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 reconciliation errors, occured for one shard,
215 : // after which the shard is ignored when considering to run optimizations.
216 : const MAX_CONSECUTIVE_RECONCILIATION_ERRORS: usize = 5;
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 : struct MutationLocation {
702 : node: Node,
703 : generation: Generation,
704 : }
705 :
706 : #[derive(Clone)]
707 : struct ShardMutationLocations {
708 : latest: MutationLocation,
709 : other: Vec<MutationLocation>,
710 : }
711 :
712 : #[derive(Default, Clone)]
713 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
714 :
715 : struct ReconcileAllResult {
716 : spawned_reconciles: usize,
717 : keep_failing_reconciles: usize,
718 : has_delayed_reconciles: bool,
719 : }
720 :
721 : impl ReconcileAllResult {
722 0 : fn new(
723 0 : spawned_reconciles: usize,
724 0 : keep_failing_reconciles: usize,
725 0 : has_delayed_reconciles: bool,
726 0 : ) -> Self {
727 0 : assert!(
728 0 : spawned_reconciles >= keep_failing_reconciles,
729 0 : "It is impossible to have more keep-failing reconciles than spawned reconciles"
730 : );
731 0 : Self {
732 0 : spawned_reconciles,
733 0 : keep_failing_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 keep-failing reconciles.
740 0 : fn can_run_optimizations(&self) -> bool {
741 0 : !self.has_delayed_reconciles && self.spawned_reconciles == self.keep_failing_reconciles
742 0 : }
743 : }
744 :
745 : impl Service {
746 0 : pub fn get_config(&self) -> &Config {
747 0 : &self.config
748 0 : }
749 :
750 0 : pub fn get_http_client(&self) -> &reqwest::Client {
751 0 : &self.http_client
752 0 : }
753 :
754 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
755 : /// view of the world, and determine which pageservers are responsive.
756 : #[instrument(skip_all)]
757 : async fn startup_reconcile(
758 : self: &Arc<Service>,
759 : current_leader: Option<ControllerPersistence>,
760 : leader_step_down_state: Option<GlobalObservedState>,
761 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
762 : Result<(), (TenantShardId, NotifyError)>,
763 : >,
764 : ) {
765 : // Startup reconciliation does I/O to other services: whether they
766 : // are responsive or not, we should aim to finish within our deadline, because:
767 : // - If we don't, a k8s readiness hook watching /ready will kill us.
768 : // - While we're waiting for startup reconciliation, we are not fully
769 : // available for end user operations like creating/deleting tenants and timelines.
770 : //
771 : // We set multiple deadlines to break up the time available between the phases of work: this is
772 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
773 : let start_at = Instant::now();
774 : let node_scan_deadline = start_at
775 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
776 : .expect("Reconcile timeout is a modest constant");
777 :
778 : let observed = if let Some(state) = leader_step_down_state {
779 : tracing::info!(
780 : "Using observed state received from leader at {}",
781 : current_leader.as_ref().unwrap().address
782 : );
783 :
784 : state
785 : } else {
786 : self.build_global_observed_state(node_scan_deadline).await
787 : };
788 :
789 : // Accumulate a list of any tenant locations that ought to be detached
790 : let mut cleanup = Vec::new();
791 :
792 : // Send initial heartbeat requests to all nodes loaded from the database
793 : let all_nodes = {
794 : let locked = self.inner.read().unwrap();
795 : locked.nodes.clone()
796 : };
797 : let (mut nodes_online, mut sks_online) =
798 : self.initial_heartbeat_round(all_nodes.keys()).await;
799 :
800 : // List of tenants for which we will attempt to notify compute of their location at startup
801 : let mut compute_notifications = Vec::new();
802 :
803 : // Populate intent and observed states for all tenants, based on reported state on pageservers
804 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
805 : let shard_count = {
806 : let mut locked = self.inner.write().unwrap();
807 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
808 :
809 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
810 : let mut new_nodes = (**nodes).clone();
811 : for (node_id, node) in new_nodes.iter_mut() {
812 : if let Some(utilization) = nodes_online.remove(node_id) {
813 : node.set_availability(NodeAvailability::Active(utilization));
814 : scheduler.node_upsert(node);
815 : }
816 : }
817 : *nodes = Arc::new(new_nodes);
818 :
819 : let mut new_sks = (**safekeepers).clone();
820 : for (node_id, node) in new_sks.iter_mut() {
821 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
822 : node.set_availability(SafekeeperState::Available {
823 : utilization,
824 : last_seen_at,
825 : });
826 : }
827 : }
828 : *safekeepers = Arc::new(new_sks);
829 :
830 : for (tenant_shard_id, observed_state) in observed.0 {
831 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
832 : for node_id in observed_state.locations.keys() {
833 : cleanup.push((tenant_shard_id, *node_id));
834 : }
835 :
836 : continue;
837 : };
838 :
839 : tenant_shard.observed = observed_state;
840 : }
841 :
842 : // Populate each tenant's intent state
843 : let mut schedule_context = ScheduleContext::default();
844 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
845 : if tenant_shard_id.shard_number == ShardNumber(0) {
846 : // Reset scheduling context each time we advance to the next Tenant
847 : schedule_context = ScheduleContext::default();
848 : }
849 :
850 : tenant_shard.intent_from_observed(scheduler);
851 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
852 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
853 : // not enough pageservers are available. The tenant may well still be available
854 : // to clients.
855 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
856 : } else {
857 : // If we're both intending and observed to be attached at a particular node, we will
858 : // emit a compute notification for this. In the case where our observed state does not
859 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
860 : if let Some(attached_at) = tenant_shard.stably_attached() {
861 : compute_notifications.push(compute_hook::ShardUpdate {
862 : tenant_shard_id: *tenant_shard_id,
863 : node_id: attached_at,
864 : stripe_size: tenant_shard.shard.stripe_size,
865 : preferred_az: tenant_shard
866 : .preferred_az()
867 0 : .map(|az| Cow::Owned(az.clone())),
868 : });
869 : }
870 : }
871 : }
872 :
873 : tenants.len()
874 : };
875 :
876 : // Before making any obeservable changes to the cluster, persist self
877 : // as leader in database and memory.
878 : let leadership = Leadership::new(
879 : self.persistence.clone(),
880 : self.config.clone(),
881 : self.cancel.child_token(),
882 : );
883 :
884 : if let Err(e) = leadership.become_leader(current_leader).await {
885 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
886 : std::process::exit(1);
887 : }
888 :
889 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
890 : let sk_schedule_requests =
891 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
892 : Ok(v) => v,
893 : Err(e) => {
894 : tracing::warn!(
895 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
896 : );
897 : // std::process::exit(1);
898 : Vec::new()
899 : }
900 : };
901 :
902 : {
903 : let mut locked = self.inner.write().unwrap();
904 : locked.become_leader();
905 :
906 : for (sk_id, _sk) in locked.safekeepers.clone().iter() {
907 : locked.safekeeper_reconcilers.start_reconciler(*sk_id, self);
908 : }
909 :
910 : locked
911 : .safekeeper_reconcilers
912 : .schedule_request_vec(sk_schedule_requests);
913 : }
914 :
915 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
916 : // generation_pageserver in the database.
917 :
918 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
919 : // will emit compute hook notifications when they reconcile.
920 : //
921 : // Ordering: our calls to notify_attach_background synchronously establish a relative order for these notifications vs. any later
922 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
923 : // calls will be correctly ordered wrt these.
924 : //
925 : // Concurrency: we call notify_attach_background for all tenants, which will create O(N) tokio tasks, but almost all of them
926 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
927 : // unit and start doing I/O.
928 : tracing::info!(
929 : "Sending {} compute notifications",
930 : compute_notifications.len()
931 : );
932 : self.compute_hook.notify_attach_background(
933 : compute_notifications,
934 : bg_compute_notify_result_tx.clone(),
935 : &self.cancel,
936 : );
937 :
938 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
939 : // which require it: under normal circumstances this should only include tenants that were in some
940 : // transient state before we restarted, or any tenants whose compute hooks failed above.
941 : tracing::info!("Checking for shards in need of reconciliation...");
942 : let reconcile_all_result = self.reconcile_all();
943 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
944 : // normal operations may proceed.
945 :
946 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
947 : // background because it does not need to complete in order to proceed with other work.
948 : if !cleanup.is_empty() {
949 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
950 : tokio::task::spawn({
951 : let cleanup_self = self.clone();
952 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
953 : });
954 : }
955 :
956 : // Reconcile the timeline imports:
957 : // 1. Mark each tenant shard of tenants with an importing timeline as importing.
958 : // 2. Finalize the completed imports in the background. This handles the case where
959 : // the previous storage controller instance shut down whilst finalizing imports.
960 : let imports = self.persistence.list_timeline_imports().await;
961 : match imports {
962 : Ok(mut imports) => {
963 : {
964 : let mut locked = self.inner.write().unwrap();
965 : for import in &imports {
966 : locked
967 : .tenants
968 : .range_mut(TenantShardId::tenant_range(import.tenant_id))
969 0 : .for_each(|(_id, shard)| {
970 0 : shard.importing = TimelineImportState::Importing
971 0 : });
972 : }
973 : }
974 :
975 0 : imports.retain(|import| import.is_complete());
976 : tokio::task::spawn({
977 : let finalize_imports_self = self.clone();
978 0 : async move {
979 0 : finalize_imports_self
980 0 : .finalize_timeline_imports(imports)
981 0 : .await
982 0 : }
983 : });
984 : }
985 : Err(err) => {
986 : tracing::error!("Could not retrieve completed imports from database: {err}");
987 : }
988 : }
989 :
990 : let spawned_reconciles = reconcile_all_result.spawned_reconciles;
991 : tracing::info!(
992 : "Startup complete, spawned {spawned_reconciles} reconciliation tasks ({shard_count} shards total)"
993 : );
994 : }
995 :
996 0 : async fn initial_heartbeat_round<'a>(
997 0 : &self,
998 0 : node_ids: impl Iterator<Item = &'a NodeId>,
999 0 : ) -> (
1000 0 : HashMap<NodeId, PageserverUtilization>,
1001 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
1002 0 : ) {
1003 0 : assert!(!self.startup_complete.is_ready());
1004 :
1005 0 : let all_nodes = {
1006 0 : let locked = self.inner.read().unwrap();
1007 0 : locked.nodes.clone()
1008 : };
1009 :
1010 0 : let mut nodes_to_heartbeat = HashMap::new();
1011 0 : for node_id in node_ids {
1012 0 : match all_nodes.get(node_id) {
1013 0 : Some(node) => {
1014 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
1015 0 : }
1016 : None => {
1017 0 : tracing::warn!("Node {node_id} was removed during start-up");
1018 : }
1019 : }
1020 : }
1021 :
1022 0 : let all_sks = {
1023 0 : let locked = self.inner.read().unwrap();
1024 0 : locked.safekeepers.clone()
1025 : };
1026 :
1027 0 : tracing::info!("Sending initial heartbeats...");
1028 0 : let (res_ps, res_sk) = tokio::join!(
1029 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
1030 0 : self.heartbeater_sk.heartbeat(all_sks)
1031 : );
1032 :
1033 0 : let mut online_nodes = HashMap::new();
1034 0 : if let Ok(deltas) = res_ps {
1035 0 : for (node_id, status) in deltas.0 {
1036 0 : match status {
1037 0 : PageserverState::Available { utilization, .. } => {
1038 0 : online_nodes.insert(node_id, utilization);
1039 0 : }
1040 0 : PageserverState::Offline => {}
1041 : PageserverState::WarmingUp { .. } => {
1042 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
1043 : }
1044 : }
1045 : }
1046 0 : }
1047 :
1048 0 : let mut online_sks = HashMap::new();
1049 0 : if let Ok(deltas) = res_sk {
1050 0 : for (node_id, status) in deltas.0 {
1051 0 : match status {
1052 : SafekeeperState::Available {
1053 0 : utilization,
1054 0 : last_seen_at,
1055 0 : } => {
1056 0 : online_sks.insert(node_id, (utilization, last_seen_at));
1057 0 : }
1058 0 : SafekeeperState::Offline => {}
1059 : }
1060 : }
1061 0 : }
1062 :
1063 0 : (online_nodes, online_sks)
1064 0 : }
1065 :
1066 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
1067 : ///
1068 : /// The result includes only nodes which responded within the deadline
1069 0 : async fn scan_node_locations(
1070 0 : &self,
1071 0 : deadline: Instant,
1072 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
1073 0 : let nodes = {
1074 0 : let locked = self.inner.read().unwrap();
1075 0 : locked.nodes.clone()
1076 : };
1077 :
1078 0 : let mut node_results = HashMap::new();
1079 :
1080 0 : let mut node_list_futs = FuturesUnordered::new();
1081 :
1082 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
1083 0 : for node in nodes.values() {
1084 0 : node_list_futs.push({
1085 0 : async move {
1086 0 : tracing::info!("Scanning shards on node {node}...");
1087 0 : let timeout = Duration::from_secs(5);
1088 0 : let response = node
1089 0 : .with_client_retries(
1090 0 : |client| async move { client.list_location_config().await },
1091 0 : &self.http_client,
1092 0 : &self.config.pageserver_jwt_token,
1093 : 1,
1094 : 5,
1095 0 : timeout,
1096 0 : &self.cancel,
1097 : )
1098 0 : .await;
1099 0 : (node.get_id(), response)
1100 0 : }
1101 : });
1102 : }
1103 :
1104 : loop {
1105 0 : let (node_id, result) = tokio::select! {
1106 0 : next = node_list_futs.next() => {
1107 0 : match next {
1108 0 : Some(result) => result,
1109 : None =>{
1110 : // We got results for all our nodes
1111 0 : break;
1112 : }
1113 :
1114 : }
1115 : },
1116 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
1117 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
1118 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
1119 0 : break;
1120 : }
1121 : };
1122 :
1123 0 : let Some(list_response) = result else {
1124 0 : tracing::info!("Shutdown during startup_reconcile");
1125 0 : break;
1126 : };
1127 :
1128 0 : match list_response {
1129 0 : Err(e) => {
1130 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
1131 : }
1132 0 : Ok(listing) => {
1133 0 : node_results.insert(node_id, listing);
1134 0 : }
1135 : }
1136 : }
1137 :
1138 0 : node_results
1139 0 : }
1140 :
1141 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
1142 0 : let node_listings = self.scan_node_locations(deadline).await;
1143 0 : let mut observed = GlobalObservedState::default();
1144 :
1145 0 : for (node_id, location_confs) in node_listings {
1146 0 : tracing::info!(
1147 0 : "Received {} shard statuses from pageserver {}",
1148 0 : location_confs.tenant_shards.len(),
1149 : node_id
1150 : );
1151 :
1152 0 : for (tid, location_conf) in location_confs.tenant_shards {
1153 0 : let entry = observed.0.entry(tid).or_default();
1154 0 : entry.locations.insert(
1155 0 : node_id,
1156 0 : ObservedStateLocation {
1157 0 : conf: location_conf,
1158 0 : },
1159 0 : );
1160 0 : }
1161 : }
1162 :
1163 0 : observed
1164 0 : }
1165 :
1166 : /// Used during [`Self::startup_reconcile`] and shard splits: detach a list of unknown-to-us
1167 : /// tenants from pageservers.
1168 : ///
1169 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1170 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1171 : /// other task trying to attach it.
1172 : #[instrument(skip_all)]
1173 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1174 : let nodes = self.inner.read().unwrap().nodes.clone();
1175 :
1176 : for (tenant_shard_id, node_id) in cleanup {
1177 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1178 : let Some(node) = nodes.get(&node_id) else {
1179 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1180 : // a location to clean up on a node that has since been removed.
1181 : tracing::info!(
1182 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1183 : );
1184 : continue;
1185 : };
1186 :
1187 : if self.cancel.is_cancelled() {
1188 : break;
1189 : }
1190 :
1191 : let client = PageserverClient::new(
1192 : node.get_id(),
1193 : self.http_client.clone(),
1194 : node.base_url(),
1195 : self.config.pageserver_jwt_token.as_deref(),
1196 : );
1197 : match client
1198 : .location_config(
1199 : tenant_shard_id,
1200 : LocationConfig {
1201 : mode: LocationConfigMode::Detached,
1202 : generation: None,
1203 : secondary_conf: None,
1204 : shard_number: tenant_shard_id.shard_number.0,
1205 : shard_count: tenant_shard_id.shard_count.literal(),
1206 : shard_stripe_size: 0,
1207 : tenant_conf: models::TenantConfig::default(),
1208 : },
1209 : None,
1210 : false,
1211 : )
1212 : .await
1213 : {
1214 : Ok(()) => {
1215 : tracing::info!(
1216 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1217 : );
1218 : }
1219 : Err(e) => {
1220 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1221 : // break anything.
1222 : tracing::error!(
1223 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1224 : );
1225 : }
1226 : }
1227 : }
1228 : }
1229 :
1230 : /// Long running background task that periodically wakes up and looks for shards that need
1231 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1232 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1233 : /// for those retries.
1234 : #[instrument(skip_all)]
1235 : async fn background_reconcile(self: &Arc<Self>) {
1236 : self.startup_complete.clone().wait().await;
1237 :
1238 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1239 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1240 : while !self.reconcilers_cancel.is_cancelled() {
1241 : tokio::select! {
1242 : _ = interval.tick() => {
1243 : let reconcile_all_result = self.reconcile_all();
1244 : if reconcile_all_result.can_run_optimizations() {
1245 : // Run optimizer only when we didn't find any other work to do
1246 : self.optimize_all().await;
1247 : }
1248 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1249 : // must be responsive when new projects begin ingesting and reach the threshold.
1250 : self.autosplit_tenants().await;
1251 : }
1252 : _ = self.reconcilers_cancel.cancelled() => return
1253 : }
1254 : }
1255 : }
1256 : /// Heartbeat all storage nodes once in a while.
1257 : #[instrument(skip_all)]
1258 : async fn spawn_heartbeat_driver(self: &Arc<Self>) {
1259 : self.startup_complete.clone().wait().await;
1260 :
1261 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1262 : while !self.cancel.is_cancelled() {
1263 : tokio::select! {
1264 : _ = interval.tick() => { }
1265 : _ = self.cancel.cancelled() => return
1266 : };
1267 :
1268 : let nodes = {
1269 : let locked = self.inner.read().unwrap();
1270 : locked.nodes.clone()
1271 : };
1272 :
1273 : let safekeepers = {
1274 : let locked = self.inner.read().unwrap();
1275 : locked.safekeepers.clone()
1276 : };
1277 :
1278 : let (res_ps, res_sk) = tokio::join!(
1279 : self.heartbeater_ps.heartbeat(nodes),
1280 : self.heartbeater_sk.heartbeat(safekeepers)
1281 : );
1282 :
1283 : if let Ok(deltas) = res_ps {
1284 : let mut to_handle = Vec::default();
1285 :
1286 : for (node_id, state) in deltas.0 {
1287 : let new_availability = match state {
1288 : PageserverState::Available { utilization, .. } => {
1289 : NodeAvailability::Active(utilization)
1290 : }
1291 : PageserverState::WarmingUp { started_at } => {
1292 : NodeAvailability::WarmingUp(started_at)
1293 : }
1294 : PageserverState::Offline => {
1295 : // The node might have been placed in the WarmingUp state
1296 : // while the heartbeat round was on-going. Hence, filter out
1297 : // offline transitions for WarmingUp nodes that are still within
1298 : // their grace period.
1299 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1300 : .get_node(node_id)
1301 : .await
1302 : .as_ref()
1303 0 : .map(|n| n.get_availability())
1304 : {
1305 : let now = Instant::now();
1306 : if now - *started_at >= self.config.max_warming_up_interval {
1307 : NodeAvailability::Offline
1308 : } else {
1309 : NodeAvailability::WarmingUp(*started_at)
1310 : }
1311 : } else {
1312 : NodeAvailability::Offline
1313 : }
1314 : }
1315 : };
1316 :
1317 : let node_lock = trace_exclusive_lock(
1318 : &self.node_op_locks,
1319 : node_id,
1320 : NodeOperations::Configure,
1321 : )
1322 : .await;
1323 :
1324 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1325 :
1326 : // This is the code path for geniune availability transitions (i.e node
1327 : // goes unavailable and/or comes back online).
1328 : let res = self
1329 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1330 : .await;
1331 :
1332 : match res {
1333 : Ok(transition) => {
1334 : // Keep hold of the lock until the availability transitions
1335 : // have been handled in
1336 : // [`Service::handle_node_availability_transitions`] in order avoid
1337 : // racing with [`Service::external_node_configure`].
1338 : to_handle.push((node_id, node_lock, transition));
1339 : }
1340 : Err(ApiError::NotFound(_)) => {
1341 : // This should be rare, but legitimate since the heartbeats are done
1342 : // on a snapshot of the nodes.
1343 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1344 : }
1345 : Err(ApiError::ShuttingDown) => {
1346 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1347 : }
1348 : Err(err) => {
1349 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1350 : // becomes unavailable again, we may get an error here.
1351 : tracing::error!(
1352 : "Failed to update node state {} after heartbeat round: {}",
1353 : node_id,
1354 : err
1355 : );
1356 : }
1357 : }
1358 : }
1359 :
1360 : // We collected all the transitions above and now we handle them.
1361 : let res = self.handle_node_availability_transitions(to_handle).await;
1362 : if let Err(errs) = res {
1363 : for (node_id, err) in errs {
1364 : match err {
1365 : ApiError::NotFound(_) => {
1366 : // This should be rare, but legitimate since the heartbeats are done
1367 : // on a snapshot of the nodes.
1368 : tracing::info!(
1369 : "Node {} was not found after heartbeat round",
1370 : node_id
1371 : );
1372 : }
1373 : err => {
1374 : tracing::error!(
1375 : "Failed to handle availability transition for {} after heartbeat round: {}",
1376 : node_id,
1377 : err
1378 : );
1379 : }
1380 : }
1381 : }
1382 : }
1383 : }
1384 : if let Ok(deltas) = res_sk {
1385 : let mut to_activate = Vec::new();
1386 : {
1387 : let mut locked = self.inner.write().unwrap();
1388 : let mut safekeepers = (*locked.safekeepers).clone();
1389 :
1390 : for (id, state) in deltas.0 {
1391 : let Some(sk) = safekeepers.get_mut(&id) else {
1392 : tracing::info!(
1393 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1394 : );
1395 : continue;
1396 : };
1397 : if sk.scheduling_policy() == SkSchedulingPolicy::Activating
1398 : && let SafekeeperState::Available { .. } = state
1399 : {
1400 : to_activate.push(id);
1401 : }
1402 : sk.set_availability(state);
1403 : }
1404 : locked.safekeepers = Arc::new(safekeepers);
1405 : }
1406 : for sk_id in to_activate {
1407 : // TODO this can race with set_scheduling_policy (can create disjoint DB <-> in-memory state)
1408 : tracing::info!("Activating safekeeper {sk_id}");
1409 : match self.persistence.activate_safekeeper(sk_id.0 as i64).await {
1410 : Ok(Some(())) => {}
1411 : Ok(None) => {
1412 : tracing::info!(
1413 : "safekeeper {sk_id} has been removed from db or has different scheduling policy than active or activating"
1414 : );
1415 : }
1416 : Err(e) => {
1417 : tracing::warn!("couldn't apply activation of {sk_id} to db: {e}");
1418 : continue;
1419 : }
1420 : }
1421 : if let Err(e) = self
1422 : .set_safekeeper_scheduling_policy_in_mem(sk_id, SkSchedulingPolicy::Active)
1423 : .await
1424 : {
1425 : tracing::info!("couldn't activate safekeeper {sk_id} in memory: {e}");
1426 : continue;
1427 : }
1428 : tracing::info!("Activation of safekeeper {sk_id} done");
1429 : }
1430 : }
1431 : }
1432 : }
1433 :
1434 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1435 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1436 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1437 : /// will indicate that reconciliation is not needed.
1438 : #[instrument(skip_all, fields(
1439 : seq=%result.sequence,
1440 : tenant_id=%result.tenant_shard_id.tenant_id,
1441 : shard_id=%result.tenant_shard_id.shard_slug(),
1442 : ))]
1443 : fn process_result(&self, result: ReconcileResult) {
1444 : let mut locked = self.inner.write().unwrap();
1445 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1446 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1447 : // A reconciliation result might race with removing a tenant: drop results for
1448 : // tenants that aren't in our map.
1449 : return;
1450 : };
1451 :
1452 : // Usually generation should only be updated via this path, so the max() isn't
1453 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1454 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1455 :
1456 : // If the reconciler signals that it failed to notify compute, set this state on
1457 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1458 : tenant.pending_compute_notification = result.pending_compute_notification;
1459 :
1460 : // Let the TenantShard know it is idle.
1461 : tenant.reconcile_complete(result.sequence);
1462 :
1463 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1464 : // make to the tenant
1465 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1466 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1467 : // make to the tenant
1468 0 : let node = nodes.get(delta.node_id())?;
1469 :
1470 0 : if node.is_available() {
1471 0 : return Some(delta);
1472 0 : }
1473 :
1474 : // In case a node became unavailable concurrently with the reconcile, observed
1475 : // locations on it are now uncertain. By convention, set them to None in order
1476 : // for them to get refreshed when the node comes back online.
1477 0 : Some(ObservedStateDelta::Upsert(Box::new((
1478 0 : node.get_id(),
1479 0 : ObservedStateLocation { conf: None },
1480 0 : ))))
1481 0 : });
1482 :
1483 : match result.result {
1484 : Ok(()) => {
1485 : tenant.consecutive_errors_count = 0;
1486 : tenant.apply_observed_deltas(deltas);
1487 : tenant.waiter.advance(result.sequence);
1488 : }
1489 : Err(e) => {
1490 : match e {
1491 : ReconcileError::Cancel => {
1492 : tracing::info!("Reconciler was cancelled");
1493 : }
1494 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1495 : // This might be due to the reconciler getting cancelled, or it might
1496 : // be due to the `Node` being marked offline.
1497 : tracing::info!("Reconciler cancelled during pageserver API call");
1498 : }
1499 : _ => {
1500 : tracing::warn!("Reconcile error: {}", e);
1501 : }
1502 : }
1503 :
1504 : tenant.consecutive_errors_count = tenant.consecutive_errors_count.saturating_add(1);
1505 :
1506 : // Ordering: populate last_error before advancing error_seq,
1507 : // so that waiters will see the correct error after waiting.
1508 : tenant.set_last_error(result.sequence, e);
1509 :
1510 : // Skip deletions on reconcile failures
1511 : let upsert_deltas =
1512 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1513 : tenant.apply_observed_deltas(upsert_deltas);
1514 : }
1515 : }
1516 :
1517 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1518 : if tenant.policy == PlacementPolicy::Detached {
1519 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1520 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1521 : // duration of the request.
1522 : let guard = self.tenant_op_locks.try_exclusive(
1523 : tenant.tenant_shard_id.tenant_id,
1524 : TenantOperations::DropDetached,
1525 : );
1526 : if let Some(guard) = guard {
1527 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1528 : }
1529 : }
1530 :
1531 : // Maybe some other work can proceed now that this job finished.
1532 : //
1533 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1534 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1535 : if self.reconciler_concurrency.available_permits() > 0 {
1536 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1537 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1538 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1539 : shard.delayed_reconcile = false;
1540 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1541 : }
1542 :
1543 : if self.reconciler_concurrency.available_permits() == 0 {
1544 : break;
1545 : }
1546 : }
1547 : }
1548 : }
1549 :
1550 0 : async fn process_results(
1551 0 : &self,
1552 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1553 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1554 0 : Result<(), (TenantShardId, NotifyError)>,
1555 0 : >,
1556 0 : ) {
1557 : loop {
1558 : // Wait for the next result, or for cancellation
1559 0 : tokio::select! {
1560 0 : r = result_rx.recv() => {
1561 0 : match r {
1562 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1563 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1564 : }
1565 : }
1566 0 : _ = async{
1567 0 : match bg_compute_hook_result_rx.recv().await {
1568 0 : Some(result) => {
1569 0 : if let Err((tenant_shard_id, notify_error)) = result {
1570 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1571 0 : let mut locked = self.inner.write().unwrap();
1572 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1573 0 : shard.pending_compute_notification = true;
1574 0 : }
1575 :
1576 0 : }
1577 : },
1578 : None => {
1579 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1580 0 : self.cancel.cancelled().await;
1581 : }
1582 : }
1583 0 : } => {},
1584 0 : _ = self.cancel.cancelled() => {
1585 0 : break;
1586 : }
1587 : };
1588 : }
1589 0 : }
1590 :
1591 0 : async fn process_aborts(
1592 0 : &self,
1593 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1594 0 : ) {
1595 : loop {
1596 : // Wait for the next result, or for cancellation
1597 0 : let op = tokio::select! {
1598 0 : r = abort_rx.recv() => {
1599 0 : match r {
1600 0 : Some(op) => {op},
1601 0 : None => {break;}
1602 : }
1603 : }
1604 0 : _ = self.cancel.cancelled() => {
1605 0 : break;
1606 : }
1607 : };
1608 :
1609 : // Retry until shutdown: we must keep this request object alive until it is properly
1610 : // processed, as it holds a lock guard that prevents other operations trying to do things
1611 : // to the tenant while it is in a weird part-split state.
1612 0 : while !self.reconcilers_cancel.is_cancelled() {
1613 0 : match self.abort_tenant_shard_split(&op).await {
1614 0 : Ok(_) => break,
1615 0 : Err(e) => {
1616 0 : tracing::warn!(
1617 0 : "Failed to abort shard split on {}, will retry: {e}",
1618 : op.tenant_id
1619 : );
1620 :
1621 : // If a node is unavailable, we hope that it has been properly marked Offline
1622 : // when we retry, so that the abort op will succeed. If the abort op is failing
1623 : // for some other reason, we will keep retrying forever, or until a human notices
1624 : // and does something about it (either fixing a pageserver or restarting the controller).
1625 0 : tokio::time::timeout(
1626 0 : Duration::from_secs(5),
1627 0 : self.reconcilers_cancel.cancelled(),
1628 0 : )
1629 0 : .await
1630 0 : .ok();
1631 : }
1632 : }
1633 : }
1634 : }
1635 0 : }
1636 :
1637 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1638 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1639 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1640 :
1641 0 : let leadership_cancel = CancellationToken::new();
1642 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1643 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1644 :
1645 : // Apply the migrations **after** the current leader has stepped down
1646 : // (or we've given up waiting for it), but **before** reading from the
1647 : // database. The only exception is reading the current leader before
1648 : // migrating.
1649 0 : persistence.migration_run().await?;
1650 :
1651 0 : tracing::info!("Loading nodes from database...");
1652 0 : let nodes = persistence
1653 0 : .list_nodes()
1654 0 : .await?
1655 0 : .into_iter()
1656 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1657 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1658 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1659 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1660 0 : metrics::METRICS_REGISTRY
1661 0 : .metrics_group
1662 0 : .storage_controller_pageserver_nodes
1663 0 : .set(nodes.len() as i64);
1664 0 : metrics::METRICS_REGISTRY
1665 0 : .metrics_group
1666 0 : .storage_controller_https_pageserver_nodes
1667 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
1668 :
1669 0 : tracing::info!("Loading safekeepers from database...");
1670 0 : let safekeepers = persistence
1671 0 : .list_safekeepers()
1672 0 : .await?
1673 0 : .into_iter()
1674 0 : .map(|skp| {
1675 0 : Safekeeper::from_persistence(
1676 0 : skp,
1677 0 : CancellationToken::new(),
1678 0 : config.use_https_safekeeper_api,
1679 : )
1680 0 : })
1681 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1682 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1683 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1684 0 : let count_policy = |policy| {
1685 0 : safekeepers
1686 0 : .iter()
1687 0 : .filter(|sk| sk.1.scheduling_policy() == policy)
1688 0 : .count()
1689 0 : };
1690 0 : let active_sk_count = count_policy(SkSchedulingPolicy::Active);
1691 0 : let activating_sk_count = count_policy(SkSchedulingPolicy::Activating);
1692 0 : let pause_sk_count = count_policy(SkSchedulingPolicy::Pause);
1693 0 : let decom_sk_count = count_policy(SkSchedulingPolicy::Decomissioned);
1694 0 : tracing::info!(
1695 0 : "Loaded {} safekeepers from database. Active {active_sk_count}, activating {activating_sk_count}, \
1696 0 : paused {pause_sk_count}, decomissioned {decom_sk_count}.",
1697 0 : safekeepers.len()
1698 : );
1699 0 : metrics::METRICS_REGISTRY
1700 0 : .metrics_group
1701 0 : .storage_controller_safekeeper_nodes
1702 0 : .set(safekeepers.len() as i64);
1703 0 : metrics::METRICS_REGISTRY
1704 0 : .metrics_group
1705 0 : .storage_controller_https_safekeeper_nodes
1706 0 : .set(safekeepers.values().filter(|s| s.has_https_port()).count() as i64);
1707 :
1708 0 : tracing::info!("Loading shards from database...");
1709 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1710 0 : tracing::info!(
1711 0 : "Loaded {} shards from database.",
1712 0 : tenant_shard_persistence.len()
1713 : );
1714 :
1715 : // If any shard splits were in progress, reset the database state to abort them
1716 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1717 0 : HashMap::new();
1718 0 : for tsp in &mut tenant_shard_persistence {
1719 0 : let shard = tsp.get_shard_identity()?;
1720 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1721 0 : let entry = tenant_shard_count_min_max
1722 0 : .entry(tenant_shard_id.tenant_id)
1723 0 : .or_insert_with(|| (shard.count, shard.count));
1724 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1725 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1726 : }
1727 :
1728 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1729 0 : if count_min != count_max {
1730 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1731 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1732 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1733 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1734 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1735 :
1736 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1737 : // identified this tenant has having mismatching min/max counts.
1738 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1739 :
1740 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1741 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1742 : // Set idle split state on those shards that we will retain.
1743 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1744 0 : if tsp_tenant_id == tenant_id
1745 0 : && tsp.get_shard_identity().unwrap().count == count_min
1746 0 : {
1747 0 : tsp.splitting = SplitState::Idle;
1748 0 : } else if tsp_tenant_id == tenant_id {
1749 : // Leave the splitting state on the child shards: this will be used next to
1750 : // drop them.
1751 0 : tracing::info!(
1752 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1753 : );
1754 0 : }
1755 0 : });
1756 :
1757 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1758 0 : tenant_shard_persistence.retain(|tsp| {
1759 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1760 0 : || tsp.splitting == SplitState::Idle
1761 0 : });
1762 0 : }
1763 : }
1764 :
1765 0 : let mut tenants = BTreeMap::new();
1766 :
1767 0 : let mut scheduler = Scheduler::new(nodes.values());
1768 :
1769 : #[cfg(feature = "testing")]
1770 : {
1771 : use pageserver_api::controller_api::AvailabilityZone;
1772 :
1773 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1774 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1775 : // after when pageservers start up and register.
1776 0 : let mut node_ids = HashSet::new();
1777 0 : for tsp in &tenant_shard_persistence {
1778 0 : if let Some(node_id) = tsp.generation_pageserver {
1779 0 : node_ids.insert(node_id);
1780 0 : }
1781 : }
1782 0 : for node_id in node_ids {
1783 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1784 0 : let node = Node::new(
1785 0 : NodeId(node_id as u64),
1786 0 : "".to_string(),
1787 : 123,
1788 0 : None,
1789 0 : "".to_string(),
1790 : 123,
1791 0 : None,
1792 0 : None,
1793 0 : AvailabilityZone("test_az".to_string()),
1794 : false,
1795 : )
1796 0 : .unwrap();
1797 :
1798 0 : scheduler.node_upsert(&node);
1799 : }
1800 : }
1801 0 : for tsp in tenant_shard_persistence {
1802 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1803 :
1804 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1805 : // it with what we can infer: the node for which a generation was most recently issued.
1806 0 : let mut intent = IntentState::new(
1807 0 : tsp.preferred_az_id
1808 0 : .as_ref()
1809 0 : .map(|az| AvailabilityZone(az.clone())),
1810 : );
1811 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1812 : {
1813 0 : if nodes.contains_key(&generation_pageserver) {
1814 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1815 0 : } else {
1816 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1817 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1818 : // on different pageservers.
1819 0 : tracing::warn!(
1820 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1821 : );
1822 : }
1823 0 : }
1824 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1825 :
1826 0 : tenants.insert(tenant_shard_id, new_tenant);
1827 : }
1828 :
1829 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1830 :
1831 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1832 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1833 0 : tokio::sync::mpsc::channel(512);
1834 :
1835 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1836 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1837 :
1838 0 : let cancel = CancellationToken::new();
1839 0 : let reconcilers_cancel = cancel.child_token();
1840 :
1841 0 : let mut http_client = reqwest::Client::builder();
1842 : // We intentionally disable the connection pool, so every request will create its own TCP connection.
1843 : // It's especially important for heartbeaters to notice more network problems.
1844 : //
1845 : // TODO: It makes sense to use this client only in heartbeaters and create a second one with
1846 : // connection pooling for everything else. But reqwest::Client may create a connection without
1847 : // ever using it (it uses hyper's Client under the hood):
1848 : // https://github.com/hyperium/hyper-util/blob/d51318df3461d40e5f5e5ca163cb3905ac960209/src/client/legacy/client.rs#L415
1849 : //
1850 : // Because of a bug in hyper0::Connection::graceful_shutdown such connections hang during
1851 : // graceful server shutdown: https://github.com/hyperium/hyper/issues/2730
1852 : //
1853 : // The bug has been fixed in hyper v1, so keep alive may be enabled only after we migrate to hyper1.
1854 0 : http_client = http_client.pool_max_idle_per_host(0);
1855 0 : for ssl_ca_cert in &config.ssl_ca_certs {
1856 0 : http_client = http_client.add_root_certificate(ssl_ca_cert.clone());
1857 0 : }
1858 0 : let http_client = http_client.build()?;
1859 :
1860 0 : let heartbeater_ps = Heartbeater::new(
1861 0 : http_client.clone(),
1862 0 : config.pageserver_jwt_token.clone(),
1863 0 : config.max_offline_interval,
1864 0 : config.max_warming_up_interval,
1865 0 : cancel.clone(),
1866 : );
1867 :
1868 0 : let heartbeater_sk = Heartbeater::new(
1869 0 : http_client.clone(),
1870 0 : config.safekeeper_jwt_token.clone(),
1871 0 : config.max_offline_interval,
1872 0 : config.max_warming_up_interval,
1873 0 : cancel.clone(),
1874 : );
1875 :
1876 0 : let initial_leadership_status = if config.start_as_candidate {
1877 0 : LeadershipStatus::Candidate
1878 : } else {
1879 0 : LeadershipStatus::Leader
1880 : };
1881 :
1882 0 : let this = Arc::new(Self {
1883 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1884 0 : nodes,
1885 0 : safekeepers,
1886 0 : tenants,
1887 0 : scheduler,
1888 0 : delayed_reconcile_rx,
1889 0 : initial_leadership_status,
1890 0 : reconcilers_cancel.clone(),
1891 : ))),
1892 0 : config: config.clone(),
1893 0 : persistence,
1894 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())?),
1895 0 : result_tx,
1896 0 : heartbeater_ps,
1897 0 : heartbeater_sk,
1898 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1899 0 : config.reconciler_concurrency,
1900 : )),
1901 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1902 0 : config.priority_reconciler_concurrency,
1903 : )),
1904 0 : delayed_reconcile_tx,
1905 0 : abort_tx,
1906 0 : startup_complete: startup_complete.clone(),
1907 0 : cancel,
1908 0 : reconcilers_cancel,
1909 0 : gate: Gate::default(),
1910 0 : reconcilers_gate: Gate::default(),
1911 0 : tenant_op_locks: Default::default(),
1912 0 : node_op_locks: Default::default(),
1913 0 : http_client,
1914 0 : step_down_barrier: Default::default(),
1915 : });
1916 :
1917 0 : let result_task_this = this.clone();
1918 0 : tokio::task::spawn(async move {
1919 : // Block shutdown until we're done (we must respect self.cancel)
1920 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1921 0 : result_task_this
1922 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1923 0 : .await
1924 0 : }
1925 0 : });
1926 :
1927 0 : tokio::task::spawn({
1928 0 : let this = this.clone();
1929 0 : async move {
1930 : // Block shutdown until we're done (we must respect self.cancel)
1931 0 : if let Ok(_gate) = this.gate.enter() {
1932 0 : this.process_aborts(abort_rx).await
1933 0 : }
1934 0 : }
1935 : });
1936 :
1937 0 : tokio::task::spawn({
1938 0 : let this = this.clone();
1939 0 : async move {
1940 0 : if let Ok(_gate) = this.gate.enter() {
1941 : loop {
1942 0 : tokio::select! {
1943 0 : _ = this.cancel.cancelled() => {
1944 0 : break;
1945 : },
1946 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1947 : };
1948 0 : this.tenant_op_locks.housekeeping();
1949 : }
1950 0 : }
1951 0 : }
1952 : });
1953 :
1954 0 : tokio::task::spawn({
1955 0 : let this = this.clone();
1956 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1957 : // is done.
