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,
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::generation::Generation;
64 : use utils::id::{NodeId, TenantId, TimelineId};
65 : use utils::lsn::Lsn;
66 : use utils::shard::ShardIndex;
67 : use utils::sync::gate::{Gate, GateGuard};
68 : use utils::{failpoint_support, pausable_failpoint};
69 :
70 : use crate::background_node_operations::{
71 : Delete, Drain, Fill, MAX_RECONCILES_PER_OPERATION, Operation, OperationError, OperationHandler,
72 : };
73 : use crate::compute_hook::{self, ComputeHook, NotifyError};
74 : use crate::heartbeater::{Heartbeater, PageserverState, SafekeeperState};
75 : use crate::id_lock_map::{
76 : IdLockMap, TracingExclusiveGuard, trace_exclusive_lock, trace_shared_lock,
77 : };
78 : use crate::leadership::Leadership;
79 : use crate::metrics;
80 : use crate::node::{AvailabilityTransition, Node};
81 : use crate::operation_utils::{self, TenantShardDrain};
82 : use crate::pageserver_client::PageserverClient;
83 : use crate::peer_client::GlobalObservedState;
84 : use crate::persistence::split_state::SplitState;
85 : use crate::persistence::{
86 : AbortShardSplitStatus, ControllerPersistence, DatabaseError, DatabaseResult,
87 : MetadataHealthPersistence, Persistence, ShardGenerationState, TenantFilter,
88 : TenantShardPersistence,
89 : };
90 : use crate::reconciler::{
91 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder, ReconcilerPriority,
92 : attached_location_conf,
93 : };
94 : use crate::safekeeper::Safekeeper;
95 : use crate::scheduler::{
96 : AttachedShardTag, MaySchedule, ScheduleContext, ScheduleError, ScheduleMode, Scheduler,
97 : };
98 : use crate::tenant_shard::{
99 : IntentState, MigrateAttachment, ObservedState, ObservedStateDelta, ObservedStateLocation,
100 : ReconcileNeeded, ReconcileResult, ReconcileWaitError, ReconcilerStatus, ReconcilerWaiter,
101 : ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
102 : };
103 : use crate::timeline_import::{
104 : FinalizingImport, ImportResult, ShardImportStatuses, TimelineImport,
105 : TimelineImportFinalizeError, TimelineImportState, UpcallClient,
106 : };
107 :
108 : const WAITER_OPERATION_POLL_TIMEOUT: Duration = Duration::from_millis(500);
109 :
110 : // For operations that should be quick, like attaching a new tenant
111 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
112 :
113 : // For operations that might be slow, like migrating a tenant with
114 : // some data in it.
115 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
116 :
117 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
118 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
119 : // input generation from future requests as authoritative.
120 : const INITIAL_GENERATION: Generation = Generation::new(0);
121 :
122 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
123 : /// up on unresponsive pageservers and proceed.
124 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
125 :
126 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
127 : /// This must be long enough to cover node restarts as well as normal operations: in future
128 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
129 :
130 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
131 : /// offline.
132 : ///
133 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
134 : /// handling of the re-attach response may take a long time and blocks heartbeats from
135 : /// being handled on the pageserver side.
136 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
137 :
138 : /// How often to send heartbeats to registered nodes?
139 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
140 :
141 : /// How long is too long for a reconciliation?
142 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
143 :
144 : #[derive(Clone, strum_macros::Display)]
145 : enum TenantOperations {
146 : Create,
147 : LocationConfig,
148 : ConfigSet,
149 : ConfigPatch,
150 : TimeTravelRemoteStorage,
151 : Delete,
152 : UpdatePolicy,
153 : ShardSplit,
154 : SecondaryDownload,
155 : TimelineCreate,
156 : TimelineDelete,
157 : AttachHook,
158 : TimelineArchivalConfig,
159 : TimelineDetachAncestor,
160 : TimelineGcBlockUnblock,
161 : DropDetached,
162 : DownloadHeatmapLayers,
163 : TimelineLsnLease,
164 : TimelineSafekeeperMigrate,
165 : }
166 :
167 : #[derive(Clone, strum_macros::Display)]
168 : enum NodeOperations {
169 : Register,
170 : Configure,
171 : Delete,
172 : DeleteTombstone,
173 : }
174 :
175 : /// The leadership status for the storage controller process.
176 : /// Allowed transitions are:
177 : /// 1. Leader -> SteppedDown
178 : /// 2. Candidate -> Leader
179 : #[derive(
180 : Eq,
181 : PartialEq,
182 : Copy,
183 : Clone,
184 : strum_macros::Display,
185 : strum_macros::EnumIter,
186 : measured::FixedCardinalityLabel,
187 : )]
188 : #[strum(serialize_all = "snake_case")]
189 : pub(crate) enum LeadershipStatus {
190 : /// This is the steady state where the storage controller can produce
191 : /// side effects in the cluster.
192 : Leader,
193 : /// We've been notified to step down by another candidate. No reconciliations
194 : /// take place in this state.
195 : SteppedDown,
196 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
197 : #[allow(unused)]
198 : Candidate,
199 : }
200 :
201 : enum ShardGenerationValidity {
202 : Valid,
203 : Mismatched {
204 : claimed: Generation,
205 : actual: Option<Generation>,
206 : },
207 : }
208 :
209 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
210 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
211 : pub const SAFEKEEPER_RECONCILER_CONCURRENCY_DEFAULT: usize = 32;
212 :
213 : // Number of consecutive reconciliation errors, occured for one shard,
214 : // after which the shard is ignored when considering to run optimizations.
215 : const MAX_CONSECUTIVE_RECONCILIATION_ERRORS: usize = 5;
216 :
217 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
218 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
219 : // than they're being pushed onto the queue.
220 : const MAX_DELAYED_RECONCILES: usize = 10000;
221 :
222 : // Top level state available to all HTTP handlers
223 : struct ServiceState {
224 : leadership_status: LeadershipStatus,
225 :
226 : tenants: BTreeMap<TenantShardId, TenantShard>,
227 :
228 : nodes: Arc<HashMap<NodeId, Node>>,
229 :
230 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
231 :
232 : safekeeper_reconcilers: SafekeeperReconcilers,
233 :
234 : scheduler: Scheduler,
235 :
236 : /// Ongoing background operation on the cluster if any is running.
237 : /// Note that only one such operation may run at any given time,
238 : /// hence the type choice.
239 : ongoing_operation: Option<OperationHandler>,
240 :
241 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
242 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
243 :
244 : /// Tracks ongoing timeline import finalization tasks
245 : imports_finalizing: BTreeMap<(TenantId, TimelineId), FinalizingImport>,
246 : }
247 :
248 : /// Transform an error from a pageserver into an error to return to callers of a storage
249 : /// controller API.
250 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
251 0 : match e {
252 0 : mgmt_api::Error::SendRequest(e) => {
253 : // Presume errors sending requests are connectivity/availability issues
254 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
255 : }
256 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
257 : // Presume errors receiving body are connectivity/availability issues
258 0 : ApiError::ResourceUnavailable(
259 0 : format!("{node} error receiving error body: {str}").into(),
260 0 : )
261 : }
262 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
263 : // Return 500 for decoding errors.
264 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
265 : }
266 0 : mgmt_api::Error::ReceiveBody(err) => {
267 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
268 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
269 0 : ApiError::ResourceUnavailable(
270 0 : format!("{node} error receiving error body: {err} {src_str}").into(),
271 0 : )
272 : }
273 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
274 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
275 : }
276 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
277 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
278 : }
279 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
280 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
281 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
282 : // internal server errors, showing that something is wrong with the pageserver or
283 : // storage controller's auth configuration.
284 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
285 : }
286 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
287 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
288 : // turn ask our clients to wait + retry
289 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
290 : }
291 0 : mgmt_api::Error::ApiError(status, msg) => {
292 : // Presume general case of pageserver API errors is that we tried to do something
293 : // that can't be done right now.
294 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
295 : }
296 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
297 0 : mgmt_api::Error::Timeout(e) => ApiError::Timeout(e.into()),
298 : }
299 0 : }
300 :
301 : impl ServiceState {
302 0 : fn new(
303 0 : nodes: HashMap<NodeId, Node>,
304 0 : safekeepers: HashMap<NodeId, Safekeeper>,
305 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
306 0 : scheduler: Scheduler,
307 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
308 0 : initial_leadership_status: LeadershipStatus,
309 0 : reconcilers_cancel: CancellationToken,
310 0 : ) -> Self {
311 0 : metrics::update_leadership_status(initial_leadership_status);
312 :
313 0 : Self {
314 0 : leadership_status: initial_leadership_status,
315 0 : tenants,
316 0 : nodes: Arc::new(nodes),
317 0 : safekeepers: Arc::new(safekeepers),
318 0 : safekeeper_reconcilers: SafekeeperReconcilers::new(reconcilers_cancel),
319 0 : scheduler,
320 0 : ongoing_operation: None,
321 0 : delayed_reconcile_rx,
322 0 : imports_finalizing: Default::default(),
323 0 : }
324 0 : }
325 :
326 0 : fn parts_mut(
327 0 : &mut self,
328 0 : ) -> (
329 0 : &mut Arc<HashMap<NodeId, Node>>,
330 0 : &mut BTreeMap<TenantShardId, TenantShard>,
331 0 : &mut Scheduler,
332 0 : ) {
333 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
334 0 : }
335 :
336 : #[allow(clippy::type_complexity)]
337 0 : fn parts_mut_sk(
338 0 : &mut self,
339 0 : ) -> (
340 0 : &mut Arc<HashMap<NodeId, Node>>,
341 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
342 0 : &mut BTreeMap<TenantShardId, TenantShard>,
343 0 : &mut Scheduler,
344 0 : ) {
345 0 : (
346 0 : &mut self.nodes,
347 0 : &mut self.safekeepers,
348 0 : &mut self.tenants,
349 0 : &mut self.scheduler,
350 0 : )
351 0 : }
352 :
353 0 : fn get_leadership_status(&self) -> LeadershipStatus {
354 0 : self.leadership_status
355 0 : }
356 :
357 0 : fn step_down(&mut self) {
358 0 : self.leadership_status = LeadershipStatus::SteppedDown;
359 0 : metrics::update_leadership_status(self.leadership_status);
360 0 : }
361 :
362 0 : fn become_leader(&mut self) {
363 0 : self.leadership_status = LeadershipStatus::Leader;
364 0 : metrics::update_leadership_status(self.leadership_status);
365 0 : }
366 : }
367 :
368 : #[derive(Clone)]
369 : pub struct Config {
370 : // All pageservers managed by one instance of this service must have
371 : // the same public key. This JWT token will be used to authenticate
372 : // this service to the pageservers it manages.
373 : pub pageserver_jwt_token: Option<String>,
374 :
375 : // All safekeepers managed by one instance of this service must have
376 : // the same public key. This JWT token will be used to authenticate
377 : // this service to the safekeepers it manages.
378 : pub safekeeper_jwt_token: Option<String>,
379 :
380 : // This JWT token will be used to authenticate this service to the control plane.
381 : pub control_plane_jwt_token: Option<String>,
382 :
383 : // This JWT token will be used to authenticate with other storage controller instances
384 : pub peer_jwt_token: Option<String>,
385 :
386 : /// Prefix for storage API endpoints of the control plane. We use this prefix to compute
387 : /// URLs that we use to send pageserver and safekeeper attachment locations.
388 : /// If this is None, the compute hook will assume it is running in a test environment
389 : /// and try to invoke neon_local instead.
390 : pub control_plane_url: Option<String>,
391 :
392 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
393 : /// considered active. Once the grace period elapses, the next heartbeat failure will
394 : /// mark the pagseserver offline.
395 : pub max_offline_interval: Duration,
396 :
397 : /// Extended grace period within which pageserver may not respond to heartbeats.
398 : /// This extended grace period kicks in after the node has been drained for restart
399 : /// and/or upon handling the re-attach request from a node.
400 : pub max_warming_up_interval: Duration,
401 :
402 : /// How many normal-priority Reconcilers may be spawned concurrently
403 : pub reconciler_concurrency: usize,
404 :
405 : /// How many high-priority Reconcilers may be spawned concurrently
406 : pub priority_reconciler_concurrency: usize,
407 :
408 : /// How many safekeeper reconciles may happen concurrently (per safekeeper)
409 : pub safekeeper_reconciler_concurrency: usize,
410 :
411 : /// How many API requests per second to allow per tenant, across all
412 : /// tenant-scoped API endpoints. Further API requests queue until ready.
413 : pub tenant_rate_limit: NonZeroU32,
414 :
415 : /// If a tenant shard's largest timeline (max_logical_size) exceeds this value, all tenant
416 : /// shards will be split in 2 until they fall below split_threshold (up to max_split_shards).
417 : ///
418 : /// This will greedily split into as many shards as necessary to fall below split_threshold, as
419 : /// powers of 2: if a tenant shard is 7 times larger than split_threshold, it will split into 8
420 : /// immediately, rather than first 2 then 4 then 8.
421 : ///
422 : /// None or 0 disables auto-splitting.
423 : ///
424 : /// TODO: consider using total logical size of all timelines instead.
425 : pub split_threshold: Option<u64>,
426 :
427 : /// The maximum number of shards a tenant can be split into during autosplits. Does not affect
428 : /// manual split requests. 0 or 1 disables autosplits, as we already have 1 shard.
429 : pub max_split_shards: u8,
430 :
431 : /// The size at which an unsharded tenant should initially split. Ingestion is significantly
432 : /// faster with multiple shards, so eagerly splitting below split_threshold will typically speed
433 : /// up initial ingestion of large tenants.
434 : ///
435 : /// This should be below split_threshold, but it is not required. If both split_threshold and
436 : /// initial_split_threshold qualify, the largest number of target shards will be used.
437 : ///
438 : /// Does not apply to already sharded tenants: changing initial_split_threshold or
439 : /// initial_split_shards is not retroactive for already-sharded tenants.
440 : ///
441 : /// None or 0 disables initial splits.
442 : pub initial_split_threshold: Option<u64>,
443 :
444 : /// The number of shards to split into when reaching initial_split_threshold. Will
445 : /// be clamped to max_split_shards.
446 : ///
447 : /// 0 or 1 disables initial splits. Has no effect if initial_split_threshold is disabled.
448 : pub initial_split_shards: u8,
449 :
450 : // TODO: make this cfg(feature = "testing")
451 : pub neon_local_repo_dir: Option<PathBuf>,
452 :
453 : // Maximum acceptable download lag for the secondary location
454 : // while draining a node. If the secondary location is lagging
455 : // by more than the configured amount, then the secondary is not
456 : // upgraded to primary.
457 : pub max_secondary_lag_bytes: Option<u64>,
458 :
459 : pub heartbeat_interval: Duration,
460 :
461 : pub address_for_peers: Option<Uri>,
462 :
463 : pub start_as_candidate: bool,
464 :
465 : pub long_reconcile_threshold: Duration,
466 :
467 : pub use_https_pageserver_api: bool,
468 :
469 : pub use_https_safekeeper_api: bool,
470 :
471 : pub ssl_ca_certs: Vec<Certificate>,
472 :
473 : pub timelines_onto_safekeepers: bool,
474 :
475 : pub use_local_compute_notifications: bool,
476 :
477 : /// Number of safekeepers to choose for a timeline when creating it.
478 : /// Safekeepers will be choosen from different availability zones.
479 : pub timeline_safekeeper_count: usize,
480 :
481 : /// PostHog integration config
482 : pub posthog_config: Option<PostHogConfig>,
483 :
484 : /// When set, actively checks and initiates heatmap downloads/uploads.
485 : pub kick_secondary_downloads: bool,
486 : }
487 :
488 : impl From<DatabaseError> for ApiError {
489 0 : fn from(err: DatabaseError) -> ApiError {
490 0 : match err {
491 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
492 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
493 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
494 0 : ApiError::ShuttingDown
495 : }
496 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
497 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
498 : }
499 0 : DatabaseError::Cas(reason) => ApiError::Conflict(reason),
500 : }
501 0 : }
502 : }
503 :
504 : enum InitialShardScheduleOutcome {
505 : Scheduled(TenantCreateResponseShard),
506 : NotScheduled,
507 : ShardScheduleError(ScheduleError),
508 : }
509 :
510 : pub struct Service {
511 : inner: Arc<std::sync::RwLock<ServiceState>>,
512 : config: Config,
513 : persistence: Arc<Persistence>,
514 : compute_hook: Arc<ComputeHook>,
515 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
516 :
517 : heartbeater_ps: Heartbeater<Node, PageserverState>,
518 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
519 :
520 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
521 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
522 :
523 : // Locking on a tenant granularity (covers all shards in the tenant):
524 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
525 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
526 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
527 :
528 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
529 : // that transition it to/from Active.
530 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
531 :
532 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
533 : // and reconciliation on startup.
534 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
535 :
536 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
537 : // a human user might be waiting for.
538 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
539 :
540 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
541 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
542 : ///
543 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
544 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
545 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
546 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
547 :
548 : // Process shutdown will fire this token
549 : cancel: CancellationToken,
550 :
551 : // Child token of [`Service::cancel`] used by reconcilers
552 : reconcilers_cancel: CancellationToken,
553 :
554 : // Background tasks will hold this gate
555 : gate: Gate,
556 :
557 : // Reconcilers background tasks will hold this gate
558 : reconcilers_gate: Gate,
559 :
560 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
561 : /// passes, it isn't safe to do any actions that mutate tenants.
562 : pub(crate) startup_complete: Barrier,
563 :
564 : /// HTTP client with proper CA certs.
565 : http_client: reqwest::Client,
566 :
567 : /// Handle for the step down background task if one was ever requested
568 : step_down_barrier: OnceLock<tokio::sync::watch::Receiver<Option<GlobalObservedState>>>,
569 : }
570 :
571 : impl From<ReconcileWaitError> for ApiError {
572 0 : fn from(value: ReconcileWaitError) -> Self {
573 0 : match value {
574 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
575 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
576 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
577 : }
578 0 : }
579 : }
580 :
581 : impl From<OperationError> for ApiError {
582 0 : fn from(value: OperationError) -> Self {
583 0 : match value {
584 0 : OperationError::NodeStateChanged(err)
585 0 : | OperationError::FinalizeError(err)
586 0 : | OperationError::ImpossibleConstraint(err) => {
587 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
588 : }
589 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
590 : }
591 0 : }
592 : }
593 :
594 : #[allow(clippy::large_enum_variant)]
595 : enum TenantCreateOrUpdate {
596 : Create(TenantCreateRequest),
597 : Update(Vec<ShardUpdate>),
598 : }
599 :
600 : struct ShardSplitParams {
601 : old_shard_count: ShardCount,
602 : new_shard_count: ShardCount,
603 : new_stripe_size: Option<ShardStripeSize>,
604 : targets: Vec<ShardSplitTarget>,
605 : policy: PlacementPolicy,
606 : config: TenantConfig,
607 : shard_ident: ShardIdentity,
608 : preferred_az_id: Option<AvailabilityZone>,
609 : }
610 :
611 : // When preparing for a shard split, we may either choose to proceed with the split,
612 : // or find that the work is already done and return NoOp.
613 : enum ShardSplitAction {
614 : Split(Box<ShardSplitParams>),
615 : NoOp(TenantShardSplitResponse),
616 : }
617 :
618 : // A parent shard which will be split
619 : struct ShardSplitTarget {
620 : parent_id: TenantShardId,
621 : node: Node,
622 : child_ids: Vec<TenantShardId>,
623 : }
624 :
625 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
626 : /// might not be available. We therefore use a queue of abort operations processed in the background.
627 : struct TenantShardSplitAbort {
628 : tenant_id: TenantId,
629 : /// The target values from the request that failed
630 : new_shard_count: ShardCount,
631 : new_stripe_size: Option<ShardStripeSize>,
632 : /// Until this abort op is complete, no other operations may be done on the tenant
633 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
634 : /// The reconciler gate for the duration of the split operation, and any included abort.
635 : _gate: GateGuard,
636 : }
637 :
638 : #[derive(thiserror::Error, Debug)]
639 : enum TenantShardSplitAbortError {
640 : #[error(transparent)]
641 : Database(#[from] DatabaseError),
642 : #[error(transparent)]
643 : Remote(#[from] mgmt_api::Error),
644 : #[error("Unavailable")]
645 : Unavailable,
646 : }
647 :
648 : /// Inputs for computing a target shard count for a tenant.
649 : struct ShardSplitInputs {
650 : /// Current shard count.
651 : shard_count: ShardCount,
652 : /// Total size of largest timeline summed across all shards.
653 : max_logical_size: u64,
654 : /// Size-based split threshold. Zero if size-based splits are disabled.
655 : split_threshold: u64,
656 : /// Upper bound on target shards. 0 or 1 disables splits.
657 : max_split_shards: u8,
658 : /// Initial split threshold. Zero if initial splits are disabled.
659 : initial_split_threshold: u64,
660 : /// Number of shards for initial splits. 0 or 1 disables initial splits.
661 : initial_split_shards: u8,
662 : }
663 :
664 : struct ShardUpdate {
665 : tenant_shard_id: TenantShardId,
666 : placement_policy: PlacementPolicy,
667 : tenant_config: TenantConfig,
668 :
669 : /// If this is None, generation is not updated.
670 : generation: Option<Generation>,
671 :
672 : /// If this is None, scheduling policy is not updated.
673 : scheduling_policy: Option<ShardSchedulingPolicy>,
674 : }
675 :
676 : enum StopReconciliationsReason {
677 : ShuttingDown,
678 : SteppingDown,
679 : }
680 :
681 : impl std::fmt::Display for StopReconciliationsReason {
682 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
683 0 : let s = match self {
684 0 : Self::ShuttingDown => "Shutting down",
685 0 : Self::SteppingDown => "Stepping down",
686 : };
687 0 : write!(writer, "{s}")
688 0 : }
689 : }
690 :
691 : pub(crate) enum ReconcileResultRequest {
692 : ReconcileResult(ReconcileResult),
693 : Stop,
694 : }
695 :
696 : #[derive(Clone)]
697 : struct MutationLocation {
698 : node: Node,
699 : generation: Generation,
700 : }
701 :
702 : #[derive(Clone)]
703 : struct ShardMutationLocations {
704 : latest: MutationLocation,
705 : other: Vec<MutationLocation>,
706 : }
707 :
708 : #[derive(Default, Clone)]
709 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
710 :
711 : struct ReconcileAllResult {
712 : spawned_reconciles: usize,
713 : keep_failing_reconciles: usize,
714 : has_delayed_reconciles: bool,
715 : }
716 :
717 : impl ReconcileAllResult {
718 0 : fn new(
719 0 : spawned_reconciles: usize,
720 0 : keep_failing_reconciles: usize,
721 0 : has_delayed_reconciles: bool,
722 0 : ) -> Self {
723 0 : assert!(
724 0 : spawned_reconciles >= keep_failing_reconciles,
725 0 : "It is impossible to have more keep-failing reconciles than spawned reconciles"
726 : );
727 0 : Self {
728 0 : spawned_reconciles,
729 0 : keep_failing_reconciles,
730 0 : has_delayed_reconciles,
731 0 : }
732 0 : }
733 :
734 : /// We can run optimizations only if we don't have any delayed reconciles and
735 : /// all spawned reconciles are also keep-failing reconciles.
736 0 : fn can_run_optimizations(&self) -> bool {
737 0 : !self.has_delayed_reconciles && self.spawned_reconciles == self.keep_failing_reconciles
738 0 : }
739 : }
740 :
741 : impl Service {
742 0 : pub fn get_config(&self) -> &Config {
743 0 : &self.config
744 0 : }
745 :
746 0 : pub fn get_http_client(&self) -> &reqwest::Client {
747 0 : &self.http_client
748 0 : }
749 :
750 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
751 : /// view of the world, and determine which pageservers are responsive.
752 : #[instrument(skip_all)]
753 : async fn startup_reconcile(
754 : self: &Arc<Service>,
755 : current_leader: Option<ControllerPersistence>,
756 : leader_step_down_state: Option<GlobalObservedState>,
757 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
758 : Result<(), (TenantShardId, NotifyError)>,
759 : >,
760 : ) {
761 : // Startup reconciliation does I/O to other services: whether they
762 : // are responsive or not, we should aim to finish within our deadline, because:
763 : // - If we don't, a k8s readiness hook watching /ready will kill us.
764 : // - While we're waiting for startup reconciliation, we are not fully
765 : // available for end user operations like creating/deleting tenants and timelines.
766 : //
767 : // We set multiple deadlines to break up the time available between the phases of work: this is
768 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
769 : let start_at = Instant::now();
770 : let node_scan_deadline = start_at
771 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
772 : .expect("Reconcile timeout is a modest constant");
773 :
774 : let observed = if let Some(state) = leader_step_down_state {
775 : tracing::info!(
776 : "Using observed state received from leader at {}",
777 : current_leader.as_ref().unwrap().address
778 : );
779 :
780 : state
781 : } else {
782 : self.build_global_observed_state(node_scan_deadline).await
783 : };
784 :
785 : // Accumulate a list of any tenant locations that ought to be detached
786 : let mut cleanup = Vec::new();
787 :
788 : // Send initial heartbeat requests to all nodes loaded from the database
789 : let all_nodes = {
790 : let locked = self.inner.read().unwrap();
791 : locked.nodes.clone()
792 : };
793 : let (mut nodes_online, mut sks_online) =
794 : self.initial_heartbeat_round(all_nodes.keys()).await;
795 :
796 : // List of tenants for which we will attempt to notify compute of their location at startup
797 : let mut compute_notifications = Vec::new();
798 :
799 : // Populate intent and observed states for all tenants, based on reported state on pageservers
800 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
801 : let shard_count = {
802 : let mut locked = self.inner.write().unwrap();
803 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
804 :
805 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
806 : let mut new_nodes = (**nodes).clone();
807 : for (node_id, node) in new_nodes.iter_mut() {
808 : if let Some(utilization) = nodes_online.remove(node_id) {
809 : node.set_availability(NodeAvailability::Active(utilization));
810 : scheduler.node_upsert(node);
811 : }
812 : }
813 : *nodes = Arc::new(new_nodes);
814 :
815 : let mut new_sks = (**safekeepers).clone();
816 : for (node_id, node) in new_sks.iter_mut() {
817 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
818 : node.set_availability(SafekeeperState::Available {
819 : utilization,
820 : last_seen_at,
821 : });
822 : }
823 : }
824 : *safekeepers = Arc::new(new_sks);
825 :
826 : for (tenant_shard_id, observed_state) in observed.0 {
827 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
828 : for node_id in observed_state.locations.keys() {
829 : cleanup.push((tenant_shard_id, *node_id));
830 : }
831 :
832 : continue;
833 : };
834 :
835 : tenant_shard.observed = observed_state;
836 : }
837 :
838 : // Populate each tenant's intent state
839 : let mut schedule_context = ScheduleContext::default();
840 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
841 : if tenant_shard_id.shard_number == ShardNumber(0) {
842 : // Reset scheduling context each time we advance to the next Tenant
843 : schedule_context = ScheduleContext::default();
844 : }
845 :
846 : tenant_shard.intent_from_observed(scheduler);
847 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
848 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
849 : // not enough pageservers are available. The tenant may well still be available
850 : // to clients.
851 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
852 : } else {
853 : // If we're both intending and observed to be attached at a particular node, we will
854 : // emit a compute notification for this. In the case where our observed state does not
855 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
856 : if let Some(attached_at) = tenant_shard.stably_attached() {
857 : compute_notifications.push(compute_hook::ShardUpdate {
858 : tenant_shard_id: *tenant_shard_id,
859 : node_id: attached_at,
860 : stripe_size: tenant_shard.shard.stripe_size,
861 : preferred_az: tenant_shard
862 : .preferred_az()
863 0 : .map(|az| Cow::Owned(az.clone())),
864 : });
865 : }
866 : }
867 : }
868 :
869 : tenants.len()
870 : };
871 :
872 : // Before making any obeservable changes to the cluster, persist self
873 : // as leader in database and memory.
874 : let leadership = Leadership::new(
875 : self.persistence.clone(),
876 : self.config.clone(),
877 : self.cancel.child_token(),
878 : );
879 :
880 : if let Err(e) = leadership.become_leader(current_leader).await {
881 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
882 : std::process::exit(1);
883 : }
884 :
885 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
886 : let sk_schedule_requests =
887 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
888 : Ok(v) => v,
889 : Err(e) => {
890 : tracing::warn!(
891 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
892 : );
893 : // std::process::exit(1);
894 : Vec::new()
895 : }
896 : };
897 :
898 : {
899 : let mut locked = self.inner.write().unwrap();
900 : locked.become_leader();
901 :
902 : for (sk_id, _sk) in locked.safekeepers.clone().iter() {
903 : locked.safekeeper_reconcilers.start_reconciler(*sk_id, self);
904 : }
905 :
906 : locked
907 : .safekeeper_reconcilers
908 : .schedule_request_vec(sk_schedule_requests);
909 : }
910 :
911 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
912 : // generation_pageserver in the database.
913 :
914 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
915 : // will emit compute hook notifications when they reconcile.
916 : //
917 : // Ordering: our calls to notify_attach_background synchronously establish a relative order for these notifications vs. any later
918 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
919 : // calls will be correctly ordered wrt these.
920 : //
921 : // Concurrency: we call notify_attach_background for all tenants, which will create O(N) tokio tasks, but almost all of them
922 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
923 : // unit and start doing I/O.
924 : tracing::info!(
925 : "Sending {} compute notifications",
926 : compute_notifications.len()
927 : );
928 : self.compute_hook.notify_attach_background(
929 : compute_notifications,
930 : bg_compute_notify_result_tx.clone(),
931 : &self.cancel,
932 : );
933 :
934 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
935 : // which require it: under normal circumstances this should only include tenants that were in some
936 : // transient state before we restarted, or any tenants whose compute hooks failed above.
937 : tracing::info!("Checking for shards in need of reconciliation...");
938 : let reconcile_all_result = self.reconcile_all();
939 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
940 : // normal operations may proceed.
941 :
942 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
943 : // background because it does not need to complete in order to proceed with other work.
944 : if !cleanup.is_empty() {
945 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
946 : tokio::task::spawn({
947 : let cleanup_self = self.clone();
948 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
949 : });
950 : }
951 :
952 : // Reconcile the timeline imports:
953 : // 1. Mark each tenant shard of tenants with an importing timeline as importing.
954 : // 2. Finalize the completed imports in the background. This handles the case where
955 : // the previous storage controller instance shut down whilst finalizing imports.
956 : let imports = self.persistence.list_timeline_imports().await;
957 : match imports {
958 : Ok(mut imports) => {
959 : {
960 : let mut locked = self.inner.write().unwrap();
961 : for import in &imports {
962 : locked
963 : .tenants
964 : .range_mut(TenantShardId::tenant_range(import.tenant_id))
965 0 : .for_each(|(_id, shard)| {
966 0 : shard.importing = TimelineImportState::Importing
967 0 : });
968 : }
969 : }
970 :
971 0 : imports.retain(|import| import.is_complete());
972 : tokio::task::spawn({
973 : let finalize_imports_self = self.clone();
974 0 : async move {
975 0 : finalize_imports_self
976 0 : .finalize_timeline_imports(imports)
977 0 : .await
978 0 : }
979 : });
980 : }
981 : Err(err) => {
982 : tracing::error!("Could not retrieve completed imports from database: {err}");
983 : }
984 : }
985 :
986 : let spawned_reconciles = reconcile_all_result.spawned_reconciles;
987 : tracing::info!(
988 : "Startup complete, spawned {spawned_reconciles} reconciliation tasks ({shard_count} shards total)"
989 : );
990 : }
991 :
992 0 : async fn initial_heartbeat_round<'a>(
993 0 : &self,
994 0 : node_ids: impl Iterator<Item = &'a NodeId>,
995 0 : ) -> (
996 0 : HashMap<NodeId, PageserverUtilization>,
997 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
998 0 : ) {
999 0 : assert!(!self.startup_complete.is_ready());
1000 :
1001 0 : let all_nodes = {
1002 0 : let locked = self.inner.read().unwrap();
1003 0 : locked.nodes.clone()
1004 : };
1005 :
1006 0 : let mut nodes_to_heartbeat = HashMap::new();
1007 0 : for node_id in node_ids {
1008 0 : match all_nodes.get(node_id) {
1009 0 : Some(node) => {
1010 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
1011 0 : }
1012 : None => {
1013 0 : tracing::warn!("Node {node_id} was removed during start-up");
1014 : }
1015 : }
1016 : }
1017 :
1018 0 : let all_sks = {
1019 0 : let locked = self.inner.read().unwrap();
1020 0 : locked.safekeepers.clone()
1021 : };
1022 :
1023 0 : tracing::info!("Sending initial heartbeats...");
1024 0 : let (res_ps, res_sk) = tokio::join!(
1025 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
1026 0 : self.heartbeater_sk.heartbeat(all_sks)
1027 : );
1028 :
1029 0 : let mut online_nodes = HashMap::new();
1030 0 : if let Ok(deltas) = res_ps {
1031 0 : for (node_id, status) in deltas.0 {
1032 0 : match status {
1033 0 : PageserverState::Available { utilization, .. } => {
1034 0 : online_nodes.insert(node_id, utilization);
1035 0 : }
1036 0 : PageserverState::Offline => {}
1037 : PageserverState::WarmingUp { .. } => {
1038 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
1039 : }
1040 : }
1041 : }
1042 0 : }
1043 :
1044 0 : let mut online_sks = HashMap::new();
1045 0 : if let Ok(deltas) = res_sk {
1046 0 : for (node_id, status) in deltas.0 {
1047 0 : match status {
1048 : SafekeeperState::Available {
1049 0 : utilization,
1050 0 : last_seen_at,
1051 0 : } => {
1052 0 : online_sks.insert(node_id, (utilization, last_seen_at));
1053 0 : }
1054 0 : SafekeeperState::Offline => {}
1055 : }
1056 : }
1057 0 : }
1058 :
1059 0 : (online_nodes, online_sks)
1060 0 : }
1061 :
1062 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
1063 : ///
1064 : /// The result includes only nodes which responded within the deadline
1065 0 : async fn scan_node_locations(
1066 0 : &self,
1067 0 : deadline: Instant,
1068 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
1069 0 : let nodes = {
1070 0 : let locked = self.inner.read().unwrap();
1071 0 : locked.nodes.clone()
1072 : };
1073 :
1074 0 : let mut node_results = HashMap::new();
1075 :
1076 0 : let mut node_list_futs = FuturesUnordered::new();
1077 :
1078 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
1079 0 : for node in nodes.values() {
1080 0 : node_list_futs.push({
1081 0 : async move {
1082 0 : tracing::info!("Scanning shards on node {node}...");
1083 0 : let timeout = Duration::from_secs(5);
1084 0 : let response = node
1085 0 : .with_client_retries(
1086 0 : |client| async move { client.list_location_config().await },
1087 0 : &self.http_client,
1088 0 : &self.config.pageserver_jwt_token,
1089 : 1,
1090 : 5,
1091 0 : timeout,
1092 0 : &self.cancel,
1093 : )
1094 0 : .await;
1095 0 : (node.get_id(), response)
1096 0 : }
1097 : });
1098 : }
1099 :
1100 : loop {
1101 0 : let (node_id, result) = tokio::select! {
1102 0 : next = node_list_futs.next() => {
1103 0 : match next {
1104 0 : Some(result) => result,
1105 : None =>{
1106 : // We got results for all our nodes
1107 0 : break;
1108 : }
1109 :
1110 : }
1111 : },
1112 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
1113 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
1114 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
1115 0 : break;
1116 : }
1117 : };
1118 :
1119 0 : let Some(list_response) = result else {
1120 0 : tracing::info!("Shutdown during startup_reconcile");
1121 0 : break;
1122 : };
1123 :
1124 0 : match list_response {
1125 0 : Err(e) => {
1126 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
1127 : }
1128 0 : Ok(listing) => {
1129 0 : node_results.insert(node_id, listing);
1130 0 : }
1131 : }
1132 : }
1133 :
1134 0 : node_results
1135 0 : }
1136 :
1137 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
1138 0 : let node_listings = self.scan_node_locations(deadline).await;
1139 0 : let mut observed = GlobalObservedState::default();
1140 :
1141 0 : for (node_id, location_confs) in node_listings {
1142 0 : tracing::info!(
1143 0 : "Received {} shard statuses from pageserver {}",
1144 0 : location_confs.tenant_shards.len(),
1145 : node_id
1146 : );
1147 :
1148 0 : for (tid, location_conf) in location_confs.tenant_shards {
1149 0 : let entry = observed.0.entry(tid).or_default();
1150 0 : entry.locations.insert(
1151 0 : node_id,
1152 0 : ObservedStateLocation {
1153 0 : conf: location_conf,
1154 0 : },
1155 0 : );
1156 0 : }
1157 : }
1158 :
1159 0 : observed
1160 0 : }
1161 :
1162 : /// Used during [`Self::startup_reconcile`] and shard splits: detach a list of unknown-to-us
1163 : /// tenants from pageservers.
1164 : ///
1165 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1166 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1167 : /// other task trying to attach it.
1168 : #[instrument(skip_all)]
1169 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1170 : let nodes = self.inner.read().unwrap().nodes.clone();
1171 :
1172 : for (tenant_shard_id, node_id) in cleanup {
1173 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1174 : let Some(node) = nodes.get(&node_id) else {
1175 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1176 : // a location to clean up on a node that has since been removed.
1177 : tracing::info!(
1178 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1179 : );
1180 : continue;
1181 : };
1182 :
1183 : if self.cancel.is_cancelled() {
1184 : break;
1185 : }
1186 :
1187 : let client = PageserverClient::new(
1188 : node.get_id(),
1189 : self.http_client.clone(),
1190 : node.base_url(),
1191 : self.config.pageserver_jwt_token.as_deref(),
1192 : );
1193 : match client
1194 : .location_config(
1195 : tenant_shard_id,
1196 : LocationConfig {
1197 : mode: LocationConfigMode::Detached,
1198 : generation: None,
1199 : secondary_conf: None,
1200 : shard_number: tenant_shard_id.shard_number.0,
1201 : shard_count: tenant_shard_id.shard_count.literal(),
1202 : shard_stripe_size: 0,
1203 : tenant_conf: models::TenantConfig::default(),
1204 : },
1205 : None,
1206 : false,
1207 : )
1208 : .await
1209 : {
1210 : Ok(()) => {
1211 : tracing::info!(
1212 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1213 : );
1214 : }
1215 : Err(e) => {
1216 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1217 : // break anything.
1218 : tracing::error!(
1219 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1220 : );
1221 : }
1222 : }
1223 : }
1224 : }
1225 :
1226 : /// Long running background task that periodically wakes up and looks for shards that need
1227 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1228 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1229 : /// for those retries.
1230 : #[instrument(skip_all)]
1231 : async fn background_reconcile(self: &Arc<Self>) {
1232 : self.startup_complete.clone().wait().await;
1233 :
1234 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1235 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1236 : while !self.reconcilers_cancel.is_cancelled() {
1237 : tokio::select! {
1238 : _ = interval.tick() => {
1239 : let reconcile_all_result = self.reconcile_all();
1240 : if reconcile_all_result.can_run_optimizations() {
1241 : // Run optimizer only when we didn't find any other work to do
1242 : self.optimize_all().await;
1243 : }
1244 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1245 : // must be responsive when new projects begin ingesting and reach the threshold.
1246 : self.autosplit_tenants().await;
1247 : }
1248 : _ = self.reconcilers_cancel.cancelled() => return
1249 : }
1250 : }
1251 : }
1252 : /// Heartbeat all storage nodes once in a while.
1253 : #[instrument(skip_all)]
1254 : async fn spawn_heartbeat_driver(self: &Arc<Self>) {
1255 : self.startup_complete.clone().wait().await;
1256 :
1257 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1258 : while !self.cancel.is_cancelled() {
1259 : tokio::select! {
1260 : _ = interval.tick() => { }
1261 : _ = self.cancel.cancelled() => return
1262 : };
1263 :
1264 : let nodes = {
1265 : let locked = self.inner.read().unwrap();
1266 : locked.nodes.clone()
1267 : };
1268 :
1269 : let safekeepers = {
1270 : let locked = self.inner.read().unwrap();
1271 : locked.safekeepers.clone()
1272 : };
1273 :
1274 : let (res_ps, res_sk) = tokio::join!(
1275 : self.heartbeater_ps.heartbeat(nodes),
1276 : self.heartbeater_sk.heartbeat(safekeepers)
1277 : );
1278 :
1279 : if let Ok(deltas) = res_ps {
1280 : let mut to_handle = Vec::default();
1281 :
1282 : for (node_id, state) in deltas.0 {
1283 : let new_availability = match state {
1284 : PageserverState::Available { utilization, .. } => {
1285 : NodeAvailability::Active(utilization)
1286 : }
1287 : PageserverState::WarmingUp { started_at } => {
1288 : NodeAvailability::WarmingUp(started_at)
1289 : }
1290 : PageserverState::Offline => {
1291 : // The node might have been placed in the WarmingUp state
1292 : // while the heartbeat round was on-going. Hence, filter out
1293 : // offline transitions for WarmingUp nodes that are still within
1294 : // their grace period.
1295 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1296 : .get_node(node_id)
1297 : .await
1298 : .as_ref()
1299 0 : .map(|n| n.get_availability())
1300 : {
1301 : let now = Instant::now();
1302 : if now - *started_at >= self.config.max_warming_up_interval {
1303 : NodeAvailability::Offline
1304 : } else {
1305 : NodeAvailability::WarmingUp(*started_at)
1306 : }
1307 : } else {
1308 : NodeAvailability::Offline
1309 : }
1310 : }
1311 : };
1312 :
1313 : let node_lock = trace_exclusive_lock(
1314 : &self.node_op_locks,
1315 : node_id,
1316 : NodeOperations::Configure,
1317 : )
1318 : .await;
1319 :
1320 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1321 :
1322 : // This is the code path for geniune availability transitions (i.e node
1323 : // goes unavailable and/or comes back online).
1324 : let res = self
1325 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1326 : .await;
1327 :
1328 : match res {
1329 : Ok(transition) => {
1330 : // Keep hold of the lock until the availability transitions
1331 : // have been handled in
1332 : // [`Service::handle_node_availability_transitions`] in order avoid
1333 : // racing with [`Service::external_node_configure`].
1334 : to_handle.push((node_id, node_lock, transition));
1335 : }
1336 : Err(ApiError::NotFound(_)) => {
1337 : // This should be rare, but legitimate since the heartbeats are done
1338 : // on a snapshot of the nodes.
