Line data Source code
1 : pub mod chaos_injector;
2 : mod context_iterator;
3 :
4 : use hyper::Uri;
5 : use safekeeper_api::models::SafekeeperUtilization;
6 : use std::{
7 : borrow::Cow,
8 : cmp::Ordering,
9 : collections::{BTreeMap, HashMap, HashSet},
10 : error::Error,
11 : ops::Deref,
12 : path::PathBuf,
13 : str::FromStr,
14 : sync::Arc,
15 : time::{Duration, Instant},
16 : };
17 :
18 : use crate::{
19 : background_node_operations::{
20 : Drain, Fill, Operation, OperationError, OperationHandler, MAX_RECONCILES_PER_OPERATION,
21 : },
22 : compute_hook::{self, NotifyError},
23 : drain_utils::{self, TenantShardDrain, TenantShardIterator},
24 : heartbeater::SafekeeperState,
25 : id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, TracingExclusiveGuard},
26 : leadership::Leadership,
27 : metrics,
28 : peer_client::GlobalObservedState,
29 : persistence::{
30 : AbortShardSplitStatus, ControllerPersistence, DatabaseResult, MetadataHealthPersistence,
31 : ShardGenerationState, TenantFilter,
32 : },
33 : reconciler::{ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder},
34 : safekeeper::Safekeeper,
35 : scheduler::{MaySchedule, ScheduleContext, ScheduleError, ScheduleMode},
36 : tenant_shard::{
37 : MigrateAttachment, ObservedStateDelta, ReconcileNeeded, ReconcilerStatus,
38 : ScheduleOptimization, ScheduleOptimizationAction,
39 : },
40 : };
41 : use anyhow::Context;
42 : use control_plane::storage_controller::{
43 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
44 : };
45 : use diesel::result::DatabaseErrorKind;
46 : use futures::{stream::FuturesUnordered, StreamExt};
47 : use itertools::Itertools;
48 : use pageserver_api::{
49 : controller_api::{
50 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
51 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
52 : SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest,
53 : ShardsPreferredAzsResponse, SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse,
54 : TenantCreateResponseShard, TenantDescribeResponse, TenantDescribeResponseShard,
55 : TenantLocateResponse, TenantPolicyRequest, TenantShardMigrateRequest,
56 : TenantShardMigrateResponse,
57 : },
58 : models::{
59 : SecondaryProgress, TenantConfigPatchRequest, TenantConfigRequest,
60 : TimelineArchivalConfigRequest, TopTenantShardsRequest,
61 : },
62 : };
63 : use reqwest::StatusCode;
64 : use tracing::{instrument, Instrument};
65 :
66 : use crate::pageserver_client::PageserverClient;
67 : use http_utils::error::ApiError;
68 : use pageserver_api::{
69 : models::{
70 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
71 : PageserverUtilization, ShardParameters, TenantConfig, TenantLocationConfigRequest,
72 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
73 : TenantShardSplitResponse, TenantTimeTravelRequest, TimelineCreateRequest, TimelineInfo,
74 : },
75 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
76 : upcall_api::{
77 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
78 : ValidateResponse, ValidateResponseTenant,
79 : },
80 : };
81 : use pageserver_client::{mgmt_api, BlockUnblock};
82 : use tokio::sync::mpsc::error::TrySendError;
83 : use tokio_util::sync::CancellationToken;
84 : use utils::{
85 : completion::Barrier,
86 : failpoint_support,
87 : generation::Generation,
88 : id::{NodeId, TenantId, TimelineId},
89 : pausable_failpoint,
90 : sync::gate::Gate,
91 : };
92 :
93 : use crate::{
94 : compute_hook::ComputeHook,
95 : heartbeater::{Heartbeater, PageserverState},
96 : node::{AvailabilityTransition, Node},
97 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
98 : reconciler::attached_location_conf,
99 : scheduler::Scheduler,
100 : tenant_shard::{
101 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
102 : ReconcilerWaiter, TenantShard,
103 : },
104 : };
105 :
106 : use context_iterator::TenantShardContextIterator;
107 :
108 : const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
109 :
110 : // For operations that should be quick, like attaching a new tenant
111 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
112 :
113 : // For operations that might be slow, like migrating a tenant with
114 : // some data in it.
115 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
116 :
117 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
118 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
119 : // input generation from future requests as authoritative.
120 : const INITIAL_GENERATION: Generation = Generation::new(0);
121 :
122 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
123 : /// up on unresponsive pageservers and proceed.
124 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
125 :
126 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
127 : /// This must be long enough to cover node restarts as well as normal operations: in future
128 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
129 :
130 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
131 : /// offline.
132 : ///
133 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
134 : /// handling of the re-attach response may take a long time and blocks heartbeats from
135 : /// being handled on the pageserver side.
136 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
137 :
138 : /// How often to send heartbeats to registered nodes?
139 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
140 :
141 : /// How long is too long for a reconciliation?
142 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
143 :
144 : #[derive(Clone, strum_macros::Display)]
145 : enum TenantOperations {
146 : Create,
147 : LocationConfig,
148 : ConfigSet,
149 : ConfigPatch,
150 : TimeTravelRemoteStorage,
151 : Delete,
152 : UpdatePolicy,
153 : ShardSplit,
154 : SecondaryDownload,
155 : TimelineCreate,
156 : TimelineDelete,
157 : AttachHook,
158 : TimelineArchivalConfig,
159 : TimelineDetachAncestor,
160 : TimelineGcBlockUnblock,
161 : DropDetached,
162 : }
163 :
164 : #[derive(Clone, strum_macros::Display)]
165 : enum NodeOperations {
166 : Register,
167 : Configure,
168 : Delete,
169 : }
170 :
171 : /// The leadership status for the storage controller process.
172 : /// Allowed transitions are:
173 : /// 1. Leader -> SteppedDown
174 : /// 2. Candidate -> Leader
175 : #[derive(
176 : Eq,
177 : PartialEq,
178 : Copy,
179 : Clone,
180 : strum_macros::Display,
181 0 : strum_macros::EnumIter,
182 : measured::FixedCardinalityLabel,
183 : )]
184 : #[strum(serialize_all = "snake_case")]
185 : pub(crate) enum LeadershipStatus {
186 : /// This is the steady state where the storage controller can produce
187 : /// side effects in the cluster.
188 : Leader,
189 : /// We've been notified to step down by another candidate. No reconciliations
190 : /// take place in this state.
191 : SteppedDown,
192 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
193 : #[allow(unused)]
194 : Candidate,
195 : }
196 :
197 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
198 :
199 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
200 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
201 : // than they're being pushed onto the queue.
202 : const MAX_DELAYED_RECONCILES: usize = 10000;
203 :
204 : // Top level state available to all HTTP handlers
205 : struct ServiceState {
206 : leadership_status: LeadershipStatus,
207 :
208 : tenants: BTreeMap<TenantShardId, TenantShard>,
209 :
210 : nodes: Arc<HashMap<NodeId, Node>>,
211 :
212 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
213 :
214 : scheduler: Scheduler,
215 :
216 : /// Ongoing background operation on the cluster if any is running.
217 : /// Note that only one such operation may run at any given time,
218 : /// hence the type choice.
219 : ongoing_operation: Option<OperationHandler>,
220 :
221 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
222 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
223 : }
224 :
225 : /// Transform an error from a pageserver into an error to return to callers of a storage
226 : /// controller API.
227 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
228 0 : match e {
229 0 : mgmt_api::Error::SendRequest(e) => {
230 0 : // Presume errors sending requests are connectivity/availability issues
231 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
232 : }
233 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
234 0 : // Presume errors receiving body are connectivity/availability issues
235 0 : ApiError::ResourceUnavailable(
236 0 : format!("{node} error receiving error body: {str}").into(),
237 0 : )
238 : }
239 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
240 0 : // Return 500 for decoding errors.
241 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
242 : }
243 0 : mgmt_api::Error::ReceiveBody(err) => {
244 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
245 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
246 0 : ApiError::ResourceUnavailable(
247 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
248 0 : )
249 : }
250 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
251 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
252 : }
253 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
254 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
255 : }
256 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
257 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
258 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
259 : // internal server errors, showing that something is wrong with the pageserver or
260 : // storage controller's auth configuration.
261 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
262 : }
263 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
264 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
265 0 : // turn ask our clients to wait + retry
266 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
267 : }
268 0 : mgmt_api::Error::ApiError(status, msg) => {
269 0 : // Presume general case of pageserver API errors is that we tried to do something
270 0 : // that can't be done right now.
271 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
272 : }
273 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
274 : }
275 0 : }
276 :
277 : impl ServiceState {
278 0 : fn new(
279 0 : nodes: HashMap<NodeId, Node>,
280 0 : safekeepers: HashMap<NodeId, Safekeeper>,
281 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
282 0 : scheduler: Scheduler,
283 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
284 0 : initial_leadership_status: LeadershipStatus,
285 0 : ) -> Self {
286 0 : metrics::update_leadership_status(initial_leadership_status);
287 0 :
288 0 : Self {
289 0 : leadership_status: initial_leadership_status,
290 0 : tenants,
291 0 : nodes: Arc::new(nodes),
292 0 : safekeepers: Arc::new(safekeepers),
293 0 : scheduler,
294 0 : ongoing_operation: None,
295 0 : delayed_reconcile_rx,
296 0 : }
297 0 : }
298 :
299 0 : fn parts_mut(
300 0 : &mut self,
301 0 : ) -> (
302 0 : &mut Arc<HashMap<NodeId, Node>>,
303 0 : &mut BTreeMap<TenantShardId, TenantShard>,
304 0 : &mut Scheduler,
305 0 : ) {
306 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
307 0 : }
308 :
309 : #[allow(clippy::type_complexity)]
310 0 : fn parts_mut_sk(
311 0 : &mut self,
312 0 : ) -> (
313 0 : &mut Arc<HashMap<NodeId, Node>>,
314 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
315 0 : &mut BTreeMap<TenantShardId, TenantShard>,
316 0 : &mut Scheduler,
317 0 : ) {
318 0 : (
319 0 : &mut self.nodes,
320 0 : &mut self.safekeepers,
321 0 : &mut self.tenants,
322 0 : &mut self.scheduler,
323 0 : )
324 0 : }
325 :
326 0 : fn get_leadership_status(&self) -> LeadershipStatus {
327 0 : self.leadership_status
328 0 : }
329 :
330 0 : fn step_down(&mut self) {
331 0 : self.leadership_status = LeadershipStatus::SteppedDown;
332 0 : metrics::update_leadership_status(self.leadership_status);
333 0 : }
334 :
335 0 : fn become_leader(&mut self) {
336 0 : self.leadership_status = LeadershipStatus::Leader;
337 0 : metrics::update_leadership_status(self.leadership_status);
338 0 : }
339 : }
340 :
341 : #[derive(Clone)]
342 : pub struct Config {
343 : // All pageservers managed by one instance of this service must have
344 : // the same public key. This JWT token will be used to authenticate
345 : // this service to the pageservers it manages.
346 : pub jwt_token: Option<String>,
347 :
348 : // This JWT token will be used to authenticate this service to the control plane.
349 : pub control_plane_jwt_token: Option<String>,
350 :
351 : // This JWT token will be used to authenticate with other storage controller instances
352 : pub peer_jwt_token: Option<String>,
353 :
354 : /// Where the compute hook should send notifications of pageserver attachment locations
355 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
356 : /// assume it is running in a test environment and try to update neon_local.
357 : pub compute_hook_url: Option<String>,
358 :
359 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
360 : /// considered active. Once the grace period elapses, the next heartbeat failure will
361 : /// mark the pagseserver offline.
362 : pub max_offline_interval: Duration,
363 :
364 : /// Extended grace period within which pageserver may not respond to heartbeats.
365 : /// This extended grace period kicks in after the node has been drained for restart
366 : /// and/or upon handling the re-attach request from a node.
367 : pub max_warming_up_interval: Duration,
368 :
369 : /// How many Reconcilers may be spawned concurrently
370 : pub reconciler_concurrency: usize,
371 :
372 : /// How large must a shard grow in bytes before we split it?
373 : /// None disables auto-splitting.
374 : pub split_threshold: Option<u64>,
375 :
376 : // TODO: make this cfg(feature = "testing")
377 : pub neon_local_repo_dir: Option<PathBuf>,
378 :
379 : // Maximum acceptable download lag for the secondary location
380 : // while draining a node. If the secondary location is lagging
381 : // by more than the configured amount, then the secondary is not
382 : // upgraded to primary.