1958 0 : let startup_completion = startup_completion.clone();
1959 0 : async move {
1960 : // Block shutdown until we're done (we must respect self.cancel)
1961 0 : let Ok(_gate) = this.gate.enter() else {
1962 0 : return;
1963 : };
1964 :
1965 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1966 0 : .await;
1967 :
1968 0 : drop(startup_completion);
1969 0 : }
1970 : });
1971 :
1972 0 : tokio::task::spawn({
1973 0 : let this = this.clone();
1974 0 : let startup_complete = startup_complete.clone();
1975 0 : async move {
1976 0 : startup_complete.wait().await;
1977 0 : this.background_reconcile().await;
1978 0 : }
1979 : });
1980 :
1981 0 : tokio::task::spawn({
1982 0 : let this = this.clone();
1983 0 : let startup_complete = startup_complete.clone();
1984 0 : async move {
1985 0 : startup_complete.wait().await;
1986 0 : this.spawn_heartbeat_driver().await;
1987 0 : }
1988 : });
1989 :
1990 : // Check that there is enough safekeepers configured that we can create new timelines
1991 0 : let test_sk_res_str = match this.safekeepers_for_new_timeline().await {
1992 0 : Ok(v) => format!("Ok({v:?})"),
1993 0 : Err(v) => format!("Err({v:})"),
1994 : };
1995 0 : tracing::info!(
1996 : timeline_safekeeper_count = config.timeline_safekeeper_count,
1997 : timelines_onto_safekeepers = config.timelines_onto_safekeepers,
1998 0 : "viability test result (test timeline creation on safekeepers): {test_sk_res_str}",
1999 : );
2000 :
2001 0 : Ok(this)
2002 0 : }
2003 :
2004 0 : pub(crate) async fn attach_hook(
2005 0 : &self,
2006 0 : attach_req: AttachHookRequest,
2007 0 : ) -> anyhow::Result<AttachHookResponse> {
2008 0 : let _tenant_lock = trace_exclusive_lock(
2009 0 : &self.tenant_op_locks,
2010 0 : attach_req.tenant_shard_id.tenant_id,
2011 0 : TenantOperations::AttachHook,
2012 0 : )
2013 0 : .await;
2014 :
2015 : // This is a test hook. To enable using it on tenants that were created directly with
2016 : // the pageserver API (not via this service), we will auto-create any missing tenant
2017 : // shards with default state.
2018 0 : let insert = {
2019 0 : match self
2020 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
2021 0 : .await
2022 : {
2023 0 : Ok(_) => false,
2024 0 : Err(ApiError::NotFound(_)) => true,
2025 0 : Err(e) => return Err(e.into()),
2026 : }
2027 : };
2028 :
2029 0 : if insert {
2030 0 : let config = attach_req.config.clone().unwrap_or_default();
2031 0 : let tsp = TenantShardPersistence {
2032 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
2033 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
2034 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
2035 0 : shard_stripe_size: 0,
2036 0 : generation: attach_req.generation_override.or(Some(0)),
2037 0 : generation_pageserver: None,
2038 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
2039 0 : config: serde_json::to_string(&config).unwrap(),
2040 0 : splitting: SplitState::default(),
2041 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2042 0 : .unwrap(),
2043 0 : preferred_az_id: None,
2044 0 : };
2045 :
2046 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
2047 0 : Err(e) => match e {
2048 : DatabaseError::Query(diesel::result::Error::DatabaseError(
2049 : DatabaseErrorKind::UniqueViolation,
2050 : _,
2051 : )) => {
2052 0 : tracing::info!(
2053 0 : "Raced with another request to insert tenant {}",
2054 : attach_req.tenant_shard_id
2055 : )
2056 : }
2057 0 : _ => return Err(e.into()),
2058 : },
2059 : Ok(()) => {
2060 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
2061 :
2062 0 : let mut shard = TenantShard::new(
2063 0 : attach_req.tenant_shard_id,
2064 0 : ShardIdentity::unsharded(),
2065 0 : PlacementPolicy::Attached(0),
2066 0 : None,
2067 : );
2068 0 : shard.config = config;
2069 :
2070 0 : let mut locked = self.inner.write().unwrap();
2071 0 : locked.tenants.insert(attach_req.tenant_shard_id, shard);
2072 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
2073 : }
2074 : }
2075 0 : }
2076 :
2077 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
2078 0 : let maybe_tenant_conf = {
2079 0 : let locked = self.inner.write().unwrap();
2080 0 : locked
2081 0 : .tenants
2082 0 : .get(&attach_req.tenant_shard_id)
2083 0 : .map(|t| t.config.clone())
2084 : };
2085 :
2086 0 : match maybe_tenant_conf {
2087 0 : Some(conf) => {
2088 0 : let new_generation = self
2089 0 : .persistence
2090 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
2091 0 : .await?;
2092 :
2093 : // Persist the placement policy update. This is required
2094 : // when we reattaching a detached tenant.
2095 0 : self.persistence
2096 0 : .update_tenant_shard(
2097 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
2098 0 : Some(PlacementPolicy::Attached(0)),
2099 0 : Some(conf),
2100 0 : None,
2101 0 : None,
2102 0 : )
2103 0 : .await?;
2104 0 : Some(new_generation)
2105 : }
2106 : None => {
2107 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
2108 : }
2109 : }
2110 : } else {
2111 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
2112 0 : None
2113 : };
2114 :
2115 0 : let mut locked = self.inner.write().unwrap();
2116 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2117 :
2118 0 : let tenant_shard = tenants
2119 0 : .get_mut(&attach_req.tenant_shard_id)
2120 0 : .expect("Checked for existence above");
2121 :
2122 0 : if let Some(new_generation) = new_generation {
2123 0 : tenant_shard.generation = Some(new_generation);
2124 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
2125 0 : } else {
2126 : // This is a detach notification. We must update placement policy to avoid re-attaching
2127 : // during background scheduling/reconciliation, or during storage controller restart.
2128 0 : assert!(attach_req.node_id.is_none());
2129 0 : tenant_shard.policy = PlacementPolicy::Detached;
2130 : }
2131 :
2132 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
2133 0 : tracing::info!(
2134 : tenant_id = %attach_req.tenant_shard_id,
2135 : ps_id = %attaching_pageserver,
2136 : generation = ?tenant_shard.generation,
2137 0 : "issuing",
2138 : );
2139 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
2140 0 : tracing::info!(
2141 : tenant_id = %attach_req.tenant_shard_id,
2142 : %ps_id,
2143 : generation = ?tenant_shard.generation,
2144 0 : "dropping",
2145 : );
2146 : } else {
2147 0 : tracing::info!(
2148 : tenant_id = %attach_req.tenant_shard_id,
2149 0 : "no-op: tenant already has no pageserver");
2150 : }
2151 0 : tenant_shard
2152 0 : .intent
2153 0 : .set_attached(scheduler, attach_req.node_id);
2154 :
2155 0 : tracing::info!(
2156 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}, config {:?}",
2157 : attach_req.tenant_shard_id,
2158 : tenant_shard.generation,
2159 : // TODO: this is an odd number of 0xf's
2160 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff)),
2161 : attach_req.config,
2162 : );
2163 :
2164 : // Trick the reconciler into not doing anything for this tenant: this helps
2165 : // tests that manually configure a tenant on the pagesrever, and then call this
2166 : // attach hook: they don't want background reconciliation to modify what they
2167 : // did to the pageserver.
2168 : #[cfg(feature = "testing")]
2169 : {
2170 0 : if let Some(node_id) = attach_req.node_id {
2171 0 : tenant_shard.observed.locations = HashMap::from([(
2172 0 : node_id,
2173 0 : ObservedStateLocation {
2174 0 : conf: Some(attached_location_conf(
2175 0 : tenant_shard.generation.unwrap(),
2176 0 : &tenant_shard.shard,
2177 0 : &tenant_shard.config,
2178 0 : &PlacementPolicy::Attached(0),
2179 0 : tenant_shard.intent.get_secondary().len(),
2180 0 : )),
2181 0 : },
2182 0 : )]);
2183 0 : } else {
2184 0 : tenant_shard.observed.locations.clear();
2185 0 : }
2186 : }
2187 :
2188 : Ok(AttachHookResponse {
2189 0 : generation: attach_req
2190 0 : .node_id
2191 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
2192 : })
2193 0 : }
2194 :
2195 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
2196 0 : let locked = self.inner.read().unwrap();
2197 :
2198 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
2199 :
2200 : InspectResponse {
2201 0 : attachment: tenant_shard.and_then(|s| {
2202 0 : s.intent
2203 0 : .get_attached()
2204 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
2205 0 : }),
2206 : }
2207 0 : }
2208 :
2209 : // When the availability state of a node transitions to active, we must do a full reconciliation
2210 : // of LocationConfigs on that node. This is because while a node was offline:
2211 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
2212 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
2213 : //
2214 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
2215 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
2216 : // this function.
2217 : //
2218 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
2219 : // for written for a single node rather than as a batch job for all nodes.
2220 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
2221 : async fn node_activate_reconcile(
2222 : &self,
2223 : mut node: Node,
2224 : _lock: &TracingExclusiveGuard<NodeOperations>,
2225 : ) -> Result<(), ApiError> {
2226 : // This Node is a mutable local copy: we will set it active so that we can use its
2227 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
2228 : // later.
2229 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
2230 :
2231 : let configs = match node
2232 : .with_client_retries(
2233 0 : |client| async move { client.list_location_config().await },
2234 : &self.http_client,
2235 : &self.config.pageserver_jwt_token,
2236 : 1,
2237 : 5,
2238 : SHORT_RECONCILE_TIMEOUT,
2239 : &self.cancel,
2240 : )
2241 : .await
2242 : {
2243 : None => {
2244 : // We're shutting down (the Node's cancellation token can't have fired, because
2245 : // we're the only scope that has a reference to it, and we didn't fire it).
2246 : return Err(ApiError::ShuttingDown);
2247 : }
2248 : Some(Err(e)) => {
2249 : // This node didn't succeed listing its locations: it may not proceed to active state
2250 : // as it is apparently unavailable.
2251 : return Err(ApiError::PreconditionFailed(
2252 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2253 : ));
2254 : }
2255 : Some(Ok(configs)) => configs,
2256 : };
2257 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2258 :
2259 : let mut cleanup = Vec::new();
2260 : let mut mismatched_locations = 0;
2261 : {
2262 : let mut locked = self.inner.write().unwrap();
2263 :
2264 : for (tenant_shard_id, reported) in configs.tenant_shards {
2265 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2266 : cleanup.push(tenant_shard_id);
2267 : continue;
2268 : };
2269 :
2270 : let on_record = &mut tenant_shard
2271 : .observed
2272 : .locations
2273 : .entry(node.get_id())
2274 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2275 : .conf;
2276 :
2277 : // If the location reported by the node does not match our observed state,
2278 : // then we mark it as uncertain and let the background reconciliation loop
2279 : // deal with it.
2280 : //
2281 : // Note that this also covers net new locations reported by the node.
2282 : if *on_record != reported {
2283 : mismatched_locations += 1;
2284 : *on_record = None;
2285 : }
2286 : }
2287 : }
2288 :
2289 : if mismatched_locations > 0 {
2290 : tracing::info!(
2291 : "Set observed state to None for {mismatched_locations} mismatched locations"
2292 : );
2293 : }
2294 :
2295 : for tenant_shard_id in cleanup {
2296 : tracing::info!("Detaching {tenant_shard_id}");
2297 : match node
2298 : .with_client_retries(
2299 0 : |client| async move {
2300 0 : let config = LocationConfig {
2301 0 : mode: LocationConfigMode::Detached,
2302 0 : generation: None,
2303 0 : secondary_conf: None,
2304 0 : shard_number: tenant_shard_id.shard_number.0,
2305 0 : shard_count: tenant_shard_id.shard_count.literal(),
2306 0 : shard_stripe_size: 0,
2307 0 : tenant_conf: models::TenantConfig::default(),
2308 0 : };
2309 0 : client
2310 0 : .location_config(tenant_shard_id, config, None, false)
2311 0 : .await
2312 0 : },
2313 : &self.http_client,
2314 : &self.config.pageserver_jwt_token,
2315 : 1,
2316 : 5,
2317 : SHORT_RECONCILE_TIMEOUT,
2318 : &self.cancel,
2319 : )
2320 : .await
2321 : {
2322 : None => {
2323 : // We're shutting down (the Node's cancellation token can't have fired, because
2324 : // we're the only scope that has a reference to it, and we didn't fire it).
2325 : return Err(ApiError::ShuttingDown);
2326 : }
2327 : Some(Err(e)) => {
2328 : // Do not let the node proceed to Active state if it is not responsive to requests
2329 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2330 : // detach completing: we should not let this node back into the set of nodes considered
2331 : // okay for scheduling.
2332 : return Err(ApiError::Conflict(format!(
2333 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2334 : )));
2335 : }
2336 : Some(Ok(_)) => {}
2337 : };
2338 : }
2339 :
2340 : Ok(())
2341 : }
2342 :
2343 0 : pub(crate) async fn re_attach(
2344 0 : &self,
2345 0 : reattach_req: ReAttachRequest,
2346 0 : ) -> Result<ReAttachResponse, ApiError> {
2347 0 : if let Some(register_req) = reattach_req.register {
2348 0 : self.node_register(register_req).await?;
2349 0 : }
2350 :
2351 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2352 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2353 :
2354 0 : tracing::info!(
2355 : node_id=%reattach_req.node_id,
2356 0 : "Incremented {} tenant shards' generations",
2357 0 : incremented_generations.len()
2358 : );
2359 :
2360 : // Apply the updated generation to our in-memory state, and
2361 : // gather discover secondary locations.
2362 0 : let mut locked = self.inner.write().unwrap();
2363 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2364 :
2365 0 : let mut response = ReAttachResponse {
2366 0 : tenants: Vec::new(),
2367 0 : };
2368 :
2369 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2370 : // to call location_conf API with an old generation. Wait for cancellation to complete
2371 : // before responding to this request. Requires well implemented CancellationToken logic
2372 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2373 : // request in flight over the network: TODO handle that by making location_conf API refuse
2374 : // to go backward in generations.
2375 :
2376 : // Scan through all shards, applying updates for ones where we updated generation
2377 : // and identifying shards that intend to have a secondary location on this node.
2378 0 : for (tenant_shard_id, shard) in tenants {
2379 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2380 0 : let new_gen = *new_gen;
2381 0 : response.tenants.push(ReAttachResponseTenant {
2382 0 : id: *tenant_shard_id,
2383 0 : r#gen: Some(new_gen.into().unwrap()),
2384 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2385 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2386 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2387 0 : // the stale/multi states at this point.
2388 0 : mode: LocationConfigMode::AttachedSingle,
2389 0 : stripe_size: shard.shard.stripe_size,
2390 0 : });
2391 :
2392 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2393 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2394 : // Why can we update `observed` even though we're not sure our response will be received
2395 : // by the pageserver? Because the pageserver will not proceed with startup until
2396 : // it has processed response: if it loses it, we'll see another request and increment
2397 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2398 0 : if let Some(conf) = observed.conf.as_mut() {
2399 0 : conf.generation = new_gen.into();
2400 0 : }
2401 0 : } else {
2402 0 : // This node has no observed state for the shard: perhaps it was offline
2403 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2404 0 : // will be prompted to learn the location's state before it makes changes.
2405 0 : shard
2406 0 : .observed
2407 0 : .locations
2408 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2409 0 : }
2410 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2411 0 : // Ordering: pageserver will not accept /location_config requests until it has
2412 0 : // finished processing the response from re-attach. So we can update our in-memory state
2413 0 : // now, and be confident that we are not stamping on the result of some later location config.
2414 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2415 0 : // so we might update observed state here, and then get over-written by some racing
2416 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2417 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2418 0 :
2419 0 : response.tenants.push(ReAttachResponseTenant {
2420 0 : id: *tenant_shard_id,
2421 0 : r#gen: None,
2422 0 : mode: LocationConfigMode::Secondary,
2423 0 : stripe_size: shard.shard.stripe_size,
2424 0 : });
2425 0 :
2426 0 : // We must not update observed, because we have no guarantee that our
2427 0 : // response will be received by the pageserver. This could leave it
2428 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2429 0 : }
2430 : }
2431 :
2432 : // We consider a node Active once we have composed a re-attach response, but we
2433 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2434 : // implicitly synchronizes the LocationConfigs on the node.
2435 : //
2436 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2437 : // but those requests will not be accepted by the node until it has finished processing
2438 : // the re-attach response.
2439 : //
2440 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2441 : // in [`Persistence::re_attach`].
2442 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2443 0 : let reset_scheduling = matches!(
2444 0 : node.get_scheduling(),
2445 : NodeSchedulingPolicy::PauseForRestart
2446 : | NodeSchedulingPolicy::Draining
2447 : | NodeSchedulingPolicy::Filling
2448 : | NodeSchedulingPolicy::Deleting
2449 : );
2450 :
2451 0 : let mut new_nodes = (**nodes).clone();
2452 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2453 0 : if reset_scheduling {
2454 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2455 0 : }
2456 :
2457 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2458 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2459 :
2460 0 : scheduler.node_upsert(node);
2461 0 : let new_nodes = Arc::new(new_nodes);
2462 0 : *nodes = new_nodes;
2463 : } else {
2464 0 : tracing::error!(
2465 0 : "Reattaching node {} was removed while processing the request",
2466 : reattach_req.node_id
2467 : );
2468 : }
2469 0 : }
2470 :
2471 0 : Ok(response)
2472 0 : }
2473 :
2474 0 : pub(crate) async fn validate(
2475 0 : &self,
2476 0 : validate_req: ValidateRequest,
2477 0 : ) -> Result<ValidateResponse, DatabaseError> {
2478 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2479 : // in-memory generation for a shard
2480 0 : let in_memory_result = {
2481 0 : let mut in_memory_result = Vec::new();
2482 0 : let locked = self.inner.read().unwrap();
2483 0 : for req_tenant in validate_req.tenants {
2484 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2485 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2486 0 : tracing::info!(
2487 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2488 : req_tenant.id,
2489 : req_tenant.r#gen,
2490 : tenant_shard.generation
2491 : );
2492 :
2493 0 : in_memory_result.push((
2494 0 : req_tenant.id,
2495 0 : Generation::new(req_tenant.r#gen),
2496 0 : valid,
2497 0 : ));
2498 : } else {
2499 : // This is legal: for example during a shard split the pageserver may still
2500 : // have deletions in its queue from the old pre-split shard, or after deletion
2501 : // of a tenant that was busy with compaction/gc while being deleted.
2502 0 : tracing::info!(
2503 0 : "Refusing deletion validation for missing shard {}",
2504 : req_tenant.id
2505 : );
2506 : }
2507 : }
2508 :
2509 0 : in_memory_result
2510 : };
2511 :
2512 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2513 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2514 0 : let db_generations = self
2515 0 : .persistence
2516 0 : .shard_generations(
2517 0 : in_memory_result
2518 0 : .iter()
2519 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2520 : )
2521 0 : .await?;
2522 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2523 :
2524 0 : let mut response = ValidateResponse {
2525 0 : tenants: Vec::new(),
2526 0 : };
2527 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2528 0 : let valid = if valid {
2529 0 : let db_generation = db_generations.get(&tenant_shard_id);
2530 0 : db_generation == Some(&Some(validate_generation))
2531 : } else {
2532 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2533 : // this prevents a pageserver from cleaning up an object in S3.
2534 0 : false
2535 : };
2536 :
2537 0 : response.tenants.push(ValidateResponseTenant {
2538 0 : id: tenant_shard_id,
2539 0 : valid,
2540 0 : })
2541 : }
2542 :
2543 0 : Ok(response)
2544 0 : }
2545 :
2546 0 : pub(crate) async fn tenant_create(
2547 0 : &self,
2548 0 : create_req: TenantCreateRequest,
2549 0 : ) -> Result<TenantCreateResponse, ApiError> {
2550 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2551 :
2552 : // Exclude any concurrent attempts to create/access the same tenant ID
2553 0 : let _tenant_lock = trace_exclusive_lock(
2554 0 : &self.tenant_op_locks,
2555 0 : create_req.new_tenant_id.tenant_id,
2556 0 : TenantOperations::Create,
2557 0 : )
2558 0 : .await;
2559 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2560 :
2561 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2562 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2563 : // accept compute notifications while it is in the process of creating. Reconciliation will
2564 : // be retried in the background.
2565 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2566 0 : }
2567 0 : Ok(response)
2568 0 : }
2569 :
2570 0 : pub(crate) async fn do_tenant_create(
2571 0 : &self,
2572 0 : create_req: TenantCreateRequest,
2573 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2574 0 : let placement_policy = create_req
2575 0 : .placement_policy
2576 0 : .clone()
2577 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2578 0 : .unwrap_or(PlacementPolicy::Attached(0));
2579 :
2580 : // This service expects to handle sharding itself: it is an error to try and directly create
2581 : // a particular shard here.
2582 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2583 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2584 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2585 0 : )));
2586 : } else {
2587 0 : create_req.new_tenant_id.tenant_id
2588 : };
2589 :
2590 0 : tracing::info!(
2591 0 : "Creating tenant {}, shard_count={:?}",
2592 : create_req.new_tenant_id,
2593 : create_req.shard_parameters.count,
2594 : );
2595 :
2596 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2597 0 : .map(|i| TenantShardId {
2598 0 : tenant_id,
2599 0 : shard_number: ShardNumber(i),
2600 0 : shard_count: create_req.shard_parameters.count,
2601 0 : })
2602 0 : .collect::<Vec<_>>();
2603 :
2604 : // If the caller specifies a None generation, it means "start from default". This is different
2605 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2606 : // an incompletely-onboarded tenant.
2607 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2608 0 : tracing::info!(
2609 0 : "tenant_create: secondary mode, generation is_some={}",
2610 0 : create_req.generation.is_some()
2611 : );
2612 0 : create_req.generation.map(Generation::new)
2613 : } else {
2614 0 : tracing::info!(
2615 0 : "tenant_create: not secondary mode, generation is_some={}",
2616 0 : create_req.generation.is_some()
2617 : );
2618 0 : Some(
2619 0 : create_req
2620 0 : .generation
2621 0 : .map(Generation::new)
2622 0 : .unwrap_or(INITIAL_GENERATION),
2623 0 : )
2624 : };
2625 :
2626 0 : let preferred_az_id = {
2627 0 : let locked = self.inner.read().unwrap();
2628 : // Idempotency: take the existing value if the tenant already exists
2629 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2630 0 : shard.preferred_az().cloned()
2631 : } else {
2632 0 : locked.scheduler.get_az_for_new_tenant()
2633 : }
2634 : };
2635 :
2636 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2637 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2638 : // during the creation, rather than risking leaving orphan objects in S3.
2639 0 : let persist_tenant_shards = create_ids
2640 0 : .iter()
2641 0 : .map(|tenant_shard_id| TenantShardPersistence {
2642 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2643 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2644 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2645 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2646 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2647 : // The pageserver is not known until scheduling happens: we will set this column when
2648 : // incrementing the generation the first time we attach to a pageserver.
2649 0 : generation_pageserver: None,
2650 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2651 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2652 0 : splitting: SplitState::default(),
2653 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2654 0 : .unwrap(),
2655 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2656 0 : })
2657 0 : .collect();
2658 :
2659 0 : match self
2660 0 : .persistence
2661 0 : .insert_tenant_shards(persist_tenant_shards)
2662 0 : .await
2663 : {
2664 0 : Ok(_) => {}
2665 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2666 : DatabaseErrorKind::UniqueViolation,
2667 : _,
2668 : ))) => {
2669 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2670 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2671 : // creation's shard count.
2672 0 : tracing::info!(
2673 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2674 : );
2675 : }
2676 : // Any other database error is unexpected and a bug.
2677 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2678 : };
2679 :
2680 0 : let mut schedule_context = ScheduleContext::default();
2681 0 : let mut schedule_error = None;
2682 0 : let mut response_shards = Vec::new();
2683 0 : for tenant_shard_id in create_ids {
2684 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2685 :
2686 0 : let outcome = self
2687 0 : .do_initial_shard_scheduling(
2688 0 : tenant_shard_id,
2689 0 : initial_generation,
2690 0 : create_req.shard_parameters,
2691 0 : create_req.config.clone(),
2692 0 : placement_policy.clone(),
2693 0 : preferred_az_id.as_ref(),
2694 0 : &mut schedule_context,
2695 0 : )
2696 0 : .await;
2697 :
2698 0 : match outcome {
2699 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2700 0 : InitialShardScheduleOutcome::NotScheduled => {}
2701 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2702 0 : schedule_error = Some(err);
2703 0 : }
2704 : }
2705 : }
2706 :
2707 : // If we failed to schedule shards, then they are still created in the controller,
2708 : // but we return an error to the requester to avoid a silent failure when someone
2709 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2710 : // are present in the system. We do this here rather than in the above loop, to
2711 : // avoid situations where we only create a subset of shards in the tenant.
2712 0 : if let Some(e) = schedule_error {
2713 0 : return Err(ApiError::Conflict(format!(
2714 0 : "Failed to schedule shard(s): {e}"
2715 0 : )));
2716 0 : }
2717 :
2718 0 : let waiters = {
2719 0 : let mut locked = self.inner.write().unwrap();
2720 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2721 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2722 0 : .tenant_creation_hint(true)
2723 0 : .build();
2724 0 : tenants
2725 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2726 0 : .filter_map(|(_shard_id, shard)| {
2727 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2728 0 : })
2729 0 : .collect::<Vec<_>>()
2730 : };
2731 :
2732 0 : Ok((
2733 0 : TenantCreateResponse {
2734 0 : shards: response_shards,
2735 0 : },
2736 0 : waiters,
2737 0 : ))
2738 0 : }
2739 :
2740 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2741 : /// case of a new tenant and a pre-existing one.
2742 : #[allow(clippy::too_many_arguments)]
2743 0 : async fn do_initial_shard_scheduling(
2744 0 : &self,
2745 0 : tenant_shard_id: TenantShardId,
2746 0 : initial_generation: Option<Generation>,
2747 0 : shard_params: ShardParameters,
2748 0 : config: TenantConfig,
2749 0 : placement_policy: PlacementPolicy,
2750 0 : preferred_az_id: Option<&AvailabilityZone>,
2751 0 : schedule_context: &mut ScheduleContext,
2752 0 : ) -> InitialShardScheduleOutcome {
2753 0 : let mut locked = self.inner.write().unwrap();
2754 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2755 :
2756 : use std::collections::btree_map::Entry;
2757 0 : match tenants.entry(tenant_shard_id) {
2758 0 : Entry::Occupied(mut entry) => {
2759 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2760 :
2761 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2762 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2763 0 : }
2764 :
2765 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2766 0 : let generation = entry
2767 0 : .get()
2768 0 : .generation
2769 0 : .expect("Generation is set when in attached mode");
2770 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2771 0 : shard_id: tenant_shard_id,
2772 0 : node_id: *node_id,
2773 0 : generation: generation.into().unwrap(),
2774 0 : })
2775 : } else {
2776 0 : InitialShardScheduleOutcome::NotScheduled
2777 : }
2778 : }
2779 0 : Entry::Vacant(entry) => {
2780 0 : let state = entry.insert(TenantShard::new(
2781 0 : tenant_shard_id,
2782 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2783 0 : placement_policy,
2784 0 : preferred_az_id.cloned(),
2785 : ));
2786 :
2787 0 : state.generation = initial_generation;
2788 0 : state.config = config;
2789 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2790 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2791 0 : }
2792 :
2793 : // Only include shards in result if we are attaching: the purpose
2794 : // of the response is to tell the caller where the shards are attached.
2795 0 : if let Some(node_id) = state.intent.get_attached() {
2796 0 : let generation = state
2797 0 : .generation
2798 0 : .expect("Generation is set when in attached mode");
2799 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2800 0 : shard_id: tenant_shard_id,
2801 0 : node_id: *node_id,
2802 0 : generation: generation.into().unwrap(),
2803 0 : })
2804 : } else {
2805 0 : InitialShardScheduleOutcome::NotScheduled
2806 : }
2807 : }
2808 : }
2809 0 : }
2810 :
2811 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2812 : /// wait for reconciliation to complete before responding.
2813 0 : async fn await_waiters(
2814 0 : &self,
2815 0 : waiters: Vec<ReconcilerWaiter>,
2816 0 : timeout: Duration,
2817 0 : ) -> Result<(), ReconcileWaitError> {
2818 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2819 0 : for waiter in waiters {
2820 0 : let timeout = deadline.duration_since(Instant::now());
2821 0 : waiter.wait_timeout(timeout).await?;
2822 : }
2823 :
2824 0 : Ok(())
2825 0 : }
2826 :
2827 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2828 : /// in progress
2829 0 : async fn await_waiters_remainder(
2830 0 : &self,
2831 0 : waiters: Vec<ReconcilerWaiter>,
2832 0 : timeout: Duration,
2833 0 : ) -> Vec<ReconcilerWaiter> {
2834 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2835 0 : for waiter in waiters.iter() {
2836 0 : let timeout = deadline.duration_since(Instant::now());
2837 0 : let _ = waiter.wait_timeout(timeout).await;
2838 : }
2839 :
2840 0 : waiters
2841 0 : .into_iter()
2842 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2843 0 : .collect::<Vec<_>>()
2844 0 : }
2845 :
2846 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2847 : /// and transform it into either a tenant creation of a series of shard updates.
2848 : ///
2849 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2850 : /// still be returned.
2851 0 : fn tenant_location_config_prepare(
2852 0 : &self,
2853 0 : tenant_id: TenantId,
2854 0 : req: TenantLocationConfigRequest,
2855 0 : ) -> TenantCreateOrUpdate {
2856 0 : let mut updates = Vec::new();
2857 0 : let mut locked = self.inner.write().unwrap();
2858 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2859 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2860 :
2861 : // Use location config mode as an indicator of policy.
2862 0 : let placement_policy = match req.config.mode {
2863 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2864 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2865 : LocationConfigMode::AttachedMulti
2866 : | LocationConfigMode::AttachedSingle
2867 : | LocationConfigMode::AttachedStale => {
2868 0 : if nodes.len() > 1 {
2869 0 : PlacementPolicy::Attached(1)
2870 : } else {
2871 : // Convenience for dev/test: if we just have one pageserver, import
2872 : // tenants into non-HA mode so that scheduling will succeed.
2873 0 : PlacementPolicy::Attached(0)
2874 : }
2875 : }
2876 : };
2877 :
2878 : // Ordinarily we do not update scheduling policy, but when making major changes
2879 : // like detaching or demoting to secondary-only, we need to force the scheduling
2880 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2881 0 : let scheduling_policy = match req.config.mode {
2882 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2883 : // Special case: when making major changes like detaching or demoting to secondary-only,
2884 : // we need to force the scheduling mode to Active, or nothing will happen.
2885 0 : Some(ShardSchedulingPolicy::Active)
2886 : }
2887 : LocationConfigMode::AttachedMulti
2888 : | LocationConfigMode::AttachedSingle
2889 : | LocationConfigMode::AttachedStale => {
2890 : // While attached, continue to respect whatever the existing scheduling mode is.
2891 0 : None
2892 : }
2893 : };
2894 :
2895 0 : let mut create = true;
2896 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2897 : // Saw an existing shard: this is not a creation
2898 0 : create = false;
2899 :
2900 : // Shards may have initially been created by a Secondary request, where we
2901 : // would have left generation as None.
2902 : //
2903 : // We only update generation the first time we see an attached-mode request,
2904 : // and if there is no existing generation set. The caller is responsible for
2905 : // ensuring that no non-storage-controller pageserver ever uses a higher
2906 : // generation than they passed in here.
2907 : use LocationConfigMode::*;
2908 0 : let set_generation = match req.config.mode {
2909 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2910 0 : req.config.generation.map(Generation::new)
2911 : }
2912 0 : _ => None,
2913 : };
2914 :
2915 0 : updates.push(ShardUpdate {
2916 0 : tenant_shard_id: *shard_id,
2917 0 : placement_policy: placement_policy.clone(),
2918 0 : tenant_config: req.config.tenant_conf.clone(),
2919 0 : generation: set_generation,
2920 0 : scheduling_policy,
2921 0 : });
2922 : }
2923 :
2924 0 : if create {
2925 : use LocationConfigMode::*;
2926 0 : let generation = match req.config.mode {
2927 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2928 : // If a caller provided a generation in a non-attached request, ignore it
2929 : // and leave our generation as None: this enables a subsequent update to set
2930 : // the generation when setting an attached mode for the first time.
2931 0 : _ => None,
2932 : };
2933 :
2934 0 : TenantCreateOrUpdate::Create(
2935 0 : // Synthesize a creation request
2936 0 : TenantCreateRequest {
2937 0 : new_tenant_id: tenant_shard_id,
2938 0 : generation,
2939 0 : shard_parameters: ShardParameters {
2940 0 : count: tenant_shard_id.shard_count,
2941 0 : // We only import un-sharded or single-sharded tenants, so stripe
2942 0 : // size can be made up arbitrarily here.
2943 0 : stripe_size: DEFAULT_STRIPE_SIZE,
2944 0 : },
2945 0 : placement_policy: Some(placement_policy),
2946 0 : config: req.config.tenant_conf,
2947 0 : },
2948 0 : )
2949 : } else {
2950 0 : assert!(!updates.is_empty());
2951 0 : TenantCreateOrUpdate::Update(updates)
2952 : }
2953 0 : }
2954 :
2955 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2956 : /// the tenant is present in memory. If not, load it from the database. If it is found
2957 : /// in neither location, return a NotFound error.
2958 : ///
2959 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2960 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2961 0 : async fn maybe_load_tenant(
2962 0 : &self,
2963 0 : tenant_id: TenantId,
2964 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2965 0 : ) -> Result<(), ApiError> {
2966 : // Check if the tenant is present in memory, and select an AZ to use when loading
2967 : // if we will load it.
2968 0 : let load_in_az = {
2969 0 : let locked = self.inner.read().unwrap();
2970 0 : let existing = locked
2971 0 : .tenants
2972 0 : .range(TenantShardId::tenant_range(tenant_id))
2973 0 : .next();
2974 :
2975 : // If the tenant is not present in memory, we expect to load it from database,
2976 : // so let's figure out what AZ to load it into while we have self.inner locked.
2977 0 : if existing.is_none() {
2978 0 : locked
2979 0 : .scheduler
2980 0 : .get_az_for_new_tenant()
2981 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2982 0 : "No AZ with nodes found to load tenant"
2983 0 : )))?
2984 : } else {
2985 : // We already have this tenant in memory
2986 0 : return Ok(());
2987 : }
2988 : };
2989 :
2990 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2991 0 : if tenant_shards.is_empty() {
2992 0 : return Err(ApiError::NotFound(
2993 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2994 0 : ));
2995 0 : }
2996 :
2997 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2998 0 : self.persistence
2999 0 : .set_tenant_shard_preferred_azs(
3000 0 : tenant_shards
3001 0 : .iter()
3002 0 : .map(|t| {
3003 0 : (
3004 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
3005 0 : Some(load_in_az.clone()),
3006 0 : )
3007 0 : })
3008 0 : .collect(),
3009 : )
3010 0 : .await?;
3011 :
3012 0 : let mut locked = self.inner.write().unwrap();
3013 0 : tracing::info!(
3014 0 : "Loaded {} shards for tenant {}",
3015 0 : tenant_shards.len(),
3016 : tenant_id
3017 : );
3018 :
3019 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
3020 0 : let intent = IntentState::new(Some(load_in_az.clone()));
3021 0 : let shard =
3022 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
3023 :
3024 : // Sanity check: when loading on-demand, we should always be loaded something Detached
3025 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
3026 0 : if shard.policy != PlacementPolicy::Detached {
3027 0 : tracing::error!(
3028 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
3029 : shard.tenant_shard_id,
3030 : shard.policy
3031 : );
3032 0 : }
3033 :
3034 0 : (shard.tenant_shard_id, shard)
3035 0 : }));
3036 :
3037 0 : Ok(())
3038 0 : }
3039 :
3040 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
3041 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
3042 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
3043 : ///
3044 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
3045 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
3046 0 : fn maybe_drop_tenant(
3047 0 : &self,
3048 0 : tenant_id: TenantId,
3049 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
3050 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
3051 0 : ) {
3052 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
3053 0 : if tenant_shards.all(|(_id, shard)| {
3054 0 : shard.policy == PlacementPolicy::Detached
3055 0 : && shard.reconciler.is_none()
3056 0 : && shard.observed.is_empty()
3057 0 : }) {
3058 0 : let keys = locked
3059 0 : .tenants
3060 0 : .range(TenantShardId::tenant_range(tenant_id))
3061 0 : .map(|(id, _)| id)
3062 0 : .copied()
3063 0 : .collect::<Vec<_>>();
3064 0 : for key in keys {
3065 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
3066 0 : locked.tenants.remove(&key);
3067 : }
3068 0 : }
3069 0 : }
3070 :
3071 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
3072 : /// directly with pageservers into this service.
3073 : ///
3074 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
3075 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
3076 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
3077 : /// tenant's source of generation numbers.
3078 : ///
3079 : /// The mode in this request coarse-grained control of tenants:
3080 : /// - Call with mode Attached* to upsert the tenant.
3081 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
3082 : /// to set an existing tenant to PolicyMode::Secondary
3083 : /// - Call with mode Detached to switch to PolicyMode::Detached
3084 0 : pub(crate) async fn tenant_location_config(
3085 0 : &self,
3086 0 : tenant_shard_id: TenantShardId,
3087 0 : req: TenantLocationConfigRequest,
3088 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
3089 : // We require an exclusive lock, because we are updating both persistent and in-memory state
3090 0 : let _tenant_lock = trace_exclusive_lock(
3091 0 : &self.tenant_op_locks,
3092 0 : tenant_shard_id.tenant_id,
3093 0 : TenantOperations::LocationConfig,
3094 0 : )
3095 0 : .await;
3096 :
3097 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
3098 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3099 0 : "This API is for importing single-sharded or unsharded tenants"
3100 0 : )));
3101 : } else {
3102 0 : tenant_shard_id.tenant_id
3103 : };
3104 :
3105 : // In case we are waking up a Detached tenant
3106 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
3107 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
3108 0 : // This is a creation or an update
3109 0 : }
3110 0 : Err(e) => {
3111 0 : return Err(e);
3112 : }
3113 : };
3114 :
3115 : // First check if this is a creation or an update
3116 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
3117 :
3118 0 : let mut result = TenantLocationConfigResponse {
3119 0 : shards: Vec::new(),
3120 0 : stripe_size: None,
3121 0 : };
3122 0 : let waiters = match create_or_update {
3123 0 : TenantCreateOrUpdate::Create(create_req) => {
3124 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
3125 0 : result.shards = create_resp
3126 0 : .shards
3127 0 : .into_iter()
3128 0 : .map(|s| TenantShardLocation {
3129 0 : node_id: s.node_id,
3130 0 : shard_id: s.shard_id,
3131 0 : })
3132 0 : .collect();
3133 0 : waiters
3134 : }
3135 0 : TenantCreateOrUpdate::Update(updates) => {
3136 : // Persist updates
3137 : // Ordering: write to the database before applying changes in-memory, so that
3138 : // we will not appear time-travel backwards on a restart.