1339 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1340 : }
1341 : Err(ApiError::ShuttingDown) => {
1342 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1343 : }
1344 : Err(err) => {
1345 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1346 : // becomes unavailable again, we may get an error here.
1347 : tracing::error!(
1348 : "Failed to update node state {} after heartbeat round: {}",
1349 : node_id,
1350 : err
1351 : );
1352 : }
1353 : }
1354 : }
1355 :
1356 : // We collected all the transitions above and now we handle them.
1357 : let res = self.handle_node_availability_transitions(to_handle).await;
1358 : if let Err(errs) = res {
1359 : for (node_id, err) in errs {
1360 : match err {
1361 : ApiError::NotFound(_) => {
1362 : // This should be rare, but legitimate since the heartbeats are done
1363 : // on a snapshot of the nodes.
1364 : tracing::info!(
1365 : "Node {} was not found after heartbeat round",
1366 : node_id
1367 : );
1368 : }
1369 : err => {
1370 : tracing::error!(
1371 : "Failed to handle availability transition for {} after heartbeat round: {}",
1372 : node_id,
1373 : err
1374 : );
1375 : }
1376 : }
1377 : }
1378 : }
1379 : }
1380 : if let Ok(deltas) = res_sk {
1381 : let mut to_activate = Vec::new();
1382 : {
1383 : let mut locked = self.inner.write().unwrap();
1384 : let mut safekeepers = (*locked.safekeepers).clone();
1385 :
1386 : for (id, state) in deltas.0 {
1387 : let Some(sk) = safekeepers.get_mut(&id) else {
1388 : tracing::info!(
1389 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1390 : );
1391 : continue;
1392 : };
1393 : if sk.scheduling_policy() == SkSchedulingPolicy::Activating
1394 : && let SafekeeperState::Available { .. } = state
1395 : {
1396 : to_activate.push(id);
1397 : }
1398 : sk.set_availability(state);
1399 : }
1400 : locked.safekeepers = Arc::new(safekeepers);
1401 : }
1402 : for sk_id in to_activate {
1403 : // TODO this can race with set_scheduling_policy (can create disjoint DB <-> in-memory state)
1404 : tracing::info!("Activating safekeeper {sk_id}");
1405 : match self.persistence.activate_safekeeper(sk_id.0 as i64).await {
1406 : Ok(Some(())) => {}
1407 : Ok(None) => {
1408 : tracing::info!(
1409 : "safekeeper {sk_id} has been removed from db or has different scheduling policy than active or activating"
1410 : );
1411 : }
1412 : Err(e) => {
1413 : tracing::warn!("couldn't apply activation of {sk_id} to db: {e}");
1414 : continue;
1415 : }
1416 : }
1417 : if let Err(e) = self
1418 : .set_safekeeper_scheduling_policy_in_mem(sk_id, SkSchedulingPolicy::Active)
1419 : .await
1420 : {
1421 : tracing::info!("couldn't activate safekeeper {sk_id} in memory: {e}");
1422 : continue;
1423 : }
1424 : tracing::info!("Activation of safekeeper {sk_id} done");
1425 : }
1426 : }
1427 : }
1428 : }
1429 :
1430 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1431 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1432 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1433 : /// will indicate that reconciliation is not needed.
1434 : #[instrument(skip_all, fields(
1435 : seq=%result.sequence,
1436 : tenant_id=%result.tenant_shard_id.tenant_id,
1437 : shard_id=%result.tenant_shard_id.shard_slug(),
1438 : ))]
1439 : fn process_result(&self, result: ReconcileResult) {
1440 : let mut locked = self.inner.write().unwrap();
1441 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1442 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1443 : // A reconciliation result might race with removing a tenant: drop results for
1444 : // tenants that aren't in our map.
1445 : return;
1446 : };
1447 :
1448 : // Usually generation should only be updated via this path, so the max() isn't
1449 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1450 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1451 :
1452 : // If the reconciler signals that it failed to notify compute, set this state on
1453 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1454 : tenant.pending_compute_notification = result.pending_compute_notification;
1455 :
1456 : // Let the TenantShard know it is idle.
1457 : tenant.reconcile_complete(result.sequence);
1458 :
1459 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1460 : // make to the tenant
1461 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1462 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1463 : // make to the tenant
1464 0 : let node = nodes.get(delta.node_id())?;
1465 :
1466 0 : if node.is_available() {
1467 0 : return Some(delta);
1468 0 : }
1469 :
1470 : // In case a node became unavailable concurrently with the reconcile, observed
1471 : // locations on it are now uncertain. By convention, set them to None in order
1472 : // for them to get refreshed when the node comes back online.
1473 0 : Some(ObservedStateDelta::Upsert(Box::new((
1474 0 : node.get_id(),
1475 0 : ObservedStateLocation { conf: None },
1476 0 : ))))
1477 0 : });
1478 :
1479 : match result.result {
1480 : Ok(()) => {
1481 : tenant.consecutive_errors_count = 0;
1482 : tenant.apply_observed_deltas(deltas);
1483 : tenant.waiter.advance(result.sequence);
1484 : }
1485 : Err(e) => {
1486 : match e {
1487 : ReconcileError::Cancel => {
1488 : tracing::info!("Reconciler was cancelled");
1489 : }
1490 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1491 : // This might be due to the reconciler getting cancelled, or it might
1492 : // be due to the `Node` being marked offline.
1493 : tracing::info!("Reconciler cancelled during pageserver API call");
1494 : }
1495 : _ => {
1496 : tracing::warn!("Reconcile error: {}", e);
1497 : }
1498 : }
1499 :
1500 : tenant.consecutive_errors_count = tenant.consecutive_errors_count.saturating_add(1);
1501 :
1502 : // Ordering: populate last_error before advancing error_seq,
1503 : // so that waiters will see the correct error after waiting.
1504 : tenant.set_last_error(result.sequence, e);
1505 :
1506 : // Skip deletions on reconcile failures
1507 : let upsert_deltas =
1508 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1509 : tenant.apply_observed_deltas(upsert_deltas);
1510 : }
1511 : }
1512 :
1513 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1514 : if tenant.policy == PlacementPolicy::Detached {
1515 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1516 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1517 : // duration of the request.
1518 : let guard = self.tenant_op_locks.try_exclusive(
1519 : tenant.tenant_shard_id.tenant_id,
1520 : TenantOperations::DropDetached,
1521 : );
1522 : if let Some(guard) = guard {
1523 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1524 : }
1525 : }
1526 :
1527 : // Maybe some other work can proceed now that this job finished.
1528 : //
1529 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1530 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1531 : if self.reconciler_concurrency.available_permits() > 0 {
1532 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1533 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1534 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1535 : shard.delayed_reconcile = false;
1536 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1537 : }
1538 :
1539 : if self.reconciler_concurrency.available_permits() == 0 {
1540 : break;
1541 : }
1542 : }
1543 : }
1544 : }
1545 :
1546 0 : async fn process_results(
1547 0 : &self,
1548 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1549 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1550 0 : Result<(), (TenantShardId, NotifyError)>,
1551 0 : >,
1552 0 : ) {
1553 : loop {
1554 : // Wait for the next result, or for cancellation
1555 0 : tokio::select! {
1556 0 : r = result_rx.recv() => {
1557 0 : match r {
1558 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1559 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1560 : }
1561 : }
1562 0 : _ = async{
1563 0 : match bg_compute_hook_result_rx.recv().await {
1564 0 : Some(result) => {
1565 0 : if let Err((tenant_shard_id, notify_error)) = result {
1566 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1567 0 : let mut locked = self.inner.write().unwrap();
1568 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1569 0 : shard.pending_compute_notification = true;
1570 0 : }
1571 :
1572 0 : }
1573 : },
1574 : None => {
1575 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1576 0 : self.cancel.cancelled().await;
1577 : }
1578 : }
1579 0 : } => {},
1580 0 : _ = self.cancel.cancelled() => {
1581 0 : break;
1582 : }
1583 : };
1584 : }
1585 0 : }
1586 :
1587 0 : async fn process_aborts(
1588 0 : &self,
1589 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1590 0 : ) {
1591 : loop {
1592 : // Wait for the next result, or for cancellation
1593 0 : let op = tokio::select! {
1594 0 : r = abort_rx.recv() => {
1595 0 : match r {
1596 0 : Some(op) => {op},
1597 0 : None => {break;}
1598 : }
1599 : }
1600 0 : _ = self.cancel.cancelled() => {
1601 0 : break;
1602 : }
1603 : };
1604 :
1605 : // Retry until shutdown: we must keep this request object alive until it is properly
1606 : // processed, as it holds a lock guard that prevents other operations trying to do things
1607 : // to the tenant while it is in a weird part-split state.
1608 0 : while !self.reconcilers_cancel.is_cancelled() {
1609 0 : match self.abort_tenant_shard_split(&op).await {
1610 0 : Ok(_) => break,
1611 0 : Err(e) => {
1612 0 : tracing::warn!(
1613 0 : "Failed to abort shard split on {}, will retry: {e}",
1614 : op.tenant_id
1615 : );
1616 :
1617 : // If a node is unavailable, we hope that it has been properly marked Offline
1618 : // when we retry, so that the abort op will succeed. If the abort op is failing
1619 : // for some other reason, we will keep retrying forever, or until a human notices
1620 : // and does something about it (either fixing a pageserver or restarting the controller).
1621 0 : tokio::time::timeout(
1622 0 : Duration::from_secs(5),
1623 0 : self.reconcilers_cancel.cancelled(),
1624 0 : )
1625 0 : .await
1626 0 : .ok();
1627 : }
1628 : }
1629 : }
1630 : }
1631 0 : }
1632 :
1633 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1634 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1635 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1636 :
1637 0 : let leadership_cancel = CancellationToken::new();
1638 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1639 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1640 :
1641 : // Apply the migrations **after** the current leader has stepped down
1642 : // (or we've given up waiting for it), but **before** reading from the
1643 : // database. The only exception is reading the current leader before
1644 : // migrating.
1645 0 : persistence.migration_run().await?;
1646 :
1647 0 : tracing::info!("Loading nodes from database...");
1648 0 : let nodes = persistence
1649 0 : .list_nodes()
1650 0 : .await?
1651 0 : .into_iter()
1652 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1653 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1654 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1655 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1656 0 : metrics::METRICS_REGISTRY
1657 0 : .metrics_group
1658 0 : .storage_controller_pageserver_nodes
1659 0 : .set(nodes.len() as i64);
1660 0 : metrics::METRICS_REGISTRY
1661 0 : .metrics_group
1662 0 : .storage_controller_https_pageserver_nodes
1663 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
1664 :
1665 0 : tracing::info!("Loading safekeepers from database...");
1666 0 : let safekeepers = persistence
1667 0 : .list_safekeepers()
1668 0 : .await?
1669 0 : .into_iter()
1670 0 : .map(|skp| {
1671 0 : Safekeeper::from_persistence(
1672 0 : skp,
1673 0 : CancellationToken::new(),
1674 0 : config.use_https_safekeeper_api,
1675 : )
1676 0 : })
1677 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1678 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1679 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1680 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1681 0 : metrics::METRICS_REGISTRY
1682 0 : .metrics_group
1683 0 : .storage_controller_safekeeper_nodes
1684 0 : .set(safekeepers.len() as i64);
1685 0 : metrics::METRICS_REGISTRY
1686 0 : .metrics_group
1687 0 : .storage_controller_https_safekeeper_nodes
1688 0 : .set(safekeepers.values().filter(|s| s.has_https_port()).count() as i64);
1689 :
1690 0 : tracing::info!("Loading shards from database...");
1691 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1692 0 : tracing::info!(
1693 0 : "Loaded {} shards from database.",
1694 0 : tenant_shard_persistence.len()
1695 : );
1696 :
1697 : // If any shard splits were in progress, reset the database state to abort them
1698 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1699 0 : HashMap::new();
1700 0 : for tsp in &mut tenant_shard_persistence {
1701 0 : let shard = tsp.get_shard_identity()?;
1702 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1703 0 : let entry = tenant_shard_count_min_max
1704 0 : .entry(tenant_shard_id.tenant_id)
1705 0 : .or_insert_with(|| (shard.count, shard.count));
1706 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1707 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1708 : }
1709 :
1710 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1711 0 : if count_min != count_max {
1712 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1713 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1714 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1715 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1716 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1717 :
1718 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1719 : // identified this tenant has having mismatching min/max counts.
1720 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1721 :
1722 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1723 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1724 : // Set idle split state on those shards that we will retain.
1725 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1726 0 : if tsp_tenant_id == tenant_id
1727 0 : && tsp.get_shard_identity().unwrap().count == count_min
1728 0 : {
1729 0 : tsp.splitting = SplitState::Idle;
1730 0 : } else if tsp_tenant_id == tenant_id {
1731 : // Leave the splitting state on the child shards: this will be used next to
1732 : // drop them.
1733 0 : tracing::info!(
1734 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1735 : );
1736 0 : }
1737 0 : });
1738 :
1739 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1740 0 : tenant_shard_persistence.retain(|tsp| {
1741 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1742 0 : || tsp.splitting == SplitState::Idle
1743 0 : });
1744 0 : }
1745 : }
1746 :
1747 0 : let mut tenants = BTreeMap::new();
1748 :
1749 0 : let mut scheduler = Scheduler::new(nodes.values());
1750 :
1751 : #[cfg(feature = "testing")]
1752 : {
1753 : use pageserver_api::controller_api::AvailabilityZone;
1754 :
1755 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1756 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1757 : // after when pageservers start up and register.
1758 0 : let mut node_ids = HashSet::new();
1759 0 : for tsp in &tenant_shard_persistence {
1760 0 : if let Some(node_id) = tsp.generation_pageserver {
1761 0 : node_ids.insert(node_id);
1762 0 : }
1763 : }
1764 0 : for node_id in node_ids {
1765 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1766 0 : let node = Node::new(
1767 0 : NodeId(node_id as u64),
1768 0 : "".to_string(),
1769 : 123,
1770 0 : None,
1771 0 : "".to_string(),
1772 : 123,
1773 0 : None,
1774 0 : None,
1775 0 : AvailabilityZone("test_az".to_string()),
1776 : false,
1777 : )
1778 0 : .unwrap();
1779 :
1780 0 : scheduler.node_upsert(&node);
1781 : }
1782 : }
1783 0 : for tsp in tenant_shard_persistence {
1784 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1785 :
1786 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1787 : // it with what we can infer: the node for which a generation was most recently issued.
1788 0 : let mut intent = IntentState::new(
1789 0 : tsp.preferred_az_id
1790 0 : .as_ref()
1791 0 : .map(|az| AvailabilityZone(az.clone())),
1792 : );
1793 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1794 : {
1795 0 : if nodes.contains_key(&generation_pageserver) {
1796 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1797 0 : } else {
1798 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1799 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1800 : // on different pageservers.
1801 0 : tracing::warn!(
1802 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1803 : );
1804 : }
1805 0 : }
1806 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1807 :
1808 0 : tenants.insert(tenant_shard_id, new_tenant);
1809 : }
1810 :
1811 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1812 :
1813 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1814 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1815 0 : tokio::sync::mpsc::channel(512);
1816 :
1817 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1818 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1819 :
1820 0 : let cancel = CancellationToken::new();
1821 0 : let reconcilers_cancel = cancel.child_token();
1822 :
1823 0 : let mut http_client = reqwest::Client::builder();
1824 : // We intentionally disable the connection pool, so every request will create its own TCP connection.
1825 : // It's especially important for heartbeaters to notice more network problems.
1826 : //
1827 : // TODO: It makes sense to use this client only in heartbeaters and create a second one with
1828 : // connection pooling for everything else. But reqwest::Client may create a connection without
1829 : // ever using it (it uses hyper's Client under the hood):
1830 : // https://github.com/hyperium/hyper-util/blob/d51318df3461d40e5f5e5ca163cb3905ac960209/src/client/legacy/client.rs#L415
1831 : //
1832 : // Because of a bug in hyper0::Connection::graceful_shutdown such connections hang during
1833 : // graceful server shutdown: https://github.com/hyperium/hyper/issues/2730
1834 : //
1835 : // The bug has been fixed in hyper v1, so keep alive may be enabled only after we migrate to hyper1.
1836 0 : http_client = http_client.pool_max_idle_per_host(0);
1837 0 : for ssl_ca_cert in &config.ssl_ca_certs {
1838 0 : http_client = http_client.add_root_certificate(ssl_ca_cert.clone());
1839 0 : }
1840 0 : let http_client = http_client.build()?;
1841 :
1842 0 : let heartbeater_ps = Heartbeater::new(
1843 0 : http_client.clone(),
1844 0 : config.pageserver_jwt_token.clone(),
1845 0 : config.max_offline_interval,
1846 0 : config.max_warming_up_interval,
1847 0 : cancel.clone(),
1848 : );
1849 :
1850 0 : let heartbeater_sk = Heartbeater::new(
1851 0 : http_client.clone(),
1852 0 : config.safekeeper_jwt_token.clone(),
1853 0 : config.max_offline_interval,
1854 0 : config.max_warming_up_interval,
1855 0 : cancel.clone(),
1856 : );
1857 :
1858 0 : let initial_leadership_status = if config.start_as_candidate {
1859 0 : LeadershipStatus::Candidate
1860 : } else {
1861 0 : LeadershipStatus::Leader
1862 : };
1863 :
1864 0 : let this = Arc::new(Self {
1865 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1866 0 : nodes,
1867 0 : safekeepers,
1868 0 : tenants,
1869 0 : scheduler,
1870 0 : delayed_reconcile_rx,
1871 0 : initial_leadership_status,
1872 0 : reconcilers_cancel.clone(),
1873 : ))),
1874 0 : config: config.clone(),
1875 0 : persistence,
1876 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())?),
1877 0 : result_tx,
1878 0 : heartbeater_ps,
1879 0 : heartbeater_sk,
1880 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1881 0 : config.reconciler_concurrency,
1882 : )),
1883 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1884 0 : config.priority_reconciler_concurrency,
1885 : )),
1886 0 : delayed_reconcile_tx,
1887 0 : abort_tx,
1888 0 : startup_complete: startup_complete.clone(),
1889 0 : cancel,
1890 0 : reconcilers_cancel,
1891 0 : gate: Gate::default(),
1892 0 : reconcilers_gate: Gate::default(),
1893 0 : tenant_op_locks: Default::default(),
1894 0 : node_op_locks: Default::default(),
1895 0 : http_client,
1896 0 : step_down_barrier: Default::default(),
1897 : });
1898 :
1899 0 : let result_task_this = this.clone();
1900 0 : tokio::task::spawn(async move {
1901 : // Block shutdown until we're done (we must respect self.cancel)
1902 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1903 0 : result_task_this
1904 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1905 0 : .await
1906 0 : }
1907 0 : });
1908 :
1909 0 : tokio::task::spawn({
1910 0 : let this = this.clone();
1911 0 : async move {
1912 : // Block shutdown until we're done (we must respect self.cancel)
1913 0 : if let Ok(_gate) = this.gate.enter() {
1914 0 : this.process_aborts(abort_rx).await
1915 0 : }
1916 0 : }
1917 : });
1918 :
1919 0 : tokio::task::spawn({
1920 0 : let this = this.clone();
1921 0 : async move {
1922 0 : if let Ok(_gate) = this.gate.enter() {
1923 : loop {
1924 0 : tokio::select! {
1925 0 : _ = this.cancel.cancelled() => {
1926 0 : break;
1927 : },
1928 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1929 : };
1930 0 : this.tenant_op_locks.housekeeping();
1931 : }
1932 0 : }
1933 0 : }
1934 : });
1935 :
1936 0 : tokio::task::spawn({
1937 0 : let this = this.clone();
1938 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1939 : // is done.
1940 0 : let startup_completion = startup_completion.clone();
1941 0 : async move {
1942 : // Block shutdown until we're done (we must respect self.cancel)
1943 0 : let Ok(_gate) = this.gate.enter() else {
1944 0 : return;
1945 : };
1946 :
1947 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1948 0 : .await;
1949 :
1950 0 : drop(startup_completion);
1951 0 : }
1952 : });
1953 :
1954 0 : tokio::task::spawn({
1955 0 : let this = this.clone();
1956 0 : let startup_complete = startup_complete.clone();
1957 0 : async move {
1958 0 : startup_complete.wait().await;
1959 0 : this.background_reconcile().await;
1960 0 : }
1961 : });
1962 :
1963 0 : tokio::task::spawn({
1964 0 : let this = this.clone();
1965 0 : let startup_complete = startup_complete.clone();
1966 0 : async move {
1967 0 : startup_complete.wait().await;
1968 0 : this.spawn_heartbeat_driver().await;
1969 0 : }
1970 : });
1971 :
1972 0 : Ok(this)
1973 0 : }
1974 :
1975 0 : pub(crate) async fn attach_hook(
1976 0 : &self,
1977 0 : attach_req: AttachHookRequest,
1978 0 : ) -> anyhow::Result<AttachHookResponse> {
1979 0 : let _tenant_lock = trace_exclusive_lock(
1980 0 : &self.tenant_op_locks,
1981 0 : attach_req.tenant_shard_id.tenant_id,
1982 0 : TenantOperations::AttachHook,
1983 0 : )
1984 0 : .await;
1985 :
1986 : // This is a test hook. To enable using it on tenants that were created directly with
1987 : // the pageserver API (not via this service), we will auto-create any missing tenant
1988 : // shards with default state.
1989 0 : let insert = {
1990 0 : match self
1991 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1992 0 : .await
1993 : {
1994 0 : Ok(_) => false,
1995 0 : Err(ApiError::NotFound(_)) => true,
1996 0 : Err(e) => return Err(e.into()),
1997 : }
1998 : };
1999 :
2000 0 : if insert {
2001 0 : let config = attach_req.config.clone().unwrap_or_default();
2002 0 : let tsp = TenantShardPersistence {
2003 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
2004 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
2005 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
2006 0 : shard_stripe_size: 0,
2007 0 : generation: attach_req.generation_override.or(Some(0)),
2008 0 : generation_pageserver: None,
2009 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
2010 0 : config: serde_json::to_string(&config).unwrap(),
2011 0 : splitting: SplitState::default(),
2012 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2013 0 : .unwrap(),
2014 0 : preferred_az_id: None,
2015 0 : };
2016 :
2017 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
2018 0 : Err(e) => match e {
2019 : DatabaseError::Query(diesel::result::Error::DatabaseError(
2020 : DatabaseErrorKind::UniqueViolation,
2021 : _,
2022 : )) => {
2023 0 : tracing::info!(
2024 0 : "Raced with another request to insert tenant {}",
2025 : attach_req.tenant_shard_id
2026 : )
2027 : }
2028 0 : _ => return Err(e.into()),
2029 : },
2030 : Ok(()) => {
2031 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
2032 :
2033 0 : let mut shard = TenantShard::new(
2034 0 : attach_req.tenant_shard_id,
2035 0 : ShardIdentity::unsharded(),
2036 0 : PlacementPolicy::Attached(0),
2037 0 : None,
2038 : );
2039 0 : shard.config = config;
2040 :
2041 0 : let mut locked = self.inner.write().unwrap();
2042 0 : locked.tenants.insert(attach_req.tenant_shard_id, shard);
2043 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
2044 : }
2045 : }
2046 0 : }
2047 :
2048 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
2049 0 : let maybe_tenant_conf = {
2050 0 : let locked = self.inner.write().unwrap();
2051 0 : locked
2052 0 : .tenants
2053 0 : .get(&attach_req.tenant_shard_id)
2054 0 : .map(|t| t.config.clone())
2055 : };
2056 :
2057 0 : match maybe_tenant_conf {
2058 0 : Some(conf) => {
2059 0 : let new_generation = self
2060 0 : .persistence
2061 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
2062 0 : .await?;
2063 :
2064 : // Persist the placement policy update. This is required
2065 : // when we reattaching a detached tenant.
2066 0 : self.persistence
2067 0 : .update_tenant_shard(
2068 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
2069 0 : Some(PlacementPolicy::Attached(0)),
2070 0 : Some(conf),
2071 0 : None,
2072 0 : None,
2073 0 : )
2074 0 : .await?;
2075 0 : Some(new_generation)
2076 : }
2077 : None => {
2078 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
2079 : }
2080 : }
2081 : } else {
2082 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
2083 0 : None
2084 : };
2085 :
2086 0 : let mut locked = self.inner.write().unwrap();
2087 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2088 :
2089 0 : let tenant_shard = tenants
2090 0 : .get_mut(&attach_req.tenant_shard_id)
2091 0 : .expect("Checked for existence above");
2092 :
2093 0 : if let Some(new_generation) = new_generation {
2094 0 : tenant_shard.generation = Some(new_generation);
2095 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
2096 0 : } else {
2097 : // This is a detach notification. We must update placement policy to avoid re-attaching
2098 : // during background scheduling/reconciliation, or during storage controller restart.
2099 0 : assert!(attach_req.node_id.is_none());
2100 0 : tenant_shard.policy = PlacementPolicy::Detached;
2101 : }
2102 :
2103 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
2104 0 : tracing::info!(
2105 : tenant_id = %attach_req.tenant_shard_id,
2106 : ps_id = %attaching_pageserver,
2107 : generation = ?tenant_shard.generation,
2108 0 : "issuing",
2109 : );
2110 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
2111 0 : tracing::info!(
2112 : tenant_id = %attach_req.tenant_shard_id,
2113 : %ps_id,
2114 : generation = ?tenant_shard.generation,
2115 0 : "dropping",
2116 : );
2117 : } else {
2118 0 : tracing::info!(
2119 : tenant_id = %attach_req.tenant_shard_id,
2120 0 : "no-op: tenant already has no pageserver");
2121 : }
2122 0 : tenant_shard
2123 0 : .intent
2124 0 : .set_attached(scheduler, attach_req.node_id);
2125 :
2126 0 : tracing::info!(
2127 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}, config {:?}",
2128 : attach_req.tenant_shard_id,
2129 : tenant_shard.generation,
2130 : // TODO: this is an odd number of 0xf's
2131 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff)),
2132 : attach_req.config,
2133 : );
2134 :
2135 : // Trick the reconciler into not doing anything for this tenant: this helps
2136 : // tests that manually configure a tenant on the pagesrever, and then call this
2137 : // attach hook: they don't want background reconciliation to modify what they
2138 : // did to the pageserver.
2139 : #[cfg(feature = "testing")]
2140 : {
2141 0 : if let Some(node_id) = attach_req.node_id {
2142 0 : tenant_shard.observed.locations = HashMap::from([(
2143 0 : node_id,
2144 0 : ObservedStateLocation {
2145 0 : conf: Some(attached_location_conf(
2146 0 : tenant_shard.generation.unwrap(),
2147 0 : &tenant_shard.shard,
2148 0 : &tenant_shard.config,
2149 0 : &PlacementPolicy::Attached(0),
2150 0 : tenant_shard.intent.get_secondary().len(),
2151 0 : )),
2152 0 : },
2153 0 : )]);
2154 0 : } else {
2155 0 : tenant_shard.observed.locations.clear();
2156 0 : }
2157 : }
2158 :
2159 : Ok(AttachHookResponse {
2160 0 : generation: attach_req
2161 0 : .node_id
2162 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
2163 : })
2164 0 : }
2165 :
2166 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
2167 0 : let locked = self.inner.read().unwrap();
2168 :
2169 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
2170 :
2171 : InspectResponse {
2172 0 : attachment: tenant_shard.and_then(|s| {
2173 0 : s.intent
2174 0 : .get_attached()
2175 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
2176 0 : }),
2177 : }
2178 0 : }
2179 :
2180 : // When the availability state of a node transitions to active, we must do a full reconciliation
2181 : // of LocationConfigs on that node. This is because while a node was offline:
2182 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
2183 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
2184 : //
2185 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
2186 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
2187 : // this function.
2188 : //
2189 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
2190 : // for written for a single node rather than as a batch job for all nodes.
2191 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
2192 : async fn node_activate_reconcile(
2193 : &self,
2194 : mut node: Node,
2195 : _lock: &TracingExclusiveGuard<NodeOperations>,
2196 : ) -> Result<(), ApiError> {
2197 : // This Node is a mutable local copy: we will set it active so that we can use its
2198 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
2199 : // later.
2200 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
2201 :
2202 : let configs = match node
2203 : .with_client_retries(
2204 0 : |client| async move { client.list_location_config().await },
2205 : &self.http_client,
2206 : &self.config.pageserver_jwt_token,
2207 : 1,
2208 : 5,
2209 : SHORT_RECONCILE_TIMEOUT,
2210 : &self.cancel,
2211 : )
2212 : .await
2213 : {
2214 : None => {
2215 : // We're shutting down (the Node's cancellation token can't have fired, because
2216 : // we're the only scope that has a reference to it, and we didn't fire it).
2217 : return Err(ApiError::ShuttingDown);
2218 : }
2219 : Some(Err(e)) => {
2220 : // This node didn't succeed listing its locations: it may not proceed to active state
2221 : // as it is apparently unavailable.
2222 : return Err(ApiError::PreconditionFailed(
2223 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2224 : ));
2225 : }
2226 : Some(Ok(configs)) => configs,
2227 : };
2228 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2229 :
2230 : let mut cleanup = Vec::new();
2231 : let mut mismatched_locations = 0;
2232 : {
2233 : let mut locked = self.inner.write().unwrap();
2234 :
2235 : for (tenant_shard_id, reported) in configs.tenant_shards {
2236 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2237 : cleanup.push(tenant_shard_id);
2238 : continue;
2239 : };
2240 :
2241 : let on_record = &mut tenant_shard
2242 : .observed
2243 : .locations
2244 : .entry(node.get_id())
2245 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2246 : .conf;
2247 :
2248 : // If the location reported by the node does not match our observed state,
2249 : // then we mark it as uncertain and let the background reconciliation loop
2250 : // deal with it.
2251 : //
2252 : // Note that this also covers net new locations reported by the node.
2253 : if *on_record != reported {
2254 : mismatched_locations += 1;
2255 : *on_record = None;
2256 : }
2257 : }
2258 : }
2259 :
2260 : if mismatched_locations > 0 {
2261 : tracing::info!(
2262 : "Set observed state to None for {mismatched_locations} mismatched locations"
2263 : );
2264 : }
2265 :
2266 : for tenant_shard_id in cleanup {
2267 : tracing::info!("Detaching {tenant_shard_id}");
2268 : match node
2269 : .with_client_retries(
2270 0 : |client| async move {
2271 0 : let config = LocationConfig {
2272 0 : mode: LocationConfigMode::Detached,
2273 0 : generation: None,
2274 0 : secondary_conf: None,
2275 0 : shard_number: tenant_shard_id.shard_number.0,
2276 0 : shard_count: tenant_shard_id.shard_count.literal(),
2277 0 : shard_stripe_size: 0,
2278 0 : tenant_conf: models::TenantConfig::default(),
2279 0 : };
2280 0 : client
2281 0 : .location_config(tenant_shard_id, config, None, false)
2282 0 : .await
2283 0 : },
2284 : &self.http_client,
2285 : &self.config.pageserver_jwt_token,
2286 : 1,
2287 : 5,
2288 : SHORT_RECONCILE_TIMEOUT,
2289 : &self.cancel,
2290 : )
2291 : .await
2292 : {
2293 : None => {
2294 : // We're shutting down (the Node's cancellation token can't have fired, because
2295 : // we're the only scope that has a reference to it, and we didn't fire it).
2296 : return Err(ApiError::ShuttingDown);
2297 : }
2298 : Some(Err(e)) => {
2299 : // Do not let the node proceed to Active state if it is not responsive to requests
2300 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2301 : // detach completing: we should not let this node back into the set of nodes considered
2302 : // okay for scheduling.
2303 : return Err(ApiError::Conflict(format!(
2304 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2305 : )));
2306 : }
2307 : Some(Ok(_)) => {}
2308 : };
2309 : }
2310 :
2311 : Ok(())
2312 : }
2313 :
2314 0 : pub(crate) async fn re_attach(
2315 0 : &self,
2316 0 : reattach_req: ReAttachRequest,
2317 0 : ) -> Result<ReAttachResponse, ApiError> {
2318 0 : if let Some(register_req) = reattach_req.register {
2319 0 : self.node_register(register_req).await?;
2320 0 : }
2321 :
2322 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2323 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2324 :
2325 0 : tracing::info!(
2326 : node_id=%reattach_req.node_id,
2327 0 : "Incremented {} tenant shards' generations",
2328 0 : incremented_generations.len()
2329 : );
2330 :
2331 : // Apply the updated generation to our in-memory state, and
2332 : // gather discover secondary locations.
2333 0 : let mut locked = self.inner.write().unwrap();
2334 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2335 :
2336 0 : let mut response = ReAttachResponse {
2337 0 : tenants: Vec::new(),
2338 0 : };
2339 :
2340 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2341 : // to call location_conf API with an old generation. Wait for cancellation to complete
2342 : // before responding to this request. Requires well implemented CancellationToken logic
2343 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2344 : // request in flight over the network: TODO handle that by making location_conf API refuse
2345 : // to go backward in generations.
2346 :
2347 : // Scan through all shards, applying updates for ones where we updated generation
2348 : // and identifying shards that intend to have a secondary location on this node.
2349 0 : for (tenant_shard_id, shard) in tenants {
2350 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2351 0 : let new_gen = *new_gen;
2352 0 : response.tenants.push(ReAttachResponseTenant {
2353 0 : id: *tenant_shard_id,
2354 0 : r#gen: Some(new_gen.into().unwrap()),
2355 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2356 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2357 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2358 0 : // the stale/multi states at this point.
2359 0 : mode: LocationConfigMode::AttachedSingle,
2360 0 : stripe_size: shard.shard.stripe_size,
2361 0 : });
2362 :
2363 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2364 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2365 : // Why can we update `observed` even though we're not sure our response will be received
2366 : // by the pageserver? Because the pageserver will not proceed with startup until
2367 : // it has processed response: if it loses it, we'll see another request and increment
2368 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2369 0 : if let Some(conf) = observed.conf.as_mut() {
2370 0 : conf.generation = new_gen.into();
2371 0 : }
2372 0 : } else {
2373 0 : // This node has no observed state for the shard: perhaps it was offline
2374 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2375 0 : // will be prompted to learn the location's state before it makes changes.
2376 0 : shard
2377 0 : .observed
2378 0 : .locations
2379 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2380 0 : }
2381 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2382 0 : // Ordering: pageserver will not accept /location_config requests until it has
2383 0 : // finished processing the response from re-attach. So we can update our in-memory state
2384 0 : // now, and be confident that we are not stamping on the result of some later location config.
2385 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2386 0 : // so we might update observed state here, and then get over-written by some racing
2387 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2388 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2389 0 :
2390 0 : response.tenants.push(ReAttachResponseTenant {
2391 0 : id: *tenant_shard_id,
2392 0 : r#gen: None,
2393 0 : mode: LocationConfigMode::Secondary,
2394 0 : stripe_size: shard.shard.stripe_size,
2395 0 : });
2396 0 :
2397 0 : // We must not update observed, because we have no guarantee that our
2398 0 : // response will be received by the pageserver. This could leave it
2399 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2400 0 : }
2401 : }
2402 :
2403 : // We consider a node Active once we have composed a re-attach response, but we
2404 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2405 : // implicitly synchronizes the LocationConfigs on the node.
2406 : //
2407 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2408 : // but those requests will not be accepted by the node until it has finished processing
2409 : // the re-attach response.
2410 : //
2411 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2412 : // in [`Persistence::re_attach`].
2413 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2414 0 : let reset_scheduling = matches!(
2415 0 : node.get_scheduling(),
2416 : NodeSchedulingPolicy::PauseForRestart
2417 : | NodeSchedulingPolicy::Draining
2418 : | NodeSchedulingPolicy::Filling
2419 : | NodeSchedulingPolicy::Deleting
2420 : );
2421 :
2422 0 : let mut new_nodes = (**nodes).clone();
2423 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2424 0 : if reset_scheduling {
2425 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2426 0 : }
2427 :
2428 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2429 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2430 :
2431 0 : scheduler.node_upsert(node);
2432 0 : let new_nodes = Arc::new(new_nodes);
2433 0 : *nodes = new_nodes;
2434 : } else {
2435 0 : tracing::error!(
2436 0 : "Reattaching node {} was removed while processing the request",
2437 : reattach_req.node_id
2438 : );
2439 : }
2440 0 : }
2441 :
2442 0 : Ok(response)
2443 0 : }
2444 :
2445 0 : pub(crate) async fn validate(
2446 0 : &self,
2447 0 : validate_req: ValidateRequest,
2448 0 : ) -> Result<ValidateResponse, DatabaseError> {
2449 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2450 : // in-memory generation for a shard
2451 0 : let in_memory_result = {
2452 0 : let mut in_memory_result = Vec::new();
2453 0 : let locked = self.inner.read().unwrap();
2454 0 : for req_tenant in validate_req.tenants {
2455 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2456 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2457 0 : tracing::info!(
2458 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2459 : req_tenant.id,
2460 : req_tenant.r#gen,
2461 : tenant_shard.generation
2462 : );
2463 :
2464 0 : in_memory_result.push((
2465 0 : req_tenant.id,
2466 0 : Generation::new(req_tenant.r#gen),
2467 0 : valid,
2468 0 : ));
2469 : } else {
2470 : // This is legal: for example during a shard split the pageserver may still
2471 : // have deletions in its queue from the old pre-split shard, or after deletion
2472 : // of a tenant that was busy with compaction/gc while being deleted.
2473 0 : tracing::info!(
2474 0 : "Refusing deletion validation for missing shard {}",
2475 : req_tenant.id
2476 : );
2477 : }
2478 : }
2479 :
2480 0 : in_memory_result
2481 : };
2482 :
2483 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2484 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2485 0 : let db_generations = self
2486 0 : .persistence
2487 0 : .shard_generations(
2488 0 : in_memory_result
2489 0 : .iter()
2490 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2491 : )
2492 0 : .await?;
2493 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2494 :
2495 0 : let mut response = ValidateResponse {
2496 0 : tenants: Vec::new(),
2497 0 : };
2498 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2499 0 : let valid = if valid {
2500 0 : let db_generation = db_generations.get(&tenant_shard_id);
2501 0 : db_generation == Some(&Some(validate_generation))
2502 : } else {
2503 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2504 : // this prevents a pageserver from cleaning up an object in S3.
2505 0 : false
2506 : };
2507 :
2508 0 : response.tenants.push(ValidateResponseTenant {
2509 0 : id: tenant_shard_id,
2510 0 : valid,
2511 0 : })
2512 : }
2513 :
2514 0 : Ok(response)
2515 0 : }
2516 :
2517 0 : pub(crate) async fn tenant_create(
2518 0 : &self,
2519 0 : create_req: TenantCreateRequest,
2520 0 : ) -> Result<TenantCreateResponse, ApiError> {
2521 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2522 :
2523 : // Exclude any concurrent attempts to create/access the same tenant ID
2524 0 : let _tenant_lock = trace_exclusive_lock(
2525 0 : &self.tenant_op_locks,
2526 0 : create_req.new_tenant_id.tenant_id,
2527 0 : TenantOperations::Create,
2528 0 : )
2529 0 : .await;
2530 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2531 :
2532 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2533 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2534 : // accept compute notifications while it is in the process of creating. Reconciliation will
2535 : // be retried in the background.
2536 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2537 0 : }
2538 0 : Ok(response)
2539 0 : }
2540 :
2541 0 : pub(crate) async fn do_tenant_create(
2542 0 : &self,
2543 0 : create_req: TenantCreateRequest,
2544 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2545 0 : let placement_policy = create_req
2546 0 : .placement_policy
2547 0 : .clone()
2548 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2549 0 : .unwrap_or(PlacementPolicy::Attached(0));
2550 :
2551 : // This service expects to handle sharding itself: it is an error to try and directly create
2552 : // a particular shard here.
2553 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2554 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2555 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2556 0 : )));
2557 : } else {
2558 0 : create_req.new_tenant_id.tenant_id
2559 : };
2560 :
2561 0 : tracing::info!(
2562 0 : "Creating tenant {}, shard_count={:?}",
2563 : create_req.new_tenant_id,
2564 : create_req.shard_parameters.count,
2565 : );
2566 :
2567 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2568 0 : .map(|i| TenantShardId {
2569 0 : tenant_id,
2570 0 : shard_number: ShardNumber(i),
2571 0 : shard_count: create_req.shard_parameters.count,
2572 0 : })
2573 0 : .collect::<Vec<_>>();
2574 :
2575 : // If the caller specifies a None generation, it means "start from default". This is different
2576 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2577 : // an incompletely-onboarded tenant.
2578 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2579 0 : tracing::info!(
2580 0 : "tenant_create: secondary mode, generation is_some={}",
2581 0 : create_req.generation.is_some()
2582 : );
2583 0 : create_req.generation.map(Generation::new)
2584 : } else {
2585 0 : tracing::info!(
2586 0 : "tenant_create: not secondary mode, generation is_some={}",
2587 0 : create_req.generation.is_some()
2588 : );
2589 0 : Some(
2590 0 : create_req
2591 0 : .generation
2592 0 : .map(Generation::new)
2593 0 : .unwrap_or(INITIAL_GENERATION),
2594 0 : )
2595 : };
2596 :
2597 0 : let preferred_az_id = {
2598 0 : let locked = self.inner.read().unwrap();
2599 : // Idempotency: take the existing value if the tenant already exists
2600 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2601 0 : shard.preferred_az().cloned()
2602 : } else {
2603 0 : locked.scheduler.get_az_for_new_tenant()
2604 : }
2605 : };
2606 :
2607 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2608 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2609 : // during the creation, rather than risking leaving orphan objects in S3.
2610 0 : let persist_tenant_shards = create_ids
2611 0 : .iter()
2612 0 : .map(|tenant_shard_id| TenantShardPersistence {
2613 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2614 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2615 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2616 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2617 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2618 : // The pageserver is not known until scheduling happens: we will set this column when
2619 : // incrementing the generation the first time we attach to a pageserver.
2620 0 : generation_pageserver: None,
2621 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2622 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2623 0 : splitting: SplitState::default(),
2624 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2625 0 : .unwrap(),
2626 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2627 0 : })
2628 0 : .collect();
2629 :
2630 0 : match self
2631 0 : .persistence
2632 0 : .insert_tenant_shards(persist_tenant_shards)
2633 0 : .await
2634 : {
2635 0 : Ok(_) => {}
2636 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2637 : DatabaseErrorKind::UniqueViolation,
2638 : _,
2639 : ))) => {
2640 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2641 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2642 : // creation's shard count.
2643 0 : tracing::info!(
2644 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2645 : );
2646 : }
2647 : // Any other database error is unexpected and a bug.
