Line data Source code
1 : pub mod chaos_injector;
2 : mod context_iterator;
3 : pub(crate) mod safekeeper_reconciler;
4 : mod safekeeper_service;
5 :
6 : use std::borrow::Cow;
7 : use std::cmp::Ordering;
8 : use std::collections::{BTreeMap, HashMap, HashSet};
9 : use std::error::Error;
10 : use std::num::NonZeroU32;
11 : use std::ops::{Deref, DerefMut};
12 : use std::path::PathBuf;
13 : use std::str::FromStr;
14 : use std::sync::{Arc, OnceLock};
15 : use std::time::{Duration, Instant, SystemTime};
16 :
17 : use anyhow::Context;
18 : use context_iterator::TenantShardContextIterator;
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::controller_api::{
29 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
30 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
31 : ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
32 : TenantCreateRequest, TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse,
33 : TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
34 : TenantShardMigrateRequest, TenantShardMigrateResponse,
35 : };
36 : use pageserver_api::models::{
37 : self, DetachBehavior, LocationConfig, LocationConfigListResponse, LocationConfigMode, LsnLease,
38 : PageserverUtilization, SecondaryProgress, ShardParameters, TenantConfig,
39 : TenantConfigPatchRequest, TenantConfigRequest, TenantLocationConfigRequest,
40 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
41 : TenantShardSplitResponse, TenantSorting, TenantTimeTravelRequest,
42 : TimelineArchivalConfigRequest, TimelineCreateRequest, TimelineCreateResponseStorcon,
43 : TimelineInfo, TimelineState, TopTenantShardItem, TopTenantShardsRequest,
44 : };
45 : use pageserver_api::shard::{
46 : DEFAULT_STRIPE_SIZE, ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
47 : };
48 : use pageserver_api::upcall_api::{
49 : PutTimelineImportStatusRequest, ReAttachRequest, ReAttachResponse, ReAttachResponseTenant,
50 : ValidateRequest, ValidateResponse, ValidateResponseTenant,
51 : };
52 : use pageserver_client::{BlockUnblock, mgmt_api};
53 : use reqwest::{Certificate, StatusCode};
54 : use safekeeper_api::models::SafekeeperUtilization;
55 : use safekeeper_reconciler::SafekeeperReconcilers;
56 : use tokio::sync::TryAcquireError;
57 : use tokio::sync::mpsc::error::TrySendError;
58 : use tokio_util::sync::CancellationToken;
59 : use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
60 : use utils::completion::Barrier;
61 : use utils::generation::Generation;
62 : use utils::id::{NodeId, TenantId, TimelineId};
63 : use utils::lsn::Lsn;
64 : use utils::sync::gate::{Gate, GateGuard};
65 : use utils::{failpoint_support, pausable_failpoint};
66 :
67 : use crate::background_node_operations::{
68 : Drain, Fill, MAX_RECONCILES_PER_OPERATION, Operation, OperationError, OperationHandler,
69 : };
70 : use crate::compute_hook::{self, ComputeHook, NotifyError};
71 : use crate::drain_utils::{self, TenantShardDrain, TenantShardIterator};
72 : use crate::heartbeater::{Heartbeater, PageserverState, SafekeeperState};
73 : use crate::id_lock_map::{
74 : IdLockMap, TracingExclusiveGuard, trace_exclusive_lock, trace_shared_lock,
75 : };
76 : use crate::leadership::Leadership;
77 : use crate::metrics;
78 : use crate::node::{AvailabilityTransition, Node};
79 : use crate::pageserver_client::PageserverClient;
80 : use crate::peer_client::GlobalObservedState;
81 : use crate::persistence::split_state::SplitState;
82 : use crate::persistence::{
83 : AbortShardSplitStatus, ControllerPersistence, DatabaseError, DatabaseResult,
84 : MetadataHealthPersistence, Persistence, ShardGenerationState, TenantFilter,
85 : TenantShardPersistence,
86 : };
87 : use crate::reconciler::{
88 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder, ReconcilerPriority,
89 : attached_location_conf,
90 : };
91 : use crate::safekeeper::Safekeeper;
92 : use crate::scheduler::{
93 : AttachedShardTag, MaySchedule, ScheduleContext, ScheduleError, ScheduleMode, Scheduler,
94 : };
95 : use crate::tenant_shard::{
96 : IntentState, MigrateAttachment, ObservedState, ObservedStateDelta, ObservedStateLocation,
97 : ReconcileNeeded, ReconcileResult, ReconcileWaitError, ReconcilerStatus, ReconcilerWaiter,
98 : ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
99 : };
100 : use crate::timeline_import::{
101 : ShardImportStatuses, TimelineImport, TimelineImportState, UpcallClient,
102 : };
103 :
104 : const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
105 :
106 : // For operations that should be quick, like attaching a new tenant
107 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
108 :
109 : // For operations that might be slow, like migrating a tenant with
110 : // some data in it.
111 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
112 :
113 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
114 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
115 : // input generation from future requests as authoritative.
116 : const INITIAL_GENERATION: Generation = Generation::new(0);
117 :
118 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
119 : /// up on unresponsive pageservers and proceed.
120 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
121 :
122 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
123 : /// This must be long enough to cover node restarts as well as normal operations: in future
124 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
125 :
126 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
127 : /// offline.
128 : ///
129 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
130 : /// handling of the re-attach response may take a long time and blocks heartbeats from
131 : /// being handled on the pageserver side.
132 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
133 :
134 : /// How often to send heartbeats to registered nodes?
135 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
136 :
137 : /// How long is too long for a reconciliation?
138 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
139 :
140 : #[derive(Clone, strum_macros::Display)]
141 : enum TenantOperations {
142 : Create,
143 : LocationConfig,
144 : ConfigSet,
145 : ConfigPatch,
146 : TimeTravelRemoteStorage,
147 : Delete,
148 : UpdatePolicy,
149 : ShardSplit,
150 : SecondaryDownload,
151 : TimelineCreate,
152 : TimelineDelete,
153 : AttachHook,
154 : TimelineArchivalConfig,
155 : TimelineDetachAncestor,
156 : TimelineGcBlockUnblock,
157 : DropDetached,
158 : DownloadHeatmapLayers,
159 : TimelineLsnLease,
160 : }
161 :
162 : #[derive(Clone, strum_macros::Display)]
163 : enum NodeOperations {
164 : Register,
165 : Configure,
166 : Delete,
167 : }
168 :
169 : /// The leadership status for the storage controller process.
170 : /// Allowed transitions are:
171 : /// 1. Leader -> SteppedDown
172 : /// 2. Candidate -> Leader
173 : #[derive(
174 : Eq,
175 : PartialEq,
176 : Copy,
177 : Clone,
178 : strum_macros::Display,
179 0 : strum_macros::EnumIter,
180 : measured::FixedCardinalityLabel,
181 : )]
182 : #[strum(serialize_all = "snake_case")]
183 : pub(crate) enum LeadershipStatus {
184 : /// This is the steady state where the storage controller can produce
185 : /// side effects in the cluster.
186 : Leader,
187 : /// We've been notified to step down by another candidate. No reconciliations
188 : /// take place in this state.
189 : SteppedDown,
190 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
191 : #[allow(unused)]
192 : Candidate,
193 : }
194 :
195 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
196 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
197 : pub const SAFEKEEPER_RECONCILER_CONCURRENCY_DEFAULT: usize = 32;
198 :
199 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
200 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
201 : // than they're being pushed onto the queue.
202 : const MAX_DELAYED_RECONCILES: usize = 10000;
203 :
204 : // Top level state available to all HTTP handlers
205 : struct ServiceState {
206 : leadership_status: LeadershipStatus,
207 :
208 : tenants: BTreeMap<TenantShardId, TenantShard>,
209 :
210 : nodes: Arc<HashMap<NodeId, Node>>,
211 :
212 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
213 :
214 : safekeeper_reconcilers: SafekeeperReconcilers,
215 :
216 : scheduler: Scheduler,
217 :
218 : /// Ongoing background operation on the cluster if any is running.
219 : /// Note that only one such operation may run at any given time,
220 : /// hence the type choice.
221 : ongoing_operation: Option<OperationHandler>,
222 :
223 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
224 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
225 : }
226 :
227 : /// Transform an error from a pageserver into an error to return to callers of a storage
228 : /// controller API.
229 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
230 0 : match e {
231 0 : mgmt_api::Error::SendRequest(e) => {
232 0 : // Presume errors sending requests are connectivity/availability issues
233 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
234 : }
235 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
236 0 : // Presume errors receiving body are connectivity/availability issues
237 0 : ApiError::ResourceUnavailable(
238 0 : format!("{node} error receiving error body: {str}").into(),
239 0 : )
240 : }
241 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
242 0 : // Return 500 for decoding errors.
243 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
244 : }
245 0 : mgmt_api::Error::ReceiveBody(err) => {
246 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
247 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
248 0 : ApiError::ResourceUnavailable(
249 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
250 0 : )
251 : }
252 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
253 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
254 : }
255 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
256 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
257 : }
258 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
259 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
260 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
261 : // internal server errors, showing that something is wrong with the pageserver or
262 : // storage controller's auth configuration.
263 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
264 : }
265 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
266 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
267 0 : // turn ask our clients to wait + retry
268 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
269 : }
270 0 : mgmt_api::Error::ApiError(status, msg) => {
271 0 : // Presume general case of pageserver API errors is that we tried to do something
272 0 : // that can't be done right now.
273 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
274 : }
275 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
276 0 : mgmt_api::Error::Timeout(e) => ApiError::Timeout(e.into()),
277 : }
278 0 : }
279 :
280 : impl ServiceState {
281 0 : fn new(
282 0 : nodes: HashMap<NodeId, Node>,
283 0 : safekeepers: HashMap<NodeId, Safekeeper>,
284 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
285 0 : scheduler: Scheduler,
286 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
287 0 : initial_leadership_status: LeadershipStatus,
288 0 : reconcilers_cancel: CancellationToken,
289 0 : ) -> Self {
290 0 : metrics::update_leadership_status(initial_leadership_status);
291 0 :
292 0 : Self {
293 0 : leadership_status: initial_leadership_status,
294 0 : tenants,
295 0 : nodes: Arc::new(nodes),
296 0 : safekeepers: Arc::new(safekeepers),
297 0 : safekeeper_reconcilers: SafekeeperReconcilers::new(reconcilers_cancel),
298 0 : scheduler,
299 0 : ongoing_operation: None,
300 0 : delayed_reconcile_rx,
301 0 : }
302 0 : }
303 :
304 0 : fn parts_mut(
305 0 : &mut self,
306 0 : ) -> (
307 0 : &mut Arc<HashMap<NodeId, Node>>,
308 0 : &mut BTreeMap<TenantShardId, TenantShard>,
309 0 : &mut Scheduler,
310 0 : ) {
311 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
312 0 : }
313 :
314 : #[allow(clippy::type_complexity)]
315 0 : fn parts_mut_sk(
316 0 : &mut self,
317 0 : ) -> (
318 0 : &mut Arc<HashMap<NodeId, Node>>,
319 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
320 0 : &mut BTreeMap<TenantShardId, TenantShard>,
321 0 : &mut Scheduler,
322 0 : ) {
323 0 : (
324 0 : &mut self.nodes,
325 0 : &mut self.safekeepers,
326 0 : &mut self.tenants,
327 0 : &mut self.scheduler,
328 0 : )
329 0 : }
330 :
331 0 : fn get_leadership_status(&self) -> LeadershipStatus {
332 0 : self.leadership_status
333 0 : }
334 :
335 0 : fn step_down(&mut self) {
336 0 : self.leadership_status = LeadershipStatus::SteppedDown;
337 0 : metrics::update_leadership_status(self.leadership_status);
338 0 : }
339 :
340 0 : fn become_leader(&mut self) {
341 0 : self.leadership_status = LeadershipStatus::Leader;
342 0 : metrics::update_leadership_status(self.leadership_status);
343 0 : }
344 : }
345 :
346 : #[derive(Clone)]
347 : pub struct Config {
348 : // All pageservers managed by one instance of this service must have
349 : // the same public key. This JWT token will be used to authenticate
350 : // this service to the pageservers it manages.
351 : pub pageserver_jwt_token: Option<String>,
352 :
353 : // All safekeepers managed by one instance of this service must have
354 : // the same public key. This JWT token will be used to authenticate
355 : // this service to the safekeepers it manages.
356 : pub safekeeper_jwt_token: Option<String>,
357 :
358 : // This JWT token will be used to authenticate this service to the control plane.
359 : pub control_plane_jwt_token: Option<String>,
360 :
361 : // This JWT token will be used to authenticate with other storage controller instances
362 : pub peer_jwt_token: Option<String>,
363 :
364 : /// Prefix for storage API endpoints of the control plane. We use this prefix to compute
365 : /// URLs that we use to send pageserver and safekeeper attachment locations.
366 : /// If this is None, the compute hook will assume it is running in a test environment
367 : /// and try to invoke neon_local instead.
368 : pub control_plane_url: Option<String>,
369 :
370 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
371 : /// considered active. Once the grace period elapses, the next heartbeat failure will
372 : /// mark the pagseserver offline.
373 : pub max_offline_interval: Duration,
374 :
375 : /// Extended grace period within which pageserver may not respond to heartbeats.
376 : /// This extended grace period kicks in after the node has been drained for restart
377 : /// and/or upon handling the re-attach request from a node.
378 : pub max_warming_up_interval: Duration,
379 :
380 : /// How many normal-priority Reconcilers may be spawned concurrently
381 : pub reconciler_concurrency: usize,
382 :
383 : /// How many high-priority Reconcilers may be spawned concurrently
384 : pub priority_reconciler_concurrency: usize,
385 :
386 : /// How many safekeeper reconciles may happen concurrently (per safekeeper)
387 : pub safekeeper_reconciler_concurrency: usize,
388 :
389 : /// How many API requests per second to allow per tenant, across all
390 : /// tenant-scoped API endpoints. Further API requests queue until ready.
391 : pub tenant_rate_limit: NonZeroU32,
392 :
393 : /// If a tenant shard's largest timeline (max_logical_size) exceeds this value, all tenant
394 : /// shards will be split in 2 until they fall below split_threshold (up to max_split_shards).
395 : ///
396 : /// This will greedily split into as many shards as necessary to fall below split_threshold, as
397 : /// powers of 2: if a tenant shard is 7 times larger than split_threshold, it will split into 8
398 : /// immediately, rather than first 2 then 4 then 8.
399 : ///
400 : /// None or 0 disables auto-splitting.
401 : ///
402 : /// TODO: consider using total logical size of all timelines instead.
403 : pub split_threshold: Option<u64>,
404 :
405 : /// The maximum number of shards a tenant can be split into during autosplits. Does not affect
406 : /// manual split requests. 0 or 1 disables autosplits, as we already have 1 shard.
407 : pub max_split_shards: u8,
408 :
409 : /// The size at which an unsharded tenant should initially split. Ingestion is significantly
410 : /// faster with multiple shards, so eagerly splitting below split_threshold will typically speed
411 : /// up initial ingestion of large tenants.
412 : ///
413 : /// This should be below split_threshold, but it is not required. If both split_threshold and
414 : /// initial_split_threshold qualify, the largest number of target shards will be used.
415 : ///
416 : /// Does not apply to already sharded tenants: changing initial_split_threshold or
417 : /// initial_split_shards is not retroactive for already-sharded tenants.
418 : ///
419 : /// None or 0 disables initial splits.
420 : pub initial_split_threshold: Option<u64>,
421 :
422 : /// The number of shards to split into when reaching initial_split_threshold. Will
423 : /// be clamped to max_split_shards.
424 : ///
425 : /// 0 or 1 disables initial splits. Has no effect if initial_split_threshold is disabled.
426 : pub initial_split_shards: u8,
427 :
428 : // TODO: make this cfg(feature = "testing")
429 : pub neon_local_repo_dir: Option<PathBuf>,
430 :
431 : // Maximum acceptable download lag for the secondary location
432 : // while draining a node. If the secondary location is lagging
433 : // by more than the configured amount, then the secondary is not
434 : // upgraded to primary.
435 : pub max_secondary_lag_bytes: Option<u64>,
436 :
437 : pub heartbeat_interval: Duration,
438 :
439 : pub address_for_peers: Option<Uri>,
440 :
441 : pub start_as_candidate: bool,
442 :
443 : pub long_reconcile_threshold: Duration,
444 :
445 : pub use_https_pageserver_api: bool,
446 :
447 : pub use_https_safekeeper_api: bool,
448 :
449 : pub ssl_ca_certs: Vec<Certificate>,
450 :
451 : pub timelines_onto_safekeepers: bool,
452 :
453 : pub use_local_compute_notifications: bool,
454 : }
455 :
456 : impl From<DatabaseError> for ApiError {
457 0 : fn from(err: DatabaseError) -> ApiError {
458 0 : match err {
459 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
460 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
461 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
462 0 : ApiError::ShuttingDown
463 : }
464 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
465 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
466 : }
467 : }
468 0 : }
469 : }
470 :
471 : enum InitialShardScheduleOutcome {
472 : Scheduled(TenantCreateResponseShard),
473 : NotScheduled,
474 : ShardScheduleError(ScheduleError),
475 : }
476 :
477 : pub struct Service {
478 : inner: Arc<std::sync::RwLock<ServiceState>>,
479 : config: Config,
480 : persistence: Arc<Persistence>,
481 : compute_hook: Arc<ComputeHook>,
482 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
483 :
484 : heartbeater_ps: Heartbeater<Node, PageserverState>,
485 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
486 :
487 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
488 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
489 :
490 : // Locking on a tenant granularity (covers all shards in the tenant):
491 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
492 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
493 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
494 :
495 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
496 : // that transition it to/from Active.
497 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
498 :
499 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
500 : // and reconciliation on startup.
501 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
502 :
503 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
504 : // a human user might be waiting for.
505 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
506 :
507 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
508 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
509 : ///
510 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
511 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
512 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
513 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
514 :
515 : // Process shutdown will fire this token
516 : cancel: CancellationToken,
517 :
518 : // Child token of [`Service::cancel`] used by reconcilers
519 : reconcilers_cancel: CancellationToken,
520 :
521 : // Background tasks will hold this gate
522 : gate: Gate,
523 :
524 : // Reconcilers background tasks will hold this gate
525 : reconcilers_gate: Gate,
526 :
527 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
528 : /// passes, it isn't safe to do any actions that mutate tenants.
529 : pub(crate) startup_complete: Barrier,
530 :
531 : /// HTTP client with proper CA certs.
532 : http_client: reqwest::Client,
533 :
534 : /// Handle for the step down background task if one was ever requested
535 : step_down_barrier: OnceLock<tokio::sync::watch::Receiver<Option<GlobalObservedState>>>,
536 : }
537 :
538 : impl From<ReconcileWaitError> for ApiError {
539 0 : fn from(value: ReconcileWaitError) -> Self {
540 0 : match value {
541 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
542 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
543 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
544 : }
545 0 : }
546 : }
547 :
548 : impl From<OperationError> for ApiError {
549 0 : fn from(value: OperationError) -> Self {
550 0 : match value {
551 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
552 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
553 : }
554 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
555 : }
556 0 : }
557 : }
558 :
559 : #[allow(clippy::large_enum_variant)]
560 : enum TenantCreateOrUpdate {
561 : Create(TenantCreateRequest),
562 : Update(Vec<ShardUpdate>),
563 : }
564 :
565 : struct ShardSplitParams {
566 : old_shard_count: ShardCount,
567 : new_shard_count: ShardCount,
568 : new_stripe_size: Option<ShardStripeSize>,
569 : targets: Vec<ShardSplitTarget>,
570 : policy: PlacementPolicy,
571 : config: TenantConfig,
572 : shard_ident: ShardIdentity,
573 : preferred_az_id: Option<AvailabilityZone>,
574 : }
575 :
576 : // When preparing for a shard split, we may either choose to proceed with the split,
577 : // or find that the work is already done and return NoOp.
578 : enum ShardSplitAction {
579 : Split(Box<ShardSplitParams>),
580 : NoOp(TenantShardSplitResponse),
581 : }
582 :
583 : // A parent shard which will be split
584 : struct ShardSplitTarget {
585 : parent_id: TenantShardId,
586 : node: Node,
587 : child_ids: Vec<TenantShardId>,
588 : }
589 :
590 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
591 : /// might not be available. We therefore use a queue of abort operations processed in the background.
592 : struct TenantShardSplitAbort {
593 : tenant_id: TenantId,
594 : /// The target values from the request that failed
595 : new_shard_count: ShardCount,
596 : new_stripe_size: Option<ShardStripeSize>,
597 : /// Until this abort op is complete, no other operations may be done on the tenant
598 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
599 : /// The reconciler gate for the duration of the split operation, and any included abort.
600 : _gate: GateGuard,
601 : }
602 :
603 : #[derive(thiserror::Error, Debug)]
604 : enum TenantShardSplitAbortError {
605 : #[error(transparent)]
606 : Database(#[from] DatabaseError),
607 : #[error(transparent)]
608 : Remote(#[from] mgmt_api::Error),
609 : #[error("Unavailable")]
610 : Unavailable,
611 : }
612 :
613 : /// Inputs for computing a target shard count for a tenant.
614 : struct ShardSplitInputs {
615 : /// Current shard count.
616 : shard_count: ShardCount,
617 : /// Total size of largest timeline summed across all shards.
618 : max_logical_size: u64,
619 : /// Size-based split threshold. Zero if size-based splits are disabled.
620 : split_threshold: u64,
621 : /// Upper bound on target shards. 0 or 1 disables splits.
622 : max_split_shards: u8,
623 : /// Initial split threshold. Zero if initial splits are disabled.
624 : initial_split_threshold: u64,
625 : /// Number of shards for initial splits. 0 or 1 disables initial splits.
626 : initial_split_shards: u8,
627 : }
628 :
629 : struct ShardUpdate {
630 : tenant_shard_id: TenantShardId,
631 : placement_policy: PlacementPolicy,
632 : tenant_config: TenantConfig,
633 :
634 : /// If this is None, generation is not updated.
635 : generation: Option<Generation>,
636 :
637 : /// If this is None, scheduling policy is not updated.
638 : scheduling_policy: Option<ShardSchedulingPolicy>,
639 : }
640 :
641 : enum StopReconciliationsReason {
642 : ShuttingDown,
643 : SteppingDown,
644 : }
645 :
646 : impl std::fmt::Display for StopReconciliationsReason {
647 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
648 0 : let s = match self {
649 0 : Self::ShuttingDown => "Shutting down",
650 0 : Self::SteppingDown => "Stepping down",
651 : };
652 0 : write!(writer, "{}", s)
653 0 : }
654 : }
655 :
656 : pub(crate) enum ReconcileResultRequest {
657 : ReconcileResult(ReconcileResult),
658 : Stop,
659 : }
660 :
661 : #[derive(Clone)]
662 : struct MutationLocation {
663 : node: Node,
664 : generation: Generation,
665 : }
666 :
667 : #[derive(Clone)]
668 : struct ShardMutationLocations {
669 : latest: MutationLocation,
670 : other: Vec<MutationLocation>,
671 : }
672 :
673 : #[derive(Default, Clone)]
674 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
675 :
676 : impl Service {
677 0 : pub fn get_config(&self) -> &Config {
678 0 : &self.config
679 0 : }
680 :
681 0 : pub fn get_http_client(&self) -> &reqwest::Client {
682 0 : &self.http_client
683 0 : }
684 :
685 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
686 : /// view of the world, and determine which pageservers are responsive.
687 : #[instrument(skip_all)]
688 : async fn startup_reconcile(
689 : self: &Arc<Service>,
690 : current_leader: Option<ControllerPersistence>,
691 : leader_step_down_state: Option<GlobalObservedState>,
692 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
693 : Result<(), (TenantShardId, NotifyError)>,
694 : >,
695 : ) {
696 : // Startup reconciliation does I/O to other services: whether they
697 : // are responsive or not, we should aim to finish within our deadline, because:
698 : // - If we don't, a k8s readiness hook watching /ready will kill us.
699 : // - While we're waiting for startup reconciliation, we are not fully
700 : // available for end user operations like creating/deleting tenants and timelines.
701 : //
702 : // We set multiple deadlines to break up the time available between the phases of work: this is
703 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
704 : let start_at = Instant::now();
705 : let node_scan_deadline = start_at
706 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
707 : .expect("Reconcile timeout is a modest constant");
708 :
709 : let observed = if let Some(state) = leader_step_down_state {
710 : tracing::info!(
711 : "Using observed state received from leader at {}",
712 : current_leader.as_ref().unwrap().address
713 : );
714 :
715 : state
716 : } else {
717 : self.build_global_observed_state(node_scan_deadline).await
718 : };
719 :
720 : // Accumulate a list of any tenant locations that ought to be detached
721 : let mut cleanup = Vec::new();
722 :
723 : // Send initial heartbeat requests to all nodes loaded from the database
724 : let all_nodes = {
725 : let locked = self.inner.read().unwrap();
726 : locked.nodes.clone()
727 : };
728 : let (mut nodes_online, mut sks_online) =
729 : self.initial_heartbeat_round(all_nodes.keys()).await;
730 :
731 : // List of tenants for which we will attempt to notify compute of their location at startup
732 : let mut compute_notifications = Vec::new();
733 :
734 : // Populate intent and observed states for all tenants, based on reported state on pageservers
735 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
736 : let shard_count = {
737 : let mut locked = self.inner.write().unwrap();
738 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
739 :
740 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
741 : let mut new_nodes = (**nodes).clone();
742 : for (node_id, node) in new_nodes.iter_mut() {
743 : if let Some(utilization) = nodes_online.remove(node_id) {
744 : node.set_availability(NodeAvailability::Active(utilization));
745 : scheduler.node_upsert(node);
746 : }
747 : }
748 : *nodes = Arc::new(new_nodes);
749 :
750 : let mut new_sks = (**safekeepers).clone();
751 : for (node_id, node) in new_sks.iter_mut() {
752 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
753 : node.set_availability(SafekeeperState::Available {
754 : utilization,
755 : last_seen_at,
756 : });
757 : }
758 : }
759 : *safekeepers = Arc::new(new_sks);
760 :
761 : for (tenant_shard_id, observed_state) in observed.0 {
762 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
763 : for node_id in observed_state.locations.keys() {
764 : cleanup.push((tenant_shard_id, *node_id));
765 : }
766 :
767 : continue;
768 : };
769 :
770 : tenant_shard.observed = observed_state;
771 : }
772 :
773 : // Populate each tenant's intent state
774 : let mut schedule_context = ScheduleContext::default();
775 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
776 : if tenant_shard_id.shard_number == ShardNumber(0) {
777 : // Reset scheduling context each time we advance to the next Tenant
778 : schedule_context = ScheduleContext::default();
779 : }
780 :
781 : tenant_shard.intent_from_observed(scheduler);
782 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
783 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
784 : // not enough pageservers are available. The tenant may well still be available
785 : // to clients.
786 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
787 : } else {
788 : // If we're both intending and observed to be attached at a particular node, we will
789 : // emit a compute notification for this. In the case where our observed state does not
790 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
791 : if let Some(attached_at) = tenant_shard.stably_attached() {
792 : compute_notifications.push(compute_hook::ShardUpdate {
793 : tenant_shard_id: *tenant_shard_id,
794 : node_id: attached_at,
795 : stripe_size: tenant_shard.shard.stripe_size,
796 : preferred_az: tenant_shard
797 : .preferred_az()
798 0 : .map(|az| Cow::Owned(az.clone())),
799 : });
800 : }
801 : }
802 : }
803 :
804 : tenants.len()
805 : };
806 :
807 : // Before making any obeservable changes to the cluster, persist self
808 : // as leader in database and memory.
809 : let leadership = Leadership::new(
810 : self.persistence.clone(),
811 : self.config.clone(),
812 : self.cancel.child_token(),
813 : );
814 :
815 : if let Err(e) = leadership.become_leader(current_leader).await {
816 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
817 : std::process::exit(1);
818 : }
819 :
820 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
821 : let sk_schedule_requests =
822 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
823 : Ok(v) => v,
824 : Err(e) => {
825 : tracing::warn!(
826 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
827 : );
828 : // std::process::exit(1);
829 : Vec::new()
830 : }
831 : };
832 :
833 : {
834 : let mut locked = self.inner.write().unwrap();
835 : locked.become_leader();
836 :
837 : for (sk_id, _sk) in locked.safekeepers.clone().iter() {
838 : locked.safekeeper_reconcilers.start_reconciler(*sk_id, self);
839 : }
840 :
841 : locked
842 : .safekeeper_reconcilers
843 : .schedule_request_vec(sk_schedule_requests);
844 : }
845 :
846 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
847 : // generation_pageserver in the database.
848 :
849 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
850 : // will emit compute hook notifications when they reconcile.
851 : //
852 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
853 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
854 : // calls will be correctly ordered wrt these.
855 : //
856 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
857 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
858 : // unit and start doing I/O.
859 : tracing::info!(
860 : "Sending {} compute notifications",
861 : compute_notifications.len()
862 : );
863 : self.compute_hook.notify_background(
864 : compute_notifications,
865 : bg_compute_notify_result_tx.clone(),
866 : &self.cancel,
867 : );
868 :
869 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
870 : // which require it: under normal circumstances this should only include tenants that were in some
871 : // transient state before we restarted, or any tenants whose compute hooks failed above.
872 : tracing::info!("Checking for shards in need of reconciliation...");
873 : let reconcile_tasks = self.reconcile_all();
874 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
875 : // normal operations may proceed.
876 :
877 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
878 : // background because it does not need to complete in order to proceed with other work.
879 : if !cleanup.is_empty() {
880 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
881 : tokio::task::spawn({
882 : let cleanup_self = self.clone();
883 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
884 : });
885 : }
886 :
887 : // Reconcile the timeline imports:
888 : // 1. Mark each tenant shard of tenants with an importing timeline as importing.
889 : // 2. Finalize the completed imports in the background. This handles the case where
890 : // the previous storage controller instance shut down whilst finalizing imports.
891 : let imports = self.persistence.list_timeline_imports().await;
892 : match imports {
893 : Ok(mut imports) => {
894 : {
895 : let mut locked = self.inner.write().unwrap();
896 : for import in &imports {
897 : locked
898 : .tenants
899 : .range_mut(TenantShardId::tenant_range(import.tenant_id))
900 0 : .for_each(|(_id, shard)| {
901 0 : shard.importing = TimelineImportState::Importing
902 0 : });
903 : }
904 : }
905 :
906 0 : imports.retain(|import| import.is_complete());
907 : tokio::task::spawn({
908 : let finalize_imports_self = self.clone();
909 0 : async move {
910 0 : finalize_imports_self
911 0 : .finalize_timeline_imports(imports)
912 0 : .await
913 0 : }
914 : });
915 : }
916 : Err(err) => {
917 : tracing::error!("Could not retrieve completed imports from database: {err}");
918 : }
919 : }
920 :
921 : tracing::info!(
922 : "Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)"
923 : );
924 : }
925 :
926 0 : async fn initial_heartbeat_round<'a>(
927 0 : &self,
928 0 : node_ids: impl Iterator<Item = &'a NodeId>,
929 0 : ) -> (
930 0 : HashMap<NodeId, PageserverUtilization>,
931 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
932 0 : ) {
933 0 : assert!(!self.startup_complete.is_ready());
934 :
935 0 : let all_nodes = {
936 0 : let locked = self.inner.read().unwrap();
937 0 : locked.nodes.clone()
938 0 : };
939 0 :
940 0 : let mut nodes_to_heartbeat = HashMap::new();
941 0 : for node_id in node_ids {
942 0 : match all_nodes.get(node_id) {
943 0 : Some(node) => {
944 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
945 0 : }
946 : None => {
947 0 : tracing::warn!("Node {node_id} was removed during start-up");
948 : }
949 : }
950 : }
951 :
952 0 : let all_sks = {
953 0 : let locked = self.inner.read().unwrap();
954 0 : locked.safekeepers.clone()
955 0 : };
956 0 :
957 0 : tracing::info!("Sending initial heartbeats...");
958 0 : let (res_ps, res_sk) = tokio::join!(
959 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
960 0 : self.heartbeater_sk.heartbeat(all_sks)
961 0 : );
962 :
963 0 : let mut online_nodes = HashMap::new();
964 0 : if let Ok(deltas) = res_ps {
965 0 : for (node_id, status) in deltas.0 {
966 0 : match status {
967 0 : PageserverState::Available { utilization, .. } => {
968 0 : online_nodes.insert(node_id, utilization);
969 0 : }
970 0 : PageserverState::Offline => {}
971 : PageserverState::WarmingUp { .. } => {
972 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
973 : }
974 : }
975 : }
976 0 : }
977 :
978 0 : let mut online_sks = HashMap::new();
979 0 : if let Ok(deltas) = res_sk {
980 0 : for (node_id, status) in deltas.0 {
981 0 : match status {
982 : SafekeeperState::Available {
983 0 : utilization,
984 0 : last_seen_at,
985 0 : } => {
986 0 : online_sks.insert(node_id, (utilization, last_seen_at));
987 0 : }
988 0 : SafekeeperState::Offline => {}
989 : }
990 : }
991 0 : }
992 :
993 0 : (online_nodes, online_sks)
994 0 : }
995 :
996 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
997 : ///
998 : /// The result includes only nodes which responded within the deadline
999 0 : async fn scan_node_locations(
1000 0 : &self,
1001 0 : deadline: Instant,
1002 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
1003 0 : let nodes = {
1004 0 : let locked = self.inner.read().unwrap();
1005 0 : locked.nodes.clone()
1006 0 : };
1007 0 :
1008 0 : let mut node_results = HashMap::new();
1009 0 :
1010 0 : let mut node_list_futs = FuturesUnordered::new();
1011 0 :
1012 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
1013 0 : for node in nodes.values() {
1014 0 : node_list_futs.push({
1015 0 : async move {
1016 0 : tracing::info!("Scanning shards on node {node}...");
1017 0 : let timeout = Duration::from_secs(5);
1018 0 : let response = node
1019 0 : .with_client_retries(
1020 0 : |client| async move { client.list_location_config().await },
1021 0 : &self.http_client,
1022 0 : &self.config.pageserver_jwt_token,
1023 0 : 1,
1024 0 : 5,
1025 0 : timeout,
1026 0 : &self.cancel,
1027 0 : )
1028 0 : .await;
1029 0 : (node.get_id(), response)
1030 0 : }
1031 0 : });
1032 0 : }
1033 :
1034 : loop {
1035 0 : let (node_id, result) = tokio::select! {
1036 0 : next = node_list_futs.next() => {
1037 0 : match next {
1038 0 : Some(result) => result,
1039 : None =>{
1040 : // We got results for all our nodes
1041 0 : break;
1042 : }
1043 :
1044 : }
1045 : },
1046 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
1047 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
1048 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
1049 0 : break;
1050 : }
1051 : };
1052 :
1053 0 : let Some(list_response) = result else {
1054 0 : tracing::info!("Shutdown during startup_reconcile");
1055 0 : break;
1056 : };
1057 :
1058 0 : match list_response {
1059 0 : Err(e) => {
1060 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
1061 : }
1062 0 : Ok(listing) => {
1063 0 : node_results.insert(node_id, listing);
1064 0 : }
1065 : }
1066 : }
1067 :
1068 0 : node_results
1069 0 : }
1070 :
1071 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
1072 0 : let node_listings = self.scan_node_locations(deadline).await;
1073 0 : let mut observed = GlobalObservedState::default();
1074 :
1075 0 : for (node_id, location_confs) in node_listings {
1076 0 : tracing::info!(
1077 0 : "Received {} shard statuses from pageserver {}",
1078 0 : location_confs.tenant_shards.len(),
1079 : node_id
1080 : );
1081 :
1082 0 : for (tid, location_conf) in location_confs.tenant_shards {
1083 0 : let entry = observed.0.entry(tid).or_default();
1084 0 : entry.locations.insert(
1085 0 : node_id,
1086 0 : ObservedStateLocation {
1087 0 : conf: location_conf,
1088 0 : },
1089 0 : );
1090 0 : }
1091 : }
1092 :
1093 0 : observed
1094 0 : }
1095 :
1096 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
1097 : ///
1098 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1099 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1100 : /// other task trying to attach it.
1101 : #[instrument(skip_all)]
1102 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1103 : let nodes = self.inner.read().unwrap().nodes.clone();
1104 :
1105 : for (tenant_shard_id, node_id) in cleanup {
1106 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1107 : let Some(node) = nodes.get(&node_id) else {
1108 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1109 : // a location to clean up on a node that has since been removed.
1110 : tracing::info!(
1111 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1112 : );
1113 : continue;
1114 : };
1115 :
1116 : if self.cancel.is_cancelled() {
1117 : break;
1118 : }
1119 :
1120 : let client = PageserverClient::new(
1121 : node.get_id(),
1122 : self.http_client.clone(),
1123 : node.base_url(),
1124 : self.config.pageserver_jwt_token.as_deref(),
1125 : );
1126 : match client
1127 : .location_config(
1128 : tenant_shard_id,
1129 : LocationConfig {
1130 : mode: LocationConfigMode::Detached,
1131 : generation: None,
1132 : secondary_conf: None,
1133 : shard_number: tenant_shard_id.shard_number.0,
1134 : shard_count: tenant_shard_id.shard_count.literal(),
1135 : shard_stripe_size: 0,
1136 : tenant_conf: models::TenantConfig::default(),
1137 : },
1138 : None,
1139 : false,
1140 : )
1141 : .await
1142 : {
1143 : Ok(()) => {
1144 : tracing::info!(
1145 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1146 : );
1147 : }
1148 : Err(e) => {
1149 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1150 : // break anything.
1151 : tracing::error!(
1152 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1153 : );
1154 : }
1155 : }
1156 : }
1157 : }
1158 :
1159 : /// Long running background task that periodically wakes up and looks for shards that need
1160 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1161 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1162 : /// for those retries.
1163 : #[instrument(skip_all)]
1164 : async fn background_reconcile(self: &Arc<Self>) {
1165 : self.startup_complete.clone().wait().await;
1166 :
1167 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1168 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1169 : while !self.reconcilers_cancel.is_cancelled() {
1170 : tokio::select! {
1171 : _ = interval.tick() => {
1172 : let reconciles_spawned = self.reconcile_all();
1173 : if reconciles_spawned == 0 {
1174 : // Run optimizer only when we didn't find any other work to do
1175 : self.optimize_all().await;
1176 : }
1177 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1178 : // must be responsive when new projects begin ingesting and reach the threshold.
1179 : self.autosplit_tenants().await;
1180 : }
1181 : _ = self.reconcilers_cancel.cancelled() => return
1182 : }
1183 : }
1184 : }
1185 : /// Heartbeat all storage nodes once in a while.