3139 :
3140 0 : let mut schedule_context = ScheduleContext::default();
3141 : for ShardUpdate {
3142 0 : tenant_shard_id,
3143 0 : placement_policy,
3144 0 : tenant_config,
3145 0 : generation,
3146 0 : scheduling_policy,
3147 0 : } in &updates
3148 : {
3149 0 : self.persistence
3150 0 : .update_tenant_shard(
3151 0 : TenantFilter::Shard(*tenant_shard_id),
3152 0 : Some(placement_policy.clone()),
3153 0 : Some(tenant_config.clone()),
3154 0 : *generation,
3155 0 : *scheduling_policy,
3156 0 : )
3157 0 : .await?;
3158 : }
3159 :
3160 : // Apply updates in-memory
3161 0 : let mut waiters = Vec::new();
3162 : {
3163 0 : let mut locked = self.inner.write().unwrap();
3164 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3165 :
3166 : for ShardUpdate {
3167 0 : tenant_shard_id,
3168 0 : placement_policy,
3169 0 : tenant_config,
3170 0 : generation: update_generation,
3171 0 : scheduling_policy,
3172 0 : } in updates
3173 : {
3174 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
3175 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
3176 0 : continue;
3177 : };
3178 :
3179 : // Update stripe size
3180 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
3181 0 : result.stripe_size = Some(shard.shard.stripe_size);
3182 0 : }
3183 :
3184 0 : shard.policy = placement_policy;
3185 0 : shard.config = tenant_config;
3186 0 : if let Some(generation) = update_generation {
3187 0 : shard.generation = Some(generation);
3188 0 : }
3189 :
3190 0 : if let Some(scheduling_policy) = scheduling_policy {
3191 0 : shard.set_scheduling_policy(scheduling_policy);
3192 0 : }
3193 :
3194 0 : shard.schedule(scheduler, &mut schedule_context)?;
3195 :
3196 0 : let maybe_waiter =
3197 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3198 0 : if let Some(waiter) = maybe_waiter {
3199 0 : waiters.push(waiter);
3200 0 : }
3201 :
3202 0 : if let Some(node_id) = shard.intent.get_attached() {
3203 0 : result.shards.push(TenantShardLocation {
3204 0 : shard_id: tenant_shard_id,
3205 0 : node_id: *node_id,
3206 0 : })
3207 0 : }
3208 : }
3209 : }
3210 0 : waiters
3211 : }
3212 : };
3213 :
3214 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3215 : // Do not treat a reconcile error as fatal: we have already applied any requested
3216 : // Intent changes, and the reconcile can fail for external reasons like unavailable
3217 : // compute notification API. In these cases, it is important that we do not
3218 : // cause the cloud control plane to retry forever on this API.
3219 0 : tracing::warn!(
3220 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
3221 : );
3222 0 : }
3223 :
3224 : // Logging the full result is useful because it lets us cross-check what the cloud control
3225 : // plane's tenant_shards table should contain.
3226 0 : tracing::info!("Complete, returning {result:?}");
3227 :
3228 0 : Ok(result)
3229 0 : }
3230 :
3231 0 : pub(crate) async fn tenant_config_patch(
3232 0 : &self,
3233 0 : req: TenantConfigPatchRequest,
3234 0 : ) -> Result<(), ApiError> {
3235 0 : let _tenant_lock = trace_exclusive_lock(
3236 0 : &self.tenant_op_locks,
3237 0 : req.tenant_id,
3238 0 : TenantOperations::ConfigPatch,
3239 0 : )
3240 0 : .await;
3241 :
3242 0 : let tenant_id = req.tenant_id;
3243 0 : let patch = req.config;
3244 :
3245 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3246 :
3247 0 : let base = {
3248 0 : let locked = self.inner.read().unwrap();
3249 0 : let shards = locked
3250 0 : .tenants
3251 0 : .range(TenantShardId::tenant_range(req.tenant_id));
3252 :
3253 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
3254 :
3255 0 : let first = match configs.peek() {
3256 0 : Some(first) => (*first).clone(),
3257 : None => {
3258 0 : return Err(ApiError::NotFound(
3259 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3260 0 : ));
3261 : }
3262 : };
3263 :
3264 0 : if !configs.all_equal() {
3265 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3266 : // This can't happen because we atomically update the database records
3267 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3268 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3269 0 : "Tenant configs for {} are mismatched",
3270 0 : req.tenant_id
3271 0 : )));
3272 0 : }
3273 :
3274 0 : first
3275 : };
3276 :
3277 0 : let updated_config = base
3278 0 : .apply_patch(patch)
3279 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3280 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3281 0 : .await
3282 0 : }
3283 :
3284 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3285 : // We require an exclusive lock, because we are updating persistent and in-memory state
3286 0 : let _tenant_lock = trace_exclusive_lock(
3287 0 : &self.tenant_op_locks,
3288 0 : req.tenant_id,
3289 0 : TenantOperations::ConfigSet,
3290 0 : )
3291 0 : .await;
3292 :
3293 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3294 :
3295 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3296 0 : .await
3297 0 : }
3298 :
3299 0 : async fn set_tenant_config_and_reconcile(
3300 0 : &self,
3301 0 : tenant_id: TenantId,
3302 0 : config: TenantConfig,
3303 0 : ) -> Result<(), ApiError> {
3304 0 : self.persistence
3305 0 : .update_tenant_shard(
3306 0 : TenantFilter::Tenant(tenant_id),
3307 0 : None,
3308 0 : Some(config.clone()),
3309 0 : None,
3310 0 : None,
3311 0 : )
3312 0 : .await?;
3313 :
3314 0 : let waiters = {
3315 0 : let mut waiters = Vec::new();
3316 0 : let mut locked = self.inner.write().unwrap();
3317 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3318 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3319 0 : shard.config = config.clone();
3320 0 : if let Some(waiter) =
3321 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3322 0 : {
3323 0 : waiters.push(waiter);
3324 0 : }
3325 : }
3326 0 : waiters
3327 : };
3328 :
3329 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3330 : // Treat this as success because we have stored the configuration. If e.g.
3331 : // a node was unavailable at this time, it should not stop us accepting a
3332 : // configuration change.
3333 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3334 0 : }
3335 :
3336 0 : Ok(())
3337 0 : }
3338 :
3339 0 : pub(crate) fn tenant_config_get(
3340 0 : &self,
3341 0 : tenant_id: TenantId,
3342 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3343 0 : let config = {
3344 0 : let locked = self.inner.read().unwrap();
3345 :
3346 0 : match locked
3347 0 : .tenants
3348 0 : .range(TenantShardId::tenant_range(tenant_id))
3349 0 : .next()
3350 : {
3351 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3352 : None => {
3353 0 : return Err(ApiError::NotFound(
3354 0 : anyhow::anyhow!("Tenant not found").into(),
3355 0 : ));
3356 : }
3357 : }
3358 : };
3359 :
3360 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3361 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3362 : // and `effective_config` in the response is meaningless, but we retain that syntax
3363 : // in order to remain compatible with the pageserver API.
3364 :
3365 0 : let response = HashMap::from([
3366 : (
3367 : "tenant_specific_overrides",
3368 0 : serde_json::to_value(&config)
3369 0 : .context("serializing tenant specific overrides")
3370 0 : .map_err(ApiError::InternalServerError)?,
3371 : ),
3372 : (
3373 0 : "effective_config",
3374 0 : serde_json::to_value(&config)
3375 0 : .context("serializing effective config")
3376 0 : .map_err(ApiError::InternalServerError)?,
3377 : ),
3378 : ]);
3379 :
3380 0 : Ok(response)
3381 0 : }
3382 :
3383 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3384 0 : &self,
3385 0 : time_travel_req: &TenantTimeTravelRequest,
3386 0 : tenant_id: TenantId,
3387 0 : timestamp: Cow<'_, str>,
3388 0 : done_if_after: Cow<'_, str>,
3389 0 : ) -> Result<(), ApiError> {
3390 0 : let _tenant_lock = trace_exclusive_lock(
3391 0 : &self.tenant_op_locks,
3392 0 : tenant_id,
3393 0 : TenantOperations::TimeTravelRemoteStorage,
3394 0 : )
3395 0 : .await;
3396 :
3397 0 : let node = {
3398 0 : let mut locked = self.inner.write().unwrap();
3399 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3400 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3401 : // but only at the start of the process, so it's really just to prevent operator
3402 : // mistakes.
3403 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3404 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3405 : {
3406 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3407 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3408 0 : )));
3409 0 : }
3410 0 : let maybe_attached = shard
3411 0 : .observed
3412 0 : .locations
3413 0 : .iter()
3414 0 : .filter_map(|(node_id, observed_location)| {
3415 0 : observed_location
3416 0 : .conf
3417 0 : .as_ref()
3418 0 : .map(|loc| (node_id, observed_location, loc.mode))
3419 0 : })
3420 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3421 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3422 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3423 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3424 0 : )));
3425 0 : }
3426 : }
3427 0 : let scheduler = &mut locked.scheduler;
3428 : // Right now we only perform the operation on a single node without parallelization
3429 : // TODO fan out the operation to multiple nodes for better performance
3430 0 : let node_id = scheduler.any_available_node()?;
3431 0 : let node = locked
3432 0 : .nodes
3433 0 : .get(&node_id)
3434 0 : .expect("Pageservers may not be deleted while lock is active");
3435 0 : node.clone()
3436 : };
3437 :
3438 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3439 0 : let mut counts = time_travel_req
3440 0 : .shard_counts
3441 0 : .iter()
3442 0 : .copied()
3443 0 : .collect::<HashSet<_>>()
3444 0 : .into_iter()
3445 0 : .collect::<Vec<_>>();
3446 0 : counts.sort_unstable();
3447 :
3448 0 : for count in counts {
3449 0 : let shard_ids = (0..count.count())
3450 0 : .map(|i| TenantShardId {
3451 0 : tenant_id,
3452 0 : shard_number: ShardNumber(i),
3453 0 : shard_count: count,
3454 0 : })
3455 0 : .collect::<Vec<_>>();
3456 0 : for tenant_shard_id in shard_ids {
3457 0 : let client = PageserverClient::new(
3458 0 : node.get_id(),
3459 0 : self.http_client.clone(),
3460 0 : node.base_url(),
3461 0 : self.config.pageserver_jwt_token.as_deref(),
3462 : );
3463 :
3464 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3465 :
3466 0 : client
3467 0 : .tenant_time_travel_remote_storage(
3468 0 : tenant_shard_id,
3469 0 : ×tamp,
3470 0 : &done_if_after,
3471 0 : )
3472 0 : .await
3473 0 : .map_err(|e| {
3474 0 : ApiError::InternalServerError(anyhow::anyhow!(
3475 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3476 0 : node
3477 0 : ))
3478 0 : })?;
3479 : }
3480 : }
3481 0 : Ok(())
3482 0 : }
3483 :
3484 0 : pub(crate) async fn tenant_secondary_download(
3485 0 : &self,
3486 0 : tenant_id: TenantId,
3487 0 : wait: Option<Duration>,
3488 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3489 0 : let _tenant_lock = trace_shared_lock(
3490 0 : &self.tenant_op_locks,
3491 0 : tenant_id,
3492 0 : TenantOperations::SecondaryDownload,
3493 0 : )
3494 0 : .await;
3495 :
3496 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3497 0 : let targets = {
3498 0 : let locked = self.inner.read().unwrap();
3499 0 : let mut targets = Vec::new();
3500 :
3501 0 : for (tenant_shard_id, shard) in
3502 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3503 : {
3504 0 : for node_id in shard.intent.get_secondary() {
3505 0 : let node = locked
3506 0 : .nodes
3507 0 : .get(node_id)
3508 0 : .expect("Pageservers may not be deleted while referenced");
3509 0 :
3510 0 : targets.push((*tenant_shard_id, node.clone()));
3511 0 : }
3512 : }
3513 0 : targets
3514 : };
3515 :
3516 : // Issue concurrent requests to all shards' locations
3517 0 : let mut futs = FuturesUnordered::new();
3518 0 : for (tenant_shard_id, node) in targets {
3519 0 : let client = PageserverClient::new(
3520 0 : node.get_id(),
3521 0 : self.http_client.clone(),
3522 0 : node.base_url(),
3523 0 : self.config.pageserver_jwt_token.as_deref(),
3524 : );
3525 0 : futs.push(async move {
3526 0 : let result = client
3527 0 : .tenant_secondary_download(tenant_shard_id, wait)
3528 0 : .await;
3529 0 : (result, node, tenant_shard_id)
3530 0 : })
3531 : }
3532 :
3533 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3534 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3535 : // well as more general cases like 503s, 500s, or timeouts.
3536 0 : let mut aggregate_progress = SecondaryProgress::default();
3537 0 : let mut aggregate_status: Option<StatusCode> = None;
3538 0 : let mut error: Option<mgmt_api::Error> = None;
3539 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3540 0 : match result {
3541 0 : Err(e) => {
3542 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3543 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3544 : // than they had hoped for.
3545 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3546 0 : error = Some(e)
3547 : }
3548 0 : Ok((status_code, progress)) => {
3549 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3550 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3551 0 : aggregate_progress.layers_total += progress.layers_total;
3552 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3553 0 : aggregate_progress.bytes_total += progress.bytes_total;
3554 0 : aggregate_progress.heatmap_mtime =
3555 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3556 0 : aggregate_status = match aggregate_status {
3557 0 : None => Some(status_code),
3558 0 : Some(StatusCode::OK) => Some(status_code),
3559 0 : Some(cur) => {
3560 : // Other status codes (e.g. 202) -- do not overwrite.
3561 0 : Some(cur)
3562 : }
3563 : };
3564 : }
3565 : }
3566 : }
3567 :
3568 : // If any of the shards return 202, indicate our result as 202.
3569 0 : match aggregate_status {
3570 : None => {
3571 0 : match error {
3572 0 : Some(e) => {
3573 : // No successes, and an error: surface it
3574 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3575 : }
3576 : None => {
3577 : // No shards found
3578 0 : Err(ApiError::NotFound(
3579 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3580 0 : ))
3581 : }
3582 : }
3583 : }
3584 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3585 : }
3586 0 : }
3587 :
3588 0 : pub(crate) async fn tenant_delete(
3589 0 : self: &Arc<Self>,
3590 0 : tenant_id: TenantId,
3591 0 : ) -> Result<StatusCode, ApiError> {
3592 0 : let _tenant_lock =
3593 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3594 :
3595 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3596 :
3597 : // Detach all shards. This also deletes local pageserver shard data.
3598 0 : let (detach_waiters, node) = {
3599 0 : let mut detach_waiters = Vec::new();
3600 0 : let mut locked = self.inner.write().unwrap();
3601 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3602 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3603 : // Update the tenant's intent to remove all attachments
3604 0 : shard.policy = PlacementPolicy::Detached;
3605 0 : shard
3606 0 : .schedule(scheduler, &mut ScheduleContext::default())
3607 0 : .expect("De-scheduling is infallible");
3608 0 : debug_assert!(shard.intent.get_attached().is_none());
3609 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3610 :
3611 0 : if let Some(waiter) =
3612 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3613 0 : {
3614 0 : detach_waiters.push(waiter);
3615 0 : }
3616 : }
3617 :
3618 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3619 : // was attached, just has to be able to see the S3 content)
3620 0 : let node_id = scheduler.any_available_node()?;
3621 0 : let node = nodes
3622 0 : .get(&node_id)
3623 0 : .expect("Pageservers may not be deleted while lock is active");
3624 0 : (detach_waiters, node.clone())
3625 : };
3626 :
3627 : // This reconcile wait can fail in a few ways:
3628 : // A there is a very long queue for the reconciler semaphore
3629 : // B some pageserver is failing to handle a detach promptly
3630 : // C some pageserver goes offline right at the moment we send it a request.
3631 : //
3632 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3633 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3634 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3635 : // deleting the underlying data).
3636 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3637 0 : .await?;
3638 :
3639 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3640 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3641 : // the tenant ID prefix, including all shards (even possibly stale ones).
3642 0 : match node
3643 0 : .with_client_retries(
3644 0 : |client| async move {
3645 0 : client
3646 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3647 0 : .await
3648 0 : },
3649 0 : &self.http_client,
3650 0 : &self.config.pageserver_jwt_token,
3651 : 1,
3652 : 3,
3653 : RECONCILE_TIMEOUT,
3654 0 : &self.cancel,
3655 : )
3656 0 : .await
3657 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3658 : {
3659 0 : Ok(_) => {}
3660 : Err(mgmt_api::Error::Cancelled) => {
3661 0 : return Err(ApiError::ShuttingDown);
3662 : }
3663 0 : Err(e) => {
3664 : // This is unexpected: remote deletion should be infallible, unless the object store
3665 : // at large is unavailable.
3666 0 : tracing::error!("Error deleting via node {node}: {e}");
3667 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3668 : }
3669 : }
3670 :
3671 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3672 : // our in-memory state and database state.
3673 :
3674 : // Ordering: we delete persistent state first: if we then
3675 : // crash, we will drop the in-memory state.
3676 :
3677 : // Drop persistent state.
3678 0 : self.persistence.delete_tenant(tenant_id).await?;
3679 :
3680 : // Drop in-memory state
3681 : {
3682 0 : let mut locked = self.inner.write().unwrap();
3683 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3684 :
3685 : // Dereference Scheduler from shards before dropping them
3686 0 : for (_tenant_shard_id, shard) in
3687 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3688 0 : {
3689 0 : shard.intent.clear(scheduler);
3690 0 : }
3691 :
3692 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3693 0 : tracing::info!(
3694 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3695 0 : locked.tenants.len()
3696 : );
3697 : };
3698 :
3699 : // Delete the tenant from safekeepers (if needed)
3700 0 : self.tenant_delete_safekeepers(tenant_id)
3701 0 : .instrument(tracing::info_span!("tenant_delete_safekeepers", %tenant_id))
3702 0 : .await?;
3703 :
3704 : // Success is represented as 404, to imitate the existing pageserver deletion API
3705 0 : Ok(StatusCode::NOT_FOUND)
3706 0 : }
3707 :
3708 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3709 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3710 : /// the tenant's policies (configuration) within the storage controller
3711 0 : pub(crate) async fn tenant_update_policy(
3712 0 : &self,
3713 0 : tenant_id: TenantId,
3714 0 : req: TenantPolicyRequest,
3715 0 : ) -> Result<(), ApiError> {
3716 : // We require an exclusive lock, because we are updating persistent and in-memory state
3717 0 : let _tenant_lock = trace_exclusive_lock(
3718 0 : &self.tenant_op_locks,
3719 0 : tenant_id,
3720 0 : TenantOperations::UpdatePolicy,
3721 0 : )
3722 0 : .await;
3723 :
3724 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3725 :
3726 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3727 :
3728 : let TenantPolicyRequest {
3729 0 : placement,
3730 0 : mut scheduling,
3731 0 : } = req;
3732 :
3733 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3734 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3735 : // this to take effect.
3736 0 : if scheduling.is_none() {
3737 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3738 0 : }
3739 0 : }
3740 :
3741 0 : self.persistence
3742 0 : .update_tenant_shard(
3743 0 : TenantFilter::Tenant(tenant_id),
3744 0 : placement.clone(),
3745 0 : None,
3746 0 : None,
3747 0 : scheduling,
3748 0 : )
3749 0 : .await?;
3750 :
3751 0 : let mut schedule_context = ScheduleContext::default();
3752 0 : let mut locked = self.inner.write().unwrap();
3753 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3754 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3755 0 : if let Some(placement) = &placement {
3756 0 : shard.policy = placement.clone();
3757 :
3758 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3759 0 : "Updated placement policy to {placement:?}");
3760 0 : }
3761 :
3762 0 : if let Some(scheduling) = &scheduling {
3763 0 : shard.set_scheduling_policy(*scheduling);
3764 :
3765 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3766 0 : "Updated scheduling policy to {scheduling:?}");
3767 0 : }
3768 :
3769 : // In case scheduling is being switched back on, try it now.
3770 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3771 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3772 : }
3773 :
3774 0 : Ok(())
3775 0 : }
3776 :
3777 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3778 0 : &self,
3779 0 : tenant_id: TenantId,
3780 0 : mut create_req: TimelineCreateRequest,
3781 0 : ) -> Result<TimelineInfo, ApiError> {
3782 0 : tracing::info!(
3783 0 : "Creating timeline {}/{}",
3784 : tenant_id,
3785 : create_req.new_timeline_id,
3786 : );
3787 :
3788 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3789 0 : if targets.0.is_empty() {
3790 0 : return Err(ApiError::NotFound(
3791 0 : anyhow::anyhow!("Tenant not found").into(),
3792 0 : ));
3793 0 : };
3794 :
3795 0 : let (shard_zero_tid, shard_zero_locations) =
3796 0 : targets.0.pop_first().expect("Must have at least one shard");
3797 0 : assert!(shard_zero_tid.is_shard_zero());
3798 :
3799 0 : async fn create_one(
3800 0 : tenant_shard_id: TenantShardId,
3801 0 : locations: ShardMutationLocations,
3802 0 : http_client: reqwest::Client,
3803 0 : jwt: Option<String>,
3804 0 : mut create_req: TimelineCreateRequest,
3805 0 : ) -> Result<TimelineInfo, ApiError> {
3806 0 : let latest = locations.latest.node;
3807 :
3808 0 : tracing::info!(
3809 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3810 : tenant_shard_id,
3811 : create_req.new_timeline_id,
3812 : locations.latest.generation
3813 : );
3814 :
3815 0 : let client =
3816 0 : PageserverClient::new(latest.get_id(), http_client.clone(), latest.base_url(), jwt.as_deref());
3817 :
3818 0 : let timeline_info = client
3819 0 : .timeline_create(tenant_shard_id, &create_req)
3820 0 : .await
3821 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3822 :
3823 : // If we are going to create the timeline on some stale locations for shard 0, then ask them to re-use
3824 : // the initdb generated by the latest location, rather than generating their own. This avoids racing uploads
3825 : // of initdb to S3 which might not be binary-identical if different pageservers have different postgres binaries.
3826 0 : if tenant_shard_id.is_shard_zero() {
3827 0 : if let models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } = &mut create_req.mode {
3828 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id);
3829 0 : }
3830 0 : }
3831 :
3832 : // We propagate timeline creations to all attached locations such that a compute
3833 : // for the new timeline is able to start regardless of the current state of the
3834 : // tenant shard reconciliation.
3835 0 : for location in locations.other {
3836 0 : tracing::info!(
3837 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3838 : tenant_shard_id,
3839 : create_req.new_timeline_id,
3840 : location.node,
3841 : location.generation
3842 : );
3843 :
3844 0 : let client = PageserverClient::new(
3845 0 : location.node.get_id(),
3846 0 : http_client.clone(),
3847 0 : location.node.base_url(),
3848 0 : jwt.as_deref(),
3849 : );
3850 :
3851 0 : let res = client
3852 0 : .timeline_create(tenant_shard_id, &create_req)
3853 0 : .await;
3854 :
3855 0 : if let Err(e) = res {
3856 0 : match e {
3857 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3858 0 : // Tenant might have been detached from the stale location,
3859 0 : // so ignore 404s.
3860 0 : },
3861 : _ => {
3862 0 : return Err(passthrough_api_error(&location.node, e));
3863 : }
3864 : }
3865 0 : }
3866 : }
3867 :
3868 0 : Ok(timeline_info)
3869 0 : }
3870 :
3871 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3872 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3873 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3874 : //
3875 : // This also enables non-zero shards to use the initdb that shard 0 generated and uploaded to S3, rather than
3876 : // independently generating their own initdb. This guarantees that shards cannot end up with different initial
3877 : // states if e.g. they have different postgres binary versions.
3878 0 : let timeline_info = create_one(
3879 0 : shard_zero_tid,
3880 0 : shard_zero_locations,
3881 0 : self.http_client.clone(),
3882 0 : self.config.pageserver_jwt_token.clone(),
3883 0 : create_req.clone(),
3884 0 : )
3885 0 : .await?;
3886 :
3887 : // Update the create request for shards >= 0
3888 0 : match &mut create_req.mode {
3889 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3890 0 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3891 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3892 0 : },
3893 0 : models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } => {
3894 : // For shards >= 0, do not run initdb: use the one that shard 0 uploaded to S3
3895 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id)
3896 : }
3897 0 : _ => {}
3898 : }
3899 :
3900 : // Create timeline on remaining shards with number >0
3901 0 : if !targets.0.is_empty() {
3902 : // If we had multiple shards, issue requests for the remainder now.
3903 0 : let jwt = &self.config.pageserver_jwt_token;
3904 0 : self.tenant_for_shards(
3905 0 : targets
3906 0 : .0
3907 0 : .iter()
3908 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3909 0 : .collect(),
3910 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3911 0 : let create_req = create_req.clone();
3912 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3913 0 : Box::pin(create_one(
3914 0 : tenant_shard_id,
3915 0 : mutation_locations,
3916 0 : self.http_client.clone(),
3917 0 : jwt.clone(),
3918 0 : create_req,
3919 0 : ))
3920 0 : },
3921 : )
3922 0 : .await?;
3923 0 : }
3924 :
3925 0 : Ok(timeline_info)
3926 0 : })
3927 0 : .await?
3928 0 : }
3929 :
3930 0 : pub(crate) async fn tenant_timeline_create(
3931 0 : self: &Arc<Self>,
3932 0 : tenant_id: TenantId,
3933 0 : create_req: TimelineCreateRequest,
3934 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
3935 0 : let safekeepers = self.config.timelines_onto_safekeepers;
3936 0 : let timeline_id = create_req.new_timeline_id;
3937 :
3938 0 : tracing::info!(
3939 0 : mode=%create_req.mode_tag(),
3940 : %safekeepers,
3941 0 : "Creating timeline {}/{}",
3942 : tenant_id,
3943 : timeline_id,
3944 : );
3945 :
3946 0 : let _tenant_lock = trace_shared_lock(
3947 0 : &self.tenant_op_locks,
3948 0 : tenant_id,
3949 0 : TenantOperations::TimelineCreate,
3950 0 : )
3951 0 : .await;
3952 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3953 0 : let is_import = create_req.is_import();
3954 0 : let read_only = matches!(
3955 0 : create_req.mode,
3956 : models::TimelineCreateRequestMode::Branch {
3957 : read_only: true,
3958 : ..
3959 : }
3960 : );
3961 :
3962 0 : if is_import {
3963 : // Ensure that there is no split on-going.
3964 : // [`Self::tenant_shard_split`] holds the exclusive tenant lock
3965 : // for the duration of the split, but here we handle the case
3966 : // where we restarted and the split is being aborted.
3967 0 : let locked = self.inner.read().unwrap();
3968 0 : let splitting = locked
3969 0 : .tenants
3970 0 : .range(TenantShardId::tenant_range(tenant_id))
3971 0 : .any(|(_id, shard)| shard.splitting != SplitState::Idle);
3972 :
3973 0 : if splitting {
3974 0 : return Err(ApiError::Conflict("Tenant is splitting shard".to_string()));
3975 0 : }
3976 0 : }
3977 :
3978 0 : let timeline_info = self
3979 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
3980 0 : .await?;
3981 :
3982 0 : let selected_safekeepers = if is_import {
3983 0 : let shards = {
3984 0 : let locked = self.inner.read().unwrap();
3985 0 : locked
3986 0 : .tenants
3987 0 : .range(TenantShardId::tenant_range(tenant_id))
3988 0 : .map(|(ts_id, _)| ts_id.to_index())
3989 0 : .collect::<Vec<_>>()
3990 : };
3991 :
3992 0 : if !shards
3993 0 : .iter()
3994 0 : .map(|shard_index| shard_index.shard_count)
3995 0 : .all_equal()
3996 : {
3997 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3998 0 : "Inconsistent shard count"
3999 0 : )));
4000 0 : }
4001 :
4002 0 : let import = TimelineImport {
4003 0 : tenant_id,
4004 0 : timeline_id,
4005 0 : shard_statuses: ShardImportStatuses::new(shards),
4006 0 : };
4007 :
4008 0 : let inserted = self
4009 0 : .persistence
4010 0 : .insert_timeline_import(import.to_persistent())
4011 0 : .await
4012 0 : .context("timeline import insert")
4013 0 : .map_err(ApiError::InternalServerError)?;
4014 :
4015 : // Set the importing flag on the tenant shards
4016 0 : self.inner
4017 0 : .write()
4018 0 : .unwrap()
4019 0 : .tenants
4020 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4021 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Importing);
4022 :
4023 0 : match inserted {
4024 : true => {
4025 0 : tracing::info!(%tenant_id, %timeline_id, "Inserted timeline import");
4026 : }
4027 : false => {
4028 0 : tracing::info!(%tenant_id, %timeline_id, "Timeline import entry already present");
4029 : }
4030 : }
4031 :
4032 0 : None
4033 0 : } else if safekeepers || read_only {
4034 : // Note that for imported timelines, we do not create the timeline on the safekeepers
4035 : // straight away. Instead, we do it once the import finalized such that we know what
4036 : // start LSN to provide for the safekeepers. This is done in
4037 : // [`Self::finalize_timeline_import`].
4038 0 : let res = self
4039 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
4040 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
4041 0 : .await?;
4042 0 : Some(res)
4043 : } else {
4044 0 : None
4045 : };
4046 :
4047 0 : Ok(TimelineCreateResponseStorcon {
4048 0 : timeline_info,
4049 0 : safekeepers: selected_safekeepers,
4050 0 : })
4051 0 : }
4052 :
4053 : #[instrument(skip_all, fields(
4054 : tenant_id=%req.tenant_shard_id.tenant_id,
4055 : shard_id=%req.tenant_shard_id.shard_slug(),
4056 : timeline_id=%req.timeline_id,
4057 : ))]
4058 : pub(crate) async fn handle_timeline_shard_import_progress(
4059 : self: &Arc<Self>,
4060 : req: TimelineImportStatusRequest,
4061 : ) -> Result<ShardImportStatus, ApiError> {
4062 : let validity = self
4063 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4064 : .await?;
4065 : match validity {
4066 : ShardGenerationValidity::Valid => {
4067 : // fallthrough
4068 : }
4069 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4070 : tracing::info!(
4071 : claimed=?claimed.into(),
4072 0 : actual=?actual.and_then(|g| g.into()),
4073 : "Rejecting import progress fetch from stale generation"
4074 : );
4075 :
4076 : return Err(ApiError::BadRequest(anyhow::anyhow!("Invalid generation")));
4077 : }
4078 : }
4079 :
4080 : let maybe_import = self
4081 : .persistence
4082 : .get_timeline_import(req.tenant_shard_id.tenant_id, req.timeline_id)
4083 : .await?;
4084 :
4085 0 : let import = maybe_import.ok_or_else(|| {
4086 0 : ApiError::NotFound(
4087 0 : format!(
4088 0 : "import for {}/{} not found",
4089 0 : req.tenant_shard_id.tenant_id, req.timeline_id
4090 0 : )
4091 0 : .into(),
4092 0 : )
4093 0 : })?;
4094 :
4095 : import
4096 : .shard_statuses
4097 : .0
4098 : .get(&req.tenant_shard_id.to_index())
4099 : .cloned()
4100 0 : .ok_or_else(|| {
4101 0 : ApiError::NotFound(
4102 0 : format!("shard {} not found", req.tenant_shard_id.shard_slug()).into(),
4103 0 : )
4104 0 : })
4105 : }
4106 :
4107 : #[instrument(skip_all, fields(
4108 : tenant_id=%req.tenant_shard_id.tenant_id,
4109 : shard_id=%req.tenant_shard_id.shard_slug(),
4110 : timeline_id=%req.timeline_id,
4111 : ))]
4112 : pub(crate) async fn handle_timeline_shard_import_progress_upcall(
4113 : self: &Arc<Self>,
4114 : req: PutTimelineImportStatusRequest,
4115 : ) -> Result<(), ApiError> {
4116 : let validity = self
4117 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4118 : .await?;
4119 : match validity {
4120 : ShardGenerationValidity::Valid => {
4121 : // fallthrough
4122 : }
4123 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4124 : tracing::info!(
4125 : claimed=?claimed.into(),
4126 0 : actual=?actual.and_then(|g| g.into()),
4127 : "Rejecting import progress update from stale generation"
4128 : );
4129 :
4130 : return Err(ApiError::PreconditionFailed("Invalid generation".into()));
4131 : }
4132 : }
4133 :
4134 : let res = self
4135 : .persistence
4136 : .update_timeline_import(req.tenant_shard_id, req.timeline_id, req.status)
4137 : .await;
4138 : let timeline_import = match res {
4139 : Ok(Ok(Some(timeline_import))) => timeline_import,
4140 : Ok(Ok(None)) => {
4141 : // Idempotency: we've already seen and handled this update.
4142 : return Ok(());
4143 : }
4144 : Ok(Err(logical_err)) => {
4145 : return Err(logical_err.into());
4146 : }
4147 : Err(db_err) => {
4148 : return Err(db_err.into());
4149 : }
4150 : };
4151 :
4152 : tracing::info!(
4153 : tenant_id=%req.tenant_shard_id.tenant_id,
4154 : timeline_id=%req.timeline_id,
4155 : shard_id=%req.tenant_shard_id.shard_slug(),
4156 : "Updated timeline import status to: {timeline_import:?}");
4157 :
4158 : if timeline_import.is_complete() {
4159 : tokio::task::spawn({
4160 : let this = self.clone();
4161 0 : async move { this.finalize_timeline_import(timeline_import).await }
4162 : });
4163 : }
4164 :
4165 : Ok(())
4166 : }
4167 :
4168 : /// Check that a provided generation for some tenant shard is the most recent one.
4169 : ///
4170 : /// Validate with the in-mem state first, and, if that passes, validate with the
4171 : /// database state which is authoritative.
4172 0 : async fn validate_shard_generation(
4173 0 : self: &Arc<Self>,
4174 0 : tenant_shard_id: TenantShardId,
4175 0 : generation: Generation,
4176 0 : ) -> Result<ShardGenerationValidity, ApiError> {
4177 : {
4178 0 : let locked = self.inner.read().unwrap();
4179 0 : let tenant_shard =
4180 0 : locked
4181 0 : .tenants
4182 0 : .get(&tenant_shard_id)
4183 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4184 0 : "{} shard not found",
4185 0 : tenant_shard_id
4186 0 : )))?;
4187 :
4188 0 : if tenant_shard.generation != Some(generation) {
4189 0 : return Ok(ShardGenerationValidity::Mismatched {
4190 0 : claimed: generation,
4191 0 : actual: tenant_shard.generation,
4192 0 : });
4193 0 : }
4194 : }
4195 :
4196 0 : let mut db_generations = self
4197 0 : .persistence
4198 0 : .shard_generations(std::iter::once(&tenant_shard_id))
4199 0 : .await?;
4200 0 : let (_tid, db_generation) =
4201 0 : db_generations
4202 0 : .pop()
4203 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4204 0 : "{} shard not found",
4205 0 : tenant_shard_id
4206 0 : )))?;
4207 :
4208 0 : if db_generation != Some(generation) {
4209 0 : return Ok(ShardGenerationValidity::Mismatched {
4210 0 : claimed: generation,
4211 0 : actual: db_generation,
4212 0 : });
4213 0 : }
4214 :
4215 0 : Ok(ShardGenerationValidity::Valid)
4216 0 : }
4217 :
4218 : /// Finalize the import of a timeline
4219 : ///
4220 : /// This method should be called once all shards have reported that the import is complete.
4221 : /// Firstly, it polls the post import timeline activation endpoint exposed by the pageserver.
4222 : /// Once the timeline is active on all shards, the timeline also gets created on the
4223 : /// safekeepers. Finally, notify cplane of the import completion (whether failed or
4224 : /// successful), and remove the import from the database and in-memory.
4225 : ///
4226 : /// If this method gets pre-empted by shut down, it will be called again at start-up (on-going
4227 : /// imports are stored in the database).
4228 : ///
4229 : /// # Cancel-Safety
4230 : /// Not cancel safe.
4231 : /// If the caller stops polling, the import will not be removed from
4232 : /// [`ServiceState::imports_finalizing`].
4233 : #[instrument(skip_all, fields(
4234 : tenant_id=%import.tenant_id,
4235 : timeline_id=%import.timeline_id,
4236 : ))]
4237 :
4238 : async fn finalize_timeline_import(
4239 : self: &Arc<Self>,
4240 : import: TimelineImport,
4241 : ) -> Result<(), TimelineImportFinalizeError> {
4242 : let tenant_timeline = (import.tenant_id, import.timeline_id);
4243 :
4244 : let (_finalize_import_guard, cancel) = {
4245 : let mut locked = self.inner.write().unwrap();
4246 : let gate = Gate::default();
4247 : let cancel = CancellationToken::default();
4248 :
4249 : let guard = gate.enter().unwrap();
4250 :
4251 : locked.imports_finalizing.insert(
4252 : tenant_timeline,
4253 : FinalizingImport {
4254 : gate,
4255 : cancel: cancel.clone(),
4256 : },
4257 : );
4258 :
4259 : (guard, cancel)
4260 : };
4261 :
4262 : let res = tokio::select! {
4263 : res = self.finalize_timeline_import_impl(import) => {
4264 : res
4265 : },
4266 : _ = cancel.cancelled() => {
4267 : Err(TimelineImportFinalizeError::Cancelled)
4268 : }
4269 : };
4270 :
4271 : let mut locked = self.inner.write().unwrap();
4272 : locked.imports_finalizing.remove(&tenant_timeline);
4273 :
4274 : res
4275 : }
4276 :
4277 0 : async fn finalize_timeline_import_impl(
4278 0 : self: &Arc<Self>,
4279 0 : import: TimelineImport,
4280 0 : ) -> Result<(), TimelineImportFinalizeError> {
4281 0 : tracing::info!("Finalizing timeline import");
4282 :
4283 0 : pausable_failpoint!("timeline-import-pre-cplane-notification");
4284 :
4285 0 : let tenant_id = import.tenant_id;
4286 0 : let timeline_id = import.timeline_id;
4287 :
4288 0 : let import_error = import.completion_error();
4289 0 : match import_error {
4290 0 : Some(err) => {
4291 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Err(err))
4292 0 : .await?;
4293 0 : tracing::warn!("Timeline import completed with shard errors");
4294 0 : Ok(())
4295 : }
4296 0 : None => match self.activate_timeline_post_import(&import).await {
4297 0 : Ok(timeline_info) => {
4298 0 : tracing::info!("Post import timeline activation complete");
4299 :
4300 0 : if self.config.timelines_onto_safekeepers {
4301 : // Now that we know the start LSN of this timeline, create it on the
4302 : // safekeepers.
4303 0 : self.tenant_timeline_create_safekeepers_until_success(
4304 0 : import.tenant_id,
4305 0 : timeline_info,
4306 0 : )
4307 0 : .await?;
4308 0 : }
4309 :
4310 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Ok(()))
4311 0 : .await?;
4312 :
4313 0 : tracing::info!("Timeline import completed successfully");
4314 0 : Ok(())
4315 : }
4316 : Err(TimelineImportFinalizeError::ShuttingDown) => {
4317 : // We got pre-empted by shut down and will resume after the restart.
4318 0 : Err(TimelineImportFinalizeError::ShuttingDown)
4319 : }
4320 0 : Err(err) => {
4321 : // Any finalize error apart from shut down is permanent and requires us to notify
4322 : // cplane such that it can clean up.
4323 0 : tracing::error!("Import finalize failed with permanent error: {err}");
4324 0 : self.notify_cplane_and_delete_import(
4325 0 : tenant_id,
4326 0 : timeline_id,
4327 0 : Err(err.to_string()),
4328 0 : )
4329 0 : .await?;
4330 0 : Err(err)
4331 : }
4332 : },
4333 : }
4334 0 : }
4335 :
4336 0 : async fn notify_cplane_and_delete_import(
4337 0 : self: &Arc<Self>,
4338 0 : tenant_id: TenantId,
4339 0 : timeline_id: TimelineId,
4340 0 : import_result: ImportResult,
4341 0 : ) -> Result<(), TimelineImportFinalizeError> {
4342 0 : let import_failed = import_result.is_err();
4343 0 : tracing::info!(%import_failed, "Notifying cplane of import completion");
4344 :
4345 0 : let client = UpcallClient::new(self.get_config(), self.cancel.child_token());
4346 0 : client
4347 0 : .notify_import_complete(tenant_id, timeline_id, import_result)
4348 0 : .await
4349 0 : .map_err(|_err| TimelineImportFinalizeError::ShuttingDown)?;
4350 :
4351 0 : if let Err(err) = self
4352 0 : .persistence
4353 0 : .delete_timeline_import(tenant_id, timeline_id)
4354 0 : .await
4355 : {
4356 0 : tracing::warn!("Failed to delete timeline import entry from database: {err}");
4357 0 : }
4358 :
4359 0 : self.inner
4360 0 : .write()
4361 0 : .unwrap()
4362 0 : .tenants
4363 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4364 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Idle);
4365 :
4366 0 : Ok(())
4367 0 : }
4368 :
4369 : /// Activate an imported timeline on all shards once the import is complete.
4370 : /// Returns the [`TimelineInfo`] reported by shard zero.