383 : pub max_secondary_lag_bytes: Option<u64>,
384 :
385 : pub heartbeat_interval: Duration,
386 :
387 : pub address_for_peers: Option<Uri>,
388 :
389 : pub start_as_candidate: bool,
390 :
391 : pub http_service_port: i32,
392 :
393 : pub long_reconcile_threshold: Duration,
394 : }
395 :
396 : impl From<DatabaseError> for ApiError {
397 0 : fn from(err: DatabaseError) -> ApiError {
398 0 : match err {
399 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
400 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
401 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
402 0 : ApiError::ShuttingDown
403 : }
404 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
405 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
406 : }
407 : }
408 0 : }
409 : }
410 :
411 : enum InitialShardScheduleOutcome {
412 : Scheduled(TenantCreateResponseShard),
413 : NotScheduled,
414 : ShardScheduleError(ScheduleError),
415 : }
416 :
417 : pub struct Service {
418 : inner: Arc<std::sync::RwLock<ServiceState>>,
419 : config: Config,
420 : persistence: Arc<Persistence>,
421 : compute_hook: Arc<ComputeHook>,
422 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
423 :
424 : heartbeater_ps: Heartbeater<Node, PageserverState>,
425 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
426 :
427 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
428 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
429 :
430 : // Locking on a tenant granularity (covers all shards in the tenant):
431 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
432 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
433 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
434 :
435 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
436 : // that transition it to/from Active.
437 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
438 :
439 : // Limit how many Reconcilers we will spawn concurrently
440 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
441 :
442 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
443 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
444 : ///
445 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
446 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
447 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
448 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
449 :
450 : // Process shutdown will fire this token
451 : cancel: CancellationToken,
452 :
453 : // Child token of [`Service::cancel`] used by reconcilers
454 : reconcilers_cancel: CancellationToken,
455 :
456 : // Background tasks will hold this gate
457 : gate: Gate,
458 :
459 : // Reconcilers background tasks will hold this gate
460 : reconcilers_gate: Gate,
461 :
462 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
463 : /// passes, it isn't safe to do any actions that mutate tenants.
464 : pub(crate) startup_complete: Barrier,
465 : }
466 :
467 : impl From<ReconcileWaitError> for ApiError {
468 0 : fn from(value: ReconcileWaitError) -> Self {
469 0 : match value {
470 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
471 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
472 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
473 : }
474 0 : }
475 : }
476 :
477 : impl From<OperationError> for ApiError {
478 0 : fn from(value: OperationError) -> Self {
479 0 : match value {
480 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
481 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
482 : }
483 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
484 : }
485 0 : }
486 : }
487 :
488 : #[allow(clippy::large_enum_variant)]
489 : enum TenantCreateOrUpdate {
490 : Create(TenantCreateRequest),
491 : Update(Vec<ShardUpdate>),
492 : }
493 :
494 : struct ShardSplitParams {
495 : old_shard_count: ShardCount,
496 : new_shard_count: ShardCount,
497 : new_stripe_size: Option<ShardStripeSize>,
498 : targets: Vec<ShardSplitTarget>,
499 : policy: PlacementPolicy,
500 : config: TenantConfig,
501 : shard_ident: ShardIdentity,
502 : preferred_az_id: Option<AvailabilityZone>,
503 : }
504 :
505 : // When preparing for a shard split, we may either choose to proceed with the split,
506 : // or find that the work is already done and return NoOp.
507 : enum ShardSplitAction {
508 : Split(Box<ShardSplitParams>),
509 : NoOp(TenantShardSplitResponse),
510 : }
511 :
512 : // A parent shard which will be split
513 : struct ShardSplitTarget {
514 : parent_id: TenantShardId,
515 : node: Node,
516 : child_ids: Vec<TenantShardId>,
517 : }
518 :
519 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
520 : /// might not be available. We therefore use a queue of abort operations processed in the background.
521 : struct TenantShardSplitAbort {
522 : tenant_id: TenantId,
523 : /// The target values from the request that failed
524 : new_shard_count: ShardCount,
525 : new_stripe_size: Option<ShardStripeSize>,
526 : /// Until this abort op is complete, no other operations may be done on the tenant
527 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
528 : }
529 :
530 : #[derive(thiserror::Error, Debug)]
531 : enum TenantShardSplitAbortError {
532 : #[error(transparent)]
533 : Database(#[from] DatabaseError),
534 : #[error(transparent)]
535 : Remote(#[from] mgmt_api::Error),
536 : #[error("Unavailable")]
537 : Unavailable,
538 : }
539 :
540 : struct ShardUpdate {
541 : tenant_shard_id: TenantShardId,
542 : placement_policy: PlacementPolicy,
543 : tenant_config: TenantConfig,
544 :
545 : /// If this is None, generation is not updated.
546 : generation: Option<Generation>,
547 :
548 : /// If this is None, scheduling policy is not updated.
549 : scheduling_policy: Option<ShardSchedulingPolicy>,
550 : }
551 :
552 : enum StopReconciliationsReason {
553 : ShuttingDown,
554 : SteppingDown,
555 : }
556 :
557 : impl std::fmt::Display for StopReconciliationsReason {
558 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
559 0 : let s = match self {
560 0 : Self::ShuttingDown => "Shutting down",
561 0 : Self::SteppingDown => "Stepping down",
562 : };
563 0 : write!(writer, "{}", s)
564 0 : }
565 : }
566 :
567 : pub(crate) enum ReconcileResultRequest {
568 : ReconcileResult(ReconcileResult),
569 : Stop,
570 : }
571 :
572 : #[derive(Clone)]
573 : struct MutationLocation {
574 : node: Node,
575 : generation: Generation,
576 : }
577 :
578 : #[derive(Clone)]
579 : struct ShardMutationLocations {
580 : latest: MutationLocation,
581 : other: Vec<MutationLocation>,
582 : }
583 :
584 : #[derive(Default, Clone)]
585 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
586 :
587 : impl Service {
588 0 : pub fn get_config(&self) -> &Config {
589 0 : &self.config
590 0 : }
591 :
592 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
593 : /// view of the world, and determine which pageservers are responsive.
594 : #[instrument(skip_all)]
595 : async fn startup_reconcile(
596 : self: &Arc<Service>,
597 : current_leader: Option<ControllerPersistence>,
598 : leader_step_down_state: Option<GlobalObservedState>,
599 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
600 : Result<(), (TenantShardId, NotifyError)>,
601 : >,
602 : ) {
603 : // Startup reconciliation does I/O to other services: whether they
604 : // are responsive or not, we should aim to finish within our deadline, because:
605 : // - If we don't, a k8s readiness hook watching /ready will kill us.
606 : // - While we're waiting for startup reconciliation, we are not fully
607 : // available for end user operations like creating/deleting tenants and timelines.
608 : //
609 : // We set multiple deadlines to break up the time available between the phases of work: this is
610 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
611 : let start_at = Instant::now();
612 : let node_scan_deadline = start_at
613 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
614 : .expect("Reconcile timeout is a modest constant");
615 :
616 : let observed = if let Some(state) = leader_step_down_state {
617 : tracing::info!(
618 : "Using observed state received from leader at {}",
619 : current_leader.as_ref().unwrap().address
620 : );
621 :
622 : state
623 : } else {
624 : self.build_global_observed_state(node_scan_deadline).await
625 : };
626 :
627 : // Accumulate a list of any tenant locations that ought to be detached
628 : let mut cleanup = Vec::new();
629 :
630 : // Send initial heartbeat requests to all nodes loaded from the database
631 : let all_nodes = {
632 : let locked = self.inner.read().unwrap();
633 : locked.nodes.clone()
634 : };
635 : let (mut nodes_online, mut sks_online) =
636 : self.initial_heartbeat_round(all_nodes.keys()).await;
637 :
638 : // List of tenants for which we will attempt to notify compute of their location at startup
639 : let mut compute_notifications = Vec::new();
640 :
641 : // Populate intent and observed states for all tenants, based on reported state on pageservers
642 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
643 : let shard_count = {
644 : let mut locked = self.inner.write().unwrap();
645 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
646 :
647 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
648 : let mut new_nodes = (**nodes).clone();
649 : for (node_id, node) in new_nodes.iter_mut() {
650 : if let Some(utilization) = nodes_online.remove(node_id) {
651 : node.set_availability(NodeAvailability::Active(utilization));
652 : scheduler.node_upsert(node);
653 : }
654 : }
655 : *nodes = Arc::new(new_nodes);
656 :
657 : let mut new_sks = (**safekeepers).clone();
658 : for (node_id, node) in new_sks.iter_mut() {
659 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
660 : node.set_availability(SafekeeperState::Available {
661 : utilization,
662 : last_seen_at,
663 : });
664 : }
665 : }
666 : *safekeepers = Arc::new(new_sks);
667 :
668 : for (tenant_shard_id, observed_state) in observed.0 {
669 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
670 : for node_id in observed_state.locations.keys() {
671 : cleanup.push((tenant_shard_id, *node_id));
672 : }
673 :
674 : continue;
675 : };
676 :
677 : tenant_shard.observed = observed_state;
678 : }
679 :
680 : // Populate each tenant's intent state
681 : let mut schedule_context = ScheduleContext::default();
682 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
683 : if tenant_shard_id.shard_number == ShardNumber(0) {
684 : // Reset scheduling context each time we advance to the next Tenant
685 : schedule_context = ScheduleContext::default();
686 : }
687 :
688 : tenant_shard.intent_from_observed(scheduler);
689 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
690 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
691 : // not enough pageservers are available. The tenant may well still be available
692 : // to clients.
693 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
694 : } else {
695 : // If we're both intending and observed to be attached at a particular node, we will
696 : // emit a compute notification for this. In the case where our observed state does not
697 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
698 : if let Some(attached_at) = tenant_shard.stably_attached() {
699 : compute_notifications.push(compute_hook::ShardUpdate {
700 : tenant_shard_id: *tenant_shard_id,
701 : node_id: attached_at,
702 : stripe_size: tenant_shard.shard.stripe_size,
703 : preferred_az: tenant_shard
704 : .preferred_az()
705 0 : .map(|az| Cow::Owned(az.clone())),
706 : });
707 : }
708 : }
709 : }
710 :
711 : tenants.len()
712 : };
713 :
714 : // Before making any obeservable changes to the cluster, persist self
715 : // as leader in database and memory.
716 : let leadership = Leadership::new(
717 : self.persistence.clone(),
718 : self.config.clone(),
719 : self.cancel.child_token(),
720 : );
721 :
722 : if let Err(e) = leadership.become_leader(current_leader).await {
723 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
724 : std::process::exit(1);
725 : }
726 :
727 : self.inner.write().unwrap().become_leader();
728 :
729 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
730 : // generation_pageserver in the database.
731 :
732 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
733 : // will emit compute hook notifications when they reconcile.
734 : //
735 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
736 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
737 : // calls will be correctly ordered wrt these.
738 : //
739 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
740 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
741 : // unit and start doing I/O.
742 : tracing::info!(
743 : "Sending {} compute notifications",
744 : compute_notifications.len()
745 : );
746 : self.compute_hook.notify_background(
747 : compute_notifications,
748 : bg_compute_notify_result_tx.clone(),
749 : &self.cancel,
750 : );
751 :
752 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
753 : // which require it: under normal circumstances this should only include tenants that were in some
754 : // transient state before we restarted, or any tenants whose compute hooks failed above.
755 : tracing::info!("Checking for shards in need of reconciliation...");
756 : let reconcile_tasks = self.reconcile_all();
757 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
758 : // normal operations may proceed.
759 :
760 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
761 : // background because it does not need to complete in order to proceed with other work.