1186 : #[instrument(skip_all)]
1187 : async fn spawn_heartbeat_driver(&self) {
1188 : self.startup_complete.clone().wait().await;
1189 :
1190 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1191 : while !self.cancel.is_cancelled() {
1192 : tokio::select! {
1193 : _ = interval.tick() => { }
1194 : _ = self.cancel.cancelled() => return
1195 : };
1196 :
1197 : let nodes = {
1198 : let locked = self.inner.read().unwrap();
1199 : locked.nodes.clone()
1200 : };
1201 :
1202 : let safekeepers = {
1203 : let locked = self.inner.read().unwrap();
1204 : locked.safekeepers.clone()
1205 : };
1206 :
1207 : let (res_ps, res_sk) = tokio::join!(
1208 : self.heartbeater_ps.heartbeat(nodes),
1209 : self.heartbeater_sk.heartbeat(safekeepers)
1210 : );
1211 :
1212 : if let Ok(deltas) = res_ps {
1213 : let mut to_handle = Vec::default();
1214 :
1215 : for (node_id, state) in deltas.0 {
1216 : let new_availability = match state {
1217 : PageserverState::Available { utilization, .. } => {
1218 : NodeAvailability::Active(utilization)
1219 : }
1220 : PageserverState::WarmingUp { started_at } => {
1221 : NodeAvailability::WarmingUp(started_at)
1222 : }
1223 : PageserverState::Offline => {
1224 : // The node might have been placed in the WarmingUp state
1225 : // while the heartbeat round was on-going. Hence, filter out
1226 : // offline transitions for WarmingUp nodes that are still within
1227 : // their grace period.
1228 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1229 : .get_node(node_id)
1230 : .await
1231 : .as_ref()
1232 0 : .map(|n| n.get_availability())
1233 : {
1234 : let now = Instant::now();
1235 : if now - *started_at >= self.config.max_warming_up_interval {
1236 : NodeAvailability::Offline
1237 : } else {
1238 : NodeAvailability::WarmingUp(*started_at)
1239 : }
1240 : } else {
1241 : NodeAvailability::Offline
1242 : }
1243 : }
1244 : };
1245 :
1246 : let node_lock = trace_exclusive_lock(
1247 : &self.node_op_locks,
1248 : node_id,
1249 : NodeOperations::Configure,
1250 : )
1251 : .await;
1252 :
1253 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1254 :
1255 : // This is the code path for geniune availability transitions (i.e node
1256 : // goes unavailable and/or comes back online).
1257 : let res = self
1258 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1259 : .await;
1260 :
1261 : match res {
1262 : Ok(transition) => {
1263 : // Keep hold of the lock until the availability transitions
1264 : // have been handled in
1265 : // [`Service::handle_node_availability_transitions`] in order avoid
1266 : // racing with [`Service::external_node_configure`].
1267 : to_handle.push((node_id, node_lock, transition));
1268 : }
1269 : Err(ApiError::NotFound(_)) => {
1270 : // This should be rare, but legitimate since the heartbeats are done
1271 : // on a snapshot of the nodes.
1272 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1273 : }
1274 : Err(ApiError::ShuttingDown) => {
1275 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1276 : }
1277 : Err(err) => {
1278 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1279 : // becomes unavailable again, we may get an error here.
1280 : tracing::error!(
1281 : "Failed to update node state {} after heartbeat round: {}",
1282 : node_id,
1283 : err
1284 : );
1285 : }
1286 : }
1287 : }
1288 :
1289 : // We collected all the transitions above and now we handle them.
1290 : let res = self.handle_node_availability_transitions(to_handle).await;
1291 : if let Err(errs) = res {
1292 : for (node_id, err) in errs {
1293 : match err {
1294 : ApiError::NotFound(_) => {
1295 : // This should be rare, but legitimate since the heartbeats are done
1296 : // on a snapshot of the nodes.
1297 : tracing::info!(
1298 : "Node {} was not found after heartbeat round",
1299 : node_id
1300 : );
1301 : }
1302 : err => {
1303 : tracing::error!(
1304 : "Failed to handle availability transition for {} after heartbeat round: {}",
1305 : node_id,
1306 : err
1307 : );
1308 : }
1309 : }
1310 : }
1311 : }
1312 : }
1313 : if let Ok(deltas) = res_sk {
1314 : let mut locked = self.inner.write().unwrap();
1315 : let mut safekeepers = (*locked.safekeepers).clone();
1316 : for (id, state) in deltas.0 {
1317 : let Some(sk) = safekeepers.get_mut(&id) else {
1318 : tracing::info!(
1319 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1320 : );
1321 : continue;
1322 : };
1323 : sk.set_availability(state);
1324 : }
1325 : locked.safekeepers = Arc::new(safekeepers);
1326 : }
1327 : }
1328 : }
1329 :
1330 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1331 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1332 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1333 : /// will indicate that reconciliation is not needed.
1334 : #[instrument(skip_all, fields(
1335 : seq=%result.sequence,
1336 : tenant_id=%result.tenant_shard_id.tenant_id,
1337 : shard_id=%result.tenant_shard_id.shard_slug(),
1338 : ))]
1339 : fn process_result(&self, result: ReconcileResult) {
1340 : let mut locked = self.inner.write().unwrap();
1341 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1342 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1343 : // A reconciliation result might race with removing a tenant: drop results for
1344 : // tenants that aren't in our map.
1345 : return;
1346 : };
1347 :
1348 : // Usually generation should only be updated via this path, so the max() isn't
1349 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1350 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1351 :
1352 : // If the reconciler signals that it failed to notify compute, set this state on
1353 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1354 : tenant.pending_compute_notification = result.pending_compute_notification;
1355 :
1356 : // Let the TenantShard know it is idle.
1357 : tenant.reconcile_complete(result.sequence);
1358 :
1359 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1360 : // make to the tenant
1361 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1362 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1363 : // make to the tenant
1364 0 : let node = nodes.get(delta.node_id())?;
1365 :
1366 0 : if node.is_available() {
1367 0 : return Some(delta);
1368 0 : }
1369 0 :
1370 0 : // In case a node became unavailable concurrently with the reconcile, observed
1371 0 : // locations on it are now uncertain. By convention, set them to None in order
1372 0 : // for them to get refreshed when the node comes back online.
1373 0 : Some(ObservedStateDelta::Upsert(Box::new((
1374 0 : node.get_id(),
1375 0 : ObservedStateLocation { conf: None },
1376 0 : ))))
1377 0 : });
1378 :
1379 : match result.result {
1380 : Ok(()) => {
1381 : tenant.apply_observed_deltas(deltas);
1382 : tenant.waiter.advance(result.sequence);
1383 : }
1384 : Err(e) => {
1385 : match e {
1386 : ReconcileError::Cancel => {
1387 : tracing::info!("Reconciler was cancelled");
1388 : }
1389 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1390 : // This might be due to the reconciler getting cancelled, or it might
1391 : // be due to the `Node` being marked offline.
1392 : tracing::info!("Reconciler cancelled during pageserver API call");
1393 : }
1394 : _ => {
1395 : tracing::warn!("Reconcile error: {}", e);
1396 : }
1397 : }
1398 :
1399 : // Ordering: populate last_error before advancing error_seq,
1400 : // so that waiters will see the correct error after waiting.
1401 : tenant.set_last_error(result.sequence, e);
1402 :
1403 : // Skip deletions on reconcile failures
1404 : let upsert_deltas =
1405 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1406 : tenant.apply_observed_deltas(upsert_deltas);
1407 : }
1408 : }
1409 :
1410 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1411 : if tenant.policy == PlacementPolicy::Detached {
1412 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1413 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1414 : // duration of the request.
1415 : let guard = self.tenant_op_locks.try_exclusive(
1416 : tenant.tenant_shard_id.tenant_id,
1417 : TenantOperations::DropDetached,
1418 : );
1419 : if let Some(guard) = guard {
1420 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1421 : }
1422 : }
1423 :
1424 : // Maybe some other work can proceed now that this job finished.
1425 : //
1426 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1427 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1428 : if self.reconciler_concurrency.available_permits() > 0 {
1429 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1430 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1431 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1432 : shard.delayed_reconcile = false;
1433 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1434 : }
1435 :
1436 : if self.reconciler_concurrency.available_permits() == 0 {
1437 : break;
1438 : }
1439 : }
1440 : }
1441 : }
1442 :
1443 0 : async fn process_results(
1444 0 : &self,
1445 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1446 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1447 0 : Result<(), (TenantShardId, NotifyError)>,
1448 0 : >,
1449 0 : ) {
1450 : loop {
1451 : // Wait for the next result, or for cancellation
1452 0 : tokio::select! {
1453 0 : r = result_rx.recv() => {
1454 0 : match r {
1455 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1456 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1457 : }
1458 : }
1459 0 : _ = async{
1460 0 : match bg_compute_hook_result_rx.recv().await {
1461 0 : Some(result) => {
1462 0 : if let Err((tenant_shard_id, notify_error)) = result {
1463 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1464 0 : let mut locked = self.inner.write().unwrap();
1465 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1466 0 : shard.pending_compute_notification = true;
1467 0 : }
1468 :
1469 0 : }
1470 : },
1471 : None => {
1472 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1473 0 : self.cancel.cancelled().await;
1474 : }
1475 : }
1476 0 : } => {},
1477 0 : _ = self.cancel.cancelled() => {
1478 0 : break;
1479 : }
1480 : };
1481 : }
1482 0 : }
1483 :
1484 0 : async fn process_aborts(
1485 0 : &self,
1486 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1487 0 : ) {
1488 : loop {
1489 : // Wait for the next result, or for cancellation
1490 0 : let op = tokio::select! {
1491 0 : r = abort_rx.recv() => {
1492 0 : match r {
1493 0 : Some(op) => {op},
1494 0 : None => {break;}
1495 : }
1496 : }
1497 0 : _ = self.cancel.cancelled() => {
1498 0 : break;
1499 : }
1500 : };
1501 :
1502 : // Retry until shutdown: we must keep this request object alive until it is properly
1503 : // processed, as it holds a lock guard that prevents other operations trying to do things
1504 : // to the tenant while it is in a weird part-split state.
1505 0 : while !self.reconcilers_cancel.is_cancelled() {
1506 0 : match self.abort_tenant_shard_split(&op).await {
1507 0 : Ok(_) => break,
1508 0 : Err(e) => {
1509 0 : tracing::warn!(
1510 0 : "Failed to abort shard split on {}, will retry: {e}",
1511 : op.tenant_id
1512 : );
1513 :
1514 : // If a node is unavailable, we hope that it has been properly marked Offline
1515 : // when we retry, so that the abort op will succeed. If the abort op is failing
1516 : // for some other reason, we will keep retrying forever, or until a human notices
1517 : // and does something about it (either fixing a pageserver or restarting the controller).
1518 0 : tokio::time::timeout(
1519 0 : Duration::from_secs(5),
1520 0 : self.reconcilers_cancel.cancelled(),
1521 0 : )
1522 0 : .await
1523 0 : .ok();
1524 : }
1525 : }
1526 : }
1527 : }
1528 0 : }
1529 :
1530 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1531 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1532 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1533 0 :
1534 0 : let leadership_cancel = CancellationToken::new();
1535 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1536 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1537 :
1538 : // Apply the migrations **after** the current leader has stepped down
1539 : // (or we've given up waiting for it), but **before** reading from the
1540 : // database. The only exception is reading the current leader before
1541 : // migrating.
1542 0 : persistence.migration_run().await?;
1543 :
1544 0 : tracing::info!("Loading nodes from database...");
1545 0 : let nodes = persistence
1546 0 : .list_nodes()
1547 0 : .await?
1548 0 : .into_iter()
1549 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1550 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1551 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1552 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1553 0 : metrics::METRICS_REGISTRY
1554 0 : .metrics_group
1555 0 : .storage_controller_pageserver_nodes
1556 0 : .set(nodes.len() as i64);
1557 0 : metrics::METRICS_REGISTRY
1558 0 : .metrics_group
1559 0 : .storage_controller_https_pageserver_nodes
1560 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
1561 0 :
1562 0 : tracing::info!("Loading safekeepers from database...");
1563 0 : let safekeepers = persistence
1564 0 : .list_safekeepers()
1565 0 : .await?
1566 0 : .into_iter()
1567 0 : .map(|skp| {
1568 0 : Safekeeper::from_persistence(
1569 0 : skp,
1570 0 : CancellationToken::new(),
1571 0 : config.use_https_safekeeper_api,
1572 0 : )
1573 0 : })
1574 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1575 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1576 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1577 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1578 0 : metrics::METRICS_REGISTRY
1579 0 : .metrics_group
1580 0 : .storage_controller_safekeeper_nodes
1581 0 : .set(safekeepers.len() as i64);
1582 0 : metrics::METRICS_REGISTRY
1583 0 : .metrics_group
1584 0 : .storage_controller_https_safekeeper_nodes
1585 0 : .set(safekeepers.values().filter(|s| s.has_https_port()).count() as i64);
1586 0 :
1587 0 : tracing::info!("Loading shards from database...");
1588 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1589 0 : tracing::info!(
1590 0 : "Loaded {} shards from database.",
1591 0 : tenant_shard_persistence.len()
1592 : );
1593 :
1594 : // If any shard splits were in progress, reset the database state to abort them
1595 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1596 0 : HashMap::new();
1597 0 : for tsp in &mut tenant_shard_persistence {
1598 0 : let shard = tsp.get_shard_identity()?;
1599 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1600 0 : let entry = tenant_shard_count_min_max
1601 0 : .entry(tenant_shard_id.tenant_id)
1602 0 : .or_insert_with(|| (shard.count, shard.count));
1603 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1604 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1605 0 : }
1606 :
1607 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1608 0 : if count_min != count_max {
1609 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1610 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1611 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1612 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1613 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1614 :
1615 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1616 : // identified this tenant has having mismatching min/max counts.
1617 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1618 :
1619 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1620 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1621 0 : // Set idle split state on those shards that we will retain.
1622 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1623 0 : if tsp_tenant_id == tenant_id
1624 0 : && tsp.get_shard_identity().unwrap().count == count_min
1625 0 : {
1626 0 : tsp.splitting = SplitState::Idle;
1627 0 : } else if tsp_tenant_id == tenant_id {
1628 : // Leave the splitting state on the child shards: this will be used next to
1629 : // drop them.
1630 0 : tracing::info!(
1631 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1632 : );
1633 0 : }
1634 0 : });
1635 0 :
1636 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1637 0 : tenant_shard_persistence.retain(|tsp| {
1638 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1639 0 : || tsp.splitting == SplitState::Idle
1640 0 : });
1641 0 : }
1642 : }
1643 :
1644 0 : let mut tenants = BTreeMap::new();
1645 0 :
1646 0 : let mut scheduler = Scheduler::new(nodes.values());
1647 :
1648 : #[cfg(feature = "testing")]
1649 : {
1650 : use pageserver_api::controller_api::AvailabilityZone;
1651 :
1652 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1653 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1654 : // after when pageservers start up and register.
1655 0 : let mut node_ids = HashSet::new();
1656 0 : for tsp in &tenant_shard_persistence {
1657 0 : if let Some(node_id) = tsp.generation_pageserver {
1658 0 : node_ids.insert(node_id);
1659 0 : }
1660 : }
1661 0 : for node_id in node_ids {
1662 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1663 0 : let node = Node::new(
1664 0 : NodeId(node_id as u64),
1665 0 : "".to_string(),
1666 0 : 123,
1667 0 : None,
1668 0 : "".to_string(),
1669 0 : 123,
1670 0 : AvailabilityZone("test_az".to_string()),
1671 0 : false,
1672 0 : )
1673 0 : .unwrap();
1674 0 :
1675 0 : scheduler.node_upsert(&node);
1676 : }
1677 : }
1678 0 : for tsp in tenant_shard_persistence {
1679 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1680 :
1681 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1682 : // it with what we can infer: the node for which a generation was most recently issued.
1683 0 : let mut intent = IntentState::new(
1684 0 : tsp.preferred_az_id
1685 0 : .as_ref()
1686 0 : .map(|az| AvailabilityZone(az.clone())),
1687 0 : );
1688 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1689 : {
1690 0 : if nodes.contains_key(&generation_pageserver) {
1691 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1692 0 : } else {
1693 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1694 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1695 : // on different pageservers.
1696 0 : tracing::warn!(
1697 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1698 : );
1699 : }
1700 0 : }
1701 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1702 :
1703 0 : tenants.insert(tenant_shard_id, new_tenant);
1704 : }
1705 :
1706 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1707 0 :
1708 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1709 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1710 0 : tokio::sync::mpsc::channel(512);
1711 0 :
1712 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1713 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1714 0 :
1715 0 : let cancel = CancellationToken::new();
1716 0 : let reconcilers_cancel = cancel.child_token();
1717 0 :
1718 0 : let mut http_client = reqwest::Client::builder();
1719 0 : // We intentionally disable the connection pool, so every request will create its own TCP connection.
1720 0 : // It's especially important for heartbeaters to notice more network problems.
1721 0 : //
1722 0 : // TODO: It makes sense to use this client only in heartbeaters and create a second one with
1723 0 : // connection pooling for everything else. But reqwest::Client may create a connection without
1724 0 : // ever using it (it uses hyper's Client under the hood):
1725 0 : // https://github.com/hyperium/hyper-util/blob/d51318df3461d40e5f5e5ca163cb3905ac960209/src/client/legacy/client.rs#L415
1726 0 : //
1727 0 : // Because of a bug in hyper0::Connection::graceful_shutdown such connections hang during
1728 0 : // graceful server shutdown: https://github.com/hyperium/hyper/issues/2730
1729 0 : //
1730 0 : // The bug has been fixed in hyper v1, so keep alive may be enabled only after we migrate to hyper1.
1731 0 : http_client = http_client.pool_max_idle_per_host(0);
1732 0 : for ssl_ca_cert in &config.ssl_ca_certs {
1733 0 : http_client = http_client.add_root_certificate(ssl_ca_cert.clone());
1734 0 : }
1735 0 : let http_client = http_client.build()?;
1736 :
1737 0 : let heartbeater_ps = Heartbeater::new(
1738 0 : http_client.clone(),
1739 0 : config.pageserver_jwt_token.clone(),
1740 0 : config.max_offline_interval,
1741 0 : config.max_warming_up_interval,
1742 0 : cancel.clone(),
1743 0 : );
1744 0 :
1745 0 : let heartbeater_sk = Heartbeater::new(
1746 0 : http_client.clone(),
1747 0 : config.safekeeper_jwt_token.clone(),
1748 0 : config.max_offline_interval,
1749 0 : config.max_warming_up_interval,
1750 0 : cancel.clone(),
1751 0 : );
1752 :
1753 0 : let initial_leadership_status = if config.start_as_candidate {
1754 0 : LeadershipStatus::Candidate
1755 : } else {
1756 0 : LeadershipStatus::Leader
1757 : };
1758 :
1759 0 : let this = Arc::new(Self {
1760 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1761 0 : nodes,
1762 0 : safekeepers,
1763 0 : tenants,
1764 0 : scheduler,
1765 0 : delayed_reconcile_rx,
1766 0 : initial_leadership_status,
1767 0 : reconcilers_cancel.clone(),
1768 0 : ))),
1769 0 : config: config.clone(),
1770 0 : persistence,
1771 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())?),
1772 0 : result_tx,
1773 0 : heartbeater_ps,
1774 0 : heartbeater_sk,
1775 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1776 0 : config.reconciler_concurrency,
1777 0 : )),
1778 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1779 0 : config.priority_reconciler_concurrency,
1780 0 : )),
1781 0 : delayed_reconcile_tx,
1782 0 : abort_tx,
1783 0 : startup_complete: startup_complete.clone(),
1784 0 : cancel,
1785 0 : reconcilers_cancel,
1786 0 : gate: Gate::default(),
1787 0 : reconcilers_gate: Gate::default(),
1788 0 : tenant_op_locks: Default::default(),
1789 0 : node_op_locks: Default::default(),
1790 0 : http_client,
1791 0 : step_down_barrier: Default::default(),
1792 0 : });
1793 0 :
1794 0 : let result_task_this = this.clone();
1795 0 : tokio::task::spawn(async move {
1796 : // Block shutdown until we're done (we must respect self.cancel)
1797 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1798 0 : result_task_this
1799 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1800 0 : .await
1801 0 : }
1802 0 : });
1803 0 :
1804 0 : tokio::task::spawn({
1805 0 : let this = this.clone();
1806 0 : async move {
1807 : // Block shutdown until we're done (we must respect self.cancel)
1808 0 : if let Ok(_gate) = this.gate.enter() {
1809 0 : this.process_aborts(abort_rx).await
1810 0 : }
1811 0 : }
1812 0 : });
1813 0 :
1814 0 : tokio::task::spawn({
1815 0 : let this = this.clone();
1816 0 : async move {
1817 0 : if let Ok(_gate) = this.gate.enter() {
1818 : loop {
1819 0 : tokio::select! {
1820 0 : _ = this.cancel.cancelled() => {
1821 0 : break;
1822 : },
1823 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1824 0 : };
1825 0 : this.tenant_op_locks.housekeeping();
1826 : }
1827 0 : }
1828 0 : }
1829 0 : });
1830 0 :
1831 0 : tokio::task::spawn({
1832 0 : let this = this.clone();
1833 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1834 0 : // is done.
1835 0 : let startup_completion = startup_completion.clone();
1836 0 : async move {
1837 : // Block shutdown until we're done (we must respect self.cancel)
1838 0 : let Ok(_gate) = this.gate.enter() else {
1839 0 : return;
1840 : };
1841 :
1842 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1843 0 : .await;
1844 :
1845 0 : drop(startup_completion);
1846 0 : }
1847 0 : });
1848 0 :
1849 0 : tokio::task::spawn({
1850 0 : let this = this.clone();
1851 0 : let startup_complete = startup_complete.clone();
1852 0 : async move {
1853 0 : startup_complete.wait().await;
1854 0 : this.background_reconcile().await;
1855 0 : }
1856 0 : });
1857 0 :
1858 0 : tokio::task::spawn({
1859 0 : let this = this.clone();
1860 0 : let startup_complete = startup_complete.clone();
1861 0 : async move {
1862 0 : startup_complete.wait().await;
1863 0 : this.spawn_heartbeat_driver().await;
1864 0 : }
1865 0 : });
1866 0 :
1867 0 : Ok(this)
1868 0 : }
1869 :
1870 0 : pub(crate) async fn attach_hook(
1871 0 : &self,
1872 0 : attach_req: AttachHookRequest,
1873 0 : ) -> anyhow::Result<AttachHookResponse> {
1874 0 : let _tenant_lock = trace_exclusive_lock(
1875 0 : &self.tenant_op_locks,
1876 0 : attach_req.tenant_shard_id.tenant_id,
1877 0 : TenantOperations::AttachHook,
1878 0 : )
1879 0 : .await;
1880 :
1881 : // This is a test hook. To enable using it on tenants that were created directly with
1882 : // the pageserver API (not via this service), we will auto-create any missing tenant
1883 : // shards with default state.
1884 0 : let insert = {
1885 0 : match self
1886 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1887 0 : .await
1888 : {
1889 0 : Ok(_) => false,
1890 0 : Err(ApiError::NotFound(_)) => true,
1891 0 : Err(e) => return Err(e.into()),
1892 : }
1893 : };
1894 :
1895 0 : if insert {
1896 0 : let config = attach_req.config.clone().unwrap_or_default();
1897 0 : let tsp = TenantShardPersistence {
1898 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1899 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1900 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1901 0 : shard_stripe_size: 0,
1902 0 : generation: attach_req.generation_override.or(Some(0)),
1903 0 : generation_pageserver: None,
1904 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1905 0 : config: serde_json::to_string(&config).unwrap(),
1906 0 : splitting: SplitState::default(),
1907 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1908 0 : .unwrap(),
1909 0 : preferred_az_id: None,
1910 0 : };
1911 0 :
1912 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1913 0 : Err(e) => match e {
1914 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1915 : DatabaseErrorKind::UniqueViolation,
1916 : _,
1917 : )) => {
1918 0 : tracing::info!(
1919 0 : "Raced with another request to insert tenant {}",
1920 : attach_req.tenant_shard_id
1921 : )
1922 : }
1923 0 : _ => return Err(e.into()),
1924 : },
1925 : Ok(()) => {
1926 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1927 :
1928 0 : let mut shard = TenantShard::new(
1929 0 : attach_req.tenant_shard_id,
1930 0 : ShardIdentity::unsharded(),
1931 0 : PlacementPolicy::Attached(0),
1932 0 : None,
1933 0 : );
1934 0 : shard.config = config;
1935 0 :
1936 0 : let mut locked = self.inner.write().unwrap();
1937 0 : locked.tenants.insert(attach_req.tenant_shard_id, shard);
1938 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1939 : }
1940 : }
1941 0 : }
1942 :
1943 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1944 0 : let maybe_tenant_conf = {
1945 0 : let locked = self.inner.write().unwrap();
1946 0 : locked
1947 0 : .tenants
1948 0 : .get(&attach_req.tenant_shard_id)
1949 0 : .map(|t| t.config.clone())
1950 0 : };
1951 0 :
1952 0 : match maybe_tenant_conf {
1953 0 : Some(conf) => {
1954 0 : let new_generation = self
1955 0 : .persistence
1956 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1957 0 : .await?;
1958 :
1959 : // Persist the placement policy update. This is required
1960 : // when we reattaching a detached tenant.
1961 0 : self.persistence
1962 0 : .update_tenant_shard(
1963 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1964 0 : Some(PlacementPolicy::Attached(0)),
1965 0 : Some(conf),
1966 0 : None,
1967 0 : None,
1968 0 : )
1969 0 : .await?;
1970 0 : Some(new_generation)
1971 : }
1972 : None => {
1973 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1974 : }
1975 : }
1976 : } else {
1977 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1978 0 : None
1979 : };
1980 :
1981 0 : let mut locked = self.inner.write().unwrap();
1982 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1983 0 :
1984 0 : let tenant_shard = tenants
1985 0 : .get_mut(&attach_req.tenant_shard_id)
1986 0 : .expect("Checked for existence above");
1987 :
1988 0 : if let Some(new_generation) = new_generation {
1989 0 : tenant_shard.generation = Some(new_generation);
1990 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1991 0 : } else {
1992 : // This is a detach notification. We must update placement policy to avoid re-attaching
1993 : // during background scheduling/reconciliation, or during storage controller restart.
1994 0 : assert!(attach_req.node_id.is_none());
1995 0 : tenant_shard.policy = PlacementPolicy::Detached;
1996 : }
1997 :
1998 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1999 0 : tracing::info!(
2000 : tenant_id = %attach_req.tenant_shard_id,
2001 : ps_id = %attaching_pageserver,
2002 : generation = ?tenant_shard.generation,
2003 0 : "issuing",
2004 : );
2005 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
2006 0 : tracing::info!(
2007 : tenant_id = %attach_req.tenant_shard_id,
2008 : %ps_id,
2009 : generation = ?tenant_shard.generation,
2010 0 : "dropping",
2011 : );
2012 : } else {
2013 0 : tracing::info!(
2014 : tenant_id = %attach_req.tenant_shard_id,
2015 0 : "no-op: tenant already has no pageserver");
2016 : }
2017 0 : tenant_shard
2018 0 : .intent
2019 0 : .set_attached(scheduler, attach_req.node_id);
2020 0 :
2021 0 : tracing::info!(
2022 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}, config {:?}",
2023 0 : attach_req.tenant_shard_id,
2024 0 : tenant_shard.generation,
2025 0 : // TODO: this is an odd number of 0xf's
2026 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff)),
2027 : attach_req.config,
2028 : );
2029 :
2030 : // Trick the reconciler into not doing anything for this tenant: this helps
2031 : // tests that manually configure a tenant on the pagesrever, and then call this
2032 : // attach hook: they don't want background reconciliation to modify what they
2033 : // did to the pageserver.
2034 : #[cfg(feature = "testing")]
2035 : {
2036 0 : if let Some(node_id) = attach_req.node_id {
2037 0 : tenant_shard.observed.locations = HashMap::from([(
2038 0 : node_id,
2039 0 : ObservedStateLocation {
2040 0 : conf: Some(attached_location_conf(
2041 0 : tenant_shard.generation.unwrap(),
2042 0 : &tenant_shard.shard,
2043 0 : &tenant_shard.config,
2044 0 : &PlacementPolicy::Attached(0),
2045 0 : )),
2046 0 : },
2047 0 : )]);
2048 0 : } else {
2049 0 : tenant_shard.observed.locations.clear();
2050 0 : }
2051 : }
2052 :
2053 0 : Ok(AttachHookResponse {
2054 0 : generation: attach_req
2055 0 : .node_id
2056 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
2057 0 : })
2058 0 : }
2059 :
2060 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
2061 0 : let locked = self.inner.read().unwrap();
2062 0 :
2063 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
2064 0 :
2065 0 : InspectResponse {
2066 0 : attachment: tenant_shard.and_then(|s| {
2067 0 : s.intent
2068 0 : .get_attached()
2069 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
2070 0 : }),
2071 0 : }
2072 0 : }
2073 :
2074 : // When the availability state of a node transitions to active, we must do a full reconciliation
2075 : // of LocationConfigs on that node. This is because while a node was offline:
2076 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
2077 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
2078 : //
2079 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
2080 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
2081 : // this function.
2082 : //
2083 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
2084 : // for written for a single node rather than as a batch job for all nodes.
2085 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
2086 : async fn node_activate_reconcile(
2087 : &self,
2088 : mut node: Node,
2089 : _lock: &TracingExclusiveGuard<NodeOperations>,
2090 : ) -> Result<(), ApiError> {
2091 : // This Node is a mutable local copy: we will set it active so that we can use its
2092 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
2093 : // later.
2094 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
2095 :
2096 : let configs = match node
2097 : .with_client_retries(
2098 0 : |client| async move { client.list_location_config().await },
2099 : &self.http_client,
2100 : &self.config.pageserver_jwt_token,
2101 : 1,
2102 : 5,
2103 : SHORT_RECONCILE_TIMEOUT,
2104 : &self.cancel,
2105 : )
2106 : .await
2107 : {
2108 : None => {
2109 : // We're shutting down (the Node's cancellation token can't have fired, because
2110 : // we're the only scope that has a reference to it, and we didn't fire it).
2111 : return Err(ApiError::ShuttingDown);
2112 : }
2113 : Some(Err(e)) => {
2114 : // This node didn't succeed listing its locations: it may not proceed to active state
2115 : // as it is apparently unavailable.
2116 : return Err(ApiError::PreconditionFailed(
2117 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2118 : ));
2119 : }
2120 : Some(Ok(configs)) => configs,
2121 : };
2122 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2123 :
2124 : let mut cleanup = Vec::new();
2125 : let mut mismatched_locations = 0;
2126 : {
2127 : let mut locked = self.inner.write().unwrap();
2128 :
2129 : for (tenant_shard_id, reported) in configs.tenant_shards {
2130 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2131 : cleanup.push(tenant_shard_id);
2132 : continue;
2133 : };
2134 :
2135 : let on_record = &mut tenant_shard
2136 : .observed
2137 : .locations
2138 : .entry(node.get_id())
2139 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2140 : .conf;
2141 :
2142 : // If the location reported by the node does not match our observed state,
2143 : // then we mark it as uncertain and let the background reconciliation loop
2144 : // deal with it.
2145 : //
2146 : // Note that this also covers net new locations reported by the node.
2147 : if *on_record != reported {
2148 : mismatched_locations += 1;
2149 : *on_record = None;
2150 : }
2151 : }
2152 : }
2153 :
2154 : if mismatched_locations > 0 {
2155 : tracing::info!(
2156 : "Set observed state to None for {mismatched_locations} mismatched locations"
2157 : );
2158 : }
2159 :
2160 : for tenant_shard_id in cleanup {
2161 : tracing::info!("Detaching {tenant_shard_id}");
2162 : match node
2163 : .with_client_retries(
2164 0 : |client| async move {
2165 0 : let config = LocationConfig {
2166 0 : mode: LocationConfigMode::Detached,
2167 0 : generation: None,
2168 0 : secondary_conf: None,
2169 0 : shard_number: tenant_shard_id.shard_number.0,
2170 0 : shard_count: tenant_shard_id.shard_count.literal(),
2171 0 : shard_stripe_size: 0,
2172 0 : tenant_conf: models::TenantConfig::default(),
2173 0 : };
2174 0 : client
2175 0 : .location_config(tenant_shard_id, config, None, false)
2176 0 : .await
2177 0 : },
2178 : &self.http_client,
2179 : &self.config.pageserver_jwt_token,
2180 : 1,
2181 : 5,
2182 : SHORT_RECONCILE_TIMEOUT,
2183 : &self.cancel,
2184 : )
2185 : .await
2186 : {
2187 : None => {
2188 : // We're shutting down (the Node's cancellation token can't have fired, because
2189 : // we're the only scope that has a reference to it, and we didn't fire it).
2190 : return Err(ApiError::ShuttingDown);
2191 : }
2192 : Some(Err(e)) => {
2193 : // Do not let the node proceed to Active state if it is not responsive to requests
2194 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2195 : // detach completing: we should not let this node back into the set of nodes considered
2196 : // okay for scheduling.
2197 : return Err(ApiError::Conflict(format!(
2198 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2199 : )));
2200 : }
2201 : Some(Ok(_)) => {}
2202 : };
2203 : }
2204 :
2205 : Ok(())
2206 : }
2207 :
2208 0 : pub(crate) async fn re_attach(
2209 0 : &self,
2210 0 : reattach_req: ReAttachRequest,
2211 0 : ) -> Result<ReAttachResponse, ApiError> {
2212 0 : if let Some(register_req) = reattach_req.register {
2213 0 : self.node_register(register_req).await?;
2214 0 : }
2215 :
2216 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2217 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2218 :
2219 0 : tracing::info!(
2220 : node_id=%reattach_req.node_id,
2221 0 : "Incremented {} tenant shards' generations",
2222 0 : incremented_generations.len()
2223 : );
2224 :
2225 : // Apply the updated generation to our in-memory state, and
2226 : // gather discover secondary locations.
2227 0 : let mut locked = self.inner.write().unwrap();
2228 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2229 0 :
2230 0 : let mut response = ReAttachResponse {
2231 0 : tenants: Vec::new(),
2232 0 : };
2233 :
2234 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2235 : // to call location_conf API with an old generation. Wait for cancellation to complete
2236 : // before responding to this request. Requires well implemented CancellationToken logic
2237 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2238 : // request in flight over the network: TODO handle that by making location_conf API refuse
2239 : // to go backward in generations.
2240 :
2241 : // Scan through all shards, applying updates for ones where we updated generation
2242 : // and identifying shards that intend to have a secondary location on this node.
2243 0 : for (tenant_shard_id, shard) in tenants {
2244 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2245 0 : let new_gen = *new_gen;
2246 0 : response.tenants.push(ReAttachResponseTenant {
2247 0 : id: *tenant_shard_id,
2248 0 : r#gen: Some(new_gen.into().unwrap()),
2249 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2250 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2251 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2252 0 : // the stale/multi states at this point.
2253 0 : mode: LocationConfigMode::AttachedSingle,
2254 0 : });
2255 0 :
2256 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2257 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2258 : // Why can we update `observed` even though we're not sure our response will be received
2259 : // by the pageserver? Because the pageserver will not proceed with startup until
2260 : // it has processed response: if it loses it, we'll see another request and increment
2261 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2262 0 : if let Some(conf) = observed.conf.as_mut() {
2263 0 : conf.generation = new_gen.into();
2264 0 : }
2265 0 : } else {
2266 0 : // This node has no observed state for the shard: perhaps it was offline
2267 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2268 0 : // will be prompted to learn the location's state before it makes changes.
2269 0 : shard
2270 0 : .observed
2271 0 : .locations
2272 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2273 0 : }
2274 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2275 0 : // Ordering: pageserver will not accept /location_config requests until it has
2276 0 : // finished processing the response from re-attach. So we can update our in-memory state
2277 0 : // now, and be confident that we are not stamping on the result of some later location config.
2278 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2279 0 : // so we might update observed state here, and then get over-written by some racing
2280 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2281 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2282 0 :
2283 0 : response.tenants.push(ReAttachResponseTenant {
2284 0 : id: *tenant_shard_id,
2285 0 : r#gen: None,
2286 0 : mode: LocationConfigMode::Secondary,
2287 0 : });
2288 0 :
2289 0 : // We must not update observed, because we have no guarantee that our
2290 0 : // response will be received by the pageserver. This could leave it
2291 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2292 0 : }
2293 : }
2294 :
2295 : // We consider a node Active once we have composed a re-attach response, but we
2296 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2297 : // implicitly synchronizes the LocationConfigs on the node.
2298 : //
2299 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2300 : // but those requests will not be accepted by the node until it has finished processing
2301 : // the re-attach response.
2302 : //
2303 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2304 : // in [`Persistence::re_attach`].
2305 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2306 0 : let reset_scheduling = matches!(
2307 0 : node.get_scheduling(),
2308 : NodeSchedulingPolicy::PauseForRestart
2309 : | NodeSchedulingPolicy::Draining
2310 : | NodeSchedulingPolicy::Filling
2311 : );
2312 :
2313 0 : let mut new_nodes = (**nodes).clone();
2314 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2315 0 : if reset_scheduling {
2316 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2317 0 : }
2318 :
2319 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2320 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2321 0 :
2322 0 : scheduler.node_upsert(node);
2323 0 : let new_nodes = Arc::new(new_nodes);
2324 0 : *nodes = new_nodes;
2325 : } else {
2326 0 : tracing::error!(
2327 0 : "Reattaching node {} was removed while processing the request",
2328 : reattach_req.node_id
2329 : );
2330 : }
2331 0 : }
2332 :
2333 0 : Ok(response)
2334 0 : }
2335 :
2336 0 : pub(crate) async fn validate(
2337 0 : &self,
2338 0 : validate_req: ValidateRequest,
2339 0 : ) -> Result<ValidateResponse, DatabaseError> {
2340 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2341 : // in-memory generation for a shard
2342 0 : let in_memory_result = {
2343 0 : let mut in_memory_result = Vec::new();
2344 0 : let locked = self.inner.read().unwrap();
2345 0 : for req_tenant in validate_req.tenants {
2346 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2347 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2348 0 : tracing::info!(
2349 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2350 : req_tenant.id,
2351 : req_tenant.r#gen,
2352 : tenant_shard.generation
2353 : );
2354 :
2355 0 : in_memory_result.push((
2356 0 : req_tenant.id,
2357 0 : Generation::new(req_tenant.r#gen),
2358 0 : valid,
2359 0 : ));
2360 : } else {
2361 : // This is legal: for example during a shard split the pageserver may still
2362 : // have deletions in its queue from the old pre-split shard, or after deletion
2363 : // of a tenant that was busy with compaction/gc while being deleted.
2364 0 : tracing::info!(
2365 0 : "Refusing deletion validation for missing shard {}",
2366 : req_tenant.id
2367 : );
2368 : }
2369 : }
2370 :
2371 0 : in_memory_result
2372 : };
2373 :
2374 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2375 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2376 0 : let db_generations = self
2377 0 : .persistence
2378 0 : .shard_generations(
2379 0 : in_memory_result
2380 0 : .iter()
2381 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2382 0 : )
2383 0 : .await?;
2384 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2385 0 :
2386 0 : let mut response = ValidateResponse {
2387 0 : tenants: Vec::new(),
2388 0 : };
2389 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2390 0 : let valid = if valid {
2391 0 : let db_generation = db_generations.get(&tenant_shard_id);
2392 0 : db_generation == Some(&Some(validate_generation))
2393 : } else {
2394 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2395 : // this prevents a pageserver from cleaning up an object in S3.