4371 0 : async fn activate_timeline_post_import(
4372 0 : self: &Arc<Self>,
4373 0 : import: &TimelineImport,
4374 0 : ) -> Result<TimelineInfo, TimelineImportFinalizeError> {
4375 : const TIMELINE_ACTIVATE_TIMEOUT: Duration = Duration::from_millis(128);
4376 :
4377 0 : let mut shards_to_activate: HashSet<ShardIndex> =
4378 0 : import.shard_statuses.0.keys().cloned().collect();
4379 0 : let mut shard_zero_timeline_info = None;
4380 :
4381 0 : while !shards_to_activate.is_empty() {
4382 0 : if self.cancel.is_cancelled() {
4383 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4384 0 : }
4385 :
4386 0 : let targets = {
4387 0 : let locked = self.inner.read().unwrap();
4388 0 : let mut targets = Vec::new();
4389 :
4390 0 : for (tenant_shard_id, shard) in locked
4391 0 : .tenants
4392 0 : .range(TenantShardId::tenant_range(import.tenant_id))
4393 : {
4394 0 : if !import
4395 0 : .shard_statuses
4396 0 : .0
4397 0 : .contains_key(&tenant_shard_id.to_index())
4398 : {
4399 0 : return Err(TimelineImportFinalizeError::MismatchedShards(
4400 0 : tenant_shard_id.to_index(),
4401 0 : ));
4402 0 : }
4403 :
4404 0 : if let Some(node_id) = shard.intent.get_attached() {
4405 0 : let node = locked
4406 0 : .nodes
4407 0 : .get(node_id)
4408 0 : .expect("Pageservers may not be deleted while referenced");
4409 0 : targets.push((*tenant_shard_id, node.clone()));
4410 0 : }
4411 : }
4412 :
4413 0 : targets
4414 : };
4415 :
4416 0 : let targeted_tenant_shards: Vec<_> = targets.iter().map(|(tid, _node)| *tid).collect();
4417 :
4418 0 : let results = self
4419 0 : .tenant_for_shards_api(
4420 0 : targets,
4421 0 : |tenant_shard_id, client| async move {
4422 0 : client
4423 0 : .activate_post_import(
4424 0 : tenant_shard_id,
4425 0 : import.timeline_id,
4426 0 : TIMELINE_ACTIVATE_TIMEOUT,
4427 0 : )
4428 0 : .await
4429 0 : },
4430 : 1,
4431 : 1,
4432 : SHORT_RECONCILE_TIMEOUT,
4433 0 : &self.cancel,
4434 : )
4435 0 : .await;
4436 :
4437 0 : let mut failed = 0;
4438 0 : for (tid, (_, result)) in targeted_tenant_shards.iter().zip(results.into_iter()) {
4439 0 : match result {
4440 0 : Ok(ok) => {
4441 0 : if tid.is_shard_zero() {
4442 0 : shard_zero_timeline_info = Some(ok);
4443 0 : }
4444 :
4445 0 : shards_to_activate.remove(&tid.to_index());
4446 : }
4447 0 : Err(_err) => {
4448 0 : failed += 1;
4449 0 : }
4450 : }
4451 : }
4452 :
4453 0 : if failed > 0 {
4454 0 : tracing::info!(
4455 0 : "Failed to activate timeline on {failed} shards post import. Will retry"
4456 : );
4457 0 : }
4458 :
4459 0 : tokio::select! {
4460 0 : _ = tokio::time::sleep(Duration::from_millis(250)) => {},
4461 0 : _ = self.cancel.cancelled() => {
4462 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4463 : }
4464 : }
4465 : }
4466 :
4467 0 : Ok(shard_zero_timeline_info.expect("All shards replied"))
4468 0 : }
4469 :
4470 0 : async fn finalize_timeline_imports(self: &Arc<Self>, imports: Vec<TimelineImport>) {
4471 0 : futures::future::join_all(
4472 0 : imports
4473 0 : .into_iter()
4474 0 : .map(|import| self.finalize_timeline_import(import)),
4475 : )
4476 0 : .await;
4477 0 : }
4478 :
4479 : /// Delete a timeline import if it exists
4480 : ///
4481 : /// Firstly, delete the entry from the database. Any updates
4482 : /// from pageservers after the update will fail with a 404, so the
4483 : /// import cannot progress into finalizing state if it's not there already.
4484 : /// Secondly, cancel the finalization if one is in progress.
4485 0 : pub(crate) async fn maybe_delete_timeline_import(
4486 0 : self: &Arc<Self>,
4487 0 : tenant_id: TenantId,
4488 0 : timeline_id: TimelineId,
4489 0 : ) -> Result<(), DatabaseError> {
4490 0 : let tenant_has_ongoing_import = {
4491 0 : let locked = self.inner.read().unwrap();
4492 0 : locked
4493 0 : .tenants
4494 0 : .range(TenantShardId::tenant_range(tenant_id))
4495 0 : .any(|(_tid, shard)| shard.importing == TimelineImportState::Importing)
4496 : };
4497 :
4498 0 : if !tenant_has_ongoing_import {
4499 0 : return Ok(());
4500 0 : }
4501 :
4502 0 : self.persistence
4503 0 : .delete_timeline_import(tenant_id, timeline_id)
4504 0 : .await?;
4505 :
4506 0 : let maybe_finalizing = {
4507 0 : let mut locked = self.inner.write().unwrap();
4508 0 : locked.imports_finalizing.remove(&(tenant_id, timeline_id))
4509 : };
4510 :
4511 0 : if let Some(finalizing) = maybe_finalizing {
4512 0 : finalizing.cancel.cancel();
4513 0 : finalizing.gate.close().await;
4514 0 : }
4515 :
4516 0 : Ok(())
4517 0 : }
4518 :
4519 0 : pub(crate) async fn tenant_timeline_archival_config(
4520 0 : &self,
4521 0 : tenant_id: TenantId,
4522 0 : timeline_id: TimelineId,
4523 0 : req: TimelineArchivalConfigRequest,
4524 0 : ) -> Result<(), ApiError> {
4525 0 : tracing::info!(
4526 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
4527 : req.state
4528 : );
4529 :
4530 0 : let _tenant_lock = trace_shared_lock(
4531 0 : &self.tenant_op_locks,
4532 0 : tenant_id,
4533 0 : TenantOperations::TimelineArchivalConfig,
4534 0 : )
4535 0 : .await;
4536 :
4537 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4538 0 : if targets.0.is_empty() {
4539 0 : return Err(ApiError::NotFound(
4540 0 : anyhow::anyhow!("Tenant not found").into(),
4541 0 : ));
4542 0 : }
4543 0 : async fn config_one(
4544 0 : tenant_shard_id: TenantShardId,
4545 0 : timeline_id: TimelineId,
4546 0 : node: Node,
4547 0 : http_client: reqwest::Client,
4548 0 : jwt: Option<String>,
4549 0 : req: TimelineArchivalConfigRequest,
4550 0 : ) -> Result<(), ApiError> {
4551 0 : tracing::info!(
4552 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4553 : );
4554 :
4555 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4556 :
4557 0 : client
4558 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
4559 0 : .await
4560 0 : .map_err(|e| match e {
4561 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
4562 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
4563 : }
4564 0 : _ => passthrough_api_error(&node, e),
4565 0 : })
4566 0 : }
4567 :
4568 : // no shard needs to go first/last; the operation should be idempotent
4569 : // TODO: it would be great to ensure that all shards return the same error
4570 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4571 0 : let results = self
4572 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4573 0 : futures::FutureExt::boxed(config_one(
4574 0 : tenant_shard_id,
4575 0 : timeline_id,
4576 0 : node,
4577 0 : self.http_client.clone(),
4578 0 : self.config.pageserver_jwt_token.clone(),
4579 0 : req.clone(),
4580 0 : ))
4581 0 : })
4582 0 : .await?;
4583 0 : assert!(!results.is_empty(), "must have at least one result");
4584 :
4585 0 : Ok(())
4586 0 : }).await?
4587 0 : }
4588 :
4589 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
4590 0 : &self,
4591 0 : tenant_id: TenantId,
4592 0 : timeline_id: TimelineId,
4593 0 : behavior: Option<DetachBehavior>,
4594 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
4595 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
4596 :
4597 0 : let _tenant_lock = trace_shared_lock(
4598 0 : &self.tenant_op_locks,
4599 0 : tenant_id,
4600 0 : TenantOperations::TimelineDetachAncestor,
4601 0 : )
4602 0 : .await;
4603 :
4604 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4605 0 : if targets.0.is_empty() {
4606 0 : return Err(ApiError::NotFound(
4607 0 : anyhow::anyhow!("Tenant not found").into(),
4608 0 : ));
4609 0 : }
4610 :
4611 0 : async fn detach_one(
4612 0 : tenant_shard_id: TenantShardId,
4613 0 : timeline_id: TimelineId,
4614 0 : node: Node,
4615 0 : http_client: reqwest::Client,
4616 0 : jwt: Option<String>,
4617 0 : behavior: Option<DetachBehavior>,
4618 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
4619 0 : tracing::info!(
4620 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4621 : );
4622 :
4623 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4624 :
4625 0 : client
4626 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id, behavior)
4627 0 : .await
4628 0 : .map_err(|e| {
4629 : use mgmt_api::Error;
4630 :
4631 0 : match e {
4632 : // no ancestor (ever)
4633 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
4634 0 : "{node}: {}",
4635 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
4636 0 : )),
4637 : // too many ancestors
4638 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
4639 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
4640 : }
4641 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
4642 : // avoid turning these into conflicts to remain compatible with
4643 : // pageservers, 500 errors are sadly retryable with timeline ancestor
4644 : // detach
4645 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
4646 : }
4647 : // rest can be mapped as usual
4648 0 : other => passthrough_api_error(&node, other),
4649 : }
4650 0 : })
4651 0 : .map(|res| (tenant_shard_id.shard_number, res))
4652 0 : }
4653 :
4654 : // no shard needs to go first/last; the operation should be idempotent
4655 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4656 0 : let mut results = self
4657 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4658 0 : futures::FutureExt::boxed(detach_one(
4659 0 : tenant_shard_id,
4660 0 : timeline_id,
4661 0 : node,
4662 0 : self.http_client.clone(),
4663 0 : self.config.pageserver_jwt_token.clone(),
4664 0 : behavior,
4665 0 : ))
4666 0 : })
4667 0 : .await?;
4668 :
4669 0 : let any = results.pop().expect("we must have at least one response");
4670 :
4671 0 : let mismatching = results
4672 0 : .iter()
4673 0 : .filter(|(_, res)| res != &any.1)
4674 0 : .collect::<Vec<_>>();
4675 0 : if !mismatching.is_empty() {
4676 : // this can be hit by races which should not happen because operation lock on cplane
4677 0 : let matching = results.len() - mismatching.len();
4678 0 : tracing::error!(
4679 : matching,
4680 : compared_against=?any,
4681 : ?mismatching,
4682 0 : "shards returned different results"
4683 : );
4684 :
4685 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
4686 0 : }
4687 :
4688 0 : Ok(any.1)
4689 0 : }).await?
4690 0 : }
4691 :
4692 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
4693 0 : &self,
4694 0 : tenant_id: TenantId,
4695 0 : timeline_id: TimelineId,
4696 0 : dir: BlockUnblock,
4697 0 : ) -> Result<(), ApiError> {
4698 0 : let _tenant_lock = trace_shared_lock(
4699 0 : &self.tenant_op_locks,
4700 0 : tenant_id,
4701 0 : TenantOperations::TimelineGcBlockUnblock,
4702 0 : )
4703 0 : .await;
4704 :
4705 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4706 0 : if targets.0.is_empty() {
4707 0 : return Err(ApiError::NotFound(
4708 0 : anyhow::anyhow!("Tenant not found").into(),
4709 0 : ));
4710 0 : }
4711 :
4712 0 : async fn do_one(
4713 0 : tenant_shard_id: TenantShardId,
4714 0 : timeline_id: TimelineId,
4715 0 : node: Node,
4716 0 : http_client: reqwest::Client,
4717 0 : jwt: Option<String>,
4718 0 : dir: BlockUnblock,
4719 0 : ) -> Result<(), ApiError> {
4720 0 : let client = PageserverClient::new(
4721 0 : node.get_id(),
4722 0 : http_client,
4723 0 : node.base_url(),
4724 0 : jwt.as_deref(),
4725 : );
4726 :
4727 0 : client
4728 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
4729 0 : .await
4730 0 : .map_err(|e| passthrough_api_error(&node, e))
4731 0 : }
4732 :
4733 : // no shard needs to go first/last; the operation should be idempotent
4734 0 : let locations = targets
4735 0 : .0
4736 0 : .iter()
4737 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
4738 0 : .collect();
4739 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
4740 0 : futures::FutureExt::boxed(do_one(
4741 0 : tenant_shard_id,
4742 0 : timeline_id,
4743 0 : node,
4744 0 : self.http_client.clone(),
4745 0 : self.config.pageserver_jwt_token.clone(),
4746 0 : dir,
4747 0 : ))
4748 0 : })
4749 0 : .await
4750 0 : })
4751 0 : .await??;
4752 0 : Ok(())
4753 0 : }
4754 :
4755 0 : pub(crate) async fn tenant_timeline_lsn_lease(
4756 0 : &self,
4757 0 : tenant_id: TenantId,
4758 0 : timeline_id: TimelineId,
4759 0 : lsn: Lsn,
4760 0 : ) -> Result<LsnLease, ApiError> {
4761 0 : let _tenant_lock = trace_shared_lock(
4762 0 : &self.tenant_op_locks,
4763 0 : tenant_id,
4764 0 : TenantOperations::TimelineLsnLease,
4765 0 : )
4766 0 : .await;
4767 :
4768 0 : let mut retry_if_not_attached = false;
4769 0 : let targets = {
4770 0 : let locked = self.inner.read().unwrap();
4771 0 : let mut targets = Vec::new();
4772 :
4773 : // If the request got an unsharded tenant id, then apply
4774 : // the operation to all shards. Otherwise, apply it to a specific shard.
4775 0 : let shards_range = TenantShardId::tenant_range(tenant_id);
4776 :
4777 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4778 0 : if let Some(node_id) = shard.intent.get_attached() {
4779 0 : let node = locked
4780 0 : .nodes
4781 0 : .get(node_id)
4782 0 : .expect("Pageservers may not be deleted while referenced");
4783 :
4784 0 : targets.push((*tenant_shard_id, node.clone()));
4785 :
4786 0 : if let Some(location) = shard.observed.locations.get(node_id) {
4787 0 : if let Some(ref conf) = location.conf {
4788 0 : if conf.mode != LocationConfigMode::AttachedSingle
4789 0 : && conf.mode != LocationConfigMode::AttachedMulti
4790 0 : {
4791 0 : // If the shard is attached as secondary, we need to retry if 404.
4792 0 : retry_if_not_attached = true;
4793 0 : }
4794 : // If the shard is attached as primary, we should succeed.
4795 0 : } else {
4796 0 : // Location conf is not available yet, retry if 404.
4797 0 : retry_if_not_attached = true;
4798 0 : }
4799 0 : } else {
4800 0 : // The shard is not attached to the intended pageserver yet, retry if 404.
4801 0 : retry_if_not_attached = true;
4802 0 : }
4803 0 : }
4804 : }
4805 0 : targets
4806 : };
4807 :
4808 0 : let res = self
4809 0 : .tenant_for_shards_api(
4810 0 : targets,
4811 0 : |tenant_shard_id, client| async move {
4812 0 : client
4813 0 : .timeline_lease_lsn(tenant_shard_id, timeline_id, lsn)
4814 0 : .await
4815 0 : },
4816 : 1,
4817 : 1,
4818 : SHORT_RECONCILE_TIMEOUT,
4819 0 : &self.cancel,
4820 : )
4821 0 : .await;
4822 :
4823 0 : let mut valid_until = None;
4824 0 : for (node, r) in res {
4825 0 : match r {
4826 0 : Ok(lease) => {
4827 0 : if let Some(ref mut valid_until) = valid_until {
4828 0 : *valid_until = std::cmp::min(*valid_until, lease.valid_until);
4829 0 : } else {
4830 0 : valid_until = Some(lease.valid_until);
4831 0 : }
4832 : }
4833 : Err(mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _))
4834 0 : if retry_if_not_attached =>
4835 : {
4836 : // This is expected if the attach is not finished yet. Return 503 so that the client can retry.
4837 0 : return Err(ApiError::ResourceUnavailable(
4838 0 : format!(
4839 0 : "Timeline is not attached to the pageserver {} yet, please retry",
4840 0 : node.get_id()
4841 0 : )
4842 0 : .into(),
4843 0 : ));
4844 : }
4845 0 : Err(e) => {
4846 0 : return Err(passthrough_api_error(&node, e));
4847 : }
4848 : }
4849 : }
4850 0 : Ok(LsnLease {
4851 0 : valid_until: valid_until.unwrap_or_else(SystemTime::now),
4852 0 : })
4853 0 : }
4854 :
4855 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4856 0 : &self,
4857 0 : tenant_shard_id: TenantShardId,
4858 0 : timeline_id: TimelineId,
4859 0 : concurrency: Option<usize>,
4860 0 : recurse: bool,
4861 0 : ) -> Result<(), ApiError> {
4862 0 : let _tenant_lock = trace_shared_lock(
4863 0 : &self.tenant_op_locks,
4864 0 : tenant_shard_id.tenant_id,
4865 0 : TenantOperations::DownloadHeatmapLayers,
4866 0 : )
4867 0 : .await;
4868 :
4869 0 : let targets = {
4870 0 : let locked = self.inner.read().unwrap();
4871 0 : let mut targets = Vec::new();
4872 :
4873 : // If the request got an unsharded tenant id, then apply
4874 : // the operation to all shards. Otherwise, apply it to a specific shard.
4875 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4876 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4877 : } else {
4878 0 : tenant_shard_id.range()
4879 : };
4880 :
4881 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4882 0 : if let Some(node_id) = shard.intent.get_attached() {
4883 0 : let node = locked
4884 0 : .nodes
4885 0 : .get(node_id)
4886 0 : .expect("Pageservers may not be deleted while referenced");
4887 0 :
4888 0 : targets.push((*tenant_shard_id, node.clone()));
4889 0 : }
4890 : }
4891 0 : targets
4892 : };
4893 :
4894 0 : self.tenant_for_shards_api(
4895 0 : targets,
4896 0 : |tenant_shard_id, client| async move {
4897 0 : client
4898 0 : .timeline_download_heatmap_layers(
4899 0 : tenant_shard_id,
4900 0 : timeline_id,
4901 0 : concurrency,
4902 0 : recurse,
4903 0 : )
4904 0 : .await
4905 0 : },
4906 : 1,
4907 : 1,
4908 : SHORT_RECONCILE_TIMEOUT,
4909 0 : &self.cancel,
4910 : )
4911 0 : .await;
4912 :
4913 0 : Ok(())
4914 0 : }
4915 :
4916 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4917 : ///
4918 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`
4919 : /// and returned element at index `i` is the result for `req_fn(op(locations[i])`.
4920 0 : async fn tenant_for_shards<F, R>(
4921 0 : &self,
4922 0 : locations: Vec<(TenantShardId, Node)>,
4923 0 : mut req_fn: F,
4924 0 : ) -> Result<Vec<R>, ApiError>
4925 0 : where
4926 0 : F: FnMut(
4927 0 : TenantShardId,
4928 0 : Node,
4929 0 : )
4930 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4931 0 : {
4932 0 : let mut futs = FuturesUnordered::new();
4933 0 : let mut results = Vec::with_capacity(locations.len());
4934 :
4935 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4936 0 : let fut = req_fn(tenant_shard_id, node);
4937 0 : futs.push(async move { (idx, fut.await) });
4938 : }
4939 :
4940 0 : while let Some((idx, r)) = futs.next().await {
4941 0 : results.push((idx, r?));
4942 : }
4943 :
4944 0 : results.sort_by_key(|(idx, _)| *idx);
4945 0 : Ok(results.into_iter().map(|(_, r)| r).collect())
4946 0 : }
4947 :
4948 : /// Concurrently invoke a pageserver API call on many shards at once.
4949 : ///
4950 : /// The returned Vec has the same length as the `locations` Vec,
4951 : /// and returned element at index `i` is the result for `op(locations[i])`.
4952 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
4953 0 : &self,
4954 0 : locations: Vec<(TenantShardId, Node)>,
4955 0 : op: O,
4956 0 : warn_threshold: u32,
4957 0 : max_retries: u32,
4958 0 : timeout: Duration,
4959 0 : cancel: &CancellationToken,
4960 0 : ) -> Vec<(Node, mgmt_api::Result<T>)>
4961 0 : where
4962 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
4963 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
4964 0 : {
4965 0 : let mut futs = FuturesUnordered::new();
4966 0 : let mut results = Vec::with_capacity(locations.len());
4967 :
4968 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4969 0 : futs.push(async move {
4970 0 : let r = node
4971 0 : .with_client_retries(
4972 0 : |client| op(tenant_shard_id, client),
4973 0 : &self.http_client,
4974 0 : &self.config.pageserver_jwt_token,
4975 0 : warn_threshold,
4976 0 : max_retries,
4977 0 : timeout,
4978 0 : cancel,
4979 : )
4980 0 : .await;
4981 0 : (idx, node, r)
4982 0 : });
4983 : }
4984 :
4985 0 : while let Some((idx, node, r)) = futs.next().await {
4986 0 : results.push((idx, node, r.unwrap_or(Err(mgmt_api::Error::Cancelled))));
4987 0 : }
4988 :
4989 0 : results.sort_by_key(|(idx, _, _)| *idx);
4990 0 : results.into_iter().map(|(_, node, r)| (node, r)).collect()
4991 0 : }
4992 :
4993 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
4994 : /// when creating and deleting timelines:
4995 : /// - Makes sure shards are attached somewhere if they weren't already
4996 : /// - Looks up the shards and the nodes where they were most recently attached
4997 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
4998 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
4999 0 : async fn tenant_remote_mutation<R, O, F>(
5000 0 : &self,
5001 0 : tenant_id: TenantId,
5002 0 : op: O,
5003 0 : ) -> Result<R, ApiError>
5004 0 : where
5005 0 : O: FnOnce(TenantMutationLocations) -> F,
5006 0 : F: std::future::Future<Output = R>,
5007 0 : {
5008 0 : let mutation_locations = {
5009 0 : let mut locations = TenantMutationLocations::default();
5010 :
5011 : // Load the currently attached pageservers for the latest generation of each shard. This can
5012 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
5013 : // will still be the latest when we're done: we will check generations again at the end of
5014 : // this function to handle that.
5015 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
5016 :
5017 0 : if generations
5018 0 : .iter()
5019 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
5020 : {
5021 0 : let shard_generations = generations
5022 0 : .into_iter()
5023 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
5024 0 : .collect::<HashMap<_, _>>();
5025 :
5026 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
5027 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
5028 0 : let locked = self.inner.read().unwrap();
5029 0 : for (shard_id, shard) in
5030 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5031 : {
5032 0 : match shard.policy {
5033 : PlacementPolicy::Attached(_) => {
5034 : // This shard is meant to be attached: the caller is not wrong to try and
5035 : // use this function, but we can't service the request right now.
5036 0 : let Some(generation) = shard_generations.get(shard_id) else {
5037 : // This can only happen if there is a split brain controller modifying the database. This should
5038 : // never happen when testing, and if it happens in production we can only log the issue.
5039 0 : debug_assert!(false);
5040 0 : tracing::error!(
5041 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
5042 : );
5043 0 : continue;
5044 : };
5045 0 : let (generation, generation_pageserver) = generation;
5046 0 : if let Some(generation) = generation {
5047 0 : if generation_pageserver.is_none() {
5048 : // This is legitimate only in a very narrow window where the shard was only just configured into
5049 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
5050 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
5051 0 : tracing::warn!(
5052 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
5053 : );
5054 0 : }
5055 : } else {
5056 : // This should never happen: a shard with no generation is only permitted when it was created in some state
5057 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
5058 0 : debug_assert!(false);
5059 0 : tracing::error!(
5060 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
5061 : );
5062 0 : continue;
5063 : }
5064 : }
5065 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
5066 0 : return Err(ApiError::Conflict(format!(
5067 0 : "Shard {shard_id} tenant has policy {:?}",
5068 0 : shard.policy
5069 0 : )));
5070 : }
5071 : }
5072 : }
5073 :
5074 0 : return Err(ApiError::ResourceUnavailable(
5075 0 : "One or more shards in tenant is not yet attached".into(),
5076 0 : ));
5077 0 : }
5078 :
5079 0 : let locked = self.inner.read().unwrap();
5080 : for ShardGenerationState {
5081 0 : tenant_shard_id,
5082 0 : generation,
5083 0 : generation_pageserver,
5084 0 : } in generations
5085 : {
5086 0 : let node_id = generation_pageserver.expect("We checked for None above");
5087 0 : let node = locked
5088 0 : .nodes
5089 0 : .get(&node_id)
5090 0 : .ok_or(ApiError::Conflict(format!(
5091 0 : "Raced with removal of node {node_id}"
5092 0 : )))?;
5093 0 : let generation = generation.expect("Checked above");
5094 :
5095 0 : let tenant = locked.tenants.get(&tenant_shard_id);
5096 :
5097 : // TODO(vlad): Abstract the logic that finds stale attached locations
5098 : // from observed state into a [`Service`] method.
5099 0 : let other_locations = match tenant {
5100 0 : Some(tenant) => {
5101 0 : let mut other = tenant.attached_locations();
5102 0 : let latest_location_index =
5103 0 : other.iter().position(|&l| l == (node.get_id(), generation));
5104 0 : if let Some(idx) = latest_location_index {
5105 0 : other.remove(idx);
5106 0 : }
5107 :
5108 0 : other
5109 : }
5110 0 : None => Vec::default(),
5111 : };
5112 :
5113 0 : let location = ShardMutationLocations {
5114 0 : latest: MutationLocation {
5115 0 : node: node.clone(),
5116 0 : generation,
5117 0 : },
5118 0 : other: other_locations
5119 0 : .into_iter()
5120 0 : .filter_map(|(node_id, generation)| {
5121 0 : let node = locked.nodes.get(&node_id)?;
5122 :
5123 0 : Some(MutationLocation {
5124 0 : node: node.clone(),
5125 0 : generation,
5126 0 : })
5127 0 : })
5128 0 : .collect(),
5129 : };
5130 0 : locations.0.insert(tenant_shard_id, location);
5131 : }
5132 :
5133 0 : locations
5134 : };
5135 :
5136 0 : let result = op(mutation_locations.clone()).await;
5137 :
5138 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
5139 : // our remote operation executed on the latest generation and is therefore persistent.
5140 : {
5141 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
5142 0 : if latest_generations
5143 0 : .into_iter()
5144 0 : .map(
5145 : |ShardGenerationState {
5146 : tenant_shard_id,
5147 : generation,
5148 : generation_pageserver: _,
5149 0 : }| (tenant_shard_id, generation),
5150 : )
5151 0 : .collect::<Vec<_>>()
5152 0 : != mutation_locations
5153 0 : .0
5154 0 : .into_iter()
5155 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
5156 0 : .collect::<Vec<_>>()
5157 : {
5158 : // We raced with something that incremented the generation, and therefore cannot be
5159 : // confident that our actions are persistent (they might have hit an old generation).
5160 : //
5161 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
5162 0 : return Err(ApiError::ResourceUnavailable(
5163 0 : "Tenant attachment changed, please retry".into(),
5164 0 : ));
5165 0 : }
5166 : }
5167 :
5168 0 : Ok(result)
5169 0 : }
5170 :
5171 0 : pub(crate) async fn tenant_timeline_delete(
5172 0 : self: &Arc<Self>,
5173 0 : tenant_id: TenantId,
5174 0 : timeline_id: TimelineId,
5175 0 : ) -> Result<StatusCode, ApiError> {
5176 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
5177 0 : let _tenant_lock = trace_shared_lock(
5178 0 : &self.tenant_op_locks,
5179 0 : tenant_id,
5180 0 : TenantOperations::TimelineDelete,
5181 0 : )
5182 0 : .await;
5183 :
5184 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
5185 0 : if targets.0.is_empty() {
5186 0 : return Err(ApiError::NotFound(
5187 0 : anyhow::anyhow!("Tenant not found").into(),
5188 0 : ));
5189 0 : }
5190 :
5191 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
5192 0 : assert!(shard_zero_tid.is_shard_zero());
5193 :
5194 0 : async fn delete_one(
5195 0 : tenant_shard_id: TenantShardId,
5196 0 : timeline_id: TimelineId,
5197 0 : node: Node,
5198 0 : http_client: reqwest::Client,
5199 0 : jwt: Option<String>,
5200 0 : ) -> Result<StatusCode, ApiError> {
5201 0 : tracing::info!(
5202 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
5203 : );
5204 :
5205 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
5206 0 : let res = client
5207 0 : .timeline_delete(tenant_shard_id, timeline_id)
5208 0 : .await;
5209 :
5210 0 : match res {
5211 0 : Ok(ok) => Ok(ok),
5212 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
5213 0 : Err(mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg)) if msg.contains("Requested tenant is missing") => {
5214 0 : Err(ApiError::ResourceUnavailable("Tenant migration in progress".into()))
5215 : },
5216 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
5217 0 : Err(e) => {
5218 0 : Err(
5219 0 : ApiError::InternalServerError(anyhow::anyhow!(
5220 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
5221 0 : ))
5222 0 : )
5223 : }
5224 : }
5225 0 : }
5226 :
5227 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
5228 0 : let statuses = self
5229 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
5230 0 : Box::pin(delete_one(
5231 0 : tenant_shard_id,
5232 0 : timeline_id,
5233 0 : node,
5234 0 : self.http_client.clone(),
5235 0 : self.config.pageserver_jwt_token.clone(),
5236 0 : ))
5237 0 : })
5238 0 : .await?;
5239 :
5240 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
5241 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
5242 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
5243 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
5244 0 : return Ok(StatusCode::CONFLICT);
5245 0 : }
5246 :
5247 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
5248 0 : return Ok(StatusCode::ACCEPTED);
5249 0 : }
5250 :
5251 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
5252 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
5253 0 : let shard_zero_status = delete_one(
5254 0 : shard_zero_tid,
5255 0 : timeline_id,
5256 0 : shard_zero_locations.latest.node,
5257 0 : self.http_client.clone(),
5258 0 : self.config.pageserver_jwt_token.clone(),
5259 0 : )
5260 0 : .await?;
5261 0 : Ok(shard_zero_status)
5262 0 : }).await?;
5263 :
5264 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
5265 0 : .await?;
5266 :
5267 0 : status_code
5268 0 : }
5269 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
5270 0 : pub(crate) async fn tenant_shard0_node(
5271 0 : &self,
5272 0 : tenant_id: TenantId,
5273 0 : ) -> Result<(Node, TenantShardId), ApiError> {
5274 0 : let tenant_shard_id = {
5275 0 : let locked = self.inner.read().unwrap();
5276 0 : let Some((tenant_shard_id, _shard)) = locked
5277 0 : .tenants
5278 0 : .range(TenantShardId::tenant_range(tenant_id))
5279 0 : .next()
5280 : else {
5281 0 : return Err(ApiError::NotFound(
5282 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
5283 0 : ));
5284 : };
5285 :
5286 0 : *tenant_shard_id
5287 : };
5288 :
5289 0 : self.tenant_shard_node(tenant_shard_id)
5290 0 : .await
5291 0 : .map(|node| (node, tenant_shard_id))
5292 0 : }
5293 :
5294 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
5295 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
5296 0 : pub(crate) async fn tenant_shard_node(
5297 0 : &self,
5298 0 : tenant_shard_id: TenantShardId,
5299 0 : ) -> Result<Node, ApiError> {
5300 : // Look up in-memory state and maybe use the node from there.
5301 : {
5302 0 : let locked = self.inner.read().unwrap();
5303 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5304 0 : return Err(ApiError::NotFound(
5305 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
5306 0 : ));
5307 : };
5308 :
5309 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
5310 0 : tracing::warn!(
5311 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
5312 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
5313 : shard.policy
5314 : );
5315 0 : return Err(ApiError::Conflict(
5316 0 : "Cannot call timeline API on non-attached tenant".to_string(),
5317 0 : ));
5318 : };
5319 :
5320 0 : if shard.reconciler.is_none() {
5321 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
5322 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
5323 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
5324 : // This should never happen
5325 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5326 0 : "Shard refers to nonexistent node"
5327 0 : )));
5328 : };
5329 0 : return Ok(node.clone());
5330 0 : }
5331 : };
5332 :
5333 : // Look up the latest attached pageserver location from the database
5334 : // generation state: this will reflect the progress of any ongoing migration.
5335 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
5336 : // the case where they call through to the pageserver and get a 404.
5337 0 : let db_result = self
5338 0 : .persistence
5339 0 : .tenant_generations(tenant_shard_id.tenant_id)
5340 0 : .await?;
5341 : let Some(ShardGenerationState {
5342 : tenant_shard_id: _,
5343 : generation: _,
5344 0 : generation_pageserver: Some(node_id),
5345 0 : }) = db_result
5346 0 : .into_iter()
5347 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
5348 : else {
5349 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
5350 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
5351 : // or a conflict response (case where tenant was detached in background)
5352 0 : return Err(ApiError::ResourceUnavailable(
5353 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
5354 0 : ));
5355 : };
5356 0 : let locked = self.inner.read().unwrap();
5357 0 : let Some(node) = locked.nodes.get(&node_id) else {
5358 : // This should never happen
5359 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5360 0 : "Shard refers to nonexistent node"
5361 0 : )));
5362 : };
5363 :
5364 0 : Ok(node.clone())
5365 0 : }
5366 :
5367 0 : pub(crate) fn tenant_locate(
5368 0 : &self,
5369 0 : tenant_id: TenantId,
5370 0 : ) -> Result<TenantLocateResponse, ApiError> {
5371 0 : let locked = self.inner.read().unwrap();
5372 0 : tracing::info!("Locating shards for tenant {tenant_id}");
5373 :
5374 0 : let mut result = Vec::new();
5375 0 : let mut shard_params: Option<ShardParameters> = None;
5376 :
5377 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5378 : {
5379 0 : let node_id =
5380 0 : shard
5381 0 : .intent
5382 0 : .get_attached()
5383 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
5384 0 : "Cannot locate a tenant that is not attached"
5385 0 : )))?;
5386 :
5387 0 : let node = locked
5388 0 : .nodes
5389 0 : .get(&node_id)
5390 0 : .expect("Pageservers may not be deleted while referenced");
5391 :
5392 0 : result.push(node.shard_location(*tenant_shard_id));
5393 :
5394 0 : match &shard_params {
5395 0 : None => {
5396 0 : shard_params = Some(ShardParameters {
5397 0 : stripe_size: shard.shard.stripe_size,
5398 0 : count: shard.shard.count,
5399 0 : });
5400 0 : }
5401 0 : Some(params) => {
5402 0 : if params.stripe_size != shard.shard.stripe_size {
5403 : // This should never happen. We enforce at runtime because it's simpler than
5404 : // adding an extra per-tenant data structure to store the things that should be the same
5405 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5406 0 : "Inconsistent shard stripe size parameters!"
5407 0 : )));
5408 0 : }
5409 : }
5410 : }
5411 : }
5412 :
5413 0 : if result.is_empty() {
5414 0 : return Err(ApiError::NotFound(
5415 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
5416 0 : ));
5417 0 : }
5418 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
5419 0 : tracing::info!(
5420 0 : "Located tenant {} with params {:?} on shards {}",
5421 : tenant_id,
5422 : shard_params,
5423 0 : result
5424 0 : .iter()
5425 0 : .map(|s| format!("{s:?}"))
5426 0 : .collect::<Vec<_>>()
5427 0 : .join(",")
5428 : );
5429 :
5430 0 : Ok(TenantLocateResponse {
5431 0 : shards: result,
5432 0 : shard_params,
5433 0 : })
5434 0 : }
5435 :
5436 : /// Returns None if the input iterator of shards does not include a shard with number=0
5437 0 : fn tenant_describe_impl<'a>(
5438 0 : &self,
5439 0 : shards: impl Iterator<Item = &'a TenantShard>,
5440 0 : ) -> Option<TenantDescribeResponse> {
5441 0 : let mut shard_zero = None;
5442 0 : let mut describe_shards = Vec::new();
5443 :
5444 0 : for shard in shards {
5445 0 : if shard.tenant_shard_id.is_shard_zero() {
5446 0 : shard_zero = Some(shard);
5447 0 : }
5448 :
5449 0 : describe_shards.push(TenantDescribeResponseShard {
5450 0 : tenant_shard_id: shard.tenant_shard_id,
5451 0 : node_attached: *shard.intent.get_attached(),
5452 0 : node_secondary: shard.intent.get_secondary().to_vec(),
5453 0 : last_error: shard
5454 0 : .last_error
5455 0 : .lock()
5456 0 : .unwrap()
5457 0 : .as_ref()
5458 0 : .map(|e| format!("{e}"))
5459 0 : .unwrap_or("".to_string())
5460 0 : .clone(),
5461 0 : is_reconciling: shard.reconciler.is_some(),
5462 0 : is_pending_compute_notification: shard.pending_compute_notification,
5463 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
5464 0 : is_importing: shard.importing == TimelineImportState::Importing,
5465 0 : scheduling_policy: shard.get_scheduling_policy(),
5466 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
5467 : })
5468 : }
5469 :
5470 0 : let shard_zero = shard_zero?;
5471 :
5472 0 : Some(TenantDescribeResponse {
5473 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
5474 0 : shards: describe_shards,
5475 0 : stripe_size: shard_zero.shard.stripe_size,
5476 0 : policy: shard_zero.policy.clone(),
5477 0 : config: shard_zero.config.clone(),
5478 0 : })
5479 0 : }
5480 :
5481 0 : pub(crate) fn tenant_describe(
5482 0 : &self,
5483 0 : tenant_id: TenantId,
5484 0 : ) -> Result<TenantDescribeResponse, ApiError> {
5485 0 : let locked = self.inner.read().unwrap();
5486 :
5487 0 : self.tenant_describe_impl(
5488 0 : locked
5489 0 : .tenants
5490 0 : .range(TenantShardId::tenant_range(tenant_id))
5491 0 : .map(|(_k, v)| v),
5492 : )
5493 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
5494 0 : }
5495 :
5496 : /* BEGIN_HADRON */
5497 0 : pub(crate) async fn tenant_timeline_describe(
5498 0 : &self,
5499 0 : tenant_id: TenantId,
5500 0 : timeline_id: TimelineId,
5501 0 : ) -> Result<TenantTimelineDescribeResponse, ApiError> {
5502 0 : self.tenant_remote_mutation(tenant_id, |locations| async move {
5503 0 : if locations.0.is_empty() {
5504 0 : return Err(ApiError::NotFound(
5505 0 : anyhow::anyhow!("Tenant not found").into(),
5506 0 : ));
5507 0 : };
5508 :
5509 0 : let locations: Vec<(TenantShardId, Node)> = locations
5510 0 : .0
5511 0 : .iter()
5512 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
5513 0 : .collect();
5514 0 : let mut futs = FuturesUnordered::new();
5515 :
5516 0 : for (shard_id, node) in locations {
5517 0 : futs.push({
5518 0 : async move {
5519 0 : let result = node
5520 0 : .with_client_retries(
5521 0 : |client| async move {
5522 0 : client
5523 0 : .tenant_timeline_describe(&shard_id, &timeline_id)
5524 0 : .await
5525 0 : },
5526 0 : &self.http_client,
5527 0 : &self.config.pageserver_jwt_token,
5528 : 3,
5529 : 3,
5530 0 : Duration::from_secs(30),
5531 0 : &self.cancel,
5532 : )
5533 0 : .await;
5534 0 : (result, shard_id, node.get_id())
5535 0 : }
5536 : });
5537 : }
5538 :
5539 0 : let mut results: Vec<TimelineInfo> = Vec::new();
5540 0 : while let Some((result, tenant_shard_id, node_id)) = futs.next().await {
5541 0 : match result {
5542 0 : Some(Ok(timeline_info)) => results.push(timeline_info),
5543 0 : Some(Err(e)) => {
5544 0 : tracing::warn!(
5545 0 : "Failed to describe tenant {} timeline {} for pageserver {}: {e}",
5546 : tenant_shard_id,
5547 : timeline_id,
5548 : node_id,
5549 : );
5550 0 : return Err(ApiError::ResourceUnavailable(format!("{e}").into()));
5551 : }
5552 0 : None => return Err(ApiError::Cancelled),
5553 : }
5554 : }
5555 0 : let mut image_consistent_lsn: Option<Lsn> = Some(Lsn::MAX);
5556 0 : for timeline_info in &results {
5557 0 : if let Some(tline_image_consistent_lsn) = timeline_info.image_consistent_lsn {
5558 0 : image_consistent_lsn = Some(std::cmp::min(
5559 0 : image_consistent_lsn.unwrap(),
5560 0 : tline_image_consistent_lsn,
5561 0 : ));
5562 0 : } else {
5563 0 : tracing::warn!(
5564 0 : "Timeline {} on shard {} does not have image consistent lsn",
5565 : timeline_info.timeline_id,
5566 : timeline_info.tenant_id
5567 : );
5568 0 : image_consistent_lsn = None;
5569 0 : break;
5570 : }
5571 : }
5572 :
5573 0 : Ok(TenantTimelineDescribeResponse {
5574 0 : shards: results,
5575 0 : image_consistent_lsn,
5576 0 : })
5577 0 : })
5578 0 : .await?