2648 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2649 : };
2650 :
2651 0 : let mut schedule_context = ScheduleContext::default();
2652 0 : let mut schedule_error = None;
2653 0 : let mut response_shards = Vec::new();
2654 0 : for tenant_shard_id in create_ids {
2655 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2656 :
2657 0 : let outcome = self
2658 0 : .do_initial_shard_scheduling(
2659 0 : tenant_shard_id,
2660 0 : initial_generation,
2661 0 : create_req.shard_parameters,
2662 0 : create_req.config.clone(),
2663 0 : placement_policy.clone(),
2664 0 : preferred_az_id.as_ref(),
2665 0 : &mut schedule_context,
2666 0 : )
2667 0 : .await;
2668 :
2669 0 : match outcome {
2670 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2671 0 : InitialShardScheduleOutcome::NotScheduled => {}
2672 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2673 0 : schedule_error = Some(err);
2674 0 : }
2675 : }
2676 : }
2677 :
2678 : // If we failed to schedule shards, then they are still created in the controller,
2679 : // but we return an error to the requester to avoid a silent failure when someone
2680 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2681 : // are present in the system. We do this here rather than in the above loop, to
2682 : // avoid situations where we only create a subset of shards in the tenant.
2683 0 : if let Some(e) = schedule_error {
2684 0 : return Err(ApiError::Conflict(format!(
2685 0 : "Failed to schedule shard(s): {e}"
2686 0 : )));
2687 0 : }
2688 :
2689 0 : let waiters = {
2690 0 : let mut locked = self.inner.write().unwrap();
2691 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2692 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2693 0 : .tenant_creation_hint(true)
2694 0 : .build();
2695 0 : tenants
2696 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2697 0 : .filter_map(|(_shard_id, shard)| {
2698 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2699 0 : })
2700 0 : .collect::<Vec<_>>()
2701 : };
2702 :
2703 0 : Ok((
2704 0 : TenantCreateResponse {
2705 0 : shards: response_shards,
2706 0 : },
2707 0 : waiters,
2708 0 : ))
2709 0 : }
2710 :
2711 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2712 : /// case of a new tenant and a pre-existing one.
2713 : #[allow(clippy::too_many_arguments)]
2714 0 : async fn do_initial_shard_scheduling(
2715 0 : &self,
2716 0 : tenant_shard_id: TenantShardId,
2717 0 : initial_generation: Option<Generation>,
2718 0 : shard_params: ShardParameters,
2719 0 : config: TenantConfig,
2720 0 : placement_policy: PlacementPolicy,
2721 0 : preferred_az_id: Option<&AvailabilityZone>,
2722 0 : schedule_context: &mut ScheduleContext,
2723 0 : ) -> InitialShardScheduleOutcome {
2724 0 : let mut locked = self.inner.write().unwrap();
2725 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2726 :
2727 : use std::collections::btree_map::Entry;
2728 0 : match tenants.entry(tenant_shard_id) {
2729 0 : Entry::Occupied(mut entry) => {
2730 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2731 :
2732 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2733 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2734 0 : }
2735 :
2736 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2737 0 : let generation = entry
2738 0 : .get()
2739 0 : .generation
2740 0 : .expect("Generation is set when in attached mode");
2741 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2742 0 : shard_id: tenant_shard_id,
2743 0 : node_id: *node_id,
2744 0 : generation: generation.into().unwrap(),
2745 0 : })
2746 : } else {
2747 0 : InitialShardScheduleOutcome::NotScheduled
2748 : }
2749 : }
2750 0 : Entry::Vacant(entry) => {
2751 0 : let state = entry.insert(TenantShard::new(
2752 0 : tenant_shard_id,
2753 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2754 0 : placement_policy,
2755 0 : preferred_az_id.cloned(),
2756 : ));
2757 :
2758 0 : state.generation = initial_generation;
2759 0 : state.config = config;
2760 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2761 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2762 0 : }
2763 :
2764 : // Only include shards in result if we are attaching: the purpose
2765 : // of the response is to tell the caller where the shards are attached.
2766 0 : if let Some(node_id) = state.intent.get_attached() {
2767 0 : let generation = state
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 : }
2780 0 : }
2781 :
2782 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2783 : /// wait for reconciliation to complete before responding.
2784 0 : async fn await_waiters(
2785 0 : &self,
2786 0 : waiters: Vec<ReconcilerWaiter>,
2787 0 : timeout: Duration,
2788 0 : ) -> Result<(), ReconcileWaitError> {
2789 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2790 0 : for waiter in waiters {
2791 0 : let timeout = deadline.duration_since(Instant::now());
2792 0 : waiter.wait_timeout(timeout).await?;
2793 : }
2794 :
2795 0 : Ok(())
2796 0 : }
2797 :
2798 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2799 : /// in progress
2800 0 : async fn await_waiters_remainder(
2801 0 : &self,
2802 0 : waiters: Vec<ReconcilerWaiter>,
2803 0 : timeout: Duration,
2804 0 : ) -> Vec<ReconcilerWaiter> {
2805 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2806 0 : for waiter in waiters.iter() {
2807 0 : let timeout = deadline.duration_since(Instant::now());
2808 0 : let _ = waiter.wait_timeout(timeout).await;
2809 : }
2810 :
2811 0 : waiters
2812 0 : .into_iter()
2813 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2814 0 : .collect::<Vec<_>>()
2815 0 : }
2816 :
2817 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2818 : /// and transform it into either a tenant creation of a series of shard updates.
2819 : ///
2820 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2821 : /// still be returned.
2822 0 : fn tenant_location_config_prepare(
2823 0 : &self,
2824 0 : tenant_id: TenantId,
2825 0 : req: TenantLocationConfigRequest,
2826 0 : ) -> TenantCreateOrUpdate {
2827 0 : let mut updates = Vec::new();
2828 0 : let mut locked = self.inner.write().unwrap();
2829 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2830 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2831 :
2832 : // Use location config mode as an indicator of policy.
2833 0 : let placement_policy = match req.config.mode {
2834 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2835 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2836 : LocationConfigMode::AttachedMulti
2837 : | LocationConfigMode::AttachedSingle
2838 : | LocationConfigMode::AttachedStale => {
2839 0 : if nodes.len() > 1 {
2840 0 : PlacementPolicy::Attached(1)
2841 : } else {
2842 : // Convenience for dev/test: if we just have one pageserver, import
2843 : // tenants into non-HA mode so that scheduling will succeed.
2844 0 : PlacementPolicy::Attached(0)
2845 : }
2846 : }
2847 : };
2848 :
2849 : // Ordinarily we do not update scheduling policy, but when making major changes
2850 : // like detaching or demoting to secondary-only, we need to force the scheduling
2851 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2852 0 : let scheduling_policy = match req.config.mode {
2853 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2854 : // Special case: when making major changes like detaching or demoting to secondary-only,
2855 : // we need to force the scheduling mode to Active, or nothing will happen.
2856 0 : Some(ShardSchedulingPolicy::Active)
2857 : }
2858 : LocationConfigMode::AttachedMulti
2859 : | LocationConfigMode::AttachedSingle
2860 : | LocationConfigMode::AttachedStale => {
2861 : // While attached, continue to respect whatever the existing scheduling mode is.
2862 0 : None
2863 : }
2864 : };
2865 :
2866 0 : let mut create = true;
2867 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2868 : // Saw an existing shard: this is not a creation
2869 0 : create = false;
2870 :
2871 : // Shards may have initially been created by a Secondary request, where we
2872 : // would have left generation as None.
2873 : //
2874 : // We only update generation the first time we see an attached-mode request,
2875 : // and if there is no existing generation set. The caller is responsible for
2876 : // ensuring that no non-storage-controller pageserver ever uses a higher
2877 : // generation than they passed in here.
2878 : use LocationConfigMode::*;
2879 0 : let set_generation = match req.config.mode {
2880 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2881 0 : req.config.generation.map(Generation::new)
2882 : }
2883 0 : _ => None,
2884 : };
2885 :
2886 0 : updates.push(ShardUpdate {
2887 0 : tenant_shard_id: *shard_id,
2888 0 : placement_policy: placement_policy.clone(),
2889 0 : tenant_config: req.config.tenant_conf.clone(),
2890 0 : generation: set_generation,
2891 0 : scheduling_policy,
2892 0 : });
2893 : }
2894 :
2895 0 : if create {
2896 : use LocationConfigMode::*;
2897 0 : let generation = match req.config.mode {
2898 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2899 : // If a caller provided a generation in a non-attached request, ignore it
2900 : // and leave our generation as None: this enables a subsequent update to set
2901 : // the generation when setting an attached mode for the first time.
2902 0 : _ => None,
2903 : };
2904 :
2905 0 : TenantCreateOrUpdate::Create(
2906 0 : // Synthesize a creation request
2907 0 : TenantCreateRequest {
2908 0 : new_tenant_id: tenant_shard_id,
2909 0 : generation,
2910 0 : shard_parameters: ShardParameters {
2911 0 : count: tenant_shard_id.shard_count,
2912 0 : // We only import un-sharded or single-sharded tenants, so stripe
2913 0 : // size can be made up arbitrarily here.
2914 0 : stripe_size: DEFAULT_STRIPE_SIZE,
2915 0 : },
2916 0 : placement_policy: Some(placement_policy),
2917 0 : config: req.config.tenant_conf,
2918 0 : },
2919 0 : )
2920 : } else {
2921 0 : assert!(!updates.is_empty());
2922 0 : TenantCreateOrUpdate::Update(updates)
2923 : }
2924 0 : }
2925 :
2926 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2927 : /// the tenant is present in memory. If not, load it from the database. If it is found
2928 : /// in neither location, return a NotFound error.
2929 : ///
2930 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2931 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2932 0 : async fn maybe_load_tenant(
2933 0 : &self,
2934 0 : tenant_id: TenantId,
2935 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2936 0 : ) -> Result<(), ApiError> {
2937 : // Check if the tenant is present in memory, and select an AZ to use when loading
2938 : // if we will load it.
2939 0 : let load_in_az = {
2940 0 : let locked = self.inner.read().unwrap();
2941 0 : let existing = locked
2942 0 : .tenants
2943 0 : .range(TenantShardId::tenant_range(tenant_id))
2944 0 : .next();
2945 :
2946 : // If the tenant is not present in memory, we expect to load it from database,
2947 : // so let's figure out what AZ to load it into while we have self.inner locked.
2948 0 : if existing.is_none() {
2949 0 : locked
2950 0 : .scheduler
2951 0 : .get_az_for_new_tenant()
2952 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2953 0 : "No AZ with nodes found to load tenant"
2954 0 : )))?
2955 : } else {
2956 : // We already have this tenant in memory
2957 0 : return Ok(());
2958 : }
2959 : };
2960 :
2961 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2962 0 : if tenant_shards.is_empty() {
2963 0 : return Err(ApiError::NotFound(
2964 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2965 0 : ));
2966 0 : }
2967 :
2968 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2969 0 : self.persistence
2970 0 : .set_tenant_shard_preferred_azs(
2971 0 : tenant_shards
2972 0 : .iter()
2973 0 : .map(|t| {
2974 0 : (
2975 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2976 0 : Some(load_in_az.clone()),
2977 0 : )
2978 0 : })
2979 0 : .collect(),
2980 : )
2981 0 : .await?;
2982 :
2983 0 : let mut locked = self.inner.write().unwrap();
2984 0 : tracing::info!(
2985 0 : "Loaded {} shards for tenant {}",
2986 0 : tenant_shards.len(),
2987 : tenant_id
2988 : );
2989 :
2990 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2991 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2992 0 : let shard =
2993 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2994 :
2995 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2996 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2997 0 : if shard.policy != PlacementPolicy::Detached {
2998 0 : tracing::error!(
2999 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
3000 : shard.tenant_shard_id,
3001 : shard.policy
3002 : );
3003 0 : }
3004 :
3005 0 : (shard.tenant_shard_id, shard)
3006 0 : }));
3007 :
3008 0 : Ok(())
3009 0 : }
3010 :
3011 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
3012 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
3013 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
3014 : ///
3015 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
3016 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
3017 0 : fn maybe_drop_tenant(
3018 0 : &self,
3019 0 : tenant_id: TenantId,
3020 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
3021 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
3022 0 : ) {
3023 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
3024 0 : if tenant_shards.all(|(_id, shard)| {
3025 0 : shard.policy == PlacementPolicy::Detached
3026 0 : && shard.reconciler.is_none()
3027 0 : && shard.observed.is_empty()
3028 0 : }) {
3029 0 : let keys = locked
3030 0 : .tenants
3031 0 : .range(TenantShardId::tenant_range(tenant_id))
3032 0 : .map(|(id, _)| id)
3033 0 : .copied()
3034 0 : .collect::<Vec<_>>();
3035 0 : for key in keys {
3036 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
3037 0 : locked.tenants.remove(&key);
3038 : }
3039 0 : }
3040 0 : }
3041 :
3042 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
3043 : /// directly with pageservers into this service.
3044 : ///
3045 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
3046 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
3047 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
3048 : /// tenant's source of generation numbers.
3049 : ///
3050 : /// The mode in this request coarse-grained control of tenants:
3051 : /// - Call with mode Attached* to upsert the tenant.
3052 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
3053 : /// to set an existing tenant to PolicyMode::Secondary
3054 : /// - Call with mode Detached to switch to PolicyMode::Detached
3055 0 : pub(crate) async fn tenant_location_config(
3056 0 : &self,
3057 0 : tenant_shard_id: TenantShardId,
3058 0 : req: TenantLocationConfigRequest,
3059 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
3060 : // We require an exclusive lock, because we are updating both persistent and in-memory state
3061 0 : let _tenant_lock = trace_exclusive_lock(
3062 0 : &self.tenant_op_locks,
3063 0 : tenant_shard_id.tenant_id,
3064 0 : TenantOperations::LocationConfig,
3065 0 : )
3066 0 : .await;
3067 :
3068 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
3069 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
3070 0 : "This API is for importing single-sharded or unsharded tenants"
3071 0 : )));
3072 : } else {
3073 0 : tenant_shard_id.tenant_id
3074 : };
3075 :
3076 : // In case we are waking up a Detached tenant
3077 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
3078 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
3079 0 : // This is a creation or an update
3080 0 : }
3081 0 : Err(e) => {
3082 0 : return Err(e);
3083 : }
3084 : };
3085 :
3086 : // First check if this is a creation or an update
3087 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
3088 :
3089 0 : let mut result = TenantLocationConfigResponse {
3090 0 : shards: Vec::new(),
3091 0 : stripe_size: None,
3092 0 : };
3093 0 : let waiters = match create_or_update {
3094 0 : TenantCreateOrUpdate::Create(create_req) => {
3095 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
3096 0 : result.shards = create_resp
3097 0 : .shards
3098 0 : .into_iter()
3099 0 : .map(|s| TenantShardLocation {
3100 0 : node_id: s.node_id,
3101 0 : shard_id: s.shard_id,
3102 0 : })
3103 0 : .collect();
3104 0 : waiters
3105 : }
3106 0 : TenantCreateOrUpdate::Update(updates) => {
3107 : // Persist updates
3108 : // Ordering: write to the database before applying changes in-memory, so that
3109 : // we will not appear time-travel backwards on a restart.
3110 :
3111 0 : let mut schedule_context = ScheduleContext::default();
3112 : for ShardUpdate {
3113 0 : tenant_shard_id,
3114 0 : placement_policy,
3115 0 : tenant_config,
3116 0 : generation,
3117 0 : scheduling_policy,
3118 0 : } in &updates
3119 : {
3120 0 : self.persistence
3121 0 : .update_tenant_shard(
3122 0 : TenantFilter::Shard(*tenant_shard_id),
3123 0 : Some(placement_policy.clone()),
3124 0 : Some(tenant_config.clone()),
3125 0 : *generation,
3126 0 : *scheduling_policy,
3127 0 : )
3128 0 : .await?;
3129 : }
3130 :
3131 : // Apply updates in-memory
3132 0 : let mut waiters = Vec::new();
3133 : {
3134 0 : let mut locked = self.inner.write().unwrap();
3135 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3136 :
3137 : for ShardUpdate {
3138 0 : tenant_shard_id,
3139 0 : placement_policy,
3140 0 : tenant_config,
3141 0 : generation: update_generation,
3142 0 : scheduling_policy,
3143 0 : } in updates
3144 : {
3145 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
3146 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
3147 0 : continue;
3148 : };
3149 :
3150 : // Update stripe size
3151 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
3152 0 : result.stripe_size = Some(shard.shard.stripe_size);
3153 0 : }
3154 :
3155 0 : shard.policy = placement_policy;
3156 0 : shard.config = tenant_config;
3157 0 : if let Some(generation) = update_generation {
3158 0 : shard.generation = Some(generation);
3159 0 : }
3160 :
3161 0 : if let Some(scheduling_policy) = scheduling_policy {
3162 0 : shard.set_scheduling_policy(scheduling_policy);
3163 0 : }
3164 :
3165 0 : shard.schedule(scheduler, &mut schedule_context)?;
3166 :
3167 0 : let maybe_waiter =
3168 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3169 0 : if let Some(waiter) = maybe_waiter {
3170 0 : waiters.push(waiter);
3171 0 : }
3172 :
3173 0 : if let Some(node_id) = shard.intent.get_attached() {
3174 0 : result.shards.push(TenantShardLocation {
3175 0 : shard_id: tenant_shard_id,
3176 0 : node_id: *node_id,
3177 0 : })
3178 0 : }
3179 : }
3180 : }
3181 0 : waiters
3182 : }
3183 : };
3184 :
3185 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3186 : // Do not treat a reconcile error as fatal: we have already applied any requested
3187 : // Intent changes, and the reconcile can fail for external reasons like unavailable
3188 : // compute notification API. In these cases, it is important that we do not
3189 : // cause the cloud control plane to retry forever on this API.
3190 0 : tracing::warn!(
3191 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
3192 : );
3193 0 : }
3194 :
3195 : // Logging the full result is useful because it lets us cross-check what the cloud control
3196 : // plane's tenant_shards table should contain.
3197 0 : tracing::info!("Complete, returning {result:?}");
3198 :
3199 0 : Ok(result)
3200 0 : }
3201 :
3202 0 : pub(crate) async fn tenant_config_patch(
3203 0 : &self,
3204 0 : req: TenantConfigPatchRequest,
3205 0 : ) -> Result<(), ApiError> {
3206 0 : let _tenant_lock = trace_exclusive_lock(
3207 0 : &self.tenant_op_locks,
3208 0 : req.tenant_id,
3209 0 : TenantOperations::ConfigPatch,
3210 0 : )
3211 0 : .await;
3212 :
3213 0 : let tenant_id = req.tenant_id;
3214 0 : let patch = req.config;
3215 :
3216 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3217 :
3218 0 : let base = {
3219 0 : let locked = self.inner.read().unwrap();
3220 0 : let shards = locked
3221 0 : .tenants
3222 0 : .range(TenantShardId::tenant_range(req.tenant_id));
3223 :
3224 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
3225 :
3226 0 : let first = match configs.peek() {
3227 0 : Some(first) => (*first).clone(),
3228 : None => {
3229 0 : return Err(ApiError::NotFound(
3230 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3231 0 : ));
3232 : }
3233 : };
3234 :
3235 0 : if !configs.all_equal() {
3236 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3237 : // This can't happen because we atomically update the database records
3238 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3239 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3240 0 : "Tenant configs for {} are mismatched",
3241 0 : req.tenant_id
3242 0 : )));
3243 0 : }
3244 :
3245 0 : first
3246 : };
3247 :
3248 0 : let updated_config = base
3249 0 : .apply_patch(patch)
3250 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3251 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3252 0 : .await
3253 0 : }
3254 :
3255 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3256 : // We require an exclusive lock, because we are updating persistent and in-memory state
3257 0 : let _tenant_lock = trace_exclusive_lock(
3258 0 : &self.tenant_op_locks,
3259 0 : req.tenant_id,
3260 0 : TenantOperations::ConfigSet,
3261 0 : )
3262 0 : .await;
3263 :
3264 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3265 :
3266 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3267 0 : .await
3268 0 : }
3269 :
3270 0 : async fn set_tenant_config_and_reconcile(
3271 0 : &self,
3272 0 : tenant_id: TenantId,
3273 0 : config: TenantConfig,
3274 0 : ) -> Result<(), ApiError> {
3275 0 : self.persistence
3276 0 : .update_tenant_shard(
3277 0 : TenantFilter::Tenant(tenant_id),
3278 0 : None,
3279 0 : Some(config.clone()),
3280 0 : None,
3281 0 : None,
3282 0 : )
3283 0 : .await?;
3284 :
3285 0 : let waiters = {
3286 0 : let mut waiters = Vec::new();
3287 0 : let mut locked = self.inner.write().unwrap();
3288 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3289 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3290 0 : shard.config = config.clone();
3291 0 : if let Some(waiter) =
3292 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3293 0 : {
3294 0 : waiters.push(waiter);
3295 0 : }
3296 : }
3297 0 : waiters
3298 : };
3299 :
3300 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3301 : // Treat this as success because we have stored the configuration. If e.g.
3302 : // a node was unavailable at this time, it should not stop us accepting a
3303 : // configuration change.
3304 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3305 0 : }
3306 :
3307 0 : Ok(())
3308 0 : }
3309 :
3310 0 : pub(crate) fn tenant_config_get(
3311 0 : &self,
3312 0 : tenant_id: TenantId,
3313 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3314 0 : let config = {
3315 0 : let locked = self.inner.read().unwrap();
3316 :
3317 0 : match locked
3318 0 : .tenants
3319 0 : .range(TenantShardId::tenant_range(tenant_id))
3320 0 : .next()
3321 : {
3322 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3323 : None => {
3324 0 : return Err(ApiError::NotFound(
3325 0 : anyhow::anyhow!("Tenant not found").into(),
3326 0 : ));
3327 : }
3328 : }
3329 : };
3330 :
3331 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3332 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3333 : // and `effective_config` in the response is meaningless, but we retain that syntax
3334 : // in order to remain compatible with the pageserver API.
3335 :
3336 0 : let response = HashMap::from([
3337 : (
3338 : "tenant_specific_overrides",
3339 0 : serde_json::to_value(&config)
3340 0 : .context("serializing tenant specific overrides")
3341 0 : .map_err(ApiError::InternalServerError)?,
3342 : ),
3343 : (
3344 0 : "effective_config",
3345 0 : serde_json::to_value(&config)
3346 0 : .context("serializing effective config")
3347 0 : .map_err(ApiError::InternalServerError)?,
3348 : ),
3349 : ]);
3350 :
3351 0 : Ok(response)
3352 0 : }
3353 :
3354 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3355 0 : &self,
3356 0 : time_travel_req: &TenantTimeTravelRequest,
3357 0 : tenant_id: TenantId,
3358 0 : timestamp: Cow<'_, str>,
3359 0 : done_if_after: Cow<'_, str>,
3360 0 : ) -> Result<(), ApiError> {
3361 0 : let _tenant_lock = trace_exclusive_lock(
3362 0 : &self.tenant_op_locks,
3363 0 : tenant_id,
3364 0 : TenantOperations::TimeTravelRemoteStorage,
3365 0 : )
3366 0 : .await;
3367 :
3368 0 : let node = {
3369 0 : let mut locked = self.inner.write().unwrap();
3370 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3371 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3372 : // but only at the start of the process, so it's really just to prevent operator
3373 : // mistakes.
3374 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3375 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3376 : {
3377 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3378 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3379 0 : )));
3380 0 : }
3381 0 : let maybe_attached = shard
3382 0 : .observed
3383 0 : .locations
3384 0 : .iter()
3385 0 : .filter_map(|(node_id, observed_location)| {
3386 0 : observed_location
3387 0 : .conf
3388 0 : .as_ref()
3389 0 : .map(|loc| (node_id, observed_location, loc.mode))
3390 0 : })
3391 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3392 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3393 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3394 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3395 0 : )));
3396 0 : }
3397 : }
3398 0 : let scheduler = &mut locked.scheduler;
3399 : // Right now we only perform the operation on a single node without parallelization
3400 : // TODO fan out the operation to multiple nodes for better performance
3401 0 : let node_id = scheduler.any_available_node()?;
3402 0 : let node = locked
3403 0 : .nodes
3404 0 : .get(&node_id)
3405 0 : .expect("Pageservers may not be deleted while lock is active");
3406 0 : node.clone()
3407 : };
3408 :
3409 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3410 0 : let mut counts = time_travel_req
3411 0 : .shard_counts
3412 0 : .iter()
3413 0 : .copied()
3414 0 : .collect::<HashSet<_>>()
3415 0 : .into_iter()
3416 0 : .collect::<Vec<_>>();
3417 0 : counts.sort_unstable();
3418 :
3419 0 : for count in counts {
3420 0 : let shard_ids = (0..count.count())
3421 0 : .map(|i| TenantShardId {
3422 0 : tenant_id,
3423 0 : shard_number: ShardNumber(i),
3424 0 : shard_count: count,
3425 0 : })
3426 0 : .collect::<Vec<_>>();
3427 0 : for tenant_shard_id in shard_ids {
3428 0 : let client = PageserverClient::new(
3429 0 : node.get_id(),
3430 0 : self.http_client.clone(),
3431 0 : node.base_url(),
3432 0 : self.config.pageserver_jwt_token.as_deref(),
3433 : );
3434 :
3435 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3436 :
3437 0 : client
3438 0 : .tenant_time_travel_remote_storage(
3439 0 : tenant_shard_id,
3440 0 : ×tamp,
3441 0 : &done_if_after,
3442 0 : )
3443 0 : .await
3444 0 : .map_err(|e| {
3445 0 : ApiError::InternalServerError(anyhow::anyhow!(
3446 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3447 0 : node
3448 0 : ))
3449 0 : })?;
3450 : }
3451 : }
3452 0 : Ok(())
3453 0 : }
3454 :
3455 0 : pub(crate) async fn tenant_secondary_download(
3456 0 : &self,
3457 0 : tenant_id: TenantId,
3458 0 : wait: Option<Duration>,
3459 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3460 0 : let _tenant_lock = trace_shared_lock(
3461 0 : &self.tenant_op_locks,
3462 0 : tenant_id,
3463 0 : TenantOperations::SecondaryDownload,
3464 0 : )
3465 0 : .await;
3466 :
3467 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3468 0 : let targets = {
3469 0 : let locked = self.inner.read().unwrap();
3470 0 : let mut targets = Vec::new();
3471 :
3472 0 : for (tenant_shard_id, shard) in
3473 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3474 : {
3475 0 : for node_id in shard.intent.get_secondary() {
3476 0 : let node = locked
3477 0 : .nodes
3478 0 : .get(node_id)
3479 0 : .expect("Pageservers may not be deleted while referenced");
3480 0 :
3481 0 : targets.push((*tenant_shard_id, node.clone()));
3482 0 : }
3483 : }
3484 0 : targets
3485 : };
3486 :
3487 : // Issue concurrent requests to all shards' locations
3488 0 : let mut futs = FuturesUnordered::new();
3489 0 : for (tenant_shard_id, node) in targets {
3490 0 : let client = PageserverClient::new(
3491 0 : node.get_id(),
3492 0 : self.http_client.clone(),
3493 0 : node.base_url(),
3494 0 : self.config.pageserver_jwt_token.as_deref(),
3495 : );
3496 0 : futs.push(async move {
3497 0 : let result = client
3498 0 : .tenant_secondary_download(tenant_shard_id, wait)
3499 0 : .await;
3500 0 : (result, node, tenant_shard_id)
3501 0 : })
3502 : }
3503 :
3504 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3505 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3506 : // well as more general cases like 503s, 500s, or timeouts.
3507 0 : let mut aggregate_progress = SecondaryProgress::default();
3508 0 : let mut aggregate_status: Option<StatusCode> = None;
3509 0 : let mut error: Option<mgmt_api::Error> = None;
3510 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3511 0 : match result {
3512 0 : Err(e) => {
3513 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3514 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3515 : // than they had hoped for.
3516 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3517 0 : error = Some(e)
3518 : }
3519 0 : Ok((status_code, progress)) => {
3520 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3521 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3522 0 : aggregate_progress.layers_total += progress.layers_total;
3523 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3524 0 : aggregate_progress.bytes_total += progress.bytes_total;
3525 0 : aggregate_progress.heatmap_mtime =
3526 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3527 0 : aggregate_status = match aggregate_status {
3528 0 : None => Some(status_code),
3529 0 : Some(StatusCode::OK) => Some(status_code),
3530 0 : Some(cur) => {
3531 : // Other status codes (e.g. 202) -- do not overwrite.
3532 0 : Some(cur)
3533 : }
3534 : };
3535 : }
3536 : }
3537 : }
3538 :
3539 : // If any of the shards return 202, indicate our result as 202.
3540 0 : match aggregate_status {
3541 : None => {
3542 0 : match error {
3543 0 : Some(e) => {
3544 : // No successes, and an error: surface it
3545 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3546 : }
3547 : None => {
3548 : // No shards found
3549 0 : Err(ApiError::NotFound(
3550 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3551 0 : ))
3552 : }
3553 : }
3554 : }
3555 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3556 : }
3557 0 : }
3558 :
3559 0 : pub(crate) async fn tenant_delete(
3560 0 : self: &Arc<Self>,
3561 0 : tenant_id: TenantId,
3562 0 : ) -> Result<StatusCode, ApiError> {
3563 0 : let _tenant_lock =
3564 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3565 :
3566 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3567 :
3568 : // Detach all shards. This also deletes local pageserver shard data.
3569 0 : let (detach_waiters, node) = {
3570 0 : let mut detach_waiters = Vec::new();
3571 0 : let mut locked = self.inner.write().unwrap();
3572 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3573 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3574 : // Update the tenant's intent to remove all attachments
3575 0 : shard.policy = PlacementPolicy::Detached;
3576 0 : shard
3577 0 : .schedule(scheduler, &mut ScheduleContext::default())
3578 0 : .expect("De-scheduling is infallible");
3579 0 : debug_assert!(shard.intent.get_attached().is_none());
3580 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3581 :
3582 0 : if let Some(waiter) =
3583 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3584 0 : {
3585 0 : detach_waiters.push(waiter);
3586 0 : }
3587 : }
3588 :
3589 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3590 : // was attached, just has to be able to see the S3 content)
3591 0 : let node_id = scheduler.any_available_node()?;
3592 0 : let node = nodes
3593 0 : .get(&node_id)
3594 0 : .expect("Pageservers may not be deleted while lock is active");
3595 0 : (detach_waiters, node.clone())
3596 : };
3597 :
3598 : // This reconcile wait can fail in a few ways:
3599 : // A there is a very long queue for the reconciler semaphore
3600 : // B some pageserver is failing to handle a detach promptly
3601 : // C some pageserver goes offline right at the moment we send it a request.
3602 : //
3603 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3604 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3605 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3606 : // deleting the underlying data).
3607 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3608 0 : .await?;
3609 :
3610 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3611 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3612 : // the tenant ID prefix, including all shards (even possibly stale ones).
3613 0 : match node
3614 0 : .with_client_retries(
3615 0 : |client| async move {
3616 0 : client
3617 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3618 0 : .await
3619 0 : },
3620 0 : &self.http_client,
3621 0 : &self.config.pageserver_jwt_token,
3622 : 1,
3623 : 3,
3624 : RECONCILE_TIMEOUT,
3625 0 : &self.cancel,
3626 : )
3627 0 : .await
3628 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3629 : {
3630 0 : Ok(_) => {}
3631 : Err(mgmt_api::Error::Cancelled) => {
3632 0 : return Err(ApiError::ShuttingDown);
3633 : }
3634 0 : Err(e) => {
3635 : // This is unexpected: remote deletion should be infallible, unless the object store
3636 : // at large is unavailable.
3637 0 : tracing::error!("Error deleting via node {node}: {e}");
3638 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3639 : }
3640 : }
3641 :
3642 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3643 : // our in-memory state and database state.
3644 :
3645 : // Ordering: we delete persistent state first: if we then
3646 : // crash, we will drop the in-memory state.
3647 :
3648 : // Drop persistent state.
3649 0 : self.persistence.delete_tenant(tenant_id).await?;
3650 :
3651 : // Drop in-memory state
3652 : {
3653 0 : let mut locked = self.inner.write().unwrap();
3654 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3655 :
3656 : // Dereference Scheduler from shards before dropping them
3657 0 : for (_tenant_shard_id, shard) in
3658 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3659 0 : {
3660 0 : shard.intent.clear(scheduler);
3661 0 : }
3662 :
3663 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3664 0 : tracing::info!(
3665 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3666 0 : locked.tenants.len()
3667 : );
3668 : };
3669 :
3670 : // Delete the tenant from safekeepers (if needed)
3671 0 : self.tenant_delete_safekeepers(tenant_id)
3672 0 : .instrument(tracing::info_span!("tenant_delete_safekeepers", %tenant_id))
3673 0 : .await?;
3674 :
3675 : // Success is represented as 404, to imitate the existing pageserver deletion API
3676 0 : Ok(StatusCode::NOT_FOUND)
3677 0 : }
3678 :
3679 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3680 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3681 : /// the tenant's policies (configuration) within the storage controller
3682 0 : pub(crate) async fn tenant_update_policy(
3683 0 : &self,
3684 0 : tenant_id: TenantId,
3685 0 : req: TenantPolicyRequest,
3686 0 : ) -> Result<(), ApiError> {
3687 : // We require an exclusive lock, because we are updating persistent and in-memory state
3688 0 : let _tenant_lock = trace_exclusive_lock(
3689 0 : &self.tenant_op_locks,
3690 0 : tenant_id,
3691 0 : TenantOperations::UpdatePolicy,
3692 0 : )
3693 0 : .await;
3694 :
3695 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3696 :
3697 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3698 :
3699 : let TenantPolicyRequest {
3700 0 : placement,
3701 0 : mut scheduling,
3702 0 : } = req;
3703 :
3704 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3705 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3706 : // this to take effect.
3707 0 : if scheduling.is_none() {
3708 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3709 0 : }
3710 0 : }
3711 :
3712 0 : self.persistence
3713 0 : .update_tenant_shard(
3714 0 : TenantFilter::Tenant(tenant_id),
3715 0 : placement.clone(),
3716 0 : None,
3717 0 : None,
3718 0 : scheduling,
3719 0 : )
3720 0 : .await?;
3721 :
3722 0 : let mut schedule_context = ScheduleContext::default();
3723 0 : let mut locked = self.inner.write().unwrap();
3724 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3725 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3726 0 : if let Some(placement) = &placement {
3727 0 : shard.policy = placement.clone();
3728 :
3729 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3730 0 : "Updated placement policy to {placement:?}");
3731 0 : }
3732 :
3733 0 : if let Some(scheduling) = &scheduling {
3734 0 : shard.set_scheduling_policy(*scheduling);
3735 :
3736 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3737 0 : "Updated scheduling policy to {scheduling:?}");
3738 0 : }
3739 :
3740 : // In case scheduling is being switched back on, try it now.
3741 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3742 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3743 : }
3744 :
3745 0 : Ok(())
3746 0 : }
3747 :
3748 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3749 0 : &self,
3750 0 : tenant_id: TenantId,
3751 0 : mut create_req: TimelineCreateRequest,
3752 0 : ) -> Result<TimelineInfo, ApiError> {
3753 0 : tracing::info!(
3754 0 : "Creating timeline {}/{}",
3755 : tenant_id,
3756 : create_req.new_timeline_id,
3757 : );
3758 :
3759 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3760 0 : if targets.0.is_empty() {
3761 0 : return Err(ApiError::NotFound(
3762 0 : anyhow::anyhow!("Tenant not found").into(),
3763 0 : ));
3764 0 : };
3765 :
3766 0 : let (shard_zero_tid, shard_zero_locations) =
3767 0 : targets.0.pop_first().expect("Must have at least one shard");
3768 0 : assert!(shard_zero_tid.is_shard_zero());
3769 :
3770 0 : async fn create_one(
3771 0 : tenant_shard_id: TenantShardId,
3772 0 : locations: ShardMutationLocations,
3773 0 : http_client: reqwest::Client,
3774 0 : jwt: Option<String>,
3775 0 : mut create_req: TimelineCreateRequest,
3776 0 : ) -> Result<TimelineInfo, ApiError> {
3777 0 : let latest = locations.latest.node;
3778 :
3779 0 : tracing::info!(
3780 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3781 : tenant_shard_id,
3782 : create_req.new_timeline_id,
3783 : locations.latest.generation
3784 : );
3785 :
3786 0 : let client =
3787 0 : PageserverClient::new(latest.get_id(), http_client.clone(), latest.base_url(), jwt.as_deref());
3788 :
3789 0 : let timeline_info = client
3790 0 : .timeline_create(tenant_shard_id, &create_req)
3791 0 : .await
3792 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3793 :
3794 : // If we are going to create the timeline on some stale locations for shard 0, then ask them to re-use
3795 : // the initdb generated by the latest location, rather than generating their own. This avoids racing uploads
3796 : // of initdb to S3 which might not be binary-identical if different pageservers have different postgres binaries.
3797 0 : if tenant_shard_id.is_shard_zero() {
3798 0 : if let models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } = &mut create_req.mode {
3799 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id);
3800 0 : }
3801 0 : }
3802 :
3803 : // We propagate timeline creations to all attached locations such that a compute
3804 : // for the new timeline is able to start regardless of the current state of the
3805 : // tenant shard reconciliation.
3806 0 : for location in locations.other {
3807 0 : tracing::info!(
3808 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3809 : tenant_shard_id,
3810 : create_req.new_timeline_id,
3811 : location.node,
3812 : location.generation
3813 : );
3814 :
3815 0 : let client = PageserverClient::new(
3816 0 : location.node.get_id(),
3817 0 : http_client.clone(),
3818 0 : location.node.base_url(),
3819 0 : jwt.as_deref(),
3820 : );
3821 :
3822 0 : let res = client
3823 0 : .timeline_create(tenant_shard_id, &create_req)
3824 0 : .await;
3825 :
3826 0 : if let Err(e) = res {
3827 0 : match e {
3828 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3829 0 : // Tenant might have been detached from the stale location,
3830 0 : // so ignore 404s.
3831 0 : },
3832 : _ => {
3833 0 : return Err(passthrough_api_error(&location.node, e));
3834 : }
3835 : }
3836 0 : }
3837 : }
3838 :
3839 0 : Ok(timeline_info)
3840 0 : }
3841 :
3842 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3843 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3844 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3845 : //
3846 : // This also enables non-zero shards to use the initdb that shard 0 generated and uploaded to S3, rather than
3847 : // independently generating their own initdb. This guarantees that shards cannot end up with different initial
3848 : // states if e.g. they have different postgres binary versions.
3849 0 : let timeline_info = create_one(
3850 0 : shard_zero_tid,
3851 0 : shard_zero_locations,
3852 0 : self.http_client.clone(),
3853 0 : self.config.pageserver_jwt_token.clone(),
3854 0 : create_req.clone(),
3855 0 : )
3856 0 : .await?;
3857 :
3858 : // Update the create request for shards >= 0
3859 0 : match &mut create_req.mode {
3860 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3861 0 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3862 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3863 0 : },
3864 0 : models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } => {
3865 : // For shards >= 0, do not run initdb: use the one that shard 0 uploaded to S3
3866 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id)
3867 : }
3868 0 : _ => {}
3869 : }
3870 :
3871 : // Create timeline on remaining shards with number >0
3872 0 : if !targets.0.is_empty() {
3873 : // If we had multiple shards, issue requests for the remainder now.
3874 0 : let jwt = &self.config.pageserver_jwt_token;
3875 0 : self.tenant_for_shards(
3876 0 : targets
3877 0 : .0
3878 0 : .iter()
3879 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3880 0 : .collect(),
3881 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3882 0 : let create_req = create_req.clone();
3883 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3884 0 : Box::pin(create_one(
3885 0 : tenant_shard_id,
3886 0 : mutation_locations,
3887 0 : self.http_client.clone(),
3888 0 : jwt.clone(),
3889 0 : create_req,
3890 0 : ))
3891 0 : },
3892 : )
3893 0 : .await?;
3894 0 : }
3895 :
3896 0 : Ok(timeline_info)
3897 0 : })
3898 0 : .await?
3899 0 : }
3900 :
3901 0 : pub(crate) async fn tenant_timeline_create(
3902 0 : self: &Arc<Self>,
3903 0 : tenant_id: TenantId,
3904 0 : create_req: TimelineCreateRequest,
3905 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
3906 0 : let safekeepers = self.config.timelines_onto_safekeepers;
3907 0 : let timeline_id = create_req.new_timeline_id;
3908 :
3909 0 : tracing::info!(
3910 0 : mode=%create_req.mode_tag(),
3911 : %safekeepers,
3912 0 : "Creating timeline {}/{}",
3913 : tenant_id,
3914 : timeline_id,
3915 : );
3916 :
3917 0 : let _tenant_lock = trace_shared_lock(
3918 0 : &self.tenant_op_locks,
3919 0 : tenant_id,
3920 0 : TenantOperations::TimelineCreate,
3921 0 : )
3922 0 : .await;
3923 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3924 0 : let is_import = create_req.is_import();
3925 0 : let read_only = matches!(
3926 0 : create_req.mode,
3927 : models::TimelineCreateRequestMode::Branch {
3928 : read_only: true,
3929 : ..
3930 : }
3931 : );
3932 :
3933 0 : if is_import {
3934 : // Ensure that there is no split on-going.
3935 : // [`Self::tenant_shard_split`] holds the exclusive tenant lock
3936 : // for the duration of the split, but here we handle the case
3937 : // where we restarted and the split is being aborted.