2396 0 : false
2397 : };
2398 :
2399 0 : response.tenants.push(ValidateResponseTenant {
2400 0 : id: tenant_shard_id,
2401 0 : valid,
2402 0 : })
2403 : }
2404 :
2405 0 : Ok(response)
2406 0 : }
2407 :
2408 0 : pub(crate) async fn tenant_create(
2409 0 : &self,
2410 0 : create_req: TenantCreateRequest,
2411 0 : ) -> Result<TenantCreateResponse, ApiError> {
2412 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2413 :
2414 : // Exclude any concurrent attempts to create/access the same tenant ID
2415 0 : let _tenant_lock = trace_exclusive_lock(
2416 0 : &self.tenant_op_locks,
2417 0 : create_req.new_tenant_id.tenant_id,
2418 0 : TenantOperations::Create,
2419 0 : )
2420 0 : .await;
2421 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2422 :
2423 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2424 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2425 : // accept compute notifications while it is in the process of creating. Reconciliation will
2426 : // be retried in the background.
2427 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2428 0 : }
2429 0 : Ok(response)
2430 0 : }
2431 :
2432 0 : pub(crate) async fn do_tenant_create(
2433 0 : &self,
2434 0 : create_req: TenantCreateRequest,
2435 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2436 0 : let placement_policy = create_req
2437 0 : .placement_policy
2438 0 : .clone()
2439 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2440 0 : .unwrap_or(PlacementPolicy::Attached(0));
2441 :
2442 : // This service expects to handle sharding itself: it is an error to try and directly create
2443 : // a particular shard here.
2444 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2445 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2446 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2447 0 : )));
2448 : } else {
2449 0 : create_req.new_tenant_id.tenant_id
2450 0 : };
2451 0 :
2452 0 : tracing::info!(
2453 0 : "Creating tenant {}, shard_count={:?}",
2454 : create_req.new_tenant_id,
2455 : create_req.shard_parameters.count,
2456 : );
2457 :
2458 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2459 0 : .map(|i| TenantShardId {
2460 0 : tenant_id,
2461 0 : shard_number: ShardNumber(i),
2462 0 : shard_count: create_req.shard_parameters.count,
2463 0 : })
2464 0 : .collect::<Vec<_>>();
2465 :
2466 : // If the caller specifies a None generation, it means "start from default". This is different
2467 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2468 : // an incompletely-onboarded tenant.
2469 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2470 0 : tracing::info!(
2471 0 : "tenant_create: secondary mode, generation is_some={}",
2472 0 : create_req.generation.is_some()
2473 : );
2474 0 : create_req.generation.map(Generation::new)
2475 : } else {
2476 0 : tracing::info!(
2477 0 : "tenant_create: not secondary mode, generation is_some={}",
2478 0 : create_req.generation.is_some()
2479 : );
2480 0 : Some(
2481 0 : create_req
2482 0 : .generation
2483 0 : .map(Generation::new)
2484 0 : .unwrap_or(INITIAL_GENERATION),
2485 0 : )
2486 : };
2487 :
2488 0 : let preferred_az_id = {
2489 0 : let locked = self.inner.read().unwrap();
2490 : // Idempotency: take the existing value if the tenant already exists
2491 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2492 0 : shard.preferred_az().cloned()
2493 : } else {
2494 0 : locked.scheduler.get_az_for_new_tenant()
2495 : }
2496 : };
2497 :
2498 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2499 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2500 : // during the creation, rather than risking leaving orphan objects in S3.
2501 0 : let persist_tenant_shards = create_ids
2502 0 : .iter()
2503 0 : .map(|tenant_shard_id| TenantShardPersistence {
2504 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2505 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2506 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2507 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2508 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2509 0 : // The pageserver is not known until scheduling happens: we will set this column when
2510 0 : // incrementing the generation the first time we attach to a pageserver.
2511 0 : generation_pageserver: None,
2512 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2513 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2514 0 : splitting: SplitState::default(),
2515 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2516 0 : .unwrap(),
2517 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2518 0 : })
2519 0 : .collect();
2520 0 :
2521 0 : match self
2522 0 : .persistence
2523 0 : .insert_tenant_shards(persist_tenant_shards)
2524 0 : .await
2525 : {
2526 0 : Ok(_) => {}
2527 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2528 : DatabaseErrorKind::UniqueViolation,
2529 : _,
2530 : ))) => {
2531 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2532 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2533 : // creation's shard count.
2534 0 : tracing::info!(
2535 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2536 : );
2537 : }
2538 : // Any other database error is unexpected and a bug.
2539 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2540 : };
2541 :
2542 0 : let mut schedule_context = ScheduleContext::default();
2543 0 : let mut schedule_error = None;
2544 0 : let mut response_shards = Vec::new();
2545 0 : for tenant_shard_id in create_ids {
2546 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2547 :
2548 0 : let outcome = self
2549 0 : .do_initial_shard_scheduling(
2550 0 : tenant_shard_id,
2551 0 : initial_generation,
2552 0 : &create_req.shard_parameters,
2553 0 : create_req.config.clone(),
2554 0 : placement_policy.clone(),
2555 0 : preferred_az_id.as_ref(),
2556 0 : &mut schedule_context,
2557 0 : )
2558 0 : .await;
2559 :
2560 0 : match outcome {
2561 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2562 0 : InitialShardScheduleOutcome::NotScheduled => {}
2563 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2564 0 : schedule_error = Some(err);
2565 0 : }
2566 : }
2567 : }
2568 :
2569 : // If we failed to schedule shards, then they are still created in the controller,
2570 : // but we return an error to the requester to avoid a silent failure when someone
2571 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2572 : // are present in the system. We do this here rather than in the above loop, to
2573 : // avoid situations where we only create a subset of shards in the tenant.
2574 0 : if let Some(e) = schedule_error {
2575 0 : return Err(ApiError::Conflict(format!(
2576 0 : "Failed to schedule shard(s): {e}"
2577 0 : )));
2578 0 : }
2579 0 :
2580 0 : let waiters = {
2581 0 : let mut locked = self.inner.write().unwrap();
2582 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2583 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2584 0 : .tenant_creation_hint(true)
2585 0 : .build();
2586 0 : tenants
2587 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2588 0 : .filter_map(|(_shard_id, shard)| {
2589 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2590 0 : })
2591 0 : .collect::<Vec<_>>()
2592 0 : };
2593 0 :
2594 0 : Ok((
2595 0 : TenantCreateResponse {
2596 0 : shards: response_shards,
2597 0 : },
2598 0 : waiters,
2599 0 : ))
2600 0 : }
2601 :
2602 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2603 : /// case of a new tenant and a pre-existing one.
2604 : #[allow(clippy::too_many_arguments)]
2605 0 : async fn do_initial_shard_scheduling(
2606 0 : &self,
2607 0 : tenant_shard_id: TenantShardId,
2608 0 : initial_generation: Option<Generation>,
2609 0 : shard_params: &ShardParameters,
2610 0 : config: TenantConfig,
2611 0 : placement_policy: PlacementPolicy,
2612 0 : preferred_az_id: Option<&AvailabilityZone>,
2613 0 : schedule_context: &mut ScheduleContext,
2614 0 : ) -> InitialShardScheduleOutcome {
2615 0 : let mut locked = self.inner.write().unwrap();
2616 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2617 :
2618 : use std::collections::btree_map::Entry;
2619 0 : match tenants.entry(tenant_shard_id) {
2620 0 : Entry::Occupied(mut entry) => {
2621 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2622 :
2623 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2624 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2625 0 : }
2626 :
2627 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2628 0 : let generation = entry
2629 0 : .get()
2630 0 : .generation
2631 0 : .expect("Generation is set when in attached mode");
2632 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2633 0 : shard_id: tenant_shard_id,
2634 0 : node_id: *node_id,
2635 0 : generation: generation.into().unwrap(),
2636 0 : })
2637 : } else {
2638 0 : InitialShardScheduleOutcome::NotScheduled
2639 : }
2640 : }
2641 0 : Entry::Vacant(entry) => {
2642 0 : let state = entry.insert(TenantShard::new(
2643 0 : tenant_shard_id,
2644 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2645 0 : placement_policy,
2646 0 : preferred_az_id.cloned(),
2647 0 : ));
2648 0 :
2649 0 : state.generation = initial_generation;
2650 0 : state.config = config;
2651 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2652 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2653 0 : }
2654 :
2655 : // Only include shards in result if we are attaching: the purpose
2656 : // of the response is to tell the caller where the shards are attached.
2657 0 : if let Some(node_id) = state.intent.get_attached() {
2658 0 : let generation = state
2659 0 : .generation
2660 0 : .expect("Generation is set when in attached mode");
2661 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2662 0 : shard_id: tenant_shard_id,
2663 0 : node_id: *node_id,
2664 0 : generation: generation.into().unwrap(),
2665 0 : })
2666 : } else {
2667 0 : InitialShardScheduleOutcome::NotScheduled
2668 : }
2669 : }
2670 : }
2671 0 : }
2672 :
2673 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2674 : /// wait for reconciliation to complete before responding.
2675 0 : async fn await_waiters(
2676 0 : &self,
2677 0 : waiters: Vec<ReconcilerWaiter>,
2678 0 : timeout: Duration,
2679 0 : ) -> Result<(), ReconcileWaitError> {
2680 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2681 0 : for waiter in waiters {
2682 0 : let timeout = deadline.duration_since(Instant::now());
2683 0 : waiter.wait_timeout(timeout).await?;
2684 : }
2685 :
2686 0 : Ok(())
2687 0 : }
2688 :
2689 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2690 : /// in progress
2691 0 : async fn await_waiters_remainder(
2692 0 : &self,
2693 0 : waiters: Vec<ReconcilerWaiter>,
2694 0 : timeout: Duration,
2695 0 : ) -> Vec<ReconcilerWaiter> {
2696 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2697 0 : for waiter in waiters.iter() {
2698 0 : let timeout = deadline.duration_since(Instant::now());
2699 0 : let _ = waiter.wait_timeout(timeout).await;
2700 : }
2701 :
2702 0 : waiters
2703 0 : .into_iter()
2704 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2705 0 : .collect::<Vec<_>>()
2706 0 : }
2707 :
2708 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2709 : /// and transform it into either a tenant creation of a series of shard updates.
2710 : ///
2711 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2712 : /// still be returned.
2713 0 : fn tenant_location_config_prepare(
2714 0 : &self,
2715 0 : tenant_id: TenantId,
2716 0 : req: TenantLocationConfigRequest,
2717 0 : ) -> TenantCreateOrUpdate {
2718 0 : let mut updates = Vec::new();
2719 0 : let mut locked = self.inner.write().unwrap();
2720 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2721 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2722 :
2723 : // Use location config mode as an indicator of policy.
2724 0 : let placement_policy = match req.config.mode {
2725 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2726 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2727 : LocationConfigMode::AttachedMulti
2728 : | LocationConfigMode::AttachedSingle
2729 : | LocationConfigMode::AttachedStale => {
2730 0 : if nodes.len() > 1 {
2731 0 : PlacementPolicy::Attached(1)
2732 : } else {
2733 : // Convenience for dev/test: if we just have one pageserver, import
2734 : // tenants into non-HA mode so that scheduling will succeed.
2735 0 : PlacementPolicy::Attached(0)
2736 : }
2737 : }
2738 : };
2739 :
2740 : // Ordinarily we do not update scheduling policy, but when making major changes
2741 : // like detaching or demoting to secondary-only, we need to force the scheduling
2742 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2743 0 : let scheduling_policy = match req.config.mode {
2744 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2745 : // Special case: when making major changes like detaching or demoting to secondary-only,
2746 : // we need to force the scheduling mode to Active, or nothing will happen.
2747 0 : Some(ShardSchedulingPolicy::Active)
2748 : }
2749 : LocationConfigMode::AttachedMulti
2750 : | LocationConfigMode::AttachedSingle
2751 : | LocationConfigMode::AttachedStale => {
2752 : // While attached, continue to respect whatever the existing scheduling mode is.
2753 0 : None
2754 : }
2755 : };
2756 :
2757 0 : let mut create = true;
2758 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2759 : // Saw an existing shard: this is not a creation
2760 0 : create = false;
2761 :
2762 : // Shards may have initially been created by a Secondary request, where we
2763 : // would have left generation as None.
2764 : //
2765 : // We only update generation the first time we see an attached-mode request,
2766 : // and if there is no existing generation set. The caller is responsible for
2767 : // ensuring that no non-storage-controller pageserver ever uses a higher
2768 : // generation than they passed in here.
2769 : use LocationConfigMode::*;
2770 0 : let set_generation = match req.config.mode {
2771 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2772 0 : req.config.generation.map(Generation::new)
2773 : }
2774 0 : _ => None,
2775 : };
2776 :
2777 0 : updates.push(ShardUpdate {
2778 0 : tenant_shard_id: *shard_id,
2779 0 : placement_policy: placement_policy.clone(),
2780 0 : tenant_config: req.config.tenant_conf.clone(),
2781 0 : generation: set_generation,
2782 0 : scheduling_policy,
2783 0 : });
2784 : }
2785 :
2786 0 : if create {
2787 : use LocationConfigMode::*;
2788 0 : let generation = match req.config.mode {
2789 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2790 : // If a caller provided a generation in a non-attached request, ignore it
2791 : // and leave our generation as None: this enables a subsequent update to set
2792 : // the generation when setting an attached mode for the first time.
2793 0 : _ => None,
2794 : };
2795 :
2796 0 : TenantCreateOrUpdate::Create(
2797 0 : // Synthesize a creation request
2798 0 : TenantCreateRequest {
2799 0 : new_tenant_id: tenant_shard_id,
2800 0 : generation,
2801 0 : shard_parameters: ShardParameters {
2802 0 : count: tenant_shard_id.shard_count,
2803 0 : // We only import un-sharded or single-sharded tenants, so stripe
2804 0 : // size can be made up arbitrarily here.
2805 0 : stripe_size: DEFAULT_STRIPE_SIZE,
2806 0 : },
2807 0 : placement_policy: Some(placement_policy),
2808 0 : config: req.config.tenant_conf,
2809 0 : },
2810 0 : )
2811 : } else {
2812 0 : assert!(!updates.is_empty());
2813 0 : TenantCreateOrUpdate::Update(updates)
2814 : }
2815 0 : }
2816 :
2817 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2818 : /// the tenant is present in memory. If not, load it from the database. If it is found
2819 : /// in neither location, return a NotFound error.
2820 : ///
2821 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2822 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2823 0 : async fn maybe_load_tenant(
2824 0 : &self,
2825 0 : tenant_id: TenantId,
2826 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2827 0 : ) -> Result<(), ApiError> {
2828 : // Check if the tenant is present in memory, and select an AZ to use when loading
2829 : // if we will load it.
2830 0 : let load_in_az = {
2831 0 : let locked = self.inner.read().unwrap();
2832 0 : let existing = locked
2833 0 : .tenants
2834 0 : .range(TenantShardId::tenant_range(tenant_id))
2835 0 : .next();
2836 0 :
2837 0 : // If the tenant is not present in memory, we expect to load it from database,
2838 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2839 0 : if existing.is_none() {
2840 0 : locked
2841 0 : .scheduler
2842 0 : .get_az_for_new_tenant()
2843 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2844 0 : "No AZ with nodes found to load tenant"
2845 0 : )))?
2846 : } else {
2847 : // We already have this tenant in memory
2848 0 : return Ok(());
2849 : }
2850 : };
2851 :
2852 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2853 0 : if tenant_shards.is_empty() {
2854 0 : return Err(ApiError::NotFound(
2855 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2856 0 : ));
2857 0 : }
2858 0 :
2859 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2860 0 : self.persistence
2861 0 : .set_tenant_shard_preferred_azs(
2862 0 : tenant_shards
2863 0 : .iter()
2864 0 : .map(|t| {
2865 0 : (
2866 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2867 0 : Some(load_in_az.clone()),
2868 0 : )
2869 0 : })
2870 0 : .collect(),
2871 0 : )
2872 0 : .await?;
2873 :
2874 0 : let mut locked = self.inner.write().unwrap();
2875 0 : tracing::info!(
2876 0 : "Loaded {} shards for tenant {}",
2877 0 : tenant_shards.len(),
2878 : tenant_id
2879 : );
2880 :
2881 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2882 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2883 0 : let shard =
2884 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2885 0 :
2886 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2887 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2888 0 : if shard.policy != PlacementPolicy::Detached {
2889 0 : tracing::error!(
2890 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2891 : shard.tenant_shard_id,
2892 : shard.policy
2893 : );
2894 0 : }
2895 :
2896 0 : (shard.tenant_shard_id, shard)
2897 0 : }));
2898 0 :
2899 0 : Ok(())
2900 0 : }
2901 :
2902 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2903 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2904 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2905 : ///
2906 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2907 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2908 0 : fn maybe_drop_tenant(
2909 0 : &self,
2910 0 : tenant_id: TenantId,
2911 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2912 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2913 0 : ) {
2914 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2915 0 : if tenant_shards.all(|(_id, shard)| {
2916 0 : shard.policy == PlacementPolicy::Detached
2917 0 : && shard.reconciler.is_none()
2918 0 : && shard.observed.is_empty()
2919 0 : }) {
2920 0 : let keys = locked
2921 0 : .tenants
2922 0 : .range(TenantShardId::tenant_range(tenant_id))
2923 0 : .map(|(id, _)| id)
2924 0 : .copied()
2925 0 : .collect::<Vec<_>>();
2926 0 : for key in keys {
2927 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2928 0 : locked.tenants.remove(&key);
2929 : }
2930 0 : }
2931 0 : }
2932 :
2933 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2934 : /// directly with pageservers into this service.
2935 : ///
2936 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2937 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2938 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2939 : /// tenant's source of generation numbers.
2940 : ///
2941 : /// The mode in this request coarse-grained control of tenants:
2942 : /// - Call with mode Attached* to upsert the tenant.
2943 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2944 : /// to set an existing tenant to PolicyMode::Secondary
2945 : /// - Call with mode Detached to switch to PolicyMode::Detached
2946 0 : pub(crate) async fn tenant_location_config(
2947 0 : &self,
2948 0 : tenant_shard_id: TenantShardId,
2949 0 : req: TenantLocationConfigRequest,
2950 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2951 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2952 0 : let _tenant_lock = trace_exclusive_lock(
2953 0 : &self.tenant_op_locks,
2954 0 : tenant_shard_id.tenant_id,
2955 0 : TenantOperations::LocationConfig,
2956 0 : )
2957 0 : .await;
2958 :
2959 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2960 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2961 0 : "This API is for importing single-sharded or unsharded tenants"
2962 0 : )));
2963 : } else {
2964 0 : tenant_shard_id.tenant_id
2965 0 : };
2966 0 :
2967 0 : // In case we are waking up a Detached tenant
2968 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2969 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2970 0 : // This is a creation or an update
2971 0 : }
2972 0 : Err(e) => {
2973 0 : return Err(e);
2974 : }
2975 : };
2976 :
2977 : // First check if this is a creation or an update
2978 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2979 0 :
2980 0 : let mut result = TenantLocationConfigResponse {
2981 0 : shards: Vec::new(),
2982 0 : stripe_size: None,
2983 0 : };
2984 0 : let waiters = match create_or_update {
2985 0 : TenantCreateOrUpdate::Create(create_req) => {
2986 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2987 0 : result.shards = create_resp
2988 0 : .shards
2989 0 : .into_iter()
2990 0 : .map(|s| TenantShardLocation {
2991 0 : node_id: s.node_id,
2992 0 : shard_id: s.shard_id,
2993 0 : })
2994 0 : .collect();
2995 0 : waiters
2996 : }
2997 0 : TenantCreateOrUpdate::Update(updates) => {
2998 0 : // Persist updates
2999 0 : // Ordering: write to the database before applying changes in-memory, so that
3000 0 : // we will not appear time-travel backwards on a restart.
3001 0 :
3002 0 : let mut schedule_context = ScheduleContext::default();
3003 : for ShardUpdate {
3004 0 : tenant_shard_id,
3005 0 : placement_policy,
3006 0 : tenant_config,
3007 0 : generation,
3008 0 : scheduling_policy,
3009 0 : } in &updates
3010 : {
3011 0 : self.persistence
3012 0 : .update_tenant_shard(
3013 0 : TenantFilter::Shard(*tenant_shard_id),
3014 0 : Some(placement_policy.clone()),
3015 0 : Some(tenant_config.clone()),
3016 0 : *generation,
3017 0 : *scheduling_policy,
3018 0 : )
3019 0 : .await?;
3020 : }
3021 :
3022 : // Apply updates in-memory
3023 0 : let mut waiters = Vec::new();
3024 0 : {
3025 0 : let mut locked = self.inner.write().unwrap();
3026 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3027 :
3028 : for ShardUpdate {
3029 0 : tenant_shard_id,
3030 0 : placement_policy,
3031 0 : tenant_config,
3032 0 : generation: update_generation,
3033 0 : scheduling_policy,
3034 0 : } in updates
3035 : {
3036 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
3037 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
3038 0 : continue;
3039 : };
3040 :
3041 : // Update stripe size
3042 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
3043 0 : result.stripe_size = Some(shard.shard.stripe_size);
3044 0 : }
3045 :
3046 0 : shard.policy = placement_policy;
3047 0 : shard.config = tenant_config;
3048 0 : if let Some(generation) = update_generation {
3049 0 : shard.generation = Some(generation);
3050 0 : }
3051 :
3052 0 : if let Some(scheduling_policy) = scheduling_policy {
3053 0 : shard.set_scheduling_policy(scheduling_policy);
3054 0 : }
3055 :
3056 0 : shard.schedule(scheduler, &mut schedule_context)?;
3057 :
3058 0 : let maybe_waiter =
3059 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3060 0 : if let Some(waiter) = maybe_waiter {
3061 0 : waiters.push(waiter);
3062 0 : }
3063 :
3064 0 : if let Some(node_id) = shard.intent.get_attached() {
3065 0 : result.shards.push(TenantShardLocation {
3066 0 : shard_id: tenant_shard_id,
3067 0 : node_id: *node_id,
3068 0 : })
3069 0 : }
3070 : }
3071 : }
3072 0 : waiters
3073 : }
3074 : };
3075 :
3076 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3077 : // Do not treat a reconcile error as fatal: we have already applied any requested
3078 : // Intent changes, and the reconcile can fail for external reasons like unavailable
3079 : // compute notification API. In these cases, it is important that we do not
3080 : // cause the cloud control plane to retry forever on this API.
3081 0 : tracing::warn!(
3082 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
3083 : );
3084 0 : }
3085 :
3086 : // Logging the full result is useful because it lets us cross-check what the cloud control
3087 : // plane's tenant_shards table should contain.
3088 0 : tracing::info!("Complete, returning {result:?}");
3089 :
3090 0 : Ok(result)
3091 0 : }
3092 :
3093 0 : pub(crate) async fn tenant_config_patch(
3094 0 : &self,
3095 0 : req: TenantConfigPatchRequest,
3096 0 : ) -> Result<(), ApiError> {
3097 0 : let _tenant_lock = trace_exclusive_lock(
3098 0 : &self.tenant_op_locks,
3099 0 : req.tenant_id,
3100 0 : TenantOperations::ConfigPatch,
3101 0 : )
3102 0 : .await;
3103 :
3104 0 : let tenant_id = req.tenant_id;
3105 0 : let patch = req.config;
3106 0 :
3107 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3108 :
3109 0 : let base = {
3110 0 : let locked = self.inner.read().unwrap();
3111 0 : let shards = locked
3112 0 : .tenants
3113 0 : .range(TenantShardId::tenant_range(req.tenant_id));
3114 0 :
3115 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
3116 :
3117 0 : let first = match configs.peek() {
3118 0 : Some(first) => (*first).clone(),
3119 : None => {
3120 0 : return Err(ApiError::NotFound(
3121 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3122 0 : ));
3123 : }
3124 : };
3125 :
3126 0 : if !configs.all_equal() {
3127 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3128 : // This can't happen because we atomically update the database records
3129 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3130 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3131 0 : "Tenant configs for {} are mismatched",
3132 0 : req.tenant_id
3133 0 : )));
3134 0 : }
3135 0 :
3136 0 : first
3137 : };
3138 :
3139 0 : let updated_config = base
3140 0 : .apply_patch(patch)
3141 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3142 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3143 0 : .await
3144 0 : }
3145 :
3146 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3147 : // We require an exclusive lock, because we are updating persistent and in-memory state
3148 0 : let _tenant_lock = trace_exclusive_lock(
3149 0 : &self.tenant_op_locks,
3150 0 : req.tenant_id,
3151 0 : TenantOperations::ConfigSet,
3152 0 : )
3153 0 : .await;
3154 :
3155 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3156 :
3157 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3158 0 : .await
3159 0 : }
3160 :
3161 0 : async fn set_tenant_config_and_reconcile(
3162 0 : &self,
3163 0 : tenant_id: TenantId,
3164 0 : config: TenantConfig,
3165 0 : ) -> Result<(), ApiError> {
3166 0 : self.persistence
3167 0 : .update_tenant_shard(
3168 0 : TenantFilter::Tenant(tenant_id),
3169 0 : None,
3170 0 : Some(config.clone()),
3171 0 : None,
3172 0 : None,
3173 0 : )
3174 0 : .await?;
3175 :
3176 0 : let waiters = {
3177 0 : let mut waiters = Vec::new();
3178 0 : let mut locked = self.inner.write().unwrap();
3179 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3180 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3181 0 : shard.config = config.clone();
3182 0 : if let Some(waiter) =
3183 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3184 0 : {
3185 0 : waiters.push(waiter);
3186 0 : }
3187 : }
3188 0 : waiters
3189 : };
3190 :
3191 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3192 : // Treat this as success because we have stored the configuration. If e.g.
3193 : // a node was unavailable at this time, it should not stop us accepting a
3194 : // configuration change.
3195 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3196 0 : }
3197 :
3198 0 : Ok(())
3199 0 : }
3200 :
3201 0 : pub(crate) fn tenant_config_get(
3202 0 : &self,
3203 0 : tenant_id: TenantId,
3204 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3205 0 : let config = {
3206 0 : let locked = self.inner.read().unwrap();
3207 0 :
3208 0 : match locked
3209 0 : .tenants
3210 0 : .range(TenantShardId::tenant_range(tenant_id))
3211 0 : .next()
3212 : {
3213 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3214 : None => {
3215 0 : return Err(ApiError::NotFound(
3216 0 : anyhow::anyhow!("Tenant not found").into(),
3217 0 : ));
3218 : }
3219 : }
3220 : };
3221 :
3222 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3223 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3224 : // and `effective_config` in the response is meaningless, but we retain that syntax
3225 : // in order to remain compatible with the pageserver API.
3226 :
3227 0 : let response = HashMap::from([
3228 : (
3229 : "tenant_specific_overrides",
3230 0 : serde_json::to_value(&config)
3231 0 : .context("serializing tenant specific overrides")
3232 0 : .map_err(ApiError::InternalServerError)?,
3233 : ),
3234 : (
3235 0 : "effective_config",
3236 0 : serde_json::to_value(&config)
3237 0 : .context("serializing effective config")
3238 0 : .map_err(ApiError::InternalServerError)?,
3239 : ),
3240 : ]);
3241 :
3242 0 : Ok(response)
3243 0 : }
3244 :
3245 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3246 0 : &self,
3247 0 : time_travel_req: &TenantTimeTravelRequest,
3248 0 : tenant_id: TenantId,
3249 0 : timestamp: Cow<'_, str>,
3250 0 : done_if_after: Cow<'_, str>,
3251 0 : ) -> Result<(), ApiError> {
3252 0 : let _tenant_lock = trace_exclusive_lock(
3253 0 : &self.tenant_op_locks,
3254 0 : tenant_id,
3255 0 : TenantOperations::TimeTravelRemoteStorage,
3256 0 : )
3257 0 : .await;
3258 :
3259 0 : let node = {
3260 0 : let mut locked = self.inner.write().unwrap();
3261 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3262 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3263 : // but only at the start of the process, so it's really just to prevent operator
3264 : // mistakes.
3265 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3266 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3267 : {
3268 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3269 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3270 0 : )));
3271 0 : }
3272 0 : let maybe_attached = shard
3273 0 : .observed
3274 0 : .locations
3275 0 : .iter()
3276 0 : .filter_map(|(node_id, observed_location)| {
3277 0 : observed_location
3278 0 : .conf
3279 0 : .as_ref()
3280 0 : .map(|loc| (node_id, observed_location, loc.mode))
3281 0 : })
3282 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3283 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3284 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3285 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3286 0 : )));
3287 0 : }
3288 : }
3289 0 : let scheduler = &mut locked.scheduler;
3290 : // Right now we only perform the operation on a single node without parallelization
3291 : // TODO fan out the operation to multiple nodes for better performance
3292 0 : let node_id = scheduler.any_available_node()?;
3293 0 : let node = locked
3294 0 : .nodes
3295 0 : .get(&node_id)
3296 0 : .expect("Pageservers may not be deleted while lock is active");
3297 0 : node.clone()
3298 0 : };
3299 0 :
3300 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3301 0 : let mut counts = time_travel_req
3302 0 : .shard_counts
3303 0 : .iter()
3304 0 : .copied()
3305 0 : .collect::<HashSet<_>>()
3306 0 : .into_iter()
3307 0 : .collect::<Vec<_>>();
3308 0 : counts.sort_unstable();
3309 :
3310 0 : for count in counts {
3311 0 : let shard_ids = (0..count.count())
3312 0 : .map(|i| TenantShardId {
3313 0 : tenant_id,
3314 0 : shard_number: ShardNumber(i),
3315 0 : shard_count: count,
3316 0 : })
3317 0 : .collect::<Vec<_>>();
3318 0 : for tenant_shard_id in shard_ids {
3319 0 : let client = PageserverClient::new(
3320 0 : node.get_id(),
3321 0 : self.http_client.clone(),
3322 0 : node.base_url(),
3323 0 : self.config.pageserver_jwt_token.as_deref(),
3324 0 : );
3325 0 :
3326 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3327 :
3328 0 : client
3329 0 : .tenant_time_travel_remote_storage(
3330 0 : tenant_shard_id,
3331 0 : ×tamp,
3332 0 : &done_if_after,
3333 0 : )
3334 0 : .await
3335 0 : .map_err(|e| {
3336 0 : ApiError::InternalServerError(anyhow::anyhow!(
3337 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3338 0 : node
3339 0 : ))
3340 0 : })?;
3341 : }
3342 : }
3343 0 : Ok(())
3344 0 : }
3345 :
3346 0 : pub(crate) async fn tenant_secondary_download(
3347 0 : &self,
3348 0 : tenant_id: TenantId,
3349 0 : wait: Option<Duration>,
3350 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3351 0 : let _tenant_lock = trace_shared_lock(
3352 0 : &self.tenant_op_locks,
3353 0 : tenant_id,
3354 0 : TenantOperations::SecondaryDownload,
3355 0 : )
3356 0 : .await;
3357 :
3358 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3359 0 : let targets = {
3360 0 : let locked = self.inner.read().unwrap();
3361 0 : let mut targets = Vec::new();
3362 :
3363 0 : for (tenant_shard_id, shard) in
3364 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3365 : {
3366 0 : for node_id in shard.intent.get_secondary() {
3367 0 : let node = locked
3368 0 : .nodes
3369 0 : .get(node_id)
3370 0 : .expect("Pageservers may not be deleted while referenced");
3371 0 :
3372 0 : targets.push((*tenant_shard_id, node.clone()));
3373 0 : }
3374 : }
3375 0 : targets
3376 0 : };
3377 0 :
3378 0 : // Issue concurrent requests to all shards' locations
3379 0 : let mut futs = FuturesUnordered::new();
3380 0 : for (tenant_shard_id, node) in targets {
3381 0 : let client = PageserverClient::new(
3382 0 : node.get_id(),
3383 0 : self.http_client.clone(),
3384 0 : node.base_url(),
3385 0 : self.config.pageserver_jwt_token.as_deref(),
3386 0 : );
3387 0 : futs.push(async move {
3388 0 : let result = client
3389 0 : .tenant_secondary_download(tenant_shard_id, wait)
3390 0 : .await;
3391 0 : (result, node, tenant_shard_id)
3392 0 : })
3393 : }
3394 :
3395 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3396 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3397 : // well as more general cases like 503s, 500s, or timeouts.
3398 0 : let mut aggregate_progress = SecondaryProgress::default();
3399 0 : let mut aggregate_status: Option<StatusCode> = None;
3400 0 : let mut error: Option<mgmt_api::Error> = None;
3401 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3402 0 : match result {
3403 0 : Err(e) => {
3404 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3405 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3406 0 : // than they had hoped for.
3407 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3408 0 : error = Some(e)
3409 : }
3410 0 : Ok((status_code, progress)) => {
3411 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3412 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3413 0 : aggregate_progress.layers_total += progress.layers_total;
3414 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3415 0 : aggregate_progress.bytes_total += progress.bytes_total;
3416 0 : aggregate_progress.heatmap_mtime =
3417 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3418 0 : aggregate_status = match aggregate_status {
3419 0 : None => Some(status_code),
3420 0 : Some(StatusCode::OK) => Some(status_code),
3421 0 : Some(cur) => {
3422 0 : // Other status codes (e.g. 202) -- do not overwrite.
3423 0 : Some(cur)
3424 : }
3425 : };
3426 : }
3427 : }
3428 : }
3429 :
3430 : // If any of the shards return 202, indicate our result as 202.
3431 0 : match aggregate_status {
3432 : None => {
3433 0 : match error {
3434 0 : Some(e) => {
3435 0 : // No successes, and an error: surface it
3436 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3437 : }
3438 : None => {
3439 : // No shards found
3440 0 : Err(ApiError::NotFound(
3441 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3442 0 : ))
3443 : }
3444 : }
3445 : }
3446 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3447 : }
3448 0 : }
3449 :
3450 0 : pub(crate) async fn tenant_delete(
3451 0 : self: &Arc<Self>,
3452 0 : tenant_id: TenantId,
3453 0 : ) -> Result<StatusCode, ApiError> {
3454 0 : let _tenant_lock =
3455 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3456 :
3457 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3458 :
3459 : // Detach all shards. This also deletes local pageserver shard data.
3460 0 : let (detach_waiters, node) = {
3461 0 : let mut detach_waiters = Vec::new();
3462 0 : let mut locked = self.inner.write().unwrap();
3463 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3464 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3465 : // Update the tenant's intent to remove all attachments
3466 0 : shard.policy = PlacementPolicy::Detached;
3467 0 : shard
3468 0 : .schedule(scheduler, &mut ScheduleContext::default())
3469 0 : .expect("De-scheduling is infallible");
3470 0 : debug_assert!(shard.intent.get_attached().is_none());
3471 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3472 :
3473 0 : if let Some(waiter) =
3474 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3475 0 : {
3476 0 : detach_waiters.push(waiter);
3477 0 : }
3478 : }
3479 :
3480 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3481 : // was attached, just has to be able to see the S3 content)
3482 0 : let node_id = scheduler.any_available_node()?;
3483 0 : let node = nodes
3484 0 : .get(&node_id)
3485 0 : .expect("Pageservers may not be deleted while lock is active");
3486 0 : (detach_waiters, node.clone())
3487 0 : };
3488 0 :
3489 0 : // This reconcile wait can fail in a few ways:
3490 0 : // A there is a very long queue for the reconciler semaphore
3491 0 : // B some pageserver is failing to handle a detach promptly
3492 0 : // C some pageserver goes offline right at the moment we send it a request.
3493 0 : //
3494 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3495 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3496 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3497 0 : // deleting the underlying data).
3498 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3499 0 : .await?;
3500 :
3501 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3502 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3503 : // the tenant ID prefix, including all shards (even possibly stale ones).
3504 0 : match node
3505 0 : .with_client_retries(
3506 0 : |client| async move {
3507 0 : client
3508 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3509 0 : .await
3510 0 : },
3511 0 : &self.http_client,
3512 0 : &self.config.pageserver_jwt_token,
3513 0 : 1,
3514 0 : 3,
3515 0 : RECONCILE_TIMEOUT,
3516 0 : &self.cancel,
3517 0 : )
3518 0 : .await
3519 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3520 : {
3521 0 : Ok(_) => {}
3522 : Err(mgmt_api::Error::Cancelled) => {
3523 0 : return Err(ApiError::ShuttingDown);
3524 : }
3525 0 : Err(e) => {
3526 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3527 0 : // at large is unavailable.
3528 0 : tracing::error!("Error deleting via node {node}: {e}");
3529 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3530 : }
3531 : }
3532 :
3533 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3534 : // our in-memory state and database state.
3535 :
3536 : // Ordering: we delete persistent state first: if we then
3537 : // crash, we will drop the in-memory state.
3538 :
3539 : // Drop persistent state.
3540 0 : self.persistence.delete_tenant(tenant_id).await?;
3541 :
3542 : // Drop in-memory state
3543 : {
3544 0 : let mut locked = self.inner.write().unwrap();
3545 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3546 :
3547 : // Dereference Scheduler from shards before dropping them
3548 0 : for (_tenant_shard_id, shard) in
3549 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3550 0 : {
3551 0 : shard.intent.clear(scheduler);
3552 0 : }
3553 :
3554 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3555 0 : tracing::info!(
3556 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3557 0 : locked.tenants.len()
3558 : );
3559 : };
3560 :
3561 : // Delete the tenant from safekeepers (if needed)
3562 0 : self.tenant_delete_safekeepers(tenant_id)
3563 0 : .instrument(tracing::info_span!("tenant_delete_safekeepers", %tenant_id))
3564 0 : .await?;
3565 :
3566 : // Success is represented as 404, to imitate the existing pageserver deletion API
3567 0 : Ok(StatusCode::NOT_FOUND)
3568 0 : }
3569 :
3570 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3571 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3572 : /// the tenant's policies (configuration) within the storage controller
3573 0 : pub(crate) async fn tenant_update_policy(
3574 0 : &self,
3575 0 : tenant_id: TenantId,
3576 0 : req: TenantPolicyRequest,
3577 0 : ) -> Result<(), ApiError> {
3578 : // We require an exclusive lock, because we are updating persistent and in-memory state
3579 0 : let _tenant_lock = trace_exclusive_lock(
3580 0 : &self.tenant_op_locks,
3581 0 : tenant_id,
3582 0 : TenantOperations::UpdatePolicy,
3583 0 : )
3584 0 : .await;
3585 :
3586 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3587 :
3588 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3589 :
3590 : let TenantPolicyRequest {
3591 0 : placement,
3592 0 : mut scheduling,
3593 0 : } = req;
3594 :
3595 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3596 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3597 : // this to take effect.