5579 0 : }
5580 : /* END_HADRON */
5581 :
5582 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
5583 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
5584 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
5585 : /// in our external API.
5586 0 : pub(crate) fn tenant_list(
5587 0 : &self,
5588 0 : limit: Option<usize>,
5589 0 : start_after: Option<TenantId>,
5590 0 : ) -> Vec<TenantDescribeResponse> {
5591 0 : let locked = self.inner.read().unwrap();
5592 :
5593 : // Apply start_from parameter
5594 0 : let shard_range = match start_after {
5595 0 : None => locked.tenants.range(..),
5596 0 : Some(tenant_id) => locked.tenants.range(
5597 0 : TenantShardId {
5598 0 : tenant_id,
5599 0 : shard_number: ShardNumber(u8::MAX),
5600 0 : shard_count: ShardCount(u8::MAX),
5601 0 : }..,
5602 : ),
5603 : };
5604 :
5605 0 : let mut result = Vec::new();
5606 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
5607 0 : result.push(
5608 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
5609 0 : .expect("Groups are always non-empty"),
5610 : );
5611 :
5612 : // Enforce `limit` parameter
5613 0 : if let Some(limit) = limit {
5614 0 : if result.len() >= limit {
5615 0 : break;
5616 0 : }
5617 0 : }
5618 : }
5619 :
5620 0 : result
5621 0 : }
5622 :
5623 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
5624 : async fn abort_tenant_shard_split(
5625 : &self,
5626 : op: &TenantShardSplitAbort,
5627 : ) -> Result<(), TenantShardSplitAbortError> {
5628 : // Cleaning up a split:
5629 : // - Parent shards are not destroyed during a split, just detached.
5630 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
5631 : // just the children attached, or both.
5632 : //
5633 : // Therefore our work to do is to:
5634 : // 1. Clean up storage controller's internal state to just refer to parents, no children
5635 : // 2. Call out to pageservers to ensure that children are detached
5636 : // 3. Call out to pageservers to ensure that parents are attached.
5637 : //
5638 : // Crash safety:
5639 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
5640 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
5641 : // and detach them.
5642 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
5643 : // from our database, then we will re-enter this cleanup routine on startup.
5644 :
5645 : let TenantShardSplitAbort {
5646 : tenant_id,
5647 : new_shard_count,
5648 : new_stripe_size,
5649 : ..
5650 : } = op;
5651 :
5652 : // First abort persistent state, if any exists.
5653 : match self
5654 : .persistence
5655 : .abort_shard_split(*tenant_id, *new_shard_count)
5656 : .await?
5657 : {
5658 : AbortShardSplitStatus::Aborted => {
5659 : // Proceed to roll back any child shards created on pageservers
5660 : }
5661 : AbortShardSplitStatus::Complete => {
5662 : // The split completed (we might hit that path if e.g. our database transaction
5663 : // to write the completion landed in the database, but we dropped connection
5664 : // before seeing the result).
5665 : //
5666 : // We must update in-memory state to reflect the successful split.
5667 : self.tenant_shard_split_commit_inmem(
5668 : *tenant_id,
5669 : *new_shard_count,
5670 : *new_stripe_size,
5671 : );
5672 : return Ok(());
5673 : }
5674 : }
5675 :
5676 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
5677 : let detach_locations: Vec<(Node, TenantShardId)> = {
5678 : let mut detach_locations = Vec::new();
5679 : let mut locked = self.inner.write().unwrap();
5680 : let (nodes, tenants, scheduler) = locked.parts_mut();
5681 :
5682 : for (tenant_shard_id, shard) in
5683 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
5684 : {
5685 : if shard.shard.count == op.new_shard_count {
5686 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
5687 : // is infallible, so if we got an error we shouldn't have got that far.
5688 : tracing::warn!(
5689 : "During split abort, child shard {tenant_shard_id} found in-memory"
5690 : );
5691 : continue;
5692 : }
5693 :
5694 : // Add the children of this shard to this list of things to detach
5695 : if let Some(node_id) = shard.intent.get_attached() {
5696 : for child_id in tenant_shard_id.split(*new_shard_count) {
5697 : detach_locations.push((
5698 : nodes
5699 : .get(node_id)
5700 : .expect("Intent references nonexistent node")
5701 : .clone(),
5702 : child_id,
5703 : ));
5704 : }
5705 : } else {
5706 : tracing::warn!(
5707 : "During split abort, shard {tenant_shard_id} has no attached location"
5708 : );
5709 : }
5710 :
5711 : tracing::info!("Restoring parent shard {tenant_shard_id}");
5712 :
5713 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
5714 : // if the original attachment location is offline.
5715 : if let Some(node_id) = shard.intent.get_attached() {
5716 : if !nodes.get(node_id).unwrap().is_available() {
5717 : tracing::info!(
5718 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
5719 : );
5720 : shard.intent.demote_attached(scheduler, *node_id);
5721 : }
5722 : }
5723 : for node_id in shard.intent.get_secondary().clone() {
5724 : if !nodes.get(&node_id).unwrap().is_available() {
5725 : tracing::info!(
5726 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
5727 : );
5728 : shard.intent.remove_secondary(scheduler, node_id);
5729 : }
5730 : }
5731 :
5732 : shard.splitting = SplitState::Idle;
5733 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
5734 : // If this shard can't be scheduled now (perhaps due to offline nodes or
5735 : // capacity issues), that must not prevent us rolling back a split. In this
5736 : // case it should be eventually scheduled in the background.
5737 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
5738 : }
5739 :
5740 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
5741 : }
5742 :
5743 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
5744 : tenants
5745 0 : .retain(|id, s| !(id.tenant_id == *tenant_id && s.shard.count == *new_shard_count));
5746 :
5747 : detach_locations
5748 : };
5749 :
5750 : for (node, child_id) in detach_locations {
5751 : if !node.is_available() {
5752 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
5753 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
5754 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
5755 : // them from the node.
5756 : tracing::warn!(
5757 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
5758 : );
5759 : continue;
5760 : }
5761 :
5762 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
5763 : // a 503 and retry, up to our limit.
5764 : tracing::info!("Detaching {child_id} on {node}...");
5765 : match node
5766 : .with_client_retries(
5767 0 : |client| async move {
5768 0 : let config = LocationConfig {
5769 0 : mode: LocationConfigMode::Detached,
5770 0 : generation: None,
5771 0 : secondary_conf: None,
5772 0 : shard_number: child_id.shard_number.0,
5773 0 : shard_count: child_id.shard_count.literal(),
5774 0 : // Stripe size and tenant config don't matter when detaching
5775 0 : shard_stripe_size: 0,
5776 0 : tenant_conf: TenantConfig::default(),
5777 0 : };
5778 :
5779 0 : client.location_config(child_id, config, None, false).await
5780 0 : },
5781 : &self.http_client,
5782 : &self.config.pageserver_jwt_token,
5783 : 1,
5784 : 10,
5785 : Duration::from_secs(5),
5786 : &self.reconcilers_cancel,
5787 : )
5788 : .await
5789 : {
5790 : Some(Ok(_)) => {}
5791 : Some(Err(e)) => {
5792 : // We failed to communicate with the remote node. This is problematic: we may be
5793 : // leaving it with a rogue child shard.
5794 : tracing::warn!(
5795 : "Failed to detach child {child_id} from node {node} during abort"
5796 : );
5797 : return Err(e.into());
5798 : }
5799 : None => {
5800 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
5801 : // clean up on restart. The node going offline requires a retry.
5802 : return Err(TenantShardSplitAbortError::Unavailable);
5803 : }
5804 : };
5805 : }
5806 :
5807 : tracing::info!("Successfully aborted split");
5808 : Ok(())
5809 : }
5810 :
5811 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
5812 : /// of the tenant map to reflect the child shards that exist after the split.
5813 0 : fn tenant_shard_split_commit_inmem(
5814 0 : &self,
5815 0 : tenant_id: TenantId,
5816 0 : new_shard_count: ShardCount,
5817 0 : new_stripe_size: Option<ShardStripeSize>,
5818 0 : ) -> (
5819 0 : TenantShardSplitResponse,
5820 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
5821 0 : Vec<ReconcilerWaiter>,
5822 0 : ) {
5823 0 : let mut response = TenantShardSplitResponse {
5824 0 : new_shards: Vec::new(),
5825 0 : };
5826 0 : let mut child_locations = Vec::new();
5827 0 : let mut waiters = Vec::new();
5828 :
5829 : {
5830 0 : let mut locked = self.inner.write().unwrap();
5831 :
5832 0 : let parent_ids = locked
5833 0 : .tenants
5834 0 : .range(TenantShardId::tenant_range(tenant_id))
5835 0 : .map(|(shard_id, _)| *shard_id)
5836 0 : .collect::<Vec<_>>();
5837 :
5838 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5839 0 : for parent_id in parent_ids {
5840 0 : let child_ids = parent_id.split(new_shard_count);
5841 :
5842 : let (
5843 0 : pageserver,
5844 0 : generation,
5845 0 : policy,
5846 0 : parent_ident,
5847 0 : config,
5848 0 : preferred_az,
5849 0 : secondary_count,
5850 : ) = {
5851 0 : let mut old_state = tenants
5852 0 : .remove(&parent_id)
5853 0 : .expect("It was present, we just split it");
5854 :
5855 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
5856 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
5857 : // nothing else can clear this.
5858 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
5859 :
5860 0 : let old_attached = old_state.intent.get_attached().unwrap();
5861 0 : old_state.intent.clear(scheduler);
5862 0 : let generation = old_state.generation.expect("Shard must have been attached");
5863 0 : (
5864 0 : old_attached,
5865 0 : generation,
5866 0 : old_state.policy.clone(),
5867 0 : old_state.shard,
5868 0 : old_state.config.clone(),
5869 0 : old_state.preferred_az().cloned(),
5870 0 : old_state.intent.get_secondary().len(),
5871 0 : )
5872 : };
5873 :
5874 0 : let mut schedule_context = ScheduleContext::default();
5875 0 : for child in child_ids {
5876 0 : let mut child_shard = parent_ident;
5877 0 : child_shard.number = child.shard_number;
5878 0 : child_shard.count = child.shard_count;
5879 0 : if let Some(stripe_size) = new_stripe_size {
5880 0 : child_shard.stripe_size = stripe_size;
5881 0 : }
5882 :
5883 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5884 0 : child_observed.insert(
5885 0 : pageserver,
5886 0 : ObservedStateLocation {
5887 0 : conf: Some(attached_location_conf(
5888 0 : generation,
5889 0 : &child_shard,
5890 0 : &config,
5891 0 : &policy,
5892 0 : secondary_count,
5893 0 : )),
5894 0 : },
5895 : );
5896 :
5897 0 : let mut child_state =
5898 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
5899 0 : child_state.intent =
5900 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
5901 0 : child_state.observed = ObservedState {
5902 0 : locations: child_observed,
5903 0 : };
5904 0 : child_state.generation = Some(generation);
5905 0 : child_state.config = config.clone();
5906 :
5907 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
5908 : // as at this point in the split process we have succeeded and this part is infallible:
5909 : // we will never need to do any special recovery from this state.
5910 :
5911 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
5912 :
5913 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
5914 : // This is not fatal, because we've implicitly already got an attached
5915 : // location for the child shard. Failure here just means we couldn't
5916 : // find a secondary (e.g. because cluster is overloaded).
5917 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
5918 0 : }
5919 : // In the background, attach secondary locations for the new shards
5920 0 : if let Some(waiter) = self.maybe_reconcile_shard(
5921 0 : &mut child_state,
5922 0 : nodes,
5923 0 : ReconcilerPriority::High,
5924 0 : ) {
5925 0 : waiters.push(waiter);
5926 0 : }
5927 :
5928 0 : tenants.insert(child, child_state);
5929 0 : response.new_shards.push(child);
5930 : }
5931 : }
5932 0 : (response, child_locations, waiters)
5933 : }
5934 0 : }
5935 :
5936 0 : async fn tenant_shard_split_start_secondaries(
5937 0 : &self,
5938 0 : tenant_id: TenantId,
5939 0 : waiters: Vec<ReconcilerWaiter>,
5940 0 : ) {
5941 : // Wait for initial reconcile of child shards, this creates the secondary locations
5942 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5943 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
5944 : // their secondaries couldn't be attached.
5945 0 : tracing::warn!("Failed to reconcile after split: {e}");
5946 0 : return;
5947 0 : }
5948 :
5949 : // Take the state lock to discover the attached & secondary intents for all shards
5950 0 : let (attached, secondary) = {
5951 0 : let locked = self.inner.read().unwrap();
5952 0 : let mut attached = Vec::new();
5953 0 : let mut secondary = Vec::new();
5954 :
5955 0 : for (tenant_shard_id, shard) in
5956 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5957 : {
5958 0 : let Some(node_id) = shard.intent.get_attached() else {
5959 : // Unexpected. Race with a PlacementPolicy change?
5960 0 : tracing::warn!(
5961 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
5962 : );
5963 0 : continue;
5964 : };
5965 :
5966 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
5967 : // No secondary location. Nothing for us to do.
5968 0 : continue;
5969 : };
5970 :
5971 0 : let attached_node = locked
5972 0 : .nodes
5973 0 : .get(node_id)
5974 0 : .expect("Pageservers may not be deleted while referenced");
5975 :
5976 0 : let secondary_node = locked
5977 0 : .nodes
5978 0 : .get(secondary_node_id)
5979 0 : .expect("Pageservers may not be deleted while referenced");
5980 :
5981 0 : attached.push((*tenant_shard_id, attached_node.clone()));
5982 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
5983 : }
5984 0 : (attached, secondary)
5985 : };
5986 :
5987 0 : if secondary.is_empty() {
5988 : // No secondary locations; nothing for us to do
5989 0 : return;
5990 0 : }
5991 :
5992 0 : for (_, result) in self
5993 0 : .tenant_for_shards_api(
5994 0 : attached,
5995 0 : |tenant_shard_id, client| async move {
5996 0 : client.tenant_heatmap_upload(tenant_shard_id).await
5997 0 : },
5998 : 1,
5999 : 1,
6000 : SHORT_RECONCILE_TIMEOUT,
6001 0 : &self.cancel,
6002 : )
6003 0 : .await
6004 : {
6005 0 : if let Err(e) = result {
6006 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
6007 0 : return;
6008 0 : }
6009 : }
6010 :
6011 0 : for (_, result) in self
6012 0 : .tenant_for_shards_api(
6013 0 : secondary,
6014 0 : |tenant_shard_id, client| async move {
6015 0 : client
6016 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
6017 0 : .await
6018 0 : },
6019 : 1,
6020 : 1,
6021 : SHORT_RECONCILE_TIMEOUT,
6022 0 : &self.cancel,
6023 : )
6024 0 : .await
6025 : {
6026 0 : if let Err(e) = result {
6027 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
6028 0 : return;
6029 0 : }
6030 : }
6031 0 : }
6032 :
6033 0 : pub(crate) async fn tenant_shard_split(
6034 0 : &self,
6035 0 : tenant_id: TenantId,
6036 0 : split_req: TenantShardSplitRequest,
6037 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
6038 : // TODO: return 503 if we get stuck waiting for this lock
6039 : // (issue https://github.com/neondatabase/neon/issues/7108)
6040 0 : let _tenant_lock = trace_exclusive_lock(
6041 0 : &self.tenant_op_locks,
6042 0 : tenant_id,
6043 0 : TenantOperations::ShardSplit,
6044 0 : )
6045 0 : .await;
6046 :
6047 0 : let _gate = self
6048 0 : .reconcilers_gate
6049 0 : .enter()
6050 0 : .map_err(|_| ApiError::ShuttingDown)?;
6051 :
6052 : // Timeline imports on the pageserver side can't handle shard-splits.
6053 : // If the tenant is importing a timeline, dont't shard split it.
6054 0 : match self
6055 0 : .persistence
6056 0 : .is_tenant_importing_timeline(tenant_id)
6057 0 : .await
6058 : {
6059 0 : Ok(importing) => {
6060 0 : if importing {
6061 0 : return Err(ApiError::Conflict(
6062 0 : "Cannot shard split during timeline import".to_string(),
6063 0 : ));
6064 0 : }
6065 : }
6066 0 : Err(err) => {
6067 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6068 0 : "Failed to check for running imports: {err}"
6069 0 : )));
6070 : }
6071 : }
6072 :
6073 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
6074 0 : let new_stripe_size = split_req.new_stripe_size;
6075 :
6076 : // Validate the request and construct parameters. This phase is fallible, but does not require
6077 : // rollback on errors, as it does no I/O and mutates no state.
6078 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
6079 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
6080 0 : ShardSplitAction::Split(params) => params,
6081 : };
6082 :
6083 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
6084 : // we must roll back.
6085 0 : let r = self
6086 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
6087 0 : .await;
6088 :
6089 0 : let (response, waiters) = match r {
6090 0 : Ok(r) => r,
6091 0 : Err(e) => {
6092 : // Split might be part-done, we must do work to abort it.
6093 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
6094 0 : self.abort_tx
6095 0 : .send(TenantShardSplitAbort {
6096 0 : tenant_id,
6097 0 : new_shard_count,
6098 0 : new_stripe_size,
6099 0 : _tenant_lock,
6100 0 : _gate,
6101 0 : })
6102 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
6103 0 : .ok();
6104 0 : return Err(e);
6105 : }
6106 : };
6107 :
6108 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
6109 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
6110 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
6111 : // in [`Self::optimize_all`]
6112 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
6113 0 : .await;
6114 0 : Ok(response)
6115 0 : }
6116 :
6117 0 : fn prepare_tenant_shard_split(
6118 0 : &self,
6119 0 : tenant_id: TenantId,
6120 0 : split_req: TenantShardSplitRequest,
6121 0 : ) -> Result<ShardSplitAction, ApiError> {
6122 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
6123 0 : anyhow::anyhow!("failpoint")
6124 0 : )));
6125 :
6126 0 : let mut policy = None;
6127 0 : let mut config = None;
6128 0 : let mut shard_ident = None;
6129 0 : let mut preferred_az_id = None;
6130 : // Validate input, and calculate which shards we will create
6131 0 : let (old_shard_count, targets) =
6132 : {
6133 0 : let locked = self.inner.read().unwrap();
6134 :
6135 0 : let pageservers = locked.nodes.clone();
6136 :
6137 0 : let mut targets = Vec::new();
6138 :
6139 : // In case this is a retry, count how many already-split shards we found
6140 0 : let mut children_found = Vec::new();
6141 0 : let mut old_shard_count = None;
6142 :
6143 0 : for (tenant_shard_id, shard) in
6144 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
6145 : {
6146 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
6147 : Ordering::Equal => {
6148 : // Already split this
6149 0 : children_found.push(*tenant_shard_id);
6150 0 : continue;
6151 : }
6152 : Ordering::Greater => {
6153 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6154 0 : "Requested count {} but already have shards at count {}",
6155 0 : split_req.new_shard_count,
6156 0 : shard.shard.count.count()
6157 0 : )));
6158 : }
6159 0 : Ordering::Less => {
6160 0 : // Fall through: this shard has lower count than requested,
6161 0 : // is a candidate for splitting.
6162 0 : }
6163 : }
6164 :
6165 0 : match old_shard_count {
6166 0 : None => old_shard_count = Some(shard.shard.count),
6167 0 : Some(old_shard_count) => {
6168 0 : if old_shard_count != shard.shard.count {
6169 : // We may hit this case if a caller asked for two splits to
6170 : // different sizes, before the first one is complete.
6171 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
6172 : // of shard_count=1 and shard_count=2 shards in the map.
6173 0 : return Err(ApiError::Conflict(
6174 0 : "Cannot split, currently mid-split".to_string(),
6175 0 : ));
6176 0 : }
6177 : }
6178 : }
6179 0 : if policy.is_none() {
6180 0 : policy = Some(shard.policy.clone());
6181 0 : }
6182 0 : if shard_ident.is_none() {
6183 0 : shard_ident = Some(shard.shard);
6184 0 : }
6185 0 : if config.is_none() {
6186 0 : config = Some(shard.config.clone());
6187 0 : }
6188 0 : if preferred_az_id.is_none() {
6189 0 : preferred_az_id = shard.preferred_az().cloned();
6190 0 : }
6191 :
6192 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
6193 0 : tracing::info!(
6194 0 : "Tenant shard {} already has shard count {}",
6195 : tenant_shard_id,
6196 : split_req.new_shard_count
6197 : );
6198 0 : continue;
6199 0 : }
6200 :
6201 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
6202 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
6203 0 : ))?;
6204 :
6205 0 : let node = pageservers
6206 0 : .get(&node_id)
6207 0 : .expect("Pageservers may not be deleted while referenced");
6208 :
6209 0 : targets.push(ShardSplitTarget {
6210 0 : parent_id: *tenant_shard_id,
6211 0 : node: node.clone(),
6212 0 : child_ids: tenant_shard_id
6213 0 : .split(ShardCount::new(split_req.new_shard_count)),
6214 0 : });
6215 : }
6216 :
6217 0 : if targets.is_empty() {
6218 0 : if children_found.len() == split_req.new_shard_count as usize {
6219 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
6220 0 : new_shards: children_found,
6221 0 : }));
6222 : } else {
6223 : // No shards found to split, and no existing children found: the
6224 : // tenant doesn't exist at all.
6225 0 : return Err(ApiError::NotFound(
6226 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
6227 0 : ));
6228 : }
6229 0 : }
6230 :
6231 0 : (old_shard_count, targets)
6232 : };
6233 :
6234 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
6235 0 : let old_shard_count = old_shard_count.unwrap();
6236 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
6237 : // This ShardIdentity will be used as the template for all children, so this implicitly
6238 : // applies the new stripe size to the children.
6239 0 : let mut shard_ident = shard_ident.unwrap();
6240 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
6241 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6242 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
6243 0 : shard_ident.stripe_size
6244 0 : )));
6245 0 : }
6246 :
6247 0 : shard_ident.stripe_size = new_stripe_size;
6248 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
6249 0 : shard_ident
6250 : } else {
6251 0 : shard_ident.unwrap()
6252 : };
6253 0 : let policy = policy.unwrap();
6254 0 : let config = config.unwrap();
6255 :
6256 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
6257 0 : old_shard_count,
6258 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
6259 0 : new_stripe_size: split_req.new_stripe_size,
6260 0 : targets,
6261 0 : policy,
6262 0 : config,
6263 0 : shard_ident,
6264 0 : preferred_az_id,
6265 0 : })))
6266 0 : }
6267 :
6268 0 : async fn do_tenant_shard_split(
6269 0 : &self,
6270 0 : tenant_id: TenantId,
6271 0 : params: Box<ShardSplitParams>,
6272 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
6273 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
6274 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
6275 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
6276 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
6277 : // (https://github.com/neondatabase/neon/issues/6676)
6278 :
6279 : let ShardSplitParams {
6280 0 : old_shard_count,
6281 0 : new_shard_count,
6282 0 : new_stripe_size,
6283 0 : mut targets,
6284 0 : policy,
6285 0 : config,
6286 0 : shard_ident,
6287 0 : preferred_az_id,
6288 0 : } = *params;
6289 :
6290 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
6291 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
6292 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
6293 : // at the time of split.
6294 0 : let waiters = {
6295 0 : let mut locked = self.inner.write().unwrap();
6296 0 : let mut waiters = Vec::new();
6297 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6298 0 : for target in &mut targets {
6299 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
6300 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6301 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6302 0 : "Shard {} not found",
6303 0 : target.parent_id
6304 0 : )));
6305 : };
6306 :
6307 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
6308 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6309 0 : return Err(ApiError::Conflict(format!(
6310 0 : "Shard {} unexpectedly rescheduled during split",
6311 0 : target.parent_id
6312 0 : )));
6313 0 : }
6314 :
6315 : // Irrespective of PlacementPolicy, clear secondary locations from intent
6316 0 : shard.intent.clear_secondary(scheduler);
6317 :
6318 : // Run Reconciler to execute detach fo secondary locations.
6319 0 : if let Some(waiter) =
6320 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6321 0 : {
6322 0 : waiters.push(waiter);
6323 0 : }
6324 : }
6325 0 : waiters
6326 : };
6327 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
6328 :
6329 : // Before creating any new child shards in memory or on the pageservers, persist them: this
6330 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
6331 : // acts as the protection against two concurrent attempts to split: one of them will get a database
6332 : // error trying to insert the child shards.
6333 0 : let mut child_tsps = Vec::new();
6334 0 : for target in &targets {
6335 0 : let mut this_child_tsps = Vec::new();
6336 0 : for child in &target.child_ids {
6337 0 : let mut child_shard = shard_ident;
6338 0 : child_shard.number = child.shard_number;
6339 0 : child_shard.count = child.shard_count;
6340 :
6341 0 : tracing::info!(
6342 0 : "Create child shard persistence with stripe size {}",
6343 : shard_ident.stripe_size.0
6344 : );
6345 :
6346 0 : this_child_tsps.push(TenantShardPersistence {
6347 0 : tenant_id: child.tenant_id.to_string(),
6348 0 : shard_number: child.shard_number.0 as i32,
6349 0 : shard_count: child.shard_count.literal() as i32,
6350 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
6351 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
6352 : // populate the correct generation as part of its transaction, to protect us
6353 : // against racing with changes in the state of the parent.
6354 0 : generation: None,
6355 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
6356 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
6357 0 : config: serde_json::to_string(&config).unwrap(),
6358 0 : splitting: SplitState::Splitting,
6359 :
6360 : // Scheduling policies and preferred AZ do not carry through to children
6361 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
6362 0 : .unwrap(),
6363 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
6364 : });
6365 : }
6366 :
6367 0 : child_tsps.push((target.parent_id, this_child_tsps));
6368 : }
6369 :
6370 0 : if let Err(e) = self
6371 0 : .persistence
6372 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
6373 0 : .await
6374 : {
6375 0 : match e {
6376 : DatabaseError::Query(diesel::result::Error::DatabaseError(
6377 : DatabaseErrorKind::UniqueViolation,
6378 : _,
6379 : )) => {
6380 : // Inserting a child shard violated a unique constraint: we raced with another call to
6381 : // this function
6382 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
6383 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
6384 : }
6385 0 : _ => return Err(ApiError::InternalServerError(e.into())),
6386 : }
6387 0 : }
6388 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
6389 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6390 : ));
6391 :
6392 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
6393 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
6394 : // is not set in memory, then it was not persisted.
6395 : {
6396 0 : let mut locked = self.inner.write().unwrap();
6397 0 : for target in &targets {
6398 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
6399 0 : parent_shard.splitting = SplitState::Splitting;
6400 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
6401 0 : // split operation.
6402 0 : parent_shard
6403 0 : .observed
6404 0 : .locations
6405 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
6406 0 : }
6407 : }
6408 : }
6409 :
6410 : // TODO: issue split calls concurrently (this only matters once we're splitting
6411 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
6412 :
6413 : // HADRON: set a timeout for splitting individual shards on page servers.
6414 : // Currently we do not perform any retry because it's not clear if page server can handle
6415 : // partially split shards correctly.
6416 0 : let shard_split_timeout =
6417 0 : if let Some(env::DeploymentMode::Local) = env::get_deployment_mode() {
6418 0 : Duration::from_secs(30)
6419 : } else {
6420 0 : self.config.shard_split_request_timeout
6421 : };
6422 0 : let mut http_client_builder = reqwest::ClientBuilder::new()
6423 0 : .pool_max_idle_per_host(0)
6424 0 : .timeout(shard_split_timeout);
6425 :
6426 0 : for ssl_ca_cert in &self.config.ssl_ca_certs {
6427 0 : http_client_builder = http_client_builder.add_root_certificate(ssl_ca_cert.clone());
6428 0 : }
6429 0 : let http_client = http_client_builder
6430 0 : .build()
6431 0 : .expect("Failed to construct HTTP client");
6432 0 : for target in &targets {
6433 : let ShardSplitTarget {
6434 0 : parent_id,
6435 0 : node,
6436 0 : child_ids,
6437 0 : } = target;
6438 :
6439 0 : let client = PageserverClient::new(
6440 0 : node.get_id(),
6441 0 : http_client.clone(),
6442 0 : node.base_url(),
6443 0 : self.config.pageserver_jwt_token.as_deref(),
6444 : );
6445 :
6446 0 : let response = client
6447 0 : .tenant_shard_split(
6448 0 : *parent_id,
6449 0 : TenantShardSplitRequest {
6450 0 : new_shard_count: new_shard_count.literal(),
6451 0 : new_stripe_size,
6452 0 : },
6453 0 : )
6454 0 : .await
6455 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {parent_id}: {e}")))?;
6456 :
6457 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
6458 0 : "failpoint".to_string()
6459 0 : )));
6460 :
6461 0 : failpoint_support::sleep_millis_async!(
6462 : "shard-split-post-remote-sleep",
6463 0 : &self.reconcilers_cancel
6464 : );
6465 :
6466 0 : tracing::info!(
6467 0 : "Split {} into {}",
6468 : parent_id,
6469 0 : response
6470 0 : .new_shards
6471 0 : .iter()
6472 0 : .map(|s| format!("{s:?}"))
6473 0 : .collect::<Vec<_>>()
6474 0 : .join(",")
6475 : );
6476 :
6477 0 : if &response.new_shards != child_ids {
6478 : // This should never happen: the pageserver should agree with us on how shard splits work.
6479 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6480 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
6481 0 : parent_id,
6482 0 : response.new_shards,
6483 0 : child_ids
6484 0 : )));
6485 0 : }
6486 : }
6487 :
6488 0 : fail::fail_point!("shard-split-pre-complete", |_| Err(ApiError::Conflict(
6489 0 : "failpoint".to_string()
6490 0 : )));
6491 :
6492 0 : pausable_failpoint!("shard-split-pre-complete-pause");
6493 :
6494 : // TODO: if the pageserver restarted concurrently with our split API call,
6495 : // the actual generation of the child shard might differ from the generation
6496 : // we expect it to have. In order for our in-database generation to end up
6497 : // correct, we should carry the child generation back in the response and apply it here
6498 : // in complete_shard_split (and apply the correct generation in memory)
6499 : // (or, we can carry generation in the request and reject the request if
6500 : // it doesn't match, but that requires more retry logic on this side)
6501 :
6502 0 : self.persistence
6503 0 : .complete_shard_split(tenant_id, old_shard_count, new_shard_count)
6504 0 : .await?;
6505 :
6506 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
6507 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6508 : ));
6509 :
6510 : // Replace all the shards we just split with their children: this phase is infallible.
6511 0 : let (response, child_locations, waiters) =
6512 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
6513 :
6514 : // Notify all page servers to detach and clean up the old shards because they will no longer
6515 : // be needed. This is best-effort: if it fails, it will be cleaned up on a subsequent
6516 : // Pageserver re-attach/startup.
6517 0 : let shards_to_cleanup = targets
6518 0 : .iter()
6519 0 : .map(|target| (target.parent_id, target.node.get_id()))
6520 0 : .collect();
6521 0 : self.cleanup_locations(shards_to_cleanup).await;
6522 :
6523 : // Send compute notifications for all the new shards
6524 0 : let mut failed_notifications = Vec::new();
6525 0 : for (child_id, child_ps, stripe_size) in child_locations {
6526 0 : if let Err(e) = self
6527 0 : .compute_hook
6528 0 : .notify_attach(
6529 0 : compute_hook::ShardUpdate {
6530 0 : tenant_shard_id: child_id,
6531 0 : node_id: child_ps,
6532 0 : stripe_size,
6533 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
6534 0 : },
6535 0 : &self.reconcilers_cancel,
6536 0 : )
6537 0 : .await
6538 : {
6539 0 : tracing::warn!(
6540 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
6541 : child_id,
6542 : child_ps
6543 : );
6544 0 : failed_notifications.push(child_id);
6545 0 : }
6546 : }
6547 :
6548 : // If we failed any compute notifications, make a note to retry later.
6549 0 : if !failed_notifications.is_empty() {
6550 0 : let mut locked = self.inner.write().unwrap();
6551 0 : for failed in failed_notifications {
6552 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
6553 0 : shard.pending_compute_notification = true;
6554 0 : }
6555 : }
6556 0 : }
6557 :
6558 0 : Ok((response, waiters))
6559 0 : }
6560 :
6561 : /// A graceful migration: update the preferred node and let optimisation handle the migration
6562 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
6563 : ///
6564 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
6565 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
6566 0 : fn tenant_shard_migrate_with_prewarm(
6567 0 : &self,
6568 0 : migrate_req: &TenantShardMigrateRequest,
6569 0 : shard: &mut TenantShard,
6570 0 : scheduler: &mut Scheduler,
6571 0 : schedule_context: ScheduleContext,
6572 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
6573 0 : shard.set_preferred_node(Some(migrate_req.node_id));
6574 :
6575 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
6576 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
6577 : // e.g. by checking secondary is warm enough.
6578 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
6579 0 : }
6580 :
6581 : /// Immediate migration: directly update the intent state and kick off a reconciler
6582 0 : fn tenant_shard_migrate_immediate(
6583 0 : &self,
6584 0 : migrate_req: &TenantShardMigrateRequest,
6585 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6586 0 : shard: &mut TenantShard,
6587 0 : scheduler: &mut Scheduler,
6588 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
6589 : // Non-graceful migration: update the intent state immediately
6590 0 : let old_attached = *shard.intent.get_attached();
6591 0 : match shard.policy {
6592 0 : PlacementPolicy::Attached(n) => {
6593 : // If our new attached node was a secondary, it no longer should be.
6594 0 : shard
6595 0 : .intent
6596 0 : .remove_secondary(scheduler, migrate_req.node_id);
6597 :
6598 0 : shard
6599 0 : .intent
6600 0 : .set_attached(scheduler, Some(migrate_req.node_id));
6601 :
6602 : // If we were already attached to something, demote that to a secondary
6603 0 : if let Some(old_attached) = old_attached {
6604 0 : if n > 0 {
6605 : // Remove other secondaries to make room for the location we'll demote
6606 0 : while shard.intent.get_secondary().len() >= n {
6607 0 : shard.intent.pop_secondary(scheduler);
6608 0 : }
6609 :
6610 0 : shard.intent.push_secondary(scheduler, old_attached);
6611 0 : }
6612 0 : }
6613 : }
6614 0 : PlacementPolicy::Secondary => {
6615 0 : shard.intent.clear(scheduler);
6616 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6617 0 : }
6618 : PlacementPolicy::Detached => {
6619 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6620 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
6621 0 : )));
6622 : }
6623 : }
6624 :
6625 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
6626 0 : shard.sequence = shard.sequence.next();
6627 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
6628 0 : Ok(self.maybe_configured_reconcile_shard(
6629 0 : shard,
6630 0 : nodes,
6631 0 : (&migrate_req.migration_config).into(),
6632 0 : ))
6633 0 : }
6634 :
6635 0 : pub(crate) async fn tenant_shard_migrate(
6636 0 : &self,
6637 0 : tenant_shard_id: TenantShardId,
6638 0 : migrate_req: TenantShardMigrateRequest,
6639 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6640 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
6641 : // get a different outcome to handle
6642 : enum MigrationOutcome {
6643 : Optimization(Option<ScheduleOptimization>),
6644 : Reconcile(Option<ReconcilerWaiter>),
6645 : }
6646 :
6647 0 : let outcome = {
6648 0 : let mut locked = self.inner.write().unwrap();
6649 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6650 :
6651 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6652 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6653 0 : "Node {} not found",
6654 0 : migrate_req.node_id
6655 0 : )));
6656 : };
6657 :
6658 : // Migration to unavavailable node requires force flag
6659 0 : if !node.is_available() {
6660 0 : if migrate_req.migration_config.override_scheduler {
6661 : // Warn but proceed: the caller may intend to manually adjust the placement of
6662 : // a shard even if the node is down, e.g. if intervening during an incident.
6663 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
6664 : } else {
6665 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
6666 0 : return Err(ApiError::PreconditionFailed(
6667 0 : format!("Node {node} is unavailable").into_boxed_str(),
6668 0 : ));
6669 : }
6670 0 : }
6671 :
6672 : // Calculate the ScheduleContext for this tenant
6673 0 : let mut schedule_context = ScheduleContext::default();
6674 0 : for (_shard_id, shard) in
6675 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
6676 0 : {
6677 0 : schedule_context.avoid(&shard.intent.all_pageservers());
6678 0 : }
6679 :
6680 : // Look up the specific shard we will migrate
6681 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6682 0 : return Err(ApiError::NotFound(
6683 0 : anyhow::anyhow!("Tenant shard not found").into(),
6684 0 : ));
6685 : };
6686 :
6687 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
6688 : // be migrated back by the optimiser.
6689 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
6690 0 : scheduler,
6691 0 : &schedule_context,
6692 0 : migrate_req.node_id,
6693 0 : &[],
6694 0 : ) {
6695 0 : if !migrate_req.migration_config.override_scheduler {
6696 0 : return Err(ApiError::PreconditionFailed(
6697 0 : "Migration to a worse-scoring node".into(),
6698 0 : ));
6699 : } else {
6700 0 : tracing::info!(
6701 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
6702 : migrate_req.node_id
6703 : );
6704 : }
6705 0 : }
6706 :
6707 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
6708 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
6709 0 : return Err(ApiError::PreconditionFailed(
6710 0 : format!(
6711 0 : "Migration expected to originate from {} but shard is on {:?}",
6712 0 : origin_node_id,
6713 0 : shard.intent.get_attached()
6714 0 : )
6715 0 : .into(),
6716 0 : ));
6717 0 : }
6718 0 : }
6719 :
6720 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6721 : // No-op case: we will still proceed to wait for reconciliation in case it is
6722 : // incomplete from an earlier update to the intent.
6723 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
6724 :
6725 : // An instruction to migrate to the currently attached node should
6726 : // cancel any pending graceful migration
6727 0 : shard.set_preferred_node(None);
6728 :
6729 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
6730 0 : shard,
6731 0 : nodes,
6732 0 : (&migrate_req.migration_config).into(),
6733 0 : ))
6734 0 : } else if migrate_req.migration_config.prewarm {
6735 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
6736 0 : &migrate_req,
6737 0 : shard,
6738 0 : scheduler,
6739 0 : schedule_context,
6740 0 : )?)
6741 : } else {
6742 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
6743 0 : &migrate_req,
6744 0 : nodes,
6745 0 : shard,
6746 0 : scheduler,
6747 0 : )?)
6748 : }
6749 : };
6750 :
6751 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
6752 0 : let waiter = match outcome {
6753 0 : MigrationOutcome::Optimization(Some(optimization)) => {
6754 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
6755 : // we might as well do it more promptly as this is a direct external request.
6756 0 : let mut validated = self
6757 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
6758 0 : .await;
6759 0 : if let Some((_shard_id, optimization)) = validated.pop() {
6760 0 : let mut locked = self.inner.write().unwrap();
6761 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6762 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6763 : // Rare but possible: tenant is removed between generating optimisation and validating it.
6764 0 : return Err(ApiError::NotFound(
6765 0 : anyhow::anyhow!("Tenant shard not found").into(),
6766 0 : ));
6767 : };
6768 :
6769 0 : if !shard.apply_optimization(scheduler, optimization) {
6770 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
6771 : // and therefore not applied.
6772 0 : tracing::warn!(
6773 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
6774 : );
6775 0 : }
6776 0 : self.maybe_configured_reconcile_shard(
6777 0 : shard,
6778 0 : nodes,
6779 0 : (&migrate_req.migration_config).into(),
6780 : )
6781 : } else {
6782 0 : None
6783 : }
6784 : }
6785 0 : MigrationOutcome::Optimization(None) => None,
6786 0 : MigrationOutcome::Reconcile(waiter) => waiter,
6787 : };
6788 :
6789 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
6790 : // locations, this has a good chance of timing out.
6791 0 : if let Some(waiter) = waiter {
6792 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6793 : } else {
6794 0 : tracing::info!("Migration is a no-op");
6795 : }
6796 :
6797 0 : Ok(TenantShardMigrateResponse {})
6798 0 : }
6799 :
6800 0 : pub(crate) async fn tenant_shard_migrate_secondary(
6801 0 : &self,
6802 0 : tenant_shard_id: TenantShardId,
6803 0 : migrate_req: TenantShardMigrateRequest,
6804 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6805 0 : let waiter = {
6806 0 : let mut locked = self.inner.write().unwrap();
6807 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6808 :
6809 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6810 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6811 0 : "Node {} not found",
6812 0 : migrate_req.node_id
6813 0 : )));
6814 : };
6815 :
6816 0 : if !node.is_available() {
6817 : // Warn but proceed: the caller may intend to manually adjust the placement of
6818 : // a shard even if the node is down, e.g. if intervening during an incident.