3938 0 : let locked = self.inner.read().unwrap();
3939 0 : let splitting = locked
3940 0 : .tenants
3941 0 : .range(TenantShardId::tenant_range(tenant_id))
3942 0 : .any(|(_id, shard)| shard.splitting != SplitState::Idle);
3943 :
3944 0 : if splitting {
3945 0 : return Err(ApiError::Conflict("Tenant is splitting shard".to_string()));
3946 0 : }
3947 0 : }
3948 :
3949 0 : let timeline_info = self
3950 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
3951 0 : .await?;
3952 :
3953 0 : let selected_safekeepers = if is_import {
3954 0 : let shards = {
3955 0 : let locked = self.inner.read().unwrap();
3956 0 : locked
3957 0 : .tenants
3958 0 : .range(TenantShardId::tenant_range(tenant_id))
3959 0 : .map(|(ts_id, _)| ts_id.to_index())
3960 0 : .collect::<Vec<_>>()
3961 : };
3962 :
3963 0 : if !shards
3964 0 : .iter()
3965 0 : .map(|shard_index| shard_index.shard_count)
3966 0 : .all_equal()
3967 : {
3968 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3969 0 : "Inconsistent shard count"
3970 0 : )));
3971 0 : }
3972 :
3973 0 : let import = TimelineImport {
3974 0 : tenant_id,
3975 0 : timeline_id,
3976 0 : shard_statuses: ShardImportStatuses::new(shards),
3977 0 : };
3978 :
3979 0 : let inserted = self
3980 0 : .persistence
3981 0 : .insert_timeline_import(import.to_persistent())
3982 0 : .await
3983 0 : .context("timeline import insert")
3984 0 : .map_err(ApiError::InternalServerError)?;
3985 :
3986 : // Set the importing flag on the tenant shards
3987 0 : self.inner
3988 0 : .write()
3989 0 : .unwrap()
3990 0 : .tenants
3991 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
3992 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Importing);
3993 :
3994 0 : match inserted {
3995 : true => {
3996 0 : tracing::info!(%tenant_id, %timeline_id, "Inserted timeline import");
3997 : }
3998 : false => {
3999 0 : tracing::info!(%tenant_id, %timeline_id, "Timeline import entry already present");
4000 : }
4001 : }
4002 :
4003 0 : None
4004 0 : } else if safekeepers || read_only {
4005 : // Note that for imported timelines, we do not create the timeline on the safekeepers
4006 : // straight away. Instead, we do it once the import finalized such that we know what
4007 : // start LSN to provide for the safekeepers. This is done in
4008 : // [`Self::finalize_timeline_import`].
4009 0 : let res = self
4010 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
4011 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
4012 0 : .await?;
4013 0 : Some(res)
4014 : } else {
4015 0 : None
4016 : };
4017 :
4018 0 : Ok(TimelineCreateResponseStorcon {
4019 0 : timeline_info,
4020 0 : safekeepers: selected_safekeepers,
4021 0 : })
4022 0 : }
4023 :
4024 : #[instrument(skip_all, fields(
4025 : tenant_id=%req.tenant_shard_id.tenant_id,
4026 : shard_id=%req.tenant_shard_id.shard_slug(),
4027 : timeline_id=%req.timeline_id,
4028 : ))]
4029 : pub(crate) async fn handle_timeline_shard_import_progress(
4030 : self: &Arc<Self>,
4031 : req: TimelineImportStatusRequest,
4032 : ) -> Result<ShardImportStatus, ApiError> {
4033 : let validity = self
4034 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4035 : .await?;
4036 : match validity {
4037 : ShardGenerationValidity::Valid => {
4038 : // fallthrough
4039 : }
4040 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4041 : tracing::info!(
4042 : claimed=?claimed.into(),
4043 0 : actual=?actual.and_then(|g| g.into()),
4044 : "Rejecting import progress fetch from stale generation"
4045 : );
4046 :
4047 : return Err(ApiError::BadRequest(anyhow::anyhow!("Invalid generation")));
4048 : }
4049 : }
4050 :
4051 : let maybe_import = self
4052 : .persistence
4053 : .get_timeline_import(req.tenant_shard_id.tenant_id, req.timeline_id)
4054 : .await?;
4055 :
4056 0 : let import = maybe_import.ok_or_else(|| {
4057 0 : ApiError::NotFound(
4058 0 : format!(
4059 0 : "import for {}/{} not found",
4060 0 : req.tenant_shard_id.tenant_id, req.timeline_id
4061 0 : )
4062 0 : .into(),
4063 0 : )
4064 0 : })?;
4065 :
4066 : import
4067 : .shard_statuses
4068 : .0
4069 : .get(&req.tenant_shard_id.to_index())
4070 : .cloned()
4071 0 : .ok_or_else(|| {
4072 0 : ApiError::NotFound(
4073 0 : format!("shard {} not found", req.tenant_shard_id.shard_slug()).into(),
4074 0 : )
4075 0 : })
4076 : }
4077 :
4078 : #[instrument(skip_all, fields(
4079 : tenant_id=%req.tenant_shard_id.tenant_id,
4080 : shard_id=%req.tenant_shard_id.shard_slug(),
4081 : timeline_id=%req.timeline_id,
4082 : ))]
4083 : pub(crate) async fn handle_timeline_shard_import_progress_upcall(
4084 : self: &Arc<Self>,
4085 : req: PutTimelineImportStatusRequest,
4086 : ) -> Result<(), ApiError> {
4087 : let validity = self
4088 : .validate_shard_generation(req.tenant_shard_id, req.generation)
4089 : .await?;
4090 : match validity {
4091 : ShardGenerationValidity::Valid => {
4092 : // fallthrough
4093 : }
4094 : ShardGenerationValidity::Mismatched { claimed, actual } => {
4095 : tracing::info!(
4096 : claimed=?claimed.into(),
4097 0 : actual=?actual.and_then(|g| g.into()),
4098 : "Rejecting import progress update from stale generation"
4099 : );
4100 :
4101 : return Err(ApiError::PreconditionFailed("Invalid generation".into()));
4102 : }
4103 : }
4104 :
4105 : let res = self
4106 : .persistence
4107 : .update_timeline_import(req.tenant_shard_id, req.timeline_id, req.status)
4108 : .await;
4109 : let timeline_import = match res {
4110 : Ok(Ok(Some(timeline_import))) => timeline_import,
4111 : Ok(Ok(None)) => {
4112 : // Idempotency: we've already seen and handled this update.
4113 : return Ok(());
4114 : }
4115 : Ok(Err(logical_err)) => {
4116 : return Err(logical_err.into());
4117 : }
4118 : Err(db_err) => {
4119 : return Err(db_err.into());
4120 : }
4121 : };
4122 :
4123 : tracing::info!(
4124 : tenant_id=%req.tenant_shard_id.tenant_id,
4125 : timeline_id=%req.timeline_id,
4126 : shard_id=%req.tenant_shard_id.shard_slug(),
4127 : "Updated timeline import status to: {timeline_import:?}");
4128 :
4129 : if timeline_import.is_complete() {
4130 : tokio::task::spawn({
4131 : let this = self.clone();
4132 0 : async move { this.finalize_timeline_import(timeline_import).await }
4133 : });
4134 : }
4135 :
4136 : Ok(())
4137 : }
4138 :
4139 : /// Check that a provided generation for some tenant shard is the most recent one.
4140 : ///
4141 : /// Validate with the in-mem state first, and, if that passes, validate with the
4142 : /// database state which is authoritative.
4143 0 : async fn validate_shard_generation(
4144 0 : self: &Arc<Self>,
4145 0 : tenant_shard_id: TenantShardId,
4146 0 : generation: Generation,
4147 0 : ) -> Result<ShardGenerationValidity, ApiError> {
4148 : {
4149 0 : let locked = self.inner.read().unwrap();
4150 0 : let tenant_shard =
4151 0 : locked
4152 0 : .tenants
4153 0 : .get(&tenant_shard_id)
4154 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4155 0 : "{} shard not found",
4156 0 : tenant_shard_id
4157 0 : )))?;
4158 :
4159 0 : if tenant_shard.generation != Some(generation) {
4160 0 : return Ok(ShardGenerationValidity::Mismatched {
4161 0 : claimed: generation,
4162 0 : actual: tenant_shard.generation,
4163 0 : });
4164 0 : }
4165 : }
4166 :
4167 0 : let mut db_generations = self
4168 0 : .persistence
4169 0 : .shard_generations(std::iter::once(&tenant_shard_id))
4170 0 : .await?;
4171 0 : let (_tid, db_generation) =
4172 0 : db_generations
4173 0 : .pop()
4174 0 : .ok_or(ApiError::InternalServerError(anyhow::anyhow!(
4175 0 : "{} shard not found",
4176 0 : tenant_shard_id
4177 0 : )))?;
4178 :
4179 0 : if db_generation != Some(generation) {
4180 0 : return Ok(ShardGenerationValidity::Mismatched {
4181 0 : claimed: generation,
4182 0 : actual: db_generation,
4183 0 : });
4184 0 : }
4185 :
4186 0 : Ok(ShardGenerationValidity::Valid)
4187 0 : }
4188 :
4189 : /// Finalize the import of a timeline
4190 : ///
4191 : /// This method should be called once all shards have reported that the import is complete.
4192 : /// Firstly, it polls the post import timeline activation endpoint exposed by the pageserver.
4193 : /// Once the timeline is active on all shards, the timeline also gets created on the
4194 : /// safekeepers. Finally, notify cplane of the import completion (whether failed or
4195 : /// successful), and remove the import from the database and in-memory.
4196 : ///
4197 : /// If this method gets pre-empted by shut down, it will be called again at start-up (on-going
4198 : /// imports are stored in the database).
4199 : ///
4200 : /// # Cancel-Safety
4201 : /// Not cancel safe.
4202 : /// If the caller stops polling, the import will not be removed from
4203 : /// [`ServiceState::imports_finalizing`].
4204 : #[instrument(skip_all, fields(
4205 : tenant_id=%import.tenant_id,
4206 : timeline_id=%import.timeline_id,
4207 : ))]
4208 :
4209 : async fn finalize_timeline_import(
4210 : self: &Arc<Self>,
4211 : import: TimelineImport,
4212 : ) -> Result<(), TimelineImportFinalizeError> {
4213 : let tenant_timeline = (import.tenant_id, import.timeline_id);
4214 :
4215 : let (_finalize_import_guard, cancel) = {
4216 : let mut locked = self.inner.write().unwrap();
4217 : let gate = Gate::default();
4218 : let cancel = CancellationToken::default();
4219 :
4220 : let guard = gate.enter().unwrap();
4221 :
4222 : locked.imports_finalizing.insert(
4223 : tenant_timeline,
4224 : FinalizingImport {
4225 : gate,
4226 : cancel: cancel.clone(),
4227 : },
4228 : );
4229 :
4230 : (guard, cancel)
4231 : };
4232 :
4233 : let res = tokio::select! {
4234 : res = self.finalize_timeline_import_impl(import) => {
4235 : res
4236 : },
4237 : _ = cancel.cancelled() => {
4238 : Err(TimelineImportFinalizeError::Cancelled)
4239 : }
4240 : };
4241 :
4242 : let mut locked = self.inner.write().unwrap();
4243 : locked.imports_finalizing.remove(&tenant_timeline);
4244 :
4245 : res
4246 : }
4247 :
4248 0 : async fn finalize_timeline_import_impl(
4249 0 : self: &Arc<Self>,
4250 0 : import: TimelineImport,
4251 0 : ) -> Result<(), TimelineImportFinalizeError> {
4252 0 : tracing::info!("Finalizing timeline import");
4253 :
4254 0 : pausable_failpoint!("timeline-import-pre-cplane-notification");
4255 :
4256 0 : let tenant_id = import.tenant_id;
4257 0 : let timeline_id = import.timeline_id;
4258 :
4259 0 : let import_error = import.completion_error();
4260 0 : match import_error {
4261 0 : Some(err) => {
4262 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Err(err))
4263 0 : .await?;
4264 0 : tracing::warn!("Timeline import completed with shard errors");
4265 0 : Ok(())
4266 : }
4267 0 : None => match self.activate_timeline_post_import(&import).await {
4268 0 : Ok(timeline_info) => {
4269 0 : tracing::info!("Post import timeline activation complete");
4270 :
4271 0 : if self.config.timelines_onto_safekeepers {
4272 : // Now that we know the start LSN of this timeline, create it on the
4273 : // safekeepers.
4274 0 : self.tenant_timeline_create_safekeepers_until_success(
4275 0 : import.tenant_id,
4276 0 : timeline_info,
4277 0 : )
4278 0 : .await?;
4279 0 : }
4280 :
4281 0 : self.notify_cplane_and_delete_import(tenant_id, timeline_id, Ok(()))
4282 0 : .await?;
4283 :
4284 0 : tracing::info!("Timeline import completed successfully");
4285 0 : Ok(())
4286 : }
4287 : Err(TimelineImportFinalizeError::ShuttingDown) => {
4288 : // We got pre-empted by shut down and will resume after the restart.
4289 0 : Err(TimelineImportFinalizeError::ShuttingDown)
4290 : }
4291 0 : Err(err) => {
4292 : // Any finalize error apart from shut down is permanent and requires us to notify
4293 : // cplane such that it can clean up.
4294 0 : tracing::error!("Import finalize failed with permanent error: {err}");
4295 0 : self.notify_cplane_and_delete_import(
4296 0 : tenant_id,
4297 0 : timeline_id,
4298 0 : Err(err.to_string()),
4299 0 : )
4300 0 : .await?;
4301 0 : Err(err)
4302 : }
4303 : },
4304 : }
4305 0 : }
4306 :
4307 0 : async fn notify_cplane_and_delete_import(
4308 0 : self: &Arc<Self>,
4309 0 : tenant_id: TenantId,
4310 0 : timeline_id: TimelineId,
4311 0 : import_result: ImportResult,
4312 0 : ) -> Result<(), TimelineImportFinalizeError> {
4313 0 : let import_failed = import_result.is_err();
4314 0 : tracing::info!(%import_failed, "Notifying cplane of import completion");
4315 :
4316 0 : let client = UpcallClient::new(self.get_config(), self.cancel.child_token());
4317 0 : client
4318 0 : .notify_import_complete(tenant_id, timeline_id, import_result)
4319 0 : .await
4320 0 : .map_err(|_err| TimelineImportFinalizeError::ShuttingDown)?;
4321 :
4322 0 : if let Err(err) = self
4323 0 : .persistence
4324 0 : .delete_timeline_import(tenant_id, timeline_id)
4325 0 : .await
4326 : {
4327 0 : tracing::warn!("Failed to delete timeline import entry from database: {err}");
4328 0 : }
4329 :
4330 0 : self.inner
4331 0 : .write()
4332 0 : .unwrap()
4333 0 : .tenants
4334 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
4335 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Idle);
4336 :
4337 0 : Ok(())
4338 0 : }
4339 :
4340 : /// Activate an imported timeline on all shards once the import is complete.
4341 : /// Returns the [`TimelineInfo`] reported by shard zero.
4342 0 : async fn activate_timeline_post_import(
4343 0 : self: &Arc<Self>,
4344 0 : import: &TimelineImport,
4345 0 : ) -> Result<TimelineInfo, TimelineImportFinalizeError> {
4346 : const TIMELINE_ACTIVATE_TIMEOUT: Duration = Duration::from_millis(128);
4347 :
4348 0 : let mut shards_to_activate: HashSet<ShardIndex> =
4349 0 : import.shard_statuses.0.keys().cloned().collect();
4350 0 : let mut shard_zero_timeline_info = None;
4351 :
4352 0 : while !shards_to_activate.is_empty() {
4353 0 : if self.cancel.is_cancelled() {
4354 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4355 0 : }
4356 :
4357 0 : let targets = {
4358 0 : let locked = self.inner.read().unwrap();
4359 0 : let mut targets = Vec::new();
4360 :
4361 0 : for (tenant_shard_id, shard) in locked
4362 0 : .tenants
4363 0 : .range(TenantShardId::tenant_range(import.tenant_id))
4364 : {
4365 0 : if !import
4366 0 : .shard_statuses
4367 0 : .0
4368 0 : .contains_key(&tenant_shard_id.to_index())
4369 : {
4370 0 : return Err(TimelineImportFinalizeError::MismatchedShards(
4371 0 : tenant_shard_id.to_index(),
4372 0 : ));
4373 0 : }
4374 :
4375 0 : if let Some(node_id) = shard.intent.get_attached() {
4376 0 : let node = locked
4377 0 : .nodes
4378 0 : .get(node_id)
4379 0 : .expect("Pageservers may not be deleted while referenced");
4380 0 : targets.push((*tenant_shard_id, node.clone()));
4381 0 : }
4382 : }
4383 :
4384 0 : targets
4385 : };
4386 :
4387 0 : let targeted_tenant_shards: Vec<_> = targets.iter().map(|(tid, _node)| *tid).collect();
4388 :
4389 0 : let results = self
4390 0 : .tenant_for_shards_api(
4391 0 : targets,
4392 0 : |tenant_shard_id, client| async move {
4393 0 : client
4394 0 : .activate_post_import(
4395 0 : tenant_shard_id,
4396 0 : import.timeline_id,
4397 0 : TIMELINE_ACTIVATE_TIMEOUT,
4398 0 : )
4399 0 : .await
4400 0 : },
4401 : 1,
4402 : 1,
4403 : SHORT_RECONCILE_TIMEOUT,
4404 0 : &self.cancel,
4405 : )
4406 0 : .await;
4407 :
4408 0 : let mut failed = 0;
4409 0 : for (tid, result) in targeted_tenant_shards.iter().zip(results.into_iter()) {
4410 0 : match result {
4411 0 : Ok(ok) => {
4412 0 : if tid.is_shard_zero() {
4413 0 : shard_zero_timeline_info = Some(ok);
4414 0 : }
4415 :
4416 0 : shards_to_activate.remove(&tid.to_index());
4417 : }
4418 0 : Err(_err) => {
4419 0 : failed += 1;
4420 0 : }
4421 : }
4422 : }
4423 :
4424 0 : if failed > 0 {
4425 0 : tracing::info!(
4426 0 : "Failed to activate timeline on {failed} shards post import. Will retry"
4427 : );
4428 0 : }
4429 :
4430 0 : tokio::select! {
4431 0 : _ = tokio::time::sleep(Duration::from_millis(250)) => {},
4432 0 : _ = self.cancel.cancelled() => {
4433 0 : return Err(TimelineImportFinalizeError::ShuttingDown);
4434 : }
4435 : }
4436 : }
4437 :
4438 0 : Ok(shard_zero_timeline_info.expect("All shards replied"))
4439 0 : }
4440 :
4441 0 : async fn finalize_timeline_imports(self: &Arc<Self>, imports: Vec<TimelineImport>) {
4442 0 : futures::future::join_all(
4443 0 : imports
4444 0 : .into_iter()
4445 0 : .map(|import| self.finalize_timeline_import(import)),
4446 : )
4447 0 : .await;
4448 0 : }
4449 :
4450 : /// Delete a timeline import if it exists
4451 : ///
4452 : /// Firstly, delete the entry from the database. Any updates
4453 : /// from pageservers after the update will fail with a 404, so the
4454 : /// import cannot progress into finalizing state if it's not there already.
4455 : /// Secondly, cancel the finalization if one is in progress.
4456 0 : pub(crate) async fn maybe_delete_timeline_import(
4457 0 : self: &Arc<Self>,
4458 0 : tenant_id: TenantId,
4459 0 : timeline_id: TimelineId,
4460 0 : ) -> Result<(), DatabaseError> {
4461 0 : let tenant_has_ongoing_import = {
4462 0 : let locked = self.inner.read().unwrap();
4463 0 : locked
4464 0 : .tenants
4465 0 : .range(TenantShardId::tenant_range(tenant_id))
4466 0 : .any(|(_tid, shard)| shard.importing == TimelineImportState::Importing)
4467 : };
4468 :
4469 0 : if !tenant_has_ongoing_import {
4470 0 : return Ok(());
4471 0 : }
4472 :
4473 0 : self.persistence
4474 0 : .delete_timeline_import(tenant_id, timeline_id)
4475 0 : .await?;
4476 :
4477 0 : let maybe_finalizing = {
4478 0 : let mut locked = self.inner.write().unwrap();
4479 0 : locked.imports_finalizing.remove(&(tenant_id, timeline_id))
4480 : };
4481 :
4482 0 : if let Some(finalizing) = maybe_finalizing {
4483 0 : finalizing.cancel.cancel();
4484 0 : finalizing.gate.close().await;
4485 0 : }
4486 :
4487 0 : Ok(())
4488 0 : }
4489 :
4490 0 : pub(crate) async fn tenant_timeline_archival_config(
4491 0 : &self,
4492 0 : tenant_id: TenantId,
4493 0 : timeline_id: TimelineId,
4494 0 : req: TimelineArchivalConfigRequest,
4495 0 : ) -> Result<(), ApiError> {
4496 0 : tracing::info!(
4497 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
4498 : req.state
4499 : );
4500 :
4501 0 : let _tenant_lock = trace_shared_lock(
4502 0 : &self.tenant_op_locks,
4503 0 : tenant_id,
4504 0 : TenantOperations::TimelineArchivalConfig,
4505 0 : )
4506 0 : .await;
4507 :
4508 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4509 0 : if targets.0.is_empty() {
4510 0 : return Err(ApiError::NotFound(
4511 0 : anyhow::anyhow!("Tenant not found").into(),
4512 0 : ));
4513 0 : }
4514 0 : async fn config_one(
4515 0 : tenant_shard_id: TenantShardId,
4516 0 : timeline_id: TimelineId,
4517 0 : node: Node,
4518 0 : http_client: reqwest::Client,
4519 0 : jwt: Option<String>,
4520 0 : req: TimelineArchivalConfigRequest,
4521 0 : ) -> Result<(), ApiError> {
4522 0 : tracing::info!(
4523 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4524 : );
4525 :
4526 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4527 :
4528 0 : client
4529 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
4530 0 : .await
4531 0 : .map_err(|e| match e {
4532 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
4533 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
4534 : }
4535 0 : _ => passthrough_api_error(&node, e),
4536 0 : })
4537 0 : }
4538 :
4539 : // no shard needs to go first/last; the operation should be idempotent
4540 : // TODO: it would be great to ensure that all shards return the same error
4541 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4542 0 : let results = self
4543 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4544 0 : futures::FutureExt::boxed(config_one(
4545 0 : tenant_shard_id,
4546 0 : timeline_id,
4547 0 : node,
4548 0 : self.http_client.clone(),
4549 0 : self.config.pageserver_jwt_token.clone(),
4550 0 : req.clone(),
4551 0 : ))
4552 0 : })
4553 0 : .await?;
4554 0 : assert!(!results.is_empty(), "must have at least one result");
4555 :
4556 0 : Ok(())
4557 0 : }).await?
4558 0 : }
4559 :
4560 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
4561 0 : &self,
4562 0 : tenant_id: TenantId,
4563 0 : timeline_id: TimelineId,
4564 0 : behavior: Option<DetachBehavior>,
4565 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
4566 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
4567 :
4568 0 : let _tenant_lock = trace_shared_lock(
4569 0 : &self.tenant_op_locks,
4570 0 : tenant_id,
4571 0 : TenantOperations::TimelineDetachAncestor,
4572 0 : )
4573 0 : .await;
4574 :
4575 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4576 0 : if targets.0.is_empty() {
4577 0 : return Err(ApiError::NotFound(
4578 0 : anyhow::anyhow!("Tenant not found").into(),
4579 0 : ));
4580 0 : }
4581 :
4582 0 : async fn detach_one(
4583 0 : tenant_shard_id: TenantShardId,
4584 0 : timeline_id: TimelineId,
4585 0 : node: Node,
4586 0 : http_client: reqwest::Client,
4587 0 : jwt: Option<String>,
4588 0 : behavior: Option<DetachBehavior>,
4589 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
4590 0 : tracing::info!(
4591 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4592 : );
4593 :
4594 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4595 :
4596 0 : client
4597 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id, behavior)
4598 0 : .await
4599 0 : .map_err(|e| {
4600 : use mgmt_api::Error;
4601 :
4602 0 : match e {
4603 : // no ancestor (ever)
4604 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
4605 0 : "{node}: {}",
4606 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
4607 0 : )),
4608 : // too many ancestors
4609 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
4610 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
4611 : }
4612 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
4613 : // avoid turning these into conflicts to remain compatible with
4614 : // pageservers, 500 errors are sadly retryable with timeline ancestor
4615 : // detach
4616 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
4617 : }
4618 : // rest can be mapped as usual
4619 0 : other => passthrough_api_error(&node, other),
4620 : }
4621 0 : })
4622 0 : .map(|res| (tenant_shard_id.shard_number, res))
4623 0 : }
4624 :
4625 : // no shard needs to go first/last; the operation should be idempotent
4626 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4627 0 : let mut results = self
4628 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4629 0 : futures::FutureExt::boxed(detach_one(
4630 0 : tenant_shard_id,
4631 0 : timeline_id,
4632 0 : node,
4633 0 : self.http_client.clone(),
4634 0 : self.config.pageserver_jwt_token.clone(),
4635 0 : behavior,
4636 0 : ))
4637 0 : })
4638 0 : .await?;
4639 :
4640 0 : let any = results.pop().expect("we must have at least one response");
4641 :
4642 0 : let mismatching = results
4643 0 : .iter()
4644 0 : .filter(|(_, res)| res != &any.1)
4645 0 : .collect::<Vec<_>>();
4646 0 : if !mismatching.is_empty() {
4647 : // this can be hit by races which should not happen because operation lock on cplane
4648 0 : let matching = results.len() - mismatching.len();
4649 0 : tracing::error!(
4650 : matching,
4651 : compared_against=?any,
4652 : ?mismatching,
4653 0 : "shards returned different results"
4654 : );
4655 :
4656 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
4657 0 : }
4658 :
4659 0 : Ok(any.1)
4660 0 : }).await?
4661 0 : }
4662 :
4663 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
4664 0 : &self,
4665 0 : tenant_id: TenantId,
4666 0 : timeline_id: TimelineId,
4667 0 : dir: BlockUnblock,
4668 0 : ) -> Result<(), ApiError> {
4669 0 : let _tenant_lock = trace_shared_lock(
4670 0 : &self.tenant_op_locks,
4671 0 : tenant_id,
4672 0 : TenantOperations::TimelineGcBlockUnblock,
4673 0 : )
4674 0 : .await;
4675 :
4676 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4677 0 : if targets.0.is_empty() {
4678 0 : return Err(ApiError::NotFound(
4679 0 : anyhow::anyhow!("Tenant not found").into(),
4680 0 : ));
4681 0 : }
4682 :
4683 0 : async fn do_one(
4684 0 : tenant_shard_id: TenantShardId,
4685 0 : timeline_id: TimelineId,
4686 0 : node: Node,
4687 0 : http_client: reqwest::Client,
4688 0 : jwt: Option<String>,
4689 0 : dir: BlockUnblock,
4690 0 : ) -> Result<(), ApiError> {
4691 0 : let client = PageserverClient::new(
4692 0 : node.get_id(),
4693 0 : http_client,
4694 0 : node.base_url(),
4695 0 : jwt.as_deref(),
4696 : );
4697 :
4698 0 : client
4699 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
4700 0 : .await
4701 0 : .map_err(|e| passthrough_api_error(&node, e))
4702 0 : }
4703 :
4704 : // no shard needs to go first/last; the operation should be idempotent
4705 0 : let locations = targets
4706 0 : .0
4707 0 : .iter()
4708 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
4709 0 : .collect();
4710 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
4711 0 : futures::FutureExt::boxed(do_one(
4712 0 : tenant_shard_id,
4713 0 : timeline_id,
4714 0 : node,
4715 0 : self.http_client.clone(),
4716 0 : self.config.pageserver_jwt_token.clone(),
4717 0 : dir,
4718 0 : ))
4719 0 : })
4720 0 : .await
4721 0 : })
4722 0 : .await??;
4723 0 : Ok(())
4724 0 : }
4725 :
4726 0 : pub(crate) async fn tenant_timeline_lsn_lease(
4727 0 : &self,
4728 0 : tenant_id: TenantId,
4729 0 : timeline_id: TimelineId,
4730 0 : lsn: Lsn,
4731 0 : ) -> Result<LsnLease, ApiError> {
4732 0 : let _tenant_lock = trace_shared_lock(
4733 0 : &self.tenant_op_locks,
4734 0 : tenant_id,
4735 0 : TenantOperations::TimelineLsnLease,
4736 0 : )
4737 0 : .await;
4738 :
4739 0 : let targets = {
4740 0 : let locked = self.inner.read().unwrap();
4741 0 : let mut targets = Vec::new();
4742 :
4743 : // If the request got an unsharded tenant id, then apply
4744 : // the operation to all shards. Otherwise, apply it to a specific shard.
4745 0 : let shards_range = TenantShardId::tenant_range(tenant_id);
4746 :
4747 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4748 0 : if let Some(node_id) = shard.intent.get_attached() {
4749 0 : let node = locked
4750 0 : .nodes
4751 0 : .get(node_id)
4752 0 : .expect("Pageservers may not be deleted while referenced");
4753 0 :
4754 0 : targets.push((*tenant_shard_id, node.clone()));
4755 0 : }
4756 : }
4757 0 : targets
4758 : };
4759 :
4760 0 : let res = self
4761 0 : .tenant_for_shards_api(
4762 0 : targets,
4763 0 : |tenant_shard_id, client| async move {
4764 0 : client
4765 0 : .timeline_lease_lsn(tenant_shard_id, timeline_id, lsn)
4766 0 : .await
4767 0 : },
4768 : 1,
4769 : 1,
4770 : SHORT_RECONCILE_TIMEOUT,
4771 0 : &self.cancel,
4772 : )
4773 0 : .await;
4774 :
4775 0 : let mut valid_until = None;
4776 0 : for r in res {
4777 0 : match r {
4778 0 : Ok(lease) => {
4779 0 : if let Some(ref mut valid_until) = valid_until {
4780 0 : *valid_until = std::cmp::min(*valid_until, lease.valid_until);
4781 0 : } else {
4782 0 : valid_until = Some(lease.valid_until);
4783 0 : }
4784 : }
4785 0 : Err(e) => {
4786 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
4787 : }
4788 : }
4789 : }
4790 0 : Ok(LsnLease {
4791 0 : valid_until: valid_until.unwrap_or_else(SystemTime::now),
4792 0 : })
4793 0 : }
4794 :
4795 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4796 0 : &self,
4797 0 : tenant_shard_id: TenantShardId,
4798 0 : timeline_id: TimelineId,
4799 0 : concurrency: Option<usize>,
4800 0 : recurse: bool,
4801 0 : ) -> Result<(), ApiError> {
4802 0 : let _tenant_lock = trace_shared_lock(
4803 0 : &self.tenant_op_locks,
4804 0 : tenant_shard_id.tenant_id,
4805 0 : TenantOperations::DownloadHeatmapLayers,
4806 0 : )
4807 0 : .await;
4808 :
4809 0 : let targets = {
4810 0 : let locked = self.inner.read().unwrap();
4811 0 : let mut targets = Vec::new();
4812 :
4813 : // If the request got an unsharded tenant id, then apply
4814 : // the operation to all shards. Otherwise, apply it to a specific shard.
4815 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4816 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4817 : } else {
4818 0 : tenant_shard_id.range()
4819 : };
4820 :
4821 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4822 0 : if let Some(node_id) = shard.intent.get_attached() {
4823 0 : let node = locked
4824 0 : .nodes
4825 0 : .get(node_id)
4826 0 : .expect("Pageservers may not be deleted while referenced");
4827 0 :
4828 0 : targets.push((*tenant_shard_id, node.clone()));
4829 0 : }
4830 : }
4831 0 : targets
4832 : };
4833 :
4834 0 : self.tenant_for_shards_api(
4835 0 : targets,
4836 0 : |tenant_shard_id, client| async move {
4837 0 : client
4838 0 : .timeline_download_heatmap_layers(
4839 0 : tenant_shard_id,
4840 0 : timeline_id,
4841 0 : concurrency,
4842 0 : recurse,
4843 0 : )
4844 0 : .await
4845 0 : },
4846 : 1,
4847 : 1,
4848 : SHORT_RECONCILE_TIMEOUT,
4849 0 : &self.cancel,
4850 : )
4851 0 : .await;
4852 :
4853 0 : Ok(())
4854 0 : }
4855 :
4856 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4857 : ///
4858 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`
4859 : /// and returned element at index `i` is the result for `req_fn(op(locations[i])`.
4860 0 : async fn tenant_for_shards<F, R>(
4861 0 : &self,
4862 0 : locations: Vec<(TenantShardId, Node)>,
4863 0 : mut req_fn: F,
4864 0 : ) -> Result<Vec<R>, ApiError>
4865 0 : where
4866 0 : F: FnMut(
4867 0 : TenantShardId,
4868 0 : Node,
4869 0 : )
4870 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4871 0 : {
4872 0 : let mut futs = FuturesUnordered::new();
4873 0 : let mut results = Vec::with_capacity(locations.len());
4874 :
4875 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4876 0 : let fut = req_fn(tenant_shard_id, node);
4877 0 : futs.push(async move { (idx, fut.await) });
4878 : }
4879 :
4880 0 : while let Some((idx, r)) = futs.next().await {
4881 0 : results.push((idx, r?));
4882 : }
4883 :
4884 0 : results.sort_by_key(|(idx, _)| *idx);
4885 0 : Ok(results.into_iter().map(|(_, r)| r).collect())
4886 0 : }
4887 :
4888 : /// Concurrently invoke a pageserver API call on many shards at once.
4889 : ///
4890 : /// The returned Vec has the same length as the `locations` Vec,
4891 : /// and returned element at index `i` is the result for `op(locations[i])`.
4892 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
4893 0 : &self,
4894 0 : locations: Vec<(TenantShardId, Node)>,
4895 0 : op: O,
4896 0 : warn_threshold: u32,
4897 0 : max_retries: u32,
4898 0 : timeout: Duration,
4899 0 : cancel: &CancellationToken,
4900 0 : ) -> Vec<mgmt_api::Result<T>>
4901 0 : where
4902 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
4903 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
4904 0 : {
4905 0 : let mut futs = FuturesUnordered::new();
4906 0 : let mut results = Vec::with_capacity(locations.len());
4907 :
4908 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4909 0 : futs.push(async move {
4910 0 : let r = node
4911 0 : .with_client_retries(
4912 0 : |client| op(tenant_shard_id, client),
4913 0 : &self.http_client,
4914 0 : &self.config.pageserver_jwt_token,
4915 0 : warn_threshold,
4916 0 : max_retries,
4917 0 : timeout,
4918 0 : cancel,
4919 : )
4920 0 : .await;
4921 0 : (idx, r)
4922 0 : });
4923 : }
4924 :
4925 0 : while let Some((idx, r)) = futs.next().await {
4926 0 : results.push((idx, r.unwrap_or(Err(mgmt_api::Error::Cancelled))));
4927 0 : }
4928 :
4929 0 : results.sort_by_key(|(idx, _)| *idx);
4930 0 : results.into_iter().map(|(_, r)| r).collect()
4931 0 : }
4932 :
4933 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
4934 : /// when creating and deleting timelines:
4935 : /// - Makes sure shards are attached somewhere if they weren't already
4936 : /// - Looks up the shards and the nodes where they were most recently attached
4937 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
4938 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
4939 0 : async fn tenant_remote_mutation<R, O, F>(
4940 0 : &self,
4941 0 : tenant_id: TenantId,
4942 0 : op: O,
4943 0 : ) -> Result<R, ApiError>
4944 0 : where
4945 0 : O: FnOnce(TenantMutationLocations) -> F,
4946 0 : F: std::future::Future<Output = R>,
4947 0 : {
4948 0 : let mutation_locations = {
4949 0 : let mut locations = TenantMutationLocations::default();
4950 :
4951 : // Load the currently attached pageservers for the latest generation of each shard. This can
4952 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
4953 : // will still be the latest when we're done: we will check generations again at the end of
4954 : // this function to handle that.
4955 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
4956 :
4957 0 : if generations
4958 0 : .iter()
4959 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
4960 : {
4961 0 : let shard_generations = generations
4962 0 : .into_iter()
4963 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
4964 0 : .collect::<HashMap<_, _>>();
4965 :
4966 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
4967 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
4968 0 : let locked = self.inner.read().unwrap();
4969 0 : for (shard_id, shard) in
4970 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4971 : {
4972 0 : match shard.policy {
4973 : PlacementPolicy::Attached(_) => {
4974 : // This shard is meant to be attached: the caller is not wrong to try and
4975 : // use this function, but we can't service the request right now.
4976 0 : let Some(generation) = shard_generations.get(shard_id) else {
4977 : // This can only happen if there is a split brain controller modifying the database. This should
4978 : // never happen when testing, and if it happens in production we can only log the issue.
4979 0 : debug_assert!(false);
4980 0 : tracing::error!(
4981 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
4982 : );
4983 0 : continue;
4984 : };
4985 0 : let (generation, generation_pageserver) = generation;
4986 0 : if let Some(generation) = generation {
4987 0 : if generation_pageserver.is_none() {
4988 : // This is legitimate only in a very narrow window where the shard was only just configured into
4989 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
4990 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
4991 0 : tracing::warn!(
4992 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
4993 : );
4994 0 : }
4995 : } else {
4996 : // This should never happen: a shard with no generation is only permitted when it was created in some state
4997 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
4998 0 : debug_assert!(false);
4999 0 : tracing::error!(
5000 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
5001 : );
5002 0 : continue;
5003 : }
5004 : }
5005 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
5006 0 : return Err(ApiError::Conflict(format!(
5007 0 : "Shard {shard_id} tenant has policy {:?}",
5008 0 : shard.policy
5009 0 : )));
5010 : }
5011 : }
5012 : }
5013 :
5014 0 : return Err(ApiError::ResourceUnavailable(
5015 0 : "One or more shards in tenant is not yet attached".into(),
5016 0 : ));
5017 0 : }
5018 :
5019 0 : let locked = self.inner.read().unwrap();
5020 : for ShardGenerationState {
5021 0 : tenant_shard_id,
5022 0 : generation,
5023 0 : generation_pageserver,
5024 0 : } in generations
5025 : {
5026 0 : let node_id = generation_pageserver.expect("We checked for None above");
5027 0 : let node = locked
5028 0 : .nodes
5029 0 : .get(&node_id)
5030 0 : .ok_or(ApiError::Conflict(format!(
5031 0 : "Raced with removal of node {node_id}"
5032 0 : )))?;
5033 0 : let generation = generation.expect("Checked above");
5034 :
5035 0 : let tenant = locked.tenants.get(&tenant_shard_id);
5036 :
5037 : // TODO(vlad): Abstract the logic that finds stale attached locations
5038 : // from observed state into a [`Service`] method.
5039 0 : let other_locations = match tenant {
5040 0 : Some(tenant) => {
5041 0 : let mut other = tenant.attached_locations();
5042 0 : let latest_location_index =
5043 0 : other.iter().position(|&l| l == (node.get_id(), generation));
5044 0 : if let Some(idx) = latest_location_index {
5045 0 : other.remove(idx);
5046 0 : }
5047 :
5048 0 : other
5049 : }
5050 0 : None => Vec::default(),
5051 : };
5052 :
5053 0 : let location = ShardMutationLocations {
5054 0 : latest: MutationLocation {
5055 0 : node: node.clone(),
5056 0 : generation,
5057 0 : },
5058 0 : other: other_locations
5059 0 : .into_iter()
5060 0 : .filter_map(|(node_id, generation)| {
5061 0 : let node = locked.nodes.get(&node_id)?;
5062 :
5063 0 : Some(MutationLocation {
5064 0 : node: node.clone(),
5065 0 : generation,
5066 0 : })
5067 0 : })
5068 0 : .collect(),
5069 : };
5070 0 : locations.0.insert(tenant_shard_id, location);
5071 : }
5072 :
5073 0 : locations
5074 : };
5075 :
5076 0 : let result = op(mutation_locations.clone()).await;
5077 :
5078 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
5079 : // our remote operation executed on the latest generation and is therefore persistent.
5080 : {
5081 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
5082 0 : if latest_generations
5083 0 : .into_iter()
5084 0 : .map(
5085 : |ShardGenerationState {
5086 : tenant_shard_id,
5087 : generation,
5088 : generation_pageserver: _,
5089 0 : }| (tenant_shard_id, generation),
5090 : )
5091 0 : .collect::<Vec<_>>()
5092 0 : != mutation_locations
5093 0 : .0
5094 0 : .into_iter()
5095 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
5096 0 : .collect::<Vec<_>>()
5097 : {
5098 : // We raced with something that incremented the generation, and therefore cannot be
5099 : // confident that our actions are persistent (they might have hit an old generation).
5100 : //
5101 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
5102 0 : return Err(ApiError::ResourceUnavailable(
5103 0 : "Tenant attachment changed, please retry".into(),
5104 0 : ));
5105 0 : }
5106 : }
5107 :
5108 0 : Ok(result)
5109 0 : }
5110 :
5111 0 : pub(crate) async fn tenant_timeline_delete(
5112 0 : self: &Arc<Self>,
5113 0 : tenant_id: TenantId,
5114 0 : timeline_id: TimelineId,
5115 0 : ) -> Result<StatusCode, ApiError> {
5116 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
5117 0 : let _tenant_lock = trace_shared_lock(
5118 0 : &self.tenant_op_locks,
5119 0 : tenant_id,
5120 0 : TenantOperations::TimelineDelete,
5121 0 : )
5122 0 : .await;
5123 :
5124 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
5125 0 : if targets.0.is_empty() {
5126 0 : return Err(ApiError::NotFound(
5127 0 : anyhow::anyhow!("Tenant not found").into(),
5128 0 : ));
5129 0 : }
5130 :
5131 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
5132 0 : assert!(shard_zero_tid.is_shard_zero());
5133 :
5134 0 : async fn delete_one(
5135 0 : tenant_shard_id: TenantShardId,
5136 0 : timeline_id: TimelineId,
5137 0 : node: Node,
5138 0 : http_client: reqwest::Client,
5139 0 : jwt: Option<String>,
5140 0 : ) -> Result<StatusCode, ApiError> {
5141 0 : tracing::info!(
5142 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
5143 : );
5144 :
5145 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
5146 0 : let res = client
5147 0 : .timeline_delete(tenant_shard_id, timeline_id)
5148 0 : .await;
5149 :
5150 0 : match res {
5151 0 : Ok(ok) => Ok(ok),
5152 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
5153 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
5154 0 : Err(e) => {
5155 0 : Err(
5156 0 : ApiError::InternalServerError(anyhow::anyhow!(
5157 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
5158 0 : ))
5159 0 : )
5160 : }
5161 : }
5162 0 : }
5163 :
5164 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
5165 0 : let statuses = self
5166 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
5167 0 : Box::pin(delete_one(
5168 0 : tenant_shard_id,
5169 0 : timeline_id,
5170 0 : node,
5171 0 : self.http_client.clone(),
5172 0 : self.config.pageserver_jwt_token.clone(),
5173 0 : ))
5174 0 : })
5175 0 : .await?;
5176 :
5177 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
5178 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
5179 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
5180 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
5181 0 : return Ok(StatusCode::CONFLICT);
5182 0 : }
5183 :
5184 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
5185 0 : return Ok(StatusCode::ACCEPTED);
5186 0 : }
5187 :
5188 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
5189 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
5190 0 : let shard_zero_status = delete_one(
5191 0 : shard_zero_tid,
5192 0 : timeline_id,
5193 0 : shard_zero_locations.latest.node,
5194 0 : self.http_client.clone(),
5195 0 : self.config.pageserver_jwt_token.clone(),
5196 0 : )
5197 0 : .await?;
5198 0 : Ok(shard_zero_status)
5199 0 : }).await?;
5200 :
5201 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
5202 0 : .await?;
5203 :
5204 0 : status_code
5205 0 : }
5206 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
5207 0 : pub(crate) async fn tenant_shard0_node(
5208 0 : &self,
5209 0 : tenant_id: TenantId,
5210 0 : ) -> Result<(Node, TenantShardId), ApiError> {
5211 0 : let tenant_shard_id = {
5212 0 : let locked = self.inner.read().unwrap();
5213 0 : let Some((tenant_shard_id, _shard)) = locked
5214 0 : .tenants
5215 0 : .range(TenantShardId::tenant_range(tenant_id))
5216 0 : .next()
5217 : else {
5218 0 : return Err(ApiError::NotFound(
5219 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
5220 0 : ));
5221 : };
5222 :
5223 0 : *tenant_shard_id
5224 : };
5225 :
5226 0 : self.tenant_shard_node(tenant_shard_id)
5227 0 : .await
5228 0 : .map(|node| (node, tenant_shard_id))
5229 0 : }
5230 :
5231 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
5232 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
5233 0 : pub(crate) async fn tenant_shard_node(
5234 0 : &self,
5235 0 : tenant_shard_id: TenantShardId,
5236 0 : ) -> Result<Node, ApiError> {
5237 : // Look up in-memory state and maybe use the node from there.