3598 0 : if scheduling.is_none() {
3599 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3600 0 : }
3601 0 : }
3602 :
3603 0 : self.persistence
3604 0 : .update_tenant_shard(
3605 0 : TenantFilter::Tenant(tenant_id),
3606 0 : placement.clone(),
3607 0 : None,
3608 0 : None,
3609 0 : scheduling,
3610 0 : )
3611 0 : .await?;
3612 :
3613 0 : let mut schedule_context = ScheduleContext::default();
3614 0 : let mut locked = self.inner.write().unwrap();
3615 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3616 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3617 0 : if let Some(placement) = &placement {
3618 0 : shard.policy = placement.clone();
3619 0 :
3620 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3621 0 : "Updated placement policy to {placement:?}");
3622 0 : }
3623 :
3624 0 : if let Some(scheduling) = &scheduling {
3625 0 : shard.set_scheduling_policy(*scheduling);
3626 0 :
3627 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3628 0 : "Updated scheduling policy to {scheduling:?}");
3629 0 : }
3630 :
3631 : // In case scheduling is being switched back on, try it now.
3632 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3633 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3634 : }
3635 :
3636 0 : Ok(())
3637 0 : }
3638 :
3639 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3640 0 : &self,
3641 0 : tenant_id: TenantId,
3642 0 : mut create_req: TimelineCreateRequest,
3643 0 : ) -> Result<TimelineInfo, ApiError> {
3644 0 : tracing::info!(
3645 0 : "Creating timeline {}/{}",
3646 : tenant_id,
3647 : create_req.new_timeline_id,
3648 : );
3649 :
3650 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3651 0 : if targets.0.is_empty() {
3652 0 : return Err(ApiError::NotFound(
3653 0 : anyhow::anyhow!("Tenant not found").into(),
3654 0 : ));
3655 0 : };
3656 0 :
3657 0 : let (shard_zero_tid, shard_zero_locations) =
3658 0 : targets.0.pop_first().expect("Must have at least one shard");
3659 0 : assert!(shard_zero_tid.is_shard_zero());
3660 :
3661 0 : async fn create_one(
3662 0 : tenant_shard_id: TenantShardId,
3663 0 : locations: ShardMutationLocations,
3664 0 : http_client: reqwest::Client,
3665 0 : jwt: Option<String>,
3666 0 : mut create_req: TimelineCreateRequest,
3667 0 : ) -> Result<TimelineInfo, ApiError> {
3668 0 : let latest = locations.latest.node;
3669 0 :
3670 0 : tracing::info!(
3671 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3672 : tenant_shard_id,
3673 : create_req.new_timeline_id,
3674 : locations.latest.generation
3675 : );
3676 :
3677 0 : let client =
3678 0 : PageserverClient::new(latest.get_id(), http_client.clone(), latest.base_url(), jwt.as_deref());
3679 :
3680 0 : let timeline_info = client
3681 0 : .timeline_create(tenant_shard_id, &create_req)
3682 0 : .await
3683 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3684 :
3685 : // If we are going to create the timeline on some stale locations for shard 0, then ask them to re-use
3686 : // the initdb generated by the latest location, rather than generating their own. This avoids racing uploads
3687 : // of initdb to S3 which might not be binary-identical if different pageservers have different postgres binaries.
3688 0 : if tenant_shard_id.is_shard_zero() {
3689 0 : if let models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } = &mut create_req.mode {
3690 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id);
3691 0 : }
3692 0 : }
3693 :
3694 : // We propagate timeline creations to all attached locations such that a compute
3695 : // for the new timeline is able to start regardless of the current state of the
3696 : // tenant shard reconciliation.
3697 0 : for location in locations.other {
3698 0 : tracing::info!(
3699 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3700 : tenant_shard_id,
3701 : create_req.new_timeline_id,
3702 : location.node,
3703 : location.generation
3704 : );
3705 :
3706 0 : let client = PageserverClient::new(
3707 0 : location.node.get_id(),
3708 0 : http_client.clone(),
3709 0 : location.node.base_url(),
3710 0 : jwt.as_deref(),
3711 0 : );
3712 :
3713 0 : let res = client
3714 0 : .timeline_create(tenant_shard_id, &create_req)
3715 0 : .await;
3716 :
3717 0 : if let Err(e) = res {
3718 0 : match e {
3719 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3720 0 : // Tenant might have been detached from the stale location,
3721 0 : // so ignore 404s.
3722 0 : },
3723 : _ => {
3724 0 : return Err(passthrough_api_error(&location.node, e));
3725 : }
3726 : }
3727 0 : }
3728 : }
3729 :
3730 0 : Ok(timeline_info)
3731 0 : }
3732 :
3733 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3734 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3735 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3736 : //
3737 : // This also enables non-zero shards to use the initdb that shard 0 generated and uploaded to S3, rather than
3738 : // independently generating their own initdb. This guarantees that shards cannot end up with different initial
3739 : // states if e.g. they have different postgres binary versions.
3740 0 : let timeline_info = create_one(
3741 0 : shard_zero_tid,
3742 0 : shard_zero_locations,
3743 0 : self.http_client.clone(),
3744 0 : self.config.pageserver_jwt_token.clone(),
3745 0 : create_req.clone(),
3746 0 : )
3747 0 : .await?;
3748 :
3749 : // Update the create request for shards >= 0
3750 0 : match &mut create_req.mode {
3751 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3752 0 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3753 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3754 0 : },
3755 0 : models::TimelineCreateRequestMode::Bootstrap { existing_initdb_timeline_id, .. } => {
3756 0 : // For shards >= 0, do not run initdb: use the one that shard 0 uploaded to S3
3757 0 : *existing_initdb_timeline_id = Some(create_req.new_timeline_id)
3758 : }
3759 0 : _ => {}
3760 : }
3761 :
3762 : // Create timeline on remaining shards with number >0
3763 0 : if !targets.0.is_empty() {
3764 : // If we had multiple shards, issue requests for the remainder now.
3765 0 : let jwt = &self.config.pageserver_jwt_token;
3766 0 : self.tenant_for_shards(
3767 0 : targets
3768 0 : .0
3769 0 : .iter()
3770 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3771 0 : .collect(),
3772 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3773 0 : let create_req = create_req.clone();
3774 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3775 0 : Box::pin(create_one(
3776 0 : tenant_shard_id,
3777 0 : mutation_locations,
3778 0 : self.http_client.clone(),
3779 0 : jwt.clone(),
3780 0 : create_req,
3781 0 : ))
3782 0 : },
3783 0 : )
3784 0 : .await?;
3785 0 : }
3786 :
3787 0 : Ok(timeline_info)
3788 0 : })
3789 0 : .await?
3790 0 : }
3791 :
3792 0 : pub(crate) async fn tenant_timeline_create(
3793 0 : self: &Arc<Self>,
3794 0 : tenant_id: TenantId,
3795 0 : create_req: TimelineCreateRequest,
3796 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
3797 0 : let safekeepers = self.config.timelines_onto_safekeepers;
3798 0 : let timeline_id = create_req.new_timeline_id;
3799 0 :
3800 0 : tracing::info!(
3801 0 : mode=%create_req.mode_tag(),
3802 0 : %safekeepers,
3803 0 : "Creating timeline {}/{}",
3804 : tenant_id,
3805 : timeline_id,
3806 : );
3807 :
3808 0 : let _tenant_lock = trace_shared_lock(
3809 0 : &self.tenant_op_locks,
3810 0 : tenant_id,
3811 0 : TenantOperations::TimelineCreate,
3812 0 : )
3813 0 : .await;
3814 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3815 0 : let is_import = create_req.is_import();
3816 0 :
3817 0 : if is_import {
3818 : // Ensure that there is no split on-going.
3819 : // [`Self::tenant_shard_split`] holds the exclusive tenant lock
3820 : // for the duration of the split, but here we handle the case
3821 : // where we restarted and the split is being aborted.
3822 0 : let locked = self.inner.read().unwrap();
3823 0 : let splitting = locked
3824 0 : .tenants
3825 0 : .range(TenantShardId::tenant_range(tenant_id))
3826 0 : .any(|(_id, shard)| shard.splitting != SplitState::Idle);
3827 0 :
3828 0 : if splitting {
3829 0 : return Err(ApiError::Conflict("Tenant is splitting shard".to_string()));
3830 0 : }
3831 0 : }
3832 :
3833 0 : let timeline_info = self
3834 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
3835 0 : .await?;
3836 :
3837 0 : let selected_safekeepers = if is_import {
3838 0 : let shards = {
3839 0 : let locked = self.inner.read().unwrap();
3840 0 : locked
3841 0 : .tenants
3842 0 : .range(TenantShardId::tenant_range(tenant_id))
3843 0 : .map(|(ts_id, _)| ts_id.to_index())
3844 0 : .collect::<Vec<_>>()
3845 0 : };
3846 0 :
3847 0 : if !shards
3848 0 : .iter()
3849 0 : .map(|shard_index| shard_index.shard_count)
3850 0 : .all_equal()
3851 : {
3852 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3853 0 : "Inconsistent shard count"
3854 0 : )));
3855 0 : }
3856 0 :
3857 0 : let import = TimelineImport {
3858 0 : tenant_id,
3859 0 : timeline_id,
3860 0 : shard_statuses: ShardImportStatuses::new(shards),
3861 0 : };
3862 :
3863 0 : let inserted = self
3864 0 : .persistence
3865 0 : .insert_timeline_import(import.to_persistent())
3866 0 : .await
3867 0 : .context("timeline import insert")
3868 0 : .map_err(ApiError::InternalServerError)?;
3869 :
3870 : // Set the importing flag on the tenant shards
3871 0 : self.inner
3872 0 : .write()
3873 0 : .unwrap()
3874 0 : .tenants
3875 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
3876 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Importing);
3877 0 :
3878 0 : match inserted {
3879 : true => {
3880 0 : tracing::info!(%tenant_id, %timeline_id, "Inserted timeline import");
3881 : }
3882 : false => {
3883 0 : tracing::info!(%tenant_id, %timeline_id, "Timeline import entry already present");
3884 : }
3885 : }
3886 :
3887 0 : None
3888 0 : } else if safekeepers {
3889 : // Note that we do not support creating the timeline on the safekeepers
3890 : // for imported timelines. The `start_lsn` of the timeline is not known
3891 : // until the import finshes.
3892 : // https://github.com/neondatabase/neon/issues/11569
3893 0 : let res = self
3894 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info)
3895 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
3896 0 : .await?;
3897 0 : Some(res)
3898 : } else {
3899 0 : None
3900 : };
3901 :
3902 0 : Ok(TimelineCreateResponseStorcon {
3903 0 : timeline_info,
3904 0 : safekeepers: selected_safekeepers,
3905 0 : })
3906 0 : }
3907 :
3908 0 : pub(crate) async fn handle_timeline_shard_import_progress_upcall(
3909 0 : self: &Arc<Self>,
3910 0 : req: PutTimelineImportStatusRequest,
3911 0 : ) -> Result<(), ApiError> {
3912 0 : let res = self
3913 0 : .persistence
3914 0 : .update_timeline_import(req.tenant_shard_id, req.timeline_id, req.status)
3915 0 : .await;
3916 0 : let timeline_import = match res {
3917 0 : Ok(Ok(Some(timeline_import))) => timeline_import,
3918 : Ok(Ok(None)) => {
3919 : // Idempotency: we've already seen and handled this update.
3920 0 : return Ok(());
3921 : }
3922 0 : Ok(Err(logical_err)) => {
3923 0 : return Err(logical_err.into());
3924 : }
3925 0 : Err(db_err) => {
3926 0 : return Err(db_err.into());
3927 : }
3928 : };
3929 :
3930 0 : tracing::info!(
3931 : tenant_id=%req.tenant_shard_id.tenant_id,
3932 : timeline_id=%req.timeline_id,
3933 0 : shard_id=%req.tenant_shard_id.shard_slug(),
3934 0 : "Updated timeline import status to: {timeline_import:?}");
3935 :
3936 0 : if timeline_import.is_complete() {
3937 0 : tokio::task::spawn({
3938 0 : let this = self.clone();
3939 0 : async move { this.finalize_timeline_import(timeline_import).await }
3940 0 : });
3941 0 : }
3942 :
3943 0 : Ok(())
3944 0 : }
3945 :
3946 : #[instrument(skip_all, fields(
3947 : tenant_id=%import.tenant_id,
3948 : shard_id=%import.timeline_id,
3949 : ))]
3950 : async fn finalize_timeline_import(
3951 : self: &Arc<Self>,
3952 : import: TimelineImport,
3953 : ) -> anyhow::Result<()> {
3954 : tracing::info!("Finalizing timeline import");
3955 :
3956 : pausable_failpoint!("timeline-import-pre-cplane-notification");
3957 :
3958 : let import_failed = import.completion_error().is_some();
3959 :
3960 : if !import_failed {
3961 : loop {
3962 : if self.cancel.is_cancelled() {
3963 : anyhow::bail!("Shut down requested while finalizing import");
3964 : }
3965 :
3966 : let active = self.timeline_active_on_all_shards(&import).await?;
3967 :
3968 : match active {
3969 : true => {
3970 : tracing::info!("Timeline became active on all shards");
3971 : break;
3972 : }
3973 : false => {
3974 : tracing::info!("Timeline not active on all shards yet");
3975 :
3976 : tokio::select! {
3977 : _ = self.cancel.cancelled() => {
3978 : anyhow::bail!("Shut down requested while finalizing import");
3979 : },
3980 : _ = tokio::time::sleep(Duration::from_secs(5)) => {}
3981 : };
3982 : }
3983 : }
3984 : }
3985 : }
3986 :
3987 : tracing::info!(%import_failed, "Notifying cplane of import completion");
3988 :
3989 : let client = UpcallClient::new(self.get_config(), self.cancel.child_token());
3990 : client.notify_import_complete(&import).await?;
3991 :
3992 : if let Err(err) = self
3993 : .persistence
3994 : .delete_timeline_import(import.tenant_id, import.timeline_id)
3995 : .await
3996 : {
3997 : tracing::warn!("Failed to delete timeline import entry from database: {err}");
3998 : }
3999 :
4000 : self.inner
4001 : .write()
4002 : .unwrap()
4003 : .tenants
4004 : .range_mut(TenantShardId::tenant_range(import.tenant_id))
4005 0 : .for_each(|(_id, shard)| shard.importing = TimelineImportState::Idle);
4006 :
4007 : // TODO(vlad): Timeline creations in import mode do not return a correct initdb lsn,
4008 : // so we can't create the timeline on the safekeepers. Fix by moving creation here.
4009 : // https://github.com/neondatabase/neon/issues/11569
4010 : tracing::info!(%import_failed, "Timeline import complete");
4011 :
4012 : Ok(())
4013 : }
4014 :
4015 0 : async fn finalize_timeline_imports(self: &Arc<Self>, imports: Vec<TimelineImport>) {
4016 0 : futures::future::join_all(
4017 0 : imports
4018 0 : .into_iter()
4019 0 : .map(|import| self.finalize_timeline_import(import)),
4020 0 : )
4021 0 : .await;
4022 0 : }
4023 :
4024 0 : async fn timeline_active_on_all_shards(
4025 0 : self: &Arc<Self>,
4026 0 : import: &TimelineImport,
4027 0 : ) -> anyhow::Result<bool> {
4028 0 : let targets = {
4029 0 : let locked = self.inner.read().unwrap();
4030 0 : let mut targets = Vec::new();
4031 :
4032 0 : for (tenant_shard_id, shard) in locked
4033 0 : .tenants
4034 0 : .range(TenantShardId::tenant_range(import.tenant_id))
4035 : {
4036 0 : if !import
4037 0 : .shard_statuses
4038 0 : .0
4039 0 : .contains_key(&tenant_shard_id.to_index())
4040 : {
4041 0 : anyhow::bail!("Shard layout change detected on completion");
4042 0 : }
4043 :
4044 0 : if let Some(node_id) = shard.intent.get_attached() {
4045 0 : let node = locked
4046 0 : .nodes
4047 0 : .get(node_id)
4048 0 : .expect("Pageservers may not be deleted while referenced");
4049 0 : targets.push((*tenant_shard_id, node.clone()));
4050 0 : } else {
4051 0 : return Ok(false);
4052 : }
4053 : }
4054 :
4055 0 : targets
4056 : };
4057 :
4058 0 : let results = self
4059 0 : .tenant_for_shards_api(
4060 0 : targets,
4061 0 : |tenant_shard_id, client| async move {
4062 0 : client
4063 0 : .timeline_detail(tenant_shard_id, import.timeline_id)
4064 0 : .await
4065 0 : },
4066 0 : 1,
4067 0 : 1,
4068 0 : SHORT_RECONCILE_TIMEOUT,
4069 0 : &self.cancel,
4070 0 : )
4071 0 : .await;
4072 :
4073 0 : Ok(results.into_iter().all(|res| match res {
4074 0 : Ok(info) => info.state == TimelineState::Active,
4075 0 : Err(_) => false,
4076 0 : }))
4077 0 : }
4078 :
4079 0 : pub(crate) async fn tenant_timeline_archival_config(
4080 0 : &self,
4081 0 : tenant_id: TenantId,
4082 0 : timeline_id: TimelineId,
4083 0 : req: TimelineArchivalConfigRequest,
4084 0 : ) -> Result<(), ApiError> {
4085 0 : tracing::info!(
4086 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
4087 : req.state
4088 : );
4089 :
4090 0 : let _tenant_lock = trace_shared_lock(
4091 0 : &self.tenant_op_locks,
4092 0 : tenant_id,
4093 0 : TenantOperations::TimelineArchivalConfig,
4094 0 : )
4095 0 : .await;
4096 :
4097 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4098 0 : if targets.0.is_empty() {
4099 0 : return Err(ApiError::NotFound(
4100 0 : anyhow::anyhow!("Tenant not found").into(),
4101 0 : ));
4102 0 : }
4103 0 : async fn config_one(
4104 0 : tenant_shard_id: TenantShardId,
4105 0 : timeline_id: TimelineId,
4106 0 : node: Node,
4107 0 : http_client: reqwest::Client,
4108 0 : jwt: Option<String>,
4109 0 : req: TimelineArchivalConfigRequest,
4110 0 : ) -> Result<(), ApiError> {
4111 0 : tracing::info!(
4112 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4113 : );
4114 :
4115 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4116 0 :
4117 0 : client
4118 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
4119 0 : .await
4120 0 : .map_err(|e| match e {
4121 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
4122 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
4123 : }
4124 0 : _ => passthrough_api_error(&node, e),
4125 0 : })
4126 0 : }
4127 :
4128 : // no shard needs to go first/last; the operation should be idempotent
4129 : // TODO: it would be great to ensure that all shards return the same error
4130 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4131 0 : let results = self
4132 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4133 0 : futures::FutureExt::boxed(config_one(
4134 0 : tenant_shard_id,
4135 0 : timeline_id,
4136 0 : node,
4137 0 : self.http_client.clone(),
4138 0 : self.config.pageserver_jwt_token.clone(),
4139 0 : req.clone(),
4140 0 : ))
4141 0 : })
4142 0 : .await?;
4143 0 : assert!(!results.is_empty(), "must have at least one result");
4144 :
4145 0 : Ok(())
4146 0 : }).await?
4147 0 : }
4148 :
4149 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
4150 0 : &self,
4151 0 : tenant_id: TenantId,
4152 0 : timeline_id: TimelineId,
4153 0 : behavior: Option<DetachBehavior>,
4154 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
4155 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
4156 :
4157 0 : let _tenant_lock = trace_shared_lock(
4158 0 : &self.tenant_op_locks,
4159 0 : tenant_id,
4160 0 : TenantOperations::TimelineDetachAncestor,
4161 0 : )
4162 0 : .await;
4163 :
4164 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4165 0 : if targets.0.is_empty() {
4166 0 : return Err(ApiError::NotFound(
4167 0 : anyhow::anyhow!("Tenant not found").into(),
4168 0 : ));
4169 0 : }
4170 :
4171 0 : async fn detach_one(
4172 0 : tenant_shard_id: TenantShardId,
4173 0 : timeline_id: TimelineId,
4174 0 : node: Node,
4175 0 : http_client: reqwest::Client,
4176 0 : jwt: Option<String>,
4177 0 : behavior: Option<DetachBehavior>,
4178 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
4179 0 : tracing::info!(
4180 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4181 : );
4182 :
4183 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4184 0 :
4185 0 : client
4186 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id, behavior)
4187 0 : .await
4188 0 : .map_err(|e| {
4189 : use mgmt_api::Error;
4190 :
4191 0 : match e {
4192 : // no ancestor (ever)
4193 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
4194 0 : "{node}: {}",
4195 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
4196 0 : )),
4197 : // too many ancestors
4198 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
4199 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
4200 : }
4201 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
4202 0 : // avoid turning these into conflicts to remain compatible with
4203 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
4204 0 : // detach
4205 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
4206 : }
4207 : // rest can be mapped as usual
4208 0 : other => passthrough_api_error(&node, other),
4209 : }
4210 0 : })
4211 0 : .map(|res| (tenant_shard_id.shard_number, res))
4212 0 : }
4213 :
4214 : // no shard needs to go first/last; the operation should be idempotent
4215 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4216 0 : let mut results = self
4217 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4218 0 : futures::FutureExt::boxed(detach_one(
4219 0 : tenant_shard_id,
4220 0 : timeline_id,
4221 0 : node,
4222 0 : self.http_client.clone(),
4223 0 : self.config.pageserver_jwt_token.clone(),
4224 0 : behavior,
4225 0 : ))
4226 0 : })
4227 0 : .await?;
4228 :
4229 0 : let any = results.pop().expect("we must have at least one response");
4230 0 :
4231 0 : let mismatching = results
4232 0 : .iter()
4233 0 : .filter(|(_, res)| res != &any.1)
4234 0 : .collect::<Vec<_>>();
4235 0 : if !mismatching.is_empty() {
4236 : // this can be hit by races which should not happen because operation lock on cplane
4237 0 : let matching = results.len() - mismatching.len();
4238 0 : tracing::error!(
4239 : matching,
4240 : compared_against=?any,
4241 : ?mismatching,
4242 0 : "shards returned different results"
4243 : );
4244 :
4245 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
4246 0 : }
4247 0 :
4248 0 : Ok(any.1)
4249 0 : }).await?
4250 0 : }
4251 :
4252 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
4253 0 : &self,
4254 0 : tenant_id: TenantId,
4255 0 : timeline_id: TimelineId,
4256 0 : dir: BlockUnblock,
4257 0 : ) -> Result<(), ApiError> {
4258 0 : let _tenant_lock = trace_shared_lock(
4259 0 : &self.tenant_op_locks,
4260 0 : tenant_id,
4261 0 : TenantOperations::TimelineGcBlockUnblock,
4262 0 : )
4263 0 : .await;
4264 :
4265 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4266 0 : if targets.0.is_empty() {
4267 0 : return Err(ApiError::NotFound(
4268 0 : anyhow::anyhow!("Tenant not found").into(),
4269 0 : ));
4270 0 : }
4271 :
4272 0 : async fn do_one(
4273 0 : tenant_shard_id: TenantShardId,
4274 0 : timeline_id: TimelineId,
4275 0 : node: Node,
4276 0 : http_client: reqwest::Client,
4277 0 : jwt: Option<String>,
4278 0 : dir: BlockUnblock,
4279 0 : ) -> Result<(), ApiError> {
4280 0 : let client = PageserverClient::new(
4281 0 : node.get_id(),
4282 0 : http_client,
4283 0 : node.base_url(),
4284 0 : jwt.as_deref(),
4285 0 : );
4286 0 :
4287 0 : client
4288 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
4289 0 : .await
4290 0 : .map_err(|e| passthrough_api_error(&node, e))
4291 0 : }
4292 :
4293 : // no shard needs to go first/last; the operation should be idempotent
4294 0 : let locations = targets
4295 0 : .0
4296 0 : .iter()
4297 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
4298 0 : .collect();
4299 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
4300 0 : futures::FutureExt::boxed(do_one(
4301 0 : tenant_shard_id,
4302 0 : timeline_id,
4303 0 : node,
4304 0 : self.http_client.clone(),
4305 0 : self.config.pageserver_jwt_token.clone(),
4306 0 : dir,
4307 0 : ))
4308 0 : })
4309 0 : .await
4310 0 : })
4311 0 : .await??;
4312 0 : Ok(())
4313 0 : }
4314 :
4315 0 : pub(crate) async fn tenant_timeline_lsn_lease(
4316 0 : &self,
4317 0 : tenant_id: TenantId,
4318 0 : timeline_id: TimelineId,
4319 0 : lsn: Lsn,
4320 0 : ) -> Result<LsnLease, ApiError> {
4321 0 : let _tenant_lock = trace_shared_lock(
4322 0 : &self.tenant_op_locks,
4323 0 : tenant_id,
4324 0 : TenantOperations::TimelineLsnLease,
4325 0 : )
4326 0 : .await;
4327 :
4328 0 : let targets = {
4329 0 : let locked = self.inner.read().unwrap();
4330 0 : let mut targets = Vec::new();
4331 0 :
4332 0 : // If the request got an unsharded tenant id, then apply
4333 0 : // the operation to all shards. Otherwise, apply it to a specific shard.
4334 0 : let shards_range = TenantShardId::tenant_range(tenant_id);
4335 :
4336 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4337 0 : if let Some(node_id) = shard.intent.get_attached() {
4338 0 : let node = locked
4339 0 : .nodes
4340 0 : .get(node_id)
4341 0 : .expect("Pageservers may not be deleted while referenced");
4342 0 :
4343 0 : targets.push((*tenant_shard_id, node.clone()));
4344 0 : }
4345 : }
4346 0 : targets
4347 : };
4348 :
4349 0 : let res = self
4350 0 : .tenant_for_shards_api(
4351 0 : targets,
4352 0 : |tenant_shard_id, client| async move {
4353 0 : client
4354 0 : .timeline_lease_lsn(tenant_shard_id, timeline_id, lsn)
4355 0 : .await
4356 0 : },
4357 0 : 1,
4358 0 : 1,
4359 0 : SHORT_RECONCILE_TIMEOUT,
4360 0 : &self.cancel,
4361 0 : )
4362 0 : .await;
4363 :
4364 0 : let mut valid_until = None;
4365 0 : for r in res {
4366 0 : match r {
4367 0 : Ok(lease) => {
4368 0 : if let Some(ref mut valid_until) = valid_until {
4369 0 : *valid_until = std::cmp::min(*valid_until, lease.valid_until);
4370 0 : } else {
4371 0 : valid_until = Some(lease.valid_until);
4372 0 : }
4373 : }
4374 0 : Err(e) => {
4375 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
4376 : }
4377 : }
4378 : }
4379 0 : Ok(LsnLease {
4380 0 : valid_until: valid_until.unwrap_or_else(SystemTime::now),
4381 0 : })
4382 0 : }
4383 :
4384 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4385 0 : &self,
4386 0 : tenant_shard_id: TenantShardId,
4387 0 : timeline_id: TimelineId,
4388 0 : concurrency: Option<usize>,
4389 0 : recurse: bool,
4390 0 : ) -> Result<(), ApiError> {
4391 0 : let _tenant_lock = trace_shared_lock(
4392 0 : &self.tenant_op_locks,
4393 0 : tenant_shard_id.tenant_id,
4394 0 : TenantOperations::DownloadHeatmapLayers,
4395 0 : )
4396 0 : .await;
4397 :
4398 0 : let targets = {
4399 0 : let locked = self.inner.read().unwrap();
4400 0 : let mut targets = Vec::new();
4401 :
4402 : // If the request got an unsharded tenant id, then apply
4403 : // the operation to all shards. Otherwise, apply it to a specific shard.
4404 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4405 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4406 : } else {
4407 0 : tenant_shard_id.range()
4408 : };
4409 :
4410 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4411 0 : if let Some(node_id) = shard.intent.get_attached() {
4412 0 : let node = locked
4413 0 : .nodes
4414 0 : .get(node_id)
4415 0 : .expect("Pageservers may not be deleted while referenced");
4416 0 :
4417 0 : targets.push((*tenant_shard_id, node.clone()));
4418 0 : }
4419 : }
4420 0 : targets
4421 0 : };
4422 0 :
4423 0 : self.tenant_for_shards_api(
4424 0 : targets,
4425 0 : |tenant_shard_id, client| async move {
4426 0 : client
4427 0 : .timeline_download_heatmap_layers(
4428 0 : tenant_shard_id,
4429 0 : timeline_id,
4430 0 : concurrency,
4431 0 : recurse,
4432 0 : )
4433 0 : .await
4434 0 : },
4435 0 : 1,
4436 0 : 1,
4437 0 : SHORT_RECONCILE_TIMEOUT,
4438 0 : &self.cancel,
4439 0 : )
4440 0 : .await;
4441 :
4442 0 : Ok(())
4443 0 : }
4444 :
4445 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4446 : ///
4447 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`
4448 : /// and returned element at index `i` is the result for `req_fn(op(locations[i])`.
4449 0 : async fn tenant_for_shards<F, R>(
4450 0 : &self,
4451 0 : locations: Vec<(TenantShardId, Node)>,
4452 0 : mut req_fn: F,
4453 0 : ) -> Result<Vec<R>, ApiError>
4454 0 : where
4455 0 : F: FnMut(
4456 0 : TenantShardId,
4457 0 : Node,
4458 0 : )
4459 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4460 0 : {
4461 0 : let mut futs = FuturesUnordered::new();
4462 0 : let mut results = Vec::with_capacity(locations.len());
4463 :
4464 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4465 0 : let fut = req_fn(tenant_shard_id, node);
4466 0 : futs.push(async move { (idx, fut.await) });
4467 0 : }
4468 :
4469 0 : while let Some((idx, r)) = futs.next().await {
4470 0 : results.push((idx, r?));
4471 : }
4472 :
4473 0 : results.sort_by_key(|(idx, _)| *idx);
4474 0 : Ok(results.into_iter().map(|(_, r)| r).collect())
4475 0 : }
4476 :
4477 : /// Concurrently invoke a pageserver API call on many shards at once.
4478 : ///
4479 : /// The returned Vec has the same length as the `locations` Vec,
4480 : /// and returned element at index `i` is the result for `op(locations[i])`.
4481 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
4482 0 : &self,
4483 0 : locations: Vec<(TenantShardId, Node)>,
4484 0 : op: O,
4485 0 : warn_threshold: u32,
4486 0 : max_retries: u32,
4487 0 : timeout: Duration,
4488 0 : cancel: &CancellationToken,
4489 0 : ) -> Vec<mgmt_api::Result<T>>
4490 0 : where
4491 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
4492 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
4493 0 : {
4494 0 : let mut futs = FuturesUnordered::new();
4495 0 : let mut results = Vec::with_capacity(locations.len());
4496 :
4497 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4498 0 : futs.push(async move {
4499 0 : let r = node
4500 0 : .with_client_retries(
4501 0 : |client| op(tenant_shard_id, client),
4502 0 : &self.http_client,
4503 0 : &self.config.pageserver_jwt_token,
4504 0 : warn_threshold,
4505 0 : max_retries,
4506 0 : timeout,
4507 0 : cancel,
4508 0 : )
4509 0 : .await;
4510 0 : (idx, r)
4511 0 : });
4512 0 : }
4513 :
4514 0 : while let Some((idx, r)) = futs.next().await {
4515 0 : results.push((idx, r.unwrap_or(Err(mgmt_api::Error::Cancelled))));
4516 0 : }
4517 :
4518 0 : results.sort_by_key(|(idx, _)| *idx);
4519 0 : results.into_iter().map(|(_, r)| r).collect()
4520 0 : }
4521 :
4522 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
4523 : /// when creating and deleting timelines:
4524 : /// - Makes sure shards are attached somewhere if they weren't already
4525 : /// - Looks up the shards and the nodes where they were most recently attached
4526 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
4527 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
4528 0 : async fn tenant_remote_mutation<R, O, F>(
4529 0 : &self,
4530 0 : tenant_id: TenantId,
4531 0 : op: O,
4532 0 : ) -> Result<R, ApiError>
4533 0 : where
4534 0 : O: FnOnce(TenantMutationLocations) -> F,
4535 0 : F: std::future::Future<Output = R>,
4536 0 : {
4537 0 : let mutation_locations = {
4538 0 : let mut locations = TenantMutationLocations::default();
4539 :
4540 : // Load the currently attached pageservers for the latest generation of each shard. This can
4541 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
4542 : // will still be the latest when we're done: we will check generations again at the end of
4543 : // this function to handle that.
4544 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
4545 :
4546 0 : if generations
4547 0 : .iter()
4548 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
4549 : {
4550 0 : let shard_generations = generations
4551 0 : .into_iter()
4552 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
4553 0 : .collect::<HashMap<_, _>>();
4554 0 :
4555 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
4556 0 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
4557 0 : let locked = self.inner.read().unwrap();
4558 0 : for (shard_id, shard) in
4559 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4560 : {
4561 0 : match shard.policy {
4562 : PlacementPolicy::Attached(_) => {
4563 : // This shard is meant to be attached: the caller is not wrong to try and
4564 : // use this function, but we can't service the request right now.
4565 0 : let Some(generation) = shard_generations.get(shard_id) else {
4566 : // This can only happen if there is a split brain controller modifying the database. This should
4567 : // never happen when testing, and if it happens in production we can only log the issue.
4568 0 : debug_assert!(false);
4569 0 : tracing::error!(
4570 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
4571 : );
4572 0 : continue;
4573 : };
4574 0 : let (generation, generation_pageserver) = generation;
4575 0 : if let Some(generation) = generation {
4576 0 : if generation_pageserver.is_none() {
4577 : // This is legitimate only in a very narrow window where the shard was only just configured into
4578 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
4579 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
4580 0 : tracing::warn!(
4581 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
4582 : );
4583 0 : }
4584 : } else {
4585 : // This should never happen: a shard with no generation is only permitted when it was created in some state
4586 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
4587 0 : debug_assert!(false);
4588 0 : tracing::error!(
4589 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
4590 : );
4591 0 : continue;
4592 : }
4593 : }
4594 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
4595 0 : return Err(ApiError::Conflict(format!(
4596 0 : "Shard {shard_id} tenant has policy {:?}",
4597 0 : shard.policy
4598 0 : )));
4599 : }
4600 : }
4601 : }
4602 :
4603 0 : return Err(ApiError::ResourceUnavailable(
4604 0 : "One or more shards in tenant is not yet attached".into(),
4605 0 : ));
4606 0 : }
4607 0 :
4608 0 : let locked = self.inner.read().unwrap();
4609 : for ShardGenerationState {
4610 0 : tenant_shard_id,
4611 0 : generation,
4612 0 : generation_pageserver,
4613 0 : } in generations
4614 : {
4615 0 : let node_id = generation_pageserver.expect("We checked for None above");
4616 0 : let node = locked
4617 0 : .nodes
4618 0 : .get(&node_id)
4619 0 : .ok_or(ApiError::Conflict(format!(
4620 0 : "Raced with removal of node {node_id}"
4621 0 : )))?;
4622 0 : let generation = generation.expect("Checked above");
4623 0 :
4624 0 : let tenant = locked.tenants.get(&tenant_shard_id);
4625 :
4626 : // TODO(vlad): Abstract the logic that finds stale attached locations
4627 : // from observed state into a [`Service`] method.
4628 0 : let other_locations = match tenant {
4629 0 : Some(tenant) => {
4630 0 : let mut other = tenant.attached_locations();
4631 0 : let latest_location_index =
4632 0 : other.iter().position(|&l| l == (node.get_id(), generation));
4633 0 : if let Some(idx) = latest_location_index {
4634 0 : other.remove(idx);
4635 0 : }
4636 :
4637 0 : other
4638 : }
4639 0 : None => Vec::default(),
4640 : };
4641 :
4642 0 : let location = ShardMutationLocations {
4643 0 : latest: MutationLocation {
4644 0 : node: node.clone(),
4645 0 : generation,
4646 0 : },
4647 0 : other: other_locations
4648 0 : .into_iter()
4649 0 : .filter_map(|(node_id, generation)| {
4650 0 : let node = locked.nodes.get(&node_id)?;
4651 :
4652 0 : Some(MutationLocation {
4653 0 : node: node.clone(),
4654 0 : generation,
4655 0 : })
4656 0 : })
4657 0 : .collect(),
4658 0 : };
4659 0 : locations.0.insert(tenant_shard_id, location);
4660 0 : }
4661 :
4662 0 : locations
4663 : };
4664 :
4665 0 : let result = op(mutation_locations.clone()).await;
4666 :
4667 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
4668 : // our remote operation executed on the latest generation and is therefore persistent.
4669 : {
4670 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
4671 0 : if latest_generations
4672 0 : .into_iter()
4673 0 : .map(
4674 0 : |ShardGenerationState {
4675 : tenant_shard_id,
4676 : generation,
4677 : generation_pageserver: _,
4678 0 : }| (tenant_shard_id, generation),
4679 0 : )
4680 0 : .collect::<Vec<_>>()
4681 0 : != mutation_locations
4682 0 : .0
4683 0 : .into_iter()
4684 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
4685 0 : .collect::<Vec<_>>()
4686 : {
4687 : // We raced with something that incremented the generation, and therefore cannot be
4688 : // confident that our actions are persistent (they might have hit an old generation).