6819 0 : tracing::warn!("Migrating to unavailable node {node}");
6820 0 : }
6821 :
6822 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6823 0 : return Err(ApiError::NotFound(
6824 0 : anyhow::anyhow!("Tenant shard not found").into(),
6825 0 : ));
6826 : };
6827 :
6828 0 : if shard.intent.get_secondary().len() == 1
6829 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
6830 : {
6831 0 : tracing::info!(
6832 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
6833 : shard.intent
6834 : );
6835 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6836 0 : tracing::info!(
6837 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
6838 : );
6839 : } else {
6840 0 : let old_secondaries = shard.intent.get_secondary().clone();
6841 0 : for secondary in old_secondaries {
6842 0 : shard.intent.remove_secondary(scheduler, secondary);
6843 0 : }
6844 :
6845 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6846 0 : shard.sequence = shard.sequence.next();
6847 0 : tracing::info!(
6848 0 : "Migrating secondary to {node}: new intent {:?}",
6849 : shard.intent
6850 : );
6851 : }
6852 :
6853 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6854 : };
6855 :
6856 0 : if let Some(waiter) = waiter {
6857 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6858 : } else {
6859 0 : tracing::info!("Migration is a no-op");
6860 : }
6861 :
6862 0 : Ok(TenantShardMigrateResponse {})
6863 0 : }
6864 :
6865 : /// 'cancel' in this context means cancel any ongoing reconcile
6866 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
6867 0 : &self,
6868 0 : tenant_shard_id: TenantShardId,
6869 0 : ) -> Result<(), ApiError> {
6870 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
6871 0 : let waiter = {
6872 0 : let locked = self.inner.write().unwrap();
6873 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6874 0 : return Err(ApiError::NotFound(
6875 0 : anyhow::anyhow!("Tenant shard not found").into(),
6876 0 : ));
6877 : };
6878 :
6879 0 : let waiter = shard.get_waiter();
6880 0 : match waiter {
6881 : None => {
6882 0 : tracing::info!("Shard does not have an ongoing Reconciler");
6883 0 : return Ok(());
6884 : }
6885 0 : Some(waiter) => {
6886 0 : tracing::info!("Cancelling Reconciler");
6887 0 : shard.cancel_reconciler();
6888 0 : waiter
6889 : }
6890 : }
6891 : };
6892 :
6893 : // Cancellation should be prompt. If this fails we have still done our job of firing the
6894 : // cancellation token, but by returning an ApiError we will indicate to the caller that
6895 : // the Reconciler is misbehaving and not respecting the cancellation token
6896 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
6897 0 : .await?;
6898 :
6899 0 : Ok(())
6900 0 : }
6901 :
6902 : /// This is for debug/support only: we simply drop all state for a tenant, without
6903 : /// detaching or deleting it on pageservers.
6904 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
6905 0 : self.persistence.delete_tenant(tenant_id).await?;
6906 :
6907 0 : let mut locked = self.inner.write().unwrap();
6908 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6909 0 : let mut shards = Vec::new();
6910 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
6911 0 : shards.push(*tenant_shard_id);
6912 0 : }
6913 :
6914 0 : for shard_id in shards {
6915 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
6916 0 : shard.intent.clear(scheduler);
6917 0 : }
6918 : }
6919 :
6920 0 : Ok(())
6921 0 : }
6922 :
6923 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
6924 : /// tenant with a very high generation number so that it will see the existing data.
6925 : /// It does not create timelines on safekeepers, because they might already exist on some
6926 : /// safekeeper set. So, the timelines are not storcon-managed after the import.
6927 0 : pub(crate) async fn tenant_import(
6928 0 : &self,
6929 0 : tenant_id: TenantId,
6930 0 : ) -> Result<TenantCreateResponse, ApiError> {
6931 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
6932 0 : let maybe_node = {
6933 0 : self.inner
6934 0 : .read()
6935 0 : .unwrap()
6936 0 : .nodes
6937 0 : .values()
6938 0 : .find(|n| n.is_available())
6939 0 : .cloned()
6940 : };
6941 0 : let Some(node) = maybe_node else {
6942 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
6943 : };
6944 :
6945 0 : let client = PageserverClient::new(
6946 0 : node.get_id(),
6947 0 : self.http_client.clone(),
6948 0 : node.base_url(),
6949 0 : self.config.pageserver_jwt_token.as_deref(),
6950 : );
6951 :
6952 0 : let scan_result = client
6953 0 : .tenant_scan_remote_storage(tenant_id)
6954 0 : .await
6955 0 : .map_err(|e| passthrough_api_error(&node, e))?;
6956 :
6957 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
6958 0 : let Some(shard_count) = scan_result
6959 0 : .shards
6960 0 : .iter()
6961 0 : .map(|s| s.tenant_shard_id.shard_count)
6962 0 : .max()
6963 : else {
6964 0 : return Err(ApiError::NotFound(
6965 0 : anyhow::anyhow!("No shards found").into(),
6966 0 : ));
6967 : };
6968 :
6969 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
6970 : // to
6971 0 : let generation = scan_result
6972 0 : .shards
6973 0 : .iter()
6974 0 : .map(|s| s.generation)
6975 0 : .max()
6976 0 : .expect("We already validated >0 shards");
6977 :
6978 : // Find the tenant's stripe size. This wasn't always persisted in the tenant manifest, so
6979 : // fall back to the original default stripe size of 32768 (256 MB) if it's not specified.
6980 : const ORIGINAL_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(32768);
6981 0 : let stripe_size = scan_result
6982 0 : .shards
6983 0 : .iter()
6984 0 : .find(|s| s.tenant_shard_id.shard_count == shard_count && s.generation == generation)
6985 0 : .expect("we validated >0 shards above")
6986 : .stripe_size
6987 0 : .unwrap_or_else(|| {
6988 0 : if shard_count.count() > 1 {
6989 0 : warn!("unknown stripe size, assuming {ORIGINAL_STRIPE_SIZE}");
6990 0 : }
6991 0 : ORIGINAL_STRIPE_SIZE
6992 0 : });
6993 :
6994 0 : let (response, waiters) = self
6995 0 : .do_tenant_create(TenantCreateRequest {
6996 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
6997 0 : generation,
6998 0 :
6999 0 : shard_parameters: ShardParameters {
7000 0 : count: shard_count,
7001 0 : stripe_size,
7002 0 : },
7003 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
7004 0 : config: TenantConfig::default(),
7005 0 : })
7006 0 : .await?;
7007 :
7008 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
7009 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
7010 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
7011 : // reconcile, as reconciliation includes notifying compute.
7012 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
7013 0 : }
7014 :
7015 0 : Ok(response)
7016 0 : }
7017 :
7018 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
7019 : /// we don't have to make TenantShard clonable in the return path.
7020 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7021 0 : let serialized = {
7022 0 : let locked = self.inner.read().unwrap();
7023 0 : let result = locked.tenants.values().collect::<Vec<_>>();
7024 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
7025 : };
7026 :
7027 0 : hyper::Response::builder()
7028 0 : .status(hyper::StatusCode::OK)
7029 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7030 0 : .body(hyper::Body::from(serialized))
7031 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7032 0 : }
7033 :
7034 : /// Check the consistency of in-memory state vs. persistent state, and check that the
7035 : /// scheduler's statistics are up to date.
7036 : ///
7037 : /// These consistency checks expect an **idle** system. If changes are going on while
7038 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
7039 : /// checks, but not suitable for running continuously in the background in the field.
7040 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
7041 0 : let (mut expect_nodes, mut expect_shards) = {
7042 0 : let locked = self.inner.read().unwrap();
7043 :
7044 0 : locked
7045 0 : .scheduler
7046 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
7047 0 : .context("Scheduler checks")
7048 0 : .map_err(ApiError::InternalServerError)?;
7049 :
7050 0 : let expect_nodes = locked
7051 0 : .nodes
7052 0 : .values()
7053 0 : .map(|n| n.to_persistent())
7054 0 : .collect::<Vec<_>>();
7055 :
7056 0 : let expect_shards = locked
7057 0 : .tenants
7058 0 : .values()
7059 0 : .map(|t| t.to_persistent())
7060 0 : .collect::<Vec<_>>();
7061 :
7062 : // This method can only validate the state of an idle system: if a reconcile is in
7063 : // progress, fail out early to avoid giving false errors on state that won't match
7064 : // between database and memory under a ReconcileResult is processed.
7065 0 : for t in locked.tenants.values() {
7066 0 : if t.reconciler.is_some() {
7067 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7068 0 : "Shard {} reconciliation in progress",
7069 0 : t.tenant_shard_id
7070 0 : )));
7071 0 : }
7072 : }
7073 :
7074 0 : (expect_nodes, expect_shards)
7075 : };
7076 :
7077 0 : let mut nodes = self.persistence.list_nodes().await?;
7078 0 : expect_nodes.sort_by_key(|n| n.node_id);
7079 0 : nodes.sort_by_key(|n| n.node_id);
7080 :
7081 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
7082 0 : let node_result = if nodes != expect_nodes {
7083 0 : tracing::error!("Consistency check failed on nodes.");
7084 0 : tracing::error!(
7085 0 : "Nodes in memory: {}",
7086 0 : serde_json::to_string(&expect_nodes)
7087 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7088 : );
7089 0 : tracing::error!(
7090 0 : "Nodes in database: {}",
7091 0 : serde_json::to_string(&nodes)
7092 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7093 : );
7094 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
7095 0 : "Node consistency failure"
7096 0 : )))
7097 : } else {
7098 0 : Ok(())
7099 : };
7100 :
7101 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
7102 0 : persistent_shards
7103 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7104 :
7105 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
7106 :
7107 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
7108 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
7109 : // fields are added, before doing a comparison.
7110 0 : for tsp in &mut persistent_shards {
7111 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
7112 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
7113 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
7114 : }
7115 :
7116 0 : if persistent_shards != expect_shards {
7117 0 : tracing::error!("Consistency check failed on shards.");
7118 :
7119 0 : tracing::error!(
7120 0 : "Shards in memory: {}",
7121 0 : serde_json::to_string(&expect_shards)
7122 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7123 : );
7124 0 : tracing::error!(
7125 0 : "Shards in database: {}",
7126 0 : serde_json::to_string(&persistent_shards)
7127 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7128 : );
7129 :
7130 : // The total dump log lines above are useful in testing but in the field grafana will
7131 : // usually just drop them because they're so large. So we also do some explicit logging
7132 : // of just the diffs.
7133 0 : let persistent_shards = persistent_shards
7134 0 : .into_iter()
7135 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7136 0 : .collect::<HashMap<_, _>>();
7137 0 : let expect_shards = expect_shards
7138 0 : .into_iter()
7139 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
7140 0 : .collect::<HashMap<_, _>>();
7141 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
7142 0 : match expect_shards.get(tenant_shard_id) {
7143 : None => {
7144 0 : tracing::error!(
7145 0 : "Shard {} found in database but not in memory",
7146 : tenant_shard_id
7147 : );
7148 : }
7149 0 : Some(expect_tsp) => {
7150 0 : if expect_tsp != persistent_tsp {
7151 0 : tracing::error!(
7152 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
7153 : tenant_shard_id,
7154 0 : serde_json::to_string(expect_tsp).unwrap(),
7155 0 : serde_json::to_string(&persistent_tsp).unwrap()
7156 : );
7157 0 : }
7158 : }
7159 : }
7160 : }
7161 :
7162 : // Having already logged any differences, log any shards that simply aren't present in the database
7163 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
7164 0 : if !persistent_shards.contains_key(tenant_shard_id) {
7165 0 : tracing::error!(
7166 0 : "Shard {} found in memory but not in database: {}",
7167 : tenant_shard_id,
7168 0 : serde_json::to_string(memory_tsp)
7169 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7170 : );
7171 0 : }
7172 : }
7173 :
7174 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7175 0 : "Shard consistency failure"
7176 0 : )));
7177 0 : }
7178 :
7179 0 : node_result
7180 0 : }
7181 :
7182 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
7183 : /// we don't have to make TenantShard clonable in the return path.
7184 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7185 0 : let serialized = {
7186 0 : let locked = self.inner.read().unwrap();
7187 0 : serde_json::to_string(&locked.scheduler)
7188 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7189 : };
7190 :
7191 0 : hyper::Response::builder()
7192 0 : .status(hyper::StatusCode::OK)
7193 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7194 0 : .body(hyper::Body::from(serialized))
7195 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7196 0 : }
7197 :
7198 : /// This is for debug/support only: we simply drop all state for a tenant, without
7199 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
7200 : /// tenants that were on this node.
7201 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
7202 0 : self.persistence.set_tombstone(node_id).await?;
7203 :
7204 0 : let mut locked = self.inner.write().unwrap();
7205 :
7206 0 : for shard in locked.tenants.values_mut() {
7207 0 : shard.deref_node(node_id);
7208 0 : shard.observed.locations.remove(&node_id);
7209 0 : }
7210 :
7211 0 : let mut nodes = (*locked.nodes).clone();
7212 0 : nodes.remove(&node_id);
7213 0 : locked.nodes = Arc::new(nodes);
7214 0 : metrics::METRICS_REGISTRY
7215 0 : .metrics_group
7216 0 : .storage_controller_pageserver_nodes
7217 0 : .set(locked.nodes.len() as i64);
7218 0 : metrics::METRICS_REGISTRY
7219 0 : .metrics_group
7220 0 : .storage_controller_https_pageserver_nodes
7221 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7222 :
7223 0 : locked.scheduler.node_remove(node_id);
7224 :
7225 0 : Ok(())
7226 0 : }
7227 :
7228 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
7229 : /// that we don't leave any bad state behind in the storage controller, but unclean
7230 : /// in the sense that we are not carefully draining the node.
7231 0 : pub(crate) async fn node_delete_old(&self, node_id: NodeId) -> Result<(), ApiError> {
7232 0 : let _node_lock =
7233 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
7234 :
7235 : // 1. Atomically update in-memory state:
7236 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
7237 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
7238 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
7239 : // re-insert references to this node into the ObservedState of shards
7240 : // - drop the node from the scheduler
7241 : {
7242 0 : let mut locked = self.inner.write().unwrap();
7243 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7244 :
7245 : {
7246 0 : let mut nodes_mut = (*nodes).deref().clone();
7247 0 : match nodes_mut.get_mut(&node_id) {
7248 0 : Some(node) => {
7249 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
7250 0 : // if we crash it would not be desirable to leave the node paused after a restart.
7251 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
7252 0 : }
7253 : None => {
7254 0 : tracing::info!(
7255 0 : "Node not found: presuming this is a retry and returning success"
7256 : );
7257 0 : return Ok(());
7258 : }
7259 : }
7260 :
7261 0 : *nodes = Arc::new(nodes_mut);
7262 : }
7263 :
7264 0 : for (_tenant_id, mut schedule_context, shards) in
7265 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7266 : {
7267 0 : for shard in shards {
7268 0 : if shard.deref_node(node_id) {
7269 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
7270 : // TODO: implement force flag to remove a node even if we can't reschedule
7271 : // a tenant
7272 0 : tracing::error!(
7273 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7274 : shard.tenant_shard_id
7275 : );
7276 0 : return Err(e.into());
7277 : } else {
7278 0 : tracing::info!(
7279 0 : "Rescheduled shard {} away from node during deletion",
7280 : shard.tenant_shard_id
7281 : )
7282 : }
7283 :
7284 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
7285 0 : }
7286 :
7287 : // Here we remove an existing observed location for the node we're removing, and it will
7288 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7289 : // process_result.
7290 : //
7291 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
7292 : // means any reconciles we spawned will know about the node we're deleting, enabling them
7293 : // to do live migrations if it's still online.
7294 0 : shard.observed.locations.remove(&node_id);
7295 : }
7296 : }
7297 :
7298 0 : scheduler.node_remove(node_id);
7299 :
7300 : {
7301 0 : let mut nodes_mut = (**nodes).clone();
7302 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7303 0 : // Ensure that any reconciler holding an Arc<> to this node will
7304 0 : // drop out when trying to RPC to it (setting Offline state sets the
7305 0 : // cancellation token on the Node object).
7306 0 : removed_node.set_availability(NodeAvailability::Offline);
7307 0 : }
7308 0 : *nodes = Arc::new(nodes_mut);
7309 0 : metrics::METRICS_REGISTRY
7310 0 : .metrics_group
7311 0 : .storage_controller_pageserver_nodes
7312 0 : .set(nodes.len() as i64);
7313 0 : metrics::METRICS_REGISTRY
7314 0 : .metrics_group
7315 0 : .storage_controller_https_pageserver_nodes
7316 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7317 : }
7318 : }
7319 :
7320 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
7321 : // the removed node, as this column means "The pageserver to which this generation was issued", and
7322 : // their generations won't get updated until the reconcilers moving them away from this node complete.
7323 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
7324 : // that exists.
7325 :
7326 : // 2. Actually delete the node from in-memory state and set tombstone to the database
7327 : // for preventing the node to register again.
7328 0 : tracing::info!("Deleting node from database");
7329 0 : self.persistence.set_tombstone(node_id).await?;
7330 :
7331 0 : Ok(())
7332 0 : }
7333 :
7334 0 : pub(crate) async fn delete_node(
7335 0 : self: &Arc<Self>,
7336 0 : node_id: NodeId,
7337 0 : policy_on_start: NodeSchedulingPolicy,
7338 0 : cancel: CancellationToken,
7339 0 : ) -> Result<(), OperationError> {
7340 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal).build();
7341 :
7342 0 : let mut waiters: Vec<ReconcilerWaiter> = Vec::new();
7343 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
7344 :
7345 0 : while !tid_iter.finished() {
7346 0 : if cancel.is_cancelled() {
7347 0 : match self
7348 0 : .node_configure(node_id, None, Some(policy_on_start))
7349 0 : .await
7350 : {
7351 0 : Ok(()) => return Err(OperationError::Cancelled),
7352 0 : Err(err) => {
7353 0 : return Err(OperationError::FinalizeError(
7354 0 : format!(
7355 0 : "Failed to finalise delete cancel of {} by setting scheduling policy to {}: {}",
7356 0 : node_id, String::from(policy_on_start), err
7357 0 : )
7358 0 : .into(),
7359 0 : ));
7360 : }
7361 : }
7362 0 : }
7363 :
7364 0 : operation_utils::validate_node_state(
7365 0 : &node_id,
7366 0 : self.inner.read().unwrap().nodes.clone(),
7367 0 : NodeSchedulingPolicy::Deleting,
7368 0 : )?;
7369 :
7370 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7371 0 : let tid = match tid_iter.next() {
7372 0 : Some(tid) => tid,
7373 : None => {
7374 0 : break;
7375 : }
7376 : };
7377 :
7378 0 : let mut locked = self.inner.write().unwrap();
7379 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7380 :
7381 : // Calculate a schedule context here to avoid borrow checker issues.
7382 0 : let mut schedule_context = ScheduleContext::default();
7383 0 : for (_, shard) in tenants.range(TenantShardId::tenant_range(tid.tenant_id)) {
7384 0 : schedule_context.avoid(&shard.intent.all_pageservers());
7385 0 : }
7386 :
7387 0 : let tenant_shard = match tenants.get_mut(&tid) {
7388 0 : Some(tenant_shard) => tenant_shard,
7389 : None => {
7390 : // Tenant shard was deleted by another operation. Skip it.
7391 0 : continue;
7392 : }
7393 : };
7394 :
7395 0 : match tenant_shard.get_scheduling_policy() {
7396 0 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {
7397 0 : // A migration during delete is classed as 'essential' because it is required to
7398 0 : // uphold our availability goals for the tenant: this shard is elegible for migration.
7399 0 : }
7400 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
7401 : // If we have been asked to avoid rescheduling this shard, then do not migrate it during a deletion
7402 0 : tracing::warn!(
7403 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7404 0 : "Skip migration during deletion because shard scheduling policy {:?} disallows it",
7405 0 : tenant_shard.get_scheduling_policy(),
7406 : );
7407 0 : continue;
7408 : }
7409 : }
7410 :
7411 0 : if tenant_shard.deref_node(node_id) {
7412 0 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
7413 0 : tracing::error!(
7414 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7415 : tenant_shard.tenant_shard_id
7416 : );
7417 0 : return Err(OperationError::ImpossibleConstraint(e.to_string().into()));
7418 : } else {
7419 0 : tracing::info!(
7420 0 : "Rescheduled shard {} away from node during deletion",
7421 : tenant_shard.tenant_shard_id
7422 : )
7423 : }
7424 :
7425 0 : let waiter = self.maybe_configured_reconcile_shard(
7426 0 : tenant_shard,
7427 0 : nodes,
7428 0 : reconciler_config,
7429 0 : );
7430 0 : if let Some(some) = waiter {
7431 0 : waiters.push(some);
7432 0 : }
7433 0 : }
7434 : }
7435 :
7436 0 : waiters = self
7437 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
7438 0 : .await;
7439 :
7440 0 : failpoint_support::sleep_millis_async!("sleepy-delete-loop", &cancel);
7441 : }
7442 :
7443 0 : while !waiters.is_empty() {
7444 0 : if cancel.is_cancelled() {
7445 0 : match self
7446 0 : .node_configure(node_id, None, Some(policy_on_start))
7447 0 : .await
7448 : {
7449 0 : Ok(()) => return Err(OperationError::Cancelled),
7450 0 : Err(err) => {
7451 0 : return Err(OperationError::FinalizeError(
7452 0 : format!(
7453 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to {}: {}",
7454 0 : node_id, String::from(policy_on_start), err
7455 0 : )
7456 0 : .into(),
7457 0 : ));
7458 : }
7459 : }
7460 0 : }
7461 :
7462 0 : tracing::info!("Awaiting {} pending delete reconciliations", waiters.len());
7463 :
7464 0 : waiters = self
7465 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7466 0 : .await;
7467 : }
7468 :
7469 0 : self.persistence
7470 0 : .set_tombstone(node_id)
7471 0 : .await
7472 0 : .map_err(|e| OperationError::FinalizeError(e.to_string().into()))?;
7473 :
7474 : {
7475 0 : let mut locked = self.inner.write().unwrap();
7476 0 : let (nodes, _, scheduler) = locked.parts_mut();
7477 :
7478 0 : scheduler.node_remove(node_id);
7479 :
7480 0 : let mut nodes_mut = (**nodes).clone();
7481 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7482 0 : // Ensure that any reconciler holding an Arc<> to this node will
7483 0 : // drop out when trying to RPC to it (setting Offline state sets the
7484 0 : // cancellation token on the Node object).
7485 0 : removed_node.set_availability(NodeAvailability::Offline);
7486 0 : }
7487 0 : *nodes = Arc::new(nodes_mut);
7488 :
7489 0 : metrics::METRICS_REGISTRY
7490 0 : .metrics_group
7491 0 : .storage_controller_pageserver_nodes
7492 0 : .set(nodes.len() as i64);
7493 0 : metrics::METRICS_REGISTRY
7494 0 : .metrics_group
7495 0 : .storage_controller_https_pageserver_nodes
7496 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7497 : }
7498 :
7499 0 : Ok(())
7500 0 : }
7501 :
7502 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
7503 0 : let nodes = {
7504 0 : self.inner
7505 0 : .read()
7506 0 : .unwrap()
7507 0 : .nodes
7508 0 : .values()
7509 0 : .cloned()
7510 0 : .collect::<Vec<_>>()
7511 : };
7512 :
7513 0 : Ok(nodes)
7514 0 : }
7515 :
7516 0 : pub(crate) async fn tombstone_list(&self) -> Result<Vec<Node>, ApiError> {
7517 0 : self.persistence
7518 0 : .list_tombstones()
7519 0 : .await?
7520 0 : .into_iter()
7521 0 : .map(|np| Node::from_persistent(np, false))
7522 0 : .collect::<Result<Vec<_>, _>>()
7523 0 : .map_err(ApiError::InternalServerError)
7524 0 : }
7525 :
7526 0 : pub(crate) async fn tombstone_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
7527 0 : let _node_lock = trace_exclusive_lock(
7528 0 : &self.node_op_locks,
7529 0 : node_id,
7530 0 : NodeOperations::DeleteTombstone,
7531 0 : )
7532 0 : .await;
7533 :
7534 0 : if matches!(self.get_node(node_id).await, Err(ApiError::NotFound(_))) {
7535 0 : self.persistence.delete_node(node_id).await?;
7536 0 : Ok(())
7537 : } else {
7538 0 : Err(ApiError::Conflict(format!(
7539 0 : "Node {node_id} is in use, consider using tombstone API first"
7540 0 : )))
7541 : }
7542 0 : }
7543 :
7544 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
7545 0 : self.inner
7546 0 : .read()
7547 0 : .unwrap()
7548 0 : .nodes
7549 0 : .get(&node_id)
7550 0 : .cloned()
7551 0 : .ok_or(ApiError::NotFound(
7552 0 : format!("Node {node_id} not registered").into(),
7553 0 : ))
7554 0 : }
7555 :
7556 0 : pub(crate) async fn get_node_shards(
7557 0 : &self,
7558 0 : node_id: NodeId,
7559 0 : ) -> Result<NodeShardResponse, ApiError> {
7560 0 : let locked = self.inner.read().unwrap();
7561 0 : let mut shards = Vec::new();
7562 0 : for (tid, tenant) in locked.tenants.iter() {
7563 0 : let is_intended_secondary = match (
7564 0 : tenant.intent.get_attached() == &Some(node_id),
7565 0 : tenant.intent.get_secondary().contains(&node_id),
7566 0 : ) {
7567 : (true, true) => {
7568 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7569 0 : "{} attached as primary+secondary on the same node",
7570 0 : tid
7571 0 : )));
7572 : }
7573 0 : (true, false) => Some(false),
7574 0 : (false, true) => Some(true),
7575 0 : (false, false) => None,
7576 : };
7577 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
7578 0 : tenant.observed.locations.get(&node_id)
7579 : {
7580 0 : Some(conf.secondary_conf.is_some())
7581 : } else {
7582 0 : None
7583 : };
7584 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
7585 0 : shards.push(NodeShard {
7586 0 : tenant_shard_id: *tid,
7587 0 : is_intended_secondary,
7588 0 : is_observed_secondary,
7589 0 : });
7590 0 : }
7591 : }
7592 0 : Ok(NodeShardResponse { node_id, shards })
7593 0 : }
7594 :
7595 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
7596 0 : self.persistence.get_leader().await
7597 0 : }
7598 :
7599 0 : pub(crate) async fn node_register(
7600 0 : &self,
7601 0 : register_req: NodeRegisterRequest,
7602 0 : ) -> Result<(), ApiError> {
7603 0 : let _node_lock = trace_exclusive_lock(
7604 0 : &self.node_op_locks,
7605 0 : register_req.node_id,
7606 0 : NodeOperations::Register,
7607 0 : )
7608 0 : .await;
7609 :
7610 : #[derive(PartialEq)]
7611 : enum RegistrationStatus {
7612 : UpToDate,
7613 : NeedUpdate,
7614 : Mismatched,
7615 : New,
7616 : }
7617 :
7618 0 : let registration_status = {
7619 0 : let locked = self.inner.read().unwrap();
7620 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
7621 0 : if node.registration_match(®ister_req) {
7622 0 : if node.need_update(®ister_req) {
7623 0 : RegistrationStatus::NeedUpdate
7624 : } else {
7625 0 : RegistrationStatus::UpToDate
7626 : }
7627 : } else {
7628 0 : RegistrationStatus::Mismatched
7629 : }
7630 : } else {
7631 0 : RegistrationStatus::New
7632 : }
7633 : };
7634 :
7635 0 : match registration_status {
7636 : RegistrationStatus::UpToDate => {
7637 0 : tracing::info!(
7638 0 : "Node {} re-registered with matching address and is up to date",
7639 : register_req.node_id
7640 : );
7641 :
7642 0 : return Ok(());
7643 : }
7644 : RegistrationStatus::Mismatched => {
7645 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
7646 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
7647 : // a fixed address through the lifetime of a node.
7648 0 : tracing::warn!(
7649 0 : "Node {} tried to register with different address",
7650 : register_req.node_id
7651 : );
7652 0 : return Err(ApiError::Conflict(
7653 0 : "Node is already registered with different address".to_string(),
7654 0 : ));
7655 : }
7656 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
7657 0 : // fallthrough
7658 0 : }
7659 : }
7660 :
7661 : // We do not require that a node is actually online when registered (it will start life
7662 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
7663 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
7664 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
7665 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
7666 0 : if tokio::net::lookup_host(format!(
7667 0 : "{}:{}",
7668 : register_req.listen_http_addr, register_req.listen_http_port
7669 : ))
7670 0 : .await
7671 0 : .is_err()
7672 : {
7673 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
7674 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
7675 : // we return a soft 503 error, to encourage callers to retry past transient issues.
7676 0 : return Err(ApiError::ResourceUnavailable(
7677 0 : format!(
7678 0 : "Node {} tried to register with unknown DNS name '{}'",
7679 0 : register_req.node_id, register_req.listen_http_addr
7680 0 : )
7681 0 : .into(),
7682 0 : ));
7683 0 : }
7684 :
7685 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
7686 0 : return Err(ApiError::PreconditionFailed(
7687 0 : format!(
7688 0 : "Node {} has no https port, but use_https is enabled",
7689 0 : register_req.node_id
7690 0 : )
7691 0 : .into(),
7692 0 : ));
7693 0 : }
7694 :
7695 0 : if register_req.listen_grpc_addr.is_some() != register_req.listen_grpc_port.is_some() {
7696 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
7697 0 : "must specify both gRPC address and port"
7698 0 : )));
7699 0 : }
7700 :
7701 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
7702 : // This ensures that before we use it for anything or expose it via any external
7703 : // API, it is guaranteed to be available after a restart.
7704 0 : let new_node = Node::new(
7705 0 : register_req.node_id,
7706 0 : register_req.listen_http_addr,
7707 0 : register_req.listen_http_port,
7708 0 : register_req.listen_https_port,
7709 0 : register_req.listen_pg_addr,
7710 0 : register_req.listen_pg_port,
7711 0 : register_req.listen_grpc_addr,
7712 0 : register_req.listen_grpc_port,
7713 0 : register_req.availability_zone_id.clone(),
7714 0 : self.config.use_https_pageserver_api,
7715 : );
7716 0 : let new_node = match new_node {
7717 0 : Ok(new_node) => new_node,
7718 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
7719 : };
7720 :
7721 0 : match registration_status {
7722 : RegistrationStatus::New => {
7723 0 : self.persistence.insert_node(&new_node).await.map_err(|e| {
7724 0 : if matches!(
7725 0 : e,
7726 : crate::persistence::DatabaseError::Query(
7727 : diesel::result::Error::DatabaseError(
7728 : diesel::result::DatabaseErrorKind::UniqueViolation,
7729 : _,
7730 : )
7731 : )
7732 : ) {
7733 : // The node can be deleted by tombstone API, and not show up in the list of nodes.
7734 : // If you see this error, check tombstones first.
7735 0 : ApiError::Conflict(format!("Node {} is already exists", new_node.get_id()))
7736 : } else {
7737 0 : ApiError::from(e)
7738 : }
7739 0 : })?;
7740 : }
7741 : RegistrationStatus::NeedUpdate => {
7742 0 : self.persistence
7743 0 : .update_node_on_registration(
7744 0 : register_req.node_id,
7745 0 : register_req.listen_https_port,
7746 0 : )
7747 0 : .await?
7748 : }
7749 0 : _ => unreachable!("Other statuses have been processed earlier"),
7750 : }
7751 :
7752 0 : let mut locked = self.inner.write().unwrap();
7753 0 : let mut new_nodes = (*locked.nodes).clone();
7754 :
7755 0 : locked.scheduler.node_upsert(&new_node);
7756 0 : new_nodes.insert(register_req.node_id, new_node);
7757 :
7758 0 : locked.nodes = Arc::new(new_nodes);
7759 :
7760 0 : metrics::METRICS_REGISTRY
7761 0 : .metrics_group
7762 0 : .storage_controller_pageserver_nodes
7763 0 : .set(locked.nodes.len() as i64);
7764 0 : metrics::METRICS_REGISTRY
7765 0 : .metrics_group
7766 0 : .storage_controller_https_pageserver_nodes
7767 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7768 :
7769 0 : match registration_status {
7770 : RegistrationStatus::New => {
7771 0 : tracing::info!(
7772 0 : "Registered pageserver {} ({}), now have {} pageservers",
7773 : register_req.node_id,
7774 : register_req.availability_zone_id,
7775 0 : locked.nodes.len()
7776 : );
7777 : }
7778 : RegistrationStatus::NeedUpdate => {
7779 0 : tracing::info!(
7780 0 : "Re-registered and updated node {} ({})",
7781 : register_req.node_id,
7782 : register_req.availability_zone_id,
7783 : );
7784 : }
7785 0 : _ => unreachable!("Other statuses have been processed earlier"),
7786 : }
7787 0 : Ok(())
7788 0 : }
7789 :
7790 : /// Configure in-memory and persistent state of a node as requested
7791 : ///
7792 : /// Note that this function does not trigger any immediate side effects in response
7793 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
7794 0 : async fn node_state_configure(
7795 0 : &self,
7796 0 : node_id: NodeId,
7797 0 : availability: Option<NodeAvailability>,
7798 0 : scheduling: Option<NodeSchedulingPolicy>,
7799 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
7800 0 : ) -> Result<AvailabilityTransition, ApiError> {
7801 0 : if let Some(scheduling) = scheduling {
7802 : // Scheduling is a persistent part of Node: we must write updates to the database before
7803 : // applying them in memory
7804 0 : self.persistence
7805 0 : .update_node_scheduling_policy(node_id, scheduling)
7806 0 : .await?;
7807 0 : }
7808 :
7809 : // If we're activating a node, then before setting it active we must reconcile any shard locations
7810 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
7811 : // by calling [`Self::node_activate_reconcile`]
7812 : //
7813 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
7814 : // nothing else can mutate its availability while we run.
7815 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
7816 0 : let (activate_node, availability_transition) = {
7817 0 : let locked = self.inner.read().unwrap();
7818 0 : let Some(node) = locked.nodes.get(&node_id) else {
7819 0 : return Err(ApiError::NotFound(
7820 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7821 0 : ));
7822 : };
7823 :
7824 0 : (
7825 0 : node.clone(),
7826 0 : node.get_availability_transition(input_availability),
7827 0 : )
7828 : };
7829 :
7830 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
7831 0 : self.node_activate_reconcile(activate_node, node_lock)
7832 0 : .await?;
7833 0 : }
7834 0 : availability_transition
7835 : } else {
7836 0 : AvailabilityTransition::Unchanged
7837 : };
7838 :
7839 : // Apply changes from the request to our in-memory state for the Node
7840 0 : let mut locked = self.inner.write().unwrap();
7841 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
7842 :
7843 0 : let mut new_nodes = (**nodes).clone();
7844 :
7845 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
7846 0 : return Err(ApiError::NotFound(
7847 0 : anyhow::anyhow!("Node not registered").into(),
7848 0 : ));
7849 : };
7850 :
7851 0 : if let Some(availability) = availability {
7852 0 : node.set_availability(availability);
7853 0 : }
7854 :
7855 0 : if let Some(scheduling) = scheduling {
7856 0 : node.set_scheduling(scheduling);
7857 0 : }
7858 :
7859 : // Update the scheduler, in case the elegibility of the node for new shards has changed
7860 0 : scheduler.node_upsert(node);
7861 :
7862 0 : let new_nodes = Arc::new(new_nodes);
7863 0 : locked.nodes = new_nodes;
7864 :
7865 0 : Ok(availability_transition)
7866 0 : }
7867 :
7868 : /// Handle availability transition of one node
7869 : ///
7870 : /// Note that you should first call [`Self::node_state_configure`] to update
7871 : /// the in-memory state referencing that node. If you need to handle more than one transition
7872 : /// consider using [`Self::handle_node_availability_transitions`].
7873 0 : async fn handle_node_availability_transition(
7874 0 : &self,
7875 0 : node_id: NodeId,
7876 0 : transition: AvailabilityTransition,
7877 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
7878 0 : ) -> Result<(), ApiError> {
7879 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
7880 0 : match transition {
7881 : AvailabilityTransition::ToOffline => {
7882 0 : tracing::info!("Node {} transition to offline", node_id);
7883 :
7884 0 : let mut locked = self.inner.write().unwrap();
7885 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7886 :
7887 0 : let mut tenants_affected: usize = 0;
7888 :
7889 0 : for (_tenant_id, mut schedule_context, shards) in
7890 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7891 : {
7892 0 : for tenant_shard in shards {
7893 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
7894 0 : if let Some(observed_loc) =
7895 0 : tenant_shard.observed.locations.get_mut(&node_id)
7896 0 : {
7897 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
7898 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
7899 0 : observed_loc.conf = None;
7900 0 : }
7901 :
7902 0 : if nodes.len() == 1 {
7903 : // Special case for single-node cluster: there is no point trying to reschedule
7904 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
7905 : // failures to schedule them.
7906 0 : continue;
7907 0 : }
7908 :
7909 0 : if !nodes
7910 0 : .values()
7911 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
7912 : {
7913 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
7914 : // trying to reschedule since there's nowhere else to go. Without this
7915 : // branch we incorrectly detach tenants in response to node unavailability.
7916 0 : continue;
7917 0 : }
7918 :
7919 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
7920 0 : tenant_shard.sequence = tenant_shard.sequence.next();
7921 :
7922 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
7923 0 : Err(e) => {
7924 : // It is possible that some tenants will become unschedulable when too many pageservers
7925 : // go offline: in this case there isn't much we can do other than make the issue observable.
7926 : // TODO: give TenantShard a scheduling error attribute to be queried later.
7927 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
7928 : }
7929 : Ok(()) => {
7930 0 : if self
7931 0 : .maybe_reconcile_shard(
7932 0 : tenant_shard,
7933 0 : nodes,
7934 0 : ReconcilerPriority::Normal,
7935 0 : )
7936 0 : .is_some()
7937 0 : {
7938 0 : tenants_affected += 1;
7939 0 : };
7940 : }
7941 : }
7942 0 : }
7943 : }
7944 : }
7945 0 : tracing::info!(
7946 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
7947 : tenants_affected,
7948 : node_id
7949 : )
7950 : }
7951 : AvailabilityTransition::ToActive => {
7952 0 : tracing::info!("Node {} transition to active", node_id);
7953 :
7954 0 : let mut locked = self.inner.write().unwrap();
7955 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7956 :
7957 : // When a node comes back online, we must reconcile any tenant that has a None observed
7958 : // location on the node.
7959 0 : for tenant_shard in tenants.values_mut() {
7960 : // If a reconciliation is already in progress, rely on the previous scheduling
7961 : // decision and skip triggering a new reconciliation.
7962 0 : if tenant_shard.reconciler.is_some() {
7963 0 : continue;
7964 0 : }
7965 :
7966 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
7967 0 : if observed_loc.conf.is_none() {
7968 0 : self.maybe_reconcile_shard(
7969 0 : tenant_shard,
7970 0 : nodes,
7971 0 : ReconcilerPriority::Normal,
7972 0 : );
7973 0 : }
7974 0 : }
7975 : }
7976 :
7977 : // TODO: in the background, we should balance work back onto this pageserver
7978 : }
7979 : // No action required for the intermediate unavailable state.
7980 : // When we transition into active or offline from the unavailable state,
7981 : // the correct handling above will kick in.
7982 : AvailabilityTransition::ToWarmingUpFromActive => {
7983 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
7984 : }
7985 : AvailabilityTransition::ToWarmingUpFromOffline => {
7986 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
7987 : }
7988 : AvailabilityTransition::Unchanged => {
7989 0 : tracing::debug!("Node {} no availability change during config", node_id);
7990 : }
7991 : }
7992 :
7993 0 : Ok(())
7994 0 : }
7995 :
7996 : /// Handle availability transition for multiple nodes
7997 : ///
7998 : /// Note that you should first call [`Self::node_state_configure`] for
7999 : /// all nodes being handled here for the handling to use fresh in-memory state.
8000 0 : async fn handle_node_availability_transitions(
8001 0 : &self,
8002 0 : transitions: Vec<(
8003 0 : NodeId,
8004 0 : TracingExclusiveGuard<NodeOperations>,
8005 0 : AvailabilityTransition,
8006 0 : )>,
8007 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
8008 0 : let mut errors = Vec::default();
8009 0 : for (node_id, node_lock, transition) in transitions {
8010 0 : let res = self
8011 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
8012 0 : .await;
8013 0 : if let Err(err) = res {
8014 0 : errors.push((node_id, err));
8015 0 : }
8016 : }
8017 :
8018 0 : if errors.is_empty() {
8019 0 : Ok(())
8020 : } else {
8021 0 : Err(errors)
8022 : }
8023 0 : }
8024 :
8025 0 : pub(crate) async fn node_configure(
8026 0 : &self,
8027 0 : node_id: NodeId,
8028 0 : availability: Option<NodeAvailability>,
8029 0 : scheduling: Option<NodeSchedulingPolicy>,
8030 0 : ) -> Result<(), ApiError> {
8031 0 : let node_lock =
8032 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
8033 :
8034 0 : let transition = self
8035 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
8036 0 : .await?;
8037 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
8038 0 : .await
8039 0 : }
8040 :
8041 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
8042 : /// operation for HTTP api.