5238 : {
5239 0 : let locked = self.inner.read().unwrap();
5240 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5241 0 : return Err(ApiError::NotFound(
5242 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
5243 0 : ));
5244 : };
5245 :
5246 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
5247 0 : tracing::warn!(
5248 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
5249 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
5250 : shard.policy
5251 : );
5252 0 : return Err(ApiError::Conflict(
5253 0 : "Cannot call timeline API on non-attached tenant".to_string(),
5254 0 : ));
5255 : };
5256 :
5257 0 : if shard.reconciler.is_none() {
5258 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
5259 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
5260 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
5261 : // This should never happen
5262 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5263 0 : "Shard refers to nonexistent node"
5264 0 : )));
5265 : };
5266 0 : return Ok(node.clone());
5267 0 : }
5268 : };
5269 :
5270 : // Look up the latest attached pageserver location from the database
5271 : // generation state: this will reflect the progress of any ongoing migration.
5272 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
5273 : // the case where they call through to the pageserver and get a 404.
5274 0 : let db_result = self
5275 0 : .persistence
5276 0 : .tenant_generations(tenant_shard_id.tenant_id)
5277 0 : .await?;
5278 : let Some(ShardGenerationState {
5279 : tenant_shard_id: _,
5280 : generation: _,
5281 0 : generation_pageserver: Some(node_id),
5282 0 : }) = db_result
5283 0 : .into_iter()
5284 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
5285 : else {
5286 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
5287 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
5288 : // or a conflict response (case where tenant was detached in background)
5289 0 : return Err(ApiError::ResourceUnavailable(
5290 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
5291 0 : ));
5292 : };
5293 0 : let locked = self.inner.read().unwrap();
5294 0 : let Some(node) = locked.nodes.get(&node_id) else {
5295 : // This should never happen
5296 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5297 0 : "Shard refers to nonexistent node"
5298 0 : )));
5299 : };
5300 :
5301 0 : Ok(node.clone())
5302 0 : }
5303 :
5304 0 : pub(crate) fn tenant_locate(
5305 0 : &self,
5306 0 : tenant_id: TenantId,
5307 0 : ) -> Result<TenantLocateResponse, ApiError> {
5308 0 : let locked = self.inner.read().unwrap();
5309 0 : tracing::info!("Locating shards for tenant {tenant_id}");
5310 :
5311 0 : let mut result = Vec::new();
5312 0 : let mut shard_params: Option<ShardParameters> = None;
5313 :
5314 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5315 : {
5316 0 : let node_id =
5317 0 : shard
5318 0 : .intent
5319 0 : .get_attached()
5320 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
5321 0 : "Cannot locate a tenant that is not attached"
5322 0 : )))?;
5323 :
5324 0 : let node = locked
5325 0 : .nodes
5326 0 : .get(&node_id)
5327 0 : .expect("Pageservers may not be deleted while referenced");
5328 :
5329 0 : result.push(node.shard_location(*tenant_shard_id));
5330 :
5331 0 : match &shard_params {
5332 0 : None => {
5333 0 : shard_params = Some(ShardParameters {
5334 0 : stripe_size: shard.shard.stripe_size,
5335 0 : count: shard.shard.count,
5336 0 : });
5337 0 : }
5338 0 : Some(params) => {
5339 0 : if params.stripe_size != shard.shard.stripe_size {
5340 : // This should never happen. We enforce at runtime because it's simpler than
5341 : // adding an extra per-tenant data structure to store the things that should be the same
5342 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5343 0 : "Inconsistent shard stripe size parameters!"
5344 0 : )));
5345 0 : }
5346 : }
5347 : }
5348 : }
5349 :
5350 0 : if result.is_empty() {
5351 0 : return Err(ApiError::NotFound(
5352 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
5353 0 : ));
5354 0 : }
5355 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
5356 0 : tracing::info!(
5357 0 : "Located tenant {} with params {:?} on shards {}",
5358 : tenant_id,
5359 : shard_params,
5360 0 : result
5361 0 : .iter()
5362 0 : .map(|s| format!("{s:?}"))
5363 0 : .collect::<Vec<_>>()
5364 0 : .join(",")
5365 : );
5366 :
5367 0 : Ok(TenantLocateResponse {
5368 0 : shards: result,
5369 0 : shard_params,
5370 0 : })
5371 0 : }
5372 :
5373 : /// Returns None if the input iterator of shards does not include a shard with number=0
5374 0 : fn tenant_describe_impl<'a>(
5375 0 : &self,
5376 0 : shards: impl Iterator<Item = &'a TenantShard>,
5377 0 : ) -> Option<TenantDescribeResponse> {
5378 0 : let mut shard_zero = None;
5379 0 : let mut describe_shards = Vec::new();
5380 :
5381 0 : for shard in shards {
5382 0 : if shard.tenant_shard_id.is_shard_zero() {
5383 0 : shard_zero = Some(shard);
5384 0 : }
5385 :
5386 0 : describe_shards.push(TenantDescribeResponseShard {
5387 0 : tenant_shard_id: shard.tenant_shard_id,
5388 0 : node_attached: *shard.intent.get_attached(),
5389 0 : node_secondary: shard.intent.get_secondary().to_vec(),
5390 0 : last_error: shard
5391 0 : .last_error
5392 0 : .lock()
5393 0 : .unwrap()
5394 0 : .as_ref()
5395 0 : .map(|e| format!("{e}"))
5396 0 : .unwrap_or("".to_string())
5397 0 : .clone(),
5398 0 : is_reconciling: shard.reconciler.is_some(),
5399 0 : is_pending_compute_notification: shard.pending_compute_notification,
5400 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
5401 0 : is_importing: shard.importing == TimelineImportState::Importing,
5402 0 : scheduling_policy: shard.get_scheduling_policy(),
5403 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
5404 : })
5405 : }
5406 :
5407 0 : let shard_zero = shard_zero?;
5408 :
5409 0 : Some(TenantDescribeResponse {
5410 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
5411 0 : shards: describe_shards,
5412 0 : stripe_size: shard_zero.shard.stripe_size,
5413 0 : policy: shard_zero.policy.clone(),
5414 0 : config: shard_zero.config.clone(),
5415 0 : })
5416 0 : }
5417 :
5418 0 : pub(crate) fn tenant_describe(
5419 0 : &self,
5420 0 : tenant_id: TenantId,
5421 0 : ) -> Result<TenantDescribeResponse, ApiError> {
5422 0 : let locked = self.inner.read().unwrap();
5423 :
5424 0 : self.tenant_describe_impl(
5425 0 : locked
5426 0 : .tenants
5427 0 : .range(TenantShardId::tenant_range(tenant_id))
5428 0 : .map(|(_k, v)| v),
5429 : )
5430 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
5431 0 : }
5432 :
5433 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
5434 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
5435 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
5436 : /// in our external API.
5437 0 : pub(crate) fn tenant_list(
5438 0 : &self,
5439 0 : limit: Option<usize>,
5440 0 : start_after: Option<TenantId>,
5441 0 : ) -> Vec<TenantDescribeResponse> {
5442 0 : let locked = self.inner.read().unwrap();
5443 :
5444 : // Apply start_from parameter
5445 0 : let shard_range = match start_after {
5446 0 : None => locked.tenants.range(..),
5447 0 : Some(tenant_id) => locked.tenants.range(
5448 0 : TenantShardId {
5449 0 : tenant_id,
5450 0 : shard_number: ShardNumber(u8::MAX),
5451 0 : shard_count: ShardCount(u8::MAX),
5452 0 : }..,
5453 : ),
5454 : };
5455 :
5456 0 : let mut result = Vec::new();
5457 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
5458 0 : result.push(
5459 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
5460 0 : .expect("Groups are always non-empty"),
5461 : );
5462 :
5463 : // Enforce `limit` parameter
5464 0 : if let Some(limit) = limit {
5465 0 : if result.len() >= limit {
5466 0 : break;
5467 0 : }
5468 0 : }
5469 : }
5470 :
5471 0 : result
5472 0 : }
5473 :
5474 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
5475 : async fn abort_tenant_shard_split(
5476 : &self,
5477 : op: &TenantShardSplitAbort,
5478 : ) -> Result<(), TenantShardSplitAbortError> {
5479 : // Cleaning up a split:
5480 : // - Parent shards are not destroyed during a split, just detached.
5481 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
5482 : // just the children attached, or both.
5483 : //
5484 : // Therefore our work to do is to:
5485 : // 1. Clean up storage controller's internal state to just refer to parents, no children
5486 : // 2. Call out to pageservers to ensure that children are detached
5487 : // 3. Call out to pageservers to ensure that parents are attached.
5488 : //
5489 : // Crash safety:
5490 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
5491 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
5492 : // and detach them.
5493 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
5494 : // from our database, then we will re-enter this cleanup routine on startup.
5495 :
5496 : let TenantShardSplitAbort {
5497 : tenant_id,
5498 : new_shard_count,
5499 : new_stripe_size,
5500 : ..
5501 : } = op;
5502 :
5503 : // First abort persistent state, if any exists.
5504 : match self
5505 : .persistence
5506 : .abort_shard_split(*tenant_id, *new_shard_count)
5507 : .await?
5508 : {
5509 : AbortShardSplitStatus::Aborted => {
5510 : // Proceed to roll back any child shards created on pageservers
5511 : }
5512 : AbortShardSplitStatus::Complete => {
5513 : // The split completed (we might hit that path if e.g. our database transaction
5514 : // to write the completion landed in the database, but we dropped connection
5515 : // before seeing the result).
5516 : //
5517 : // We must update in-memory state to reflect the successful split.
5518 : self.tenant_shard_split_commit_inmem(
5519 : *tenant_id,
5520 : *new_shard_count,
5521 : *new_stripe_size,
5522 : );
5523 : return Ok(());
5524 : }
5525 : }
5526 :
5527 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
5528 : let detach_locations: Vec<(Node, TenantShardId)> = {
5529 : let mut detach_locations = Vec::new();
5530 : let mut locked = self.inner.write().unwrap();
5531 : let (nodes, tenants, scheduler) = locked.parts_mut();
5532 :
5533 : for (tenant_shard_id, shard) in
5534 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
5535 : {
5536 : if shard.shard.count == op.new_shard_count {
5537 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
5538 : // is infallible, so if we got an error we shouldn't have got that far.
5539 : tracing::warn!(
5540 : "During split abort, child shard {tenant_shard_id} found in-memory"
5541 : );
5542 : continue;
5543 : }
5544 :
5545 : // Add the children of this shard to this list of things to detach
5546 : if let Some(node_id) = shard.intent.get_attached() {
5547 : for child_id in tenant_shard_id.split(*new_shard_count) {
5548 : detach_locations.push((
5549 : nodes
5550 : .get(node_id)
5551 : .expect("Intent references nonexistent node")
5552 : .clone(),
5553 : child_id,
5554 : ));
5555 : }
5556 : } else {
5557 : tracing::warn!(
5558 : "During split abort, shard {tenant_shard_id} has no attached location"
5559 : );
5560 : }
5561 :
5562 : tracing::info!("Restoring parent shard {tenant_shard_id}");
5563 :
5564 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
5565 : // if the original attachment location is offline.
5566 : if let Some(node_id) = shard.intent.get_attached() {
5567 : if !nodes.get(node_id).unwrap().is_available() {
5568 : tracing::info!(
5569 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
5570 : );
5571 : shard.intent.demote_attached(scheduler, *node_id);
5572 : }
5573 : }
5574 : for node_id in shard.intent.get_secondary().clone() {
5575 : if !nodes.get(&node_id).unwrap().is_available() {
5576 : tracing::info!(
5577 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
5578 : );
5579 : shard.intent.remove_secondary(scheduler, node_id);
5580 : }
5581 : }
5582 :
5583 : shard.splitting = SplitState::Idle;
5584 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
5585 : // If this shard can't be scheduled now (perhaps due to offline nodes or
5586 : // capacity issues), that must not prevent us rolling back a split. In this
5587 : // case it should be eventually scheduled in the background.
5588 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
5589 : }
5590 :
5591 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
5592 : }
5593 :
5594 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
5595 : tenants
5596 0 : .retain(|id, s| !(id.tenant_id == *tenant_id && s.shard.count == *new_shard_count));
5597 :
5598 : detach_locations
5599 : };
5600 :
5601 : for (node, child_id) in detach_locations {
5602 : if !node.is_available() {
5603 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
5604 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
5605 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
5606 : // them from the node.
5607 : tracing::warn!(
5608 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
5609 : );
5610 : continue;
5611 : }
5612 :
5613 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
5614 : // a 503 and retry, up to our limit.
5615 : tracing::info!("Detaching {child_id} on {node}...");
5616 : match node
5617 : .with_client_retries(
5618 0 : |client| async move {
5619 0 : let config = LocationConfig {
5620 0 : mode: LocationConfigMode::Detached,
5621 0 : generation: None,
5622 0 : secondary_conf: None,
5623 0 : shard_number: child_id.shard_number.0,
5624 0 : shard_count: child_id.shard_count.literal(),
5625 0 : // Stripe size and tenant config don't matter when detaching
5626 0 : shard_stripe_size: 0,
5627 0 : tenant_conf: TenantConfig::default(),
5628 0 : };
5629 :
5630 0 : client.location_config(child_id, config, None, false).await
5631 0 : },
5632 : &self.http_client,
5633 : &self.config.pageserver_jwt_token,
5634 : 1,
5635 : 10,
5636 : Duration::from_secs(5),
5637 : &self.reconcilers_cancel,
5638 : )
5639 : .await
5640 : {
5641 : Some(Ok(_)) => {}
5642 : Some(Err(e)) => {
5643 : // We failed to communicate with the remote node. This is problematic: we may be
5644 : // leaving it with a rogue child shard.
5645 : tracing::warn!(
5646 : "Failed to detach child {child_id} from node {node} during abort"
5647 : );
5648 : return Err(e.into());
5649 : }
5650 : None => {
5651 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
5652 : // clean up on restart. The node going offline requires a retry.
5653 : return Err(TenantShardSplitAbortError::Unavailable);
5654 : }
5655 : };
5656 : }
5657 :
5658 : tracing::info!("Successfully aborted split");
5659 : Ok(())
5660 : }
5661 :
5662 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
5663 : /// of the tenant map to reflect the child shards that exist after the split.
5664 0 : fn tenant_shard_split_commit_inmem(
5665 0 : &self,
5666 0 : tenant_id: TenantId,
5667 0 : new_shard_count: ShardCount,
5668 0 : new_stripe_size: Option<ShardStripeSize>,
5669 0 : ) -> (
5670 0 : TenantShardSplitResponse,
5671 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
5672 0 : Vec<ReconcilerWaiter>,
5673 0 : ) {
5674 0 : let mut response = TenantShardSplitResponse {
5675 0 : new_shards: Vec::new(),
5676 0 : };
5677 0 : let mut child_locations = Vec::new();
5678 0 : let mut waiters = Vec::new();
5679 :
5680 : {
5681 0 : let mut locked = self.inner.write().unwrap();
5682 :
5683 0 : let parent_ids = locked
5684 0 : .tenants
5685 0 : .range(TenantShardId::tenant_range(tenant_id))
5686 0 : .map(|(shard_id, _)| *shard_id)
5687 0 : .collect::<Vec<_>>();
5688 :
5689 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5690 0 : for parent_id in parent_ids {
5691 0 : let child_ids = parent_id.split(new_shard_count);
5692 :
5693 : let (
5694 0 : pageserver,
5695 0 : generation,
5696 0 : policy,
5697 0 : parent_ident,
5698 0 : config,
5699 0 : preferred_az,
5700 0 : secondary_count,
5701 : ) = {
5702 0 : let mut old_state = tenants
5703 0 : .remove(&parent_id)
5704 0 : .expect("It was present, we just split it");
5705 :
5706 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
5707 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
5708 : // nothing else can clear this.
5709 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
5710 :
5711 0 : let old_attached = old_state.intent.get_attached().unwrap();
5712 0 : old_state.intent.clear(scheduler);
5713 0 : let generation = old_state.generation.expect("Shard must have been attached");
5714 0 : (
5715 0 : old_attached,
5716 0 : generation,
5717 0 : old_state.policy.clone(),
5718 0 : old_state.shard,
5719 0 : old_state.config.clone(),
5720 0 : old_state.preferred_az().cloned(),
5721 0 : old_state.intent.get_secondary().len(),
5722 0 : )
5723 : };
5724 :
5725 0 : let mut schedule_context = ScheduleContext::default();
5726 0 : for child in child_ids {
5727 0 : let mut child_shard = parent_ident;
5728 0 : child_shard.number = child.shard_number;
5729 0 : child_shard.count = child.shard_count;
5730 0 : if let Some(stripe_size) = new_stripe_size {
5731 0 : child_shard.stripe_size = stripe_size;
5732 0 : }
5733 :
5734 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5735 0 : child_observed.insert(
5736 0 : pageserver,
5737 0 : ObservedStateLocation {
5738 0 : conf: Some(attached_location_conf(
5739 0 : generation,
5740 0 : &child_shard,
5741 0 : &config,
5742 0 : &policy,
5743 0 : secondary_count,
5744 0 : )),
5745 0 : },
5746 : );
5747 :
5748 0 : let mut child_state =
5749 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
5750 0 : child_state.intent =
5751 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
5752 0 : child_state.observed = ObservedState {
5753 0 : locations: child_observed,
5754 0 : };
5755 0 : child_state.generation = Some(generation);
5756 0 : child_state.config = config.clone();
5757 :
5758 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
5759 : // as at this point in the split process we have succeeded and this part is infallible:
5760 : // we will never need to do any special recovery from this state.
5761 :
5762 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
5763 :
5764 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
5765 : // This is not fatal, because we've implicitly already got an attached
5766 : // location for the child shard. Failure here just means we couldn't
5767 : // find a secondary (e.g. because cluster is overloaded).
5768 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
5769 0 : }
5770 : // In the background, attach secondary locations for the new shards
5771 0 : if let Some(waiter) = self.maybe_reconcile_shard(
5772 0 : &mut child_state,
5773 0 : nodes,
5774 0 : ReconcilerPriority::High,
5775 0 : ) {
5776 0 : waiters.push(waiter);
5777 0 : }
5778 :
5779 0 : tenants.insert(child, child_state);
5780 0 : response.new_shards.push(child);
5781 : }
5782 : }
5783 0 : (response, child_locations, waiters)
5784 : }
5785 0 : }
5786 :
5787 0 : async fn tenant_shard_split_start_secondaries(
5788 0 : &self,
5789 0 : tenant_id: TenantId,
5790 0 : waiters: Vec<ReconcilerWaiter>,
5791 0 : ) {
5792 : // Wait for initial reconcile of child shards, this creates the secondary locations
5793 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5794 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
5795 : // their secondaries couldn't be attached.
5796 0 : tracing::warn!("Failed to reconcile after split: {e}");
5797 0 : return;
5798 0 : }
5799 :
5800 : // Take the state lock to discover the attached & secondary intents for all shards
5801 0 : let (attached, secondary) = {
5802 0 : let locked = self.inner.read().unwrap();
5803 0 : let mut attached = Vec::new();
5804 0 : let mut secondary = Vec::new();
5805 :
5806 0 : for (tenant_shard_id, shard) in
5807 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5808 : {
5809 0 : let Some(node_id) = shard.intent.get_attached() else {
5810 : // Unexpected. Race with a PlacementPolicy change?
5811 0 : tracing::warn!(
5812 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
5813 : );
5814 0 : continue;
5815 : };
5816 :
5817 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
5818 : // No secondary location. Nothing for us to do.
5819 0 : continue;
5820 : };
5821 :
5822 0 : let attached_node = locked
5823 0 : .nodes
5824 0 : .get(node_id)
5825 0 : .expect("Pageservers may not be deleted while referenced");
5826 :
5827 0 : let secondary_node = locked
5828 0 : .nodes
5829 0 : .get(secondary_node_id)
5830 0 : .expect("Pageservers may not be deleted while referenced");
5831 :
5832 0 : attached.push((*tenant_shard_id, attached_node.clone()));
5833 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
5834 : }
5835 0 : (attached, secondary)
5836 : };
5837 :
5838 0 : if secondary.is_empty() {
5839 : // No secondary locations; nothing for us to do
5840 0 : return;
5841 0 : }
5842 :
5843 0 : for result in self
5844 0 : .tenant_for_shards_api(
5845 0 : attached,
5846 0 : |tenant_shard_id, client| async move {
5847 0 : client.tenant_heatmap_upload(tenant_shard_id).await
5848 0 : },
5849 : 1,
5850 : 1,
5851 : SHORT_RECONCILE_TIMEOUT,
5852 0 : &self.cancel,
5853 : )
5854 0 : .await
5855 : {
5856 0 : if let Err(e) = result {
5857 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
5858 0 : return;
5859 0 : }
5860 : }
5861 :
5862 0 : for result in self
5863 0 : .tenant_for_shards_api(
5864 0 : secondary,
5865 0 : |tenant_shard_id, client| async move {
5866 0 : client
5867 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
5868 0 : .await
5869 0 : },
5870 : 1,
5871 : 1,
5872 : SHORT_RECONCILE_TIMEOUT,
5873 0 : &self.cancel,
5874 : )
5875 0 : .await
5876 : {
5877 0 : if let Err(e) = result {
5878 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
5879 0 : return;
5880 0 : }
5881 : }
5882 0 : }
5883 :
5884 0 : pub(crate) async fn tenant_shard_split(
5885 0 : &self,
5886 0 : tenant_id: TenantId,
5887 0 : split_req: TenantShardSplitRequest,
5888 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
5889 : // TODO: return 503 if we get stuck waiting for this lock
5890 : // (issue https://github.com/neondatabase/neon/issues/7108)
5891 0 : let _tenant_lock = trace_exclusive_lock(
5892 0 : &self.tenant_op_locks,
5893 0 : tenant_id,
5894 0 : TenantOperations::ShardSplit,
5895 0 : )
5896 0 : .await;
5897 :
5898 0 : let _gate = self
5899 0 : .reconcilers_gate
5900 0 : .enter()
5901 0 : .map_err(|_| ApiError::ShuttingDown)?;
5902 :
5903 : // Timeline imports on the pageserver side can't handle shard-splits.
5904 : // If the tenant is importing a timeline, dont't shard split it.
5905 0 : match self
5906 0 : .persistence
5907 0 : .is_tenant_importing_timeline(tenant_id)
5908 0 : .await
5909 : {
5910 0 : Ok(importing) => {
5911 0 : if importing {
5912 0 : return Err(ApiError::Conflict(
5913 0 : "Cannot shard split during timeline import".to_string(),
5914 0 : ));
5915 0 : }
5916 : }
5917 0 : Err(err) => {
5918 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5919 0 : "Failed to check for running imports: {err}"
5920 0 : )));
5921 : }
5922 : }
5923 :
5924 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
5925 0 : let new_stripe_size = split_req.new_stripe_size;
5926 :
5927 : // Validate the request and construct parameters. This phase is fallible, but does not require
5928 : // rollback on errors, as it does no I/O and mutates no state.
5929 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
5930 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
5931 0 : ShardSplitAction::Split(params) => params,
5932 : };
5933 :
5934 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
5935 : // we must roll back.
5936 0 : let r = self
5937 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
5938 0 : .await;
5939 :
5940 0 : let (response, waiters) = match r {
5941 0 : Ok(r) => r,
5942 0 : Err(e) => {
5943 : // Split might be part-done, we must do work to abort it.
5944 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
5945 0 : self.abort_tx
5946 0 : .send(TenantShardSplitAbort {
5947 0 : tenant_id,
5948 0 : new_shard_count,
5949 0 : new_stripe_size,
5950 0 : _tenant_lock,
5951 0 : _gate,
5952 0 : })
5953 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
5954 0 : .ok();
5955 0 : return Err(e);
5956 : }
5957 : };
5958 :
5959 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
5960 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
5961 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
5962 : // in [`Self::optimize_all`]
5963 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
5964 0 : .await;
5965 0 : Ok(response)
5966 0 : }
5967 :
5968 0 : fn prepare_tenant_shard_split(
5969 0 : &self,
5970 0 : tenant_id: TenantId,
5971 0 : split_req: TenantShardSplitRequest,
5972 0 : ) -> Result<ShardSplitAction, ApiError> {
5973 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
5974 0 : anyhow::anyhow!("failpoint")
5975 0 : )));
5976 :
5977 0 : let mut policy = None;
5978 0 : let mut config = None;
5979 0 : let mut shard_ident = None;
5980 0 : let mut preferred_az_id = None;
5981 : // Validate input, and calculate which shards we will create
5982 0 : let (old_shard_count, targets) =
5983 : {
5984 0 : let locked = self.inner.read().unwrap();
5985 :
5986 0 : let pageservers = locked.nodes.clone();
5987 :
5988 0 : let mut targets = Vec::new();
5989 :
5990 : // In case this is a retry, count how many already-split shards we found
5991 0 : let mut children_found = Vec::new();
5992 0 : let mut old_shard_count = None;
5993 :
5994 0 : for (tenant_shard_id, shard) in
5995 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5996 : {
5997 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
5998 : Ordering::Equal => {
5999 : // Already split this
6000 0 : children_found.push(*tenant_shard_id);
6001 0 : continue;
6002 : }
6003 : Ordering::Greater => {
6004 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6005 0 : "Requested count {} but already have shards at count {}",
6006 0 : split_req.new_shard_count,
6007 0 : shard.shard.count.count()
6008 0 : )));
6009 : }
6010 0 : Ordering::Less => {
6011 0 : // Fall through: this shard has lower count than requested,
6012 0 : // is a candidate for splitting.
6013 0 : }
6014 : }
6015 :
6016 0 : match old_shard_count {
6017 0 : None => old_shard_count = Some(shard.shard.count),
6018 0 : Some(old_shard_count) => {
6019 0 : if old_shard_count != shard.shard.count {
6020 : // We may hit this case if a caller asked for two splits to
6021 : // different sizes, before the first one is complete.
6022 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
6023 : // of shard_count=1 and shard_count=2 shards in the map.
6024 0 : return Err(ApiError::Conflict(
6025 0 : "Cannot split, currently mid-split".to_string(),
6026 0 : ));
6027 0 : }
6028 : }
6029 : }
6030 0 : if policy.is_none() {
6031 0 : policy = Some(shard.policy.clone());
6032 0 : }
6033 0 : if shard_ident.is_none() {
6034 0 : shard_ident = Some(shard.shard);
6035 0 : }
6036 0 : if config.is_none() {
6037 0 : config = Some(shard.config.clone());
6038 0 : }
6039 0 : if preferred_az_id.is_none() {
6040 0 : preferred_az_id = shard.preferred_az().cloned();
6041 0 : }
6042 :
6043 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
6044 0 : tracing::info!(
6045 0 : "Tenant shard {} already has shard count {}",
6046 : tenant_shard_id,
6047 : split_req.new_shard_count
6048 : );
6049 0 : continue;
6050 0 : }
6051 :
6052 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
6053 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
6054 0 : ))?;
6055 :
6056 0 : let node = pageservers
6057 0 : .get(&node_id)
6058 0 : .expect("Pageservers may not be deleted while referenced");
6059 :
6060 0 : targets.push(ShardSplitTarget {
6061 0 : parent_id: *tenant_shard_id,
6062 0 : node: node.clone(),
6063 0 : child_ids: tenant_shard_id
6064 0 : .split(ShardCount::new(split_req.new_shard_count)),
6065 0 : });
6066 : }
6067 :
6068 0 : if targets.is_empty() {
6069 0 : if children_found.len() == split_req.new_shard_count as usize {
6070 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
6071 0 : new_shards: children_found,
6072 0 : }));
6073 : } else {
6074 : // No shards found to split, and no existing children found: the
6075 : // tenant doesn't exist at all.
6076 0 : return Err(ApiError::NotFound(
6077 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
6078 0 : ));
6079 : }
6080 0 : }
6081 :
6082 0 : (old_shard_count, targets)
6083 : };
6084 :
6085 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
6086 0 : let old_shard_count = old_shard_count.unwrap();
6087 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
6088 : // This ShardIdentity will be used as the template for all children, so this implicitly
6089 : // applies the new stripe size to the children.
6090 0 : let mut shard_ident = shard_ident.unwrap();
6091 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
6092 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6093 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
6094 0 : shard_ident.stripe_size
6095 0 : )));
6096 0 : }
6097 :
6098 0 : shard_ident.stripe_size = new_stripe_size;
6099 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
6100 0 : shard_ident
6101 : } else {
6102 0 : shard_ident.unwrap()
6103 : };
6104 0 : let policy = policy.unwrap();
6105 0 : let config = config.unwrap();
6106 :
6107 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
6108 0 : old_shard_count,
6109 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
6110 0 : new_stripe_size: split_req.new_stripe_size,
6111 0 : targets,
6112 0 : policy,
6113 0 : config,
6114 0 : shard_ident,
6115 0 : preferred_az_id,
6116 0 : })))
6117 0 : }
6118 :
6119 0 : async fn do_tenant_shard_split(
6120 0 : &self,
6121 0 : tenant_id: TenantId,
6122 0 : params: Box<ShardSplitParams>,
6123 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
6124 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
6125 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
6126 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
6127 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
6128 : // (https://github.com/neondatabase/neon/issues/6676)
6129 :
6130 : let ShardSplitParams {
6131 0 : old_shard_count,
6132 0 : new_shard_count,
6133 0 : new_stripe_size,
6134 0 : mut targets,
6135 0 : policy,
6136 0 : config,
6137 0 : shard_ident,
6138 0 : preferred_az_id,
6139 0 : } = *params;
6140 :
6141 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
6142 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
6143 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
6144 : // at the time of split.
6145 0 : let waiters = {
6146 0 : let mut locked = self.inner.write().unwrap();
6147 0 : let mut waiters = Vec::new();
6148 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6149 0 : for target in &mut targets {
6150 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
6151 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6152 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6153 0 : "Shard {} not found",
6154 0 : target.parent_id
6155 0 : )));
6156 : };
6157 :
6158 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
6159 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
6160 0 : return Err(ApiError::Conflict(format!(
6161 0 : "Shard {} unexpectedly rescheduled during split",
6162 0 : target.parent_id
6163 0 : )));
6164 0 : }
6165 :
6166 : // Irrespective of PlacementPolicy, clear secondary locations from intent
6167 0 : shard.intent.clear_secondary(scheduler);
6168 :
6169 : // Run Reconciler to execute detach fo secondary locations.
6170 0 : if let Some(waiter) =
6171 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6172 0 : {
6173 0 : waiters.push(waiter);
6174 0 : }
6175 : }
6176 0 : waiters
6177 : };
6178 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
6179 :
6180 : // Before creating any new child shards in memory or on the pageservers, persist them: this
6181 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
6182 : // acts as the protection against two concurrent attempts to split: one of them will get a database
6183 : // error trying to insert the child shards.
6184 0 : let mut child_tsps = Vec::new();
6185 0 : for target in &targets {
6186 0 : let mut this_child_tsps = Vec::new();
6187 0 : for child in &target.child_ids {
6188 0 : let mut child_shard = shard_ident;
6189 0 : child_shard.number = child.shard_number;
6190 0 : child_shard.count = child.shard_count;
6191 :
6192 0 : tracing::info!(
6193 0 : "Create child shard persistence with stripe size {}",
6194 : shard_ident.stripe_size.0
6195 : );
6196 :
6197 0 : this_child_tsps.push(TenantShardPersistence {
6198 0 : tenant_id: child.tenant_id.to_string(),
6199 0 : shard_number: child.shard_number.0 as i32,
6200 0 : shard_count: child.shard_count.literal() as i32,
6201 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
6202 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
6203 : // populate the correct generation as part of its transaction, to protect us
6204 : // against racing with changes in the state of the parent.
6205 0 : generation: None,
6206 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
6207 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
6208 0 : config: serde_json::to_string(&config).unwrap(),
6209 0 : splitting: SplitState::Splitting,
6210 :
6211 : // Scheduling policies and preferred AZ do not carry through to children
6212 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
6213 0 : .unwrap(),
6214 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
6215 : });
6216 : }
6217 :
6218 0 : child_tsps.push((target.parent_id, this_child_tsps));
6219 : }
6220 :
6221 0 : if let Err(e) = self
6222 0 : .persistence
6223 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
6224 0 : .await
6225 : {
6226 0 : match e {
6227 : DatabaseError::Query(diesel::result::Error::DatabaseError(
6228 : DatabaseErrorKind::UniqueViolation,
6229 : _,
6230 : )) => {
6231 : // Inserting a child shard violated a unique constraint: we raced with another call to
6232 : // this function
6233 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
6234 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
6235 : }
6236 0 : _ => return Err(ApiError::InternalServerError(e.into())),
6237 : }
6238 0 : }
6239 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
6240 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6241 : ));
6242 :
6243 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
6244 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
6245 : // is not set in memory, then it was not persisted.
6246 : {
6247 0 : let mut locked = self.inner.write().unwrap();
6248 0 : for target in &targets {
6249 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
6250 0 : parent_shard.splitting = SplitState::Splitting;
6251 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
6252 0 : // split operation.
6253 0 : parent_shard
6254 0 : .observed
6255 0 : .locations
6256 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
6257 0 : }
6258 : }
6259 : }
6260 :
6261 : // TODO: issue split calls concurrently (this only matters once we're splitting
6262 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
6263 :
6264 0 : for target in &targets {
6265 : let ShardSplitTarget {
6266 0 : parent_id,
6267 0 : node,
6268 0 : child_ids,
6269 0 : } = target;
6270 0 : let client = PageserverClient::new(
6271 0 : node.get_id(),
6272 0 : self.http_client.clone(),
6273 0 : node.base_url(),
6274 0 : self.config.pageserver_jwt_token.as_deref(),
6275 : );
6276 0 : let response = client
6277 0 : .tenant_shard_split(
6278 0 : *parent_id,
6279 0 : TenantShardSplitRequest {
6280 0 : new_shard_count: new_shard_count.literal(),
6281 0 : new_stripe_size,
6282 0 : },
6283 0 : )
6284 0 : .await
6285 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {parent_id}: {e}")))?;
6286 :
6287 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
6288 0 : "failpoint".to_string()
6289 0 : )));
6290 :
6291 0 : failpoint_support::sleep_millis_async!(
6292 : "shard-split-post-remote-sleep",
6293 0 : &self.reconcilers_cancel
6294 : );
6295 :
6296 0 : tracing::info!(
6297 0 : "Split {} into {}",
6298 : parent_id,
6299 0 : response
6300 0 : .new_shards
6301 0 : .iter()
6302 0 : .map(|s| format!("{s:?}"))
6303 0 : .collect::<Vec<_>>()
6304 0 : .join(",")
6305 : );
6306 :
6307 0 : if &response.new_shards != child_ids {
6308 : // This should never happen: the pageserver should agree with us on how shard splits work.
6309 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6310 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
6311 0 : parent_id,
6312 0 : response.new_shards,
6313 0 : child_ids
6314 0 : )));
6315 0 : }
6316 : }
6317 :
6318 0 : fail::fail_point!("shard-split-pre-complete", |_| Err(ApiError::Conflict(
6319 0 : "failpoint".to_string()
6320 0 : )));
6321 :
6322 0 : pausable_failpoint!("shard-split-pre-complete-pause");
6323 :
6324 : // TODO: if the pageserver restarted concurrently with our split API call,
6325 : // the actual generation of the child shard might differ from the generation
6326 : // we expect it to have. In order for our in-database generation to end up
6327 : // correct, we should carry the child generation back in the response and apply it here
6328 : // in complete_shard_split (and apply the correct generation in memory)
6329 : // (or, we can carry generation in the request and reject the request if
6330 : // it doesn't match, but that requires more retry logic on this side)
6331 :
6332 0 : self.persistence
6333 0 : .complete_shard_split(tenant_id, old_shard_count, new_shard_count)
6334 0 : .await?;
6335 :
6336 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
6337 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
6338 : ));
6339 :
6340 : // Replace all the shards we just split with their children: this phase is infallible.
6341 0 : let (response, child_locations, waiters) =
6342 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
6343 :
6344 : // Notify all page servers to detach and clean up the old shards because they will no longer
6345 : // be needed. This is best-effort: if it fails, it will be cleaned up on a subsequent
6346 : // Pageserver re-attach/startup.
6347 0 : let shards_to_cleanup = targets
6348 0 : .iter()
6349 0 : .map(|target| (target.parent_id, target.node.get_id()))
6350 0 : .collect();
6351 0 : self.cleanup_locations(shards_to_cleanup).await;
6352 :
6353 : // Send compute notifications for all the new shards
6354 0 : let mut failed_notifications = Vec::new();
6355 0 : for (child_id, child_ps, stripe_size) in child_locations {
6356 0 : if let Err(e) = self
6357 0 : .compute_hook
6358 0 : .notify_attach(
6359 0 : compute_hook::ShardUpdate {
6360 0 : tenant_shard_id: child_id,
6361 0 : node_id: child_ps,
6362 0 : stripe_size,
6363 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
6364 0 : },
6365 0 : &self.reconcilers_cancel,
6366 0 : )
6367 0 : .await
6368 : {
6369 0 : tracing::warn!(
6370 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
6371 : child_id,
6372 : child_ps
6373 : );
6374 0 : failed_notifications.push(child_id);
6375 0 : }
6376 : }
6377 :
6378 : // If we failed any compute notifications, make a note to retry later.
6379 0 : if !failed_notifications.is_empty() {
6380 0 : let mut locked = self.inner.write().unwrap();
6381 0 : for failed in failed_notifications {
6382 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
6383 0 : shard.pending_compute_notification = true;
6384 0 : }
6385 : }
6386 0 : }
6387 :
6388 0 : Ok((response, waiters))
6389 0 : }
6390 :
6391 : /// A graceful migration: update the preferred node and let optimisation handle the migration
6392 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
6393 : ///
6394 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
6395 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
6396 0 : fn tenant_shard_migrate_with_prewarm(
6397 0 : &self,
6398 0 : migrate_req: &TenantShardMigrateRequest,
6399 0 : shard: &mut TenantShard,
6400 0 : scheduler: &mut Scheduler,
6401 0 : schedule_context: ScheduleContext,
6402 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
6403 0 : shard.set_preferred_node(Some(migrate_req.node_id));
6404 :
6405 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
6406 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
6407 : // e.g. by checking secondary is warm enough.
6408 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
6409 0 : }
6410 :
6411 : /// Immediate migration: directly update the intent state and kick off a reconciler
6412 0 : fn tenant_shard_migrate_immediate(
6413 0 : &self,
6414 0 : migrate_req: &TenantShardMigrateRequest,
6415 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6416 0 : shard: &mut TenantShard,
6417 0 : scheduler: &mut Scheduler,
6418 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
6419 : // Non-graceful migration: update the intent state immediately
6420 0 : let old_attached = *shard.intent.get_attached();
6421 0 : match shard.policy {
6422 0 : PlacementPolicy::Attached(n) => {
6423 : // If our new attached node was a secondary, it no longer should be.
6424 0 : shard
6425 0 : .intent
6426 0 : .remove_secondary(scheduler, migrate_req.node_id);
6427 :
6428 0 : shard
6429 0 : .intent
6430 0 : .set_attached(scheduler, Some(migrate_req.node_id));
6431 :
6432 : // If we were already attached to something, demote that to a secondary
6433 0 : if let Some(old_attached) = old_attached {
6434 0 : if n > 0 {
6435 : // Remove other secondaries to make room for the location we'll demote
6436 0 : while shard.intent.get_secondary().len() >= n {
6437 0 : shard.intent.pop_secondary(scheduler);
6438 0 : }
6439 :
6440 0 : shard.intent.push_secondary(scheduler, old_attached);
6441 0 : }
6442 0 : }
6443 : }
6444 0 : PlacementPolicy::Secondary => {
6445 0 : shard.intent.clear(scheduler);
6446 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6447 0 : }
6448 : PlacementPolicy::Detached => {
6449 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6450 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
6451 0 : )));
6452 : }
6453 : }
6454 :
6455 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
6456 0 : shard.sequence = shard.sequence.next();
6457 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
6458 0 : Ok(self.maybe_configured_reconcile_shard(
6459 0 : shard,
6460 0 : nodes,
6461 0 : (&migrate_req.migration_config).into(),
6462 0 : ))
6463 0 : }
6464 :
6465 0 : pub(crate) async fn tenant_shard_migrate(
6466 0 : &self,
6467 0 : tenant_shard_id: TenantShardId,
6468 0 : migrate_req: TenantShardMigrateRequest,
6469 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6470 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
6471 : // get a different outcome to handle
6472 : enum MigrationOutcome {
6473 : Optimization(Option<ScheduleOptimization>),
6474 : Reconcile(Option<ReconcilerWaiter>),
6475 : }
6476 :
6477 0 : let outcome = {
6478 0 : let mut locked = self.inner.write().unwrap();
6479 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6480 :
6481 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6482 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6483 0 : "Node {} not found",
6484 0 : migrate_req.node_id
6485 0 : )));
6486 : };
6487 :
6488 : // Migration to unavavailable node requires force flag
6489 0 : if !node.is_available() {
6490 0 : if migrate_req.migration_config.override_scheduler {
6491 : // Warn but proceed: the caller may intend to manually adjust the placement of
6492 : // a shard even if the node is down, e.g. if intervening during an incident.