4689 : //
4690 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
4691 0 : return Err(ApiError::ResourceUnavailable(
4692 0 : "Tenant attachment changed, please retry".into(),
4693 0 : ));
4694 0 : }
4695 0 : }
4696 0 :
4697 0 : Ok(result)
4698 0 : }
4699 :
4700 0 : pub(crate) async fn tenant_timeline_delete(
4701 0 : self: &Arc<Self>,
4702 0 : tenant_id: TenantId,
4703 0 : timeline_id: TimelineId,
4704 0 : ) -> Result<StatusCode, ApiError> {
4705 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
4706 0 : let _tenant_lock = trace_shared_lock(
4707 0 : &self.tenant_op_locks,
4708 0 : tenant_id,
4709 0 : TenantOperations::TimelineDelete,
4710 0 : )
4711 0 : .await;
4712 :
4713 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
4714 0 : if targets.0.is_empty() {
4715 0 : return Err(ApiError::NotFound(
4716 0 : anyhow::anyhow!("Tenant not found").into(),
4717 0 : ));
4718 0 : }
4719 0 :
4720 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
4721 0 : assert!(shard_zero_tid.is_shard_zero());
4722 :
4723 0 : async fn delete_one(
4724 0 : tenant_shard_id: TenantShardId,
4725 0 : timeline_id: TimelineId,
4726 0 : node: Node,
4727 0 : http_client: reqwest::Client,
4728 0 : jwt: Option<String>,
4729 0 : ) -> Result<StatusCode, ApiError> {
4730 0 : tracing::info!(
4731 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4732 : );
4733 :
4734 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4735 0 : let res = client
4736 0 : .timeline_delete(tenant_shard_id, timeline_id)
4737 0 : .await;
4738 :
4739 0 : match res {
4740 0 : Ok(ok) => Ok(ok),
4741 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
4742 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
4743 0 : Err(e) => {
4744 0 : Err(
4745 0 : ApiError::InternalServerError(anyhow::anyhow!(
4746 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
4747 0 : ))
4748 0 : )
4749 : }
4750 : }
4751 0 : }
4752 :
4753 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4754 0 : let statuses = self
4755 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
4756 0 : Box::pin(delete_one(
4757 0 : tenant_shard_id,
4758 0 : timeline_id,
4759 0 : node,
4760 0 : self.http_client.clone(),
4761 0 : self.config.pageserver_jwt_token.clone(),
4762 0 : ))
4763 0 : })
4764 0 : .await?;
4765 :
4766 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
4767 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
4768 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
4769 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
4770 0 : return Ok(StatusCode::CONFLICT);
4771 0 : }
4772 0 :
4773 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
4774 0 : return Ok(StatusCode::ACCEPTED);
4775 0 : }
4776 :
4777 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
4778 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
4779 0 : let shard_zero_status = delete_one(
4780 0 : shard_zero_tid,
4781 0 : timeline_id,
4782 0 : shard_zero_locations.latest.node,
4783 0 : self.http_client.clone(),
4784 0 : self.config.pageserver_jwt_token.clone(),
4785 0 : )
4786 0 : .await?;
4787 0 : Ok(shard_zero_status)
4788 0 : }).await?;
4789 :
4790 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
4791 0 : .await?;
4792 :
4793 0 : status_code
4794 0 : }
4795 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
4796 0 : pub(crate) async fn tenant_shard0_node(
4797 0 : &self,
4798 0 : tenant_id: TenantId,
4799 0 : ) -> Result<(Node, TenantShardId), ApiError> {
4800 0 : let tenant_shard_id = {
4801 0 : let locked = self.inner.read().unwrap();
4802 0 : let Some((tenant_shard_id, _shard)) = locked
4803 0 : .tenants
4804 0 : .range(TenantShardId::tenant_range(tenant_id))
4805 0 : .next()
4806 : else {
4807 0 : return Err(ApiError::NotFound(
4808 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
4809 0 : ));
4810 : };
4811 :
4812 0 : *tenant_shard_id
4813 0 : };
4814 0 :
4815 0 : self.tenant_shard_node(tenant_shard_id)
4816 0 : .await
4817 0 : .map(|node| (node, tenant_shard_id))
4818 0 : }
4819 :
4820 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
4821 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
4822 0 : pub(crate) async fn tenant_shard_node(
4823 0 : &self,
4824 0 : tenant_shard_id: TenantShardId,
4825 0 : ) -> Result<Node, ApiError> {
4826 0 : // Look up in-memory state and maybe use the node from there.
4827 0 : {
4828 0 : let locked = self.inner.read().unwrap();
4829 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
4830 0 : return Err(ApiError::NotFound(
4831 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
4832 0 : ));
4833 : };
4834 :
4835 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
4836 0 : tracing::warn!(
4837 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
4838 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
4839 : shard.policy
4840 : );
4841 0 : return Err(ApiError::Conflict(
4842 0 : "Cannot call timeline API on non-attached tenant".to_string(),
4843 0 : ));
4844 : };
4845 :
4846 0 : if shard.reconciler.is_none() {
4847 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
4848 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
4849 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
4850 : // This should never happen
4851 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4852 0 : "Shard refers to nonexistent node"
4853 0 : )));
4854 : };
4855 0 : return Ok(node.clone());
4856 0 : }
4857 : };
4858 :
4859 : // Look up the latest attached pageserver location from the database
4860 : // generation state: this will reflect the progress of any ongoing migration.
4861 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4862 : // the case where they call through to the pageserver and get a 404.
4863 0 : let db_result = self
4864 0 : .persistence
4865 0 : .tenant_generations(tenant_shard_id.tenant_id)
4866 0 : .await?;
4867 : let Some(ShardGenerationState {
4868 : tenant_shard_id: _,
4869 : generation: _,
4870 0 : generation_pageserver: Some(node_id),
4871 0 : }) = db_result
4872 0 : .into_iter()
4873 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
4874 : else {
4875 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4876 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4877 : // or a conflict response (case where tenant was detached in background)
4878 0 : return Err(ApiError::ResourceUnavailable(
4879 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
4880 0 : ));
4881 : };
4882 0 : let locked = self.inner.read().unwrap();
4883 0 : let Some(node) = locked.nodes.get(&node_id) else {
4884 : // This should never happen
4885 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4886 0 : "Shard refers to nonexistent node"
4887 0 : )));
4888 : };
4889 :
4890 0 : Ok(node.clone())
4891 0 : }
4892 :
4893 0 : pub(crate) fn tenant_locate(
4894 0 : &self,
4895 0 : tenant_id: TenantId,
4896 0 : ) -> Result<TenantLocateResponse, ApiError> {
4897 0 : let locked = self.inner.read().unwrap();
4898 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4899 :
4900 0 : let mut result = Vec::new();
4901 0 : let mut shard_params: Option<ShardParameters> = None;
4902 :
4903 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4904 : {
4905 0 : let node_id =
4906 0 : shard
4907 0 : .intent
4908 0 : .get_attached()
4909 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4910 0 : "Cannot locate a tenant that is not attached"
4911 0 : )))?;
4912 :
4913 0 : let node = locked
4914 0 : .nodes
4915 0 : .get(&node_id)
4916 0 : .expect("Pageservers may not be deleted while referenced");
4917 0 :
4918 0 : result.push(node.shard_location(*tenant_shard_id));
4919 0 :
4920 0 : match &shard_params {
4921 0 : None => {
4922 0 : shard_params = Some(ShardParameters {
4923 0 : stripe_size: shard.shard.stripe_size,
4924 0 : count: shard.shard.count,
4925 0 : });
4926 0 : }
4927 0 : Some(params) => {
4928 0 : if params.stripe_size != shard.shard.stripe_size {
4929 : // This should never happen. We enforce at runtime because it's simpler than
4930 : // adding an extra per-tenant data structure to store the things that should be the same
4931 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4932 0 : "Inconsistent shard stripe size parameters!"
4933 0 : )));
4934 0 : }
4935 : }
4936 : }
4937 : }
4938 :
4939 0 : if result.is_empty() {
4940 0 : return Err(ApiError::NotFound(
4941 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4942 0 : ));
4943 0 : }
4944 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4945 0 : tracing::info!(
4946 0 : "Located tenant {} with params {:?} on shards {}",
4947 0 : tenant_id,
4948 0 : shard_params,
4949 0 : result
4950 0 : .iter()
4951 0 : .map(|s| format!("{:?}", s))
4952 0 : .collect::<Vec<_>>()
4953 0 : .join(",")
4954 : );
4955 :
4956 0 : Ok(TenantLocateResponse {
4957 0 : shards: result,
4958 0 : shard_params,
4959 0 : })
4960 0 : }
4961 :
4962 : /// Returns None if the input iterator of shards does not include a shard with number=0
4963 0 : fn tenant_describe_impl<'a>(
4964 0 : &self,
4965 0 : shards: impl Iterator<Item = &'a TenantShard>,
4966 0 : ) -> Option<TenantDescribeResponse> {
4967 0 : let mut shard_zero = None;
4968 0 : let mut describe_shards = Vec::new();
4969 :
4970 0 : for shard in shards {
4971 0 : if shard.tenant_shard_id.is_shard_zero() {
4972 0 : shard_zero = Some(shard);
4973 0 : }
4974 :
4975 0 : describe_shards.push(TenantDescribeResponseShard {
4976 0 : tenant_shard_id: shard.tenant_shard_id,
4977 0 : node_attached: *shard.intent.get_attached(),
4978 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4979 0 : last_error: shard
4980 0 : .last_error
4981 0 : .lock()
4982 0 : .unwrap()
4983 0 : .as_ref()
4984 0 : .map(|e| format!("{e}"))
4985 0 : .unwrap_or("".to_string())
4986 0 : .clone(),
4987 0 : is_reconciling: shard.reconciler.is_some(),
4988 0 : is_pending_compute_notification: shard.pending_compute_notification,
4989 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4990 0 : is_importing: shard.importing == TimelineImportState::Importing,
4991 0 : scheduling_policy: shard.get_scheduling_policy(),
4992 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4993 : })
4994 : }
4995 :
4996 0 : let shard_zero = shard_zero?;
4997 :
4998 0 : Some(TenantDescribeResponse {
4999 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
5000 0 : shards: describe_shards,
5001 0 : stripe_size: shard_zero.shard.stripe_size,
5002 0 : policy: shard_zero.policy.clone(),
5003 0 : config: shard_zero.config.clone(),
5004 0 : })
5005 0 : }
5006 :
5007 0 : pub(crate) fn tenant_describe(
5008 0 : &self,
5009 0 : tenant_id: TenantId,
5010 0 : ) -> Result<TenantDescribeResponse, ApiError> {
5011 0 : let locked = self.inner.read().unwrap();
5012 0 :
5013 0 : self.tenant_describe_impl(
5014 0 : locked
5015 0 : .tenants
5016 0 : .range(TenantShardId::tenant_range(tenant_id))
5017 0 : .map(|(_k, v)| v),
5018 0 : )
5019 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
5020 0 : }
5021 :
5022 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
5023 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
5024 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
5025 : /// in our external API.
5026 0 : pub(crate) fn tenant_list(
5027 0 : &self,
5028 0 : limit: Option<usize>,
5029 0 : start_after: Option<TenantId>,
5030 0 : ) -> Vec<TenantDescribeResponse> {
5031 0 : let locked = self.inner.read().unwrap();
5032 :
5033 : // Apply start_from parameter
5034 0 : let shard_range = match start_after {
5035 0 : None => locked.tenants.range(..),
5036 0 : Some(tenant_id) => locked.tenants.range(
5037 0 : TenantShardId {
5038 0 : tenant_id,
5039 0 : shard_number: ShardNumber(u8::MAX),
5040 0 : shard_count: ShardCount(u8::MAX),
5041 0 : }..,
5042 0 : ),
5043 : };
5044 :
5045 0 : let mut result = Vec::new();
5046 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
5047 0 : result.push(
5048 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
5049 0 : .expect("Groups are always non-empty"),
5050 0 : );
5051 :
5052 : // Enforce `limit` parameter
5053 0 : if let Some(limit) = limit {
5054 0 : if result.len() >= limit {
5055 0 : break;
5056 0 : }
5057 0 : }
5058 : }
5059 :
5060 0 : result
5061 0 : }
5062 :
5063 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
5064 : async fn abort_tenant_shard_split(
5065 : &self,
5066 : op: &TenantShardSplitAbort,
5067 : ) -> Result<(), TenantShardSplitAbortError> {
5068 : // Cleaning up a split:
5069 : // - Parent shards are not destroyed during a split, just detached.
5070 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
5071 : // just the children attached, or both.
5072 : //
5073 : // Therefore our work to do is to:
5074 : // 1. Clean up storage controller's internal state to just refer to parents, no children
5075 : // 2. Call out to pageservers to ensure that children are detached
5076 : // 3. Call out to pageservers to ensure that parents are attached.
5077 : //
5078 : // Crash safety:
5079 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
5080 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
5081 : // and detach them.
5082 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
5083 : // from our database, then we will re-enter this cleanup routine on startup.
5084 :
5085 : let TenantShardSplitAbort {
5086 : tenant_id,
5087 : new_shard_count,
5088 : new_stripe_size,
5089 : ..
5090 : } = op;
5091 :
5092 : // First abort persistent state, if any exists.
5093 : match self
5094 : .persistence
5095 : .abort_shard_split(*tenant_id, *new_shard_count)
5096 : .await?
5097 : {
5098 : AbortShardSplitStatus::Aborted => {
5099 : // Proceed to roll back any child shards created on pageservers
5100 : }
5101 : AbortShardSplitStatus::Complete => {
5102 : // The split completed (we might hit that path if e.g. our database transaction
5103 : // to write the completion landed in the database, but we dropped connection
5104 : // before seeing the result).
5105 : //
5106 : // We must update in-memory state to reflect the successful split.
5107 : self.tenant_shard_split_commit_inmem(
5108 : *tenant_id,
5109 : *new_shard_count,
5110 : *new_stripe_size,
5111 : );
5112 : return Ok(());
5113 : }
5114 : }
5115 :
5116 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
5117 : let detach_locations: Vec<(Node, TenantShardId)> = {
5118 : let mut detach_locations = Vec::new();
5119 : let mut locked = self.inner.write().unwrap();
5120 : let (nodes, tenants, scheduler) = locked.parts_mut();
5121 :
5122 : for (tenant_shard_id, shard) in
5123 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
5124 : {
5125 : if shard.shard.count == op.new_shard_count {
5126 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
5127 : // is infallible, so if we got an error we shouldn't have got that far.
5128 : tracing::warn!(
5129 : "During split abort, child shard {tenant_shard_id} found in-memory"
5130 : );
5131 : continue;
5132 : }
5133 :
5134 : // Add the children of this shard to this list of things to detach
5135 : if let Some(node_id) = shard.intent.get_attached() {
5136 : for child_id in tenant_shard_id.split(*new_shard_count) {
5137 : detach_locations.push((
5138 : nodes
5139 : .get(node_id)
5140 : .expect("Intent references nonexistent node")
5141 : .clone(),
5142 : child_id,
5143 : ));
5144 : }
5145 : } else {
5146 : tracing::warn!(
5147 : "During split abort, shard {tenant_shard_id} has no attached location"
5148 : );
5149 : }
5150 :
5151 : tracing::info!("Restoring parent shard {tenant_shard_id}");
5152 :
5153 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
5154 : // if the original attachment location is offline.
5155 : if let Some(node_id) = shard.intent.get_attached() {
5156 : if !nodes.get(node_id).unwrap().is_available() {
5157 : tracing::info!(
5158 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
5159 : );
5160 : shard.intent.demote_attached(scheduler, *node_id);
5161 : }
5162 : }
5163 : for node_id in shard.intent.get_secondary().clone() {
5164 : if !nodes.get(&node_id).unwrap().is_available() {
5165 : tracing::info!(
5166 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
5167 : );
5168 : shard.intent.remove_secondary(scheduler, node_id);
5169 : }
5170 : }
5171 :
5172 : shard.splitting = SplitState::Idle;
5173 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
5174 : // If this shard can't be scheduled now (perhaps due to offline nodes or
5175 : // capacity issues), that must not prevent us rolling back a split. In this
5176 : // case it should be eventually scheduled in the background.
5177 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
5178 : }
5179 :
5180 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
5181 : }
5182 :
5183 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
5184 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
5185 :
5186 : detach_locations
5187 : };
5188 :
5189 : for (node, child_id) in detach_locations {
5190 : if !node.is_available() {
5191 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
5192 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
5193 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
5194 : // them from the node.
5195 : tracing::warn!(
5196 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
5197 : );
5198 : continue;
5199 : }
5200 :
5201 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
5202 : // a 503 and retry, up to our limit.
5203 : tracing::info!("Detaching {child_id} on {node}...");
5204 : match node
5205 : .with_client_retries(
5206 0 : |client| async move {
5207 0 : let config = LocationConfig {
5208 0 : mode: LocationConfigMode::Detached,
5209 0 : generation: None,
5210 0 : secondary_conf: None,
5211 0 : shard_number: child_id.shard_number.0,
5212 0 : shard_count: child_id.shard_count.literal(),
5213 0 : // Stripe size and tenant config don't matter when detaching
5214 0 : shard_stripe_size: 0,
5215 0 : tenant_conf: TenantConfig::default(),
5216 0 : };
5217 0 :
5218 0 : client.location_config(child_id, config, None, false).await
5219 0 : },
5220 : &self.http_client,
5221 : &self.config.pageserver_jwt_token,
5222 : 1,
5223 : 10,
5224 : Duration::from_secs(5),
5225 : &self.reconcilers_cancel,
5226 : )
5227 : .await
5228 : {
5229 : Some(Ok(_)) => {}
5230 : Some(Err(e)) => {
5231 : // We failed to communicate with the remote node. This is problematic: we may be
5232 : // leaving it with a rogue child shard.
5233 : tracing::warn!(
5234 : "Failed to detach child {child_id} from node {node} during abort"
5235 : );
5236 : return Err(e.into());
5237 : }
5238 : None => {
5239 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
5240 : // clean up on restart. The node going offline requires a retry.
5241 : return Err(TenantShardSplitAbortError::Unavailable);
5242 : }
5243 : };
5244 : }
5245 :
5246 : tracing::info!("Successfully aborted split");
5247 : Ok(())
5248 : }
5249 :
5250 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
5251 : /// of the tenant map to reflect the child shards that exist after the split.
5252 0 : fn tenant_shard_split_commit_inmem(
5253 0 : &self,
5254 0 : tenant_id: TenantId,
5255 0 : new_shard_count: ShardCount,
5256 0 : new_stripe_size: Option<ShardStripeSize>,
5257 0 : ) -> (
5258 0 : TenantShardSplitResponse,
5259 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
5260 0 : Vec<ReconcilerWaiter>,
5261 0 : ) {
5262 0 : let mut response = TenantShardSplitResponse {
5263 0 : new_shards: Vec::new(),
5264 0 : };
5265 0 : let mut child_locations = Vec::new();
5266 0 : let mut waiters = Vec::new();
5267 0 :
5268 0 : {
5269 0 : let mut locked = self.inner.write().unwrap();
5270 0 :
5271 0 : let parent_ids = locked
5272 0 : .tenants
5273 0 : .range(TenantShardId::tenant_range(tenant_id))
5274 0 : .map(|(shard_id, _)| *shard_id)
5275 0 : .collect::<Vec<_>>();
5276 0 :
5277 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5278 0 : for parent_id in parent_ids {
5279 0 : let child_ids = parent_id.split(new_shard_count);
5280 :
5281 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
5282 0 : let mut old_state = tenants
5283 0 : .remove(&parent_id)
5284 0 : .expect("It was present, we just split it");
5285 0 :
5286 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
5287 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
5288 0 : // nothing else can clear this.
5289 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
5290 :
5291 0 : let old_attached = old_state.intent.get_attached().unwrap();
5292 0 : old_state.intent.clear(scheduler);
5293 0 : let generation = old_state.generation.expect("Shard must have been attached");
5294 0 : (
5295 0 : old_attached,
5296 0 : generation,
5297 0 : old_state.policy.clone(),
5298 0 : old_state.shard,
5299 0 : old_state.config.clone(),
5300 0 : old_state.preferred_az().cloned(),
5301 0 : )
5302 0 : };
5303 0 :
5304 0 : let mut schedule_context = ScheduleContext::default();
5305 0 : for child in child_ids {
5306 0 : let mut child_shard = parent_ident;
5307 0 : child_shard.number = child.shard_number;
5308 0 : child_shard.count = child.shard_count;
5309 0 : if let Some(stripe_size) = new_stripe_size {
5310 0 : child_shard.stripe_size = stripe_size;
5311 0 : }
5312 :
5313 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5314 0 : child_observed.insert(
5315 0 : pageserver,
5316 0 : ObservedStateLocation {
5317 0 : conf: Some(attached_location_conf(
5318 0 : generation,
5319 0 : &child_shard,
5320 0 : &config,
5321 0 : &policy,
5322 0 : )),
5323 0 : },
5324 0 : );
5325 0 :
5326 0 : let mut child_state =
5327 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
5328 0 : child_state.intent =
5329 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
5330 0 : child_state.observed = ObservedState {
5331 0 : locations: child_observed,
5332 0 : };
5333 0 : child_state.generation = Some(generation);
5334 0 : child_state.config = config.clone();
5335 0 :
5336 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
5337 0 : // as at this point in the split process we have succeeded and this part is infallible:
5338 0 : // we will never need to do any special recovery from this state.
5339 0 :
5340 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
5341 :
5342 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
5343 : // This is not fatal, because we've implicitly already got an attached
5344 : // location for the child shard. Failure here just means we couldn't
5345 : // find a secondary (e.g. because cluster is overloaded).
5346 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
5347 0 : }
5348 : // In the background, attach secondary locations for the new shards
5349 0 : if let Some(waiter) = self.maybe_reconcile_shard(
5350 0 : &mut child_state,
5351 0 : nodes,
5352 0 : ReconcilerPriority::High,
5353 0 : ) {
5354 0 : waiters.push(waiter);
5355 0 : }
5356 :
5357 0 : tenants.insert(child, child_state);
5358 0 : response.new_shards.push(child);
5359 : }
5360 : }
5361 0 : (response, child_locations, waiters)
5362 0 : }
5363 0 : }
5364 :
5365 0 : async fn tenant_shard_split_start_secondaries(
5366 0 : &self,
5367 0 : tenant_id: TenantId,
5368 0 : waiters: Vec<ReconcilerWaiter>,
5369 0 : ) {
5370 : // Wait for initial reconcile of child shards, this creates the secondary locations
5371 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5372 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
5373 : // their secondaries couldn't be attached.
5374 0 : tracing::warn!("Failed to reconcile after split: {e}");
5375 0 : return;
5376 0 : }
5377 :
5378 : // Take the state lock to discover the attached & secondary intents for all shards
5379 0 : let (attached, secondary) = {
5380 0 : let locked = self.inner.read().unwrap();
5381 0 : let mut attached = Vec::new();
5382 0 : let mut secondary = Vec::new();
5383 :
5384 0 : for (tenant_shard_id, shard) in
5385 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5386 : {
5387 0 : let Some(node_id) = shard.intent.get_attached() else {
5388 : // Unexpected. Race with a PlacementPolicy change?
5389 0 : tracing::warn!(
5390 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
5391 : );
5392 0 : continue;
5393 : };
5394 :
5395 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
5396 : // No secondary location. Nothing for us to do.
5397 0 : continue;
5398 : };
5399 :
5400 0 : let attached_node = locked
5401 0 : .nodes
5402 0 : .get(node_id)
5403 0 : .expect("Pageservers may not be deleted while referenced");
5404 0 :
5405 0 : let secondary_node = locked
5406 0 : .nodes
5407 0 : .get(secondary_node_id)
5408 0 : .expect("Pageservers may not be deleted while referenced");
5409 0 :
5410 0 : attached.push((*tenant_shard_id, attached_node.clone()));
5411 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
5412 : }
5413 0 : (attached, secondary)
5414 0 : };
5415 0 :
5416 0 : if secondary.is_empty() {
5417 : // No secondary locations; nothing for us to do
5418 0 : return;
5419 0 : }
5420 :
5421 0 : for result in self
5422 0 : .tenant_for_shards_api(
5423 0 : attached,
5424 0 : |tenant_shard_id, client| async move {
5425 0 : client.tenant_heatmap_upload(tenant_shard_id).await
5426 0 : },
5427 0 : 1,
5428 0 : 1,
5429 0 : SHORT_RECONCILE_TIMEOUT,
5430 0 : &self.cancel,
5431 0 : )
5432 0 : .await
5433 : {
5434 0 : if let Err(e) = result {
5435 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
5436 0 : return;
5437 0 : }
5438 : }
5439 :
5440 0 : for result in self
5441 0 : .tenant_for_shards_api(
5442 0 : secondary,
5443 0 : |tenant_shard_id, client| async move {
5444 0 : client
5445 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
5446 0 : .await
5447 0 : },
5448 0 : 1,
5449 0 : 1,
5450 0 : SHORT_RECONCILE_TIMEOUT,
5451 0 : &self.cancel,
5452 0 : )
5453 0 : .await
5454 : {
5455 0 : if let Err(e) = result {
5456 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
5457 0 : return;
5458 0 : }
5459 : }
5460 0 : }
5461 :
5462 0 : pub(crate) async fn tenant_shard_split(
5463 0 : &self,
5464 0 : tenant_id: TenantId,
5465 0 : split_req: TenantShardSplitRequest,
5466 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
5467 : // TODO: return 503 if we get stuck waiting for this lock
5468 : // (issue https://github.com/neondatabase/neon/issues/7108)
5469 0 : let _tenant_lock = trace_exclusive_lock(
5470 0 : &self.tenant_op_locks,
5471 0 : tenant_id,
5472 0 : TenantOperations::ShardSplit,
5473 0 : )
5474 0 : .await;
5475 :
5476 0 : let _gate = self
5477 0 : .reconcilers_gate
5478 0 : .enter()
5479 0 : .map_err(|_| ApiError::ShuttingDown)?;
5480 :
5481 : // Timeline imports on the pageserver side can't handle shard-splits.
5482 : // If the tenant is importing a timeline, dont't shard split it.
5483 0 : match self
5484 0 : .persistence
5485 0 : .is_tenant_importing_timeline(tenant_id)
5486 0 : .await
5487 : {
5488 0 : Ok(importing) => {
5489 0 : if importing {
5490 0 : return Err(ApiError::Conflict(
5491 0 : "Cannot shard split during timeline import".to_string(),
5492 0 : ));
5493 0 : }
5494 : }
5495 0 : Err(err) => {
5496 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5497 0 : "Failed to check for running imports: {err}"
5498 0 : )));
5499 : }
5500 : }
5501 :
5502 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
5503 0 : let new_stripe_size = split_req.new_stripe_size;
5504 :
5505 : // Validate the request and construct parameters. This phase is fallible, but does not require
5506 : // rollback on errors, as it does no I/O and mutates no state.
5507 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
5508 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
5509 0 : ShardSplitAction::Split(params) => params,
5510 : };
5511 :
5512 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
5513 : // we must roll back.
5514 0 : let r = self
5515 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
5516 0 : .await;
5517 :
5518 0 : let (response, waiters) = match r {
5519 0 : Ok(r) => r,
5520 0 : Err(e) => {
5521 0 : // Split might be part-done, we must do work to abort it.
5522 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
5523 0 : self.abort_tx
5524 0 : .send(TenantShardSplitAbort {
5525 0 : tenant_id,
5526 0 : new_shard_count,
5527 0 : new_stripe_size,
5528 0 : _tenant_lock,
5529 0 : _gate,
5530 0 : })
5531 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
5532 0 : .ok();
5533 0 : return Err(e);
5534 : }
5535 : };
5536 :
5537 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
5538 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
5539 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
5540 : // in [`Self::optimize_all`]
5541 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
5542 0 : .await;
5543 0 : Ok(response)
5544 0 : }
5545 :
5546 0 : fn prepare_tenant_shard_split(
5547 0 : &self,
5548 0 : tenant_id: TenantId,
5549 0 : split_req: TenantShardSplitRequest,
5550 0 : ) -> Result<ShardSplitAction, ApiError> {
5551 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
5552 0 : anyhow::anyhow!("failpoint")
5553 0 : )));
5554 :
5555 0 : let mut policy = None;
5556 0 : let mut config = None;
5557 0 : let mut shard_ident = None;
5558 0 : let mut preferred_az_id = None;
5559 : // Validate input, and calculate which shards we will create
5560 0 : let (old_shard_count, targets) =
5561 : {
5562 0 : let locked = self.inner.read().unwrap();
5563 0 :
5564 0 : let pageservers = locked.nodes.clone();
5565 0 :
5566 0 : let mut targets = Vec::new();
5567 0 :
5568 0 : // In case this is a retry, count how many already-split shards we found
5569 0 : let mut children_found = Vec::new();
5570 0 : let mut old_shard_count = None;
5571 :
5572 0 : for (tenant_shard_id, shard) in
5573 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5574 : {
5575 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
5576 : Ordering::Equal => {
5577 : // Already split this
5578 0 : children_found.push(*tenant_shard_id);
5579 0 : continue;
5580 : }
5581 : Ordering::Greater => {
5582 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5583 0 : "Requested count {} but already have shards at count {}",
5584 0 : split_req.new_shard_count,
5585 0 : shard.shard.count.count()
5586 0 : )));
5587 : }
5588 0 : Ordering::Less => {
5589 0 : // Fall through: this shard has lower count than requested,
5590 0 : // is a candidate for splitting.
5591 0 : }
5592 0 : }
5593 0 :
5594 0 : match old_shard_count {
5595 0 : None => old_shard_count = Some(shard.shard.count),
5596 0 : Some(old_shard_count) => {
5597 0 : if old_shard_count != shard.shard.count {
5598 : // We may hit this case if a caller asked for two splits to
5599 : // different sizes, before the first one is complete.
5600 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
5601 : // of shard_count=1 and shard_count=2 shards in the map.
5602 0 : return Err(ApiError::Conflict(
5603 0 : "Cannot split, currently mid-split".to_string(),
5604 0 : ));
5605 0 : }
5606 : }
5607 : }
5608 0 : if policy.is_none() {
5609 0 : policy = Some(shard.policy.clone());
5610 0 : }
5611 0 : if shard_ident.is_none() {
5612 0 : shard_ident = Some(shard.shard);
5613 0 : }
5614 0 : if config.is_none() {
5615 0 : config = Some(shard.config.clone());
5616 0 : }
5617 0 : if preferred_az_id.is_none() {
5618 0 : preferred_az_id = shard.preferred_az().cloned();
5619 0 : }
5620 :
5621 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
5622 0 : tracing::info!(
5623 0 : "Tenant shard {} already has shard count {}",
5624 : tenant_shard_id,
5625 : split_req.new_shard_count
5626 : );
5627 0 : continue;
5628 0 : }
5629 :
5630 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
5631 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
5632 0 : ))?;
5633 :
5634 0 : let node = pageservers
5635 0 : .get(&node_id)
5636 0 : .expect("Pageservers may not be deleted while referenced");
5637 0 :
5638 0 : targets.push(ShardSplitTarget {
5639 0 : parent_id: *tenant_shard_id,
5640 0 : node: node.clone(),
5641 0 : child_ids: tenant_shard_id
5642 0 : .split(ShardCount::new(split_req.new_shard_count)),
5643 0 : });
5644 : }
5645 :
5646 0 : if targets.is_empty() {
5647 0 : if children_found.len() == split_req.new_shard_count as usize {
5648 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
5649 0 : new_shards: children_found,
5650 0 : }));
5651 : } else {
5652 : // No shards found to split, and no existing children found: the
5653 : // tenant doesn't exist at all.
5654 0 : return Err(ApiError::NotFound(
5655 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
5656 0 : ));
5657 : }
5658 0 : }
5659 0 :
5660 0 : (old_shard_count, targets)
5661 0 : };
5662 0 :
5663 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
5664 0 : let old_shard_count = old_shard_count.unwrap();
5665 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
5666 : // This ShardIdentity will be used as the template for all children, so this implicitly
5667 : // applies the new stripe size to the children.
5668 0 : let mut shard_ident = shard_ident.unwrap();
5669 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
5670 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5671 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
5672 0 : shard_ident.stripe_size
5673 0 : )));
5674 0 : }
5675 0 :
5676 0 : shard_ident.stripe_size = new_stripe_size;
5677 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
5678 0 : shard_ident
5679 : } else {
5680 0 : shard_ident.unwrap()
5681 : };
5682 0 : let policy = policy.unwrap();
5683 0 : let config = config.unwrap();
5684 0 :
5685 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
5686 0 : old_shard_count,
5687 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
5688 0 : new_stripe_size: split_req.new_stripe_size,
5689 0 : targets,
5690 0 : policy,
5691 0 : config,
5692 0 : shard_ident,
5693 0 : preferred_az_id,
5694 0 : })))
5695 0 : }
5696 :
5697 0 : async fn do_tenant_shard_split(
5698 0 : &self,
5699 0 : tenant_id: TenantId,
5700 0 : params: Box<ShardSplitParams>,
5701 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
5702 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
5703 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
5704 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
5705 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
5706 0 : // (https://github.com/neondatabase/neon/issues/6676)
5707 0 :
5708 0 : let ShardSplitParams {
5709 0 : old_shard_count,
5710 0 : new_shard_count,
5711 0 : new_stripe_size,
5712 0 : mut targets,
5713 0 : policy,
5714 0 : config,
5715 0 : shard_ident,
5716 0 : preferred_az_id,
5717 0 : } = *params;
5718 :
5719 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
5720 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
5721 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
5722 : // at the time of split.
5723 0 : let waiters = {
5724 0 : let mut locked = self.inner.write().unwrap();
5725 0 : let mut waiters = Vec::new();
5726 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5727 0 : for target in &mut targets {
5728 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
5729 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5730 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5731 0 : "Shard {} not found",
5732 0 : target.parent_id
5733 0 : )));
5734 : };
5735 :
5736 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
5737 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5738 0 : return Err(ApiError::Conflict(format!(
5739 0 : "Shard {} unexpectedly rescheduled during split",
5740 0 : target.parent_id
5741 0 : )));
5742 0 : }
5743 0 :
5744 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
5745 0 : shard.intent.clear_secondary(scheduler);
5746 :
5747 : // Run Reconciler to execute detach fo secondary locations.
5748 0 : if let Some(waiter) =
5749 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5750 0 : {
5751 0 : waiters.push(waiter);
5752 0 : }
5753 : }
5754 0 : waiters
5755 0 : };
5756 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
5757 :
5758 : // Before creating any new child shards in memory or on the pageservers, persist them: this
5759 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
5760 : // acts as the protection against two concurrent attempts to split: one of them will get a database
5761 : // error trying to insert the child shards.
5762 0 : let mut child_tsps = Vec::new();
5763 0 : for target in &targets {
5764 0 : let mut this_child_tsps = Vec::new();
5765 0 : for child in &target.child_ids {
5766 0 : let mut child_shard = shard_ident;
5767 0 : child_shard.number = child.shard_number;
5768 0 : child_shard.count = child.shard_count;
5769 0 :
5770 0 : tracing::info!(
5771 0 : "Create child shard persistence with stripe size {}",
5772 : shard_ident.stripe_size.0
5773 : );
5774 :
5775 0 : this_child_tsps.push(TenantShardPersistence {
5776 0 : tenant_id: child.tenant_id.to_string(),
5777 0 : shard_number: child.shard_number.0 as i32,
5778 0 : shard_count: child.shard_count.literal() as i32,
5779 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
5780 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
5781 0 : // populate the correct generation as part of its transaction, to protect us
5782 0 : // against racing with changes in the state of the parent.
5783 0 : generation: None,
5784 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
5785 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
5786 0 : config: serde_json::to_string(&config).unwrap(),
5787 0 : splitting: SplitState::Splitting,
5788 0 :
5789 0 : // Scheduling policies and preferred AZ do not carry through to children
5790 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
5791 0 : .unwrap(),
5792 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
5793 0 : });
5794 0 : }
5795 :
5796 0 : child_tsps.push((target.parent_id, this_child_tsps));
5797 : }
5798 :
5799 0 : if let Err(e) = self
5800 0 : .persistence
5801 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
5802 0 : .await
5803 : {
5804 0 : match e {
5805 : DatabaseError::Query(diesel::result::Error::DatabaseError(
5806 : DatabaseErrorKind::UniqueViolation,
5807 : _,
5808 : )) => {
5809 : // Inserting a child shard violated a unique constraint: we raced with another call to
5810 : // this function
5811 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
5812 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
5813 : }
5814 0 : _ => return Err(ApiError::InternalServerError(e.into())),
5815 : }
5816 0 : }
5817 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
5818 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5819 0 : ));
5820 :
5821 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
5822 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
5823 : // is not set in memory, then it was not persisted.
5824 : {
5825 0 : let mut locked = self.inner.write().unwrap();
5826 0 : for target in &targets {
5827 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
5828 0 : parent_shard.splitting = SplitState::Splitting;
5829 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
5830 0 : // split operation.
5831 0 : parent_shard
5832 0 : .observed
5833 0 : .locations
5834 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
5835 0 : }
5836 : }
5837 : }
5838 :
5839 : // TODO: issue split calls concurrently (this only matters once we're splitting
5840 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
5841 :
5842 0 : for target in &targets {
5843 : let ShardSplitTarget {
5844 0 : parent_id,
5845 0 : node,
5846 0 : child_ids,
5847 0 : } = target;
5848 0 : let client = PageserverClient::new(
5849 0 : node.get_id(),
5850 0 : self.http_client.clone(),
5851 0 : node.base_url(),
5852 0 : self.config.pageserver_jwt_token.as_deref(),
5853 0 : );
5854 0 : let response = client
5855 0 : .tenant_shard_split(
5856 0 : *parent_id,
5857 0 : TenantShardSplitRequest {
5858 0 : new_shard_count: new_shard_count.literal(),
5859 0 : new_stripe_size,
5860 0 : },
5861 0 : )
5862 0 : .await
5863 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
5864 :
5865 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
5866 0 : "failpoint".to_string()
5867 0 : )));
5868 :
5869 0 : failpoint_support::sleep_millis_async!(
5870 : "shard-split-post-remote-sleep",
5871 0 : &self.reconcilers_cancel
5872 : );
5873 :
5874 0 : tracing::info!(
5875 0 : "Split {} into {}",
5876 0 : parent_id,
5877 0 : response
5878 0 : .new_shards
5879 0 : .iter()
5880 0 : .map(|s| format!("{:?}", s))
5881 0 : .collect::<Vec<_>>()
5882 0 : .join(",")
5883 : );
5884 :
5885 0 : if &response.new_shards != child_ids {
5886 : // This should never happen: the pageserver should agree with us on how shard splits work.
5887 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5888 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
5889 0 : parent_id,
5890 0 : response.new_shards,
5891 0 : child_ids
5892 0 : )));
5893 0 : }
5894 : }
5895 :
5896 0 : pausable_failpoint!("shard-split-pre-complete");
5897 :
5898 : // TODO: if the pageserver restarted concurrently with our split API call,
5899 : // the actual generation of the child shard might differ from the generation
5900 : // we expect it to have. In order for our in-database generation to end up
5901 : // correct, we should carry the child generation back in the response and apply it here
5902 : // in complete_shard_split (and apply the correct generation in memory)
5903 : // (or, we can carry generation in the request and reject the request if
5904 : // it doesn't match, but that requires more retry logic on this side)
5905 :
5906 0 : self.persistence
5907 0 : .complete_shard_split(tenant_id, old_shard_count, new_shard_count)
5908 0 : .await?;
5909 :
5910 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5911 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5912 0 : ));
5913 :
5914 : // Replace all the shards we just split with their children: this phase is infallible.
5915 0 : let (response, child_locations, waiters) =
5916 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5917 0 :
5918 0 : // Send compute notifications for all the new shards
5919 0 : let mut failed_notifications = Vec::new();
5920 0 : for (child_id, child_ps, stripe_size) in child_locations {
5921 0 : if let Err(e) = self
5922 0 : .compute_hook
5923 0 : .notify(
5924 0 : compute_hook::ShardUpdate {
5925 0 : tenant_shard_id: child_id,
5926 0 : node_id: child_ps,
5927 0 : stripe_size,
5928 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5929 0 : },
5930 0 : &self.reconcilers_cancel,
5931 0 : )
5932 0 : .await
5933 : {
5934 0 : tracing::warn!(
5935 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5936 : child_id,
5937 : child_ps
5938 : );
5939 0 : failed_notifications.push(child_id);
5940 0 : }
5941 : }
5942 :
5943 : // If we failed any compute notifications, make a note to retry later.
5944 0 : if !failed_notifications.is_empty() {
5945 0 : let mut locked = self.inner.write().unwrap();
5946 0 : for failed in failed_notifications {
5947 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5948 0 : shard.pending_compute_notification = true;
5949 0 : }
5950 : }
5951 0 : }
5952 :
5953 0 : Ok((response, waiters))
5954 0 : }
5955 :
5956 : /// A graceful migration: update the preferred node and let optimisation handle the migration
5957 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
5958 : ///
5959 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
5960 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
5961 0 : fn tenant_shard_migrate_with_prewarm(
5962 0 : &self,
5963 0 : migrate_req: &TenantShardMigrateRequest,
5964 0 : shard: &mut TenantShard,
5965 0 : scheduler: &mut Scheduler,
5966 0 : schedule_context: ScheduleContext,
5967 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
5968 0 : shard.set_preferred_node(Some(migrate_req.node_id));
5969 0 :
5970 0 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
5971 0 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
5972 0 : // e.g. by checking secondary is warm enough.
