {
10248 1 : shard_count: ShardCount(16),
10249 1 : max_logical_size: 10000,
10250 1 : split_threshold: 64,
10251 1 : max_split_shards: 16,
10252 1 : initial_split_threshold: 0,
10253 1 : initial_split_shards: 0,
10254 1 : }),
10255 : None
10256 : );
10257 :
10258 1 : assert_eq!(
10259 1 : Service::compute_split_shards(ShardSplitInputs {
10260 1 : shard_count: ShardCount(32),
10261 1 : max_logical_size: 10000,
10262 1 : split_threshold: 64,
10263 1 : max_split_shards: 16,
10264 1 : initial_split_threshold: 0,
10265 1 : initial_split_shards: 0,
10266 1 : }),
10267 : None
10268 : );
10269 :
10270 : // Size-based split: a non-power-of-2 shard count is normalized to power-of-2 if it
10271 : // exceeds split_threshold (i.e. a 3-shard tenant splits into 8, not 6).
10272 1 : assert_eq!(
10273 1 : Service::compute_split_shards(ShardSplitInputs {
10274 1 : shard_count: ShardCount(3),
10275 1 : max_logical_size: 320,
10276 1 : split_threshold: 64,
10277 1 : max_split_shards: 16,
10278 1 : initial_split_threshold: 0,
10279 1 : initial_split_shards: 0,
10280 1 : }),
10281 : Some(ShardCount(8))
10282 : );
10283 :
10284 : // Size-based split: a non-power-of-2 shard count is not normalized to power-of-2 if the
10285 : // existing shards are below or at split_threshold, but splits into 4 if it exceeds it.
10286 1 : assert_eq!(
10287 1 : Service::compute_split_shards(ShardSplitInputs {
10288 1 : shard_count: ShardCount(3),
10289 1 : max_logical_size: 191,
10290 1 : split_threshold: 64,
10291 1 : max_split_shards: 16,
10292 1 : initial_split_threshold: 0,
10293 1 : initial_split_shards: 0,
10294 1 : }),
10295 : None
10296 : );
10297 1 : assert_eq!(
10298 1 : Service::compute_split_shards(ShardSplitInputs {
10299 1 : shard_count: ShardCount(3),
10300 1 : max_logical_size: 192,
10301 1 : split_threshold: 64,
10302 1 : max_split_shards: 16,
10303 1 : initial_split_threshold: 0,
10304 1 : initial_split_shards: 0,
10305 1 : }),
10306 : None
10307 : );
10308 1 : assert_eq!(
10309 1 : Service::compute_split_shards(ShardSplitInputs {
10310 1 : shard_count: ShardCount(3),
10311 1 : max_logical_size: 193,
10312 1 : split_threshold: 64,
10313 1 : max_split_shards: 16,
10314 1 : initial_split_threshold: 0,
10315 1 : initial_split_shards: 0,
10316 1 : }),
10317 : Some(ShardCount(4))
10318 : );
10319 :
10320 : // Initial split: tenant has a 10 GB timeline, split into 4 shards.
10321 1 : assert_eq!(
10322 1 : Service::compute_split_shards(ShardSplitInputs {
10323 1 : shard_count: ShardCount(1),
10324 1 : max_logical_size: 10,
10325 1 : split_threshold: 0,
10326 1 : max_split_shards: 16,
10327 1 : initial_split_threshold: 8,
10328 1 : initial_split_shards: 4,
10329 1 : }),
10330 : Some(ShardCount(4))
10331 : );
10332 :
10333 : // Initial split: 0 ShardCount is equivalent to 1.
10334 1 : assert_eq!(
10335 1 : Service::compute_split_shards(ShardSplitInputs {
10336 1 : shard_count: ShardCount(0),
10337 1 : max_logical_size: 10,
10338 1 : split_threshold: 0,
10339 1 : max_split_shards: 16,
10340 1 : initial_split_threshold: 8,
10341 1 : initial_split_shards: 4,
10342 1 : }),
10343 : Some(ShardCount(4))
10344 : );
10345 :
10346 : // Initial split: at or below threshold is noop.
10347 1 : assert_eq!(
10348 1 : Service::compute_split_shards(ShardSplitInputs {
10349 1 : shard_count: ShardCount(1),
10350 1 : max_logical_size: 7,
10351 1 : split_threshold: 0,
10352 1 : max_split_shards: 16,
10353 1 : initial_split_threshold: 8,
10354 1 : initial_split_shards: 4,
10355 1 : }),
10356 : None,
10357 : );
10358 1 : assert_eq!(
10359 1 : Service::compute_split_shards(ShardSplitInputs {
10360 1 : shard_count: ShardCount(1),
10361 1 : max_logical_size: 8,
10362 1 : split_threshold: 0,
10363 1 : max_split_shards: 16,
10364 1 : initial_split_threshold: 8,
10365 1 : initial_split_shards: 4,
10366 1 : }),
10367 : None,
10368 : );
10369 1 : assert_eq!(
10370 1 : Service::compute_split_shards(ShardSplitInputs {
10371 1 : shard_count: ShardCount(1),
10372 1 : max_logical_size: 9,
10373 1 : split_threshold: 0,
10374 1 : max_split_shards: 16,
10375 1 : initial_split_threshold: 8,
10376 1 : initial_split_shards: 4,
10377 1 : }),
10378 : Some(ShardCount(4))
10379 : );
10380 :
10381 : // Initial split: already sharded tenant is not affected, even if above threshold and below
10382 : // shard count.
