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