5973 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
5974 0 : }
5975 :
5976 : /// Immediate migration: directly update the intent state and kick off a reconciler
5977 0 : fn tenant_shard_migrate_immediate(
5978 0 : &self,
5979 0 : migrate_req: &TenantShardMigrateRequest,
5980 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5981 0 : shard: &mut TenantShard,
5982 0 : scheduler: &mut Scheduler,
5983 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
5984 0 : // Non-graceful migration: update the intent state immediately
5985 0 : let old_attached = *shard.intent.get_attached();
5986 0 : match shard.policy {
5987 0 : PlacementPolicy::Attached(n) => {
5988 0 : // If our new attached node was a secondary, it no longer should be.
5989 0 : shard
5990 0 : .intent
5991 0 : .remove_secondary(scheduler, migrate_req.node_id);
5992 0 :
5993 0 : shard
5994 0 : .intent
5995 0 : .set_attached(scheduler, Some(migrate_req.node_id));
5996 :
5997 : // If we were already attached to something, demote that to a secondary
5998 0 : if let Some(old_attached) = old_attached {
5999 0 : if n > 0 {
6000 : // Remove other secondaries to make room for the location we'll demote
6001 0 : while shard.intent.get_secondary().len() >= n {
6002 0 : shard.intent.pop_secondary(scheduler);
6003 0 : }
6004 :
6005 0 : shard.intent.push_secondary(scheduler, old_attached);
6006 0 : }
6007 0 : }
6008 : }
6009 0 : PlacementPolicy::Secondary => {
6010 0 : shard.intent.clear(scheduler);
6011 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6012 0 : }
6013 : PlacementPolicy::Detached => {
6014 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6015 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
6016 0 : )));
6017 : }
6018 : }
6019 :
6020 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
6021 0 : shard.sequence = shard.sequence.next();
6022 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
6023 0 : Ok(self.maybe_configured_reconcile_shard(
6024 0 : shard,
6025 0 : nodes,
6026 0 : (&migrate_req.migration_config).into(),
6027 0 : ))
6028 0 : }
6029 :
6030 0 : pub(crate) async fn tenant_shard_migrate(
6031 0 : &self,
6032 0 : tenant_shard_id: TenantShardId,
6033 0 : migrate_req: TenantShardMigrateRequest,
6034 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6035 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
6036 : // get a different outcome to handle
6037 : enum MigrationOutcome {
6038 : Optimization(Option<ScheduleOptimization>),
6039 : Reconcile(Option<ReconcilerWaiter>),
6040 : }
6041 :
6042 0 : let outcome = {
6043 0 : let mut locked = self.inner.write().unwrap();
6044 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6045 :
6046 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6047 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6048 0 : "Node {} not found",
6049 0 : migrate_req.node_id
6050 0 : )));
6051 : };
6052 :
6053 : // Migration to unavavailable node requires force flag
6054 0 : if !node.is_available() {
6055 0 : if migrate_req.migration_config.override_scheduler {
6056 : // Warn but proceed: the caller may intend to manually adjust the placement of
6057 : // a shard even if the node is down, e.g. if intervening during an incident.
6058 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
6059 : } else {
6060 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
6061 0 : return Err(ApiError::PreconditionFailed(
6062 0 : format!("Node {node} is unavailable").into_boxed_str(),
6063 0 : ));
6064 : }
6065 0 : }
6066 :
6067 : // Calculate the ScheduleContext for this tenant
6068 0 : let mut schedule_context = ScheduleContext::default();
6069 0 : for (_shard_id, shard) in
6070 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
6071 0 : {
6072 0 : schedule_context.avoid(&shard.intent.all_pageservers());
6073 0 : }
6074 :
6075 : // Look up the specific shard we will migrate
6076 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6077 0 : return Err(ApiError::NotFound(
6078 0 : anyhow::anyhow!("Tenant shard not found").into(),
6079 0 : ));
6080 : };
6081 :
6082 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
6083 : // be migrated back by the optimiser.
6084 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
6085 0 : scheduler,
6086 0 : &schedule_context,
6087 0 : migrate_req.node_id,
6088 0 : &[],
6089 0 : ) {
6090 0 : if !migrate_req.migration_config.override_scheduler {
6091 0 : return Err(ApiError::PreconditionFailed(
6092 0 : "Migration to a worse-scoring node".into(),
6093 0 : ));
6094 : } else {
6095 0 : tracing::info!(
6096 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
6097 : migrate_req.node_id
6098 : );
6099 : }
6100 0 : }
6101 :
6102 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
6103 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
6104 0 : return Err(ApiError::PreconditionFailed(
6105 0 : format!(
6106 0 : "Migration expected to originate from {} but shard is on {:?}",
6107 0 : origin_node_id,
6108 0 : shard.intent.get_attached()
6109 0 : )
6110 0 : .into(),
6111 0 : ));
6112 0 : }
6113 0 : }
6114 :
6115 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6116 : // No-op case: we will still proceed to wait for reconciliation in case it is
6117 : // incomplete from an earlier update to the intent.
6118 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
6119 :
6120 : // An instruction to migrate to the currently attached node should
6121 : // cancel any pending graceful migration
6122 0 : shard.set_preferred_node(None);
6123 0 :
6124 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
6125 0 : shard,
6126 0 : nodes,
6127 0 : (&migrate_req.migration_config).into(),
6128 0 : ))
6129 0 : } else if migrate_req.migration_config.prewarm {
6130 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
6131 0 : &migrate_req,
6132 0 : shard,
6133 0 : scheduler,
6134 0 : schedule_context,
6135 0 : )?)
6136 : } else {
6137 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
6138 0 : &migrate_req,
6139 0 : nodes,
6140 0 : shard,
6141 0 : scheduler,
6142 0 : )?)
6143 : }
6144 : };
6145 :
6146 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
6147 0 : let waiter = match outcome {
6148 0 : MigrationOutcome::Optimization(Some(optimization)) => {
6149 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
6150 : // we might as well do it more promptly as this is a direct external request.
6151 0 : let mut validated = self
6152 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
6153 0 : .await;
6154 0 : if let Some((_shard_id, optimization)) = validated.pop() {
6155 0 : let mut locked = self.inner.write().unwrap();
6156 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6157 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6158 : // Rare but possible: tenant is removed between generating optimisation and validating it.
6159 0 : return Err(ApiError::NotFound(
6160 0 : anyhow::anyhow!("Tenant shard not found").into(),
6161 0 : ));
6162 : };
6163 :
6164 0 : if !shard.apply_optimization(scheduler, optimization) {
6165 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
6166 : // and therefore not applied.
6167 0 : tracing::warn!(
6168 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
6169 : );
6170 0 : }
6171 0 : self.maybe_configured_reconcile_shard(
6172 0 : shard,
6173 0 : nodes,
6174 0 : (&migrate_req.migration_config).into(),
6175 0 : )
6176 : } else {
6177 0 : None
6178 : }
6179 : }
6180 0 : MigrationOutcome::Optimization(None) => None,
6181 0 : MigrationOutcome::Reconcile(waiter) => waiter,
6182 : };
6183 :
6184 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
6185 : // locations, this has a good chance of timing out.
6186 0 : if let Some(waiter) = waiter {
6187 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6188 : } else {
6189 0 : tracing::info!("Migration is a no-op");
6190 : }
6191 :
6192 0 : Ok(TenantShardMigrateResponse {})
6193 0 : }
6194 :
6195 0 : pub(crate) async fn tenant_shard_migrate_secondary(
6196 0 : &self,
6197 0 : tenant_shard_id: TenantShardId,
6198 0 : migrate_req: TenantShardMigrateRequest,
6199 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6200 0 : let waiter = {
6201 0 : let mut locked = self.inner.write().unwrap();
6202 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6203 :
6204 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6205 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6206 0 : "Node {} not found",
6207 0 : migrate_req.node_id
6208 0 : )));
6209 : };
6210 :
6211 0 : if !node.is_available() {
6212 : // Warn but proceed: the caller may intend to manually adjust the placement of
6213 : // a shard even if the node is down, e.g. if intervening during an incident.
6214 0 : tracing::warn!("Migrating to unavailable node {node}");
6215 0 : }
6216 :
6217 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6218 0 : return Err(ApiError::NotFound(
6219 0 : anyhow::anyhow!("Tenant shard not found").into(),
6220 0 : ));
6221 : };
6222 :
6223 0 : if shard.intent.get_secondary().len() == 1
6224 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
6225 : {
6226 0 : tracing::info!(
6227 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
6228 : shard.intent
6229 : );
6230 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6231 0 : tracing::info!(
6232 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
6233 : );
6234 : } else {
6235 0 : let old_secondaries = shard.intent.get_secondary().clone();
6236 0 : for secondary in old_secondaries {
6237 0 : shard.intent.remove_secondary(scheduler, secondary);
6238 0 : }
6239 :
6240 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6241 0 : shard.sequence = shard.sequence.next();
6242 0 : tracing::info!(
6243 0 : "Migrating secondary to {node}: new intent {:?}",
6244 : shard.intent
6245 : );
6246 : }
6247 :
6248 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6249 : };
6250 :
6251 0 : if let Some(waiter) = waiter {
6252 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6253 : } else {
6254 0 : tracing::info!("Migration is a no-op");
6255 : }
6256 :
6257 0 : Ok(TenantShardMigrateResponse {})
6258 0 : }
6259 :
6260 : /// 'cancel' in this context means cancel any ongoing reconcile
6261 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
6262 0 : &self,
6263 0 : tenant_shard_id: TenantShardId,
6264 0 : ) -> Result<(), ApiError> {
6265 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
6266 0 : let waiter = {
6267 0 : let locked = self.inner.write().unwrap();
6268 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6269 0 : return Err(ApiError::NotFound(
6270 0 : anyhow::anyhow!("Tenant shard not found").into(),
6271 0 : ));
6272 : };
6273 :
6274 0 : let waiter = shard.get_waiter();
6275 0 : match waiter {
6276 : None => {
6277 0 : tracing::info!("Shard does not have an ongoing Reconciler");
6278 0 : return Ok(());
6279 : }
6280 0 : Some(waiter) => {
6281 0 : tracing::info!("Cancelling Reconciler");
6282 0 : shard.cancel_reconciler();
6283 0 : waiter
6284 0 : }
6285 0 : }
6286 0 : };
6287 0 :
6288 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
6289 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
6290 0 : // the Reconciler is misbehaving and not respecting the cancellation token
6291 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
6292 0 : .await?;
6293 :
6294 0 : Ok(())
6295 0 : }
6296 :
6297 : /// This is for debug/support only: we simply drop all state for a tenant, without
6298 : /// detaching or deleting it on pageservers.
6299 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
6300 0 : self.persistence.delete_tenant(tenant_id).await?;
6301 :
6302 0 : let mut locked = self.inner.write().unwrap();
6303 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6304 0 : let mut shards = Vec::new();
6305 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
6306 0 : shards.push(*tenant_shard_id);
6307 0 : }
6308 :
6309 0 : for shard_id in shards {
6310 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
6311 0 : shard.intent.clear(scheduler);
6312 0 : }
6313 : }
6314 :
6315 0 : Ok(())
6316 0 : }
6317 :
6318 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
6319 : /// tenant with a very high generation number so that it will see the existing data.
6320 0 : pub(crate) async fn tenant_import(
6321 0 : &self,
6322 0 : tenant_id: TenantId,
6323 0 : ) -> Result<TenantCreateResponse, ApiError> {
6324 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
6325 0 : let maybe_node = {
6326 0 : self.inner
6327 0 : .read()
6328 0 : .unwrap()
6329 0 : .nodes
6330 0 : .values()
6331 0 : .find(|n| n.is_available())
6332 0 : .cloned()
6333 : };
6334 0 : let Some(node) = maybe_node else {
6335 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
6336 : };
6337 :
6338 0 : let client = PageserverClient::new(
6339 0 : node.get_id(),
6340 0 : self.http_client.clone(),
6341 0 : node.base_url(),
6342 0 : self.config.pageserver_jwt_token.as_deref(),
6343 0 : );
6344 :
6345 0 : let scan_result = client
6346 0 : .tenant_scan_remote_storage(tenant_id)
6347 0 : .await
6348 0 : .map_err(|e| passthrough_api_error(&node, e))?;
6349 :
6350 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
6351 0 : let Some(shard_count) = scan_result
6352 0 : .shards
6353 0 : .iter()
6354 0 : .map(|s| s.tenant_shard_id.shard_count)
6355 0 : .max()
6356 : else {
6357 0 : return Err(ApiError::NotFound(
6358 0 : anyhow::anyhow!("No shards found").into(),
6359 0 : ));
6360 : };
6361 :
6362 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
6363 : // to
6364 0 : let generation = scan_result
6365 0 : .shards
6366 0 : .iter()
6367 0 : .map(|s| s.generation)
6368 0 : .max()
6369 0 : .expect("We already validated >0 shards");
6370 :
6371 : // Find the tenant's stripe size. This wasn't always persisted in the tenant manifest, so
6372 : // fall back to the original default stripe size of 32768 (256 MB) if it's not specified.
6373 : const ORIGINAL_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(32768);
6374 0 : let stripe_size = scan_result
6375 0 : .shards
6376 0 : .iter()
6377 0 : .find(|s| s.tenant_shard_id.shard_count == shard_count && s.generation == generation)
6378 0 : .expect("we validated >0 shards above")
6379 0 : .stripe_size
6380 0 : .unwrap_or_else(|| {
6381 0 : if shard_count.count() > 1 {
6382 0 : warn!("unknown stripe size, assuming {ORIGINAL_STRIPE_SIZE}");
6383 0 : }
6384 0 : ORIGINAL_STRIPE_SIZE
6385 0 : });
6386 :
6387 0 : let (response, waiters) = self
6388 0 : .do_tenant_create(TenantCreateRequest {
6389 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
6390 0 : generation,
6391 0 :
6392 0 : shard_parameters: ShardParameters {
6393 0 : count: shard_count,
6394 0 : stripe_size,
6395 0 : },
6396 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
6397 0 : config: TenantConfig::default(),
6398 0 : })
6399 0 : .await?;
6400 :
6401 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
6402 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
6403 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
6404 : // reconcile, as reconciliation includes notifying compute.
6405 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
6406 0 : }
6407 :
6408 0 : Ok(response)
6409 0 : }
6410 :
6411 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
6412 : /// we don't have to make TenantShard clonable in the return path.
6413 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
6414 0 : let serialized = {
6415 0 : let locked = self.inner.read().unwrap();
6416 0 : let result = locked.tenants.values().collect::<Vec<_>>();
6417 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
6418 : };
6419 :
6420 0 : hyper::Response::builder()
6421 0 : .status(hyper::StatusCode::OK)
6422 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
6423 0 : .body(hyper::Body::from(serialized))
6424 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
6425 0 : }
6426 :
6427 : /// Check the consistency of in-memory state vs. persistent state, and check that the
6428 : /// scheduler's statistics are up to date.
6429 : ///
6430 : /// These consistency checks expect an **idle** system. If changes are going on while
6431 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
6432 : /// checks, but not suitable for running continuously in the background in the field.
6433 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
6434 0 : let (mut expect_nodes, mut expect_shards) = {
6435 0 : let locked = self.inner.read().unwrap();
6436 0 :
6437 0 : locked
6438 0 : .scheduler
6439 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
6440 0 : .context("Scheduler checks")
6441 0 : .map_err(ApiError::InternalServerError)?;
6442 :
6443 0 : let expect_nodes = locked
6444 0 : .nodes
6445 0 : .values()
6446 0 : .map(|n| n.to_persistent())
6447 0 : .collect::<Vec<_>>();
6448 0 :
6449 0 : let expect_shards = locked
6450 0 : .tenants
6451 0 : .values()
6452 0 : .map(|t| t.to_persistent())
6453 0 : .collect::<Vec<_>>();
6454 :
6455 : // This method can only validate the state of an idle system: if a reconcile is in
6456 : // progress, fail out early to avoid giving false errors on state that won't match
6457 : // between database and memory under a ReconcileResult is processed.
6458 0 : for t in locked.tenants.values() {
6459 0 : if t.reconciler.is_some() {
6460 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6461 0 : "Shard {} reconciliation in progress",
6462 0 : t.tenant_shard_id
6463 0 : )));
6464 0 : }
6465 : }
6466 :
6467 0 : (expect_nodes, expect_shards)
6468 : };
6469 :
6470 0 : let mut nodes = self.persistence.list_nodes().await?;
6471 0 : expect_nodes.sort_by_key(|n| n.node_id);
6472 0 : nodes.sort_by_key(|n| n.node_id);
6473 :
6474 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
6475 0 : let node_result = if nodes != expect_nodes {
6476 0 : tracing::error!("Consistency check failed on nodes.");
6477 0 : tracing::error!(
6478 0 : "Nodes in memory: {}",
6479 0 : serde_json::to_string(&expect_nodes)
6480 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6481 : );
6482 0 : tracing::error!(
6483 0 : "Nodes in database: {}",
6484 0 : serde_json::to_string(&nodes)
6485 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6486 : );
6487 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
6488 0 : "Node consistency failure"
6489 0 : )))
6490 : } else {
6491 0 : Ok(())
6492 : };
6493 :
6494 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
6495 0 : persistent_shards
6496 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6497 0 :
6498 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6499 :
6500 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
6501 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
6502 : // fields are added, before doing a comparison.
6503 0 : for tsp in &mut persistent_shards {
6504 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
6505 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
6506 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
6507 : }
6508 :
6509 0 : if persistent_shards != expect_shards {
6510 0 : tracing::error!("Consistency check failed on shards.");
6511 :
6512 0 : tracing::error!(
6513 0 : "Shards in memory: {}",
6514 0 : serde_json::to_string(&expect_shards)
6515 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6516 : );
6517 0 : tracing::error!(
6518 0 : "Shards in database: {}",
6519 0 : serde_json::to_string(&persistent_shards)
6520 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6521 : );
6522 :
6523 : // The total dump log lines above are useful in testing but in the field grafana will
6524 : // usually just drop them because they're so large. So we also do some explicit logging
6525 : // of just the diffs.
6526 0 : let persistent_shards = persistent_shards
6527 0 : .into_iter()
6528 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6529 0 : .collect::<HashMap<_, _>>();
6530 0 : let expect_shards = expect_shards
6531 0 : .into_iter()
6532 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6533 0 : .collect::<HashMap<_, _>>();
6534 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
6535 0 : match expect_shards.get(tenant_shard_id) {
6536 : None => {
6537 0 : tracing::error!(
6538 0 : "Shard {} found in database but not in memory",
6539 : tenant_shard_id
6540 : );
6541 : }
6542 0 : Some(expect_tsp) => {
6543 0 : if expect_tsp != persistent_tsp {
6544 0 : tracing::error!(
6545 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
6546 0 : tenant_shard_id,
6547 0 : serde_json::to_string(expect_tsp).unwrap(),
6548 0 : serde_json::to_string(&persistent_tsp).unwrap()
6549 : );
6550 0 : }
6551 : }
6552 : }
6553 : }
6554 :
6555 : // Having already logged any differences, log any shards that simply aren't present in the database
6556 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
6557 0 : if !persistent_shards.contains_key(tenant_shard_id) {
6558 0 : tracing::error!(
6559 0 : "Shard {} found in memory but not in database: {}",
6560 0 : tenant_shard_id,
6561 0 : serde_json::to_string(memory_tsp)
6562 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6563 : );
6564 0 : }
6565 : }
6566 :
6567 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6568 0 : "Shard consistency failure"
6569 0 : )));
6570 0 : }
6571 0 :
6572 0 : node_result
6573 0 : }
6574 :
6575 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
6576 : /// we don't have to make TenantShard clonable in the return path.
6577 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
6578 0 : let serialized = {
6579 0 : let locked = self.inner.read().unwrap();
6580 0 : serde_json::to_string(&locked.scheduler)
6581 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6582 : };
6583 :
6584 0 : hyper::Response::builder()
6585 0 : .status(hyper::StatusCode::OK)
6586 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
6587 0 : .body(hyper::Body::from(serialized))
6588 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
6589 0 : }
6590 :
6591 : /// This is for debug/support only: we simply drop all state for a tenant, without
6592 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
6593 : /// tenants that were on this node.
6594 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
6595 0 : self.persistence.delete_node(node_id).await?;
6596 :
6597 0 : let mut locked = self.inner.write().unwrap();
6598 :
6599 0 : for shard in locked.tenants.values_mut() {
6600 0 : shard.deref_node(node_id);
6601 0 : shard.observed.locations.remove(&node_id);
6602 0 : }
6603 :
6604 0 : let mut nodes = (*locked.nodes).clone();
6605 0 : nodes.remove(&node_id);
6606 0 : locked.nodes = Arc::new(nodes);
6607 0 : metrics::METRICS_REGISTRY
6608 0 : .metrics_group
6609 0 : .storage_controller_pageserver_nodes
6610 0 : .set(locked.nodes.len() as i64);
6611 0 : metrics::METRICS_REGISTRY
6612 0 : .metrics_group
6613 0 : .storage_controller_https_pageserver_nodes
6614 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
6615 0 :
6616 0 : locked.scheduler.node_remove(node_id);
6617 0 :
6618 0 : Ok(())
6619 0 : }
6620 :
6621 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
6622 : /// that we don't leave any bad state behind in the storage controller, but unclean
6623 : /// in the sense that we are not carefully draining the node.
6624 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
6625 0 : let _node_lock =
6626 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
6627 :
6628 : // 1. Atomically update in-memory state:
6629 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
6630 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
6631 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
6632 : // re-insert references to this node into the ObservedState of shards
6633 : // - drop the node from the scheduler
6634 : {
6635 0 : let mut locked = self.inner.write().unwrap();
6636 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6637 0 :
6638 0 : {
6639 0 : let mut nodes_mut = (*nodes).deref().clone();
6640 0 : match nodes_mut.get_mut(&node_id) {
6641 0 : Some(node) => {
6642 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
6643 0 : // if we crash it would not be desirable to leave the node paused after a restart.
6644 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
6645 0 : }
6646 : None => {
6647 0 : tracing::info!(
6648 0 : "Node not found: presuming this is a retry and returning success"
6649 : );
6650 0 : return Ok(());
6651 : }
6652 : }
6653 :
6654 0 : *nodes = Arc::new(nodes_mut);
6655 : }
6656 :
6657 0 : for (_tenant_id, mut schedule_context, shards) in
6658 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6659 : {
6660 0 : for shard in shards {
6661 0 : if shard.deref_node(node_id) {
6662 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
6663 : // TODO: implement force flag to remove a node even if we can't reschedule
6664 : // a tenant
6665 0 : tracing::error!(
6666 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
6667 : shard.tenant_shard_id
6668 : );
6669 0 : return Err(e.into());
6670 : } else {
6671 0 : tracing::info!(
6672 0 : "Rescheduled shard {} away from node during deletion",
6673 : shard.tenant_shard_id
6674 : )
6675 : }
6676 :
6677 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
6678 0 : }
6679 :
6680 : // Here we remove an existing observed location for the node we're removing, and it will
6681 : // not be re-added by a reconciler's completion because we filter out removed nodes in
6682 : // process_result.
6683 : //
6684 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
6685 : // means any reconciles we spawned will know about the node we're deleting, enabling them
6686 : // to do live migrations if it's still online.
6687 0 : shard.observed.locations.remove(&node_id);
6688 : }
6689 : }
6690 :
6691 0 : scheduler.node_remove(node_id);
6692 0 :
6693 0 : {
6694 0 : let mut nodes_mut = (**nodes).clone();
6695 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
6696 0 : // Ensure that any reconciler holding an Arc<> to this node will
6697 0 : // drop out when trying to RPC to it (setting Offline state sets the
6698 0 : // cancellation token on the Node object).
6699 0 : removed_node.set_availability(NodeAvailability::Offline);
6700 0 : }
6701 0 : *nodes = Arc::new(nodes_mut);
6702 0 : metrics::METRICS_REGISTRY
6703 0 : .metrics_group
6704 0 : .storage_controller_pageserver_nodes
6705 0 : .set(nodes.len() as i64);
6706 0 : metrics::METRICS_REGISTRY
6707 0 : .metrics_group
6708 0 : .storage_controller_https_pageserver_nodes
6709 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
6710 0 : }
6711 0 : }
6712 0 :
6713 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
6714 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
6715 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
6716 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
6717 0 : // that exists.
6718 0 :
6719 0 : // 2. Actually delete the node from the database and from in-memory state
6720 0 : tracing::info!("Deleting node from database");
6721 0 : self.persistence.delete_node(node_id).await?;
6722 :
6723 0 : Ok(())
6724 0 : }
6725 :
6726 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
6727 0 : let nodes = {
6728 0 : self.inner
6729 0 : .read()
6730 0 : .unwrap()
6731 0 : .nodes
6732 0 : .values()
6733 0 : .cloned()
6734 0 : .collect::<Vec<_>>()
6735 0 : };
6736 0 :
6737 0 : Ok(nodes)
6738 0 : }
6739 :
6740 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
6741 0 : self.inner
6742 0 : .read()
6743 0 : .unwrap()
6744 0 : .nodes
6745 0 : .get(&node_id)
6746 0 : .cloned()
6747 0 : .ok_or(ApiError::NotFound(
6748 0 : format!("Node {node_id} not registered").into(),
6749 0 : ))
6750 0 : }
6751 :
6752 0 : pub(crate) async fn get_node_shards(
6753 0 : &self,
6754 0 : node_id: NodeId,
6755 0 : ) -> Result<NodeShardResponse, ApiError> {
6756 0 : let locked = self.inner.read().unwrap();
6757 0 : let mut shards = Vec::new();
6758 0 : for (tid, tenant) in locked.tenants.iter() {
6759 0 : let is_intended_secondary = match (
6760 0 : tenant.intent.get_attached() == &Some(node_id),
6761 0 : tenant.intent.get_secondary().contains(&node_id),
6762 0 : ) {
6763 : (true, true) => {
6764 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6765 0 : "{} attached as primary+secondary on the same node",
6766 0 : tid
6767 0 : )));
6768 : }
6769 0 : (true, false) => Some(false),
6770 0 : (false, true) => Some(true),
6771 0 : (false, false) => None,
6772 : };
6773 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
6774 0 : tenant.observed.locations.get(&node_id)
6775 : {
6776 0 : Some(conf.secondary_conf.is_some())
6777 : } else {
6778 0 : None
6779 : };
6780 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
6781 0 : shards.push(NodeShard {
6782 0 : tenant_shard_id: *tid,
6783 0 : is_intended_secondary,
6784 0 : is_observed_secondary,
6785 0 : });
6786 0 : }
6787 : }
6788 0 : Ok(NodeShardResponse { node_id, shards })
6789 0 : }
6790 :
6791 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
6792 0 : self.persistence.get_leader().await
6793 0 : }
6794 :
6795 0 : pub(crate) async fn node_register(
6796 0 : &self,
6797 0 : register_req: NodeRegisterRequest,
6798 0 : ) -> Result<(), ApiError> {
6799 0 : let _node_lock = trace_exclusive_lock(
6800 0 : &self.node_op_locks,
6801 0 : register_req.node_id,
6802 0 : NodeOperations::Register,
6803 0 : )
6804 0 : .await;
6805 :
6806 : #[derive(PartialEq)]
6807 : enum RegistrationStatus {
6808 : UpToDate,
6809 : NeedUpdate,
6810 : Mismatched,
6811 : New,
6812 : }
6813 :
6814 0 : let registration_status = {
6815 0 : let locked = self.inner.read().unwrap();
6816 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
6817 0 : if node.registration_match(®ister_req) {
6818 0 : if node.need_update(®ister_req) {
6819 0 : RegistrationStatus::NeedUpdate
6820 : } else {
6821 0 : RegistrationStatus::UpToDate
6822 : }
6823 : } else {
6824 0 : RegistrationStatus::Mismatched
6825 : }
6826 : } else {
6827 0 : RegistrationStatus::New
6828 : }
6829 : };
6830 :
6831 0 : match registration_status {
6832 : RegistrationStatus::UpToDate => {
6833 0 : tracing::info!(
6834 0 : "Node {} re-registered with matching address and is up to date",
6835 : register_req.node_id
6836 : );
6837 :
6838 0 : return Ok(());
6839 : }
6840 : RegistrationStatus::Mismatched => {
6841 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
6842 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
6843 : // a fixed address through the lifetime of a node.
6844 0 : tracing::warn!(
6845 0 : "Node {} tried to register with different address",
6846 : register_req.node_id
6847 : );
6848 0 : return Err(ApiError::Conflict(
6849 0 : "Node is already registered with different address".to_string(),
6850 0 : ));
6851 : }
6852 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
6853 0 : // fallthrough
6854 0 : }
6855 0 : }
6856 0 :
6857 0 : // We do not require that a node is actually online when registered (it will start life
6858 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
6859 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
6860 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
6861 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
6862 0 : if tokio::net::lookup_host(format!(
6863 0 : "{}:{}",
6864 0 : register_req.listen_http_addr, register_req.listen_http_port
6865 0 : ))
6866 0 : .await
6867 0 : .is_err()
6868 : {
6869 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
6870 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
6871 : // we return a soft 503 error, to encourage callers to retry past transient issues.
6872 0 : return Err(ApiError::ResourceUnavailable(
6873 0 : format!(
6874 0 : "Node {} tried to register with unknown DNS name '{}'",
6875 0 : register_req.node_id, register_req.listen_http_addr
6876 0 : )
6877 0 : .into(),
6878 0 : ));
6879 0 : }
6880 0 :
6881 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
6882 0 : return Err(ApiError::PreconditionFailed(
6883 0 : format!(
6884 0 : "Node {} has no https port, but use_https is enabled",
6885 0 : register_req.node_id
6886 0 : )
6887 0 : .into(),
6888 0 : ));
6889 0 : }
6890 0 :
6891 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
6892 0 : // This ensures that before we use it for anything or expose it via any external
6893 0 : // API, it is guaranteed to be available after a restart.
6894 0 : let new_node = Node::new(
6895 0 : register_req.node_id,
6896 0 : register_req.listen_http_addr,
6897 0 : register_req.listen_http_port,
6898 0 : register_req.listen_https_port,
6899 0 : register_req.listen_pg_addr,
6900 0 : register_req.listen_pg_port,
6901 0 : register_req.availability_zone_id.clone(),
6902 0 : self.config.use_https_pageserver_api,
6903 0 : );
6904 0 : let new_node = match new_node {
6905 0 : Ok(new_node) => new_node,
6906 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
6907 : };
6908 :
6909 0 : match registration_status {
6910 0 : RegistrationStatus::New => self.persistence.insert_node(&new_node).await?,
6911 : RegistrationStatus::NeedUpdate => {
6912 0 : self.persistence
6913 0 : .update_node_on_registration(
6914 0 : register_req.node_id,
6915 0 : register_req.listen_https_port,
6916 0 : )
6917 0 : .await?
6918 : }
6919 0 : _ => unreachable!("Other statuses have been processed earlier"),
6920 : }
6921 :
6922 0 : let mut locked = self.inner.write().unwrap();
6923 0 : let mut new_nodes = (*locked.nodes).clone();
6924 0 :
6925 0 : locked.scheduler.node_upsert(&new_node);
6926 0 : new_nodes.insert(register_req.node_id, new_node);
6927 0 :
6928 0 : locked.nodes = Arc::new(new_nodes);
6929 0 :
6930 0 : metrics::METRICS_REGISTRY
6931 0 : .metrics_group
6932 0 : .storage_controller_pageserver_nodes
6933 0 : .set(locked.nodes.len() as i64);
6934 0 : metrics::METRICS_REGISTRY
6935 0 : .metrics_group
6936 0 : .storage_controller_https_pageserver_nodes
6937 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
6938 0 :
6939 0 : match registration_status {
6940 : RegistrationStatus::New => {
6941 0 : tracing::info!(
6942 0 : "Registered pageserver {} ({}), now have {} pageservers",
6943 0 : register_req.node_id,
6944 0 : register_req.availability_zone_id,
6945 0 : locked.nodes.len()
6946 : );
6947 : }
6948 : RegistrationStatus::NeedUpdate => {
6949 0 : tracing::info!(
6950 0 : "Re-registered and updated node {} ({})",
6951 : register_req.node_id,
6952 : register_req.availability_zone_id,
6953 : );
6954 : }
6955 0 : _ => unreachable!("Other statuses have been processed earlier"),
6956 : }
6957 0 : Ok(())
6958 0 : }
6959 :
6960 : /// Configure in-memory and persistent state of a node as requested
6961 : ///
6962 : /// Note that this function does not trigger any immediate side effects in response
6963 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
6964 0 : async fn node_state_configure(
6965 0 : &self,
6966 0 : node_id: NodeId,
6967 0 : availability: Option<NodeAvailability>,
6968 0 : scheduling: Option<NodeSchedulingPolicy>,
6969 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
6970 0 : ) -> Result<AvailabilityTransition, ApiError> {
6971 0 : if let Some(scheduling) = scheduling {
6972 : // Scheduling is a persistent part of Node: we must write updates to the database before
6973 : // applying them in memory
6974 0 : self.persistence
6975 0 : .update_node_scheduling_policy(node_id, scheduling)
6976 0 : .await?;
6977 0 : }
6978 :
6979 : // If we're activating a node, then before setting it active we must reconcile any shard locations
6980 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
6981 : // by calling [`Self::node_activate_reconcile`]
6982 : //
6983 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
6984 : // nothing else can mutate its availability while we run.
6985 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
6986 0 : let (activate_node, availability_transition) = {
6987 0 : let locked = self.inner.read().unwrap();
6988 0 : let Some(node) = locked.nodes.get(&node_id) else {
6989 0 : return Err(ApiError::NotFound(
6990 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6991 0 : ));
6992 : };
6993 :
6994 0 : (
6995 0 : node.clone(),
6996 0 : node.get_availability_transition(input_availability),
6997 0 : )
6998 : };
6999 :
7000 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
7001 0 : self.node_activate_reconcile(activate_node, node_lock)
7002 0 : .await?;
7003 0 : }
7004 0 : availability_transition
7005 : } else {
7006 0 : AvailabilityTransition::Unchanged
7007 : };
7008 :
7009 : // Apply changes from the request to our in-memory state for the Node
7010 0 : let mut locked = self.inner.write().unwrap();
7011 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
7012 0 :
7013 0 : let mut new_nodes = (**nodes).clone();
7014 :
7015 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
7016 0 : return Err(ApiError::NotFound(
7017 0 : anyhow::anyhow!("Node not registered").into(),
7018 0 : ));
7019 : };
7020 :
7021 0 : if let Some(availability) = availability {
7022 0 : node.set_availability(availability);
7023 0 : }
7024 :
7025 0 : if let Some(scheduling) = scheduling {
7026 0 : node.set_scheduling(scheduling);
7027 0 : }
7028 :
7029 : // Update the scheduler, in case the elegibility of the node for new shards has changed
7030 0 : scheduler.node_upsert(node);
7031 0 :
7032 0 : let new_nodes = Arc::new(new_nodes);
7033 0 : locked.nodes = new_nodes;
7034 0 :
7035 0 : Ok(availability_transition)
7036 0 : }
7037 :
7038 : /// Handle availability transition of one node
7039 : ///
7040 : /// Note that you should first call [`Self::node_state_configure`] to update
7041 : /// the in-memory state referencing that node. If you need to handle more than one transition
7042 : /// consider using [`Self::handle_node_availability_transitions`].
7043 0 : async fn handle_node_availability_transition(
7044 0 : &self,
7045 0 : node_id: NodeId,
7046 0 : transition: AvailabilityTransition,
7047 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
7048 0 : ) -> Result<(), ApiError> {
7049 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
7050 0 : match transition {
7051 : AvailabilityTransition::ToOffline => {
7052 0 : tracing::info!("Node {} transition to offline", node_id);
7053 :
7054 0 : let mut locked = self.inner.write().unwrap();
7055 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7056 0 :
7057 0 : let mut tenants_affected: usize = 0;
7058 :
7059 0 : for (_tenant_id, mut schedule_context, shards) in
7060 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
7061 : {
7062 0 : for tenant_shard in shards {
7063 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
7064 0 : if let Some(observed_loc) =
7065 0 : tenant_shard.observed.locations.get_mut(&node_id)
7066 0 : {
7067 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
7068 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
7069 0 : observed_loc.conf = None;
7070 0 : }
7071 :
7072 0 : if nodes.len() == 1 {
7073 : // Special case for single-node cluster: there is no point trying to reschedule
7074 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
7075 : // failures to schedule them.
7076 0 : continue;
7077 0 : }
7078 0 :
7079 0 : if !nodes
7080 0 : .values()
7081 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
7082 : {
7083 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
7084 : // trying to reschedule since there's nowhere else to go. Without this
7085 : // branch we incorrectly detach tenants in response to node unavailability.
7086 0 : continue;
7087 0 : }
7088 0 :
7089 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
7090 0 : tenant_shard.sequence = tenant_shard.sequence.next();
7091 0 :
7092 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
7093 0 : Err(e) => {
7094 0 : // It is possible that some tenants will become unschedulable when too many pageservers
7095 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
7096 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
7097 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
7098 : }
7099 : Ok(()) => {
7100 0 : if self
7101 0 : .maybe_reconcile_shard(
7102 0 : tenant_shard,
7103 0 : nodes,
7104 0 : ReconcilerPriority::Normal,
7105 0 : )
7106 0 : .is_some()
7107 0 : {
7108 0 : tenants_affected += 1;
7109 0 : };
7110 : }
7111 : }
7112 0 : }
7113 : }
7114 : }
7115 0 : tracing::info!(
7116 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
7117 : tenants_affected,
7118 : node_id
7119 : )
7120 : }
7121 : AvailabilityTransition::ToActive => {
7122 0 : tracing::info!("Node {} transition to active", node_id);
7123 :
7124 0 : let mut locked = self.inner.write().unwrap();
7125 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7126 :
7127 : // When a node comes back online, we must reconcile any tenant that has a None observed
7128 : // location on the node.
7129 0 : for tenant_shard in tenants.values_mut() {
7130 : // If a reconciliation is already in progress, rely on the previous scheduling
7131 : // decision and skip triggering a new reconciliation.
7132 0 : if tenant_shard.reconciler.is_some() {
7133 0 : continue;
7134 0 : }
7135 :
7136 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
7137 0 : if observed_loc.conf.is_none() {
7138 0 : self.maybe_reconcile_shard(
7139 0 : tenant_shard,
7140 0 : nodes,
7141 0 : ReconcilerPriority::Normal,
7142 0 : );
7143 0 : }
7144 0 : }
7145 : }
7146 :
7147 : // TODO: in the background, we should balance work back onto this pageserver
7148 : }
7149 : // No action required for the intermediate unavailable state.