10383 1 : assert_eq!(
10384 1 : Service::compute_split_shards(ShardSplitInputs {
10385 1 : shard_count: ShardCount(2),
10386 1 : max_logical_size: 20,
10387 1 : split_threshold: 0,
10388 1 : max_split_shards: 16,
10389 1 : initial_split_threshold: 8,
10390 1 : initial_split_shards: 4,
10391 1 : }),
10392 : None,
10393 : );
10394 :
10395 : // Initial split: clamped to max_shards.
10396 1 : assert_eq!(
10397 1 : Service::compute_split_shards(ShardSplitInputs {
10398 1 : shard_count: ShardCount(1),
10399 1 : max_logical_size: 10,
10400 1 : split_threshold: 0,
10401 1 : max_split_shards: 3,
10402 1 : initial_split_threshold: 8,
10403 1 : initial_split_shards: 4,
10404 1 : }),
10405 : Some(ShardCount(3)),
10406 : );
10407 :
10408 : // Initial+size split: tenant eligible for both will use the larger shard count.
10409 1 : assert_eq!(
10410 1 : Service::compute_split_shards(ShardSplitInputs {
10411 1 : shard_count: ShardCount(1),
10412 1 : max_logical_size: 10,
10413 1 : split_threshold: 64,
10414 1 : max_split_shards: 16,
10415 1 : initial_split_threshold: 8,
10416 1 : initial_split_shards: 4,
10417 1 : }),
10418 : Some(ShardCount(4)),
10419 : );
10420 1 : assert_eq!(
10421 1 : Service::compute_split_shards(ShardSplitInputs {
10422 1 : shard_count: ShardCount(1),
10423 1 : max_logical_size: 500,
10424 1 : split_threshold: 64,
10425 1 : max_split_shards: 16,
10426 1 : initial_split_threshold: 8,
10427 1 : initial_split_shards: 4,
10428 1 : }),
10429 : Some(ShardCount(8)),
10430 : );
10431 :
10432 : // Initial+size split: sharded tenant is only eligible for size-based split.
10433 1 : assert_eq!(
10434 1 : Service::compute_split_shards(ShardSplitInputs {
10435 1 : shard_count: ShardCount(2),
10436 1 : max_logical_size: 200,
10437 1 : split_threshold: 64,
10438 1 : max_split_shards: 16,
10439 1 : initial_split_threshold: 8,
10440 1 : initial_split_shards: 8,
10441 1 : }),
10442 : Some(ShardCount(4)),
10443 : );
10444 :
10445 : // Initial+size split: uses the larger shard count even with initial_split_threshold above
10446 : // split_threshold.
10447 1 : assert_eq!(
10448 1 : Service::compute_split_shards(ShardSplitInputs {
10449 1 : shard_count: ShardCount(1),
10450 1 : max_logical_size: 10,
10451 1 : split_threshold: 4,
10452 1 : max_split_shards: 16,
10453 1 : initial_split_threshold: 8,
10454 1 : initial_split_shards: 8,
10455 1 : }),
10456 : Some(ShardCount(8)),
10457 : );
10458 :
10459 : // Test backwards compatibility with production settings when initial/size-based splits were
10460 : // rolled out: a single split into 8 shards at 64 GB. Any already sharded tenants with <8
10461 : // shards will split according to split_threshold.
10462 1 : assert_eq!(
10463 1 : Service::compute_split_shards(ShardSplitInputs {
10464 1 : shard_count: ShardCount(1),
10465 1 : max_logical_size: 65,
10466 1 : split_threshold: 64,
10467 1 : max_split_shards: 8,
10468 1 : initial_split_threshold: 64,
10469 1 : initial_split_shards: 8,
10470 1 : }),
10471 : Some(ShardCount(8)),
10472 : );
10473 :
10474 1 : assert_eq!(
10475 1 : Service::compute_split_shards(ShardSplitInputs {
10476 1 : shard_count: ShardCount(1),
10477 1 : max_logical_size: 64,
10478 1 : split_threshold: 64,
10479 1 : max_split_shards: 8,
10480 1 : initial_split_threshold: 64,
10481 1 : initial_split_shards: 8,
10482 1 : }),
10483 : None,
10484 : );
10485 :
10486 1 : assert_eq!(
10487 1 : Service::compute_split_shards(ShardSplitInputs {
10488 1 : shard_count: ShardCount(2),
10489 1 : max_logical_size: 129,
10490 1 : split_threshold: 64,
10491 1 : max_split_shards: 8,
10492 1 : initial_split_threshold: 64,
10493 1 : initial_split_shards: 8,
10494 1 : }),
10495 : Some(ShardCount(4)),
10496 : );
10497 1 : }
10498 : }
|