762 : if !cleanup.is_empty() {
763 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
764 : tokio::task::spawn({
765 : let cleanup_self = self.clone();
766 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
767 : });
768 : }
769 :
770 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
771 : }
772 :
773 0 : async fn initial_heartbeat_round<'a>(
774 0 : &self,
775 0 : node_ids: impl Iterator<Item = &'a NodeId>,
776 0 : ) -> (
777 0 : HashMap<NodeId, PageserverUtilization>,
778 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
779 0 : ) {
780 0 : assert!(!self.startup_complete.is_ready());
781 :
782 0 : let all_nodes = {
783 0 : let locked = self.inner.read().unwrap();
784 0 : locked.nodes.clone()
785 0 : };
786 0 :
787 0 : let mut nodes_to_heartbeat = HashMap::new();
788 0 : for node_id in node_ids {
789 0 : match all_nodes.get(node_id) {
790 0 : Some(node) => {
791 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
792 0 : }
793 : None => {
794 0 : tracing::warn!("Node {node_id} was removed during start-up");
795 : }
796 : }
797 : }
798 :
799 0 : let all_sks = {
800 0 : let locked = self.inner.read().unwrap();
801 0 : locked.safekeepers.clone()
802 0 : };
803 0 :
804 0 : tracing::info!("Sending initial heartbeats...");
805 0 : let res_ps = self
806 0 : .heartbeater_ps
807 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
808 0 : .await;
809 0 : let res_sk = self.heartbeater_sk.heartbeat(all_sks).await;
810 :
811 0 : let mut online_nodes = HashMap::new();
812 0 : if let Ok(deltas) = res_ps {
813 0 : for (node_id, status) in deltas.0 {
814 0 : match status {
815 0 : PageserverState::Available { utilization, .. } => {
816 0 : online_nodes.insert(node_id, utilization);
817 0 : }
818 0 : PageserverState::Offline => {}
819 : PageserverState::WarmingUp { .. } => {
820 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
821 : }
822 : }
823 : }
824 0 : }
825 :
826 0 : let mut online_sks = HashMap::new();
827 0 : if let Ok(deltas) = res_sk {
828 0 : for (node_id, status) in deltas.0 {
829 0 : match status {
830 : SafekeeperState::Available {
831 0 : utilization,
832 0 : last_seen_at,
833 0 : } => {
834 0 : online_sks.insert(node_id, (utilization, last_seen_at));
835 0 : }
836 0 : SafekeeperState::Offline => {}
837 : }
838 : }
839 0 : }
840 :
841 0 : (online_nodes, online_sks)
842 0 : }
843 :
844 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
845 : ///
846 : /// The result includes only nodes which responded within the deadline
847 0 : async fn scan_node_locations(
848 0 : &self,
849 0 : deadline: Instant,
850 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
851 0 : let nodes = {
852 0 : let locked = self.inner.read().unwrap();
853 0 : locked.nodes.clone()
854 0 : };
855 0 :
856 0 : let mut node_results = HashMap::new();
857 0 :
858 0 : let mut node_list_futs = FuturesUnordered::new();
859 0 :
860 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
861 0 : for node in nodes.values() {
862 0 : node_list_futs.push({
863 0 : async move {
864 0 : tracing::info!("Scanning shards on node {node}...");
865 0 : let timeout = Duration::from_secs(5);
866 0 : let response = node
867 0 : .with_client_retries(
868 0 : |client| async move { client.list_location_config().await },
869 0 : &self.config.jwt_token,
870 0 : 1,
871 0 : 5,
872 0 : timeout,
873 0 : &self.cancel,
874 0 : )
875 0 : .await;
876 0 : (node.get_id(), response)
877 0 : }
878 0 : });
879 0 : }
880 :
881 : loop {
882 0 : let (node_id, result) = tokio::select! {
883 0 : next = node_list_futs.next() => {
884 0 : match next {
885 0 : Some(result) => result,
886 : None =>{
887 : // We got results for all our nodes
888 0 : break;
889 : }
890 :
891 : }
892 : },
893 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
894 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
895 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
896 0 : break;
897 : }
898 : };
899 :
900 0 : let Some(list_response) = result else {
901 0 : tracing::info!("Shutdown during startup_reconcile");
902 0 : break;
903 : };
904 :
905 0 : match list_response {
906 0 : Err(e) => {
907 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
908 : }
909 0 : Ok(listing) => {
910 0 : node_results.insert(node_id, listing);
911 0 : }
912 : }
913 : }
914 :
915 0 : node_results
916 0 : }
917 :
918 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
919 0 : let node_listings = self.scan_node_locations(deadline).await;
920 0 : let mut observed = GlobalObservedState::default();
921 :
922 0 : for (node_id, location_confs) in node_listings {
923 0 : tracing::info!(
924 0 : "Received {} shard statuses from pageserver {}",
925 0 : location_confs.tenant_shards.len(),
926 : node_id
927 : );
928 :
929 0 : for (tid, location_conf) in location_confs.tenant_shards {
930 0 : let entry = observed.0.entry(tid).or_default();
931 0 : entry.locations.insert(
932 0 : node_id,
933 0 : ObservedStateLocation {
934 0 : conf: location_conf,
935 0 : },
936 0 : );
937 0 : }
938 : }
939 :
940 0 : observed
941 0 : }
942 :
943 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
944 : ///
945 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
946 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
947 : /// other task trying to attach it.
948 : #[instrument(skip_all)]
949 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
950 : let nodes = self.inner.read().unwrap().nodes.clone();
951 :
952 : for (tenant_shard_id, node_id) in cleanup {
953 : // A node reported a tenant_shard_id which is unknown to us: detach it.
954 : let Some(node) = nodes.get(&node_id) else {
955 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
956 : // a location to clean up on a node that has since been removed.
957 : tracing::info!(
958 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
959 : );
960 : continue;
961 : };
962 :
963 : if self.cancel.is_cancelled() {
964 : break;
965 : }
966 :
967 : let client = PageserverClient::new(
968 : node.get_id(),
969 : node.base_url(),
970 : self.config.jwt_token.as_deref(),
971 : );
972 : match client
973 : .location_config(
974 : tenant_shard_id,
975 : LocationConfig {
976 : mode: LocationConfigMode::Detached,
977 : generation: None,
978 : secondary_conf: None,
979 : shard_number: tenant_shard_id.shard_number.0,
980 : shard_count: tenant_shard_id.shard_count.literal(),
981 : shard_stripe_size: 0,
982 : tenant_conf: models::TenantConfig::default(),
983 : },
984 : None,
985 : false,
986 : )
987 : .await
988 : {
989 : Ok(()) => {
990 : tracing::info!(
991 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
992 : );
993 : }
994 : Err(e) => {
995 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
996 : // break anything.
997 : tracing::error!(
998 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
999 : );
1000 : }
1001 : }
1002 : }
1003 : }
1004 :
1005 : /// Long running background task that periodically wakes up and looks for shards that need
1006 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1007 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1008 : /// for those retries.
1009 : #[instrument(skip_all)]
1010 : async fn background_reconcile(self: &Arc<Self>) {
1011 : self.startup_complete.clone().wait().await;
1012 :
1013 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1014 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1015 : while !self.reconcilers_cancel.is_cancelled() {
1016 : tokio::select! {
1017 : _ = interval.tick() => {
1018 : let reconciles_spawned = self.reconcile_all();
1019 : if reconciles_spawned == 0 {
1020 : // Run optimizer only when we didn't find any other work to do
1021 : let optimizations = self.optimize_all().await;
1022 : if optimizations == 0 {
1023 : // Run new splits only when no optimizations are pending
1024 : self.autosplit_tenants().await;
1025 : }
1026 : }
1027 : }
1028 : _ = self.reconcilers_cancel.cancelled() => return
1029 : }
1030 : }
1031 : }
1032 : #[instrument(skip_all)]
1033 : async fn spawn_heartbeat_driver(&self) {
1034 : self.startup_complete.clone().wait().await;
1035 :
1036 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1037 : while !self.cancel.is_cancelled() {
1038 : tokio::select! {
1039 : _ = interval.tick() => { }
1040 : _ = self.cancel.cancelled() => return
1041 : };
1042 :
1043 : let nodes = {
1044 : let locked = self.inner.read().unwrap();
1045 : locked.nodes.clone()
1046 : };
1047 :
1048 : let safekeepers = {
1049 : let locked = self.inner.read().unwrap();
1050 : locked.safekeepers.clone()
1051 : };
1052 :
1053 : let res_ps = self.heartbeater_ps.heartbeat(nodes).await;
1054 : let res_sk = self.heartbeater_sk.heartbeat(safekeepers).await;
1055 : if let Ok(deltas) = res_ps {
1056 : let mut to_handle = Vec::default();
1057 :
1058 : for (node_id, state) in deltas.0 {
1059 : let new_availability = match state {
1060 : PageserverState::Available { utilization, .. } => {
1061 : NodeAvailability::Active(utilization)
1062 : }
1063 : PageserverState::WarmingUp { started_at } => {
1064 : NodeAvailability::WarmingUp(started_at)
1065 : }
1066 : PageserverState::Offline => {
1067 : // The node might have been placed in the WarmingUp state
1068 : // while the heartbeat round was on-going. Hence, filter out
1069 : // offline transitions for WarmingUp nodes that are still within
1070 : // their grace period.
1071 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1072 : .get_node(node_id)
1073 : .await
1074 : .as_ref()
1075 0 : .map(|n| n.get_availability())
1076 : {
1077 : let now = Instant::now();
1078 : if now - *started_at >= self.config.max_warming_up_interval {
1079 : NodeAvailability::Offline
1080 : } else {
1081 : NodeAvailability::WarmingUp(*started_at)
1082 : }
1083 : } else {
1084 : NodeAvailability::Offline
1085 : }
1086 : }
1087 : };
1088 :
1089 : let node_lock = trace_exclusive_lock(
1090 : &self.node_op_locks,
1091 : node_id,
1092 : NodeOperations::Configure,
1093 : )
1094 : .await;
1095 :
1096 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1097 :
1098 : // This is the code path for geniune availability transitions (i.e node
1099 : // goes unavailable and/or comes back online).
1100 : let res = self
1101 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1102 : .await;
1103 :
1104 : match res {
1105 : Ok(transition) => {
1106 : // Keep hold of the lock until the availability transitions
1107 : // have been handled in
1108 : // [`Service::handle_node_availability_transitions`] in order avoid
1109 : // racing with [`Service::external_node_configure`].
1110 : to_handle.push((node_id, node_lock, transition));
1111 : }
1112 : Err(ApiError::NotFound(_)) => {
1113 : // This should be rare, but legitimate since the heartbeats are done
1114 : // on a snapshot of the nodes.
1115 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1116 : }
1117 : Err(ApiError::ShuttingDown) => {
1118 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1119 : }
1120 : Err(err) => {
1121 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1122 : // becomes unavailable again, we may get an error here.
1123 : tracing::error!(
1124 : "Failed to update node state {} after heartbeat round: {}",
1125 : node_id,
1126 : err
1127 : );
1128 : }
1129 : }
1130 : }
1131 :
1132 : // We collected all the transitions above and now we handle them.
1133 : let res = self.handle_node_availability_transitions(to_handle).await;
1134 : if let Err(errs) = res {
1135 : for (node_id, err) in errs {
1136 : match err {
1137 : ApiError::NotFound(_) => {
1138 : // This should be rare, but legitimate since the heartbeats are done
1139 : // on a snapshot of the nodes.
1140 : tracing::info!(
1141 : "Node {} was not found after heartbeat round",
1142 : node_id
1143 : );
1144 : }
1145 : err => {
1146 : tracing::error!(
1147 : "Failed to handle availability transition for {} after heartbeat round: {}",
1148 : node_id,
1149 : err
1150 : );
1151 : }
1152 : }
1153 : }
1154 : }
1155 : }
1156 : if let Ok(deltas) = res_sk {
1157 : let mut locked = self.inner.write().unwrap();
1158 : let mut safekeepers = (*locked.safekeepers).clone();
1159 : for (id, state) in deltas.0 {
1160 : let Some(sk) = safekeepers.get_mut(&id) else {
1161 : tracing::info!("Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}");
1162 : continue;
1163 : };
1164 : sk.set_availability(state);
1165 : }
1166 : locked.safekeepers = Arc::new(safekeepers);
1167 : }
1168 : }
1169 : }
1170 :
1171 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1172 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1173 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1174 : /// will indicate that reconciliation is not needed.
1175 : #[instrument(skip_all, fields(
1176 : seq=%result.sequence,
1177 : tenant_id=%result.tenant_shard_id.tenant_id,
1178 : shard_id=%result.tenant_shard_id.shard_slug(),
1179 : ))]
1180 : fn process_result(&self, result: ReconcileResult) {
1181 : let mut locked = self.inner.write().unwrap();
1182 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1183 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1184 : // A reconciliation result might race with removing a tenant: drop results for
1185 : // tenants that aren't in our map.
1186 : return;
1187 : };
1188 :
1189 : // Usually generation should only be updated via this path, so the max() isn't
1190 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1191 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1192 :
1193 : // If the reconciler signals that it failed to notify compute, set this state on
1194 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1195 : tenant.pending_compute_notification = result.pending_compute_notification;
1196 :
1197 : // Let the TenantShard know it is idle.