7150 : // When we transition into active or offline from the unavailable state,
7151 : // the correct handling above will kick in.
7152 : AvailabilityTransition::ToWarmingUpFromActive => {
7153 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
7154 : }
7155 : AvailabilityTransition::ToWarmingUpFromOffline => {
7156 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
7157 : }
7158 : AvailabilityTransition::Unchanged => {
7159 0 : tracing::debug!("Node {} no availability change during config", node_id);
7160 : }
7161 : }
7162 :
7163 0 : Ok(())
7164 0 : }
7165 :
7166 : /// Handle availability transition for multiple nodes
7167 : ///
7168 : /// Note that you should first call [`Self::node_state_configure`] for
7169 : /// all nodes being handled here for the handling to use fresh in-memory state.
7170 0 : async fn handle_node_availability_transitions(
7171 0 : &self,
7172 0 : transitions: Vec<(
7173 0 : NodeId,
7174 0 : TracingExclusiveGuard<NodeOperations>,
7175 0 : AvailabilityTransition,
7176 0 : )>,
7177 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
7178 0 : let mut errors = Vec::default();
7179 0 : for (node_id, node_lock, transition) in transitions {
7180 0 : let res = self
7181 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
7182 0 : .await;
7183 0 : if let Err(err) = res {
7184 0 : errors.push((node_id, err));
7185 0 : }
7186 : }
7187 :
7188 0 : if errors.is_empty() {
7189 0 : Ok(())
7190 : } else {
7191 0 : Err(errors)
7192 : }
7193 0 : }
7194 :
7195 0 : pub(crate) async fn node_configure(
7196 0 : &self,
7197 0 : node_id: NodeId,
7198 0 : availability: Option<NodeAvailability>,
7199 0 : scheduling: Option<NodeSchedulingPolicy>,
7200 0 : ) -> Result<(), ApiError> {
7201 0 : let node_lock =
7202 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
7203 :
7204 0 : let transition = self
7205 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
7206 0 : .await?;
7207 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
7208 0 : .await
7209 0 : }
7210 :
7211 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
7212 : /// operation for HTTP api.
7213 0 : pub(crate) async fn external_node_configure(
7214 0 : &self,
7215 0 : node_id: NodeId,
7216 0 : availability: Option<NodeAvailability>,
7217 0 : scheduling: Option<NodeSchedulingPolicy>,
7218 0 : ) -> Result<(), ApiError> {
7219 0 : {
7220 0 : let locked = self.inner.read().unwrap();
7221 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
7222 0 : return Err(ApiError::PreconditionFailed(
7223 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
7224 0 : ));
7225 0 : }
7226 0 : }
7227 0 :
7228 0 : self.node_configure(node_id, availability, scheduling).await
7229 0 : }
7230 :
7231 0 : pub(crate) async fn start_node_drain(
7232 0 : self: &Arc<Self>,
7233 0 : node_id: NodeId,
7234 0 : ) -> Result<(), ApiError> {
7235 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
7236 0 : let locked = self.inner.read().unwrap();
7237 0 : let nodes = &locked.nodes;
7238 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7239 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7240 0 : ))?;
7241 0 : let schedulable_nodes_count = nodes
7242 0 : .iter()
7243 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
7244 0 : .count();
7245 0 :
7246 0 : (
7247 0 : locked
7248 0 : .ongoing_operation
7249 0 : .as_ref()
7250 0 : .map(|ongoing| ongoing.operation),
7251 0 : node.is_available(),
7252 0 : node.get_scheduling(),
7253 0 : schedulable_nodes_count,
7254 0 : )
7255 0 : };
7256 :
7257 0 : if let Some(ongoing) = ongoing_op {
7258 0 : return Err(ApiError::PreconditionFailed(
7259 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
7260 0 : ));
7261 0 : }
7262 0 :
7263 0 : if !node_available {
7264 0 : return Err(ApiError::ResourceUnavailable(
7265 0 : format!("Node {node_id} is currently unavailable").into(),
7266 0 : ));
7267 0 : }
7268 0 :
7269 0 : if schedulable_nodes_count == 0 {
7270 0 : return Err(ApiError::PreconditionFailed(
7271 0 : "No other schedulable nodes to drain to".into(),
7272 0 : ));
7273 0 : }
7274 0 :
7275 0 : match node_policy {
7276 : NodeSchedulingPolicy::Active => {
7277 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
7278 0 : .await?;
7279 :
7280 0 : let cancel = self.cancel.child_token();
7281 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
7282 :
7283 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
7284 0 : operation: Operation::Drain(Drain { node_id }),
7285 0 : cancel: cancel.clone(),
7286 0 : });
7287 :
7288 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
7289 :
7290 0 : tokio::task::spawn({
7291 0 : let service = self.clone();
7292 0 : let cancel = cancel.clone();
7293 0 : async move {
7294 0 : let _gate_guard = gate_guard;
7295 0 :
7296 0 : scopeguard::defer! {
7297 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
7298 0 :
7299 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
7300 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
7301 0 : } else {
7302 0 : panic!("We always remove the same operation")
7303 0 : }
7304 0 : }
7305 0 :
7306 0 : tracing::info!("Drain background operation starting");
7307 0 : let res = service.drain_node(node_id, cancel).await;
7308 0 : match res {
7309 : Ok(()) => {
7310 0 : tracing::info!("Drain background operation completed successfully");
7311 : }
7312 : Err(OperationError::Cancelled) => {
7313 0 : tracing::info!("Drain background operation was cancelled");
7314 : }
7315 0 : Err(err) => {
7316 0 : tracing::error!("Drain background operation encountered: {err}")
7317 : }
7318 : }
7319 0 : }
7320 0 : }.instrument(span));
7321 0 : }
7322 : NodeSchedulingPolicy::Draining => {
7323 0 : return Err(ApiError::Conflict(format!(
7324 0 : "Node {node_id} has drain in progress"
7325 0 : )));
7326 : }
7327 0 : policy => {
7328 0 : return Err(ApiError::PreconditionFailed(
7329 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
7330 0 : ));
7331 : }
7332 : }
7333 :
7334 0 : Ok(())
7335 0 : }
7336 :
7337 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
7338 0 : let node_available = {
7339 0 : let locked = self.inner.read().unwrap();
7340 0 : let nodes = &locked.nodes;
7341 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7342 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7343 0 : ))?;
7344 :
7345 0 : node.is_available()
7346 0 : };
7347 0 :
7348 0 : if !node_available {
7349 0 : return Err(ApiError::ResourceUnavailable(
7350 0 : format!("Node {node_id} is currently unavailable").into(),
7351 0 : ));
7352 0 : }
7353 :
7354 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
7355 0 : if let Operation::Drain(drain) = op_handler.operation {
7356 0 : if drain.node_id == node_id {
7357 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
7358 0 : op_handler.cancel.cancel();
7359 0 : return Ok(());
7360 0 : }
7361 0 : }
7362 0 : }
7363 :
7364 0 : Err(ApiError::PreconditionFailed(
7365 0 : format!("Node {node_id} has no drain in progress").into(),
7366 0 : ))
7367 0 : }
7368 :
7369 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
7370 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
7371 0 : let locked = self.inner.read().unwrap();
7372 0 : let nodes = &locked.nodes;
7373 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7374 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7375 0 : ))?;
7376 :
7377 0 : (
7378 0 : locked
7379 0 : .ongoing_operation
7380 0 : .as_ref()
7381 0 : .map(|ongoing| ongoing.operation),
7382 0 : node.is_available(),
7383 0 : node.get_scheduling(),
7384 0 : nodes.len(),
7385 0 : )
7386 0 : };
7387 :
7388 0 : if let Some(ongoing) = ongoing_op {
7389 0 : return Err(ApiError::PreconditionFailed(
7390 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
7391 0 : ));
7392 0 : }
7393 0 :
7394 0 : if !node_available {
7395 0 : return Err(ApiError::ResourceUnavailable(
7396 0 : format!("Node {node_id} is currently unavailable").into(),
7397 0 : ));
7398 0 : }
7399 0 :
7400 0 : if total_nodes_count <= 1 {
7401 0 : return Err(ApiError::PreconditionFailed(
7402 0 : "No other nodes to fill from".into(),
7403 0 : ));
7404 0 : }
7405 0 :
7406 0 : match node_policy {
7407 : NodeSchedulingPolicy::Active => {
7408 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
7409 0 : .await?;
7410 :
7411 0 : let cancel = self.cancel.child_token();
7412 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
7413 :
7414 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
7415 0 : operation: Operation::Fill(Fill { node_id }),
7416 0 : cancel: cancel.clone(),
7417 0 : });
7418 :
7419 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
7420 :
7421 0 : tokio::task::spawn({
7422 0 : let service = self.clone();
7423 0 : let cancel = cancel.clone();
7424 0 : async move {
7425 0 : let _gate_guard = gate_guard;
7426 0 :
7427 0 : scopeguard::defer! {
7428 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
7429 0 :
7430 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
7431 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
7432 0 : } else {
7433 0 : panic!("We always remove the same operation")
7434 0 : }
7435 0 : }
7436 0 :
7437 0 : tracing::info!("Fill background operation starting");
7438 0 : let res = service.fill_node(node_id, cancel).await;
7439 0 : match res {
7440 : Ok(()) => {
7441 0 : tracing::info!("Fill background operation completed successfully");
7442 : }
7443 : Err(OperationError::Cancelled) => {
7444 0 : tracing::info!("Fill background operation was cancelled");
7445 : }
7446 0 : Err(err) => {
7447 0 : tracing::error!("Fill background operation encountered: {err}")
7448 : }
7449 : }
7450 0 : }
7451 0 : }.instrument(span));
7452 0 : }
7453 : NodeSchedulingPolicy::Filling => {
7454 0 : return Err(ApiError::Conflict(format!(
7455 0 : "Node {node_id} has fill in progress"
7456 0 : )));
7457 : }
7458 0 : policy => {
7459 0 : return Err(ApiError::PreconditionFailed(
7460 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
7461 0 : ));
7462 : }
7463 : }
7464 :
7465 0 : Ok(())
7466 0 : }
7467 :
7468 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
7469 0 : let node_available = {
7470 0 : let locked = self.inner.read().unwrap();
7471 0 : let nodes = &locked.nodes;
7472 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7473 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7474 0 : ))?;
7475 :
7476 0 : node.is_available()
7477 0 : };
7478 0 :
7479 0 : if !node_available {
7480 0 : return Err(ApiError::ResourceUnavailable(
7481 0 : format!("Node {node_id} is currently unavailable").into(),
7482 0 : ));
7483 0 : }
7484 :
7485 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
7486 0 : if let Operation::Fill(fill) = op_handler.operation {
7487 0 : if fill.node_id == node_id {
7488 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
7489 0 : op_handler.cancel.cancel();
7490 0 : return Ok(());
7491 0 : }
7492 0 : }
7493 0 : }
7494 :
7495 0 : Err(ApiError::PreconditionFailed(
7496 0 : format!("Node {node_id} has no fill in progress").into(),
7497 0 : ))
7498 0 : }
7499 :
7500 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
7501 : /// configuration
7502 0 : fn maybe_reconcile_shard(
7503 0 : &self,
7504 0 : shard: &mut TenantShard,
7505 0 : nodes: &Arc<HashMap<NodeId, Node>>,
7506 0 : priority: ReconcilerPriority,
7507 0 : ) -> Option<ReconcilerWaiter> {
7508 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
7509 0 : }
7510 :
7511 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
7512 0 : fn get_reconciler_units(
7513 0 : &self,
7514 0 : priority: ReconcilerPriority,
7515 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
7516 0 : let units = match priority {
7517 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
7518 : ReconcilerPriority::High => {
7519 0 : match self
7520 0 : .priority_reconciler_concurrency
7521 0 : .clone()
7522 0 : .try_acquire_owned()
7523 : {
7524 0 : Ok(u) => Ok(u),
7525 : Err(TryAcquireError::NoPermits) => {
7526 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
7527 : // the normal priority semaphore.
7528 0 : self.reconciler_concurrency.clone().try_acquire_owned()
7529 : }
7530 0 : Err(e) => Err(e),
7531 : }
7532 : }
7533 : };
7534 :
7535 0 : units.map(ReconcileUnits::new)
7536 0 : }
7537 :
7538 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
7539 0 : fn maybe_configured_reconcile_shard(
7540 0 : &self,
7541 0 : shard: &mut TenantShard,
7542 0 : nodes: &Arc<HashMap<NodeId, Node>>,
7543 0 : reconciler_config: ReconcilerConfig,
7544 0 : ) -> Option<ReconcilerWaiter> {
7545 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
7546 :
7547 0 : let reconcile_reason = match reconcile_needed {
7548 0 : ReconcileNeeded::No => return None,
7549 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
7550 0 : ReconcileNeeded::Yes(reason) => {
7551 0 : // Fall through to try and acquire units for spawning reconciler
7552 0 : reason
7553 : }
7554 : };
7555 :
7556 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
7557 0 : Ok(u) => u,
7558 : Err(_) => {
7559 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
7560 0 : "Concurrency limited: enqueued for reconcile later");
7561 0 : if !shard.delayed_reconcile {
7562 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
7563 0 : Err(TrySendError::Closed(_)) => {
7564 0 : // Weird mid-shutdown case?
7565 0 : }
7566 : Err(TrySendError::Full(_)) => {
7567 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
7568 0 : tracing::warn!(
7569 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
7570 : );
7571 : }
7572 0 : Ok(()) => {
7573 0 : shard.delayed_reconcile = true;
7574 0 : }
7575 : }
7576 0 : }
7577 :
7578 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
7579 : // number to advance. When this function is eventually called again and succeeds in getting units,
7580 : // it will spawn a reconciler that makes this waiter complete.
7581 0 : return Some(shard.future_reconcile_waiter());
7582 : }
7583 : };
7584 :
7585 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
7586 : // Gate closed: we're shutting down, drop out.
7587 0 : return None;
7588 : };
7589 :
7590 0 : shard.spawn_reconciler(
7591 0 : reconcile_reason,
7592 0 : &self.result_tx,
7593 0 : nodes,
7594 0 : &self.compute_hook,
7595 0 : reconciler_config,
7596 0 : &self.config,
7597 0 : &self.persistence,
7598 0 : units,
7599 0 : gate_guard,
7600 0 : &self.reconcilers_cancel,
7601 0 : self.http_client.clone(),
7602 0 : )
7603 0 : }
7604 :
7605 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
7606 : /// Additionally, reschedule tenants that require it.
7607 : ///
7608 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
7609 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
7610 : /// available. A return value of 0 indicates that everything is fully reconciled already.
7611 0 : fn reconcile_all(&self) -> usize {
7612 0 : let mut locked = self.inner.write().unwrap();
7613 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7614 0 : let pageservers = nodes.clone();
7615 0 :
7616 0 : // This function is an efficient place to update lazy statistics, since we are walking
7617 0 : // all tenants.
7618 0 : let mut pending_reconciles = 0;
7619 0 : let mut az_violations = 0;
7620 0 :
7621 0 : // If we find any tenants to drop from memory, stash them to offload after
7622 0 : // we're done traversing the map of tenants.
7623 0 : let mut drop_detached_tenants = Vec::new();
7624 0 :
7625 0 : let mut reconciles_spawned = 0;
7626 0 : for shard in tenants.values_mut() {
7627 : // Accumulate scheduling statistics
7628 0 : if let (Some(attached), Some(preferred)) =
7629 0 : (shard.intent.get_attached(), shard.preferred_az())
7630 : {
7631 0 : let node_az = nodes
7632 0 : .get(attached)
7633 0 : .expect("Nodes exist if referenced")
7634 0 : .get_availability_zone_id();
7635 0 : if node_az != preferred {
7636 0 : az_violations += 1;
7637 0 : }
7638 0 : }
7639 :
7640 : // Skip checking if this shard is already enqueued for reconciliation
7641 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
7642 : // If there is something delayed, then return a nonzero count so that
7643 : // callers like reconcile_all_now do not incorrectly get the impression
7644 : // that the system is in a quiescent state.
7645 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
7646 0 : pending_reconciles += 1;
7647 0 : continue;
7648 0 : }
7649 0 :
7650 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
7651 0 : // dirty, spawn another one
7652 0 : if self
7653 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
7654 0 : .is_some()
7655 0 : {
7656 0 : reconciles_spawned += 1;
7657 0 : } else if shard.delayed_reconcile {
7658 0 : // Shard wanted to reconcile but for some reason couldn't.
7659 0 : pending_reconciles += 1;
7660 0 : }
7661 :
7662 : // If this tenant is detached, try dropping it from memory. This is usually done
7663 : // proactively in [`Self::process_results`], but we do it here to handle the edge
7664 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
7665 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
7666 0 : && shard.policy == PlacementPolicy::Detached
7667 : {
7668 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
7669 0 : shard.tenant_shard_id.tenant_id,
7670 0 : TenantOperations::DropDetached,
7671 0 : ) {
7672 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
7673 0 : }
7674 0 : }
7675 : }
7676 :
7677 : // Some metrics are calculated from SchedulerNode state, update these periodically
7678 0 : scheduler.update_metrics();
7679 :
7680 : // Process any deferred tenant drops
7681 0 : for (tenant_id, guard) in drop_detached_tenants {
7682 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
7683 0 : }
7684 :
7685 0 : metrics::METRICS_REGISTRY
7686 0 : .metrics_group
7687 0 : .storage_controller_schedule_az_violation
7688 0 : .set(az_violations as i64);
7689 0 :
7690 0 : metrics::METRICS_REGISTRY
7691 0 : .metrics_group
7692 0 : .storage_controller_pending_reconciles
7693 0 : .set(pending_reconciles as i64);
7694 0 :
7695 0 : reconciles_spawned
7696 0 : }
7697 :
7698 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
7699 : /// could be scheduled somewhere better:
7700 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
7701 : /// * e.g. after a node fails then recovers, to move some work back to it
7702 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
7703 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
7704 : /// we did the split, but are probably better placed elsewhere.
7705 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
7706 : /// * e.g. after a shard split, some locations will be on the same node (where the split
7707 : /// happened), and will probably be better placed elsewhere.
7708 : ///
7709 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
7710 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
7711 : /// according to those same soft constraints.
7712 0 : async fn optimize_all(&self) -> usize {
7713 : // Limit on how many shards' optmizations each call to this function will execute. Combined
7714 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
7715 : // trickle of optimizations in the background, rather than executing a large number in parallel
7716 : // when a change occurs.
7717 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
7718 :
7719 : // Synchronous prepare: scan shards for possible scheduling optimizations
7720 0 : let candidate_work = self.optimize_all_plan();
7721 0 : let candidate_work_len = candidate_work.len();
7722 :
7723 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
7724 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
7725 :
7726 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
7727 0 :
7728 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
7729 0 : let mut reconciles_spawned = 0;
7730 0 : let mut optimizations_applied = 0;
7731 0 : let mut locked = self.inner.write().unwrap();
7732 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7733 0 : for (tenant_shard_id, optimization) in validated_work {
7734 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
7735 : // Shard was dropped between planning and execution;
7736 0 : continue;
7737 : };
7738 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
7739 0 : if shard.apply_optimization(scheduler, optimization) {
7740 0 : optimizations_applied += 1;
7741 0 : if self
7742 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
7743 0 : .is_some()
7744 0 : {
7745 0 : reconciles_spawned += 1;
7746 0 : }
7747 0 : }
7748 :
7749 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
7750 0 : break;
7751 0 : }
7752 : }
7753 :
7754 0 : if was_work_filtered {
7755 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
7756 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
7757 0 : // as these validations start passing.
7758 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
7759 0 : }
7760 :
7761 0 : reconciles_spawned
7762 0 : }
7763 :
7764 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
7765 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
7766 : // this higher than the execution limit gives us a chance to execute some work even if the first
7767 : // few optimizations we find are not ready.
7768 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
7769 :
7770 0 : let mut work = Vec::new();
7771 0 : let mut locked = self.inner.write().unwrap();
7772 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
7773 :
7774 : // We are going to plan a bunch of optimisations before applying any of them, so the
7775 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
7776 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
7777 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
7778 : // to ignore the utilisation component of the score.
7779 :
7780 0 : for (_tenant_id, schedule_context, shards) in
7781 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
7782 : {
7783 0 : for shard in shards {
7784 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
7785 0 : break;
7786 0 : }
7787 0 : match shard.get_scheduling_policy() {
7788 0 : ShardSchedulingPolicy::Active => {
7789 0 : // Ok to do optimization
7790 0 : }
7791 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
7792 0 : // Ok to do optimization: we are executing a graceful migration that
7793 0 : // has set preferred_node
7794 0 : }
7795 : ShardSchedulingPolicy::Essential
7796 : | ShardSchedulingPolicy::Pause
7797 : | ShardSchedulingPolicy::Stop => {
7798 : // Policy prevents optimizing this shard.
7799 0 : continue;
7800 : }
7801 : }
7802 :
7803 0 : if !matches!(shard.splitting, SplitState::Idle)
7804 0 : || matches!(shard.policy, PlacementPolicy::Detached)
7805 0 : || shard.reconciler.is_some()
7806 : {
7807 : // Do not start any optimizations while another change to the tenant is ongoing: this
7808 : // is not necessary for correctness, but simplifies operations and implicitly throttles
7809 : // optimization changes to happen in a "trickle" over time.
7810 0 : continue;
7811 0 : }
7812 0 :
7813 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
7814 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
7815 0 : if cfg!(feature = "testing") {
7816 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
7817 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
7818 : // panic in prod if we hit this, or spend cycles on it in prod.
7819 0 : assert!(
7820 0 : shard
7821 0 : .optimize_attachment(scheduler, &schedule_context)
7822 0 : .is_none()
7823 0 : );
7824 0 : assert!(
7825 0 : shard
7826 0 : .optimize_secondary(scheduler, &schedule_context)
7827 0 : .is_none()
7828 0 : );
7829 0 : }
7830 0 : continue;
7831 0 : }
7832 :
7833 0 : if let Some(optimization) =
7834 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
7835 : // its primary location based on soft constraints, cut it over.
7836 0 : shard.optimize_attachment(scheduler, &schedule_context)
7837 : {
7838 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
7839 0 : work.push((shard.tenant_shard_id, optimization));
7840 0 : break;
7841 0 : } else if let Some(optimization) =
7842 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
7843 : // better placed on another node, based on ScheduleContext, then adjust it. This
7844 : // covers cases like after a shard split, where we might have too many shards
7845 : // in the same tenant with secondary locations on the node where they originally split.
7846 0 : shard.optimize_secondary(scheduler, &schedule_context)
7847 : {
7848 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
7849 0 : work.push((shard.tenant_shard_id, optimization));
7850 0 : break;
7851 0 : }
7852 : }
7853 : }
7854 :
7855 0 : work
7856 0 : }
7857 :
7858 0 : async fn optimize_all_validate(
7859 0 : &self,
7860 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
7861 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
7862 0 : // Take a clone of the node map to use outside the lock in async validation phase
7863 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
7864 0 :
7865 0 : let mut want_secondary_status = Vec::new();
7866 0 :
7867 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
7868 0 : // check that the state of locations is acceptable to run the optimization, such as
7869 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
7870 0 : // in a live migration.
7871 0 : let mut validated_work = Vec::new();
7872 0 : for (tenant_shard_id, optimization) in candidate_work {
7873 0 : match optimization.action {
7874 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
7875 : old_attached_node_id: _,
7876 0 : new_attached_node_id,
7877 0 : }) => {
7878 0 : match validation_nodes.get(&new_attached_node_id) {
7879 0 : None => {
7880 0 : // Node was dropped between planning and validation
7881 0 : }
7882 0 : Some(node) => {
7883 0 : if !node.is_available() {
7884 0 : tracing::info!(
7885 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
7886 : );
7887 0 : } else {
7888 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
7889 0 : // remote API requests concurrently.
7890 0 : want_secondary_status.push((
7891 0 : tenant_shard_id,
7892 0 : node.clone(),
7893 0 : optimization,
7894 0 : ));
7895 0 : }
7896 : }
7897 : }
7898 : }
7899 : ScheduleOptimizationAction::ReplaceSecondary(_)
7900 : | ScheduleOptimizationAction::CreateSecondary(_)
7901 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
7902 : // No extra checks needed to manage secondaries: this does not interrupt client access
7903 0 : validated_work.push((tenant_shard_id, optimization))
7904 : }
7905 : };
7906 : }
7907 :
7908 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
7909 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
7910 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
7911 0 : let results = self
7912 0 : .tenant_for_shards_api(
7913 0 : want_secondary_status
7914 0 : .iter()
7915 0 : .map(|i| (i.0, i.1.clone()))
7916 0 : .collect(),
7917 0 : |tenant_shard_id, client| async move {
7918 0 : client.tenant_secondary_status(tenant_shard_id).await
7919 0 : },
7920 0 : 1,
7921 0 : 1,
7922 0 : SHORT_RECONCILE_TIMEOUT,
7923 0 : &self.cancel,
7924 0 : )
7925 0 : .await;
7926 :
7927 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
7928 0 : want_secondary_status.into_iter().zip(results.into_iter())
7929 : {
7930 0 : match secondary_status {
7931 0 : Err(e) => {
7932 0 : tracing::info!(
7933 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
7934 : );
7935 : }
7936 0 : Ok(progress) => {
7937 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
7938 : // them in an optimization
7939 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
7940 :
7941 0 : if progress.heatmap_mtime.is_none()
7942 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
7943 0 : && progress.bytes_downloaded != progress.bytes_total
7944 0 : || progress.bytes_total - progress.bytes_downloaded
7945 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
7946 : {
7947 0 : tracing::info!(
7948 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
7949 : );
7950 :
7951 : #[cfg(feature = "testing")]
7952 0 : if progress.heatmap_mtime.is_none() {
7953 : // No heatmap might mean the attached location has never uploaded one, or that
7954 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
7955 : // but fairly common in tests.
7956 0 : self.kick_secondary_download(tenant_shard_id).await;
7957 0 : }
7958 : } else {
7959 : // Location looks ready: proceed
7960 0 : tracing::info!(
7961 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
7962 : );
7963 0 : validated_work.push((tenant_shard_id, optimization))
7964 : }
7965 : }
7966 : }
7967 : }
7968 :
7969 0 : validated_work
7970 0 : }
7971 :
7972 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
7973 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
7974 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
7975 : /// we have this helper to move things along faster.
7976 : #[cfg(feature = "testing")]
7977 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
7978 0 : let (attached_node, secondaries) = {
7979 0 : let locked = self.inner.read().unwrap();
7980 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
7981 0 : tracing::warn!(
7982 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
7983 : );
7984 0 : return;
7985 : };
7986 :
7987 0 : let Some(attached) = shard.intent.get_attached() else {
7988 0 : tracing::warn!(
7989 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
7990 : );
7991 0 : return;
7992 : };
7993 :
7994 0 : let secondaries = shard
7995 0 : .intent
7996 0 : .get_secondary()
7997 0 : .iter()
7998 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
7999 0 : .collect::<Vec<_>>();
8000 0 :
8001 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
8002 0 : };
8003 0 :
8004 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
8005 0 : // a 'kick' to let scheduling optimisation run more promptly.
8006 0 : match attached_node
8007 0 : .with_client_retries(
8008 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
8009 0 : &self.http_client,
8010 0 : &self.config.pageserver_jwt_token,
8011 0 : 3,
8012 0 : 10,
8013 0 : SHORT_RECONCILE_TIMEOUT,
8014 0 : &self.cancel,
8015 0 : )
8016 0 : .await
8017 : {
8018 0 : Some(Err(e)) => {
8019 0 : tracing::info!(
8020 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
8021 : );
8022 : }
8023 : None => {
8024 0 : tracing::info!(
8025 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
8026 : );
8027 : }
8028 : Some(Ok(_)) => {
8029 0 : tracing::info!(
8030 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
8031 : );
8032 : }
8033 : }
8034 :
8035 0 : for secondary_node in secondaries {
8036 0 : match secondary_node
8037 0 : .with_client_retries(
8038 0 : |client| async move {
8039 0 : client
8040 0 : .tenant_secondary_download(
8041 0 : tenant_shard_id,
8042 0 : Some(Duration::from_secs(1)),
8043 0 : )
8044 0 : .await
8045 0 : },
8046 0 : &self.http_client,
8047 0 : &self.config.pageserver_jwt_token,
8048 0 : 3,
8049 0 : 10,
8050 0 : SHORT_RECONCILE_TIMEOUT,
8051 0 : &self.cancel,
8052 0 : )
8053 0 : .await
8054 : {
8055 0 : Some(Err(e)) => {
8056 0 : tracing::info!(
8057 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
8058 : );
8059 : }
8060 : None => {
8061 0 : tracing::info!(
8062 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
8063 : );
8064 : }
8065 0 : Some(Ok(progress)) => {
8066 0 : tracing::info!(
8067 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
8068 : );
8069 : }
8070 : }
8071 : }
8072 0 : }
8073 :
8074 : /// Asynchronously split a tenant that's eligible for automatic splits. At most one tenant will
8075 : /// be split per call.
8076 : ///
8077 : /// Two sets of criteria are used: initial splits and size-based splits (in that order).
8078 : /// Initial splits are used to eagerly split unsharded tenants that may be performing initial
8079 : /// ingestion, since sharded tenants have significantly better ingestion throughput. Size-based
8080 : /// splits are used to bound the maximum shard size and balance out load.
8081 : ///
8082 : /// Splits are based on max_logical_size, i.e. the logical size of the largest timeline in a
8083 : /// tenant. We use this instead of the total logical size because branches will duplicate
8084 : /// logical size without actually using more storage. We could also use visible physical size,
8085 : /// but this might overestimate tenants that frequently churn branches.
8086 : ///
8087 : /// Initial splits (initial_split_threshold):
8088 : /// * Applies to tenants with 1 shard.
8089 : /// * The largest timeline (max_logical_size) exceeds initial_split_threshold.
8090 : /// * Splits into initial_split_shards.
8091 : ///
8092 : /// Size-based splits (split_threshold):
8093 : /// * Applies to all tenants.
8094 : /// * The largest timeline (max_logical_size) divided by shard count exceeds split_threshold.
8095 : /// * Splits such that max_logical_size / shard_count <= split_threshold, in powers of 2.
8096 : ///
8097 : /// Tenant shards are ordered by descending max_logical_size, first initial split candidates
8098 : /// then size-based split candidates. The first matching candidate is split.
8099 : ///
8100 : /// The shard count is clamped to max_split_shards. If a candidate is eligible for both initial
8101 : /// and size-based splits, the largest shard count will be used.
8102 : ///
8103 : /// An unsharded tenant will get DEFAULT_STRIPE_SIZE, regardless of what its ShardIdentity says.
8104 : /// A sharded tenant will retain its stripe size, as splits do not allow changing it.
8105 : ///
8106 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
8107 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
8108 : /// block all other splits.
8109 0 : async fn autosplit_tenants(self: &Arc<Self>) {
8110 0 : // If max_split_shards is set to 0 or 1, we can't split.
8111 0 : let max_split_shards = self.config.max_split_shards;
8112 0 : if max_split_shards <= 1 {
8113 0 : return;
8114 0 : }
8115 0 :
8116 0 : // If initial_split_shards is set to 0 or 1, disable initial splits.
8117 0 : let mut initial_split_threshold = self.config.initial_split_threshold.unwrap_or(0);
8118 0 : let initial_split_shards = self.config.initial_split_shards;
8119 0 : if initial_split_shards <= 1 {
8120 0 : initial_split_threshold = 0;
8121 0 : }
8122 :
8123 : // If no split_threshold nor initial_split_threshold, disable autosplits.
8124 0 : let split_threshold = self.config.split_threshold.unwrap_or(0);
8125 0 : if split_threshold == 0 && initial_split_threshold == 0 {
8126 0 : return;
8127 0 : }
8128 0 :
8129 0 : // Fetch split candidates in prioritized order.
8130 0 : //
8131 0 : // If initial splits are enabled, fetch eligible tenants first. We prioritize initial splits
8132 0 : // over size-based splits, since these are often performing initial ingestion and rely on
8133 0 : // splits to improve ingest throughput.
8134 0 : let mut candidates = Vec::new();
8135 0 :
8136 0 : if initial_split_threshold > 0 {
8137 : // Initial splits: fetch tenants with 1 shard where the logical size of the largest
8138 : // timeline exceeds the initial split threshold.
8139 0 : let initial_candidates = self
8140 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
8141 0 : order_by: TenantSorting::MaxLogicalSize,
8142 0 : limit: 10,
8143 0 : where_shards_lt: Some(ShardCount(2)),
8144 0 : where_gt: Some(initial_split_threshold),
8145 0 : })
8146 0 : .await;
8147 0 : candidates.extend(initial_candidates);
8148 0 : }
8149 :
8150 0 : if split_threshold > 0 {
8151 : // Size-based splits: fetch tenants where the logical size of the largest timeline
8152 : // divided by shard count exceeds the split threshold.
8153 : //
8154 : // max_logical_size is only tracked on shard 0, and contains the total logical size
8155 : // across all shards. We have to order and filter by MaxLogicalSizePerShard, i.e.
8156 : // max_logical_size / shard_count, such that we only receive tenants that are actually
8157 : // eligible for splits. But we still use max_logical_size for later split calculations.
8158 0 : let size_candidates = self
8159 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
8160 0 : order_by: TenantSorting::MaxLogicalSizePerShard,
8161 0 : limit: 10,
8162 0 : where_shards_lt: Some(ShardCount(max_split_shards)),
8163 0 : where_gt: Some(split_threshold),
8164 0 : })
8165 0 : .await;
8166 : #[cfg(feature = "testing")]
8167 0 : assert!(
8168 0 : size_candidates.iter().all(|c| c.id.is_shard_zero()),
8169 0 : "MaxLogicalSizePerShard returned non-zero shard: {size_candidates:?}",
8170 : );
8171 0 : candidates.extend(size_candidates);
8172 0 : }
8173 :
8174 : // Filter out tenants in a prohibiting scheduling modes
8175 : // and tenants with an ongoing import.
8176 : //
8177 : // Note that the import check here is oportunistic. An import might start
8178 : // after the check before we actually update [`TenantShard::splitting`].
8179 : // [`Self::tenant_shard_split`] checks the database whilst holding the exclusive
8180 : // tenant lock. Imports might take a long time, so the check here allows us
8181 : // to split something else instead of trying the same shard over and over.
8182 0 : {
8183 0 : let state = self.inner.read().unwrap();
8184 0 : candidates.retain(|i| {
8185 0 : let shard = state.tenants.get(&i.id);
8186 0 : match shard {
8187 0 : Some(t) => {
8188 0 : t.get_scheduling_policy() == ShardSchedulingPolicy::Active
8189 0 : && t.importing == TimelineImportState::Idle
8190 : }
8191 0 : None => false,
8192 : }
8193 0 : });
8194 0 : }
8195 :
8196 : // Pick the first candidate to split. This will generally always be the first one in
8197 : // candidates, but we defensively skip candidates that end up not actually splitting.
8198 0 : let Some((candidate, new_shard_count)) = candidates
8199 0 : .into_iter()
8200 0 : .filter_map(|candidate| {
8201 0 : let new_shard_count = Self::compute_split_shards(ShardSplitInputs {
8202 0 : shard_count: candidate.id.shard_count,
8203 0 : max_logical_size: candidate.max_logical_size,
8204 0 : split_threshold,
8205 0 : max_split_shards,
8206 0 : initial_split_threshold,
8207 0 : initial_split_shards,
8208 0 : });
8209 0 : new_shard_count.map(|shards| (candidate, shards.count()))
8210 0 : })
8211 0 : .next()
8212 : else {
8213 0 : debug!("no split-eligible tenants found");
8214 0 : return;
8215 : };
8216 :
8217 : // Retain the stripe size of sharded tenants, as splits don't allow changing it. Otherwise,
8218 : // use DEFAULT_STRIPE_SIZE for unsharded tenants -- their stripe size doesn't really matter,
8219 : // and if we change the default stripe size we want to use the new default rather than an
8220 : // old, persisted stripe size.
8221 0 : let new_stripe_size = match candidate.id.shard_count.count() {
8222 0 : 0 => panic!("invalid shard count 0"),
8223 0 : 1 => Some(DEFAULT_STRIPE_SIZE),
8224 0 : 2.. => None,
8225 : };
8226 :
8227 : // We spawn a task to run this, so it's exactly like some external API client requesting
8228 : // it. We don't want to block the background reconcile loop on this.
8229 0 : let old_shard_count = candidate.id.shard_count.count();
8230 0 : info!(
8231 0 : "auto-splitting tenant {old_shard_count} → {new_shard_count} shards, \
8232 0 : current size {candidate:?} (split_threshold={split_threshold} \
8233 0 : initial_split_threshold={initial_split_threshold})"
8234 : );
8235 :
8236 0 : let this = self.clone();
8237 0 : tokio::spawn(
8238 0 : async move {
8239 0 : match this
8240 0 : .tenant_shard_split(
8241 0 : candidate.id.tenant_id,
8242 0 : TenantShardSplitRequest {
8243 0 : new_shard_count,
8244 0 : new_stripe_size,
8245 0 : },
8246 0 : )
8247 0 : .await
8248 : {
8249 : Ok(_) => {
8250 0 : info!("successful auto-split {old_shard_count} → {new_shard_count} shards")
8251 : }
8252 0 : Err(err) => error!("auto-split failed: {err}"),
8253 : }
8254 0 : }
8255 0 : .instrument(info_span!("auto_split", tenant_id=%candidate.id.tenant_id)),
8256 : );
8257 0 : }
8258 :
8259 : /// Returns the number of shards to split a tenant into, or None if the tenant shouldn't split,
8260 : /// based on the total logical size of the largest timeline summed across all shards. Uses the
8261 : /// larger of size-based and initial splits, clamped to max_split_shards.
8262 : ///
8263 : /// NB: the thresholds are exclusive, since TopTenantShardsRequest uses where_gt.
8264 25 : fn compute_split_shards(inputs: ShardSplitInputs) -> Option<ShardCount> {
8265 25 : let ShardSplitInputs {
8266 25 : shard_count,
8267 25 : max_logical_size,
8268 25 : split_threshold,
8269 25 : max_split_shards,
8270 25 : initial_split_threshold,
8271 25 : initial_split_shards,
8272 25 : } = inputs;
8273 25 :
8274 25 : let mut new_shard_count: u8 = shard_count.count();
8275 25 :
8276 25 : // Size-based splits. Ensures max_logical_size / new_shard_count <= split_threshold, using
8277 25 : // power-of-two shard counts.
8278 25 : //
8279 25 : // If the current shard count is not a power of two, and does not exceed split_threshold,
8280 25 : // then we leave it alone rather than forcing a power-of-two split.