6493 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
6494 : } else {
6495 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
6496 0 : return Err(ApiError::PreconditionFailed(
6497 0 : format!("Node {node} is unavailable").into_boxed_str(),
6498 0 : ));
6499 : }
6500 0 : }
6501 :
6502 : // Calculate the ScheduleContext for this tenant
6503 0 : let mut schedule_context = ScheduleContext::default();
6504 0 : for (_shard_id, shard) in
6505 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
6506 0 : {
6507 0 : schedule_context.avoid(&shard.intent.all_pageservers());
6508 0 : }
6509 :
6510 : // Look up the specific shard we will migrate
6511 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6512 0 : return Err(ApiError::NotFound(
6513 0 : anyhow::anyhow!("Tenant shard not found").into(),
6514 0 : ));
6515 : };
6516 :
6517 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
6518 : // be migrated back by the optimiser.
6519 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
6520 0 : scheduler,
6521 0 : &schedule_context,
6522 0 : migrate_req.node_id,
6523 0 : &[],
6524 0 : ) {
6525 0 : if !migrate_req.migration_config.override_scheduler {
6526 0 : return Err(ApiError::PreconditionFailed(
6527 0 : "Migration to a worse-scoring node".into(),
6528 0 : ));
6529 : } else {
6530 0 : tracing::info!(
6531 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
6532 : migrate_req.node_id
6533 : );
6534 : }
6535 0 : }
6536 :
6537 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
6538 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
6539 0 : return Err(ApiError::PreconditionFailed(
6540 0 : format!(
6541 0 : "Migration expected to originate from {} but shard is on {:?}",
6542 0 : origin_node_id,
6543 0 : shard.intent.get_attached()
6544 0 : )
6545 0 : .into(),
6546 0 : ));
6547 0 : }
6548 0 : }
6549 :
6550 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6551 : // No-op case: we will still proceed to wait for reconciliation in case it is
6552 : // incomplete from an earlier update to the intent.
6553 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
6554 :
6555 : // An instruction to migrate to the currently attached node should
6556 : // cancel any pending graceful migration
6557 0 : shard.set_preferred_node(None);
6558 :
6559 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
6560 0 : shard,
6561 0 : nodes,
6562 0 : (&migrate_req.migration_config).into(),
6563 0 : ))
6564 0 : } else if migrate_req.migration_config.prewarm {
6565 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
6566 0 : &migrate_req,
6567 0 : shard,
6568 0 : scheduler,
6569 0 : schedule_context,
6570 0 : )?)
6571 : } else {
6572 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
6573 0 : &migrate_req,
6574 0 : nodes,
6575 0 : shard,
6576 0 : scheduler,
6577 0 : )?)
6578 : }
6579 : };
6580 :
6581 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
6582 0 : let waiter = match outcome {
6583 0 : MigrationOutcome::Optimization(Some(optimization)) => {
6584 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
6585 : // we might as well do it more promptly as this is a direct external request.
6586 0 : let mut validated = self
6587 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
6588 0 : .await;
6589 0 : if let Some((_shard_id, optimization)) = validated.pop() {
6590 0 : let mut locked = self.inner.write().unwrap();
6591 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6592 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6593 : // Rare but possible: tenant is removed between generating optimisation and validating it.
6594 0 : return Err(ApiError::NotFound(
6595 0 : anyhow::anyhow!("Tenant shard not found").into(),
6596 0 : ));
6597 : };
6598 :
6599 0 : if !shard.apply_optimization(scheduler, optimization) {
6600 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
6601 : // and therefore not applied.
6602 0 : tracing::warn!(
6603 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
6604 : );
6605 0 : }
6606 0 : self.maybe_configured_reconcile_shard(
6607 0 : shard,
6608 0 : nodes,
6609 0 : (&migrate_req.migration_config).into(),
6610 : )
6611 : } else {
6612 0 : None
6613 : }
6614 : }
6615 0 : MigrationOutcome::Optimization(None) => None,
6616 0 : MigrationOutcome::Reconcile(waiter) => waiter,
6617 : };
6618 :
6619 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
6620 : // locations, this has a good chance of timing out.
6621 0 : if let Some(waiter) = waiter {
6622 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6623 : } else {
6624 0 : tracing::info!("Migration is a no-op");
6625 : }
6626 :
6627 0 : Ok(TenantShardMigrateResponse {})
6628 0 : }
6629 :
6630 0 : pub(crate) async fn tenant_shard_migrate_secondary(
6631 0 : &self,
6632 0 : tenant_shard_id: TenantShardId,
6633 0 : migrate_req: TenantShardMigrateRequest,
6634 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6635 0 : let waiter = {
6636 0 : let mut locked = self.inner.write().unwrap();
6637 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6638 :
6639 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6640 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6641 0 : "Node {} not found",
6642 0 : migrate_req.node_id
6643 0 : )));
6644 : };
6645 :
6646 0 : if !node.is_available() {
6647 : // Warn but proceed: the caller may intend to manually adjust the placement of
6648 : // a shard even if the node is down, e.g. if intervening during an incident.
6649 0 : tracing::warn!("Migrating to unavailable node {node}");
6650 0 : }
6651 :
6652 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6653 0 : return Err(ApiError::NotFound(
6654 0 : anyhow::anyhow!("Tenant shard not found").into(),
6655 0 : ));
6656 : };
6657 :
6658 0 : if shard.intent.get_secondary().len() == 1
6659 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
6660 : {
6661 0 : tracing::info!(
6662 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
6663 : shard.intent
6664 : );
6665 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6666 0 : tracing::info!(
6667 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
6668 : );
6669 : } else {
6670 0 : let old_secondaries = shard.intent.get_secondary().clone();
6671 0 : for secondary in old_secondaries {
6672 0 : shard.intent.remove_secondary(scheduler, secondary);
6673 0 : }
6674 :
6675 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6676 0 : shard.sequence = shard.sequence.next();
6677 0 : tracing::info!(
6678 0 : "Migrating secondary to {node}: new intent {:?}",
6679 : shard.intent
6680 : );
6681 : }
6682 :
6683 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6684 : };
6685 :
6686 0 : if let Some(waiter) = waiter {
6687 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6688 : } else {
6689 0 : tracing::info!("Migration is a no-op");
6690 : }
6691 :
6692 0 : Ok(TenantShardMigrateResponse {})
6693 0 : }
6694 :
6695 : /// 'cancel' in this context means cancel any ongoing reconcile
6696 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
6697 0 : &self,
6698 0 : tenant_shard_id: TenantShardId,
6699 0 : ) -> Result<(), ApiError> {
6700 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
6701 0 : let waiter = {
6702 0 : let locked = self.inner.write().unwrap();
6703 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6704 0 : return Err(ApiError::NotFound(
6705 0 : anyhow::anyhow!("Tenant shard not found").into(),
6706 0 : ));
6707 : };
6708 :
6709 0 : let waiter = shard.get_waiter();
6710 0 : match waiter {
6711 : None => {
6712 0 : tracing::info!("Shard does not have an ongoing Reconciler");
6713 0 : return Ok(());
6714 : }
6715 0 : Some(waiter) => {
6716 0 : tracing::info!("Cancelling Reconciler");
6717 0 : shard.cancel_reconciler();
6718 0 : waiter
6719 : }
6720 : }
6721 : };
6722 :
6723 : // Cancellation should be prompt. If this fails we have still done our job of firing the
6724 : // cancellation token, but by returning an ApiError we will indicate to the caller that
6725 : // the Reconciler is misbehaving and not respecting the cancellation token
6726 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
6727 0 : .await?;
6728 :
6729 0 : Ok(())
6730 0 : }
6731 :
6732 : /// This is for debug/support only: we simply drop all state for a tenant, without
6733 : /// detaching or deleting it on pageservers.
6734 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
6735 0 : self.persistence.delete_tenant(tenant_id).await?;
6736 :
6737 0 : let mut locked = self.inner.write().unwrap();
6738 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6739 0 : let mut shards = Vec::new();
6740 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
6741 0 : shards.push(*tenant_shard_id);
6742 0 : }
6743 :
6744 0 : for shard_id in shards {
6745 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
6746 0 : shard.intent.clear(scheduler);
6747 0 : }
6748 : }
6749 :
6750 0 : Ok(())
6751 0 : }
6752 :
6753 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
6754 : /// tenant with a very high generation number so that it will see the existing data.
6755 : /// It does not create timelines on safekeepers, because they might already exist on some
6756 : /// safekeeper set. So, the timelines are not storcon-managed after the import.
6757 0 : pub(crate) async fn tenant_import(
6758 0 : &self,
6759 0 : tenant_id: TenantId,
6760 0 : ) -> Result<TenantCreateResponse, ApiError> {
6761 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
6762 0 : let maybe_node = {
6763 0 : self.inner
6764 0 : .read()
6765 0 : .unwrap()
6766 0 : .nodes
6767 0 : .values()
6768 0 : .find(|n| n.is_available())
6769 0 : .cloned()
6770 : };
6771 0 : let Some(node) = maybe_node else {
6772 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
6773 : };
6774 :
6775 0 : let client = PageserverClient::new(
6776 0 : node.get_id(),
6777 0 : self.http_client.clone(),
6778 0 : node.base_url(),
6779 0 : self.config.pageserver_jwt_token.as_deref(),
6780 : );
6781 :
6782 0 : let scan_result = client
6783 0 : .tenant_scan_remote_storage(tenant_id)
6784 0 : .await
6785 0 : .map_err(|e| passthrough_api_error(&node, e))?;
6786 :
6787 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
6788 0 : let Some(shard_count) = scan_result
6789 0 : .shards
6790 0 : .iter()
6791 0 : .map(|s| s.tenant_shard_id.shard_count)
6792 0 : .max()
6793 : else {
6794 0 : return Err(ApiError::NotFound(
6795 0 : anyhow::anyhow!("No shards found").into(),
6796 0 : ));
6797 : };
6798 :
6799 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
6800 : // to
6801 0 : let generation = scan_result
6802 0 : .shards
6803 0 : .iter()
6804 0 : .map(|s| s.generation)
6805 0 : .max()
6806 0 : .expect("We already validated >0 shards");
6807 :
6808 : // Find the tenant's stripe size. This wasn't always persisted in the tenant manifest, so
6809 : // fall back to the original default stripe size of 32768 (256 MB) if it's not specified.
6810 : const ORIGINAL_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(32768);
6811 0 : let stripe_size = scan_result
6812 0 : .shards
6813 0 : .iter()
6814 0 : .find(|s| s.tenant_shard_id.shard_count == shard_count && s.generation == generation)
6815 0 : .expect("we validated >0 shards above")
6816 : .stripe_size
6817 0 : .unwrap_or_else(|| {
6818 0 : if shard_count.count() > 1 {
6819 0 : warn!("unknown stripe size, assuming {ORIGINAL_STRIPE_SIZE}");
6820 0 : }
6821 0 : ORIGINAL_STRIPE_SIZE
6822 0 : });
6823 :
6824 0 : let (response, waiters) = self
6825 0 : .do_tenant_create(TenantCreateRequest {
6826 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
6827 0 : generation,
6828 0 :
6829 0 : shard_parameters: ShardParameters {
6830 0 : count: shard_count,
6831 0 : stripe_size,
6832 0 : },
6833 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
6834 0 : config: TenantConfig::default(),
6835 0 : })
6836 0 : .await?;
6837 :
6838 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
6839 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
6840 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
6841 : // reconcile, as reconciliation includes notifying compute.
6842 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
6843 0 : }
6844 :
6845 0 : Ok(response)
6846 0 : }
6847 :
6848 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
6849 : /// we don't have to make TenantShard clonable in the return path.
6850 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
6851 0 : let serialized = {
6852 0 : let locked = self.inner.read().unwrap();
6853 0 : let result = locked.tenants.values().collect::<Vec<_>>();
6854 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
6855 : };
6856 :
6857 0 : hyper::Response::builder()
6858 0 : .status(hyper::StatusCode::OK)
6859 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
6860 0 : .body(hyper::Body::from(serialized))
6861 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
6862 0 : }
6863 :
6864 : /// Check the consistency of in-memory state vs. persistent state, and check that the
6865 : /// scheduler's statistics are up to date.
6866 : ///
6867 : /// These consistency checks expect an **idle** system. If changes are going on while
6868 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
6869 : /// checks, but not suitable for running continuously in the background in the field.
6870 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
6871 0 : let (mut expect_nodes, mut expect_shards) = {
6872 0 : let locked = self.inner.read().unwrap();
6873 :
6874 0 : locked
6875 0 : .scheduler
6876 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
6877 0 : .context("Scheduler checks")
6878 0 : .map_err(ApiError::InternalServerError)?;
6879 :
6880 0 : let expect_nodes = locked
6881 0 : .nodes
6882 0 : .values()
6883 0 : .map(|n| n.to_persistent())
6884 0 : .collect::<Vec<_>>();
6885 :
6886 0 : let expect_shards = locked
6887 0 : .tenants
6888 0 : .values()
6889 0 : .map(|t| t.to_persistent())
6890 0 : .collect::<Vec<_>>();
6891 :
6892 : // This method can only validate the state of an idle system: if a reconcile is in
6893 : // progress, fail out early to avoid giving false errors on state that won't match
6894 : // between database and memory under a ReconcileResult is processed.
6895 0 : for t in locked.tenants.values() {
6896 0 : if t.reconciler.is_some() {
6897 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6898 0 : "Shard {} reconciliation in progress",
6899 0 : t.tenant_shard_id
6900 0 : )));
6901 0 : }
6902 : }
6903 :
6904 0 : (expect_nodes, expect_shards)
6905 : };
6906 :
6907 0 : let mut nodes = self.persistence.list_nodes().await?;
6908 0 : expect_nodes.sort_by_key(|n| n.node_id);
6909 0 : nodes.sort_by_key(|n| n.node_id);
6910 :
6911 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
6912 0 : let node_result = if nodes != expect_nodes {
6913 0 : tracing::error!("Consistency check failed on nodes.");
6914 0 : tracing::error!(
6915 0 : "Nodes in memory: {}",
6916 0 : serde_json::to_string(&expect_nodes)
6917 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6918 : );
6919 0 : tracing::error!(
6920 0 : "Nodes in database: {}",
6921 0 : serde_json::to_string(&nodes)
6922 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6923 : );
6924 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
6925 0 : "Node consistency failure"
6926 0 : )))
6927 : } else {
6928 0 : Ok(())
6929 : };
6930 :
6931 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
6932 0 : persistent_shards
6933 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6934 :
6935 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6936 :
6937 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
6938 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
6939 : // fields are added, before doing a comparison.
6940 0 : for tsp in &mut persistent_shards {
6941 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
6942 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
6943 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
6944 : }
6945 :
6946 0 : if persistent_shards != expect_shards {
6947 0 : tracing::error!("Consistency check failed on shards.");
6948 :
6949 0 : tracing::error!(
6950 0 : "Shards in memory: {}",
6951 0 : serde_json::to_string(&expect_shards)
6952 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6953 : );
6954 0 : tracing::error!(
6955 0 : "Shards in database: {}",
6956 0 : serde_json::to_string(&persistent_shards)
6957 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6958 : );
6959 :
6960 : // The total dump log lines above are useful in testing but in the field grafana will
6961 : // usually just drop them because they're so large. So we also do some explicit logging
6962 : // of just the diffs.
6963 0 : let persistent_shards = persistent_shards
6964 0 : .into_iter()
6965 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6966 0 : .collect::<HashMap<_, _>>();
6967 0 : let expect_shards = expect_shards
6968 0 : .into_iter()
6969 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6970 0 : .collect::<HashMap<_, _>>();
6971 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
6972 0 : match expect_shards.get(tenant_shard_id) {
6973 : None => {
6974 0 : tracing::error!(
6975 0 : "Shard {} found in database but not in memory",
6976 : tenant_shard_id
6977 : );
6978 : }
6979 0 : Some(expect_tsp) => {
6980 0 : if expect_tsp != persistent_tsp {
6981 0 : tracing::error!(
6982 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
6983 : tenant_shard_id,
6984 0 : serde_json::to_string(expect_tsp).unwrap(),
6985 0 : serde_json::to_string(&persistent_tsp).unwrap()
6986 : );
6987 0 : }
6988 : }
6989 : }
6990 : }
6991 :
6992 : // Having already logged any differences, log any shards that simply aren't present in the database
6993 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
6994 0 : if !persistent_shards.contains_key(tenant_shard_id) {
6995 0 : tracing::error!(
6996 0 : "Shard {} found in memory but not in database: {}",
6997 : tenant_shard_id,
6998 0 : serde_json::to_string(memory_tsp)
6999 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7000 : );
7001 0 : }
7002 : }
7003 :
7004 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7005 0 : "Shard consistency failure"
7006 0 : )));
7007 0 : }
7008 :
7009 0 : node_result
7010 0 : }
7011 :
7012 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
7013 : /// we don't have to make TenantShard clonable in the return path.
7014 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
7015 0 : let serialized = {
7016 0 : let locked = self.inner.read().unwrap();
7017 0 : serde_json::to_string(&locked.scheduler)
7018 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
7019 : };
7020 :
7021 0 : hyper::Response::builder()
7022 0 : .status(hyper::StatusCode::OK)
7023 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
7024 0 : .body(hyper::Body::from(serialized))
7025 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
7026 0 : }
7027 :
7028 : /// This is for debug/support only: we simply drop all state for a tenant, without
7029 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
7030 : /// tenants that were on this node.
7031 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
7032 0 : self.persistence.set_tombstone(node_id).await?;
7033 :
7034 0 : let mut locked = self.inner.write().unwrap();
7035 :
7036 0 : for shard in locked.tenants.values_mut() {
7037 0 : shard.deref_node(node_id);
7038 0 : shard.observed.locations.remove(&node_id);
7039 0 : }
7040 :
7041 0 : let mut nodes = (*locked.nodes).clone();
7042 0 : nodes.remove(&node_id);
7043 0 : locked.nodes = Arc::new(nodes);
7044 0 : metrics::METRICS_REGISTRY
7045 0 : .metrics_group
7046 0 : .storage_controller_pageserver_nodes
7047 0 : .set(locked.nodes.len() as i64);
7048 0 : metrics::METRICS_REGISTRY
7049 0 : .metrics_group
7050 0 : .storage_controller_https_pageserver_nodes
7051 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7052 :
7053 0 : locked.scheduler.node_remove(node_id);
7054 :
7055 0 : Ok(())
7056 0 : }
7057 :
7058 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
7059 : /// that we don't leave any bad state behind in the storage controller, but unclean
7060 : /// in the sense that we are not carefully draining the node.
7061 0 : pub(crate) async fn node_delete_old(&self, node_id: NodeId) -> Result<(), ApiError> {
7062 0 : let _node_lock =
7063 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
7064 :
7065 : // 1. Atomically update in-memory state:
7066 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
7067 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
7068 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
7069 : // re-insert references to this node into the ObservedState of shards
7070 : // - drop the node from the scheduler
7071 : {
7072 0 : let mut locked = self.inner.write().unwrap();
7073 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7074 :
7075 : {
7076 0 : let mut nodes_mut = (*nodes).deref().clone();
7077 0 : match nodes_mut.get_mut(&node_id) {
7078 0 : Some(node) => {
7079 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
7080 0 : // if we crash it would not be desirable to leave the node paused after a restart.
7081 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
7082 0 : }
7083 : None => {
7084 0 : tracing::info!(
7085 0 : "Node not found: presuming this is a retry and returning success"
7086 : );
7087 0 : return Ok(());
7088 : }
7089 : }
7090 :
7091 0 : *nodes = Arc::new(nodes_mut);
7092 : }
7093 :
7094 0 : for (_tenant_id, mut schedule_context, shards) in
7095 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7096 : {
7097 0 : for shard in shards {
7098 0 : if shard.deref_node(node_id) {
7099 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
7100 : // TODO: implement force flag to remove a node even if we can't reschedule
7101 : // a tenant
7102 0 : tracing::error!(
7103 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7104 : shard.tenant_shard_id
7105 : );
7106 0 : return Err(e.into());
7107 : } else {
7108 0 : tracing::info!(
7109 0 : "Rescheduled shard {} away from node during deletion",
7110 : shard.tenant_shard_id
7111 : )
7112 : }
7113 :
7114 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
7115 0 : }
7116 :
7117 : // Here we remove an existing observed location for the node we're removing, and it will
7118 : // not be re-added by a reconciler's completion because we filter out removed nodes in
7119 : // process_result.
7120 : //
7121 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
7122 : // means any reconciles we spawned will know about the node we're deleting, enabling them
7123 : // to do live migrations if it's still online.
7124 0 : shard.observed.locations.remove(&node_id);
7125 : }
7126 : }
7127 :
7128 0 : scheduler.node_remove(node_id);
7129 :
7130 : {
7131 0 : let mut nodes_mut = (**nodes).clone();
7132 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7133 0 : // Ensure that any reconciler holding an Arc<> to this node will
7134 0 : // drop out when trying to RPC to it (setting Offline state sets the
7135 0 : // cancellation token on the Node object).
7136 0 : removed_node.set_availability(NodeAvailability::Offline);
7137 0 : }
7138 0 : *nodes = Arc::new(nodes_mut);
7139 0 : metrics::METRICS_REGISTRY
7140 0 : .metrics_group
7141 0 : .storage_controller_pageserver_nodes
7142 0 : .set(nodes.len() as i64);
7143 0 : metrics::METRICS_REGISTRY
7144 0 : .metrics_group
7145 0 : .storage_controller_https_pageserver_nodes
7146 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7147 : }
7148 : }
7149 :
7150 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
7151 : // the removed node, as this column means "The pageserver to which this generation was issued", and
7152 : // their generations won't get updated until the reconcilers moving them away from this node complete.
7153 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
7154 : // that exists.
7155 :
7156 : // 2. Actually delete the node from in-memory state and set tombstone to the database
7157 : // for preventing the node to register again.
7158 0 : tracing::info!("Deleting node from database");
7159 0 : self.persistence.set_tombstone(node_id).await?;
7160 :
7161 0 : Ok(())
7162 0 : }
7163 :
7164 0 : pub(crate) async fn delete_node(
7165 0 : self: &Arc<Self>,
7166 0 : node_id: NodeId,
7167 0 : policy_on_start: NodeSchedulingPolicy,
7168 0 : cancel: CancellationToken,
7169 0 : ) -> Result<(), OperationError> {
7170 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal).build();
7171 :
7172 0 : let mut waiters: Vec<ReconcilerWaiter> = Vec::new();
7173 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
7174 :
7175 0 : while !tid_iter.finished() {
7176 0 : if cancel.is_cancelled() {
7177 0 : match self
7178 0 : .node_configure(node_id, None, Some(policy_on_start))
7179 0 : .await
7180 : {
7181 0 : Ok(()) => return Err(OperationError::Cancelled),
7182 0 : Err(err) => {
7183 0 : return Err(OperationError::FinalizeError(
7184 0 : format!(
7185 0 : "Failed to finalise delete cancel of {} by setting scheduling policy to {}: {}",
7186 0 : node_id, String::from(policy_on_start), err
7187 0 : )
7188 0 : .into(),
7189 0 : ));
7190 : }
7191 : }
7192 0 : }
7193 :
7194 0 : operation_utils::validate_node_state(
7195 0 : &node_id,
7196 0 : self.inner.read().unwrap().nodes.clone(),
7197 0 : NodeSchedulingPolicy::Deleting,
7198 0 : )?;
7199 :
7200 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7201 0 : let tid = match tid_iter.next() {
7202 0 : Some(tid) => tid,
7203 : None => {
7204 0 : break;
7205 : }
7206 : };
7207 :
7208 0 : let mut locked = self.inner.write().unwrap();
7209 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7210 :
7211 : // Calculate a schedule context here to avoid borrow checker issues.
7212 0 : let mut schedule_context = ScheduleContext::default();
7213 0 : for (_, shard) in tenants.range(TenantShardId::tenant_range(tid.tenant_id)) {
7214 0 : schedule_context.avoid(&shard.intent.all_pageservers());
7215 0 : }
7216 :
7217 0 : let tenant_shard = match tenants.get_mut(&tid) {
7218 0 : Some(tenant_shard) => tenant_shard,
7219 : None => {
7220 : // Tenant shard was deleted by another operation. Skip it.
7221 0 : continue;
7222 : }
7223 : };
7224 :
7225 0 : match tenant_shard.get_scheduling_policy() {
7226 0 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {
7227 0 : // A migration during delete is classed as 'essential' because it is required to
7228 0 : // uphold our availability goals for the tenant: this shard is elegible for migration.
7229 0 : }
7230 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
7231 : // If we have been asked to avoid rescheduling this shard, then do not migrate it during a deletion
7232 0 : tracing::warn!(
7233 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7234 0 : "Skip migration during deletion because shard scheduling policy {:?} disallows it",
7235 0 : tenant_shard.get_scheduling_policy(),
7236 : );
7237 0 : continue;
7238 : }
7239 : }
7240 :
7241 0 : if tenant_shard.deref_node(node_id) {
7242 0 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
7243 0 : tracing::error!(
7244 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
7245 : tenant_shard.tenant_shard_id
7246 : );
7247 0 : return Err(OperationError::ImpossibleConstraint(e.to_string().into()));
7248 : } else {
7249 0 : tracing::info!(
7250 0 : "Rescheduled shard {} away from node during deletion",
7251 : tenant_shard.tenant_shard_id
7252 : )
7253 : }
7254 :
7255 0 : let waiter = self.maybe_configured_reconcile_shard(
7256 0 : tenant_shard,
7257 0 : nodes,
7258 0 : reconciler_config,
7259 0 : );
7260 0 : if let Some(some) = waiter {
7261 0 : waiters.push(some);
7262 0 : }
7263 0 : }
7264 : }
7265 :
7266 0 : waiters = self
7267 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
7268 0 : .await;
7269 :
7270 0 : failpoint_support::sleep_millis_async!("sleepy-delete-loop", &cancel);
7271 : }
7272 :
7273 0 : while !waiters.is_empty() {
7274 0 : if cancel.is_cancelled() {
7275 0 : match self
7276 0 : .node_configure(node_id, None, Some(policy_on_start))
7277 0 : .await
7278 : {
7279 0 : Ok(()) => return Err(OperationError::Cancelled),
7280 0 : Err(err) => {
7281 0 : return Err(OperationError::FinalizeError(
7282 0 : format!(
7283 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to {}: {}",
7284 0 : node_id, String::from(policy_on_start), err
7285 0 : )
7286 0 : .into(),
7287 0 : ));
7288 : }
7289 : }
7290 0 : }
7291 :
7292 0 : tracing::info!("Awaiting {} pending delete reconciliations", waiters.len());
7293 :
7294 0 : waiters = self
7295 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7296 0 : .await;
7297 : }
7298 :
7299 0 : self.persistence
7300 0 : .set_tombstone(node_id)
7301 0 : .await
7302 0 : .map_err(|e| OperationError::FinalizeError(e.to_string().into()))?;
7303 :
7304 : {
7305 0 : let mut locked = self.inner.write().unwrap();
7306 0 : let (nodes, _, scheduler) = locked.parts_mut();
7307 :
7308 0 : scheduler.node_remove(node_id);
7309 :
7310 0 : let mut nodes_mut = (**nodes).clone();
7311 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
7312 0 : // Ensure that any reconciler holding an Arc<> to this node will
7313 0 : // drop out when trying to RPC to it (setting Offline state sets the
7314 0 : // cancellation token on the Node object).
7315 0 : removed_node.set_availability(NodeAvailability::Offline);
7316 0 : }
7317 0 : *nodes = Arc::new(nodes_mut);
7318 :
7319 0 : metrics::METRICS_REGISTRY
7320 0 : .metrics_group
7321 0 : .storage_controller_pageserver_nodes
7322 0 : .set(nodes.len() as i64);
7323 0 : metrics::METRICS_REGISTRY
7324 0 : .metrics_group
7325 0 : .storage_controller_https_pageserver_nodes
7326 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
7327 : }
7328 :
7329 0 : Ok(())
7330 0 : }
7331 :
7332 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
7333 0 : let nodes = {
7334 0 : self.inner
7335 0 : .read()
7336 0 : .unwrap()
7337 0 : .nodes
7338 0 : .values()
7339 0 : .cloned()
7340 0 : .collect::<Vec<_>>()
7341 : };
7342 :
7343 0 : Ok(nodes)
7344 0 : }
7345 :
7346 0 : pub(crate) async fn tombstone_list(&self) -> Result<Vec<Node>, ApiError> {
7347 0 : self.persistence
7348 0 : .list_tombstones()
7349 0 : .await?
7350 0 : .into_iter()
7351 0 : .map(|np| Node::from_persistent(np, false))
7352 0 : .collect::<Result<Vec<_>, _>>()
7353 0 : .map_err(ApiError::InternalServerError)
7354 0 : }
7355 :
7356 0 : pub(crate) async fn tombstone_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
7357 0 : let _node_lock = trace_exclusive_lock(
7358 0 : &self.node_op_locks,
7359 0 : node_id,
7360 0 : NodeOperations::DeleteTombstone,
7361 0 : )
7362 0 : .await;
7363 :
7364 0 : if matches!(self.get_node(node_id).await, Err(ApiError::NotFound(_))) {
7365 0 : self.persistence.delete_node(node_id).await?;
7366 0 : Ok(())
7367 : } else {
7368 0 : Err(ApiError::Conflict(format!(
7369 0 : "Node {node_id} is in use, consider using tombstone API first"
7370 0 : )))
7371 : }
7372 0 : }
7373 :
7374 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
7375 0 : self.inner
7376 0 : .read()
7377 0 : .unwrap()
7378 0 : .nodes
7379 0 : .get(&node_id)
7380 0 : .cloned()
7381 0 : .ok_or(ApiError::NotFound(
7382 0 : format!("Node {node_id} not registered").into(),
7383 0 : ))
7384 0 : }
7385 :
7386 0 : pub(crate) async fn get_node_shards(
7387 0 : &self,
7388 0 : node_id: NodeId,
7389 0 : ) -> Result<NodeShardResponse, ApiError> {
7390 0 : let locked = self.inner.read().unwrap();
7391 0 : let mut shards = Vec::new();
7392 0 : for (tid, tenant) in locked.tenants.iter() {
7393 0 : let is_intended_secondary = match (
7394 0 : tenant.intent.get_attached() == &Some(node_id),
7395 0 : tenant.intent.get_secondary().contains(&node_id),
7396 0 : ) {
7397 : (true, true) => {
7398 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
7399 0 : "{} attached as primary+secondary on the same node",
7400 0 : tid
7401 0 : )));
7402 : }
7403 0 : (true, false) => Some(false),
7404 0 : (false, true) => Some(true),
7405 0 : (false, false) => None,
7406 : };
7407 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
7408 0 : tenant.observed.locations.get(&node_id)
7409 : {
7410 0 : Some(conf.secondary_conf.is_some())
7411 : } else {
7412 0 : None
7413 : };
7414 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
7415 0 : shards.push(NodeShard {
7416 0 : tenant_shard_id: *tid,
7417 0 : is_intended_secondary,
7418 0 : is_observed_secondary,
7419 0 : });
7420 0 : }
7421 : }
7422 0 : Ok(NodeShardResponse { node_id, shards })
7423 0 : }
7424 :
7425 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
7426 0 : self.persistence.get_leader().await
7427 0 : }
7428 :
7429 0 : pub(crate) async fn node_register(
7430 0 : &self,
7431 0 : register_req: NodeRegisterRequest,
7432 0 : ) -> Result<(), ApiError> {
7433 0 : let _node_lock = trace_exclusive_lock(
7434 0 : &self.node_op_locks,
7435 0 : register_req.node_id,
7436 0 : NodeOperations::Register,
7437 0 : )
7438 0 : .await;
7439 :
7440 : #[derive(PartialEq)]
7441 : enum RegistrationStatus {
7442 : UpToDate,
7443 : NeedUpdate,
7444 : Mismatched,
7445 : New,
7446 : }
7447 :
7448 0 : let registration_status = {
7449 0 : let locked = self.inner.read().unwrap();
7450 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
7451 0 : if node.registration_match(®ister_req) {
7452 0 : if node.need_update(®ister_req) {
7453 0 : RegistrationStatus::NeedUpdate
7454 : } else {
7455 0 : RegistrationStatus::UpToDate
7456 : }
7457 : } else {
7458 0 : RegistrationStatus::Mismatched
7459 : }
7460 : } else {
7461 0 : RegistrationStatus::New
7462 : }
7463 : };
7464 :
7465 0 : match registration_status {
7466 : RegistrationStatus::UpToDate => {
7467 0 : tracing::info!(
7468 0 : "Node {} re-registered with matching address and is up to date",
7469 : register_req.node_id
7470 : );
7471 :
7472 0 : return Ok(());
7473 : }
7474 : RegistrationStatus::Mismatched => {
7475 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
7476 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
7477 : // a fixed address through the lifetime of a node.
7478 0 : tracing::warn!(
7479 0 : "Node {} tried to register with different address",
7480 : register_req.node_id
7481 : );
7482 0 : return Err(ApiError::Conflict(
7483 0 : "Node is already registered with different address".to_string(),
7484 0 : ));
7485 : }
7486 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
7487 0 : // fallthrough
7488 0 : }
7489 : }
7490 :
7491 : // We do not require that a node is actually online when registered (it will start life
7492 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
7493 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
7494 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
7495 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
7496 0 : if tokio::net::lookup_host(format!(
7497 0 : "{}:{}",
7498 : register_req.listen_http_addr, register_req.listen_http_port
7499 : ))
7500 0 : .await
7501 0 : .is_err()
7502 : {
7503 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
7504 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
7505 : // we return a soft 503 error, to encourage callers to retry past transient issues.
7506 0 : return Err(ApiError::ResourceUnavailable(
7507 0 : format!(
7508 0 : "Node {} tried to register with unknown DNS name '{}'",
7509 0 : register_req.node_id, register_req.listen_http_addr
7510 0 : )
7511 0 : .into(),
7512 0 : ));
7513 0 : }
7514 :
7515 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
7516 0 : return Err(ApiError::PreconditionFailed(
7517 0 : format!(
7518 0 : "Node {} has no https port, but use_https is enabled",
7519 0 : register_req.node_id
7520 0 : )
7521 0 : .into(),
7522 0 : ));
7523 0 : }
7524 :
7525 0 : if register_req.listen_grpc_addr.is_some() != register_req.listen_grpc_port.is_some() {
7526 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
7527 0 : "must specify both gRPC address and port"
7528 0 : )));
7529 0 : }
7530 :
7531 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
7532 : // This ensures that before we use it for anything or expose it via any external
7533 : // API, it is guaranteed to be available after a restart.
7534 0 : let new_node = Node::new(
7535 0 : register_req.node_id,
7536 0 : register_req.listen_http_addr,
7537 0 : register_req.listen_http_port,
7538 0 : register_req.listen_https_port,
7539 0 : register_req.listen_pg_addr,
7540 0 : register_req.listen_pg_port,
7541 0 : register_req.listen_grpc_addr,
7542 0 : register_req.listen_grpc_port,
7543 0 : register_req.availability_zone_id.clone(),
7544 0 : self.config.use_https_pageserver_api,
7545 : );
7546 0 : let new_node = match new_node {
7547 0 : Ok(new_node) => new_node,
7548 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
7549 : };
7550 :
7551 0 : match registration_status {
7552 : RegistrationStatus::New => {
7553 0 : self.persistence.insert_node(&new_node).await.map_err(|e| {
7554 0 : if matches!(
7555 0 : e,
7556 : crate::persistence::DatabaseError::Query(
7557 : diesel::result::Error::DatabaseError(
7558 : diesel::result::DatabaseErrorKind::UniqueViolation,
7559 : _,
7560 : )
7561 : )
7562 : ) {
7563 : // The node can be deleted by tombstone API, and not show up in the list of nodes.
7564 : // If you see this error, check tombstones first.
7565 0 : ApiError::Conflict(format!("Node {} is already exists", new_node.get_id()))
7566 : } else {
7567 0 : ApiError::from(e)
7568 : }
7569 0 : })?;
7570 : }
7571 : RegistrationStatus::NeedUpdate => {
7572 0 : self.persistence
7573 0 : .update_node_on_registration(
7574 0 : register_req.node_id,
7575 0 : register_req.listen_https_port,
7576 0 : )
7577 0 : .await?
7578 : }
7579 0 : _ => unreachable!("Other statuses have been processed earlier"),
7580 : }
7581 :
7582 0 : let mut locked = self.inner.write().unwrap();
7583 0 : let mut new_nodes = (*locked.nodes).clone();
7584 :
7585 0 : locked.scheduler.node_upsert(&new_node);
7586 0 : new_nodes.insert(register_req.node_id, new_node);
7587 :
7588 0 : locked.nodes = Arc::new(new_nodes);
7589 :
7590 0 : metrics::METRICS_REGISTRY
7591 0 : .metrics_group
7592 0 : .storage_controller_pageserver_nodes
7593 0 : .set(locked.nodes.len() as i64);
7594 0 : metrics::METRICS_REGISTRY
7595 0 : .metrics_group
7596 0 : .storage_controller_https_pageserver_nodes
7597 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
7598 :
7599 0 : match registration_status {
7600 : RegistrationStatus::New => {
7601 0 : tracing::info!(
7602 0 : "Registered pageserver {} ({}), now have {} pageservers",
7603 : register_req.node_id,
7604 : register_req.availability_zone_id,
7605 0 : locked.nodes.len()
7606 : );
7607 : }
7608 : RegistrationStatus::NeedUpdate => {
7609 0 : tracing::info!(
7610 0 : "Re-registered and updated node {} ({})",
7611 : register_req.node_id,
7612 : register_req.availability_zone_id,
7613 : );
7614 : }
7615 0 : _ => unreachable!("Other statuses have been processed earlier"),
7616 : }
7617 0 : Ok(())
7618 0 : }
7619 :
7620 : /// Configure in-memory and persistent state of a node as requested
7621 : ///
7622 : /// Note that this function does not trigger any immediate side effects in response
7623 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
7624 0 : async fn node_state_configure(
7625 0 : &self,
7626 0 : node_id: NodeId,
7627 0 : availability: Option<NodeAvailability>,
7628 0 : scheduling: Option<NodeSchedulingPolicy>,
7629 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
7630 0 : ) -> Result<AvailabilityTransition, ApiError> {
7631 0 : if let Some(scheduling) = scheduling {
7632 : // Scheduling is a persistent part of Node: we must write updates to the database before
7633 : // applying them in memory
7634 0 : self.persistence
7635 0 : .update_node_scheduling_policy(node_id, scheduling)
7636 0 : .await?;
7637 0 : }
7638 :
7639 : // If we're activating a node, then before setting it active we must reconcile any shard locations
7640 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
7641 : // by calling [`Self::node_activate_reconcile`]
7642 : //
7643 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
7644 : // nothing else can mutate its availability while we run.
7645 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
7646 0 : let (activate_node, availability_transition) = {
7647 0 : let locked = self.inner.read().unwrap();
7648 0 : let Some(node) = locked.nodes.get(&node_id) else {
7649 0 : return Err(ApiError::NotFound(
7650 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7651 0 : ));
7652 : };
7653 :
7654 0 : (
7655 0 : node.clone(),
7656 0 : node.get_availability_transition(input_availability),
7657 0 : )
7658 : };
7659 :
7660 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
7661 0 : self.node_activate_reconcile(activate_node, node_lock)
7662 0 : .await?;
7663 0 : }
7664 0 : availability_transition
7665 : } else {
7666 0 : AvailabilityTransition::Unchanged
7667 : };
7668 :
7669 : // Apply changes from the request to our in-memory state for the Node
7670 0 : let mut locked = self.inner.write().unwrap();
7671 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
7672 :
7673 0 : let mut new_nodes = (**nodes).clone();
7674 :
7675 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
7676 0 : return Err(ApiError::NotFound(
7677 0 : anyhow::anyhow!("Node not registered").into(),
7678 0 : ));
7679 : };
7680 :
7681 0 : if let Some(availability) = availability {
7682 0 : node.set_availability(availability);
7683 0 : }
7684 :
7685 0 : if let Some(scheduling) = scheduling {
7686 0 : node.set_scheduling(scheduling);
7687 0 : }
7688 :
7689 : // Update the scheduler, in case the elegibility of the node for new shards has changed
7690 0 : scheduler.node_upsert(node);
7691 :
7692 0 : let new_nodes = Arc::new(new_nodes);
7693 0 : locked.nodes = new_nodes;
7694 :
7695 0 : Ok(availability_transition)
7696 0 : }
7697 :
7698 : /// Handle availability transition of one node
7699 : ///
7700 : /// Note that you should first call [`Self::node_state_configure`] to update
7701 : /// the in-memory state referencing that node. If you need to handle more than one transition
7702 : /// consider using [`Self::handle_node_availability_transitions`].
7703 0 : async fn handle_node_availability_transition(
7704 0 : &self,
7705 0 : node_id: NodeId,
7706 0 : transition: AvailabilityTransition,
7707 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
7708 0 : ) -> Result<(), ApiError> {
7709 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
7710 0 : match transition {
7711 : AvailabilityTransition::ToOffline => {
7712 0 : tracing::info!("Node {} transition to offline", node_id);
7713 :
7714 0 : let mut locked = self.inner.write().unwrap();
7715 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7716 :
7717 0 : let mut tenants_affected: usize = 0;
7718 :
7719 0 : for (_tenant_id, mut schedule_context, shards) in
7720 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Normal)
7721 : {
7722 0 : for tenant_shard in shards {
7723 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
7724 0 : if let Some(observed_loc) =
7725 0 : tenant_shard.observed.locations.get_mut(&node_id)
7726 0 : {
7727 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
7728 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
7729 0 : observed_loc.conf = None;
7730 0 : }
7731 :
7732 0 : if nodes.len() == 1 {
7733 : // Special case for single-node cluster: there is no point trying to reschedule
7734 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
7735 : // failures to schedule them.
7736 0 : continue;
7737 0 : }
7738 :
7739 0 : if !nodes
7740 0 : .values()
7741 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
7742 : {
7743 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
7744 : // trying to reschedule since there's nowhere else to go. Without this
7745 : // branch we incorrectly detach tenants in response to node unavailability.
7746 0 : continue;
7747 0 : }
7748 :
7749 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
7750 0 : tenant_shard.sequence = tenant_shard.sequence.next();
7751 :
7752 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
7753 0 : Err(e) => {
7754 : // It is possible that some tenants will become unschedulable when too many pageservers
7755 : // go offline: in this case there isn't much we can do other than make the issue observable.
7756 : // TODO: give TenantShard a scheduling error attribute to be queried later.