1198 : tenant.reconcile_complete(result.sequence);
1199 :
1200 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1201 : // make to the tenant
1202 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1203 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1204 : // make to the tenant
1205 0 : let node = nodes.get(delta.node_id())?;
1206 :
1207 0 : if node.is_available() {
1208 0 : return Some(delta);
1209 0 : }
1210 0 :
1211 0 : // In case a node became unavailable concurrently with the reconcile, observed
1212 0 : // locations on it are now uncertain. By convention, set them to None in order
1213 0 : // for them to get refreshed when the node comes back online.
1214 0 : Some(ObservedStateDelta::Upsert(Box::new((
1215 0 : node.get_id(),
1216 0 : ObservedStateLocation { conf: None },
1217 0 : ))))
1218 0 : });
1219 :
1220 : match result.result {
1221 : Ok(()) => {
1222 : tenant.apply_observed_deltas(deltas);
1223 : tenant.waiter.advance(result.sequence);
1224 : }
1225 : Err(e) => {
1226 : match e {
1227 : ReconcileError::Cancel => {
1228 : tracing::info!("Reconciler was cancelled");
1229 : }
1230 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1231 : // This might be due to the reconciler getting cancelled, or it might
1232 : // be due to the `Node` being marked offline.
1233 : tracing::info!("Reconciler cancelled during pageserver API call");
1234 : }
1235 : _ => {
1236 : tracing::warn!("Reconcile error: {}", e);
1237 : }
1238 : }
1239 :
1240 : // Ordering: populate last_error before advancing error_seq,
1241 : // so that waiters will see the correct error after waiting.
1242 : tenant.set_last_error(result.sequence, e);
1243 :
1244 : // Skip deletions on reconcile failures
1245 : let upsert_deltas =
1246 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1247 : tenant.apply_observed_deltas(upsert_deltas);
1248 : }
1249 : }
1250 :
1251 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1252 : if tenant.policy == PlacementPolicy::Detached {
1253 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1254 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1255 : // duration of the request.
1256 : let guard = self.tenant_op_locks.try_exclusive(
1257 : tenant.tenant_shard_id.tenant_id,
1258 : TenantOperations::DropDetached,
1259 : );
1260 : if let Some(guard) = guard {
1261 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1262 : }
1263 : }
1264 :
1265 : // Maybe some other work can proceed now that this job finished.
1266 : if self.reconciler_concurrency.available_permits() > 0 {
1267 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1268 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1269 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1270 : shard.delayed_reconcile = false;
1271 : self.maybe_reconcile_shard(shard, nodes);
1272 : }
1273 :
1274 : if self.reconciler_concurrency.available_permits() == 0 {
1275 : break;
1276 : }
1277 : }
1278 : }
1279 : }
1280 :
1281 0 : async fn process_results(
1282 0 : &self,
1283 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1284 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1285 0 : Result<(), (TenantShardId, NotifyError)>,
1286 0 : >,
1287 0 : ) {
1288 : loop {
1289 : // Wait for the next result, or for cancellation
1290 0 : tokio::select! {
1291 0 : r = result_rx.recv() => {
1292 0 : match r {
1293 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1294 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1295 : }
1296 : }
1297 0 : _ = async{
1298 0 : match bg_compute_hook_result_rx.recv().await {
1299 0 : Some(result) => {
1300 0 : if let Err((tenant_shard_id, notify_error)) = result {
1301 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1302 0 : let mut locked = self.inner.write().unwrap();
1303 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1304 0 : shard.pending_compute_notification = true;
1305 0 : }
1306 :
1307 0 : }
1308 : },
1309 : None => {
1310 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1311 0 : self.cancel.cancelled().await;
1312 : }
1313 : }
1314 0 : } => {},
1315 0 : _ = self.cancel.cancelled() => {
1316 0 : break;
1317 : }
1318 : };
1319 : }
1320 0 : }
1321 :
1322 0 : async fn process_aborts(
1323 0 : &self,
1324 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1325 0 : ) {
1326 : loop {
1327 : // Wait for the next result, or for cancellation
1328 0 : let op = tokio::select! {
1329 0 : r = abort_rx.recv() => {
1330 0 : match r {
1331 0 : Some(op) => {op},
1332 0 : None => {break;}
1333 : }
1334 : }
1335 0 : _ = self.cancel.cancelled() => {
1336 0 : break;
1337 : }
1338 : };
1339 :
1340 : // Retry until shutdown: we must keep this request object alive until it is properly
1341 : // processed, as it holds a lock guard that prevents other operations trying to do things
1342 : // to the tenant while it is in a weird part-split state.
1343 0 : while !self.cancel.is_cancelled() {
1344 0 : match self.abort_tenant_shard_split(&op).await {
1345 0 : Ok(_) => break,
1346 0 : Err(e) => {
1347 0 : tracing::warn!(
1348 0 : "Failed to abort shard split on {}, will retry: {e}",
1349 : op.tenant_id
1350 : );
1351 :
1352 : // If a node is unavailable, we hope that it has been properly marked Offline
1353 : // when we retry, so that the abort op will succeed. If the abort op is failing
1354 : // for some other reason, we will keep retrying forever, or until a human notices
1355 : // and does something about it (either fixing a pageserver or restarting the controller).
1356 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1357 0 : .await
1358 0 : .ok();
1359 : }
1360 : }
1361 : }
1362 : }
1363 0 : }
1364 :
1365 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1366 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1367 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1368 0 :
1369 0 : let leadership_cancel = CancellationToken::new();
1370 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1371 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1372 :
1373 : // Apply the migrations **after** the current leader has stepped down
1374 : // (or we've given up waiting for it), but **before** reading from the
1375 : // database. The only exception is reading the current leader before
1376 : // migrating.
1377 0 : persistence.migration_run().await?;
1378 :
1379 0 : tracing::info!("Loading nodes from database...");
1380 0 : let nodes = persistence
1381 0 : .list_nodes()
1382 0 : .await?
1383 0 : .into_iter()
1384 0 : .map(Node::from_persistent)
1385 0 : .collect::<Vec<_>>();
1386 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1387 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1388 0 : metrics::METRICS_REGISTRY
1389 0 : .metrics_group
1390 0 : .storage_controller_pageserver_nodes
1391 0 : .set(nodes.len() as i64);
1392 0 :
1393 0 : tracing::info!("Loading safekeepers from database...");
1394 0 : let safekeepers = persistence
1395 0 : .list_safekeepers()
1396 0 : .await?
1397 0 : .into_iter()
1398 0 : .map(|skp| Safekeeper::from_persistence(skp, CancellationToken::new()))
1399 0 : .collect::<Vec<_>>();
1400 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1401 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1402 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1403 :
1404 0 : tracing::info!("Loading shards from database...");
1405 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1406 0 : tracing::info!(
1407 0 : "Loaded {} shards from database.",
1408 0 : tenant_shard_persistence.len()
1409 : );
1410 :
1411 : // If any shard splits were in progress, reset the database state to abort them
1412 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1413 0 : HashMap::new();
1414 0 : for tsp in &mut tenant_shard_persistence {
1415 0 : let shard = tsp.get_shard_identity()?;
1416 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1417 0 : let entry = tenant_shard_count_min_max
1418 0 : .entry(tenant_shard_id.tenant_id)
1419 0 : .or_insert_with(|| (shard.count, shard.count));
1420 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1421 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1422 0 : }
1423 :
1424 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1425 0 : if count_min != count_max {
1426 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1427 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1428 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1429 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1430 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1431 :
1432 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1433 : // identified this tenant has having mismatching min/max counts.
1434 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1435 :
1436 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1437 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1438 0 : // Set idle split state on those shards that we will retain.
1439 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1440 0 : if tsp_tenant_id == tenant_id
1441 0 : && tsp.get_shard_identity().unwrap().count == count_min
1442 0 : {
1443 0 : tsp.splitting = SplitState::Idle;
1444 0 : } else if tsp_tenant_id == tenant_id {
1445 : // Leave the splitting state on the child shards: this will be used next to
1446 : // drop them.
1447 0 : tracing::info!(
1448 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1449 : );
1450 0 : }
1451 0 : });
1452 0 :
1453 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1454 0 : tenant_shard_persistence.retain(|tsp| {
1455 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1456 0 : || tsp.splitting == SplitState::Idle
1457 0 : });
1458 0 : }
1459 : }
1460 :
1461 0 : let mut tenants = BTreeMap::new();
1462 0 :
1463 0 : let mut scheduler = Scheduler::new(nodes.values());
1464 :
1465 : #[cfg(feature = "testing")]
1466 : {
1467 : use pageserver_api::controller_api::AvailabilityZone;
1468 :
1469 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1470 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1471 : // after when pageservers start up and register.
1472 0 : let mut node_ids = HashSet::new();
1473 0 : for tsp in &tenant_shard_persistence {
1474 0 : if let Some(node_id) = tsp.generation_pageserver {
1475 0 : node_ids.insert(node_id);
1476 0 : }
1477 : }
1478 0 : for node_id in node_ids {
1479 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1480 0 : let node = Node::new(
1481 0 : NodeId(node_id as u64),
1482 0 : "".to_string(),
1483 0 : 123,
1484 0 : "".to_string(),
1485 0 : 123,
1486 0 : AvailabilityZone("test_az".to_string()),
1487 0 : );
1488 0 :
1489 0 : scheduler.node_upsert(&node);
1490 : }
1491 : }
1492 0 : for tsp in tenant_shard_persistence {
1493 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1494 :
1495 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1496 : // it with what we can infer: the node for which a generation was most recently issued.
1497 0 : let mut intent = IntentState::new(
1498 0 : tsp.preferred_az_id
1499 0 : .as_ref()
1500 0 : .map(|az| AvailabilityZone(az.clone())),
1501 0 : );
1502 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1503 : {
1504 0 : if nodes.contains_key(&generation_pageserver) {
1505 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1506 0 : } else {
1507 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1508 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1509 : // on different pageservers.
1510 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1511 : }
1512 0 : }
1513 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1514 :
1515 0 : tenants.insert(tenant_shard_id, new_tenant);
1516 : }
1517 :
1518 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1519 0 :
1520 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1521 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1522 0 : tokio::sync::mpsc::channel(512);
1523 0 :
1524 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1525 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1526 0 :
1527 0 : let cancel = CancellationToken::new();
1528 0 : let reconcilers_cancel = cancel.child_token();
1529 0 :
1530 0 : let heartbeater_ps = Heartbeater::new(
1531 0 : config.jwt_token.clone(),
1532 0 : config.max_offline_interval,
1533 0 : config.max_warming_up_interval,
1534 0 : cancel.clone(),
1535 0 : );
1536 0 :
1537 0 : let heartbeater_sk = Heartbeater::new(
1538 0 : config.jwt_token.clone(),
1539 0 : config.max_offline_interval,
1540 0 : config.max_warming_up_interval,
1541 0 : cancel.clone(),
1542 0 : );
1543 :
1544 0 : let initial_leadership_status = if config.start_as_candidate {
1545 0 : LeadershipStatus::Candidate
1546 : } else {
1547 0 : LeadershipStatus::Leader
1548 : };
1549 :
1550 0 : let this = Arc::new(Self {
1551 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1552 0 : nodes,
1553 0 : safekeepers,
1554 0 : tenants,
1555 0 : scheduler,
1556 0 : delayed_reconcile_rx,
1557 0 : initial_leadership_status,
1558 0 : ))),
1559 0 : config: config.clone(),
1560 0 : persistence,
1561 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1562 0 : result_tx,
1563 0 : heartbeater_ps,
1564 0 : heartbeater_sk,
1565 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1566 0 : config.reconciler_concurrency,
1567 0 : )),
1568 0 : delayed_reconcile_tx,
1569 0 : abort_tx,
1570 0 : startup_complete: startup_complete.clone(),
1571 0 : cancel,
1572 0 : reconcilers_cancel,
1573 0 : gate: Gate::default(),
1574 0 : reconcilers_gate: Gate::default(),
1575 0 : tenant_op_locks: Default::default(),
1576 0 : node_op_locks: Default::default(),
1577 0 : });
1578 0 :
1579 0 : let result_task_this = this.clone();
1580 0 : tokio::task::spawn(async move {
1581 : // Block shutdown until we're done (we must respect self.cancel)
1582 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1583 0 : result_task_this
1584 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1585 0 : .await
1586 0 : }
1587 0 : });
1588 0 :
1589 0 : tokio::task::spawn({
1590 0 : let this = this.clone();
1591 0 : async move {
1592 : // Block shutdown until we're done (we must respect self.cancel)
1593 0 : if let Ok(_gate) = this.gate.enter() {
1594 0 : this.process_aborts(abort_rx).await
1595 0 : }
1596 0 : }
1597 0 : });
1598 0 :
1599 0 : tokio::task::spawn({
1600 0 : let this = this.clone();
1601 0 : async move {
1602 0 : if let Ok(_gate) = this.gate.enter() {
1603 : loop {
1604 0 : tokio::select! {
1605 0 : _ = this.cancel.cancelled() => {
1606 0 : break;
1607 : },
1608 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1609 0 : };
1610 0 : this.tenant_op_locks.housekeeping();
1611 : }
1612 0 : }
1613 0 : }
1614 0 : });
1615 0 :
1616 0 : tokio::task::spawn({
1617 0 : let this = this.clone();
1618 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1619 0 : // is done.