8281 25 : if split_threshold > 0
8282 18 : && max_logical_size.div_ceil(split_threshold) > shard_count.count() as u64
8283 12 : {
8284 12 : new_shard_count = max_logical_size
8285 12 : .div_ceil(split_threshold)
8286 12 : .checked_next_power_of_two()
8287 12 : .unwrap_or(u8::MAX as u64)
8288 12 : .try_into()
8289 12 : .unwrap_or(u8::MAX);
8290 13 : }
8291 :
8292 : // Initial splits. Use the larger of size-based and initial split shard counts. This only
8293 : // applies to unsharded tenants, i.e. changes to initial_split_threshold or
8294 : // initial_split_shards are not retroactive for sharded tenants.
8295 25 : if initial_split_threshold > 0
8296 14 : && shard_count.count() <= 1
8297 11 : && max_logical_size > initial_split_threshold
8298 8 : {
8299 8 : new_shard_count = new_shard_count.max(initial_split_shards);
8300 17 : }
8301 :
8302 : // Clamp to max shards.
8303 25 : new_shard_count = new_shard_count.min(max_split_shards);
8304 25 :
8305 25 : // Don't split if we're not increasing the shard count.
8306 25 : if new_shard_count <= shard_count.count() {
8307 10 : return None;
8308 15 : }
8309 15 :
8310 15 : Some(ShardCount(new_shard_count))
8311 25 : }
8312 :
8313 : /// Fetches the top tenant shards from every node, in descending order of
8314 : /// max logical size. Any node errors will be logged and ignored.
8315 0 : async fn get_top_tenant_shards(
8316 0 : &self,
8317 0 : request: &TopTenantShardsRequest,
8318 0 : ) -> Vec<TopTenantShardItem> {
8319 0 : let nodes = self
8320 0 : .inner
8321 0 : .read()
8322 0 : .unwrap()
8323 0 : .nodes
8324 0 : .values()
8325 0 : .cloned()
8326 0 : .collect_vec();
8327 0 :
8328 0 : let mut futures = FuturesUnordered::new();
8329 0 : for node in nodes {
8330 0 : futures.push(async move {
8331 0 : node.with_client_retries(
8332 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
8333 0 : &self.http_client,
8334 0 : &self.config.pageserver_jwt_token,
8335 0 : 3,
8336 0 : 3,
8337 0 : Duration::from_secs(5),
8338 0 : &self.cancel,
8339 0 : )
8340 0 : .await
8341 0 : });
8342 0 : }
8343 :
8344 0 : let mut top = Vec::new();
8345 0 : while let Some(output) = futures.next().await {
8346 0 : match output {
8347 0 : Some(Ok(response)) => top.extend(response.shards),
8348 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
8349 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
8350 0 : None => {} // node is shutting down
8351 : }
8352 : }
8353 :
8354 0 : top.sort_by_key(|i| i.max_logical_size);
8355 0 : top.reverse();
8356 0 : top
8357 0 : }
8358 :
8359 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
8360 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
8361 : /// put the system into a quiescent state where future background reconciliations won't do anything.
8362 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
8363 0 : let reconciles_spawned = self.reconcile_all();
8364 0 : let reconciles_spawned = if reconciles_spawned == 0 {
8365 : // Only optimize when we are otherwise idle
8366 0 : self.optimize_all().await
8367 : } else {
8368 0 : reconciles_spawned
8369 : };
8370 :
8371 0 : let waiters = {
8372 0 : let mut waiters = Vec::new();
8373 0 : let locked = self.inner.read().unwrap();
8374 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
8375 0 : if let Some(waiter) = shard.get_waiter() {
8376 0 : waiters.push(waiter);
8377 0 : }
8378 : }
8379 0 : waiters
8380 0 : };
8381 0 :
8382 0 : let waiter_count = waiters.len();
8383 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
8384 0 : Ok(()) => {}
8385 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
8386 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
8387 0 : {
8388 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
8389 0 : // because some other change superceded it. We will return a nonzero number,
8390 0 : // so the caller knows they might have to call again to quiesce the system.
8391 0 : }
8392 0 : Err(e) => {
8393 0 : return Err(e);
8394 : }
8395 : };
8396 :
8397 0 : tracing::info!(
8398 0 : "{} reconciles in reconcile_all, {} waiters",
8399 : reconciles_spawned,
8400 : waiter_count
8401 : );
8402 :
8403 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
8404 0 : }
8405 :
8406 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
8407 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
8408 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
8409 0 : self.reconcilers_cancel.cancel();
8410 0 : self.reconcilers_gate.close().await;
8411 :
8412 : // Signal the background loop in [`Service::process_results`] to exit once
8413 : // it has proccessed the results from all the reconciles we cancelled earlier.
8414 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
8415 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
8416 0 : self.result_tx.closed().await;
8417 0 : }
8418 :
8419 0 : pub async fn shutdown(&self) {
8420 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
8421 0 : .await;
8422 :
8423 : // Background tasks hold gate guards: this notifies them of the cancellation and
8424 : // waits for them all to complete.
8425 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
8426 0 : self.cancel.cancel();
8427 0 : self.gate.close().await;
8428 0 : }
8429 :
8430 : /// Spot check the download lag for a secondary location of a shard.
8431 : /// Should be used as a heuristic, since it's not always precise: the
8432 : /// secondary might have not downloaded the new heat map yet and, hence,
8433 : /// is not aware of the lag.
8434 : ///
8435 : /// Returns:
8436 : /// * Ok(None) if the lag could not be determined from the status,
8437 : /// * Ok(Some(_)) if the lag could be determind
8438 : /// * Err on failures to query the pageserver.
8439 0 : async fn secondary_lag(
8440 0 : &self,
8441 0 : secondary: &NodeId,
8442 0 : tenant_shard_id: TenantShardId,
8443 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
8444 0 : let nodes = self.inner.read().unwrap().nodes.clone();
8445 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
8446 0 : StatusCode::NOT_FOUND,
8447 0 : format!("Node with id {} not found", secondary),
8448 0 : ))?;
8449 :
8450 0 : match node
8451 0 : .with_client_retries(
8452 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
8453 0 : &self.http_client,
8454 0 : &self.config.pageserver_jwt_token,
8455 0 : 1,
8456 0 : 3,
8457 0 : Duration::from_millis(250),
8458 0 : &self.cancel,
8459 0 : )
8460 0 : .await
8461 : {
8462 0 : Some(Ok(status)) => match status.heatmap_mtime {
8463 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
8464 0 : None => Ok(None),
8465 : },
8466 0 : Some(Err(e)) => Err(e),
8467 0 : None => Err(mgmt_api::Error::Cancelled),
8468 : }
8469 0 : }
8470 :
8471 : /// Drain a node by moving the shards attached to it as primaries.
8472 : /// This is a long running operation and it should run as a separate Tokio task.
8473 0 : pub(crate) async fn drain_node(
8474 0 : self: &Arc<Self>,
8475 0 : node_id: NodeId,
8476 0 : cancel: CancellationToken,
8477 0 : ) -> Result<(), OperationError> {
8478 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
8479 0 : let max_secondary_lag_bytes = self
8480 0 : .config
8481 0 : .max_secondary_lag_bytes
8482 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
8483 :
8484 : // By default, live migrations are generous about the wait time for getting
8485 : // the secondary location up to speed. When draining, give up earlier in order
8486 : // to not stall the operation when a cold secondary is encountered.
8487 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
8488 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
8489 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
8490 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
8491 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
8492 0 : .build();
8493 0 :
8494 0 : let mut waiters = Vec::new();
8495 0 :
8496 0 : let mut tid_iter = TenantShardIterator::new({
8497 0 : let service = self.clone();
8498 0 : move |last_inspected_shard: Option<TenantShardId>| {
8499 0 : let locked = &service.inner.read().unwrap();
8500 0 : let tenants = &locked.tenants;
8501 0 : let entry = match last_inspected_shard {
8502 0 : Some(skip_past) => {
8503 0 : // Skip to the last seen tenant shard id
8504 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
8505 0 :
8506 0 : // Skip past the last seen
8507 0 : cursor.nth(1)
8508 : }
8509 0 : None => tenants.first_key_value(),
8510 : };
8511 :
8512 0 : entry.map(|(tid, _)| tid).copied()
8513 0 : }
8514 0 : });
8515 :
8516 0 : while !tid_iter.finished() {
8517 0 : if cancel.is_cancelled() {
8518 0 : match self
8519 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8520 0 : .await
8521 : {
8522 0 : Ok(()) => return Err(OperationError::Cancelled),
8523 0 : Err(err) => {
8524 0 : return Err(OperationError::FinalizeError(
8525 0 : format!(
8526 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8527 0 : node_id, err
8528 0 : )
8529 0 : .into(),
8530 0 : ));
8531 : }
8532 : }
8533 0 : }
8534 0 :
8535 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
8536 :
8537 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
8538 0 : let tid = match tid_iter.next() {
8539 0 : Some(tid) => tid,
8540 : None => {
8541 0 : break;
8542 : }
8543 : };
8544 :
8545 0 : let tid_drain = TenantShardDrain {
8546 0 : drained_node: node_id,
8547 0 : tenant_shard_id: tid,
8548 0 : };
8549 :
8550 0 : let dest_node_id = {
8551 0 : let locked = self.inner.read().unwrap();
8552 0 :
8553 0 : match tid_drain
8554 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
8555 : {
8556 0 : Some(node_id) => node_id,
8557 : None => {
8558 0 : continue;
8559 : }
8560 : }
8561 : };
8562 :
8563 0 : match self.secondary_lag(&dest_node_id, tid).await {
8564 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
8565 0 : // The secondary is reasonably up to date.
8566 0 : // Migrate to it
8567 0 : }
8568 0 : Ok(Some(lag)) => {
8569 0 : tracing::info!(
8570 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8571 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
8572 : );
8573 0 : continue;
8574 : }
8575 : Ok(None) => {
8576 0 : tracing::info!(
8577 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8578 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
8579 : );
8580 0 : continue;
8581 : }
8582 0 : Err(err) => {
8583 0 : tracing::warn!(
8584 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8585 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
8586 : );
8587 0 : continue;
8588 : }
8589 : }
8590 :
8591 : {
8592 0 : let mut locked = self.inner.write().unwrap();
8593 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8594 0 : let rescheduled = tid_drain.reschedule_to_secondary(
8595 0 : dest_node_id,
8596 0 : tenants,
8597 0 : scheduler,
8598 0 : nodes,
8599 0 : )?;
8600 :
8601 0 : if let Some(tenant_shard) = rescheduled {
8602 0 : let waiter = self.maybe_configured_reconcile_shard(
8603 0 : tenant_shard,
8604 0 : nodes,
8605 0 : reconciler_config,
8606 0 : );
8607 0 : if let Some(some) = waiter {
8608 0 : waiters.push(some);
8609 0 : }
8610 0 : }
8611 : }
8612 : }
8613 :
8614 0 : waiters = self
8615 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
8616 0 : .await;
8617 :
8618 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
8619 : }
8620 :
8621 0 : while !waiters.is_empty() {
8622 0 : if cancel.is_cancelled() {
8623 0 : match self
8624 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8625 0 : .await
8626 : {
8627 0 : Ok(()) => return Err(OperationError::Cancelled),
8628 0 : Err(err) => {
8629 0 : return Err(OperationError::FinalizeError(
8630 0 : format!(
8631 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8632 0 : node_id, err
8633 0 : )
8634 0 : .into(),
8635 0 : ));
8636 : }
8637 : }
8638 0 : }
8639 0 :
8640 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
8641 :
8642 0 : waiters = self
8643 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
8644 0 : .await;
8645 : }
8646 :
8647 : // At this point we have done the best we could to drain shards from this node.
8648 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
8649 : // to complete the drain.
8650 0 : if let Err(err) = self
8651 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
8652 0 : .await
8653 : {
8654 : // This is not fatal. Anything that is polling the node scheduling policy to detect
8655 : // the end of the drain operations will hang, but all such places should enforce an
8656 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
8657 : // by the counterpart fill operation.
8658 0 : return Err(OperationError::FinalizeError(
8659 0 : format!(
8660 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
8661 0 : )
8662 0 : .into(),
8663 0 : ));
8664 0 : }
8665 0 :
8666 0 : Ok(())
8667 0 : }
8668 :
8669 : /// Create a node fill plan (pick secondaries to promote), based on:
8670 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
8671 : /// outside their home AZ, should be migrated back here.
8672 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
8673 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
8674 : /// those shards have a home AZ that doesn't match the node we're filling.
8675 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
8676 0 : let mut locked = self.inner.write().unwrap();
8677 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
8678 0 :
8679 0 : let node_az = nodes
8680 0 : .get(&node_id)
8681 0 : .expect("Node must exist")
8682 0 : .get_availability_zone_id()
8683 0 : .clone();
8684 0 :
8685 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
8686 0 : let mut plan = Vec::new();
8687 0 :
8688 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
8689 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
8690 0 :
8691 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
8692 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
8693 0 : let respect_azs = nodes
8694 0 : .values()
8695 0 : .map(|n| n.get_availability_zone_id())
8696 0 : .unique()
8697 0 : .count()
8698 0 : > 1;
8699 :
8700 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
8701 : // requires it.
8702 0 : for (tsid, tenant_shard) in tenants {
8703 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
8704 : // Shard doesn't have a secondary on this node, ignore it.
8705 0 : continue;
8706 0 : }
8707 0 :
8708 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
8709 0 : // shards which belong on this node, not to promote shards whose scheduling preference
8710 0 : // would be on their currently attached node. So will avoid promoting shards whose
8711 0 : // home AZ doesn't match the AZ of the node we're filling.
8712 0 : match tenant_shard.preferred_az() {
8713 0 : _ if !respect_azs => {
8714 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8715 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
8716 0 : }
8717 : }
8718 : None => {
8719 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
8720 : // will only do so if our target shard count requires it.
8721 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8722 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
8723 0 : }
8724 : }
8725 0 : Some(az) if az == &node_az => {
8726 : // This shard's home AZ is equal to the node we're filling: it should
8727 : // be moved back to this node as part of filling, unless its currently
8728 : // attached location is also in its home AZ.
8729 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8730 0 : if nodes
8731 0 : .get(primary)
8732 0 : .expect("referenced node must exist")
8733 0 : .get_availability_zone_id()
8734 0 : != tenant_shard
8735 0 : .preferred_az()
8736 0 : .expect("tenant must have an AZ preference")
8737 : {
8738 0 : plan.push(*tsid)
8739 0 : }
8740 : } else {
8741 0 : plan.push(*tsid)
8742 : }
8743 : }
8744 0 : Some(_) => {
8745 0 : // This shard's home AZ is somewhere other than the node we're filling,
8746 0 : // it may not be moved back to this node as part of filling. Ignore it
8747 0 : }
8748 : }
8749 : }
8750 :
8751 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
8752 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
8753 0 :
8754 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
8755 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
8756 0 :
8757 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
8758 :
8759 0 : for (node_id, attached) in nodes_by_load {
8760 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
8761 0 : if !available {
8762 0 : continue;
8763 0 : }
8764 0 :
8765 0 : if plan.len() >= fill_requirement
8766 0 : || free_tids_by_node.is_empty()
8767 0 : || attached <= expected_attached
8768 : {
8769 0 : break;
8770 0 : }
8771 0 :
8772 0 : let can_take = attached - expected_attached;
8773 0 : let needed = fill_requirement - plan.len();
8774 0 : let mut take = std::cmp::min(can_take, needed);
8775 0 :
8776 0 : let mut remove_node = false;
8777 0 : while take > 0 {
8778 0 : match free_tids_by_node.get_mut(&node_id) {
8779 0 : Some(tids) => match tids.pop() {
8780 0 : Some(tid) => {
8781 0 : let max_promote_for_tenant = std::cmp::max(
8782 0 : tid.shard_count.count() as usize / locked.nodes.len(),
8783 0 : 1,
8784 0 : );
8785 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
8786 0 : if *promoted < max_promote_for_tenant {
8787 0 : plan.push(tid);
8788 0 : *promoted += 1;
8789 0 : take -= 1;
8790 0 : }
8791 : }
8792 : None => {
8793 0 : remove_node = true;
8794 0 : break;
8795 : }
8796 : },
8797 : None => {
8798 0 : break;
8799 : }
8800 : }
8801 : }
8802 :
8803 0 : if remove_node {
8804 0 : free_tids_by_node.remove(&node_id);
8805 0 : }
8806 : }
8807 :
8808 0 : plan
8809 0 : }
8810 :
8811 : /// Fill a node by promoting its secondaries until the cluster is balanced
8812 : /// with regards to attached shard counts. Note that this operation only
8813 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
8814 : /// This is a long running operation and it should run as a separate Tokio task.
8815 0 : pub(crate) async fn fill_node(
8816 0 : &self,
8817 0 : node_id: NodeId,
8818 0 : cancel: CancellationToken,
8819 0 : ) -> Result<(), OperationError> {
8820 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
8821 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
8822 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
8823 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
8824 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
8825 0 : .build();
8826 0 :
8827 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
8828 0 : let mut waiters = Vec::new();
8829 :
8830 : // Execute the plan we've composed above. Before aplying each move from the plan,
8831 : // we validate to ensure that it has not gone stale in the meantime.
8832 0 : while !tids_to_promote.is_empty() {
8833 0 : if cancel.is_cancelled() {
8834 0 : match self
8835 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8836 0 : .await
8837 : {
8838 0 : Ok(()) => return Err(OperationError::Cancelled),
8839 0 : Err(err) => {
8840 0 : return Err(OperationError::FinalizeError(
8841 0 : format!(
8842 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8843 0 : node_id, err
8844 0 : )
8845 0 : .into(),
8846 0 : ));
8847 : }
8848 : }
8849 0 : }
8850 0 :
8851 0 : {
8852 0 : let mut locked = self.inner.write().unwrap();
8853 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8854 :
8855 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
8856 0 : format!("node {node_id} was removed").into(),
8857 0 : ))?;
8858 :
8859 0 : let current_policy = node.get_scheduling();
8860 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
8861 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
8862 : // about it
8863 0 : return Err(OperationError::NodeStateChanged(
8864 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
8865 0 : ));
8866 0 : }
8867 :
8868 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
8869 0 : if let Some(tid) = tids_to_promote.pop() {
8870 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
8871 : // If the node being filled is not a secondary anymore,
8872 : // skip the promotion.
8873 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
8874 0 : continue;
8875 0 : }
8876 0 :
8877 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
8878 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
8879 0 : Err(e) => {
8880 0 : tracing::warn!(
8881 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8882 0 : "Scheduling error when filling pageserver {} : {e}", node_id
8883 : );
8884 : }
8885 : Ok(()) => {
8886 0 : tracing::info!(
8887 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8888 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
8889 : node_id,
8890 : previously_attached_to,
8891 : node_id
8892 : );
8893 :
8894 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
8895 0 : tenant_shard,
8896 0 : nodes,
8897 0 : reconciler_config,
8898 0 : ) {
8899 0 : waiters.push(waiter);
8900 0 : }
8901 : }
8902 : }
8903 0 : }
8904 : } else {
8905 0 : break;
8906 : }
8907 : }
8908 : }
8909 :
8910 0 : waiters = self
8911 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
8912 0 : .await;
8913 : }
8914 :
8915 0 : while !waiters.is_empty() {
8916 0 : if cancel.is_cancelled() {
8917 0 : match self
8918 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8919 0 : .await
8920 : {
8921 0 : Ok(()) => return Err(OperationError::Cancelled),
8922 0 : Err(err) => {
8923 0 : return Err(OperationError::FinalizeError(
8924 0 : format!(
8925 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8926 0 : node_id, err
8927 0 : )
8928 0 : .into(),
8929 0 : ));
8930 : }
8931 : }
8932 0 : }
8933 0 :
8934 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
8935 :
8936 0 : waiters = self
8937 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
8938 0 : .await;
8939 : }
8940 :
8941 0 : if let Err(err) = self
8942 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8943 0 : .await
8944 : {
8945 : // This isn't a huge issue since the filling process starts upon request. However, it
8946 : // will prevent the next drain from starting. The only case in which this can fail
8947 : // is database unavailability. Such a case will require manual intervention.
8948 0 : return Err(OperationError::FinalizeError(
8949 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
8950 0 : .into(),
8951 0 : ));
8952 0 : }
8953 0 :
8954 0 : Ok(())
8955 0 : }
8956 :
8957 : /// Updates scrubber metadata health check results.
8958 0 : pub(crate) async fn metadata_health_update(
8959 0 : &self,
8960 0 : update_req: MetadataHealthUpdateRequest,
8961 0 : ) -> Result<(), ApiError> {
8962 0 : let now = chrono::offset::Utc::now();
8963 0 : let (healthy_records, unhealthy_records) = {
8964 0 : let locked = self.inner.read().unwrap();
8965 0 : let healthy_records = update_req
8966 0 : .healthy_tenant_shards
8967 0 : .into_iter()
8968 0 : // Retain only health records associated with tenant shards managed by storage controller.
8969 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
8970 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
8971 0 : .collect();
8972 0 : let unhealthy_records = update_req
8973 0 : .unhealthy_tenant_shards
8974 0 : .into_iter()
8975 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
8976 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
8977 0 : .collect();
8978 0 :
8979 0 : (healthy_records, unhealthy_records)
8980 0 : };
8981 0 :
8982 0 : self.persistence
8983 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
8984 0 : .await?;
8985 0 : Ok(())
8986 0 : }
8987 :
8988 : /// Lists the tenant shards that has unhealthy metadata status.
8989 0 : pub(crate) async fn metadata_health_list_unhealthy(
8990 0 : &self,
8991 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
8992 0 : let result = self
8993 0 : .persistence
8994 0 : .list_unhealthy_metadata_health_records()
8995 0 : .await?
8996 0 : .iter()
8997 0 : .map(|p| p.get_tenant_shard_id().unwrap())
8998 0 : .collect();
8999 0 :
9000 0 : Ok(result)
9001 0 : }
9002 :
9003 : /// Lists the tenant shards that have not been scrubbed for some duration.
9004 0 : pub(crate) async fn metadata_health_list_outdated(
9005 0 : &self,
9006 0 : not_scrubbed_for: Duration,
9007 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
9008 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
9009 0 : let result = self
9010 0 : .persistence
9011 0 : .list_outdated_metadata_health_records(earlier)
9012 0 : .await?
9013 0 : .into_iter()
9014 0 : .map(|record| record.into())
9015 0 : .collect();
9016 0 : Ok(result)
9017 0 : }
9018 :
9019 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
9020 0 : self.inner.read().unwrap().get_leadership_status()
9021 0 : }
9022 :
9023 : /// Handler for step down requests
9024 : ///
9025 : /// Step down runs in separate task since once it's called it should
9026 : /// be driven to completion. Subsequent requests will wait on the same
9027 : /// step down task.
9028 0 : pub(crate) async fn step_down(self: &Arc<Self>) -> GlobalObservedState {
9029 0 : let handle = self.step_down_barrier.get_or_init(|| {
9030 0 : let step_down_self = self.clone();
9031 0 : let (tx, rx) = tokio::sync::watch::channel::<Option<GlobalObservedState>>(None);
9032 0 : tokio::spawn(async move {
9033 0 : let state = step_down_self.step_down_task().await;
9034 0 : tx.send(Some(state))
9035 0 : .expect("Task Arc<Service> keeps receiver alive");
9036 0 : });
9037 0 :
9038 0 : rx
9039 0 : });
9040 0 :
9041 0 : handle
9042 0 : .clone()
9043 0 : .wait_for(|observed_state| observed_state.is_some())
9044 0 : .await
9045 0 : .expect("Task Arc<Service> keeps sender alive")
9046 0 : .deref()
9047 0 : .clone()
9048 0 : .expect("Checked above")
9049 0 : }
9050 :
9051 0 : async fn step_down_task(&self) -> GlobalObservedState {
9052 0 : tracing::info!("Received step down request from peer");
9053 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
9054 :
9055 0 : self.inner.write().unwrap().step_down();
9056 0 :
9057 0 : let stop_reconciliations =
9058 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown);
9059 0 : let mut stop_reconciliations = std::pin::pin!(stop_reconciliations);
9060 0 :
9061 0 : let started_at = Instant::now();
9062 :
9063 : // Wait for reconciliations to stop and warn if that's taking a long time
9064 : loop {
9065 0 : tokio::select! {
9066 0 : _ = &mut stop_reconciliations => {
9067 0 : tracing::info!("Reconciliations stopped, proceeding with step down");
9068 0 : break;
9069 : }
9070 0 : _ = tokio::time::sleep(Duration::from_secs(10)) => {
9071 0 : tracing::warn!(
9072 0 : elapsed_sec=%started_at.elapsed().as_secs(),
9073 0 : "Stopping reconciliations during step down is taking too long"
9074 : );
9075 : }
9076 : }
9077 : }
9078 :
9079 0 : let mut global_observed = GlobalObservedState::default();
9080 0 : let locked = self.inner.read().unwrap();
9081 0 : for (tid, tenant_shard) in locked.tenants.iter() {
9082 0 : global_observed
9083 0 : .0
9084 0 : .insert(*tid, tenant_shard.observed.clone());
9085 0 : }
9086 :
9087 0 : global_observed
9088 0 : }
9089 :
9090 0 : pub(crate) async fn update_shards_preferred_azs(
9091 0 : &self,
9092 0 : req: ShardsPreferredAzsRequest,
9093 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
9094 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
9095 0 : let updated = self
9096 0 : .persistence
9097 0 : .set_tenant_shard_preferred_azs(preferred_azs)
9098 0 : .await
9099 0 : .map_err(|err| {
9100 0 : ApiError::InternalServerError(anyhow::anyhow!(
9101 0 : "Failed to persist preferred AZs: {err}"
9102 0 : ))
9103 0 : })?;
9104 :
9105 0 : let mut updated_in_mem_and_db = Vec::default();
9106 0 :
9107 0 : let mut locked = self.inner.write().unwrap();
9108 0 : let state = locked.deref_mut();
9109 0 : for (tid, az_id) in updated {
9110 0 : let shard = state.tenants.get_mut(&tid);
9111 0 : if let Some(shard) = shard {
9112 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
9113 0 : updated_in_mem_and_db.push(tid);
9114 0 : }
9115 : }
9116 :
9117 0 : Ok(ShardsPreferredAzsResponse {
9118 0 : updated: updated_in_mem_and_db,
9119 0 : })
9120 0 : }
9121 : }
9122 :
9123 : #[cfg(test)]
9124 : mod tests {
9125 : use super::*;
9126 :
9127 : /// Tests Service::compute_split_shards. For readability, this specifies sizes in GBs rather
9128 : /// than bytes. Note that max_logical_size is the total logical size of the largest timeline
9129 : /// summed across all shards.
9130 : #[test]
9131 1 : fn compute_split_shards() {
9132 1 : // Size-based split: two shards have a 500 GB timeline, which need to split into 8 shards
9133 1 : // that are <= 64 GB,
9134 1 : assert_eq!(
9135 1 : Service::compute_split_shards(ShardSplitInputs {
9136 1 : shard_count: ShardCount(2),
9137 1 : max_logical_size: 500,
9138 1 : split_threshold: 64,
9139 1 : max_split_shards: 16,
9140 1 : initial_split_threshold: 0,
9141 1 : initial_split_shards: 0,
9142 1 : }),
9143 1 : Some(ShardCount(8))
9144 1 : );
9145 :
9146 : // Size-based split: noop at or below threshold, fires above.
9147 1 : assert_eq!(
9148 1 : Service::compute_split_shards(ShardSplitInputs {
9149 1 : shard_count: ShardCount(2),
9150 1 : max_logical_size: 127,
9151 1 : split_threshold: 64,
9152 1 : max_split_shards: 16,
9153 1 : initial_split_threshold: 0,
9154 1 : initial_split_shards: 0,
9155 1 : }),
9156 1 : None,
9157 1 : );
9158 1 : assert_eq!(
9159 1 : Service::compute_split_shards(ShardSplitInputs {
9160 1 : shard_count: ShardCount(2),
9161 1 : max_logical_size: 128,
9162 1 : split_threshold: 64,
9163 1 : max_split_shards: 16,
9164 1 : initial_split_threshold: 0,
9165 1 : initial_split_shards: 0,
9166 1 : }),
9167 1 : None,
9168 1 : );
9169 1 : assert_eq!(
9170 1 : Service::compute_split_shards(ShardSplitInputs {
9171 1 : shard_count: ShardCount(2),
9172 1 : max_logical_size: 129,
9173 1 : split_threshold: 64,
9174 1 : max_split_shards: 16,
9175 1 : initial_split_threshold: 0,
9176 1 : initial_split_shards: 0,
9177 1 : }),
9178 1 : Some(ShardCount(4)),
9179 1 : );
9180 :
9181 : // Size-based split: clamped to max_split_shards.
9182 1 : assert_eq!(
9183 1 : Service::compute_split_shards(ShardSplitInputs {
9184 1 : shard_count: ShardCount(2),
9185 1 : max_logical_size: 10000,
9186 1 : split_threshold: 64,
9187 1 : max_split_shards: 16,
9188 1 : initial_split_threshold: 0,
9189 1 : initial_split_shards: 0,
9190 1 : }),
9191 1 : Some(ShardCount(16))
9192 1 : );
9193 :
9194 : // Size-based split: tenant already at or beyond max_split_shards is not split.
9195 1 : assert_eq!(
9196 1 : Service::compute_split_shards(ShardSplitInputs {
9197 1 : shard_count: ShardCount(16),
9198 1 : max_logical_size: 10000,
9199 1 : split_threshold: 64,
9200 1 : max_split_shards: 16,
9201 1 : initial_split_threshold: 0,
9202 1 : initial_split_shards: 0,
9203 1 : }),
9204 1 : None
9205 1 : );
9206 :
9207 1 : assert_eq!(
9208 1 : Service::compute_split_shards(ShardSplitInputs {
9209 1 : shard_count: ShardCount(32),
9210 1 : max_logical_size: 10000,
9211 1 : split_threshold: 64,
9212 1 : max_split_shards: 16,
9213 1 : initial_split_threshold: 0,
9214 1 : initial_split_shards: 0,
9215 1 : }),
9216 1 : None
9217 1 : );
9218 :
9219 : // Size-based split: a non-power-of-2 shard count is normalized to power-of-2 if it
9220 : // exceeds split_threshold (i.e. a 3-shard tenant splits into 8, not 6).
9221 1 : assert_eq!(
9222 1 : Service::compute_split_shards(ShardSplitInputs {
9223 1 : shard_count: ShardCount(3),
9224 1 : max_logical_size: 320,
9225 1 : split_threshold: 64,
9226 1 : max_split_shards: 16,
9227 1 : initial_split_threshold: 0,
9228 1 : initial_split_shards: 0,
9229 1 : }),
9230 1 : Some(ShardCount(8))
9231 1 : );
9232 :
9233 : // Size-based split: a non-power-of-2 shard count is not normalized to power-of-2 if the
9234 : // existing shards are below or at split_threshold, but splits into 4 if it exceeds it.
9235 1 : assert_eq!(
9236 1 : Service::compute_split_shards(ShardSplitInputs {
9237 1 : shard_count: ShardCount(3),
9238 1 : max_logical_size: 191,
9239 1 : split_threshold: 64,
9240 1 : max_split_shards: 16,
9241 1 : initial_split_threshold: 0,
9242 1 : initial_split_shards: 0,
9243 1 : }),
9244 1 : None
9245 1 : );
9246 1 : assert_eq!(
9247 1 : Service::compute_split_shards(ShardSplitInputs {
9248 1 : shard_count: ShardCount(3),
9249 1 : max_logical_size: 192,
9250 1 : split_threshold: 64,
9251 1 : max_split_shards: 16,
9252 1 : initial_split_threshold: 0,
9253 1 : initial_split_shards: 0,
9254 1 : }),
9255 1 : None
9256 1 : );
9257 1 : assert_eq!(
9258 1 : Service::compute_split_shards(ShardSplitInputs {
9259 1 : shard_count: ShardCount(3),
9260 1 : max_logical_size: 193,
9261 1 : split_threshold: 64,
9262 1 : max_split_shards: 16,
9263 1 : initial_split_threshold: 0,
9264 1 : initial_split_shards: 0,
9265 1 : }),
9266 1 : Some(ShardCount(4))
9267 1 : );
9268 :
9269 : // Initial split: tenant has a 10 GB timeline, split into 4 shards.
9270 1 : assert_eq!(
9271 1 : Service::compute_split_shards(ShardSplitInputs {
9272 1 : shard_count: ShardCount(1),
9273 1 : max_logical_size: 10,
9274 1 : split_threshold: 0,
9275 1 : max_split_shards: 16,
9276 1 : initial_split_threshold: 8,
9277 1 : initial_split_shards: 4,
9278 1 : }),
9279 1 : Some(ShardCount(4))
9280 1 : );
9281 :
9282 : // Initial split: 0 ShardCount is equivalent to 1.
9283 1 : assert_eq!(
9284 1 : Service::compute_split_shards(ShardSplitInputs {
9285 1 : shard_count: ShardCount(0),
9286 1 : max_logical_size: 10,
9287 1 : split_threshold: 0,
9288 1 : max_split_shards: 16,
9289 1 : initial_split_threshold: 8,
9290 1 : initial_split_shards: 4,
9291 1 : }),
9292 1 : Some(ShardCount(4))
9293 1 : );
9294 :
9295 : // Initial split: at or below threshold is noop.
9296 1 : assert_eq!(
9297 1 : Service::compute_split_shards(ShardSplitInputs {
9298 1 : shard_count: ShardCount(1),
9299 1 : max_logical_size: 7,
9300 1 : split_threshold: 0,
9301 1 : max_split_shards: 16,
9302 1 : initial_split_threshold: 8,
9303 1 : initial_split_shards: 4,
9304 1 : }),
9305 1 : None,
9306 1 : );
9307 1 : assert_eq!(
9308 1 : Service::compute_split_shards(ShardSplitInputs {
9309 1 : shard_count: ShardCount(1),
9310 1 : max_logical_size: 8,
9311 1 : split_threshold: 0,
9312 1 : max_split_shards: 16,
9313 1 : initial_split_threshold: 8,
9314 1 : initial_split_shards: 4,
9315 1 : }),
9316 1 : None,
9317 1 : );
9318 1 : assert_eq!(
9319 1 : Service::compute_split_shards(ShardSplitInputs {
9320 1 : shard_count: ShardCount(1),
9321 1 : max_logical_size: 9,
9322 1 : split_threshold: 0,
9323 1 : max_split_shards: 16,
9324 1 : initial_split_threshold: 8,
9325 1 : initial_split_shards: 4,
9326 1 : }),
9327 1 : Some(ShardCount(4))
9328 1 : );
9329 :
9330 : // Initial split: already sharded tenant is not affected, even if above threshold and below
9331 : // shard count.
9332 1 : assert_eq!(
9333 1 : Service::compute_split_shards(ShardSplitInputs {
9334 1 : shard_count: ShardCount(2),
9335 1 : max_logical_size: 20,
9336 1 : split_threshold: 0,
9337 1 : max_split_shards: 16,
9338 1 : initial_split_threshold: 8,
9339 1 : initial_split_shards: 4,
9340 1 : }),
9341 1 : None,
9342 1 : );
9343 :
9344 : // Initial split: clamped to max_shards.
9345 1 : assert_eq!(
9346 1 : Service::compute_split_shards(ShardSplitInputs {
9347 1 : shard_count: ShardCount(1),
9348 1 : max_logical_size: 10,
9349 1 : split_threshold: 0,
9350 1 : max_split_shards: 3,
9351 1 : initial_split_threshold: 8,
9352 1 : initial_split_shards: 4,
9353 1 : }),
9354 1 : Some(ShardCount(3)),
9355 1 : );
9356 :
9357 : // Initial+size split: tenant eligible for both will use the larger shard count.
9358 1 : assert_eq!(
9359 1 : Service::compute_split_shards(ShardSplitInputs {
9360 1 : shard_count: ShardCount(1),
9361 1 : max_logical_size: 10,
9362 1 : split_threshold: 64,
9363 1 : max_split_shards: 16,
9364 1 : initial_split_threshold: 8,
9365 1 : initial_split_shards: 4,
9366 1 : }),
9367 1 : Some(ShardCount(4)),
9368 1 : );
9369 1 : assert_eq!(
9370 1 : Service::compute_split_shards(ShardSplitInputs {
9371 1 : shard_count: ShardCount(1),
9372 1 : max_logical_size: 500,
9373 1 : split_threshold: 64,
9374 1 : max_split_shards: 16,
9375 1 : initial_split_threshold: 8,
9376 1 : initial_split_shards: 4,
9377 1 : }),
9378 1 : Some(ShardCount(8)),
9379 1 : );
9380 :
9381 : // Initial+size split: sharded tenant is only eligible for size-based split.
9382 1 : assert_eq!(
9383 1 : Service::compute_split_shards(ShardSplitInputs {
9384 1 : shard_count: ShardCount(2),
9385 1 : max_logical_size: 200,
9386 1 : split_threshold: 64,
9387 1 : max_split_shards: 16,
9388 1 : initial_split_threshold: 8,
9389 1 : initial_split_shards: 8,
9390 1 : }),
9391 1 : Some(ShardCount(4)),
9392 1 : );
9393 :
9394 : // Initial+size split: uses the larger shard count even with initial_split_threshold above
9395 : // split_threshold.
9396 1 : assert_eq!(
9397 1 : Service::compute_split_shards(ShardSplitInputs {
9398 1 : shard_count: ShardCount(1),
9399 1 : max_logical_size: 10,
9400 1 : split_threshold: 4,
9401 1 : max_split_shards: 16,
9402 1 : initial_split_threshold: 8,
9403 1 : initial_split_shards: 8,
9404 1 : }),
9405 1 : Some(ShardCount(8)),
9406 1 : );
9407 :
9408 : // Test backwards compatibility with production settings when initial/size-based splits were
9409 : // rolled out: a single split into 8 shards at 64 GB. Any already sharded tenants with <8
9410 : // shards will split according to split_threshold.
9411 1 : assert_eq!(
9412 1 : Service::compute_split_shards(ShardSplitInputs {
9413 1 : shard_count: ShardCount(1),
9414 1 : max_logical_size: 65,
9415 1 : split_threshold: 64,
9416 1 : max_split_shards: 8,
9417 1 : initial_split_threshold: 64,
9418 1 : initial_split_shards: 8,
9419 1 : }),
9420 1 : Some(ShardCount(8)),
9421 1 : );
9422 :
9423 1 : assert_eq!(
9424 1 : Service::compute_split_shards(ShardSplitInputs {
9425 1 : shard_count: ShardCount(1),
9426 1 : max_logical_size: 64,
9427 1 : split_threshold: 64,
9428 1 : max_split_shards: 8,
9429 1 : initial_split_threshold: 64,
9430 1 : initial_split_shards: 8,
9431 1 : }),
9432 1 : None,
9433 1 : );
9434 :
9435 1 : assert_eq!(
9436 1 : Service::compute_split_shards(ShardSplitInputs {
9437 1 : shard_count: ShardCount(2),
9438 1 : max_logical_size: 129,
9439 1 : split_threshold: 64,
9440 1 : max_split_shards: 8,
9441 1 : initial_split_threshold: 64,
9442 1 : initial_split_shards: 8,
9443 1 : }),
9444 1 : Some(ShardCount(4)),
9445 1 : );
9446 1 : }
9447 : }
|