7757 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
7758 : }
7759 : Ok(()) => {
7760 0 : if self
7761 0 : .maybe_reconcile_shard(
7762 0 : tenant_shard,
7763 0 : nodes,
7764 0 : ReconcilerPriority::Normal,
7765 0 : )
7766 0 : .is_some()
7767 0 : {
7768 0 : tenants_affected += 1;
7769 0 : };
7770 : }
7771 : }
7772 0 : }
7773 : }
7774 : }
7775 0 : tracing::info!(
7776 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
7777 : tenants_affected,
7778 : node_id
7779 : )
7780 : }
7781 : AvailabilityTransition::ToActive => {
7782 0 : tracing::info!("Node {} transition to active", node_id);
7783 :
7784 0 : let mut locked = self.inner.write().unwrap();
7785 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7786 :
7787 : // When a node comes back online, we must reconcile any tenant that has a None observed
7788 : // location on the node.
7789 0 : for tenant_shard in tenants.values_mut() {
7790 : // If a reconciliation is already in progress, rely on the previous scheduling
7791 : // decision and skip triggering a new reconciliation.
7792 0 : if tenant_shard.reconciler.is_some() {
7793 0 : continue;
7794 0 : }
7795 :
7796 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
7797 0 : if observed_loc.conf.is_none() {
7798 0 : self.maybe_reconcile_shard(
7799 0 : tenant_shard,
7800 0 : nodes,
7801 0 : ReconcilerPriority::Normal,
7802 0 : );
7803 0 : }
7804 0 : }
7805 : }
7806 :
7807 : // TODO: in the background, we should balance work back onto this pageserver
7808 : }
7809 : // No action required for the intermediate unavailable state.
7810 : // When we transition into active or offline from the unavailable state,
7811 : // the correct handling above will kick in.
7812 : AvailabilityTransition::ToWarmingUpFromActive => {
7813 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
7814 : }
7815 : AvailabilityTransition::ToWarmingUpFromOffline => {
7816 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
7817 : }
7818 : AvailabilityTransition::Unchanged => {
7819 0 : tracing::debug!("Node {} no availability change during config", node_id);
7820 : }
7821 : }
7822 :
7823 0 : Ok(())
7824 0 : }
7825 :
7826 : /// Handle availability transition for multiple nodes
7827 : ///
7828 : /// Note that you should first call [`Self::node_state_configure`] for
7829 : /// all nodes being handled here for the handling to use fresh in-memory state.
7830 0 : async fn handle_node_availability_transitions(
7831 0 : &self,
7832 0 : transitions: Vec<(
7833 0 : NodeId,
7834 0 : TracingExclusiveGuard<NodeOperations>,
7835 0 : AvailabilityTransition,
7836 0 : )>,
7837 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
7838 0 : let mut errors = Vec::default();
7839 0 : for (node_id, node_lock, transition) in transitions {
7840 0 : let res = self
7841 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
7842 0 : .await;
7843 0 : if let Err(err) = res {
7844 0 : errors.push((node_id, err));
7845 0 : }
7846 : }
7847 :
7848 0 : if errors.is_empty() {
7849 0 : Ok(())
7850 : } else {
7851 0 : Err(errors)
7852 : }
7853 0 : }
7854 :
7855 0 : pub(crate) async fn node_configure(
7856 0 : &self,
7857 0 : node_id: NodeId,
7858 0 : availability: Option<NodeAvailability>,
7859 0 : scheduling: Option<NodeSchedulingPolicy>,
7860 0 : ) -> Result<(), ApiError> {
7861 0 : let node_lock =
7862 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
7863 :
7864 0 : let transition = self
7865 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
7866 0 : .await?;
7867 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
7868 0 : .await
7869 0 : }
7870 :
7871 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
7872 : /// operation for HTTP api.
7873 0 : pub(crate) async fn external_node_configure(
7874 0 : &self,
7875 0 : node_id: NodeId,
7876 0 : availability: Option<NodeAvailability>,
7877 0 : scheduling: Option<NodeSchedulingPolicy>,
7878 0 : ) -> Result<(), ApiError> {
7879 : {
7880 0 : let locked = self.inner.read().unwrap();
7881 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
7882 0 : return Err(ApiError::PreconditionFailed(
7883 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
7884 0 : ));
7885 0 : }
7886 : }
7887 :
7888 0 : self.node_configure(node_id, availability, scheduling).await
7889 0 : }
7890 :
7891 0 : pub(crate) async fn start_node_delete(
7892 0 : self: &Arc<Self>,
7893 0 : node_id: NodeId,
7894 0 : ) -> Result<(), ApiError> {
7895 0 : let (ongoing_op, node_policy, schedulable_nodes_count) = {
7896 0 : let locked = self.inner.read().unwrap();
7897 0 : let nodes = &locked.nodes;
7898 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7899 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7900 0 : ))?;
7901 0 : let schedulable_nodes_count = nodes
7902 0 : .iter()
7903 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
7904 0 : .count();
7905 :
7906 : (
7907 0 : locked
7908 0 : .ongoing_operation
7909 0 : .as_ref()
7910 0 : .map(|ongoing| ongoing.operation),
7911 0 : node.get_scheduling(),
7912 0 : schedulable_nodes_count,
7913 : )
7914 : };
7915 :
7916 0 : if let Some(ongoing) = ongoing_op {
7917 0 : return Err(ApiError::PreconditionFailed(
7918 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
7919 0 : ));
7920 0 : }
7921 :
7922 0 : if schedulable_nodes_count == 0 {
7923 0 : return Err(ApiError::PreconditionFailed(
7924 0 : "No other schedulable nodes to move shards".into(),
7925 0 : ));
7926 0 : }
7927 :
7928 0 : match node_policy {
7929 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
7930 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Deleting))
7931 0 : .await?;
7932 :
7933 0 : let cancel = self.cancel.child_token();
7934 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
7935 0 : let policy_on_start = node_policy;
7936 :
7937 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
7938 0 : operation: Operation::Delete(Delete { node_id }),
7939 0 : cancel: cancel.clone(),
7940 0 : });
7941 :
7942 0 : let span = tracing::info_span!(parent: None, "delete_node", %node_id);
7943 :
7944 0 : tokio::task::spawn(
7945 : {
7946 0 : let service = self.clone();
7947 0 : let cancel = cancel.clone();
7948 0 : async move {
7949 0 : let _gate_guard = gate_guard;
7950 :
7951 0 : scopeguard::defer! {
7952 : let prev = service.inner.write().unwrap().ongoing_operation.take();
7953 :
7954 : if let Some(Operation::Delete(removed_delete)) = prev.map(|h| h.operation) {
7955 : assert_eq!(removed_delete.node_id, node_id, "We always take the same operation");
7956 : } else {
7957 : panic!("We always remove the same operation")
7958 : }
7959 : }
7960 :
7961 0 : tracing::info!("Delete background operation starting");
7962 0 : let res = service
7963 0 : .delete_node(node_id, policy_on_start, cancel)
7964 0 : .await;
7965 0 : match res {
7966 : Ok(()) => {
7967 0 : tracing::info!(
7968 0 : "Delete background operation completed successfully"
7969 : );
7970 : }
7971 : Err(OperationError::Cancelled) => {
7972 0 : tracing::info!("Delete background operation was cancelled");
7973 : }
7974 0 : Err(err) => {
7975 0 : tracing::error!(
7976 0 : "Delete background operation encountered: {err}"
7977 : )
7978 : }
7979 : }
7980 0 : }
7981 : }
7982 0 : .instrument(span),
7983 : );
7984 : }
7985 : NodeSchedulingPolicy::Deleting => {
7986 0 : return Err(ApiError::Conflict(format!(
7987 0 : "Node {node_id} has delete in progress"
7988 0 : )));
7989 : }
7990 0 : policy => {
7991 0 : return Err(ApiError::PreconditionFailed(
7992 0 : format!("Node {node_id} cannot be deleted due to {policy:?} policy").into(),
7993 0 : ));
7994 : }
7995 : }
7996 :
7997 0 : Ok(())
7998 0 : }
7999 :
8000 0 : pub(crate) async fn cancel_node_delete(
8001 0 : self: &Arc<Self>,
8002 0 : node_id: NodeId,
8003 0 : ) -> Result<(), ApiError> {
8004 : {
8005 0 : let locked = self.inner.read().unwrap();
8006 0 : let nodes = &locked.nodes;
8007 0 : nodes.get(&node_id).ok_or(ApiError::NotFound(
8008 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8009 0 : ))?;
8010 : }
8011 :
8012 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8013 0 : if let Operation::Delete(delete) = op_handler.operation {
8014 0 : if delete.node_id == node_id {
8015 0 : tracing::info!("Cancelling background delete operation for node {node_id}");
8016 0 : op_handler.cancel.cancel();
8017 0 : return Ok(());
8018 0 : }
8019 0 : }
8020 0 : }
8021 :
8022 0 : Err(ApiError::PreconditionFailed(
8023 0 : format!("Node {node_id} has no delete in progress").into(),
8024 0 : ))
8025 0 : }
8026 :
8027 0 : pub(crate) async fn start_node_drain(
8028 0 : self: &Arc<Self>,
8029 0 : node_id: NodeId,
8030 0 : ) -> Result<(), ApiError> {
8031 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
8032 0 : let locked = self.inner.read().unwrap();
8033 0 : let nodes = &locked.nodes;
8034 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8035 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8036 0 : ))?;
8037 0 : let schedulable_nodes_count = nodes
8038 0 : .iter()
8039 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
8040 0 : .count();
8041 :
8042 : (
8043 0 : locked
8044 0 : .ongoing_operation
8045 0 : .as_ref()
8046 0 : .map(|ongoing| ongoing.operation),
8047 0 : node.is_available(),
8048 0 : node.get_scheduling(),
8049 0 : schedulable_nodes_count,
8050 : )
8051 : };
8052 :
8053 0 : if let Some(ongoing) = ongoing_op {
8054 0 : return Err(ApiError::PreconditionFailed(
8055 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8056 0 : ));
8057 0 : }
8058 :
8059 0 : if !node_available {
8060 0 : return Err(ApiError::ResourceUnavailable(
8061 0 : format!("Node {node_id} is currently unavailable").into(),
8062 0 : ));
8063 0 : }
8064 :
8065 0 : if schedulable_nodes_count == 0 {
8066 0 : return Err(ApiError::PreconditionFailed(
8067 0 : "No other schedulable nodes to drain to".into(),
8068 0 : ));
8069 0 : }
8070 :
8071 0 : match node_policy {
8072 : NodeSchedulingPolicy::Active => {
8073 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
8074 0 : .await?;
8075 :
8076 0 : let cancel = self.cancel.child_token();
8077 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8078 :
8079 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8080 0 : operation: Operation::Drain(Drain { node_id }),
8081 0 : cancel: cancel.clone(),
8082 0 : });
8083 :
8084 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
8085 :
8086 0 : tokio::task::spawn({
8087 0 : let service = self.clone();
8088 0 : let cancel = cancel.clone();
8089 0 : async move {
8090 0 : let _gate_guard = gate_guard;
8091 :
8092 0 : scopeguard::defer! {
8093 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8094 :
8095 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
8096 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
8097 : } else {
8098 : panic!("We always remove the same operation")
8099 : }
8100 : }
8101 :
8102 0 : tracing::info!("Drain background operation starting");
8103 0 : let res = service.drain_node(node_id, cancel).await;
8104 0 : match res {
8105 : Ok(()) => {
8106 0 : tracing::info!("Drain background operation completed successfully");
8107 : }
8108 : Err(OperationError::Cancelled) => {
8109 0 : tracing::info!("Drain background operation was cancelled");
8110 : }
8111 0 : Err(err) => {
8112 0 : tracing::error!("Drain background operation encountered: {err}")
8113 : }
8114 : }
8115 0 : }
8116 0 : }.instrument(span));
8117 : }
8118 : NodeSchedulingPolicy::Draining => {
8119 0 : return Err(ApiError::Conflict(format!(
8120 0 : "Node {node_id} has drain in progress"
8121 0 : )));
8122 : }
8123 0 : policy => {
8124 0 : return Err(ApiError::PreconditionFailed(
8125 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
8126 0 : ));
8127 : }
8128 : }
8129 :
8130 0 : Ok(())
8131 0 : }
8132 :
8133 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
8134 0 : let node_available = {
8135 0 : let locked = self.inner.read().unwrap();
8136 0 : let nodes = &locked.nodes;
8137 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8138 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8139 0 : ))?;
8140 :
8141 0 : node.is_available()
8142 : };
8143 :
8144 0 : if !node_available {
8145 0 : return Err(ApiError::ResourceUnavailable(
8146 0 : format!("Node {node_id} is currently unavailable").into(),
8147 0 : ));
8148 0 : }
8149 :
8150 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8151 0 : if let Operation::Drain(drain) = op_handler.operation {
8152 0 : if drain.node_id == node_id {
8153 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8154 0 : op_handler.cancel.cancel();
8155 0 : return Ok(());
8156 0 : }
8157 0 : }
8158 0 : }
8159 :
8160 0 : Err(ApiError::PreconditionFailed(
8161 0 : format!("Node {node_id} has no drain in progress").into(),
8162 0 : ))
8163 0 : }
8164 :
8165 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
8166 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
8167 0 : let locked = self.inner.read().unwrap();
8168 0 : let nodes = &locked.nodes;
8169 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8170 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8171 0 : ))?;
8172 :
8173 : (
8174 0 : locked
8175 0 : .ongoing_operation
8176 0 : .as_ref()
8177 0 : .map(|ongoing| ongoing.operation),
8178 0 : node.is_available(),
8179 0 : node.get_scheduling(),
8180 0 : nodes.len(),
8181 : )
8182 : };
8183 :
8184 0 : if let Some(ongoing) = ongoing_op {
8185 0 : return Err(ApiError::PreconditionFailed(
8186 0 : format!("Background operation already ongoing for node: {ongoing}").into(),
8187 0 : ));
8188 0 : }
8189 :
8190 0 : if !node_available {
8191 0 : return Err(ApiError::ResourceUnavailable(
8192 0 : format!("Node {node_id} is currently unavailable").into(),
8193 0 : ));
8194 0 : }
8195 :
8196 0 : if total_nodes_count <= 1 {
8197 0 : return Err(ApiError::PreconditionFailed(
8198 0 : "No other nodes to fill from".into(),
8199 0 : ));
8200 0 : }
8201 :
8202 0 : match node_policy {
8203 : NodeSchedulingPolicy::Active => {
8204 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
8205 0 : .await?;
8206 :
8207 0 : let cancel = self.cancel.child_token();
8208 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
8209 :
8210 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
8211 0 : operation: Operation::Fill(Fill { node_id }),
8212 0 : cancel: cancel.clone(),
8213 0 : });
8214 :
8215 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
8216 :
8217 0 : tokio::task::spawn({
8218 0 : let service = self.clone();
8219 0 : let cancel = cancel.clone();
8220 0 : async move {
8221 0 : let _gate_guard = gate_guard;
8222 :
8223 0 : scopeguard::defer! {
8224 : let prev = service.inner.write().unwrap().ongoing_operation.take();
8225 :
8226 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
8227 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
8228 : } else {
8229 : panic!("We always remove the same operation")
8230 : }
8231 : }
8232 :
8233 0 : tracing::info!("Fill background operation starting");
8234 0 : let res = service.fill_node(node_id, cancel).await;
8235 0 : match res {
8236 : Ok(()) => {
8237 0 : tracing::info!("Fill background operation completed successfully");
8238 : }
8239 : Err(OperationError::Cancelled) => {
8240 0 : tracing::info!("Fill background operation was cancelled");
8241 : }
8242 0 : Err(err) => {
8243 0 : tracing::error!("Fill background operation encountered: {err}")
8244 : }
8245 : }
8246 0 : }
8247 0 : }.instrument(span));
8248 : }
8249 : NodeSchedulingPolicy::Filling => {
8250 0 : return Err(ApiError::Conflict(format!(
8251 0 : "Node {node_id} has fill in progress"
8252 0 : )));
8253 : }
8254 0 : policy => {
8255 0 : return Err(ApiError::PreconditionFailed(
8256 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
8257 0 : ));
8258 : }
8259 : }
8260 :
8261 0 : Ok(())
8262 0 : }
8263 :
8264 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
8265 0 : let node_available = {
8266 0 : let locked = self.inner.read().unwrap();
8267 0 : let nodes = &locked.nodes;
8268 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
8269 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
8270 0 : ))?;
8271 :
8272 0 : node.is_available()
8273 : };
8274 :
8275 0 : if !node_available {
8276 0 : return Err(ApiError::ResourceUnavailable(
8277 0 : format!("Node {node_id} is currently unavailable").into(),
8278 0 : ));
8279 0 : }
8280 :
8281 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
8282 0 : if let Operation::Fill(fill) = op_handler.operation {
8283 0 : if fill.node_id == node_id {
8284 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
8285 0 : op_handler.cancel.cancel();
8286 0 : return Ok(());
8287 0 : }
8288 0 : }
8289 0 : }
8290 :
8291 0 : Err(ApiError::PreconditionFailed(
8292 0 : format!("Node {node_id} has no fill in progress").into(),
8293 0 : ))
8294 0 : }
8295 :
8296 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
8297 : /// configuration
8298 0 : fn maybe_reconcile_shard(
8299 0 : &self,
8300 0 : shard: &mut TenantShard,
8301 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8302 0 : priority: ReconcilerPriority,
8303 0 : ) -> Option<ReconcilerWaiter> {
8304 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
8305 0 : }
8306 :
8307 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
8308 0 : fn get_reconciler_units(
8309 0 : &self,
8310 0 : priority: ReconcilerPriority,
8311 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
8312 0 : let units = match priority {
8313 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
8314 : ReconcilerPriority::High => {
8315 0 : match self
8316 0 : .priority_reconciler_concurrency
8317 0 : .clone()
8318 0 : .try_acquire_owned()
8319 : {
8320 0 : Ok(u) => Ok(u),
8321 : Err(TryAcquireError::NoPermits) => {
8322 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
8323 : // the normal priority semaphore.
8324 0 : self.reconciler_concurrency.clone().try_acquire_owned()
8325 : }
8326 0 : Err(e) => Err(e),
8327 : }
8328 : }
8329 : };
8330 :
8331 0 : units.map(ReconcileUnits::new)
8332 0 : }
8333 :
8334 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
8335 0 : fn maybe_configured_reconcile_shard(
8336 0 : &self,
8337 0 : shard: &mut TenantShard,
8338 0 : nodes: &Arc<HashMap<NodeId, Node>>,
8339 0 : reconciler_config: ReconcilerConfig,
8340 0 : ) -> Option<ReconcilerWaiter> {
8341 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
8342 :
8343 0 : let reconcile_reason = match reconcile_needed {
8344 0 : ReconcileNeeded::No => return None,
8345 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
8346 0 : ReconcileNeeded::Yes(reason) => {
8347 : // Fall through to try and acquire units for spawning reconciler
8348 0 : reason
8349 : }
8350 : };
8351 :
8352 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
8353 0 : Ok(u) => u,
8354 : Err(_) => {
8355 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
8356 0 : "Concurrency limited: enqueued for reconcile later");
8357 0 : if !shard.delayed_reconcile {
8358 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
8359 0 : Err(TrySendError::Closed(_)) => {
8360 0 : // Weird mid-shutdown case?
8361 0 : }
8362 : Err(TrySendError::Full(_)) => {
8363 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
8364 0 : tracing::warn!(
8365 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
8366 : );
8367 : }
8368 0 : Ok(()) => {
8369 0 : shard.delayed_reconcile = true;
8370 0 : }
8371 : }
8372 0 : }
8373 :
8374 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
8375 : // number to advance. When this function is eventually called again and succeeds in getting units,
8376 : // it will spawn a reconciler that makes this waiter complete.
8377 0 : return Some(shard.future_reconcile_waiter());
8378 : }
8379 : };
8380 :
8381 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
8382 : // Gate closed: we're shutting down, drop out.
8383 0 : return None;
8384 : };
8385 :
8386 0 : shard.spawn_reconciler(
8387 0 : reconcile_reason,
8388 0 : &self.result_tx,
8389 0 : nodes,
8390 0 : &self.compute_hook,
8391 0 : reconciler_config,
8392 0 : &self.config,
8393 0 : &self.persistence,
8394 0 : units,
8395 0 : gate_guard,
8396 0 : &self.reconcilers_cancel,
8397 0 : self.http_client.clone(),
8398 : )
8399 0 : }
8400 :
8401 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
8402 : /// Additionally, reschedule tenants that require it.
8403 : ///
8404 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
8405 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
8406 : /// available. A return value of 0 indicates that everything is fully reconciled already.
8407 0 : fn reconcile_all(&self) -> ReconcileAllResult {
8408 0 : let mut locked = self.inner.write().unwrap();
8409 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8410 0 : let pageservers = nodes.clone();
8411 :
8412 : // This function is an efficient place to update lazy statistics, since we are walking
8413 : // all tenants.
8414 0 : let mut pending_reconciles = 0;
8415 0 : let mut keep_failing_reconciles = 0;
8416 0 : let mut az_violations = 0;
8417 :
8418 : // If we find any tenants to drop from memory, stash them to offload after
8419 : // we're done traversing the map of tenants.
8420 0 : let mut drop_detached_tenants = Vec::new();
8421 :
8422 0 : let mut spawned_reconciles = 0;
8423 0 : let mut has_delayed_reconciles = false;
8424 :
8425 0 : for shard in tenants.values_mut() {
8426 : // Accumulate scheduling statistics
8427 0 : if let (Some(attached), Some(preferred)) =
8428 0 : (shard.intent.get_attached(), shard.preferred_az())
8429 : {
8430 0 : let node_az = nodes
8431 0 : .get(attached)
8432 0 : .expect("Nodes exist if referenced")
8433 0 : .get_availability_zone_id();
8434 0 : if node_az != preferred {
8435 0 : az_violations += 1;
8436 0 : }
8437 0 : }
8438 :
8439 : // Skip checking if this shard is already enqueued for reconciliation
8440 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
8441 : // If there is something delayed, then return a nonzero count so that
8442 : // callers like reconcile_all_now do not incorrectly get the impression
8443 : // that the system is in a quiescent state.
8444 0 : has_delayed_reconciles = true;
8445 0 : pending_reconciles += 1;
8446 0 : continue;
8447 0 : }
8448 :
8449 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
8450 : // dirty, spawn another one
8451 0 : let consecutive_errors_count = shard.consecutive_errors_count;
8452 0 : if self
8453 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
8454 0 : .is_some()
8455 : {
8456 0 : spawned_reconciles += 1;
8457 :
8458 : // Count shards that are keep-failing. We still want to reconcile them
8459 : // to avoid a situation where a shard is stuck.
8460 : // But we don't want to consider them when deciding to run optimizations.
8461 0 : if consecutive_errors_count >= MAX_CONSECUTIVE_RECONCILIATION_ERRORS {
8462 0 : tracing::warn!(
8463 : tenant_id=%shard.tenant_shard_id.tenant_id,
8464 0 : shard_id=%shard.tenant_shard_id.shard_slug(),
8465 0 : "Shard reconciliation is keep-failing: {} errors",
8466 : consecutive_errors_count
8467 : );
8468 0 : keep_failing_reconciles += 1;
8469 0 : }
8470 0 : } else if shard.delayed_reconcile {
8471 0 : // Shard wanted to reconcile but for some reason couldn't.
8472 0 : pending_reconciles += 1;
8473 0 : }
8474 :
8475 : // If this tenant is detached, try dropping it from memory. This is usually done
8476 : // proactively in [`Self::process_results`], but we do it here to handle the edge
8477 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
8478 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
8479 0 : && shard.policy == PlacementPolicy::Detached
8480 : {
8481 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
8482 0 : shard.tenant_shard_id.tenant_id,
8483 0 : TenantOperations::DropDetached,
8484 0 : ) {
8485 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
8486 0 : }
8487 0 : }
8488 : }
8489 :
8490 : // Some metrics are calculated from SchedulerNode state, update these periodically
8491 0 : scheduler.update_metrics();
8492 :
8493 : // Process any deferred tenant drops
8494 0 : for (tenant_id, guard) in drop_detached_tenants {
8495 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
8496 0 : }
8497 :
8498 0 : metrics::METRICS_REGISTRY
8499 0 : .metrics_group
8500 0 : .storage_controller_schedule_az_violation
8501 0 : .set(az_violations as i64);
8502 :
8503 0 : metrics::METRICS_REGISTRY
8504 0 : .metrics_group
8505 0 : .storage_controller_pending_reconciles
8506 0 : .set(pending_reconciles as i64);
8507 :
8508 0 : metrics::METRICS_REGISTRY
8509 0 : .metrics_group
8510 0 : .storage_controller_keep_failing_reconciles
8511 0 : .set(keep_failing_reconciles as i64);
8512 :
8513 0 : ReconcileAllResult::new(
8514 0 : spawned_reconciles,
8515 0 : keep_failing_reconciles,
8516 0 : has_delayed_reconciles,
8517 : )
8518 0 : }
8519 :
8520 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
8521 : /// could be scheduled somewhere better:
8522 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
8523 : /// * e.g. after a node fails then recovers, to move some work back to it
8524 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
8525 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
8526 : /// we did the split, but are probably better placed elsewhere.
8527 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
8528 : /// * e.g. after a shard split, some locations will be on the same node (where the split
8529 : /// happened), and will probably be better placed elsewhere.
8530 : ///
8531 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
8532 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
8533 : /// according to those same soft constraints.
8534 0 : async fn optimize_all(&self) -> usize {
8535 : // Limit on how many shards' optmizations each call to this function will execute. Combined
8536 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
8537 : // trickle of optimizations in the background, rather than executing a large number in parallel
8538 : // when a change occurs.
8539 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
8540 :
8541 : // Synchronous prepare: scan shards for possible scheduling optimizations
8542 0 : let candidate_work = self.optimize_all_plan();
8543 0 : let candidate_work_len = candidate_work.len();
8544 :
8545 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
8546 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
8547 :
8548 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
8549 :
8550 : // Synchronous apply: update the shards' intent states according to validated optimisations
8551 0 : let mut reconciles_spawned = 0;
8552 0 : let mut optimizations_applied = 0;
8553 0 : let mut locked = self.inner.write().unwrap();
8554 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8555 0 : for (tenant_shard_id, optimization) in validated_work {
8556 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
8557 : // Shard was dropped between planning and execution;
8558 0 : continue;
8559 : };
8560 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
8561 0 : if shard.apply_optimization(scheduler, optimization) {
8562 0 : optimizations_applied += 1;
8563 0 : if self
8564 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
8565 0 : .is_some()
8566 0 : {
8567 0 : reconciles_spawned += 1;
8568 0 : }
8569 0 : }
8570 :
8571 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
8572 0 : break;
8573 0 : }
8574 : }
8575 :
8576 0 : if was_work_filtered {
8577 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
8578 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
8579 0 : // as these validations start passing.
8580 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
8581 0 : }
8582 :
8583 0 : reconciles_spawned
8584 0 : }
8585 :
8586 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
8587 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
8588 : // this higher than the execution limit gives us a chance to execute some work even if the first
8589 : // few optimizations we find are not ready.
8590 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
8591 :
8592 0 : let mut work = Vec::new();
8593 0 : let mut locked = self.inner.write().unwrap();
8594 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
8595 :
8596 : // We are going to plan a bunch of optimisations before applying any of them, so the
8597 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
8598 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
8599 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
8600 : // to ignore the utilisation component of the score.
8601 :
8602 0 : for (_tenant_id, schedule_context, shards) in
8603 0 : TenantShardExclusiveIterator::new(tenants, ScheduleMode::Speculative)
8604 : {
8605 0 : for shard in shards {
8606 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
8607 0 : break;
8608 0 : }
8609 0 : match shard.get_scheduling_policy() {
8610 0 : ShardSchedulingPolicy::Active => {
8611 0 : // Ok to do optimization
8612 0 : }
8613 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
8614 0 : // Ok to do optimization: we are executing a graceful migration that
8615 0 : // has set preferred_node
8616 0 : }
8617 : ShardSchedulingPolicy::Essential
8618 : | ShardSchedulingPolicy::Pause
8619 : | ShardSchedulingPolicy::Stop => {
8620 : // Policy prevents optimizing this shard.
8621 0 : continue;
8622 : }
8623 : }
8624 :
8625 0 : if !matches!(shard.splitting, SplitState::Idle)
8626 0 : || matches!(shard.policy, PlacementPolicy::Detached)
8627 0 : || shard.reconciler.is_some()
8628 : {
8629 : // Do not start any optimizations while another change to the tenant is ongoing: this
8630 : // is not necessary for correctness, but simplifies operations and implicitly throttles
8631 : // optimization changes to happen in a "trickle" over time.
8632 0 : continue;
8633 0 : }
8634 :
8635 : // Fast path: we may quickly identify shards that don't have any possible optimisations
8636 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
8637 0 : if cfg!(feature = "testing") {
8638 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
8639 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
8640 : // panic in prod if we hit this, or spend cycles on it in prod.
8641 0 : assert!(
8642 0 : shard
8643 0 : .optimize_attachment(scheduler, &schedule_context)
8644 0 : .is_none()
8645 : );
8646 0 : assert!(
8647 0 : shard
8648 0 : .optimize_secondary(scheduler, &schedule_context)
8649 0 : .is_none()
8650 : );
8651 0 : }
8652 0 : continue;
8653 0 : }
8654 :
8655 0 : if let Some(optimization) =
8656 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
8657 : // its primary location based on soft constraints, cut it over.
8658 0 : shard.optimize_attachment(scheduler, &schedule_context)
8659 : {
8660 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
8661 0 : work.push((shard.tenant_shard_id, optimization));
8662 0 : break;
8663 0 : } else if let Some(optimization) =
8664 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
8665 : // better placed on another node, based on ScheduleContext, then adjust it. This
8666 : // covers cases like after a shard split, where we might have too many shards
8667 : // in the same tenant with secondary locations on the node where they originally split.
8668 0 : shard.optimize_secondary(scheduler, &schedule_context)
8669 : {
8670 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
8671 0 : work.push((shard.tenant_shard_id, optimization));
8672 0 : break;
8673 0 : }
8674 : }
8675 : }
8676 :
8677 0 : work
8678 0 : }
8679 :
8680 0 : async fn optimize_all_validate(
8681 0 : &self,
8682 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
8683 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
8684 : // Take a clone of the node map to use outside the lock in async validation phase
8685 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
8686 :
8687 0 : let mut want_secondary_status = Vec::new();
8688 :
8689 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
8690 : // check that the state of locations is acceptable to run the optimization, such as
8691 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
8692 : // in a live migration.
8693 0 : let mut validated_work = Vec::new();
8694 0 : for (tenant_shard_id, optimization) in candidate_work {
8695 0 : match optimization.action {
8696 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
8697 : old_attached_node_id: _,
8698 0 : new_attached_node_id,
8699 : }) => {
8700 0 : match validation_nodes.get(&new_attached_node_id) {
8701 0 : None => {
8702 0 : // Node was dropped between planning and validation
8703 0 : }
8704 0 : Some(node) => {
8705 0 : if !node.is_available() {
8706 0 : tracing::info!(
8707 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
8708 : );
8709 0 : } else {
8710 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
8711 0 : // remote API requests concurrently.
8712 0 : want_secondary_status.push((
8713 0 : tenant_shard_id,
8714 0 : node.clone(),
8715 0 : optimization,
8716 0 : ));
8717 0 : }
8718 : }
8719 : }
8720 : }
8721 : ScheduleOptimizationAction::ReplaceSecondary(_)
8722 : | ScheduleOptimizationAction::CreateSecondary(_)
8723 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
8724 : // No extra checks needed to manage secondaries: this does not interrupt client access
8725 0 : validated_work.push((tenant_shard_id, optimization))
8726 : }
8727 : };
8728 : }
8729 :
8730 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
8731 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
8732 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
8733 0 : let results = self
8734 0 : .tenant_for_shards_api(
8735 0 : want_secondary_status
8736 0 : .iter()
8737 0 : .map(|i| (i.0, i.1.clone()))
8738 0 : .collect(),
8739 0 : |tenant_shard_id, client| async move {
8740 0 : client.tenant_secondary_status(tenant_shard_id).await
8741 0 : },
8742 : 1,
8743 : 1,
8744 : SHORT_RECONCILE_TIMEOUT,
8745 0 : &self.cancel,
8746 : )
8747 0 : .await;
8748 :
8749 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
8750 0 : want_secondary_status.into_iter().zip(results.into_iter())
8751 : {
8752 0 : match secondary_status {
8753 0 : Err(e) => {
8754 0 : tracing::info!(
8755 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
8756 : );
8757 : }
8758 0 : Ok(progress) => {
8759 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
8760 : // them in an optimization
8761 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
8762 :
8763 0 : if progress.heatmap_mtime.is_none()
8764 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
8765 0 : && progress.bytes_downloaded != progress.bytes_total
8766 0 : || progress.bytes_total - progress.bytes_downloaded
8767 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
8768 : {
8769 0 : tracing::info!(
8770 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
8771 : );
8772 :
8773 0 : if progress.heatmap_mtime.is_none() {
8774 : // No heatmap might mean the attached location has never uploaded one, or that
8775 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
8776 : // but fairly common in tests.
8777 0 : self.kick_secondary_download(tenant_shard_id).await;
8778 0 : }
8779 : } else {
8780 : // Location looks ready: proceed
8781 0 : tracing::info!(
8782 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
8783 : );
8784 0 : validated_work.push((tenant_shard_id, optimization))
8785 : }
8786 : }
8787 : }
8788 : }
8789 :
8790 0 : validated_work
8791 0 : }
8792 :
8793 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
8794 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
8795 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
8796 : /// we have this helper to move things along faster.
8797 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
8798 0 : if !self.config.kick_secondary_downloads {
8799 : // No-op if kick_secondary_downloads functionaliuty is not configured
8800 0 : return;
8801 0 : }
8802 :
8803 0 : let (attached_node, secondaries) = {
8804 0 : let locked = self.inner.read().unwrap();
8805 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
8806 0 : tracing::warn!(
8807 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
8808 : );
8809 0 : return;
8810 : };
8811 :
8812 0 : let Some(attached) = shard.intent.get_attached() else {
8813 0 : tracing::warn!(
8814 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
8815 : );
8816 0 : return;
8817 : };
8818 :
8819 0 : let secondaries = shard
8820 0 : .intent
8821 0 : .get_secondary()
8822 0 : .iter()
8823 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
8824 0 : .collect::<Vec<_>>();
8825 :
8826 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
8827 : };
8828 :
8829 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
8830 : // a 'kick' to let scheduling optimisation run more promptly.
8831 0 : match attached_node
8832 0 : .with_client_retries(
8833 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
8834 0 : &self.http_client,
8835 0 : &self.config.pageserver_jwt_token,
8836 : 3,
8837 : 10,
8838 : SHORT_RECONCILE_TIMEOUT,
8839 0 : &self.cancel,
8840 : )
8841 0 : .await
8842 : {
8843 0 : Some(Err(e)) => {
8844 0 : tracing::info!(
8845 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
8846 : );
8847 : }
8848 : None => {
8849 0 : tracing::info!(
8850 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
8851 : );
8852 : }
8853 : Some(Ok(_)) => {
8854 0 : tracing::info!(
8855 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
8856 : );
8857 : }
8858 : }
8859 :
8860 0 : for secondary_node in secondaries {
8861 0 : match secondary_node
8862 0 : .with_client_retries(
8863 0 : |client| async move {
8864 0 : client
8865 0 : .tenant_secondary_download(
8866 0 : tenant_shard_id,
8867 0 : Some(Duration::from_secs(1)),
8868 0 : )
8869 0 : .await
8870 0 : },
8871 0 : &self.http_client,
8872 0 : &self.config.pageserver_jwt_token,
8873 : 3,
8874 : 10,
8875 : SHORT_RECONCILE_TIMEOUT,
8876 0 : &self.cancel,
8877 : )
8878 0 : .await
8879 : {
8880 0 : Some(Err(e)) => {
8881 0 : tracing::info!(
8882 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
8883 : );
8884 : }
8885 : None => {
8886 0 : tracing::info!(
8887 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
8888 : );
8889 : }
8890 0 : Some(Ok(progress)) => {
8891 0 : tracing::info!(
8892 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
8893 : );
8894 : }
8895 : }
8896 : }
8897 0 : }
8898 :
8899 : /// Asynchronously split a tenant that's eligible for automatic splits. At most one tenant will
8900 : /// be split per call.
8901 : ///
8902 : /// Two sets of criteria are used: initial splits and size-based splits (in that order).
8903 : /// Initial splits are used to eagerly split unsharded tenants that may be performing initial
8904 : /// ingestion, since sharded tenants have significantly better ingestion throughput. Size-based
8905 : /// splits are used to bound the maximum shard size and balance out load.
8906 : ///
8907 : /// Splits are based on max_logical_size, i.e. the logical size of the largest timeline in a
8908 : /// tenant. We use this instead of the total logical size because branches will duplicate
8909 : /// logical size without actually using more storage. We could also use visible physical size,
8910 : /// but this might overestimate tenants that frequently churn branches.
8911 : ///
8912 : /// Initial splits (initial_split_threshold):
8913 : /// * Applies to tenants with 1 shard.
8914 : /// * The largest timeline (max_logical_size) exceeds initial_split_threshold.
8915 : /// * Splits into initial_split_shards.
8916 : ///
8917 : /// Size-based splits (split_threshold):
8918 : /// * Applies to all tenants.
8919 : /// * The largest timeline (max_logical_size) divided by shard count exceeds split_threshold.
8920 : /// * Splits such that max_logical_size / shard_count <= split_threshold, in powers of 2.
8921 : ///
8922 : /// Tenant shards are ordered by descending max_logical_size, first initial split candidates
8923 : /// then size-based split candidates. The first matching candidate is split.
8924 : ///
8925 : /// The shard count is clamped to max_split_shards. If a candidate is eligible for both initial
8926 : /// and size-based splits, the largest shard count will be used.
8927 : ///
8928 : /// An unsharded tenant will get DEFAULT_STRIPE_SIZE, regardless of what its ShardIdentity says.
8929 : /// A sharded tenant will retain its stripe size, as splits do not allow changing it.
8930 : ///
8931 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
8932 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
8933 : /// block all other splits.
8934 0 : async fn autosplit_tenants(self: &Arc<Self>) {
8935 : // If max_split_shards is set to 0 or 1, we can't split.
8936 0 : let max_split_shards = self.config.max_split_shards;
8937 0 : if max_split_shards <= 1 {
8938 0 : return;
8939 0 : }
8940 :
8941 : // If initial_split_shards is set to 0 or 1, disable initial splits.
8942 0 : let mut initial_split_threshold = self.config.initial_split_threshold.unwrap_or(0);
8943 0 : let initial_split_shards = self.config.initial_split_shards;
8944 0 : if initial_split_shards <= 1 {
8945 0 : initial_split_threshold = 0;
8946 0 : }
8947 :
8948 : // If no split_threshold nor initial_split_threshold, disable autosplits.
8949 0 : let split_threshold = self.config.split_threshold.unwrap_or(0);
8950 0 : if split_threshold == 0 && initial_split_threshold == 0 {
8951 0 : return;
8952 0 : }
8953 :
8954 : // Fetch split candidates in prioritized order.
8955 : //
8956 : // If initial splits are enabled, fetch eligible tenants first. We prioritize initial splits
8957 : // over size-based splits, since these are often performing initial ingestion and rely on
8958 : // splits to improve ingest throughput.
8959 0 : let mut candidates = Vec::new();
8960 :
8961 0 : if initial_split_threshold > 0 {
8962 : // Initial splits: fetch tenants with 1 shard where the logical size of the largest
8963 : // timeline exceeds the initial split threshold.
8964 0 : let initial_candidates = self
8965 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
8966 0 : order_by: TenantSorting::MaxLogicalSize,
8967 0 : limit: 10,
8968 0 : where_shards_lt: Some(ShardCount(2)),
8969 0 : where_gt: Some(initial_split_threshold),
8970 0 : })
8971 0 : .await;
8972 0 : candidates.extend(initial_candidates);
8973 0 : }
8974 :
8975 0 : if split_threshold > 0 {
8976 : // Size-based splits: fetch tenants where the logical size of the largest timeline
8977 : // divided by shard count exceeds the split threshold.
8978 : //
8979 : // max_logical_size is only tracked on shard 0, and contains the total logical size
8980 : // across all shards. We have to order and filter by MaxLogicalSizePerShard, i.e.
8981 : // max_logical_size / shard_count, such that we only receive tenants that are actually
8982 : // eligible for splits. But we still use max_logical_size for later split calculations.
8983 0 : let size_candidates = self
8984 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
8985 0 : order_by: TenantSorting::MaxLogicalSizePerShard,
8986 0 : limit: 10,
8987 0 : where_shards_lt: Some(ShardCount(max_split_shards)),
8988 0 : where_gt: Some(split_threshold),
8989 0 : })
8990 0 : .await;
8991 : #[cfg(feature = "testing")]
8992 0 : assert!(
8993 0 : size_candidates.iter().all(|c| c.id.is_shard_zero()),
8994 0 : "MaxLogicalSizePerShard returned non-zero shard: {size_candidates:?}",
8995 : );
8996 0 : candidates.extend(size_candidates);
8997 0 : }
8998 :
8999 : // Filter out tenants in a prohibiting scheduling modes
9000 : // and tenants with an ongoing import.
9001 : //
9002 : // Note that the import check here is oportunistic. An import might start
9003 : // after the check before we actually update [`TenantShard::splitting`].
9004 : // [`Self::tenant_shard_split`] checks the database whilst holding the exclusive
9005 : // tenant lock. Imports might take a long time, so the check here allows us
9006 : // to split something else instead of trying the same shard over and over.
9007 : {
9008 0 : let state = self.inner.read().unwrap();
9009 0 : candidates.retain(|i| {
9010 0 : let shard = state.tenants.get(&i.id);
9011 0 : match shard {
9012 0 : Some(t) => {
9013 0 : t.get_scheduling_policy() == ShardSchedulingPolicy::Active
9014 0 : && t.importing == TimelineImportState::Idle
9015 : }
9016 0 : None => false,
9017 : }
9018 0 : });
9019 : }
9020 :
9021 : // Pick the first candidate to split. This will generally always be the first one in
9022 : // candidates, but we defensively skip candidates that end up not actually splitting.