1620 0 : let startup_completion = startup_completion.clone();
1621 0 : async move {
1622 : // Block shutdown until we're done (we must respect self.cancel)
1623 0 : let Ok(_gate) = this.gate.enter() else {
1624 0 : return;
1625 : };
1626 :
1627 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1628 0 : .await;
1629 :
1630 0 : drop(startup_completion);
1631 0 : }
1632 0 : });
1633 0 :
1634 0 : tokio::task::spawn({
1635 0 : let this = this.clone();
1636 0 : let startup_complete = startup_complete.clone();
1637 0 : async move {
1638 0 : startup_complete.wait().await;
1639 0 : this.background_reconcile().await;
1640 0 : }
1641 0 : });
1642 0 :
1643 0 : tokio::task::spawn({
1644 0 : let this = this.clone();
1645 0 : let startup_complete = startup_complete.clone();
1646 0 : async move {
1647 0 : startup_complete.wait().await;
1648 0 : this.spawn_heartbeat_driver().await;
1649 0 : }
1650 0 : });
1651 0 :
1652 0 : Ok(this)
1653 0 : }
1654 :
1655 0 : pub(crate) async fn attach_hook(
1656 0 : &self,
1657 0 : attach_req: AttachHookRequest,
1658 0 : ) -> anyhow::Result<AttachHookResponse> {
1659 0 : let _tenant_lock = trace_exclusive_lock(
1660 0 : &self.tenant_op_locks,
1661 0 : attach_req.tenant_shard_id.tenant_id,
1662 0 : TenantOperations::AttachHook,
1663 0 : )
1664 0 : .await;
1665 :
1666 : // This is a test hook. To enable using it on tenants that were created directly with
1667 : // the pageserver API (not via this service), we will auto-create any missing tenant
1668 : // shards with default state.
1669 0 : let insert = {
1670 0 : match self
1671 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1672 0 : .await
1673 : {
1674 0 : Ok(_) => false,
1675 0 : Err(ApiError::NotFound(_)) => true,
1676 0 : Err(e) => return Err(e.into()),
1677 : }
1678 : };
1679 :
1680 0 : if insert {
1681 0 : let tsp = TenantShardPersistence {
1682 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1683 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1684 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1685 0 : shard_stripe_size: 0,
1686 0 : generation: attach_req.generation_override.or(Some(0)),
1687 0 : generation_pageserver: None,
1688 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1689 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1690 0 : splitting: SplitState::default(),
1691 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1692 0 : .unwrap(),
1693 0 : preferred_az_id: None,
1694 0 : };
1695 0 :
1696 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1697 0 : Err(e) => match e {
1698 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1699 : DatabaseErrorKind::UniqueViolation,
1700 : _,
1701 : )) => {
1702 0 : tracing::info!(
1703 0 : "Raced with another request to insert tenant {}",
1704 : attach_req.tenant_shard_id
1705 : )
1706 : }
1707 0 : _ => return Err(e.into()),
1708 : },
1709 : Ok(()) => {
1710 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1711 :
1712 0 : let mut locked = self.inner.write().unwrap();
1713 0 : locked.tenants.insert(
1714 0 : attach_req.tenant_shard_id,
1715 0 : TenantShard::new(
1716 0 : attach_req.tenant_shard_id,
1717 0 : ShardIdentity::unsharded(),
1718 0 : PlacementPolicy::Attached(0),
1719 0 : None,
1720 0 : ),
1721 0 : );
1722 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1723 : }
1724 : }
1725 0 : }
1726 :
1727 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1728 0 : let maybe_tenant_conf = {
1729 0 : let locked = self.inner.write().unwrap();
1730 0 : locked
1731 0 : .tenants
1732 0 : .get(&attach_req.tenant_shard_id)
1733 0 : .map(|t| t.config.clone())
1734 0 : };
1735 0 :
1736 0 : match maybe_tenant_conf {
1737 0 : Some(conf) => {
1738 0 : let new_generation = self
1739 0 : .persistence
1740 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1741 0 : .await?;
1742 :
1743 : // Persist the placement policy update. This is required
1744 : // when we reattaching a detached tenant.
1745 0 : self.persistence
1746 0 : .update_tenant_shard(
1747 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1748 0 : Some(PlacementPolicy::Attached(0)),
1749 0 : Some(conf),
1750 0 : None,
1751 0 : None,
1752 0 : )
1753 0 : .await?;
1754 0 : Some(new_generation)
1755 : }
1756 : None => {
1757 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1758 : }
1759 : }
1760 : } else {
1761 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1762 0 : None
1763 : };
1764 :
1765 0 : let mut locked = self.inner.write().unwrap();
1766 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1767 0 :
1768 0 : let tenant_shard = tenants
1769 0 : .get_mut(&attach_req.tenant_shard_id)
1770 0 : .expect("Checked for existence above");
1771 :
1772 0 : if let Some(new_generation) = new_generation {
1773 0 : tenant_shard.generation = Some(new_generation);
1774 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1775 0 : } else {
1776 : // This is a detach notification. We must update placement policy to avoid re-attaching
1777 : // during background scheduling/reconciliation, or during storage controller restart.
1778 0 : assert!(attach_req.node_id.is_none());
1779 0 : tenant_shard.policy = PlacementPolicy::Detached;
1780 : }
1781 :
1782 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1783 0 : tracing::info!(
1784 : tenant_id = %attach_req.tenant_shard_id,
1785 : ps_id = %attaching_pageserver,
1786 : generation = ?tenant_shard.generation,
1787 0 : "issuing",
1788 : );
1789 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1790 0 : tracing::info!(
1791 : tenant_id = %attach_req.tenant_shard_id,
1792 : %ps_id,
1793 : generation = ?tenant_shard.generation,
1794 0 : "dropping",
1795 : );
1796 : } else {
1797 0 : tracing::info!(
1798 : tenant_id = %attach_req.tenant_shard_id,
1799 0 : "no-op: tenant already has no pageserver");
1800 : }
1801 0 : tenant_shard
1802 0 : .intent
1803 0 : .set_attached(scheduler, attach_req.node_id);
1804 0 :
1805 0 : tracing::info!(
1806 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1807 0 : attach_req.tenant_shard_id,
1808 0 : tenant_shard.generation,
1809 0 : // TODO: this is an odd number of 0xf's
1810 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1811 : );
1812 :
1813 : // Trick the reconciler into not doing anything for this tenant: this helps
1814 : // tests that manually configure a tenant on the pagesrever, and then call this
1815 : // attach hook: they don't want background reconciliation to modify what they
1816 : // did to the pageserver.
1817 : #[cfg(feature = "testing")]
1818 : {
1819 0 : if let Some(node_id) = attach_req.node_id {
1820 0 : tenant_shard.observed.locations = HashMap::from([(
1821 0 : node_id,
1822 0 : ObservedStateLocation {
1823 0 : conf: Some(attached_location_conf(
1824 0 : tenant_shard.generation.unwrap(),
1825 0 : &tenant_shard.shard,
1826 0 : &tenant_shard.config,
1827 0 : &PlacementPolicy::Attached(0),
1828 0 : )),
1829 0 : },
1830 0 : )]);
1831 0 : } else {
1832 0 : tenant_shard.observed.locations.clear();
1833 0 : }
1834 : }
1835 :
1836 0 : Ok(AttachHookResponse {
1837 0 : gen: attach_req
1838 0 : .node_id
1839 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1840 0 : })
1841 0 : }
1842 :
1843 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1844 0 : let locked = self.inner.read().unwrap();
1845 0 :
1846 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1847 0 :
1848 0 : InspectResponse {
1849 0 : attachment: tenant_shard.and_then(|s| {
1850 0 : s.intent
1851 0 : .get_attached()
1852 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1853 0 : }),
1854 0 : }
1855 0 : }
1856 :
1857 : // When the availability state of a node transitions to active, we must do a full reconciliation
1858 : // of LocationConfigs on that node. This is because while a node was offline:
1859 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1860 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1861 : //
1862 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1863 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1864 : // this function.
1865 : //
1866 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1867 : // for written for a single node rather than as a batch job for all nodes.
1868 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1869 : async fn node_activate_reconcile(
1870 : &self,
1871 : mut node: Node,
1872 : _lock: &TracingExclusiveGuard<NodeOperations>,
1873 : ) -> Result<(), ApiError> {
1874 : // This Node is a mutable local copy: we will set it active so that we can use its
1875 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1876 : // later.
1877 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1878 :
1879 : let configs = match node
1880 : .with_client_retries(
1881 0 : |client| async move { client.list_location_config().await },
1882 : &self.config.jwt_token,
1883 : 1,
1884 : 5,
1885 : SHORT_RECONCILE_TIMEOUT,
1886 : &self.cancel,
1887 : )
1888 : .await
1889 : {
1890 : None => {
1891 : // We're shutting down (the Node's cancellation token can't have fired, because
1892 : // we're the only scope that has a reference to it, and we didn't fire it).
1893 : return Err(ApiError::ShuttingDown);
1894 : }
1895 : Some(Err(e)) => {
1896 : // This node didn't succeed listing its locations: it may not proceed to active state
1897 : // as it is apparently unavailable.
1898 : return Err(ApiError::PreconditionFailed(
1899 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1900 : ));
1901 : }
1902 : Some(Ok(configs)) => configs,
1903 : };
1904 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1905 :
1906 : let mut cleanup = Vec::new();
1907 : {
1908 : let mut locked = self.inner.write().unwrap();
1909 :
1910 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1911 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1912 : cleanup.push(tenant_shard_id);
1913 : continue;
1914 : };
1915 : tenant_shard
1916 : .observed
1917 : .locations
1918 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1919 : }
1920 : }
1921 :
1922 : for tenant_shard_id in cleanup {
1923 : tracing::info!("Detaching {tenant_shard_id}");
1924 : match node
1925 : .with_client_retries(
1926 0 : |client| async move {
1927 0 : let config = LocationConfig {
1928 0 : mode: LocationConfigMode::Detached,
1929 0 : generation: None,
1930 0 : secondary_conf: None,
1931 0 : shard_number: tenant_shard_id.shard_number.0,
1932 0 : shard_count: tenant_shard_id.shard_count.literal(),
1933 0 : shard_stripe_size: 0,
1934 0 : tenant_conf: models::TenantConfig::default(),
1935 0 : };
1936 0 : client
1937 0 : .location_config(tenant_shard_id, config, None, false)
1938 0 : .await
1939 0 : },
1940 : &self.config.jwt_token,
1941 : 1,
1942 : 5,
1943 : SHORT_RECONCILE_TIMEOUT,
1944 : &self.cancel,
1945 : )
1946 : .await
1947 : {
1948 : None => {
1949 : // We're shutting down (the Node's cancellation token can't have fired, because
1950 : // we're the only scope that has a reference to it, and we didn't fire it).
1951 : return Err(ApiError::ShuttingDown);
1952 : }
1953 : Some(Err(e)) => {
1954 : // Do not let the node proceed to Active state if it is not responsive to requests
1955 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1956 : // detach completing: we should not let this node back into the set of nodes considered
1957 : // okay for scheduling.
1958 : return Err(ApiError::Conflict(format!(
1959 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1960 : )));
1961 : }
1962 : Some(Ok(_)) => {}
1963 : };
1964 : }
1965 :
1966 : Ok(())
1967 : }
1968 :
1969 0 : pub(crate) async fn re_attach(
1970 0 : &self,
1971 0 : reattach_req: ReAttachRequest,
1972 0 : ) -> Result<ReAttachResponse, ApiError> {
1973 0 : if let Some(register_req) = reattach_req.register {
1974 0 : self.node_register(register_req).await?;
1975 0 : }
1976 :
1977 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1978 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1979 :
1980 0 : tracing::info!(
1981 : node_id=%reattach_req.node_id,
1982 0 : "Incremented {} tenant shards' generations",
1983 0 : incremented_generations.len()
1984 : );
1985 :
1986 : // Apply the updated generation to our in-memory state, and
1987 : // gather discover secondary locations.