8043 0 : pub(crate) async fn external_node_configure(
8044 0 : &self,
8045 0 : node_id: NodeId,
8046 0 : availability: Option<NodeAvailability>,
8047 0 : scheduling: Option<NodeSchedulingPolicy>,
8048 0 : ) -> Result<(), ApiError> {
8049 : {
8050 0 : let locked = self.inner.read().unwrap();
8051 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
8052 0 : return Err(ApiError::PreconditionFailed(
8053 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
8054 0 : ));
8055 0 : }
8056 : }
8057 :
8058 0 : self.node_configure(node_id, availability, scheduling).await
8059 0 : }
8060 :
8061 0 : pub(crate) async fn start_node_delete(
8062 0 : self: &Arc<Self>,
8063 0 : node_id: NodeId,
8064 0 : ) -> Result<(), ApiError> {
8065 0 : let (ongoing_op, node_policy, schedulable_nodes_count) = {
8066 0 : let locked = self.inner.read().unwrap();
8067 0 : let nodes = &locked.nodes;
8068 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8069 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8070 0 : ))?;
8071 0 : let schedulable_nodes_count = nodes
8072 0 : .iter()
8073 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8074 0 : .count();
8075 :
8076 : (
8077 0 : locked
8078 0 : .ongoing_operation
8079 0 : .as_ref()
8080 0 : .map(|ongoing| ongoing.operation),
8081 0 : node.get_scheduling(),
8082 0 : schedulable_nodes_count,
8083 : )
8084 : };
8085 :
8086 0 : if let Some(ongoing) = ongoing_op {
8087 0 : return Err(ApiError::PreconditionFailed(
8088 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8089 0 : ));
8090 0 : }
8091 :
8092 0 : if schedulable_nodes_count == 0 {
8093 0 : return Err(ApiError::PreconditionFailed(
8094 0 : "No other schedulable nodes to move shards".into(),
8095 0 : ));
8096 0 : }
8097 :
8098 0 : match node_policy {
8099 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
8100 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Deleting))
8101 0 : .await?;
8102 :
8103 0 : let cancel = self.cancel.child_token();
8104 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8105 0 : let policy_on_start = node_policy;
8106 :
8107 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8108 0 : operation: Operation::Delete(Delete { node_id }),
8109 0 : cancel: cancel.clone(),
8110 0 : });
8111 :
8112 0 : let span = tracing::info_span!(parent: None, "delete_node", %node_id);
8113 :
8114 0 : tokio::task::spawn(
8115 : {
8116 0 : let service = self.clone();
8117 0 : let cancel = cancel.clone();
8118 0 : async move {
8119 0 : let _gate_guard = gate_guard;
8120 :
8121 0 : scopeguard::defer! {
8122 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8123 :
8124 : if let Some(Operation::Delete(removed_delete)) = prev.map(|h| h.operation) {
8125 : assert_eq!(removed_delete.node_id, node_id, "We always take the same operation");
8126 : } else {
8127 : panic!("We always remove the same operation")
8128 : }
8129 : }
8130 :
8131 0 : tracing::info!("Delete background operation starting");
8132 0 : let res = service
8133 0 : .delete_node(node_id, policy_on_start, cancel)
8134 0 : .await;
8135 0 : match res {
8136 : Ok(()) => {
8137 0 : tracing::info!(
8138 0 : "Delete background operation completed successfully"
8139 : );
8140 : }
8141 : Err(OperationError::Cancelled) => {
8142 0 : tracing::info!("Delete background operation was cancelled");
8143 : }
8144 0 : Err(err) => {
8145 0 : tracing::error!(
8146 0 : "Delete background operation encountered: {err}"
8147 : )
8148 : }
8149 : }
8150 0 : }
8151 : }
8152 0 : .instrument(span),
8153 : );
8154 : }
8155 : NodeSchedulingPolicy::Deleting => {
8156 0 : return Err(ApiError::Conflict(format!(
8157 0 : "Node {node_id} has delete in progress"
8158 0 : )));
8159 : }
8160 0 : policy => {
8161 0 : return Err(ApiError::PreconditionFailed(
8162 0 : format!("Node {node_id} cannot be deleted due to {policy:?} policy").into(),
8163 0 : ));
8164 : }
8165 : }
8166 :
8167 0 : Ok(())
8168 0 : }
8169 :
8170 0 : pub(crate) async fn cancel_node_delete(
8171 0 : self: &Arc<Self>,
8172 0 : node_id: NodeId,
8173 0 : ) -> Result<(), ApiError> {
8174 : {
8175 0 : let locked = self.inner.read().unwrap();
8176 0 : let nodes = &locked.nodes;
8177 0 : nodes.get(&node_id).ok_or(ApiError::NotFound(
8178 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8179 0 : ))?;
8180 : }
8181 :
8182 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8183 0 : if let Operation::Delete(delete) = op_handler.operation {
8184 0 : if delete.node_id == node_id {
8185 0 : tracing::info!("Cancelling background delete operation for node {node_id}");
8186 0 : op_handler.cancel.cancel();
8187 0 : return Ok(());
8188 0 : }
8189 0 : }
8190 0 : }
8191 :
8192 0 : Err(ApiError::PreconditionFailed(
8193 0 : format!("Node {node_id} has no delete in progress").into(),
8194 0 : ))
8195 0 : }
8196 :
8197 0 : pub(crate) async fn start_node_drain(
8198 0 : self: &Arc<Self>,
8199 0 : node_id: NodeId,
8200 0 : ) -> Result<(), ApiError> {
8201 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
8202 0 : let locked = self.inner.read().unwrap();
8203 0 : let nodes = &locked.nodes;
8204 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8205 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8206 0 : ))?;
8207 0 : let schedulable_nodes_count = nodes
8208 0 : .iter()
8209 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8210 0 : .count();
8211 :
8212 : (
8213 0 : locked
8214 0 : .ongoing_operation
8215 0 : .as_ref()
8216 0 : .map(|ongoing| ongoing.operation),
8217 0 : node.is_available(),
8218 0 : node.get_scheduling(),
8219 0 : schedulable_nodes_count,
8220 : )
8221 : };
8222 :
8223 0 : if let Some(ongoing) = ongoing_op {
8224 0 : return Err(ApiError::PreconditionFailed(
8225 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8226 0 : ));
8227 0 : }
8228 :
8229 0 : if !node_available {
8230 0 : return Err(ApiError::ResourceUnavailable(
8231 0 : format!("Node {node_id} is currently unavailable").into(),
8232 0 : ));
8233 0 : }
8234 :
8235 0 : if schedulable_nodes_count == 0 {
8236 0 : return Err(ApiError::PreconditionFailed(
8237 0 : "No other schedulable nodes to drain to".into(),
8238 0 : ));
8239 0 : }
8240 :
8241 0 : match node_policy {
8242 : NodeSchedulingPolicy::Active => {
8243 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
8244 0 : .await?;
8245 :
8246 0 : let cancel = self.cancel.child_token();
8247 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8248 :
8249 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8250 0 : operation: Operation::Drain(Drain { node_id }),
8251 0 : cancel: cancel.clone(),
8252 0 : });
8253 :
8254 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
8255 :
8256 0 : tokio::task::spawn({
8257 0 : let service = self.clone();
8258 0 : let cancel = cancel.clone();
8259 0 : async move {
8260 0 : let _gate_guard = gate_guard;
8261 :
8262 0 : scopeguard::defer! {
8263 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8264 :
8265 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
8266 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
8267 : } else {
8268 : panic!("We always remove the same operation")
8269 : }
8270 : }
8271 :
8272 0 : tracing::info!("Drain background operation starting");
8273 0 : let res = service.drain_node(node_id, cancel).await;
8274 0 : match res {
8275 : Ok(()) => {
8276 0 : tracing::info!("Drain background operation completed successfully");
8277 : }
8278 : Err(OperationError::Cancelled) => {
8279 0 : tracing::info!("Drain background operation was cancelled");
8280 : }
8281 0 : Err(err) => {
8282 0 : tracing::error!("Drain background operation encountered: {err}")
8283 : }
8284 : }
8285 0 : }
8286 0 : }.instrument(span));
8287 : }
8288 : NodeSchedulingPolicy::Draining => {
8289 0 : return Err(ApiError::Conflict(format!(
8290 0 : "Node {node_id} has drain in progress"
8291 0 : )));
8292 : }
8293 0 : policy => {
8294 0 : return Err(ApiError::PreconditionFailed(
8295 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
8296 0 : ));
8297 : }
8298 : }
8299 :
8300 0 : Ok(())
8301 0 : }
8302 :
8303 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
8304 0 : let node_available = {
8305 0 : let locked = self.inner.read().unwrap();
8306 0 : let nodes = &locked.nodes;
8307 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8308 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8309 0 : ))?;
8310 :
8311 0 : node.is_available()
8312 : };
8313 :
8314 0 : if !node_available {
8315 0 : return Err(ApiError::ResourceUnavailable(
8316 0 : format!("Node {node_id} is currently unavailable").into(),
8317 0 : ));
8318 0 : }
8319 :
8320 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8321 0 : if let Operation::Drain(drain) = op_handler.operation {
8322 0 : if drain.node_id == node_id {
8323 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8324 0 : op_handler.cancel.cancel();
8325 0 : return Ok(());
8326 0 : }
8327 0 : }
8328 0 : }
8329 :
8330 0 : Err(ApiError::PreconditionFailed(
8331 0 : format!("Node {node_id} has no drain in progress").into(),
8332 0 : ))
8333 0 : }
8334 :
8335 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
8336 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
8337 0 : let locked = self.inner.read().unwrap();
8338 0 : let nodes = &locked.nodes;
8339 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8340 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8341 0 : ))?;
8342 :
8343 : (
8344 0 : locked
8345 0 : .ongoing_operation
8346 0 : .as_ref()
8347 0 : .map(|ongoing| ongoing.operation),
8348 0 : node.is_available(),
8349 0 : node.get_scheduling(),
8350 0 : nodes.len(),
8351 : )
8352 : };
8353 :
8354 0 : if let Some(ongoing) = ongoing_op {
8355 0 : return Err(ApiError::PreconditionFailed(
8356 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8357 0 : ));
8358 0 : }
8359 :
8360 0 : if !node_available {
8361 0 : return Err(ApiError::ResourceUnavailable(
8362 0 : format!("Node {node_id} is currently unavailable").into(),
8363 0 : ));
8364 0 : }
8365 :
8366 0 : if total_nodes_count <= 1 {
8367 0 : return Err(ApiError::PreconditionFailed(
8368 0 : "No other nodes to fill from".into(),
8369 0 : ));
8370 0 : }
8371 :
8372 0 : match node_policy {
8373 : NodeSchedulingPolicy::Active => {
8374 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
8375 0 : .await?;
8376 :
8377 0 : let cancel = self.cancel.child_token();
8378 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8379 :
8380 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8381 0 : operation: Operation::Fill(Fill { node_id }),
8382 0 : cancel: cancel.clone(),
8383 0 : });
8384 :
8385 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
8386 :
8387 0 : tokio::task::spawn({
8388 0 : let service = self.clone();
8389 0 : let cancel = cancel.clone();
8390 0 : async move {
8391 0 : let _gate_guard = gate_guard;
8392 :
8393 0 : scopeguard::defer! {
8394 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8395 :
8396 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
8397 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
8398 : } else {
8399 : panic!("We always remove the same operation")
8400 : }
8401 : }
8402 :
8403 0 : tracing::info!("Fill background operation starting");
8404 0 : let res = service.fill_node(node_id, cancel).await;
8405 0 : match res {
8406 : Ok(()) => {
8407 0 : tracing::info!("Fill background operation completed successfully");
8408 : }
8409 : Err(OperationError::Cancelled) => {
8410 0 : tracing::info!("Fill background operation was cancelled");
8411 : }
8412 0 : Err(err) => {
8413 0 : tracing::error!("Fill background operation encountered: {err}")
8414 : }
8415 : }
8416 0 : }
8417 0 : }.instrument(span));
8418 : }
8419 : NodeSchedulingPolicy::Filling => {
8420 0 : return Err(ApiError::Conflict(format!(
8421 0 : "Node {node_id} has fill in progress"
8422 0 : )));
8423 : }
8424 0 : policy => {
8425 0 : return Err(ApiError::PreconditionFailed(
8426 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
8427 0 : ));
8428 : }
8429 : }
8430 :
8431 0 : Ok(())
8432 0 : }
8433 :
8434 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
8435 0 : let node_available = {
8436 0 : let locked = self.inner.read().unwrap();
8437 0 : let nodes = &locked.nodes;
8438 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8439 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8440 0 : ))?;
8441 :
8442 0 : node.is_available()
8443 : };
8444 :
8445 0 : if !node_available {
8446 0 : return Err(ApiError::ResourceUnavailable(
8447 0 : format!("Node {node_id} is currently unavailable").into(),
8448 0 : ));
8449 0 : }
8450 :
8451 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8452 0 : if let Operation::Fill(fill) = op_handler.operation {
8453 0 : if fill.node_id == node_id {
8454 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8455 0 : op_handler.cancel.cancel();
8456 0 : return Ok(());
8457 0 : }
8458 0 : }
8459 0 : }
8460 :
8461 0 : Err(ApiError::PreconditionFailed(
8462 0 : format!("Node {node_id} has no fill in progress").into(),
8463 0 : ))
8464 0 : }
8465 :
8466 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
8467 : /// configuration
8468 0 : fn maybe_reconcile_shard(
8469 0 : &self,
8470 0 : shard: &mut TenantShard,
8471 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8472 0 : priority: ReconcilerPriority,
8473 0 : ) -> Option<ReconcilerWaiter> {
8474 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
8475 0 : }
8476 :
8477 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
8478 0 : fn get_reconciler_units(
8479 0 : &self,
8480 0 : priority: ReconcilerPriority,
8481 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
8482 0 : let units = match priority {
8483 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
8484 : ReconcilerPriority::High => {
8485 0 : match self
8486 0 : .priority_reconciler_concurrency
8487 0 : .clone()
8488 0 : .try_acquire_owned()
8489 : {
8490 0 : Ok(u) => Ok(u),
8491 : Err(TryAcquireError::NoPermits) => {
8492 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
8493 : // the normal priority semaphore.
8494 0 : self.reconciler_concurrency.clone().try_acquire_owned()
8495 : }
8496 0 : Err(e) => Err(e),
8497 : }
8498 : }
8499 : };
8500 :
8501 0 : units.map(ReconcileUnits::new)
8502 0 : }
8503 :
8504 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
8505 0 : fn maybe_configured_reconcile_shard(
8506 0 : &self,
8507 0 : shard: &mut TenantShard,
8508 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8509 0 : reconciler_config: ReconcilerConfig,
8510 0 : ) -> Option<ReconcilerWaiter> {
8511 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
8512 :
8513 0 : let reconcile_reason = match reconcile_needed {
8514 0 : ReconcileNeeded::No => return None,
8515 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
8516 0 : ReconcileNeeded::Yes(reason) => {
8517 : // Fall through to try and acquire units for spawning reconciler
8518 0 : reason
8519 : }
8520 : };
8521 :
8522 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
8523 0 : Ok(u) => u,
8524 : Err(_) => {
8525 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
8526 0 : "Concurrency limited: enqueued for reconcile later");
8527 0 : if !shard.delayed_reconcile {
8528 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
8529 0 : Err(TrySendError::Closed(_)) => {
8530 0 : // Weird mid-shutdown case?
8531 0 : }
8532 : Err(TrySendError::Full(_)) => {
8533 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
8534 0 : tracing::warn!(
8535 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
8536 : );
8537 : }
8538 0 : Ok(()) => {
8539 0 : shard.delayed_reconcile = true;
8540 0 : }
8541 : }
8542 0 : }
8543 :
8544 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
8545 : // number to advance. When this function is eventually called again and succeeds in getting units,
8546 : // it will spawn a reconciler that makes this waiter complete.
8547 0 : return Some(shard.future_reconcile_waiter());
8548 : }
8549 : };
8550 :
8551 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
8552 : // Gate closed: we're shutting down, drop out.
8553 0 : return None;
8554 : };
8555 :
8556 0 : shard.spawn_reconciler(
8557 0 : reconcile_reason,
8558 0 : &self.result_tx,
8559 0 : nodes,
8560 0 : &self.compute_hook,
8561 0 : reconciler_config,
8562 0 : &self.config,
8563 0 : &self.persistence,
8564 0 : units,
8565 0 : gate_guard,
8566 0 : &self.reconcilers_cancel,
8567 0 : self.http_client.clone(),
8568 : )
8569 0 : }
8570 :
8571 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
8572 : /// Additionally, reschedule tenants that require it.
8573 : ///
8574 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
8575 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
8576 : /// available. A return value of 0 indicates that everything is fully reconciled already.
8577 0 : fn reconcile_all(&self) -> ReconcileAllResult {
8578 0 : let mut locked = self.inner.write().unwrap();
8579 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8580 0 : let pageservers = nodes.clone();
8581 :
8582 : // This function is an efficient place to update lazy statistics, since we are walking
8583 : // all tenants.
8584 0 : let mut pending_reconciles = 0;
8585 0 : let mut keep_failing_reconciles = 0;
8586 0 : let mut az_violations = 0;
8587 :
8588 : // If we find any tenants to drop from memory, stash them to offload after
8589 : // we're done traversing the map of tenants.
8590 0 : let mut drop_detached_tenants = Vec::new();
8591 :
8592 0 : let mut spawned_reconciles = 0;
8593 0 : let mut has_delayed_reconciles = false;
8594 :
8595 0 : for shard in tenants.values_mut() {
8596 : // Accumulate scheduling statistics
8597 0 : if let (Some(attached), Some(preferred)) =
8598 0 : (shard.intent.get_attached(), shard.preferred_az())
8599 : {
8600 0 : let node_az = nodes
8601 0 : .get(attached)
8602 0 : .expect("Nodes exist if referenced")
8603 0 : .get_availability_zone_id();
8604 0 : if node_az != preferred {
8605 0 : az_violations += 1;
8606 0 : }
8607 0 : }
8608 :
8609 : // Skip checking if this shard is already enqueued for reconciliation
8610 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
8611 : // If there is something delayed, then return a nonzero count so that
8612 : // callers like reconcile_all_now do not incorrectly get the impression
8613 : // that the system is in a quiescent state.
8614 0 : has_delayed_reconciles = true;
8615 0 : pending_reconciles += 1;
8616 0 : continue;
8617 0 : }
8618 :
8619 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
8620 : // dirty, spawn another one
8621 0 : let consecutive_errors_count = shard.consecutive_errors_count;
8622 0 : if self
8623 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
8624 0 : .is_some()
8625 : {
8626 0 : spawned_reconciles += 1;
8627 :
8628 : // Count shards that are keep-failing. We still want to reconcile them
8629 : // to avoid a situation where a shard is stuck.
8630 : // But we don't want to consider them when deciding to run optimizations.
8631 0 : if consecutive_errors_count >= MAX_CONSECUTIVE_RECONCILIATION_ERRORS {
8632 0 : tracing::warn!(
8633 : tenant_id=%shard.tenant_shard_id.tenant_id,
8634 0 : shard_id=%shard.tenant_shard_id.shard_slug(),
8635 0 : "Shard reconciliation is keep-failing: {} errors",
8636 : consecutive_errors_count
8637 : );
8638 0 : keep_failing_reconciles += 1;
8639 0 : }
8640 0 : } else if shard.delayed_reconcile {
8641 0 : // Shard wanted to reconcile but for some reason couldn't.
8642 0 : pending_reconciles += 1;
8643 0 : }
8644 :
8645 : // If this tenant is detached, try dropping it from memory. This is usually done
8646 : // proactively in [`Self::process_results`], but we do it here to handle the edge
8647 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
8648 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
8649 0 : && shard.policy == PlacementPolicy::Detached
8650 : {
8651 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
8652 0 : shard.tenant_shard_id.tenant_id,
8653 0 : TenantOperations::DropDetached,
8654 0 : ) {
8655 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
8656 0 : }
8657 0 : }
8658 : }
8659 :
8660 : // Some metrics are calculated from SchedulerNode state, update these periodically
8661 0 : scheduler.update_metrics();
8662 :
8663 : // Process any deferred tenant drops
8664 0 : for (tenant_id, guard) in drop_detached_tenants {
8665 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
8666 0 : }
8667 :
8668 0 : metrics::METRICS_REGISTRY
8669 0 : .metrics_group
8670 0 : .storage_controller_schedule_az_violation
8671 0 : .set(az_violations as i64);
8672 :
8673 0 : metrics::METRICS_REGISTRY
8674 0 : .metrics_group
8675 0 : .storage_controller_pending_reconciles
8676 0 : .set(pending_reconciles as i64);
8677 :
8678 0 : metrics::METRICS_REGISTRY
8679 0 : .metrics_group
8680 0 : .storage_controller_keep_failing_reconciles
8681 0 : .set(keep_failing_reconciles as i64);
8682 :
8683 0 : ReconcileAllResult::new(
8684 0 : spawned_reconciles,
8685 0 : keep_failing_reconciles,
8686 0 : has_delayed_reconciles,
8687 : )
8688 0 : }
8689 :
8690 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
8691 : /// could be scheduled somewhere better:
8692 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
8693 : /// * e.g. after a node fails then recovers, to move some work back to it
8694 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
8695 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
8696 : /// we did the split, but are probably better placed elsewhere.
8697 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
8698 : /// * e.g. after a shard split, some locations will be on the same node (where the split
8699 : /// happened), and will probably be better placed elsewhere.
8700 : ///
8701 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
8702 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
8703 : /// according to those same soft constraints.
8704 0 : async fn optimize_all(&self) -> usize {
8705 : // Limit on how many shards' optmizations each call to this function will execute. Combined
8706 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
8707 : // trickle of optimizations in the background, rather than executing a large number in parallel
8708 : // when a change occurs.
8709 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
8710 :
8711 : // Synchronous prepare: scan shards for possible scheduling optimizations
8712 0 : let candidate_work = self.optimize_all_plan();
8713 0 : let candidate_work_len = candidate_work.len();
8714 :
8715 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
8716 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
8717 :
8718 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
8719 :
8720 : // Synchronous apply: update the shards' intent states according to validated optimisations
8721 0 : let mut reconciles_spawned = 0;
8722 0 : let mut optimizations_applied = 0;
8723 0 : let mut locked = self.inner.write().unwrap();
8724 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8725 0 : for (tenant_shard_id, optimization) in validated_work {
8726 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
8727 : // Shard was dropped between planning and execution;
8728 0 : continue;
8729 : };
8730 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
8731 0 : if shard.apply_optimization(scheduler, optimization) {
8732 0 : optimizations_applied += 1;
8733 0 : if self
8734 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
8735 0 : .is_some()
8736 0 : {
8737 0 : reconciles_spawned += 1;
8738 0 : }
8739 0 : }
8740 :
8741 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
8742 0 : break;
8743 0 : }
8744 : }
8745 :
8746 0 : if was_work_filtered {
8747 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
8748 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
8749 0 : // as these validations start passing.
8750 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
8751 0 : }
8752 :
8753 0 : reconciles_spawned
8754 0 : }
8755 :
8756 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
8757 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
8758 : // this higher than the execution limit gives us a chance to execute some work even if the first
8759 : // few optimizations we find are not ready.
8760 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
8761 :
8762 0 : let mut work = Vec::new();
8763 0 : let mut locked = self.inner.write().unwrap();
8764 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
8765 :
8766 : // We are going to plan a bunch of optimisations before applying any of them, so the
8767 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
8768 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
8769 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
8770 : // to ignore the utilisation component of the score.
8771 :
8772 0 : for (_tenant_id, schedule_context, shards) in
8773 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Speculative)
8774 : {
8775 0 : for shard in shards {
8776 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
8777 0 : break;
8778 0 : }
8779 0 : match shard.get_scheduling_policy() {
8780 0 : ShardSchedulingPolicy::Active => {
8781 0 : // Ok to do optimization
8782 0 : }
8783 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
8784 0 : // Ok to do optimization: we are executing a graceful migration that
8785 0 : // has set preferred_node
8786 0 : }
8787 : ShardSchedulingPolicy::Essential
8788 : | ShardSchedulingPolicy::Pause
8789 : | ShardSchedulingPolicy::Stop => {
8790 : // Policy prevents optimizing this shard.
8791 0 : continue;
8792 : }
8793 : }
8794 :
8795 0 : if !matches!(shard.splitting, SplitState::Idle)
8796 0 : || matches!(shard.policy, PlacementPolicy::Detached)
8797 0 : || shard.reconciler.is_some()
8798 : {
8799 : // Do not start any optimizations while another change to the tenant is ongoing: this
8800 : // is not necessary for correctness, but simplifies operations and implicitly throttles
8801 : // optimization changes to happen in a "trickle" over time.
8802 0 : continue;
8803 0 : }
8804 :
8805 : // Fast path: we may quickly identify shards that don't have any possible optimisations
8806 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
8807 0 : if cfg!(feature = "testing") {
8808 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
8809 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
8810 : // panic in prod if we hit this, or spend cycles on it in prod.
8811 0 : assert!(
8812 0 : shard
8813 0 : .optimize_attachment(scheduler, &schedule_context)
8814 0 : .is_none()
8815 : );
8816 0 : assert!(
8817 0 : shard
8818 0 : .optimize_secondary(scheduler, &schedule_context)
8819 0 : .is_none()
8820 : );
8821 0 : }
8822 0 : continue;
8823 0 : }
8824 :
8825 0 : if let Some(optimization) =
8826 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
8827 : // its primary location based on soft constraints, cut it over.
8828 0 : shard.optimize_attachment(scheduler, &schedule_context)
8829 : {
8830 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
8831 0 : work.push((shard.tenant_shard_id, optimization));
8832 0 : break;
8833 0 : } else if let Some(optimization) =
8834 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
8835 : // better placed on another node, based on ScheduleContext, then adjust it. This
8836 : // covers cases like after a shard split, where we might have too many shards
8837 : // in the same tenant with secondary locations on the node where they originally split.
8838 0 : shard.optimize_secondary(scheduler, &schedule_context)
8839 : {
8840 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
8841 0 : work.push((shard.tenant_shard_id, optimization));
8842 0 : break;
8843 0 : }
8844 : }
8845 : }
8846 :
8847 0 : work
8848 0 : }
8849 :
8850 0 : async fn optimize_all_validate(
8851 0 : &self,
8852 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
8853 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
8854 : // Take a clone of the node map to use outside the lock in async validation phase
8855 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
8856 :
8857 0 : let mut want_secondary_status = Vec::new();
8858 :
8859 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
8860 : // check that the state of locations is acceptable to run the optimization, such as
8861 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
8862 : // in a live migration.
8863 0 : let mut validated_work = Vec::new();
8864 0 : for (tenant_shard_id, optimization) in candidate_work {
8865 0 : match optimization.action {
8866 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
8867 : old_attached_node_id: _,
8868 0 : new_attached_node_id,
8869 : }) => {
8870 0 : match validation_nodes.get(&new_attached_node_id) {
8871 0 : None => {
8872 0 : // Node was dropped between planning and validation
8873 0 : }
8874 0 : Some(node) => {
8875 0 : if !node.is_available() {
8876 0 : tracing::info!(
8877 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
8878 : );
8879 0 : } else {
8880 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
8881 0 : // remote API requests concurrently.
8882 0 : want_secondary_status.push((
8883 0 : tenant_shard_id,
8884 0 : node.clone(),
8885 0 : optimization,
8886 0 : ));
8887 0 : }
8888 : }
8889 : }
8890 : }
8891 : ScheduleOptimizationAction::ReplaceSecondary(_)
8892 : | ScheduleOptimizationAction::CreateSecondary(_)
8893 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
8894 : // No extra checks needed to manage secondaries: this does not interrupt client access
8895 0 : validated_work.push((tenant_shard_id, optimization))
8896 : }
8897 : };
8898 : }
8899 :
8900 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
8901 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
8902 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
8903 0 : let results = self
8904 0 : .tenant_for_shards_api(
8905 0 : want_secondary_status
8906 0 : .iter()
8907 0 : .map(|i| (i.0, i.1.clone()))
8908 0 : .collect(),
8909 0 : |tenant_shard_id, client| async move {
8910 0 : client.tenant_secondary_status(tenant_shard_id).await
8911 0 : },
8912 : 1,
8913 : 1,
8914 : SHORT_RECONCILE_TIMEOUT,
8915 0 : &self.cancel,
8916 : )
8917 0 : .await;
8918 :
8919 0 : for ((tenant_shard_id, node, optimization), (_, secondary_status)) in
8920 0 : want_secondary_status.into_iter().zip(results.into_iter())
8921 : {
8922 0 : match secondary_status {
8923 0 : Err(e) => {
8924 0 : tracing::info!(
8925 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
8926 : );
8927 : }
8928 0 : Ok(progress) => {
8929 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
8930 : // them in an optimization
8931 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
8932 :
8933 0 : if progress.heatmap_mtime.is_none()
8934 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
8935 0 : && progress.bytes_downloaded != progress.bytes_total
8936 0 : || progress.bytes_total - progress.bytes_downloaded
8937 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
8938 : {
8939 0 : tracing::info!(
8940 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
8941 : );
8942 :
8943 0 : if progress.heatmap_mtime.is_none() {
8944 : // No heatmap might mean the attached location has never uploaded one, or that
8945 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
8946 : // but fairly common in tests.
8947 0 : self.kick_secondary_download(tenant_shard_id).await;
8948 0 : }
8949 : } else {
8950 : // Location looks ready: proceed
8951 0 : tracing::info!(
8952 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
8953 : );
8954 0 : validated_work.push((tenant_shard_id, optimization))
8955 : }
8956 : }
8957 : }
8958 : }
8959 :
8960 0 : validated_work
8961 0 : }
8962 :
8963 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
8964 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
8965 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
8966 : /// we have this helper to move things along faster.
8967 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
8968 0 : if !self.config.kick_secondary_downloads {
8969 : // No-op if kick_secondary_downloads functionaliuty is not configured
8970 0 : return;
8971 0 : }
8972 :
8973 0 : let (attached_node, secondaries) = {
8974 0 : let locked = self.inner.read().unwrap();
8975 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
8976 0 : tracing::warn!(
8977 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
8978 : );
8979 0 : return;
8980 : };
8981 :
8982 0 : let Some(attached) = shard.intent.get_attached() else {
8983 0 : tracing::warn!(
8984 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
8985 : );
8986 0 : return;
8987 : };
8988 :
8989 0 : let secondaries = shard
8990 0 : .intent
8991 0 : .get_secondary()
8992 0 : .iter()
8993 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
8994 0 : .collect::<Vec<_>>();
8995 :
8996 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
8997 : };
8998 :
8999 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
9000 : // a 'kick' to let scheduling optimisation run more promptly.
9001 0 : match attached_node
9002 0 : .with_client_retries(
9003 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
9004 0 : &self.http_client,
9005 0 : &self.config.pageserver_jwt_token,
9006 : 3,
9007 : 10,
9008 : SHORT_RECONCILE_TIMEOUT,
9009 0 : &self.cancel,
9010 : )
9011 0 : .await
9012 : {
9013 0 : Some(Err(e)) => {
9014 0 : tracing::info!(
9015 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
9016 : );
9017 : }
9018 : None => {
9019 0 : tracing::info!(
9020 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
9021 : );
9022 : }
9023 : Some(Ok(_)) => {
9024 0 : tracing::info!(
9025 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
9026 : );
9027 : }
9028 : }
9029 :
9030 0 : for secondary_node in secondaries {
9031 0 : match secondary_node
9032 0 : .with_client_retries(
9033 0 : |client| async move {
9034 0 : client
9035 0 : .tenant_secondary_download(
9036 0 : tenant_shard_id,
9037 0 : Some(Duration::from_secs(1)),
9038 0 : )
9039 0 : .await
9040 0 : },
9041 0 : &self.http_client,
9042 0 : &self.config.pageserver_jwt_token,
9043 : 3,
9044 : 10,
9045 : SHORT_RECONCILE_TIMEOUT,
9046 0 : &self.cancel,
9047 : )
9048 0 : .await
9049 : {
9050 0 : Some(Err(e)) => {
9051 0 : tracing::info!(
9052 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
9053 : );
9054 : }
9055 : None => {
9056 0 : tracing::info!(
9057 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
9058 : );
9059 : }
9060 0 : Some(Ok(progress)) => {
9061 0 : tracing::info!(
9062 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
9063 : );
9064 : }
9065 : }
9066 : }
9067 0 : }
9068 :
9069 : /// Asynchronously split a tenant that's eligible for automatic splits. At most one tenant will
9070 : /// be split per call.
9071 : ///
9072 : /// Two sets of criteria are used: initial splits and size-based splits (in that order).
9073 : /// Initial splits are used to eagerly split unsharded tenants that may be performing initial
9074 : /// ingestion, since sharded tenants have significantly better ingestion throughput. Size-based
9075 : /// splits are used to bound the maximum shard size and balance out load.
9076 : ///
9077 : /// Splits are based on max_logical_size, i.e. the logical size of the largest timeline in a
9078 : /// tenant. We use this instead of the total logical size because branches will duplicate
9079 : /// logical size without actually using more storage. We could also use visible physical size,
9080 : /// but this might overestimate tenants that frequently churn branches.
9081 : ///
9082 : /// Initial splits (initial_split_threshold):
9083 : /// * Applies to tenants with 1 shard.
9084 : /// * The largest timeline (max_logical_size) exceeds initial_split_threshold.
9085 : /// * Splits into initial_split_shards.
9086 : ///
9087 : /// Size-based splits (split_threshold):
9088 : /// * Applies to all tenants.
9089 : /// * The largest timeline (max_logical_size) divided by shard count exceeds split_threshold.
9090 : /// * Splits such that max_logical_size / shard_count <= split_threshold, in powers of 2.
9091 : ///
9092 : /// Tenant shards are ordered by descending max_logical_size, first initial split candidates
9093 : /// then size-based split candidates. The first matching candidate is split.
9094 : ///
9095 : /// The shard count is clamped to max_split_shards. If a candidate is eligible for both initial
9096 : /// and size-based splits, the largest shard count will be used.
9097 : ///
9098 : /// An unsharded tenant will get DEFAULT_STRIPE_SIZE, regardless of what its ShardIdentity says.
9099 : /// A sharded tenant will retain its stripe size, as splits do not allow changing it.
9100 : ///
9101 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
9102 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
9103 : /// block all other splits.
9104 0 : async fn autosplit_tenants(self: &Arc<Self>) {
9105 : // If max_split_shards is set to 0 or 1, we can't split.
9106 0 : let max_split_shards = self.config.max_split_shards;
9107 0 : if max_split_shards <= 1 {
9108 0 : return;
9109 0 : }
9110 :
9111 : // If initial_split_shards is set to 0 or 1, disable initial splits.
9112 0 : let mut initial_split_threshold = self.config.initial_split_threshold.unwrap_or(0);
9113 0 : let initial_split_shards = self.config.initial_split_shards;
9114 0 : if initial_split_shards <= 1 {
9115 0 : initial_split_threshold = 0;
9116 0 : }
9117 :
9118 : // If no split_threshold nor initial_split_threshold, disable autosplits.
9119 0 : let split_threshold = self.config.split_threshold.unwrap_or(0);
9120 0 : if split_threshold == 0 && initial_split_threshold == 0 {
9121 0 : return;
9122 0 : }
9123 :
9124 : // Fetch split candidates in prioritized order.
9125 : //
9126 : // If initial splits are enabled, fetch eligible tenants first. We prioritize initial splits
9127 : // over size-based splits, since these are often performing initial ingestion and rely on
9128 : // splits to improve ingest throughput.
9129 0 : let mut candidates = Vec::new();
9130 :
9131 0 : if initial_split_threshold > 0 {
9132 : // Initial splits: fetch tenants with 1 shard where the logical size of the largest
9133 : // timeline exceeds the initial split threshold.
9134 0 : let initial_candidates = self
9135 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9136 0 : order_by: TenantSorting::MaxLogicalSize,
9137 0 : limit: 10,
9138 0 : where_shards_lt: Some(ShardCount(2)),
9139 0 : where_gt: Some(initial_split_threshold),
9140 0 : })
9141 0 : .await;
9142 0 : candidates.extend(initial_candidates);
9143 0 : }
9144 :
9145 0 : if split_threshold > 0 {
9146 : // Size-based splits: fetch tenants where the logical size of the largest timeline
9147 : // divided by shard count exceeds the split threshold.
9148 : //
9149 : // max_logical_size is only tracked on shard 0, and contains the total logical size
9150 : // across all shards. We have to order and filter by MaxLogicalSizePerShard, i.e.
9151 : // max_logical_size / shard_count, such that we only receive tenants that are actually
9152 : // eligible for splits. But we still use max_logical_size for later split calculations.
9153 0 : let size_candidates = self
9154 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
9155 0 : order_by: TenantSorting::MaxLogicalSizePerShard,
9156 0 : limit: 10,
9157 0 : where_shards_lt: Some(ShardCount(max_split_shards)),
9158 0 : where_gt: Some(split_threshold),
9159 0 : })
9160 0 : .await;
9161 : #[cfg(feature = "testing")]
9162 0 : assert!(
9163 0 : size_candidates.iter().all(|c| c.id.is_shard_zero()),
9164 0 : "MaxLogicalSizePerShard returned non-zero shard: {size_candidates:?}",
9165 : );
9166 0 : candidates.extend(size_candidates);
9167 0 : }
9168 :
9169 : // Filter out tenants in a prohibiting scheduling modes
9170 : // and tenants with an ongoing import.
9171 : //
9172 : // Note that the import check here is oportunistic. An import might start
9173 : // after the check before we actually update [`TenantShard::splitting`].
9174 : // [`Self::tenant_shard_split`] checks the database whilst holding the exclusive
9175 : // tenant lock. Imports might take a long time, so the check here allows us
9176 : // to split something else instead of trying the same shard over and over.
9177 : {
9178 0 : let state = self.inner.read().unwrap();
9179 0 : candidates.retain(|i| {
9180 0 : let shard = state.tenants.get(&i.id);
9181 0 : match shard {
9182 0 : Some(t) => {
9183 0 : t.get_scheduling_policy() == ShardSchedulingPolicy::Active
9184 0 : && t.importing == TimelineImportState::Idle
9185 : }
9186 0 : None => false,
9187 : }
9188 0 : });
9189 : }
9190 :
9191 : // Pick the first candidate to split. This will generally always be the first one in
9192 : // candidates, but we defensively skip candidates that end up not actually splitting.
9193 0 : let Some((candidate, new_shard_count)) = candidates
9194 0 : .into_iter()
9195 0 : .filter_map(|candidate| {
9196 0 : let new_shard_count = Self::compute_split_shards(ShardSplitInputs {
9197 0 : shard_count: candidate.id.shard_count,
9198 0 : max_logical_size: candidate.max_logical_size,
9199 0 : split_threshold,
9200 0 : max_split_shards,
9201 0 : initial_split_threshold,
9202 0 : initial_split_shards,
9203 0 : });
9204 0 : new_shard_count.map(|shards| (candidate, shards.count()))
9205 0 : })
9206 0 : .next()
9207 : else {
9208 0 : debug!("no split-eligible tenants found");
9209 0 : return;
9210 : };
9211 :
9212 : // Retain the stripe size of sharded tenants, as splits don't allow changing it. Otherwise,
9213 : // use DEFAULT_STRIPE_SIZE for unsharded tenants -- their stripe size doesn't really matter,
9214 : // and if we change the default stripe size we want to use the new default rather than an
9215 : // old, persisted stripe size.
9216 0 : let new_stripe_size = match candidate.id.shard_count.count() {
9217 0 : 0 => panic!("invalid shard count 0"),
9218 0 : 1 => Some(DEFAULT_STRIPE_SIZE),
9219 0 : 2.. => None,
9220 : };
9221 :
9222 : // We spawn a task to run this, so it's exactly like some external API client requesting
9223 : // it. We don't want to block the background reconcile loop on this.