9023 0 : let Some((candidate, new_shard_count)) = candidates
9024 0 : .into_iter()
9025 0 : .filter_map(|candidate| {
9026 0 : let new_shard_count = Self::compute_split_shards(ShardSplitInputs {
9027 0 : shard_count: candidate.id.shard_count,
9028 0 : max_logical_size: candidate.max_logical_size,
9029 0 : split_threshold,
9030 0 : max_split_shards,
9031 0 : initial_split_threshold,
9032 0 : initial_split_shards,
9033 0 : });
9034 0 : new_shard_count.map(|shards| (candidate, shards.count()))
9035 0 : })
9036 0 : .next()
9037 : else {
9038 0 : debug!("no split-eligible tenants found");
9039 0 : return;
9040 : };
9041 :
9042 : // Retain the stripe size of sharded tenants, as splits don't allow changing it. Otherwise,
9043 : // use DEFAULT_STRIPE_SIZE for unsharded tenants -- their stripe size doesn't really matter,
9044 : // and if we change the default stripe size we want to use the new default rather than an
9045 : // old, persisted stripe size.
9046 0 : let new_stripe_size = match candidate.id.shard_count.count() {
9047 0 : 0 => panic!("invalid shard count 0"),
9048 0 : 1 => Some(DEFAULT_STRIPE_SIZE),
9049 0 : 2.. => None,
9050 : };
9051 :
9052 : // We spawn a task to run this, so it's exactly like some external API client requesting
9053 : // it. We don't want to block the background reconcile loop on this.
9054 0 : let old_shard_count = candidate.id.shard_count.count();
9055 0 : info!(
9056 0 : "auto-splitting tenant {old_shard_count} → {new_shard_count} shards, \
9057 0 : current size {candidate:?} (split_threshold={split_threshold} \
9058 0 : initial_split_threshold={initial_split_threshold})"
9059 : );
9060 :
9061 0 : let this = self.clone();
9062 0 : tokio::spawn(
9063 0 : async move {
9064 0 : match this
9065 0 : .tenant_shard_split(
9066 0 : candidate.id.tenant_id,
9067 0 : TenantShardSplitRequest {
9068 0 : new_shard_count,
9069 0 : new_stripe_size,
9070 0 : },
9071 0 : )
9072 0 : .await
9073 : {
9074 : Ok(_) => {
9075 0 : info!("successful auto-split {old_shard_count} → {new_shard_count} shards")
9076 : }
9077 0 : Err(err) => error!("auto-split failed: {err}"),
9078 : }
9079 0 : }
9080 0 : .instrument(info_span!("auto_split", tenant_id=%candidate.id.tenant_id)),
9081 : );
9082 0 : }
9083 :
9084 : /// Returns the number of shards to split a tenant into, or None if the tenant shouldn't split,
9085 : /// based on the total logical size of the largest timeline summed across all shards. Uses the
9086 : /// larger of size-based and initial splits, clamped to max_split_shards.
9087 : ///
9088 : /// NB: the thresholds are exclusive, since TopTenantShardsRequest uses where_gt.
9089 25 : fn compute_split_shards(inputs: ShardSplitInputs) -> Option<ShardCount> {
9090 : let ShardSplitInputs {
9091 25 : shard_count,
9092 25 : max_logical_size,
9093 25 : split_threshold,
9094 25 : max_split_shards,
9095 25 : initial_split_threshold,
9096 25 : initial_split_shards,
9097 25 : } = inputs;
9098 :
9099 25 : let mut new_shard_count: u8 = shard_count.count();
9100 :
9101 : // Size-based splits. Ensures max_logical_size / new_shard_count <= split_threshold, using
9102 : // power-of-two shard counts.
9103 : //
9104 : // If the current shard count is not a power of two, and does not exceed split_threshold,
9105 : // then we leave it alone rather than forcing a power-of-two split.
9106 25 : if split_threshold > 0
9107 18 : && max_logical_size.div_ceil(split_threshold) > shard_count.count() as u64
9108 12 : {
9109 12 : new_shard_count = max_logical_size
9110 12 : .div_ceil(split_threshold)
9111 12 : .checked_next_power_of_two()
9112 12 : .unwrap_or(u8::MAX as u64)
9113 12 : .try_into()
9114 12 : .unwrap_or(u8::MAX);
9115 13 : }
9116 :
9117 : // Initial splits. Use the larger of size-based and initial split shard counts. This only
9118 : // applies to unsharded tenants, i.e. changes to initial_split_threshold or
9119 : // initial_split_shards are not retroactive for sharded tenants.
9120 25 : if initial_split_threshold > 0
9121 14 : && shard_count.count() <= 1
9122 11 : && max_logical_size > initial_split_threshold
9123 8 : {
9124 8 : new_shard_count = new_shard_count.max(initial_split_shards);
9125 17 : }
9126 :
9127 : // Clamp to max shards.
9128 25 : new_shard_count = new_shard_count.min(max_split_shards);
9129 :
9130 : // Don't split if we're not increasing the shard count.
9131 25 : if new_shard_count <= shard_count.count() {
9132 10 : return None;
9133 15 : }
9134 :
9135 15 : Some(ShardCount(new_shard_count))
9136 25 : }
9137 :
9138 : /// Fetches the top tenant shards from every available node, in descending order of
9139 : /// max logical size. Offline nodes are skipped, and any errors from available nodes
9140 : /// will be logged and ignored.
9141 0 : async fn get_top_tenant_shards(
9142 0 : &self,
9143 0 : request: &TopTenantShardsRequest,
9144 0 : ) -> Vec<TopTenantShardItem> {
9145 0 : let nodes = self
9146 0 : .inner
9147 0 : .read()
9148 0 : .unwrap()
9149 0 : .nodes
9150 0 : .values()
9151 0 : .filter(|node| node.is_available())
9152 0 : .cloned()
9153 0 : .collect_vec();
9154 :
9155 0 : let mut futures = FuturesUnordered::new();
9156 0 : for node in nodes {
9157 0 : futures.push(async move {
9158 0 : node.with_client_retries(
9159 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
9160 0 : &self.http_client,
9161 0 : &self.config.pageserver_jwt_token,
9162 : 3,
9163 : 3,
9164 0 : Duration::from_secs(5),
9165 0 : &self.cancel,
9166 : )
9167 0 : .await
9168 0 : });
9169 : }
9170 :
9171 0 : let mut top = Vec::new();
9172 0 : while let Some(output) = futures.next().await {
9173 0 : match output {
9174 0 : Some(Ok(response)) => top.extend(response.shards),
9175 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
9176 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
9177 0 : None => {} // node is shutting down
9178 : }
9179 : }
9180 :
9181 0 : top.sort_by_key(|i| i.max_logical_size);
9182 0 : top.reverse();
9183 0 : top
9184 0 : }
9185 :
9186 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
9187 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
9188 : /// put the system into a quiescent state where future background reconciliations won't do anything.
9189 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
9190 0 : let reconcile_all_result = self.reconcile_all();
9191 0 : let mut spawned_reconciles = reconcile_all_result.spawned_reconciles;
9192 0 : if reconcile_all_result.can_run_optimizations() {
9193 : // Only optimize when we are otherwise idle
9194 0 : let optimization_reconciles = self.optimize_all().await;
9195 0 : spawned_reconciles += optimization_reconciles;
9196 0 : }
9197 :
9198 0 : let waiters = {
9199 0 : let mut waiters = Vec::new();
9200 0 : let locked = self.inner.read().unwrap();
9201 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
9202 0 : if let Some(waiter) = shard.get_waiter() {
9203 0 : waiters.push(waiter);
9204 0 : }
9205 : }
9206 0 : waiters
9207 : };
9208 :
9209 0 : let waiter_count = waiters.len();
9210 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
9211 0 : Ok(()) => {}
9212 0 : Err(e) => {
9213 0 : if let ReconcileWaitError::Failed(_, reconcile_error) = &e {
9214 0 : match **reconcile_error {
9215 : ReconcileError::Cancel
9216 0 : | ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
9217 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
9218 0 : // because some other change superceded it. We will return a nonzero number,
9219 0 : // so the caller knows they might have to call again to quiesce the system.
9220 0 : }
9221 : _ => {
9222 0 : return Err(e);
9223 : }
9224 : }
9225 : } else {
9226 0 : return Err(e);
9227 : }
9228 : }
9229 : };
9230 :
9231 0 : tracing::info!(
9232 0 : "{} reconciles in reconcile_all, {} waiters",
9233 : spawned_reconciles,
9234 : waiter_count
9235 : );
9236 :
9237 0 : Ok(std::cmp::max(waiter_count, spawned_reconciles))
9238 0 : }
9239 :
9240 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
9241 : // Cancel all on-going reconciles and wait for them to exit the gate.
9242 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
9243 0 : self.reconcilers_cancel.cancel();
9244 0 : self.reconcilers_gate.close().await;
9245 :
9246 : // Signal the background loop in [`Service::process_results`] to exit once
9247 : // it has proccessed the results from all the reconciles we cancelled earlier.
9248 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
9249 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
9250 0 : self.result_tx.closed().await;
9251 0 : }
9252 :
9253 0 : pub async fn shutdown(&self) {
9254 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
9255 0 : .await;
9256 :
9257 : // Background tasks hold gate guards: this notifies them of the cancellation and
9258 : // waits for them all to complete.
9259 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
9260 0 : self.cancel.cancel();
9261 0 : self.gate.close().await;
9262 0 : }
9263 :
9264 : /// Spot check the download lag for a secondary location of a shard.
9265 : /// Should be used as a heuristic, since it's not always precise: the
9266 : /// secondary might have not downloaded the new heat map yet and, hence,
9267 : /// is not aware of the lag.
9268 : ///
9269 : /// Returns:
9270 : /// * Ok(None) if the lag could not be determined from the status,
9271 : /// * Ok(Some(_)) if the lag could be determind
9272 : /// * Err on failures to query the pageserver.
9273 0 : async fn secondary_lag(
9274 0 : &self,
9275 0 : secondary: &NodeId,
9276 0 : tenant_shard_id: TenantShardId,
9277 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
9278 0 : let nodes = self.inner.read().unwrap().nodes.clone();
9279 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
9280 0 : StatusCode::NOT_FOUND,
9281 0 : format!("Node with id {secondary} not found"),
9282 0 : ))?;
9283 :
9284 0 : match node
9285 0 : .with_client_retries(
9286 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
9287 0 : &self.http_client,
9288 0 : &self.config.pageserver_jwt_token,
9289 : 1,
9290 : 3,
9291 0 : Duration::from_millis(250),
9292 0 : &self.cancel,
9293 : )
9294 0 : .await
9295 : {
9296 0 : Some(Ok(status)) => match status.heatmap_mtime {
9297 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
9298 0 : None => Ok(None),
9299 : },
9300 0 : Some(Err(e)) => Err(e),
9301 0 : None => Err(mgmt_api::Error::Cancelled),
9302 : }
9303 0 : }
9304 :
9305 : /// Drain a node by moving the shards attached to it as primaries.
9306 : /// This is a long running operation and it should run as a separate Tokio task.
9307 0 : pub(crate) async fn drain_node(
9308 0 : self: &Arc<Self>,
9309 0 : node_id: NodeId,
9310 0 : cancel: CancellationToken,
9311 0 : ) -> Result<(), OperationError> {
9312 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
9313 0 : let max_secondary_lag_bytes = self
9314 0 : .config
9315 0 : .max_secondary_lag_bytes
9316 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
9317 :
9318 : // By default, live migrations are generous about the wait time for getting
9319 : // the secondary location up to speed. When draining, give up earlier in order
9320 : // to not stall the operation when a cold secondary is encountered.
9321 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9322 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9323 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9324 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9325 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9326 0 : .build();
9327 :
9328 0 : let mut waiters = Vec::new();
9329 :
9330 0 : let mut tid_iter = create_shared_shard_iterator(self.clone());
9331 :
9332 0 : while !tid_iter.finished() {
9333 0 : if cancel.is_cancelled() {
9334 0 : match self
9335 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9336 0 : .await
9337 : {
9338 0 : Ok(()) => return Err(OperationError::Cancelled),
9339 0 : Err(err) => {
9340 0 : return Err(OperationError::FinalizeError(
9341 0 : format!(
9342 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9343 0 : )
9344 0 : .into(),
9345 0 : ));
9346 : }
9347 : }
9348 0 : }
9349 :
9350 0 : operation_utils::validate_node_state(
9351 0 : &node_id,
9352 0 : self.inner.read().unwrap().nodes.clone(),
9353 0 : NodeSchedulingPolicy::Draining,
9354 0 : )?;
9355 :
9356 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9357 0 : let tid = match tid_iter.next() {
9358 0 : Some(tid) => tid,
9359 : None => {
9360 0 : break;
9361 : }
9362 : };
9363 :
9364 0 : let tid_drain = TenantShardDrain {
9365 0 : drained_node: node_id,
9366 0 : tenant_shard_id: tid,
9367 0 : };
9368 :
9369 0 : let dest_node_id = {
9370 0 : let locked = self.inner.read().unwrap();
9371 :
9372 0 : match tid_drain
9373 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
9374 : {
9375 0 : Some(node_id) => node_id,
9376 : None => {
9377 0 : continue;
9378 : }
9379 : }
9380 : };
9381 :
9382 0 : match self.secondary_lag(&dest_node_id, tid).await {
9383 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
9384 0 : // The secondary is reasonably up to date.
9385 0 : // Migrate to it
9386 0 : }
9387 0 : Ok(Some(lag)) => {
9388 0 : tracing::info!(
9389 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9390 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
9391 : );
9392 0 : continue;
9393 : }
9394 : Ok(None) => {
9395 0 : tracing::info!(
9396 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9397 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
9398 : );
9399 0 : continue;
9400 : }
9401 0 : Err(err) => {
9402 0 : tracing::warn!(
9403 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9404 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
9405 : );
9406 0 : continue;
9407 : }
9408 : }
9409 :
9410 : {
9411 0 : let mut locked = self.inner.write().unwrap();
9412 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9413 0 : let rescheduled = tid_drain.reschedule_to_secondary(
9414 0 : dest_node_id,
9415 0 : tenants,
9416 0 : scheduler,
9417 0 : nodes,
9418 0 : )?;
9419 :
9420 0 : if let Some(tenant_shard) = rescheduled {
9421 0 : let waiter = self.maybe_configured_reconcile_shard(
9422 0 : tenant_shard,
9423 0 : nodes,
9424 0 : reconciler_config,
9425 0 : );
9426 0 : if let Some(some) = waiter {
9427 0 : waiters.push(some);
9428 0 : }
9429 0 : }
9430 : }
9431 : }
9432 :
9433 0 : waiters = self
9434 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9435 0 : .await;
9436 :
9437 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
9438 : }
9439 :
9440 0 : while !waiters.is_empty() {
9441 0 : if cancel.is_cancelled() {
9442 0 : match self
9443 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9444 0 : .await
9445 : {
9446 0 : Ok(()) => return Err(OperationError::Cancelled),
9447 0 : Err(err) => {
9448 0 : return Err(OperationError::FinalizeError(
9449 0 : format!(
9450 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9451 0 : )
9452 0 : .into(),
9453 0 : ));
9454 : }
9455 : }
9456 0 : }
9457 :
9458 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
9459 :
9460 0 : waiters = self
9461 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9462 0 : .await;
9463 : }
9464 :
9465 : // At this point we have done the best we could to drain shards from this node.
9466 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
9467 : // to complete the drain.
9468 0 : if let Err(err) = self
9469 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
9470 0 : .await
9471 : {
9472 : // This is not fatal. Anything that is polling the node scheduling policy to detect
9473 : // the end of the drain operations will hang, but all such places should enforce an
9474 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
9475 : // by the counterpart fill operation.
9476 0 : return Err(OperationError::FinalizeError(
9477 0 : format!(
9478 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
9479 0 : )
9480 0 : .into(),
9481 0 : ));
9482 0 : }
9483 :
9484 0 : Ok(())
9485 0 : }
9486 :
9487 : /// Create a node fill plan (pick secondaries to promote), based on:
9488 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
9489 : /// outside their home AZ, should be migrated back here.
9490 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
9491 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
9492 : /// those shards have a home AZ that doesn't match the node we're filling.
9493 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
9494 0 : let mut locked = self.inner.write().unwrap();
9495 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
9496 :
9497 0 : let node_az = nodes
9498 0 : .get(&node_id)
9499 0 : .expect("Node must exist")
9500 0 : .get_availability_zone_id()
9501 0 : .clone();
9502 :
9503 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
9504 0 : let mut plan = Vec::new();
9505 :
9506 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
9507 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
9508 :
9509 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
9510 : // conceivably come up in real life if deploying a single-AZ region intentionally.
9511 0 : let respect_azs = nodes
9512 0 : .values()
9513 0 : .map(|n| n.get_availability_zone_id())
9514 0 : .unique()
9515 0 : .count()
9516 : > 1;
9517 :
9518 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
9519 : // requires it.
9520 0 : for (tsid, tenant_shard) in tenants {
9521 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9522 : // Shard doesn't have a secondary on this node, ignore it.
9523 0 : continue;
9524 0 : }
9525 :
9526 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
9527 : // shards which belong on this node, not to promote shards whose scheduling preference
9528 : // would be on their currently attached node. So will avoid promoting shards whose
9529 : // home AZ doesn't match the AZ of the node we're filling.
9530 0 : match tenant_shard.preferred_az() {
9531 0 : _ if !respect_azs => {
9532 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9533 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9534 0 : }
9535 : }
9536 : None => {
9537 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
9538 : // will only do so if our target shard count requires it.
9539 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9540 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
9541 0 : }
9542 : }
9543 0 : Some(az) if az == &node_az => {
9544 : // This shard's home AZ is equal to the node we're filling: it should
9545 : // be moved back to this node as part of filling, unless its currently
9546 : // attached location is also in its home AZ.
9547 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
9548 0 : if nodes
9549 0 : .get(primary)
9550 0 : .expect("referenced node must exist")
9551 0 : .get_availability_zone_id()
9552 0 : != tenant_shard
9553 0 : .preferred_az()
9554 0 : .expect("tenant must have an AZ preference")
9555 : {
9556 0 : plan.push(*tsid)
9557 0 : }
9558 : } else {
9559 0 : plan.push(*tsid)
9560 : }
9561 : }
9562 0 : Some(_) => {
9563 0 : // This shard's home AZ is somewhere other than the node we're filling,
9564 0 : // it may not be moved back to this node as part of filling. Ignore it
9565 0 : }
9566 : }
9567 : }
9568 :
9569 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
9570 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
9571 :
9572 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
9573 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
9574 :
9575 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
9576 :
9577 0 : for (node_id, attached) in nodes_by_load {
9578 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
9579 0 : if !available {
9580 0 : continue;
9581 0 : }
9582 :
9583 0 : if plan.len() >= fill_requirement
9584 0 : || free_tids_by_node.is_empty()
9585 0 : || attached <= expected_attached
9586 : {
9587 0 : break;
9588 0 : }
9589 :
9590 0 : let can_take = attached - expected_attached;
9591 0 : let needed = fill_requirement - plan.len();
9592 0 : let mut take = std::cmp::min(can_take, needed);
9593 :
9594 0 : let mut remove_node = false;
9595 0 : while take > 0 {
9596 0 : match free_tids_by_node.get_mut(&node_id) {
9597 0 : Some(tids) => match tids.pop() {
9598 0 : Some(tid) => {
9599 0 : let max_promote_for_tenant = std::cmp::max(
9600 0 : tid.shard_count.count() as usize / locked.nodes.len(),
9601 : 1,
9602 : );
9603 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
9604 0 : if *promoted < max_promote_for_tenant {
9605 0 : plan.push(tid);
9606 0 : *promoted += 1;
9607 0 : take -= 1;
9608 0 : }
9609 : }
9610 : None => {
9611 0 : remove_node = true;
9612 0 : break;
9613 : }
9614 : },
9615 : None => {
9616 0 : break;
9617 : }
9618 : }
9619 : }
9620 :
9621 0 : if remove_node {
9622 0 : free_tids_by_node.remove(&node_id);
9623 0 : }
9624 : }
9625 :
9626 0 : plan
9627 0 : }
9628 :
9629 : /// Fill a node by promoting its secondaries until the cluster is balanced
9630 : /// with regards to attached shard counts. Note that this operation only
9631 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
9632 : /// This is a long running operation and it should run as a separate Tokio task.
9633 0 : pub(crate) async fn fill_node(
9634 0 : &self,
9635 0 : node_id: NodeId,
9636 0 : cancel: CancellationToken,
9637 0 : ) -> Result<(), OperationError> {
9638 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(30);
9639 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
9640 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
9641 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
9642 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
9643 0 : .build();
9644 :
9645 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
9646 0 : let mut waiters = Vec::new();
9647 :
9648 : // Execute the plan we've composed above. Before aplying each move from the plan,
9649 : // we validate to ensure that it has not gone stale in the meantime.
9650 0 : while !tids_to_promote.is_empty() {
9651 0 : if cancel.is_cancelled() {
9652 0 : match self
9653 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9654 0 : .await
9655 : {
9656 0 : Ok(()) => return Err(OperationError::Cancelled),
9657 0 : Err(err) => {
9658 0 : return Err(OperationError::FinalizeError(
9659 0 : format!(
9660 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9661 0 : )
9662 0 : .into(),
9663 0 : ));
9664 : }
9665 : }
9666 0 : }
9667 :
9668 : {
9669 0 : let mut locked = self.inner.write().unwrap();
9670 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
9671 :
9672 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
9673 0 : format!("node {node_id} was removed").into(),
9674 0 : ))?;
9675 :
9676 0 : let current_policy = node.get_scheduling();
9677 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
9678 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
9679 : // about it
9680 0 : return Err(OperationError::NodeStateChanged(
9681 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
9682 0 : ));
9683 0 : }
9684 :
9685 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
9686 0 : if let Some(tid) = tids_to_promote.pop() {
9687 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
9688 : // If the node being filled is not a secondary anymore,
9689 : // skip the promotion.
9690 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
9691 0 : continue;
9692 0 : }
9693 :
9694 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
9695 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
9696 0 : Err(e) => {
9697 0 : tracing::warn!(
9698 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9699 0 : "Scheduling error when filling pageserver {} : {e}", node_id
9700 : );
9701 : }
9702 : Ok(()) => {
9703 0 : tracing::info!(
9704 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
9705 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
9706 : node_id,
9707 : previously_attached_to,
9708 : node_id
9709 : );
9710 :
9711 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
9712 0 : tenant_shard,
9713 0 : nodes,
9714 0 : reconciler_config,
9715 0 : ) {
9716 0 : waiters.push(waiter);
9717 0 : }
9718 : }
9719 : }
9720 0 : }
9721 : } else {
9722 0 : break;
9723 : }
9724 : }
9725 : }
9726 :
9727 0 : waiters = self
9728 0 : .await_waiters_remainder(waiters, WAITER_OPERATION_POLL_TIMEOUT)
9729 0 : .await;
9730 : }
9731 :
9732 0 : while !waiters.is_empty() {
9733 0 : if cancel.is_cancelled() {
9734 0 : match self
9735 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9736 0 : .await
9737 : {
9738 0 : Ok(()) => return Err(OperationError::Cancelled),
9739 0 : Err(err) => {
9740 0 : return Err(OperationError::FinalizeError(
9741 0 : format!(
9742 0 : "Failed to finalise drain cancel of {node_id} by setting scheduling policy to Active: {err}"
9743 0 : )
9744 0 : .into(),
9745 0 : ));
9746 : }
9747 : }
9748 0 : }
9749 :
9750 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
9751 :
9752 0 : waiters = self
9753 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
9754 0 : .await;
9755 : }
9756 :
9757 0 : if let Err(err) = self
9758 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
9759 0 : .await
9760 : {
9761 : // This isn't a huge issue since the filling process starts upon request. However, it
9762 : // will prevent the next drain from starting. The only case in which this can fail
9763 : // is database unavailability. Such a case will require manual intervention.
9764 0 : return Err(OperationError::FinalizeError(
9765 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
9766 0 : .into(),
9767 0 : ));
9768 0 : }
9769 :
9770 0 : Ok(())
9771 0 : }
9772 :
9773 : /// Updates scrubber metadata health check results.
9774 0 : pub(crate) async fn metadata_health_update(
9775 0 : &self,
9776 0 : update_req: MetadataHealthUpdateRequest,
9777 0 : ) -> Result<(), ApiError> {
9778 0 : let now = chrono::offset::Utc::now();
9779 0 : let (healthy_records, unhealthy_records) = {
9780 0 : let locked = self.inner.read().unwrap();
9781 0 : let healthy_records = update_req
9782 0 : .healthy_tenant_shards
9783 0 : .into_iter()
9784 : // Retain only health records associated with tenant shards managed by storage controller.
9785 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
9786 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
9787 0 : .collect();
9788 0 : let unhealthy_records = update_req
9789 0 : .unhealthy_tenant_shards
9790 0 : .into_iter()
9791 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
9792 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
9793 0 : .collect();
9794 :
9795 0 : (healthy_records, unhealthy_records)
9796 : };
9797 :
9798 0 : self.persistence
9799 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
9800 0 : .await?;
9801 0 : Ok(())
9802 0 : }
9803 :
9804 : /// Lists the tenant shards that has unhealthy metadata status.
9805 0 : pub(crate) async fn metadata_health_list_unhealthy(
9806 0 : &self,
9807 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
9808 0 : let result = self
9809 0 : .persistence
9810 0 : .list_unhealthy_metadata_health_records()
9811 0 : .await?
9812 0 : .iter()
9813 0 : .map(|p| p.get_tenant_shard_id().unwrap())
9814 0 : .collect();
9815 :
9816 0 : Ok(result)
9817 0 : }
9818 :
9819 : /// Lists the tenant shards that have not been scrubbed for some duration.
9820 0 : pub(crate) async fn metadata_health_list_outdated(
9821 0 : &self,
9822 0 : not_scrubbed_for: Duration,
9823 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
9824 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
9825 0 : let result = self
9826 0 : .persistence
9827 0 : .list_outdated_metadata_health_records(earlier)
9828 0 : .await?
9829 0 : .into_iter()
9830 0 : .map(|record| record.into())
9831 0 : .collect();
9832 0 : Ok(result)
9833 0 : }
9834 :
9835 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
9836 0 : self.inner.read().unwrap().get_leadership_status()
9837 0 : }
9838 :
9839 : /// Handler for step down requests
9840 : ///
9841 : /// Step down runs in separate task since once it's called it should
9842 : /// be driven to completion. Subsequent requests will wait on the same
9843 : /// step down task.
9844 0 : pub(crate) async fn step_down(self: &Arc<Self>) -> GlobalObservedState {
9845 0 : let handle = self.step_down_barrier.get_or_init(|| {
9846 0 : let step_down_self = self.clone();
9847 0 : let (tx, rx) = tokio::sync::watch::channel::<Option<GlobalObservedState>>(None);
9848 0 : tokio::spawn(async move {
9849 0 : let state = step_down_self.step_down_task().await;
9850 0 : tx.send(Some(state))
9851 0 : .expect("Task Arc<Service> keeps receiver alive");
9852 0 : });
9853 :
9854 0 : rx
9855 0 : });
9856 :
9857 0 : handle
9858 0 : .clone()
9859 0 : .wait_for(|observed_state| observed_state.is_some())
9860 0 : .await
9861 0 : .expect("Task Arc<Service> keeps sender alive")
9862 0 : .deref()
9863 0 : .clone()
9864 0 : .expect("Checked above")
9865 0 : }
9866 :
9867 0 : async fn step_down_task(&self) -> GlobalObservedState {
9868 0 : tracing::info!("Received step down request from peer");
9869 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
9870 :
9871 0 : self.inner.write().unwrap().step_down();
9872 :
9873 0 : let stop_reconciliations =
9874 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown);
9875 0 : let mut stop_reconciliations = std::pin::pin!(stop_reconciliations);
9876 :
9877 0 : let started_at = Instant::now();
9878 :
9879 : // Wait for reconciliations to stop and warn if that's taking a long time
9880 : loop {
9881 0 : tokio::select! {
9882 0 : _ = &mut stop_reconciliations => {
9883 0 : tracing::info!("Reconciliations stopped, proceeding with step down");
9884 0 : break;
9885 : }
9886 0 : _ = tokio::time::sleep(Duration::from_secs(10)) => {
9887 0 : tracing::warn!(
9888 0 : elapsed_sec=%started_at.elapsed().as_secs(),
9889 0 : "Stopping reconciliations during step down is taking too long"
9890 : );
9891 : }
9892 : }
9893 : }
9894 :
9895 0 : let mut global_observed = GlobalObservedState::default();
9896 0 : let locked = self.inner.read().unwrap();
9897 0 : for (tid, tenant_shard) in locked.tenants.iter() {
9898 0 : global_observed
9899 0 : .0
9900 0 : .insert(*tid, tenant_shard.observed.clone());
9901 0 : }
9902 :
9903 0 : global_observed
9904 0 : }
9905 :
9906 0 : pub(crate) async fn update_shards_preferred_azs(
9907 0 : &self,
9908 0 : req: ShardsPreferredAzsRequest,
9909 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
9910 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
9911 0 : let updated = self
9912 0 : .persistence
9913 0 : .set_tenant_shard_preferred_azs(preferred_azs)
9914 0 : .await
9915 0 : .map_err(|err| {
9916 0 : ApiError::InternalServerError(anyhow::anyhow!(
9917 0 : "Failed to persist preferred AZs: {err}"
9918 0 : ))
9919 0 : })?;
9920 :
9921 0 : let mut updated_in_mem_and_db = Vec::default();
9922 :
9923 0 : let mut locked = self.inner.write().unwrap();
9924 0 : let state = locked.deref_mut();
9925 0 : for (tid, az_id) in updated {
9926 0 : let shard = state.tenants.get_mut(&tid);
9927 0 : if let Some(shard) = shard {
9928 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
9929 0 : updated_in_mem_and_db.push(tid);
9930 0 : }
9931 : }
9932 :
9933 0 : Ok(ShardsPreferredAzsResponse {
9934 0 : updated: updated_in_mem_and_db,
9935 0 : })
9936 0 : }
9937 : }
9938 :
9939 : #[cfg(test)]
9940 : mod tests {
9941 : use super::*;
9942 :
9943 : /// Tests Service::compute_split_shards. For readability, this specifies sizes in GBs rather
9944 : /// than bytes. Note that max_logical_size is the total logical size of the largest timeline
9945 : /// summed across all shards.
9946 : #[test]
9947 1 : fn compute_split_shards() {
9948 : // Size-based split: two shards have a 500 GB timeline, which need to split into 8 shards
9949 : // that are <= 64 GB,
9950 1 : assert_eq!(
9951 1 : Service::compute_split_shards(ShardSplitInputs {
9952 1 : shard_count: ShardCount(2),
9953 1 : max_logical_size: 500,
9954 1 : split_threshold: 64,
9955 1 : max_split_shards: 16,
9956 1 : initial_split_threshold: 0,
9957 1 : initial_split_shards: 0,
9958 1 : }),
9959 : Some(ShardCount(8))
9960 : );
9961 :
9962 : // Size-based split: noop at or below threshold, fires above.
9963 1 : assert_eq!(
9964 1 : Service::compute_split_shards(ShardSplitInputs {
9965 1 : shard_count: ShardCount(2),
9966 1 : max_logical_size: 127,
9967 1 : split_threshold: 64,
9968 1 : max_split_shards: 16,
9969 1 : initial_split_threshold: 0,
9970 1 : initial_split_shards: 0,
9971 1 : }),
9972 : None,
9973 : );
9974 1 : assert_eq!(
9975 1 : Service::compute_split_shards(ShardSplitInputs {
9976 1 : shard_count: ShardCount(2),
9977 1 : max_logical_size: 128,
9978 1 : split_threshold: 64,
9979 1 : max_split_shards: 16,
9980 1 : initial_split_threshold: 0,
9981 1 : initial_split_shards: 0,
9982 1 : }),
9983 : None,
9984 : );
9985 1 : assert_eq!(
9986 1 : Service::compute_split_shards(ShardSplitInputs {
9987 1 : shard_count: ShardCount(2),
9988 1 : max_logical_size: 129,
9989 1 : split_threshold: 64,
9990 1 : max_split_shards: 16,
9991 1 : initial_split_threshold: 0,
9992 1 : initial_split_shards: 0,
9993 1 : }),
9994 : Some(ShardCount(4)),
9995 : );
9996 :
9997 : // Size-based split: clamped to max_split_shards.
9998 1 : assert_eq!(
9999 1 : Service::compute_split_shards(ShardSplitInputs {
10000 1 : shard_count: ShardCount(2),
10001 1 : max_logical_size: 10000,
10002 1 : split_threshold: 64,
10003 1 : max_split_shards: 16,
10004 1 : initial_split_threshold: 0,
10005 1 : initial_split_shards: 0,
10006 1 : }),
10007 : Some(ShardCount(16))
10008 : );
10009 :
10010 : // Size-based split: tenant already at or beyond max_split_shards is not split.
10011 1 : assert_eq!(
10012 1 : Service::compute_split_shards(ShardSplitInputs {
10013 1 : shard_count: ShardCount(16),
10014 1 : max_logical_size: 10000,
10015 1 : split_threshold: 64,
10016 1 : max_split_shards: 16,
10017 1 : initial_split_threshold: 0,
10018 1 : initial_split_shards: 0,
10019 1 : }),
10020 : None
10021 : );
10022 :
10023 1 : assert_eq!(
10024 1 : Service::compute_split_shards(ShardSplitInputs {
10025 1 : shard_count: ShardCount(32),
10026 1 : max_logical_size: 10000,
10027 1 : split_threshold: 64,
10028 1 : max_split_shards: 16,
10029 1 : initial_split_threshold: 0,
10030 1 : initial_split_shards: 0,
10031 1 : }),
10032 : None
10033 : );
10034 :
10035 : // Size-based split: a non-power-of-2 shard count is normalized to power-of-2 if it
10036 : // exceeds split_threshold (i.e. a 3-shard tenant splits into 8, not 6).
10037 1 : assert_eq!(
10038 1 : Service::compute_split_shards(ShardSplitInputs {
10039 1 : shard_count: ShardCount(3),
10040 1 : max_logical_size: 320,
10041 1 : split_threshold: 64,
10042 1 : max_split_shards: 16,
10043 1 : initial_split_threshold: 0,
10044 1 : initial_split_shards: 0,
10045 1 : }),
10046 : Some(ShardCount(8))
10047 : );
10048 :
10049 : // Size-based split: a non-power-of-2 shard count is not normalized to power-of-2 if the
10050 : // existing shards are below or at split_threshold, but splits into 4 if it exceeds it.
10051 1 : assert_eq!(
10052 1 : Service::compute_split_shards(ShardSplitInputs {
10053 1 : shard_count: ShardCount(3),
10054 1 : max_logical_size: 191,
10055 1 : split_threshold: 64,
10056 1 : max_split_shards: 16,
10057 1 : initial_split_threshold: 0,
10058 1 : initial_split_shards: 0,
10059 1 : }),
10060 : None
10061 : );
10062 1 : assert_eq!(
10063 1 : Service::compute_split_shards(ShardSplitInputs {
10064 1 : shard_count: ShardCount(3),
10065 1 : max_logical_size: 192,
10066 1 : split_threshold: 64,
10067 1 : max_split_shards: 16,
10068 1 : initial_split_threshold: 0,
10069 1 : initial_split_shards: 0,
10070 1 : }),
10071 : None
10072 : );
10073 1 : assert_eq!(
10074 1 : Service::compute_split_shards(ShardSplitInputs {
10075 1 : shard_count: ShardCount(3),
10076 1 : max_logical_size: 193,
10077 1 : split_threshold: 64,
10078 1 : max_split_shards: 16,
10079 1 : initial_split_threshold: 0,
10080 1 : initial_split_shards: 0,
10081 1 : }),
10082 : Some(ShardCount(4))
10083 : );
10084 :
10085 : // Initial split: tenant has a 10 GB timeline, split into 4 shards.
10086 1 : assert_eq!(
10087 1 : Service::compute_split_shards(ShardSplitInputs {
10088 1 : shard_count: ShardCount(1),
10089 1 : max_logical_size: 10,
10090 1 : split_threshold: 0,
10091 1 : max_split_shards: 16,
10092 1 : initial_split_threshold: 8,
10093 1 : initial_split_shards: 4,
10094 1 : }),
10095 : Some(ShardCount(4))
10096 : );
10097 :
10098 : // Initial split: 0 ShardCount is equivalent to 1.
10099 1 : assert_eq!(
10100 1 : Service::compute_split_shards(ShardSplitInputs {
10101 1 : shard_count: ShardCount(0),
10102 1 : max_logical_size: 10,
10103 1 : split_threshold: 0,
10104 1 : max_split_shards: 16,
10105 1 : initial_split_threshold: 8,
10106 1 : initial_split_shards: 4,
10107 1 : }),
10108 : Some(ShardCount(4))
10109 : );
10110 :
10111 : // Initial split: at or below threshold is noop.
10112 1 : assert_eq!(
10113 1 : Service::compute_split_shards(ShardSplitInputs {
10114 1 : shard_count: ShardCount(1),
10115 1 : max_logical_size: 7,
10116 1 : split_threshold: 0,
10117 1 : max_split_shards: 16,
10118 1 : initial_split_threshold: 8,
10119 1 : initial_split_shards: 4,
10120 1 : }),
10121 : None,
10122 : );
10123 1 : assert_eq!(
10124 1 : Service::compute_split_shards(ShardSplitInputs {
10125 1 : shard_count: ShardCount(1),
10126 1 : max_logical_size: 8,
10127 1 : split_threshold: 0,
10128 1 : max_split_shards: 16,
10129 1 : initial_split_threshold: 8,
10130 1 : initial_split_shards: 4,
10131 1 : }),
10132 : None,
10133 : );
10134 1 : assert_eq!(
10135 1 : Service::compute_split_shards(ShardSplitInputs {
10136 1 : shard_count: ShardCount(1),
10137 1 : max_logical_size: 9,
10138 1 : split_threshold: 0,
10139 1 : max_split_shards: 16,
10140 1 : initial_split_threshold: 8,
10141 1 : initial_split_shards: 4,
10142 1 : }),
10143 : Some(ShardCount(4))
10144 : );
10145 :
10146 : // Initial split: already sharded tenant is not affected, even if above threshold and below
10147 : // shard count.
10148 1 : assert_eq!(
10149 1 : Service::compute_split_shards(ShardSplitInputs {
10150 1 : shard_count: ShardCount(2),
10151 1 : max_logical_size: 20,
10152 1 : split_threshold: 0,
10153 1 : max_split_shards: 16,
10154 1 : initial_split_threshold: 8,
10155 1 : initial_split_shards: 4,
10156 1 : }),
10157 : None,
10158 : );
10159 :
10160 : // Initial split: clamped to max_shards.
10161 1 : assert_eq!(
10162 1 : Service::compute_split_shards(ShardSplitInputs {
10163 1 : shard_count: ShardCount(1),
10164 1 : max_logical_size: 10,
10165 1 : split_threshold: 0,
10166 1 : max_split_shards: 3,
10167 1 : initial_split_threshold: 8,
10168 1 : initial_split_shards: 4,
10169 1 : }),
10170 : Some(ShardCount(3)),
10171 : );
10172 :
10173 : // Initial+size split: tenant eligible for both will use the larger shard count.
10174 1 : assert_eq!(
10175 1 : Service::compute_split_shards(ShardSplitInputs {
10176 1 : shard_count: ShardCount(1),
10177 1 : max_logical_size: 10,
10178 1 : split_threshold: 64,
10179 1 : max_split_shards: 16,
10180 1 : initial_split_threshold: 8,
10181 1 : initial_split_shards: 4,
10182 1 : }),
10183 : Some(ShardCount(4)),
10184 : );
10185 1 : assert_eq!(
10186 1 : Service::compute_split_shards(ShardSplitInputs {
10187 1 : shard_count: ShardCount(1),
10188 1 : max_logical_size: 500,
10189 1 : split_threshold: 64,
10190 1 : max_split_shards: 16,
10191 1 : initial_split_threshold: 8,
10192 1 : initial_split_shards: 4,
10193 1 : }),
10194 : Some(ShardCount(8)),
10195 : );
10196 :
10197 : // Initial+size split: sharded tenant is only eligible for size-based split.
10198 1 : assert_eq!(
10199 1 : Service::compute_split_shards(ShardSplitInputs {
10200 1 : shard_count: ShardCount(2),
10201 1 : max_logical_size: 200,
10202 1 : split_threshold: 64,
10203 1 : max_split_shards: 16,
10204 1 : initial_split_threshold: 8,
10205 1 : initial_split_shards: 8,
10206 1 : }),
10207 : Some(ShardCount(4)),
10208 : );
10209 :
10210 : // Initial+size split: uses the larger shard count even with initial_split_threshold above
10211 : // split_threshold.
10212 1 : assert_eq!(
10213 1 : Service::compute_split_shards(ShardSplitInputs {
10214 1 : shard_count: ShardCount(1),
10215 1 : max_logical_size: 10,
10216 1 : split_threshold: 4,
10217 1 : max_split_shards: 16,
10218 1 : initial_split_threshold: 8,
10219 1 : initial_split_shards: 8,
10220 1 : }),
10221 : Some(ShardCount(8)),
10222 : );
10223 :
10224 : // Test backwards compatibility with production settings when initial/size-based splits were
10225 : // rolled out: a single split into 8 shards at 64 GB. Any already sharded tenants with <8
10226 : // shards will split according to split_threshold.
10227 1 : assert_eq!(
10228 1 : Service::compute_split_shards(ShardSplitInputs {
10229 1 : shard_count: ShardCount(1),
10230 1 : max_logical_size: 65,
10231 1 : split_threshold: 64,
10232 1 : max_split_shards: 8,
10233 1 : initial_split_threshold: 64,
10234 1 : initial_split_shards: 8,
10235 1 : }),
10236 : Some(ShardCount(8)),
10237 : );
10238 :
10239 1 : assert_eq!(
10240 1 : Service::compute_split_shards(ShardSplitInputs {
10241 1 : shard_count: ShardCount(1),
10242 1 : max_logical_size: 64,
10243 1 : split_threshold: 64,
10244 1 : max_split_shards: 8,
10245 1 : initial_split_threshold: 64,
10246 1 : initial_split_shards: 8,
10247 1 : }),
10248 : None,
10249 : );
10250 :
10251 1 : assert_eq!(
10252 1 : Service::compute_split_shards(ShardSplitInputs {
10253 1 : shard_count: ShardCount(2),
10254 1 : max_logical_size: 129,
10255 1 : split_threshold: 64,
10256 1 : max_split_shards: 8,
10257 1 : initial_split_threshold: 64,
10258 1 : initial_split_shards: 8,
10259 1 : }),
10260 : Some(ShardCount(4)),
10261 : );
10262 1 : }
10263 : }
|