1988 0 : let mut locked = self.inner.write().unwrap();
1989 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1990 0 :
1991 0 : let mut response = ReAttachResponse {
1992 0 : tenants: Vec::new(),
1993 0 : };
1994 :
1995 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
1996 : // to call location_conf API with an old generation. Wait for cancellation to complete
1997 : // before responding to this request. Requires well implemented CancellationToken logic
1998 : // all the way to where we call location_conf. Even then, there can still be a location_conf
1999 : // request in flight over the network: TODO handle that by making location_conf API refuse
2000 : // to go backward in generations.
2001 :
2002 : // Scan through all shards, applying updates for ones where we updated generation
2003 : // and identifying shards that intend to have a secondary location on this node.
2004 0 : for (tenant_shard_id, shard) in tenants {
2005 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2006 0 : let new_gen = *new_gen;
2007 0 : response.tenants.push(ReAttachResponseTenant {
2008 0 : id: *tenant_shard_id,
2009 0 : gen: Some(new_gen.into().unwrap()),
2010 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2011 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2012 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2013 0 : // the stale/multi states at this point.
2014 0 : mode: LocationConfigMode::AttachedSingle,
2015 0 : });
2016 0 :
2017 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2018 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2019 : // Why can we update `observed` even though we're not sure our response will be received
2020 : // by the pageserver? Because the pageserver will not proceed with startup until
2021 : // it has processed response: if it loses it, we'll see another request and increment
2022 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2023 0 : if let Some(conf) = observed.conf.as_mut() {
2024 0 : conf.generation = new_gen.into();
2025 0 : }
2026 0 : } else {
2027 0 : // This node has no observed state for the shard: perhaps it was offline
2028 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2029 0 : // will be prompted to learn the location's state before it makes changes.
2030 0 : shard
2031 0 : .observed
2032 0 : .locations
2033 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2034 0 : }
2035 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2036 0 : // Ordering: pageserver will not accept /location_config requests until it has
2037 0 : // finished processing the response from re-attach. So we can update our in-memory state
2038 0 : // now, and be confident that we are not stamping on the result of some later location config.
2039 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2040 0 : // so we might update observed state here, and then get over-written by some racing
2041 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2042 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2043 0 :
2044 0 : response.tenants.push(ReAttachResponseTenant {
2045 0 : id: *tenant_shard_id,
2046 0 : gen: None,
2047 0 : mode: LocationConfigMode::Secondary,
2048 0 : });
2049 0 :
2050 0 : // We must not update observed, because we have no guarantee that our
2051 0 : // response will be received by the pageserver. This could leave it
2052 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2053 0 : }
2054 : }
2055 :
2056 : // We consider a node Active once we have composed a re-attach response, but we
2057 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2058 : // implicitly synchronizes the LocationConfigs on the node.
2059 : //
2060 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2061 : // but those requests will not be accepted by the node until it has finished processing
2062 : // the re-attach response.
2063 : //
2064 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2065 : // in [`Persistence::re_attach`].
2066 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2067 0 : let reset_scheduling = matches!(
2068 0 : node.get_scheduling(),
2069 : NodeSchedulingPolicy::PauseForRestart
2070 : | NodeSchedulingPolicy::Draining
2071 : | NodeSchedulingPolicy::Filling
2072 : );
2073 :
2074 0 : let mut new_nodes = (**nodes).clone();
2075 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2076 0 : if reset_scheduling {
2077 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2078 0 : }
2079 :
2080 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2081 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2082 0 :
2083 0 : scheduler.node_upsert(node);
2084 0 : let new_nodes = Arc::new(new_nodes);
2085 0 : *nodes = new_nodes;
2086 : } else {
2087 0 : tracing::error!(
2088 0 : "Reattaching node {} was removed while processing the request",
2089 : reattach_req.node_id
2090 : );
2091 : }
2092 0 : }
2093 :
2094 0 : Ok(response)
2095 0 : }
2096 :
2097 0 : pub(crate) async fn validate(
2098 0 : &self,
2099 0 : validate_req: ValidateRequest,
2100 0 : ) -> Result<ValidateResponse, DatabaseError> {
2101 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2102 : // in-memory generation for a shard
2103 0 : let in_memory_result = {
2104 0 : let mut in_memory_result = Vec::new();
2105 0 : let locked = self.inner.read().unwrap();
2106 0 : for req_tenant in validate_req.tenants {
2107 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2108 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
2109 0 : tracing::info!(
2110 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2111 : req_tenant.id,
2112 : req_tenant.gen,
2113 : tenant_shard.generation
2114 : );
2115 :
2116 0 : in_memory_result.push((req_tenant.id, Generation::new(req_tenant.gen), valid));
2117 : } else {
2118 : // This is legal: for example during a shard split the pageserver may still
2119 : // have deletions in its queue from the old pre-split shard, or after deletion
2120 : // of a tenant that was busy with compaction/gc while being deleted.
2121 0 : tracing::info!(
2122 0 : "Refusing deletion validation for missing shard {}",
2123 : req_tenant.id
2124 : );
2125 : }
2126 : }
2127 :
2128 0 : in_memory_result
2129 : };
2130 :
2131 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2132 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2133 0 : let db_generations = self
2134 0 : .persistence
2135 0 : .shard_generations(in_memory_result.iter().filter_map(|i| {
2136 0 : if i.2 {
2137 0 : Some(&i.0)
2138 : } else {
2139 0 : None
2140 : }
2141 0 : }))
2142 0 : .await?;
2143 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2144 0 :
2145 0 : let mut response = ValidateResponse {
2146 0 : tenants: Vec::new(),
2147 0 : };
2148 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2149 0 : let valid = if valid {
2150 0 : let db_generation = db_generations.get(&tenant_shard_id);
2151 0 : db_generation == Some(&Some(validate_generation))
2152 : } else {
2153 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2154 : // this prevents a pageserver from cleaning up an object in S3.
2155 0 : false
2156 : };
2157 :
2158 0 : response.tenants.push(ValidateResponseTenant {
2159 0 : id: tenant_shard_id,
2160 0 : valid,
2161 0 : })
2162 : }
2163 :
2164 0 : Ok(response)
2165 0 : }
2166 :
2167 0 : pub(crate) async fn tenant_create(
2168 0 : &self,
2169 0 : create_req: TenantCreateRequest,
2170 0 : ) -> Result<TenantCreateResponse, ApiError> {
2171 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2172 :
2173 : // Exclude any concurrent attempts to create/access the same tenant ID
2174 0 : let _tenant_lock = trace_exclusive_lock(
2175 0 : &self.tenant_op_locks,
2176 0 : create_req.new_tenant_id.tenant_id,
2177 0 : TenantOperations::Create,
2178 0 : )
2179 0 : .await;
2180 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2181 :
2182 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2183 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2184 : // accept compute notifications while it is in the process of creating. Reconciliation will
2185 : // be retried in the background.
2186 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2187 0 : }
2188 0 : Ok(response)
2189 0 : }
2190 :
2191 0 : pub(crate) async fn do_tenant_create(
2192 0 : &self,
2193 0 : create_req: TenantCreateRequest,
2194 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2195 0 : let placement_policy = create_req
2196 0 : .placement_policy
2197 0 : .clone()
2198 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2199 0 : .unwrap_or(PlacementPolicy::Attached(0));
2200 :
2201 : // This service expects to handle sharding itself: it is an error to try and directly create
2202 : // a particular shard here.
2203 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2204 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2205 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2206 0 : )));
2207 : } else {
2208 0 : create_req.new_tenant_id.tenant_id
2209 0 : };
2210 0 :
2211 0 : tracing::info!(
2212 0 : "Creating tenant {}, shard_count={:?}",
2213 : create_req.new_tenant_id,
2214 : create_req.shard_parameters.count,
2215 : );
2216 :
2217 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2218 0 : .map(|i| TenantShardId {
2219 0 : tenant_id,
2220 0 : shard_number: ShardNumber(i),
2221 0 : shard_count: create_req.shard_parameters.count,
2222 0 : })
2223 0 : .collect::<Vec<_>>();
2224 :
2225 : // If the caller specifies a None generation, it means "start from default". This is different
2226 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2227 : // an incompletely-onboarded tenant.
2228 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2229 0 : tracing::info!(
2230 0 : "tenant_create: secondary mode, generation is_some={}",
2231 0 : create_req.generation.is_some()
2232 : );
2233 0 : create_req.generation.map(Generation::new)
2234 : } else {
2235 0 : tracing::info!(
2236 0 : "tenant_create: not secondary mode, generation is_some={}",
2237 0 : create_req.generation.is_some()
2238 : );
2239 0 : Some(
2240 0 : create_req
2241 0 : .generation
2242 0 : .map(Generation::new)
2243 0 : .unwrap_or(INITIAL_GENERATION),
2244 0 : )
2245 : };
2246 :
2247 0 : let preferred_az_id = {
2248 0 : let locked = self.inner.read().unwrap();
2249 : // Idempotency: take the existing value if the tenant already exists
2250 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2251 0 : shard.preferred_az().cloned()
2252 : } else {
2253 0 : locked.scheduler.get_az_for_new_tenant()
2254 : }
2255 : };
2256 :
2257 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2258 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2259 : // during the creation, rather than risking leaving orphan objects in S3.
2260 0 : let persist_tenant_shards = create_ids
2261 0 : .iter()
2262 0 : .map(|tenant_shard_id| TenantShardPersistence {
2263 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2264 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2265 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2266 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2267 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2268 0 : // The pageserver is not known until scheduling happens: we will set this column when
2269 0 : // incrementing the generation the first time we attach to a pageserver.
2270 0 : generation_pageserver: None,
2271 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2272 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2273 0 : splitting: SplitState::default(),
2274 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2275 0 : .unwrap(),
2276 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2277 0 : })
2278 0 : .collect();
2279 0 :
2280 0 : match self
2281 0 : .persistence
2282 0 : .insert_tenant_shards(persist_tenant_shards)
2283 0 : .await
2284 : {
2285 0 : Ok(_) => {}
2286 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2287 : DatabaseErrorKind::UniqueViolation,
2288 : _,
2289 : ))) => {
2290 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2291 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2292 : // creation's shard count.
2293 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
2294 : }
2295 : // Any other database error is unexpected and a bug.
2296 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2297 : };
2298 :
2299 0 : let mut schedule_context = ScheduleContext::default();
2300 0 : let mut schedule_error = None;
2301 0 : let mut response_shards = Vec::new();
2302 0 : for tenant_shard_id in create_ids {
2303 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2304 :
2305 0 : let outcome = self
2306 0 : .do_initial_shard_scheduling(
2307 0 : tenant_shard_id,
2308 0 : initial_generation,
2309 0 : &create_req.shard_parameters,
2310 0 : create_req.config.clone(),
2311 0 : placement_policy.clone(),
2312 0 : preferred_az_id.as_ref(),
2313 0 : &mut schedule_context,
2314 0 : )
2315 0 : .await;
2316 :
2317 0 : match outcome {
2318 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2319 0 : InitialShardScheduleOutcome::NotScheduled => {}
2320 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2321 0 : schedule_error = Some(err);
2322 0 : }
2323 : }
2324 : }
2325 :
2326 : // If we failed to schedule shards, then they are still created in the controller,
2327 : // but we return an error to the requester to avoid a silent failure when someone
2328 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2329 : // are present in the system. We do this here rather than in the above loop, to
2330 : // avoid situations where we only create a subset of shards in the tenant.
2331 0 : if let Some(e) = schedule_error {
2332 0 : return Err(ApiError::Conflict(format!(
2333 0 : "Failed to schedule shard(s): {e}"
2334 0 : )));
2335 0 : }
2336 0 :
2337 0 : let waiters = {
2338 0 : let mut locked = self.inner.write().unwrap();
2339 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2340 0 : let config = ReconcilerConfigBuilder::new()
2341 0 : .tenant_creation_hint(true)
2342 0 : .build();
2343 0 : tenants
2344 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2345 0 : .filter_map(|(_shard_id, shard)| {
2346 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2347 0 : })
2348 0 : .collect::<Vec<_>>()
2349 0 : };
2350 0 :
2351 0 : Ok((
2352 0 : TenantCreateResponse {
2353 0 : shards: response_shards,
2354 0 : },
2355 0 : waiters,
2356 0 : ))
2357 0 : }
2358 :
2359 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2360 : /// case of a new tenant and a pre-existing one.