9224 0 : let old_shard_count = candidate.id.shard_count.count();
9225 0 : info!(
9226 0 : "auto-splitting tenant {old_shard_count} → {new_shard_count} shards, \
9227 0 : current size {candidate:?} (split_threshold={split_threshold} \
9228 0 : initial_split_threshold={initial_split_threshold})"
9229 : );
9230 :
9231 0 : let this = self.clone();
9232 0 : tokio::spawn(
9233 0 : async move {
9234 0 : match this
9235 0 : .tenant_shard_split(
9236 0 : candidate.id.tenant_id,
9237 0 : TenantShardSplitRequest {
9238 0 : new_shard_count,
9239 0 : new_stripe_size,
9240 0 : },
9241 0 : )
9242 0 : .await
9243 : {
9244 : Ok(_) => {
9245 0 : info!("successful auto-split {old_shard_count} → {new_shard_count} shards")
9246 : }
9247 0 : Err(err) => error!("auto-split failed: {err}"),
9248 : }
9249 0 : }
9250 0 : .instrument(info_span!("auto_split", tenant_id=%candidate.id.tenant_id)),
9251 : );
9252 0 : }
9253 :
9254 : /// Returns the number of shards to split a tenant into, or None if the tenant shouldn't split,
9255 : /// based on the total logical size of the largest timeline summed across all shards. Uses the
9256 : /// larger of size-based and initial splits, clamped to max_split_shards.
9257 : ///
9258 : /// NB: the thresholds are exclusive, since TopTenantShardsRequest uses where_gt.
9259 25 : fn compute_split_shards(inputs: ShardSplitInputs) -> Option<ShardCount> {
9260 : let ShardSplitInputs {
9261 25 : shard_count,
9262 25 : max_logical_size,
9263 25 : split_threshold,
9264 25 : max_split_shards,
9265 25 : initial_split_threshold,
9266 25 : initial_split_shards,
9267 25 : } = inputs;
9268 :
9269 25 : let mut new_shard_count: u8 = shard_count.count();
9270 :
9271 : // Size-based splits. Ensures max_logical_size / new_shard_count <= split_threshold, using
9272 : // power-of-two shard counts.
9273 : //
9274 : // If the current shard count is not a power of two, and does not exceed split_threshold,
9275 : // then we leave it alone rather than forcing a power-of-two split.
9276 25 : if split_threshold > 0
9277 18 : && max_logical_size.div_ceil(split_threshold) > shard_count.count() as u64
9278 12 : {
9279 12 : new_shard_count = max_logical_size
9280 12 : .div_ceil(split_threshold)
9281 12 : .checked_next_power_of_two()
9282 12 : .unwrap_or(u8::MAX as u64)
9283 12 : .try_into()
9284 12 : .unwrap_or(u8::MAX);
9285 13 : }
9286 :
9287 : // Initial splits. Use the larger of size-based and initial split shard counts. This only
9288 : // applies to unsharded tenants, i.e. changes to initial_split_threshold or
9289 : // initial_split_shards are not retroactive for sharded tenants.
9290 25 : if initial_split_threshold > 0
9291 14 : && shard_count.count() <= 1
9292 11 : && max_logical_size > initial_split_threshold
9293 8 : {
9294 8 : new_shard_count = new_shard_count.max(initial_split_shards);
9295 17 : }
9296 :
9297 : // Clamp to max shards.
9298 25 : new_shard_count = new_shard_count.min(max_split_shards);
9299 :
9300 : // Don't split if we're not increasing the shard count.
9301 25 : if new_shard_count <= shard_count.count() {
9302 10 : return None;
9303 15 : }
9304 :
9305 15 : Some(ShardCount(new_shard_count))
9306 25 : }
9307 :
9308 : /// Fetches the top tenant shards from every available node, in descending order of
9309 : /// max logical size. Offline nodes are skipped, and any errors from available nodes
9310 : /// will be logged and ignored.
9311 0 : async fn get_top_tenant_shards(
9312 0 : &self,
9313 0 : request: &TopTenantShardsRequest,
9314 0 : ) -> Vec<TopTenantShardItem> {
9315 0 : let nodes = self
9316 0 : .inner
9317 0 : .read()
9318 0 : .unwrap()
9319 0 : .nodes
9320 0 : .values()
9321 0 : .filter(|node| node.is_available())
9322 0 : .cloned()
9323 0 : .collect_vec();
9324 :
9325 0 : let mut futures = FuturesUnordered::new();
9326 0 : for node in nodes {
9327 0 : futures.push(async move {
9328 0 : node.with_client_retries(
9329 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
9330 0 : &self.http_client,
9331 0 : &self.config.pageserver_jwt_token,
9332 : 3,
9333 : 3,
9334 0 : Duration::from_secs(5),
9335 0 : &self.cancel,
9336 : )
9337 0 : .await
9338 0 : });
9339 : }
9340 :
9341 0 : let mut top = Vec::new();
9342 0 : while let Some(output) = futures.next().await {
9343 0 : match output {
9344 0 : Some(Ok(response)) => top.extend(response.shards),
9345 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
9346 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
9347 0 : None => {} // node is shutting down
9348 : }
9349 : }
9350 :
9351 0 : top.sort_by_key(|i| i.max_logical_size);
9352 0 : top.reverse();
9353 0 : top
9354 0 : }
9355 :
9356 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
9357 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
9358 : /// put the system into a quiescent state where future background reconciliations won't do anything.
9359 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
9360 0 : let reconcile_all_result = self.reconcile_all();
9361 0 : let mut spawned_reconciles = reconcile_all_result.spawned_reconciles;
9362 0 : if reconcile_all_result.can_run_optimizations() {
9363 : // Only optimize when we are otherwise idle
9364 0 : let optimization_reconciles = self.optimize_all().await;
9365 0 : spawned_reconciles += optimization_reconciles;
9366 0 : }
9367 :
9368 0 : let waiters = {
9369 0 : let mut waiters = Vec::new();
9370 0 : let locked = self.inner.read().unwrap();
9371 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
9372 0 : if let Some(waiter) = shard.get_waiter() {
9373 0 : waiters.push(waiter);
9374 0 : }
9375 : }
9376 0 : waiters
9377 : };
9378 :
9379 0 : let waiter_count = waiters.len();
9380 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
9381 0 : Ok(()) => {}
9382 0 : Err(e) => {
9383 0 : if let ReconcileWaitError::Failed(_, reconcile_error) = &e {
9384 0 : match **reconcile_error {
9385 : ReconcileError::Cancel
9386 0 : | ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
9387 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
9388 0 : // because some other change superceded it. We will return a nonzero number,
9389 0 : // so the caller knows they might have to call again to quiesce the system.
9390 0 : }
9391 : _ => {
9392 0 : return Err(e);
9393 : }
9394 : }
9395 : } else {
9396 0 : return Err(e);
9397 : }
9398 : }
9399 : };
9400 :
9401 0 : tracing::info!(
9402 0 : "{} reconciles in reconcile_all, {} waiters",
9403 : spawned_reconciles,
9404 : waiter_count
9405 : );
9406 :
9407 0 : Ok(std::cmp::max(waiter_count, spawned_reconciles))
9408 0 : }
9409 :
9410 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
9411 : // Cancel all on-going reconciles and wait for them to exit the gate.
9412 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
9413 0 : self.reconcilers_cancel.cancel();
9414 0 : self.reconcilers_gate.close().await;
9415 :
9416 : // Signal the background loop in [`Service::process_results`] to exit once
9417 : // it has proccessed the results from all the reconciles we cancelled earlier.
9418 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
9419 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
9420 0 : self.result_tx.closed().await;
9421 0 : }
9422 :
9423 0 : pub async fn shutdown(&self) {
9424 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
9425 0 : .await;
9426 :
9427 : // Background tasks hold gate guards: this notifies them of the cancellation and
9428 : // waits for them all to complete.
9429 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
9430 0 : self.cancel.cancel();
9431 0 : self.gate.close().await;
9432 0 : }
9433 :
9434 : /// Spot check the download lag for a secondary location of a shard.
9435 : /// Should be used as a heuristic, since it's not always precise: the
9436 : /// secondary might have not downloaded the new heat map yet and, hence,
9437 : /// is not aware of the lag.
9438 : ///
9439 : /// Returns:
9440 : /// * Ok(None) if the lag could not be determined from the status,
9441 : /// * Ok(Some(_)) if the lag could be determind
9442 : /// * Err on failures to query the pageserver.
9443 0 : async fn secondary_lag(
9444 0 : &self,
9445 0 : secondary: &NodeId,
9446 0 : tenant_shard_id: TenantShardId,
9447 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
9448 0 : let nodes = self.inner.read().unwrap().nodes.clone();
9449 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
9450 0 : StatusCode::NOT_FOUND,
9451 0 : format!("Node with id {secondary} not found"),
9452 0 : ))?;
9453 :
9454 0 : match node
9455 0 : .with_client_retries(
9456 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
9457 0 : &self.http_client,
9458 0 : &self.config.pageserver_jwt_token,
9459 : 1,
9460 : 3,
9461 0 : Duration::from_millis(250),
9462 0 : &self.cancel,
9463 : )
9464 0 : .await
9465 : {
9466 0 : Some(Ok(status)) => match status.heatmap_mtime {
9467 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
9468 0 : None => Ok(None),
9469 : },
9470 0 : Some(Err(e)) => Err(e),
9471 0 : None => Err(mgmt_api::Error::Cancelled),
9472 : }
9473 0 : }
9474 :
9475 : /// Drain a node by moving the shards attached to it as primaries.
9476 : /// This is a long running operation and it should run as a separate Tokio task.
9477 0 : pub(crate) async fn drain_node(
9478 0 : self: &Arc<Self>,
9479 0 : node_id: NodeId,
9480 0 : cancel: CancellationToken,
9481 0 : ) -> Result<(), OperationError> {
9482 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
9483 0 : let max_secondary_lag_bytes = self
9484 0 : .config
9485 0 : .max_secondary_lag_bytes
9486 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
9487 :
9488 : // By default, live migrations are generous about the wait time for getting
9489 : // the secondary location up to speed. When draining, give up earlier in order
9490 : // to not stall the operation when a cold secondary is encountered.
9491 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9492 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9493 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9494 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9495 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9496 0 : .build();
9497 :
9498 0 : let mut waiters = Vec::new();
9499 :
9500 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
9501 :
9502 0 : while !tid_iter.finished() {
9503 0 : if cancel.is_cancelled() {
9504 0 : match self
9505 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9506 0 : .await
9507 : {
9508 0 : Ok(()) => return Err(OperationError::Cancelled),
9509 0 : Err(err) => {
9510 0 : return Err(OperationError::FinalizeError(
9511 0 : format!(
9512 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9513 0 : )
9514 0 : .into(),
9515 0 : ));
9516 : }
9517 : }
9518 0 : }
9519 :
9520 0 : operation_utils::validate_node_state(
9521 0 : &node_id,
9522 0 : self.inner.read().unwrap().nodes.clone(),
9523 0 : NodeSchedulingPolicy::Draining,
9524 0 : )?;
9525 :
9526 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9527 0 : let tid = match tid_iter.next() {
9528 0 : Some(tid) => tid,
9529 : None => {
9530 0 : break;
9531 : }
9532 : };
9533 :
9534 0 : let tid_drain = TenantShardDrain {
9535 0 : drained_node: node_id,
9536 0 : tenant_shard_id: tid,
9537 0 : };
9538 :
9539 0 : let dest_node_id = {
9540 0 : let locked = self.inner.read().unwrap();
9541 :
9542 0 : match tid_drain
9543 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
9544 : {
9545 0 : Some(node_id) => node_id,
9546 : None => {
9547 0 : continue;
9548 : }
9549 : }
9550 : };
9551 :
9552 0 : match self.secondary_lag(&dest_node_id, tid).await {
9553 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
9554 0 : // The secondary is reasonably up to date.
9555 0 : // Migrate to it
9556 0 : }
9557 0 : Ok(Some(lag)) => {
9558 0 : tracing::info!(
9559 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9560 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
9561 : );
9562 0 : continue;
9563 : }
9564 : Ok(None) => {
9565 0 : tracing::info!(
9566 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9567 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
9568 : );
9569 0 : continue;
9570 : }
9571 0 : Err(err) => {
9572 0 : tracing::warn!(
9573 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9574 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
9575 : );
9576 0 : continue;
9577 : }
9578 : }
9579 :
9580 : {
9581 0 : let mut locked = self.inner.write().unwrap();
9582 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9583 0 : let rescheduled = tid_drain.reschedule_to_secondary(
9584 0 : dest_node_id,
9585 0 : tenants,
9586 0 : scheduler,
9587 0 : nodes,
9588 0 : )?;
9589 :
9590 0 : if let Some(tenant_shard) = rescheduled {
9591 0 : let waiter = self.maybe_configured_reconcile_shard(
9592 0 : tenant_shard,
9593 0 : nodes,
9594 0 : reconciler_config,
9595 0 : );
9596 0 : if let Some(some) = waiter {
9597 0 : waiters.push(some);
9598 0 : }
9599 0 : }
9600 : }
9601 : }
9602 :
9603 0 : waiters = self
9604 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9605 0 : .await;
9606 :
9607 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
9608 : }
9609 :
9610 0 : while !waiters.is_empty() {
9611 0 : if cancel.is_cancelled() {
9612 0 : match self
9613 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9614 0 : .await
9615 : {
9616 0 : Ok(()) => return Err(OperationError::Cancelled),
9617 0 : Err(err) => {
9618 0 : return Err(OperationError::FinalizeError(
9619 0 : format!(
9620 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9621 0 : )
9622 0 : .into(),
9623 0 : ));
9624 : }
9625 : }
9626 0 : }
9627 :
9628 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
9629 :
9630 0 : waiters = self
9631 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9632 0 : .await;
9633 : }
9634 :
9635 : // At this point we have done the best we could to drain shards from this node.
9636 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
9637 : // to complete the drain.
9638 0 : if let Err(err) = self
9639 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
9640 0 : .await
9641 : {
9642 : // This is not fatal. Anything that is polling the node scheduling policy to detect
9643 : // the end of the drain operations will hang, but all such places should enforce an
9644 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
9645 : // by the counterpart fill operation.
9646 0 : return Err(OperationError::FinalizeError(
9647 0 : format!(
9648 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
9649 0 : )
9650 0 : .into(),
9651 0 : ));
9652 0 : }
9653 :
9654 0 : Ok(())
9655 0 : }
9656 :
9657 : /// Create a node fill plan (pick secondaries to promote), based on:
9658 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
9659 : /// outside their home AZ, should be migrated back here.
9660 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
9661 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
9662 : /// those shards have a home AZ that doesn't match the node we're filling.
9663 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
9664 0 : let mut locked = self.inner.write().unwrap();
9665 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
9666 :
9667 0 : let node_az = nodes
9668 0 : .get(&node_id)
9669 0 : .expect("Node must exist")
9670 0 : .get_availability_zone_id()
9671 0 : .clone();
9672 :
9673 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
9674 0 : let mut plan = Vec::new();
9675 :
9676 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
9677 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
9678 :
9679 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
9680 : // conceivably come up in real life if deploying a single-AZ region intentionally.
9681 0 : let respect_azs = nodes
9682 0 : .values()
9683 0 : .map(|n| n.get_availability_zone_id())
9684 0 : .unique()
9685 0 : .count()
9686 : > 1;
9687 :
9688 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
9689 : // requires it.
9690 0 : for (tsid, tenant_shard) in tenants {
9691 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9692 : // Shard doesn't have a secondary on this node, ignore it.
9693 0 : continue;
9694 0 : }
9695 :
9696 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
9697 : // shards which belong on this node, not to promote shards whose scheduling preference
9698 : // would be on their currently attached node. So will avoid promoting shards whose
9699 : // home AZ doesn't match the AZ of the node we're filling.
9700 0 : match tenant_shard.preferred_az() {
9701 0 : _ if !respect_azs => {
9702 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9703 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9704 0 : }
9705 : }
9706 : None => {
9707 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
9708 : // will only do so if our target shard count requires it.
9709 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9710 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9711 0 : }
9712 : }
9713 0 : Some(az) if az == &node_az => {
9714 : // This shard's home AZ is equal to the node we're filling: it should
9715 : // be moved back to this node as part of filling, unless its currently
9716 : // attached location is also in its home AZ.
9717 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9718 0 : if nodes
9719 0 : .get(primary)
9720 0 : .expect("referenced node must exist")
9721 0 : .get_availability_zone_id()
9722 0 : != tenant_shard
9723 0 : .preferred_az()
9724 0 : .expect("tenant must have an AZ preference")
9725 : {
9726 0 : plan.push(*tsid)
9727 0 : }
9728 : } else {
9729 0 : plan.push(*tsid)
9730 : }
9731 : }
9732 0 : Some(_) => {
9733 0 : // This shard's home AZ is somewhere other than the node we're filling,
9734 0 : // it may not be moved back to this node as part of filling. Ignore it
9735 0 : }
9736 : }
9737 : }
9738 :
9739 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
9740 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
9741 :
9742 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
9743 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
9744 :
9745 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
9746 :
9747 0 : for (node_id, attached) in nodes_by_load {
9748 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
9749 0 : if !available {
9750 0 : continue;
9751 0 : }
9752 :
9753 0 : if plan.len() >= fill_requirement
9754 0 : || free_tids_by_node.is_empty()
9755 0 : || attached <= expected_attached
9756 : {
9757 0 : break;
9758 0 : }
9759 :
9760 0 : let can_take = attached - expected_attached;
9761 0 : let needed = fill_requirement - plan.len();
9762 0 : let mut take = std::cmp::min(can_take, needed);
9763 :
9764 0 : let mut remove_node = false;
9765 0 : while take > 0 {
9766 0 : match free_tids_by_node.get_mut(&node_id) {
9767 0 : Some(tids) => match tids.pop() {
9768 0 : Some(tid) => {
9769 0 : let max_promote_for_tenant = std::cmp::max(
9770 0 : tid.shard_count.count() as usize / locked.nodes.len(),
9771 : 1,
9772 : );
9773 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
9774 0 : if *promoted < max_promote_for_tenant {
9775 0 : plan.push(tid);
9776 0 : *promoted += 1;
9777 0 : take -= 1;
9778 0 : }
9779 : }
9780 : None => {
9781 0 : remove_node = true;
9782 0 : break;
9783 : }
9784 : },
9785 : None => {
9786 0 : break;
9787 : }
9788 : }
9789 : }
9790 :
9791 0 : if remove_node {
9792 0 : free_tids_by_node.remove(&node_id);
9793 0 : }
9794 : }
9795 :
9796 0 : plan
9797 0 : }
9798 :
9799 : /// Fill a node by promoting its secondaries until the cluster is balanced
9800 : /// with regards to attached shard counts. Note that this operation only
9801 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
9802 : /// This is a long running operation and it should run as a separate Tokio task.
9803 0 : pub(crate) async fn fill_node(
9804 0 : &self,
9805 0 : node_id: NodeId,
9806 0 : cancel: CancellationToken,
9807 0 : ) -> Result<(), OperationError> {
9808 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9809 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9810 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9811 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9812 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9813 0 : .build();
9814 :
9815 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
9816 0 : let mut waiters = Vec::new();
9817 :
9818 : // Execute the plan we've composed above. Before aplying each move from the plan,
9819 : // we validate to ensure that it has not gone stale in the meantime.
9820 0 : while !tids_to_promote.is_empty() {
9821 0 : if cancel.is_cancelled() {
9822 0 : match self
9823 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9824 0 : .await
9825 : {
9826 0 : Ok(()) => return Err(OperationError::Cancelled),
9827 0 : Err(err) => {
9828 0 : return Err(OperationError::FinalizeError(
9829 0 : format!(
9830 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9831 0 : )
9832 0 : .into(),
9833 0 : ));
9834 : }
9835 : }
9836 0 : }
9837 :
9838 : {
9839 0 : let mut locked = self.inner.write().unwrap();
9840 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9841 :
9842 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
9843 0 : format!("node {node_id} was removed").into(),
9844 0 : ))?;
9845 :
9846 0 : let current_policy = node.get_scheduling();
9847 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
9848 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
9849 : // about it
9850 0 : return Err(OperationError::NodeStateChanged(
9851 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
9852 0 : ));
9853 0 : }
9854 :
9855 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9856 0 : if let Some(tid) = tids_to_promote.pop() {
9857 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
9858 : // If the node being filled is not a secondary anymore,
9859 : // skip the promotion.
9860 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9861 0 : continue;
9862 0 : }
9863 :
9864 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
9865 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
9866 0 : Err(e) => {
9867 0 : tracing::warn!(
9868 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9869 0 : "Scheduling error when filling pageserver {} : {e}", node_id
9870 : );
9871 : }
9872 : Ok(()) => {
9873 0 : tracing::info!(
9874 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9875 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
9876 : node_id,
9877 : previously_attached_to,
9878 : node_id
9879 : );
9880 :
9881 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
9882 0 : tenant_shard,
9883 0 : nodes,
9884 0 : reconciler_config,
9885 0 : ) {
9886 0 : waiters.push(waiter);
9887 0 : }
9888 : }
9889 : }
9890 0 : }
9891 : } else {
9892 0 : break;
9893 : }
9894 : }
9895 : }
9896 :
9897 0 : waiters = self
9898 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9899 0 : .await;
9900 : }
9901 :
9902 0 : while !waiters.is_empty() {
9903 0 : if cancel.is_cancelled() {
9904 0 : match self
9905 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9906 0 : .await
9907 : {
9908 0 : Ok(()) => return Err(OperationError::Cancelled),
9909 0 : Err(err) => {
9910 0 : return Err(OperationError::FinalizeError(
9911 0 : format!(
9912 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9913 0 : )
9914 0 : .into(),
9915 0 : ));
9916 : }
9917 : }
9918 0 : }
9919 :
9920 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
9921 :
9922 0 : waiters = self
9923 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9924 0 : .await;
9925 : }
9926 :
9927 0 : if let Err(err) = self
9928 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9929 0 : .await
9930 : {
9931 : // This isn't a huge issue since the filling process starts upon request. However, it
9932 : // will prevent the next drain from starting. The only case in which this can fail
9933 : // is database unavailability. Such a case will require manual intervention.
9934 0 : return Err(OperationError::FinalizeError(
9935 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
9936 0 : .into(),
9937 0 : ));
9938 0 : }
9939 :
9940 0 : Ok(())
9941 0 : }
9942 :
9943 : /// Updates scrubber metadata health check results.
9944 0 : pub(crate) async fn metadata_health_update(
9945 0 : &self,
9946 0 : update_req: MetadataHealthUpdateRequest,
9947 0 : ) -> Result<(), ApiError> {
9948 0 : let now = chrono::offset::Utc::now();
9949 0 : let (healthy_records, unhealthy_records) = {
9950 0 : let locked = self.inner.read().unwrap();
9951 0 : let healthy_records = update_req
9952 0 : .healthy_tenant_shards
9953 0 : .into_iter()
9954 : // Retain only health records associated with tenant shards managed by storage controller.
9955 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
9956 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
9957 0 : .collect();
9958 0 : let unhealthy_records = update_req
9959 0 : .unhealthy_tenant_shards
9960 0 : .into_iter()
9961 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
9962 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
9963 0 : .collect();
9964 :
9965 0 : (healthy_records, unhealthy_records)
9966 : };
9967 :
9968 0 : self.persistence
9969 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
9970 0 : .await?;
9971 0 : Ok(())
9972 0 : }
9973 :
9974 : /// Lists the tenant shards that has unhealthy metadata status.
9975 0 : pub(crate) async fn metadata_health_list_unhealthy(
9976 0 : &self,
9977 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
9978 0 : let result = self
9979 0 : .persistence
9980 0 : .list_unhealthy_metadata_health_records()
9981 0 : .await?
9982 0 : .iter()
9983 0 : .map(|p| p.get_tenant_shard_id().unwrap())
9984 0 : .collect();
9985 :
9986 0 : Ok(result)
9987 0 : }
9988 :
9989 : /// Lists the tenant shards that have not been scrubbed for some duration.
9990 0 : pub(crate) async fn metadata_health_list_outdated(
9991 0 : &self,
9992 0 : not_scrubbed_for: Duration,
9993 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
9994 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
9995 0 : let result = self
9996 0 : .persistence
9997 0 : .list_outdated_metadata_health_records(earlier)
9998 0 : .await?
9999 0 : .into_iter()
10000 0 : .map(|record| record.into())
10001 0 : .collect();
10002 0 : Ok(result)
10003 0 : }
10004 :
10005 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
10006 0 : self.inner.read().unwrap().get_leadership_status()
10007 0 : }
10008 :
10009 : /// Handler for step down requests
10010 : ///
10011 : /// Step down runs in separate task since once it's called it should
10012 : /// be driven to completion. Subsequent requests will wait on the same
10013 : /// step down task.
10014 0 : pub(crate) async fn step_down(self: &Arc<Self>) -> GlobalObservedState {
10015 0 : let handle = self.step_down_barrier.get_or_init(|| {
10016 0 : let step_down_self = self.clone();
10017 0 : let (tx, rx) = tokio::sync::watch::channel::<Option<GlobalObservedState>>(None);
10018 0 : tokio::spawn(async move {
10019 0 : let state = step_down_self.step_down_task().await;
10020 0 : tx.send(Some(state))
10021 0 : .expect("Task Arc<Service> keeps receiver alive");
10022 0 : });
10023 :
10024 0 : rx
10025 0 : });
10026 :
10027 0 : handle
10028 0 : .clone()
10029 0 : .wait_for(|observed_state| observed_state.is_some())
10030 0 : .await
10031 0 : .expect("Task Arc<Service> keeps sender alive")
10032 0 : .deref()
10033 0 : .clone()
10034 0 : .expect("Checked above")
10035 0 : }
10036 :
10037 0 : async fn step_down_task(&self) -> GlobalObservedState {
10038 0 : tracing::info!("Received step down request from peer");
10039 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
10040 :
10041 0 : self.inner.write().unwrap().step_down();
10042 :
10043 0 : let stop_reconciliations =
10044 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown);
10045 0 : let mut stop_reconciliations = std::pin::pin!(stop_reconciliations);
10046 :
10047 0 : let started_at = Instant::now();
10048 :
10049 : // Wait for reconciliations to stop and warn if that's taking a long time
10050 : loop {
10051 0 : tokio::select! {
10052 0 : _ = &mut stop_reconciliations => {
10053 0 : tracing::info!("Reconciliations stopped, proceeding with step down");
10054 0 : break;
10055 : }
10056 0 : _ = tokio::time::sleep(Duration::from_secs(10)) => {
10057 0 : tracing::warn!(
10058 0 : elapsed_sec=%started_at.elapsed().as_secs(),
10059 0 : "Stopping reconciliations during step down is taking too long"
10060 : );
10061 : }
10062 : }
10063 : }
10064 :
10065 0 : let mut global_observed = GlobalObservedState::default();
10066 0 : let locked = self.inner.read().unwrap();
10067 0 : for (tid, tenant_shard) in locked.tenants.iter() {
10068 0 : global_observed
10069 0 : .0
10070 0 : .insert(*tid, tenant_shard.observed.clone());
10071 0 : }
10072 :
10073 0 : global_observed
10074 0 : }
10075 :
10076 0 : pub(crate) async fn update_shards_preferred_azs(
10077 0 : &self,
10078 0 : req: ShardsPreferredAzsRequest,
10079 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
10080 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
10081 0 : let updated = self
10082 0 : .persistence
10083 0 : .set_tenant_shard_preferred_azs(preferred_azs)
10084 0 : .await
10085 0 : .map_err(|err| {
10086 0 : ApiError::InternalServerError(anyhow::anyhow!(
10087 0 : "Failed to persist preferred AZs: {err}"
10088 0 : ))
10089 0 : })?;
10090 :
10091 0 : let mut updated_in_mem_and_db = Vec::default();
10092 :
10093 0 : let mut locked = self.inner.write().unwrap();
10094 0 : let state = locked.deref_mut();
10095 0 : for (tid, az_id) in updated {
10096 0 : let shard = state.tenants.get_mut(&tid);
10097 0 : if let Some(shard) = shard {
10098 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
10099 0 : updated_in_mem_and_db.push(tid);
10100 0 : }
10101 : }
10102 :
10103 0 : Ok(ShardsPreferredAzsResponse {
10104 0 : updated: updated_in_mem_and_db,
10105 0 : })
10106 0 : }
10107 : }
10108 :
10109 : #[cfg(test)]
10110 : mod tests {
10111 : use super::*;
10112 :
10113 : /// Tests Service::compute_split_shards. For readability, this specifies sizes in GBs rather
10114 : /// than bytes. Note that max_logical_size is the total logical size of the largest timeline
10115 : /// summed across all shards.
10116 : #[test]
10117 1 : fn compute_split_shards() {
10118 : // Size-based split: two shards have a 500 GB timeline, which need to split into 8 shards
10119 : // that are <= 64 GB,
10120 1 : assert_eq!(
10121 1 : Service::compute_split_shards(ShardSplitInputs {
10122 1 : shard_count: ShardCount(2),
10123 1 : max_logical_size: 500,
10124 1 : split_threshold: 64,
10125 1 : max_split_shards: 16,
10126 1 : initial_split_threshold: 0,
10127 1 : initial_split_shards: 0,
10128 1 : }),
10129 : Some(ShardCount(8))
10130 : );
10131 :
10132 : // Size-based split: noop at or below threshold, fires above.
10133 1 : assert_eq!(
10134 1 : Service::compute_split_shards(ShardSplitInputs {
10135 1 : shard_count: ShardCount(2),
10136 1 : max_logical_size: 127,
10137 1 : split_threshold: 64,
10138 1 : max_split_shards: 16,
10139 1 : initial_split_threshold: 0,
10140 1 : initial_split_shards: 0,
10141 1 : }),
10142 : None,
10143 : );
10144 1 : assert_eq!(
10145 1 : Service::compute_split_shards(ShardSplitInputs {
10146 1 : shard_count: ShardCount(2),
10147 1 : max_logical_size: 128,
10148 1 : split_threshold: 64,
10149 1 : max_split_shards: 16,
10150 1 : initial_split_threshold: 0,
10151 1 : initial_split_shards: 0,
10152 1 : }),
10153 : None,
10154 : );
10155 1 : assert_eq!(
10156 1 : Service::compute_split_shards(ShardSplitInputs {
10157 1 : shard_count: ShardCount(2),
10158 1 : max_logical_size: 129,
10159 1 : split_threshold: 64,
10160 1 : max_split_shards: 16,
10161 1 : initial_split_threshold: 0,
10162 1 : initial_split_shards: 0,
10163 1 : }),
10164 : Some(ShardCount(4)),
10165 : );
10166 :
10167 : // Size-based split: clamped to max_split_shards.
10168 1 : assert_eq!(
10169 1 : Service::compute_split_shards(ShardSplitInputs {
10170 1 : shard_count: ShardCount(2),
10171 1 : max_logical_size: 10000,
10172 1 : split_threshold: 64,
10173 1 : max_split_shards: 16,
10174 1 : initial_split_threshold: 0,
10175 1 : initial_split_shards: 0,
10176 1 : }),
10177 : Some(ShardCount(16))
10178 : );
10179 :
10180 : // Size-based split: tenant already at or beyond max_split_shards is not split.
10181 1 : assert_eq!(
10182 1 : Service::compute_split_shards(ShardSplitInputs {
10183 1 : shard_count: ShardCount(16),
10184 1 : max_logical_size: 10000,
10185 1 : split_threshold: 64,
10186 1 : max_split_shards: 16,
10187 1 : initial_split_threshold: 0,
10188 1 : initial_split_shards: 0,
10189 1 : }),
10190 : None
10191 : );
10192 :
10193 1 : assert_eq!(
10194 1 : Service::compute_split_shards(ShardSplitInputs {
10195 1 : shard_count: ShardCount(32),
10196 1 : max_logical_size: 10000,
10197 1 : split_threshold: 64,
10198 1 : max_split_shards: 16,
10199 1 : initial_split_threshold: 0,
10200 1 : initial_split_shards: 0,
10201 1 : }),
10202 : None
10203 : );
10204 :
10205 : // Size-based split: a non-power-of-2 shard count is normalized to power-of-2 if it
10206 : // exceeds split_threshold (i.e. a 3-shard tenant splits into 8, not 6).
10207 1 : assert_eq!(
10208 1 : Service::compute_split_shards(ShardSplitInputs {
10209 1 : shard_count: ShardCount(3),
10210 1 : max_logical_size: 320,
10211 1 : split_threshold: 64,
10212 1 : max_split_shards: 16,
10213 1 : initial_split_threshold: 0,
10214 1 : initial_split_shards: 0,
10215 1 : }),
10216 : Some(ShardCount(8))
10217 : );
10218 :
10219 : // Size-based split: a non-power-of-2 shard count is not normalized to power-of-2 if the
10220 : // existing shards are below or at split_threshold, but splits into 4 if it exceeds it.
10221 1 : assert_eq!(
10222 1 : Service::compute_split_shards(ShardSplitInputs {
10223 1 : shard_count: ShardCount(3),
10224 1 : max_logical_size: 191,
10225 1 : split_threshold: 64,
10226 1 : max_split_shards: 16,
10227 1 : initial_split_threshold: 0,
10228 1 : initial_split_shards: 0,
10229 1 : }),
10230 : None
10231 : );
10232 1 : assert_eq!(
10233 1 : Service::compute_split_shards(ShardSplitInputs {
10234 1 : shard_count: ShardCount(3),
10235 1 : max_logical_size: 192,
10236 1 : split_threshold: 64,
10237 1 : max_split_shards: 16,
10238 1 : initial_split_threshold: 0,
10239 1 : initial_split_shards: 0,
10240 1 : }),
10241 : None
10242 : );
10243 1 : assert_eq!(
10244 1 : Service::compute_split_shards(ShardSplitInputs {
10245 1 : shard_count: ShardCount(3),
10246 1 : max_logical_size: 193,
10247 1 : split_threshold: 64,
10248 1 : max_split_shards: 16,
10249 1 : initial_split_threshold: 0,
10250 1 : initial_split_shards: 0,
10251 1 : }),
10252 : Some(ShardCount(4))
10253 : );
10254 :
10255 : // Initial split: tenant has a 10 GB timeline, split into 4 shards.
10256 1 : assert_eq!(
10257 1 : Service::compute_split_shards(ShardSplitInputs {
10258 1 : shard_count: ShardCount(1),
10259 1 : max_logical_size: 10,
10260 1 : split_threshold: 0,
10261 1 : max_split_shards: 16,
10262 1 : initial_split_threshold: 8,
10263 1 : initial_split_shards: 4,
10264 1 : }),
10265 : Some(ShardCount(4))
10266 : );
10267 :
10268 : // Initial split: 0 ShardCount is equivalent to 1.
10269 1 : assert_eq!(
10270 1 : Service::compute_split_shards(ShardSplitInputs {
10271 1 : shard_count: ShardCount(0),
10272 1 : max_logical_size: 10,
10273 1 : split_threshold: 0,
10274 1 : max_split_shards: 16,
10275 1 : initial_split_threshold: 8,
10276 1 : initial_split_shards: 4,
10277 1 : }),
10278 : Some(ShardCount(4))
10279 : );
10280 :
10281 : // Initial split: at or below threshold is noop.
10282 1 : assert_eq!(
10283 1 : Service::compute_split_shards(ShardSplitInputs {
10284 1 : shard_count: ShardCount(1),
10285 1 : max_logical_size: 7,
10286 1 : split_threshold: 0,
10287 1 : max_split_shards: 16,
10288 1 : initial_split_threshold: 8,
10289 1 : initial_split_shards: 4,
10290 1 : }),
10291 : None,
10292 : );
10293 1 : assert_eq!(
10294 1 : Service::compute_split_shards(ShardSplitInputs {
10295 1 : shard_count: ShardCount(1),
10296 1 : max_logical_size: 8,
10297 1 : split_threshold: 0,
10298 1 : max_split_shards: 16,
10299 1 : initial_split_threshold: 8,
10300 1 : initial_split_shards: 4,
10301 1 : }),
10302 : None,
10303 : );
10304 1 : assert_eq!(
10305 1 : Service::compute_split_shards(ShardSplitInputs {
10306 1 : shard_count: ShardCount(1),
10307 1 : max_logical_size: 9,
10308 1 : split_threshold: 0,
10309 1 : max_split_shards: 16,
10310 1 : initial_split_threshold: 8,
10311 1 : initial_split_shards: 4,
10312 1 : }),
10313 : Some(ShardCount(4))
10314 : );
10315 :
10316 : // Initial split: already sharded tenant is not affected, even if above threshold and below
10317 : // shard count.
10318 1 : assert_eq!(
10319 1 : Service::compute_split_shards(ShardSplitInputs {
10320 1 : shard_count: ShardCount(2),
10321 1 : max_logical_size: 20,
10322 1 : split_threshold: 0,
10323 1 : max_split_shards: 16,
10324 1 : initial_split_threshold: 8,
10325 1 : initial_split_shards: 4,
10326 1 : }),
10327 : None,
10328 : );
10329 :
10330 : // Initial split: clamped to max_shards.
10331 1 : assert_eq!(
10332 1 : Service::compute_split_shards(ShardSplitInputs {
10333 1 : shard_count: ShardCount(1),
10334 1 : max_logical_size: 10,
10335 1 : split_threshold: 0,
10336 1 : max_split_shards: 3,
10337 1 : initial_split_threshold: 8,
10338 1 : initial_split_shards: 4,
10339 1 : }),
10340 : Some(ShardCount(3)),
10341 : );
10342 :
10343 : // Initial+size split: tenant eligible for both will use the larger shard count.
10344 1 : assert_eq!(
10345 1 : Service::compute_split_shards(ShardSplitInputs {
10346 1 : shard_count: ShardCount(1),
10347 1 : max_logical_size: 10,
10348 1 : split_threshold: 64,
10349 1 : max_split_shards: 16,
10350 1 : initial_split_threshold: 8,
10351 1 : initial_split_shards: 4,
10352 1 : }),
10353 : Some(ShardCount(4)),
10354 : );
10355 1 : assert_eq!(
10356 1 : Service::compute_split_shards(ShardSplitInputs {
10357 1 : shard_count: ShardCount(1),
10358 1 : max_logical_size: 500,
10359 1 : split_threshold: 64,
10360 1 : max_split_shards: 16,
10361 1 : initial_split_threshold: 8,
10362 1 : initial_split_shards: 4,
10363 1 : }),
10364 : Some(ShardCount(8)),
10365 : );
10366 :
10367 : // Initial+size split: sharded tenant is only eligible for size-based split.
10368 1 : assert_eq!(
10369 1 : Service::compute_split_shards(ShardSplitInputs {
10370 1 : shard_count: ShardCount(2),
10371 1 : max_logical_size: 200,
10372 1 : split_threshold: 64,
10373 1 : max_split_shards: 16,
10374 1 : initial_split_threshold: 8,
10375 1 : initial_split_shards: 8,
10376 1 : }),
10377 : Some(ShardCount(4)),
10378 : );
10379 :
10380 : // Initial+size split: uses the larger shard count even with initial_split_threshold above
10381 : // split_threshold.
10382 1 : assert_eq!(
10383 1 : Service::compute_split_shards(ShardSplitInputs {
10384 1 : shard_count: ShardCount(1),
10385 1 : max_logical_size: 10,
10386 1 : split_threshold: 4,
10387 1 : max_split_shards: 16,
10388 1 : initial_split_threshold: 8,
10389 1 : initial_split_shards: 8,
10390 1 : }),
10391 : Some(ShardCount(8)),
10392 : );
10393 :
10394 : // Test backwards compatibility with production settings when initial/size-based splits were
10395 : // rolled out: a single split into 8 shards at 64 GB. Any already sharded tenants with <8
10396 : // shards will split according to split_threshold.
10397 1 : assert_eq!(
10398 1 : Service::compute_split_shards(ShardSplitInputs {
10399 1 : shard_count: ShardCount(1),
10400 1 : max_logical_size: 65,
10401 1 : split_threshold: 64,
10402 1 : max_split_shards: 8,
10403 1 : initial_split_threshold: 64,
10404 1 : initial_split_shards: 8,
10405 1 : }),
10406 : Some(ShardCount(8)),
10407 : );
10408 :
10409 1 : assert_eq!(
10410 1 : Service::compute_split_shards(ShardSplitInputs {
10411 1 : shard_count: ShardCount(1),
10412 1 : max_logical_size: 64,
10413 1 : split_threshold: 64,
10414 1 : max_split_shards: 8,
10415 1 : initial_split_threshold: 64,
10416 1 : initial_split_shards: 8,
10417 1 : }),
10418 : None,
10419 : );
10420 :
10421 1 : assert_eq!(
10422 1 : Service::compute_split_shards(ShardSplitInputs {
10423 1 : shard_count: ShardCount(2),
10424 1 : max_logical_size: 129,
10425 1 : split_threshold: 64,
10426 1 : max_split_shards: 8,
10427 1 : initial_split_threshold: 64,
10428 1 : initial_split_shards: 8,
10429 1 : }),
10430 : Some(ShardCount(4)),
10431 : );
10432 1 : }
10433 : }
|