2361 : #[allow(clippy::too_many_arguments)]
2362 0 : async fn do_initial_shard_scheduling(
2363 0 : &self,
2364 0 : tenant_shard_id: TenantShardId,
2365 0 : initial_generation: Option<Generation>,
2366 0 : shard_params: &ShardParameters,
2367 0 : config: TenantConfig,
2368 0 : placement_policy: PlacementPolicy,
2369 0 : preferred_az_id: Option<&AvailabilityZone>,
2370 0 : schedule_context: &mut ScheduleContext,
2371 0 : ) -> InitialShardScheduleOutcome {
2372 0 : let mut locked = self.inner.write().unwrap();
2373 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2374 :
2375 : use std::collections::btree_map::Entry;
2376 0 : match tenants.entry(tenant_shard_id) {
2377 0 : Entry::Occupied(mut entry) => {
2378 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2379 :
2380 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2381 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2382 0 : }
2383 :
2384 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2385 0 : let generation = entry
2386 0 : .get()
2387 0 : .generation
2388 0 : .expect("Generation is set when in attached mode");
2389 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2390 0 : shard_id: tenant_shard_id,
2391 0 : node_id: *node_id,
2392 0 : generation: generation.into().unwrap(),
2393 0 : })
2394 : } else {
2395 0 : InitialShardScheduleOutcome::NotScheduled
2396 : }
2397 : }
2398 0 : Entry::Vacant(entry) => {
2399 0 : let state = entry.insert(TenantShard::new(
2400 0 : tenant_shard_id,
2401 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2402 0 : placement_policy,
2403 0 : preferred_az_id.cloned(),
2404 0 : ));
2405 0 :
2406 0 : state.generation = initial_generation;
2407 0 : state.config = config;
2408 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2409 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2410 0 : }
2411 :
2412 : // Only include shards in result if we are attaching: the purpose
2413 : // of the response is to tell the caller where the shards are attached.
2414 0 : if let Some(node_id) = state.intent.get_attached() {
2415 0 : let generation = state
2416 0 : .generation
2417 0 : .expect("Generation is set when in attached mode");
2418 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2419 0 : shard_id: tenant_shard_id,
2420 0 : node_id: *node_id,
2421 0 : generation: generation.into().unwrap(),
2422 0 : })
2423 : } else {
2424 0 : InitialShardScheduleOutcome::NotScheduled
2425 : }
2426 : }
2427 : }
2428 0 : }
2429 :
2430 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2431 : /// wait for reconciliation to complete before responding.
2432 0 : async fn await_waiters(
2433 0 : &self,
2434 0 : waiters: Vec<ReconcilerWaiter>,
2435 0 : timeout: Duration,
2436 0 : ) -> Result<(), ReconcileWaitError> {
2437 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2438 0 : for waiter in waiters {
2439 0 : let timeout = deadline.duration_since(Instant::now());
2440 0 : waiter.wait_timeout(timeout).await?;
2441 : }
2442 :
2443 0 : Ok(())
2444 0 : }
2445 :
2446 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2447 : /// in progress
2448 0 : async fn await_waiters_remainder(
2449 0 : &self,
2450 0 : waiters: Vec<ReconcilerWaiter>,
2451 0 : timeout: Duration,
2452 0 : ) -> Vec<ReconcilerWaiter> {
2453 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2454 0 : for waiter in waiters.iter() {
2455 0 : let timeout = deadline.duration_since(Instant::now());
2456 0 : let _ = waiter.wait_timeout(timeout).await;
2457 : }
2458 :
2459 0 : waiters
2460 0 : .into_iter()
2461 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2462 0 : .collect::<Vec<_>>()
2463 0 : }
2464 :
2465 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2466 : /// and transform it into either a tenant creation of a series of shard updates.
2467 : ///
2468 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2469 : /// still be returned.
2470 0 : fn tenant_location_config_prepare(
2471 0 : &self,
2472 0 : tenant_id: TenantId,
2473 0 : req: TenantLocationConfigRequest,
2474 0 : ) -> TenantCreateOrUpdate {
2475 0 : let mut updates = Vec::new();
2476 0 : let mut locked = self.inner.write().unwrap();
2477 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2478 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2479 :
2480 : // Use location config mode as an indicator of policy.
2481 0 : let placement_policy = match req.config.mode {
2482 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2483 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2484 : LocationConfigMode::AttachedMulti
2485 : | LocationConfigMode::AttachedSingle
2486 : | LocationConfigMode::AttachedStale => {
2487 0 : if nodes.len() > 1 {
2488 0 : PlacementPolicy::Attached(1)
2489 : } else {
2490 : // Convenience for dev/test: if we just have one pageserver, import
2491 : // tenants into non-HA mode so that scheduling will succeed.
2492 0 : PlacementPolicy::Attached(0)
2493 : }
2494 : }
2495 : };
2496 :
2497 : // Ordinarily we do not update scheduling policy, but when making major changes
2498 : // like detaching or demoting to secondary-only, we need to force the scheduling
2499 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2500 0 : let scheduling_policy = match req.config.mode {
2501 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2502 : // Special case: when making major changes like detaching or demoting to secondary-only,
2503 : // we need to force the scheduling mode to Active, or nothing will happen.
2504 0 : Some(ShardSchedulingPolicy::Active)
2505 : }
2506 : LocationConfigMode::AttachedMulti
2507 : | LocationConfigMode::AttachedSingle
2508 : | LocationConfigMode::AttachedStale => {
2509 : // While attached, continue to respect whatever the existing scheduling mode is.
2510 0 : None
2511 : }
2512 : };
2513 :
2514 0 : let mut create = true;
2515 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2516 : // Saw an existing shard: this is not a creation
2517 0 : create = false;
2518 :
2519 : // Shards may have initially been created by a Secondary request, where we
2520 : // would have left generation as None.
2521 : //
2522 : // We only update generation the first time we see an attached-mode request,
2523 : // and if there is no existing generation set. The caller is responsible for
2524 : // ensuring that no non-storage-controller pageserver ever uses a higher
2525 : // generation than they passed in here.
2526 : use LocationConfigMode::*;
2527 0 : let set_generation = match req.config.mode {
2528 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2529 0 : req.config.generation.map(Generation::new)
2530 : }
2531 0 : _ => None,
2532 : };
2533 :
2534 0 : updates.push(ShardUpdate {
2535 0 : tenant_shard_id: *shard_id,
2536 0 : placement_policy: placement_policy.clone(),
2537 0 : tenant_config: req.config.tenant_conf.clone(),
2538 0 : generation: set_generation,
2539 0 : scheduling_policy,
2540 0 : });
2541 : }
2542 :
2543 0 : if create {
2544 : use LocationConfigMode::*;
2545 0 : let generation = match req.config.mode {
2546 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2547 : // If a caller provided a generation in a non-attached request, ignore it
2548 : // and leave our generation as None: this enables a subsequent update to set
2549 : // the generation when setting an attached mode for the first time.
2550 0 : _ => None,
2551 : };
2552 :
2553 0 : TenantCreateOrUpdate::Create(
2554 0 : // Synthesize a creation request
2555 0 : TenantCreateRequest {
2556 0 : new_tenant_id: tenant_shard_id,
2557 0 : generation,
2558 0 : shard_parameters: ShardParameters {
2559 0 : count: tenant_shard_id.shard_count,
2560 0 : // We only import un-sharded or single-sharded tenants, so stripe
2561 0 : // size can be made up arbitrarily here.
2562 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2563 0 : },
2564 0 : placement_policy: Some(placement_policy),
2565 0 : config: req.config.tenant_conf,
2566 0 : },
2567 0 : )
2568 : } else {
2569 0 : assert!(!updates.is_empty());
2570 0 : TenantCreateOrUpdate::Update(updates)
2571 : }
2572 0 : }
2573 :
2574 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2575 : /// the tenant is present in memory. If not, load it from the database. If it is found
2576 : /// in neither location, return a NotFound error.
2577 : ///
2578 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2579 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2580 0 : async fn maybe_load_tenant(
2581 0 : &self,
2582 0 : tenant_id: TenantId,
2583 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2584 0 : ) -> Result<(), ApiError> {
2585 : // Check if the tenant is present in memory, and select an AZ to use when loading
2586 : // if we will load it.
2587 0 : let load_in_az = {
2588 0 : let locked = self.inner.read().unwrap();
2589 0 : let existing = locked
2590 0 : .tenants
2591 0 : .range(TenantShardId::tenant_range(tenant_id))
2592 0 : .next();
2593 0 :
2594 0 : // If the tenant is not present in memory, we expect to load it from database,
2595 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2596 0 : if existing.is_none() {
2597 0 : locked
2598 0 : .scheduler
2599 0 : .get_az_for_new_tenant()
2600 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2601 0 : "No AZ with nodes found to load tenant"
2602 0 : )))?
2603 : } else {
2604 : // We already have this tenant in memory
2605 0 : return Ok(());
2606 : }
2607 : };
2608 :
2609 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2610 0 : if tenant_shards.is_empty() {
2611 0 : return Err(ApiError::NotFound(
2612 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2613 0 : ));
2614 0 : }
2615 0 :
2616 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2617 0 : self.persistence
2618 0 : .set_tenant_shard_preferred_azs(
2619 0 : tenant_shards
2620 0 : .iter()
2621 0 : .map(|t| {
2622 0 : (
2623 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2624 0 : Some(load_in_az.clone()),
2625 0 : )
2626 0 : })
2627 0 : .collect(),
2628 0 : )
2629 0 : .await?;
2630 :
2631 0 : let mut locked = self.inner.write().unwrap();
2632 0 : tracing::info!(
2633 0 : "Loaded {} shards for tenant {}",
2634 0 : tenant_shards.len(),
2635 : tenant_id
2636 : );
2637 :
2638 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2639 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2640 0 : let shard =
2641 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2642 0 :
2643 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2644 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2645 0 : if shard.policy != PlacementPolicy::Detached {
2646 0 : tracing::error!(
2647 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2648 : shard.tenant_shard_id,
2649 : shard.policy
2650 : );
2651 0 : }
2652 :
2653 0 : (shard.tenant_shard_id, shard)
2654 0 : }));
2655 0 :
2656 0 : Ok(())
2657 0 : }
2658 :
2659 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2660 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2661 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2662 : ///
2663 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2664 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2665 0 : fn maybe_drop_tenant(
2666 0 : &self,
2667 0 : tenant_id: TenantId,
2668 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2669 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2670 0 : ) {
2671 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2672 0 : if tenant_shards.all(|(_id, shard)| {
2673 0 : shard.policy == PlacementPolicy::Detached
2674 0 : && shard.reconciler.is_none()
2675 0 : && shard.observed.is_empty()
2676 0 : }) {
2677 0 : let keys = locked
2678 0 : .tenants
2679 0 : .range(TenantShardId::tenant_range(tenant_id))
2680 0 : .map(|(id, _)| id)
2681 0 : .copied()
2682 0 : .collect::<Vec<_>>();
2683 0 : for key in keys {
2684 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2685 0 : locked.tenants.remove(&key);
2686 : }
2687 0 : }
2688 0 : }
2689 :
2690 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2691 : /// directly with pageservers into this service.
2692 : ///
2693 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2694 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2695 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2696 : /// tenant's source of generation numbers.
2697 : ///
2698 : /// The mode in this request coarse-grained control of tenants:
2699 : /// - Call with mode Attached* to upsert the tenant.
2700 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2701 : /// to set an existing tenant to PolicyMode::Secondary
2702 : /// - Call with mode Detached to switch to PolicyMode::Detached
2703 0 : pub(crate) async fn tenant_location_config(
2704 0 : &self,
2705 0 : tenant_shard_id: TenantShardId,
2706 0 : req: TenantLocationConfigRequest,
2707 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2708 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2709 0 : let _tenant_lock = trace_exclusive_lock(
2710 0 : &self.tenant_op_locks,
2711 0 : tenant_shard_id.tenant_id,
2712 0 : TenantOperations::LocationConfig,
2713 0 : )
2714 0 : .await;
2715 :
2716 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2717 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2718 0 : "This API is for importing single-sharded or unsharded tenants"
2719 0 : )));
2720 : } else {
2721 0 : tenant_shard_id.tenant_id
2722 0 : };
2723 0 :
2724 0 : // In case we are waking up a Detached tenant
2725 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2726 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2727 0 : // This is a creation or an update
2728 0 : }
2729 0 : Err(e) => {
2730 0 : return Err(e);
2731 : }
2732 : };
2733 :
2734 : // First check if this is a creation or an update
2735 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2736 0 :
2737 0 : let mut result = TenantLocationConfigResponse {
2738 0 : shards: Vec::new(),
2739 0 : stripe_size: None,
2740 0 : };
2741 0 : let waiters = match create_or_update {
2742 0 : TenantCreateOrUpdate::Create(create_req) => {
2743 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2744 0 : result.shards = create_resp
2745 0 : .shards
2746 0 : .into_iter()
2747 0 : .map(|s| TenantShardLocation {
2748 0 : node_id: s.node_id,
2749 0 : shard_id: s.shard_id,
2750 0 : })
2751 0 : .collect();
2752 0 : waiters
2753 : }
2754 0 : TenantCreateOrUpdate::Update(updates) => {
2755 0 : // Persist updates
2756 0 : // Ordering: write to the database before applying changes in-memory, so that
2757 0 : // we will not appear time-travel backwards on a restart.
2758 0 :
2759 0 : let mut schedule_context = ScheduleContext::default();
2760 : for ShardUpdate {
2761 0 : tenant_shard_id,
2762 0 : placement_policy,
2763 0 : tenant_config,
2764 0 : generation,
2765 0 : scheduling_policy,
2766 0 : } in &updates
2767 : {
2768 0 : self.persistence
2769 0 : .update_tenant_shard(
2770 0 : TenantFilter::Shard(*tenant_shard_id),
2771 0 : Some(placement_policy.clone()),
2772 0 : Some(tenant_config.clone()),
2773 0 : *generation,
2774 0 : *scheduling_policy,
2775 0 : )
2776 0 : .await?;
2777 : }
2778 :
2779 : // Apply updates in-memory
2780 0 : let mut waiters = Vec::new();
2781 0 : {
2782 0 : let mut locked = self.inner.write().unwrap();
2783 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2784 :
2785 : for ShardUpdate {
2786 0 : tenant_shard_id,
2787 0 : placement_policy,
2788 0 : tenant_config,
2789 0 : generation: update_generation,
2790 0 : scheduling_policy,
2791 0 : } in updates
2792 : {
2793 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2794 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2795 0 : continue;
2796 : };
2797 :
2798 : // Update stripe size
2799 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2800 0 : result.stripe_size = Some(shard.shard.stripe_size);
2801 0 : }
2802 :
2803 0 : shard.policy = placement_policy;
2804 0 : shard.config = tenant_config;
2805 0 : if let Some(generation) = update_generation {
2806 0 : shard.generation = Some(generation);
2807 0 : }
2808 :
2809 0 : if let Some(scheduling_policy) = scheduling_policy {
2810 0 : shard.set_scheduling_policy(scheduling_policy);
2811 0 : }
2812 :
2813 0 : shard.schedule(scheduler, &mut schedule_context)?;
2814 :
2815 0 : let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
2816 0 : if let Some(waiter) = maybe_waiter {
2817 0 : waiters.push(waiter);
2818 0 : }
2819 :
2820 0 : if let Some(node_id) = shard.intent.get_attached() {
2821 0 : result.shards.push(TenantShardLocation {
2822 0 : shard_id: tenant_shard_id,
2823 0 : node_id: *node_id,
2824 0 : })
2825 0 : }
2826 : }
2827 : }
2828 0 : waiters
2829 : }
2830 : };
2831 :
2832 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2833 : // Do not treat a reconcile error as fatal: we have already applied any requested
2834 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2835 : // compute notification API. In these cases, it is important that we do not
2836 : // cause the cloud control plane to retry forever on this API.
2837 0 : tracing::warn!(
2838 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2839 : );
2840 0 : }
2841 :
2842 : // Logging the full result is useful because it lets us cross-check what the cloud control
2843 : // plane's tenant_shards table should contain.
2844 0 : tracing::info!("Complete, returning {result:?}");
2845 :
2846 0 : Ok(result)
2847 0 : }
2848 :
2849 0 : pub(crate) async fn tenant_config_patch(
2850 0 : &self,
2851 0 : req: TenantConfigPatchRequest,
2852 0 : ) -> Result<(), ApiError> {
2853 0 : let _tenant_lock = trace_exclusive_lock(
2854 0 : &self.tenant_op_locks,
2855 0 : req.tenant_id,
2856 0 : TenantOperations::ConfigPatch,
2857 0 : )
2858 0 : .await;
2859 :
2860 0 : let tenant_id = req.tenant_id;
2861 0 : let patch = req.config;
2862 0 :
2863 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
2864 :
2865 0 : let base = {
2866 0 : let locked = self.inner.read().unwrap();
2867 0 : let shards = locked
2868 0 : .tenants
2869 0 : .range(TenantShardId::tenant_range(req.tenant_id));
2870 0 :
2871 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
2872 :
2873 0 : let first = match configs.peek() {
2874 0 : Some(first) => (*first).clone(),
2875 : None => {
2876 0 : return Err(ApiError::NotFound(
2877 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
2878 0 : ));
2879 : }
2880 : };
2881 :
2882 0 : if !configs.all_equal() {
2883 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
2884 : // This can't happen because we atomically update the database records
2885 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
2886 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2887 0 : "Tenant configs for {} are mismatched",
2888 0 : req.tenant_id
2889 0 : )));
2890 0 : }
2891 0 :
2892 0 : first
2893 0 : };
2894 0 :
2895 0 : let updated_config = base.apply_patch(patch);
2896 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
2897 0 : .await
2898 0 : }
2899 :
2900 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2901 : // We require an exclusive lock, because we are updating persistent and in-memory state
2902 0 : let _tenant_lock = trace_exclusive_lock(
2903 0 : &self.tenant_op_locks,
2904 0 : req.tenant_id,
2905 0 : TenantOperations::ConfigSet,
2906 0 : )
2907 0 : .await;
2908 :
2909 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
2910 :
2911 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
2912 0 : .await
2913 0 : }
2914 :
2915 0 : async fn set_tenant_config_and_reconcile(
2916 0 : &self,
2917 0 : tenant_id: TenantId,
2918 0 : config: TenantConfig,
2919 0 : ) -> Result<(), ApiError> {
2920 0 : self.persistence
2921 0 : .update_tenant_shard(
2922 0 : TenantFilter::Tenant(tenant_id),
2923 0 : None,
2924 0 : Some(config.clone()),
2925 0 : None,
2926 0 : None,
2927 0 : )
2928 0 : .await?;
2929 :
2930 0 : let waiters = {
2931 0 : let mut waiters = Vec::new();
2932 0 : let mut locked = self.inner.write().unwrap();
2933 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2934 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2935 0 : shard.config = config.clone();
2936 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2937 0 : waiters.push(waiter);
2938 0 : }
2939 : }
2940 0 : waiters
2941 : };
2942 :
2943 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2944 : // Treat this as success because we have stored the configuration. If e.g.
2945 : // a node was unavailable at this time, it should not stop us accepting a
2946 : // configuration change.
2947 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2948 0 : }
2949 :
2950 0 : Ok(())
2951 0 : }
2952 :
2953 0 : pub(crate) fn tenant_config_get(
2954 0 : &self,
2955 0 : tenant_id: TenantId,
2956 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2957 0 : let config = {
2958 0 : let locked = self.inner.read().unwrap();
2959 0 :
2960 0 : match locked
2961 0 : .tenants
2962 0 : .range(TenantShardId::tenant_range(tenant_id))
2963 0 : .next()
2964 : {
2965 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2966 : None => {
2967 0 : return Err(ApiError::NotFound(
2968 0 : anyhow::anyhow!("Tenant not found").into(),
2969 0 : ))
2970 : }
2971 : }
2972 : };
2973 :
2974 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2975 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2976 : // and `effective_config` in the response is meaningless, but we retain that syntax
2977 : // in order to remain compatible with the pageserver API.
2978 :
2979 0 : let response = HashMap::from([
2980 : (
2981 : "tenant_specific_overrides",
2982 0 : serde_json::to_value(&config)
2983 0 : .context("serializing tenant specific overrides")
2984 0 : .map_err(ApiError::InternalServerError)?,
2985 : ),
2986 : (
2987 0 : "effective_config",
2988 0 : serde_json::to_value(&config)
2989 0 : .context("serializing effective config")
2990 0 : .map_err(ApiError::InternalServerError)?,
2991 : ),
2992 : ]);
2993 :
2994 0 : Ok(response)
2995 0 : }
2996 :
2997 0 : pub(crate) async fn tenant_time_travel_remote_storage(
2998 0 : &self,
2999 0 : time_travel_req: &TenantTimeTravelRequest,
3000 0 : tenant_id: TenantId,
3001 0 : timestamp: Cow<'_, str>,
3002 0 : done_if_after: Cow<'_, str>,
3003 0 : ) -> Result<(), ApiError> {
3004 0 : let _tenant_lock = trace_exclusive_lock(
3005 0 : &self.tenant_op_locks,
3006 0 : tenant_id,
3007 0 : TenantOperations::TimeTravelRemoteStorage,
3008 0 : )
3009 0 : .await;
3010 :
3011 0 : let node = {
3012 0 : let mut locked = self.inner.write().unwrap();
3013 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3014 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3015 : // but only at the start of the process, so it's really just to prevent operator
3016 : // mistakes.
3017 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3018 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3019 : {
3020 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3021 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3022 0 : )));
3023 0 : }
3024 0 : let maybe_attached = shard
3025 0 : .observed
3026 0 : .locations
3027 0 : .iter()
3028 0 : .filter_map(|(node_id, observed_location)| {
3029 0 : observed_location
3030 0 : .conf
3031 0 : .as_ref()
3032 0 : .map(|loc| (node_id, observed_location, loc.mode))
3033 0 : })
3034 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3035 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3036 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
3037 0 : }
3038 : }
3039 0 : let scheduler = &mut locked.scheduler;
3040 : // Right now we only perform the operation on a single node without parallelization
3041 : // TODO fan out the operation to multiple nodes for better performance
3042 0 : let node_id = scheduler.any_available_node()?;
3043 0 : let node = locked
3044 0 : .nodes
3045 0 : .get(&node_id)
3046 0 : .expect("Pageservers may not be deleted while lock is active");
3047 0 : node.clone()
3048 0 : };
3049 0 :
3050 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3051 0 : let mut counts = time_travel_req
3052 0 : .shard_counts
3053 0 : .iter()
3054 0 : .copied()
3055 0 : .collect::<HashSet<_>>()
3056 0 : .into_iter()
3057 0 : .collect::<Vec<_>>();
3058 0 : counts.sort_unstable();
3059 :
3060 0 : for count in counts {
3061 0 : let shard_ids = (0..count.count())
3062 0 : .map(|i| TenantShardId {
3063 0 : tenant_id,
3064 0 : shard_number: ShardNumber(i),
3065 0 : shard_count: count,
3066 0 : })
3067 0 : .collect::<Vec<_>>();
3068 0 : for tenant_shard_id in shard_ids {
3069 0 : let client = PageserverClient::new(
3070 0 : node.get_id(),
3071 0 : node.base_url(),
3072 0 : self.config.jwt_token.as_deref(),
3073 0 : );
3074 0 :
3075 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3076 :
3077 0 : client
3078 0 : .tenant_time_travel_remote_storage(
3079 0 : tenant_shard_id,
3080 0 : ×tamp,
3081 0 : &done_if_after,
3082 0 : )
3083 0 : .await
3084 0 : .map_err(|e| {
3085 0 : ApiError::InternalServerError(anyhow::anyhow!(
3086 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3087 0 : node
3088 0 : ))
3089 0 : })?;
3090 : }
3091 : }
3092 0 : Ok(())
3093 0 : }
3094 :
3095 0 : pub(crate) async fn tenant_secondary_download(
3096 0 : &self,
3097 0 : tenant_id: TenantId,
3098 0 : wait: Option<Duration>,
3099 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3100 0 : let _tenant_lock = trace_shared_lock(
3101 0 : &self.tenant_op_locks,
3102 0 : tenant_id,
3103 0 : TenantOperations::SecondaryDownload,
3104 0 : )
3105 0 : .await;
3106 :
3107 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3108 0 : let targets = {
3109 0 : let locked = self.inner.read().unwrap();
3110 0 : let mut targets = Vec::new();
3111 :
3112 0 : for (tenant_shard_id, shard) in
3113 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3114 : {
3115 0 : for node_id in shard.intent.get_secondary() {
3116 0 : let node = locked
3117 0 : .nodes
3118 0 : .get(node_id)
3119 0 : .expect("Pageservers may not be deleted while referenced");
3120 0 :
3121 0 : targets.push((*tenant_shard_id, node.clone()));
3122 0 : }
3123 : }
3124 0 : targets
3125 0 : };
3126 0 :
3127 0 : // Issue concurrent requests to all shards' locations
3128 0 : let mut futs = FuturesUnordered::new();
3129 0 : for (tenant_shard_id, node) in targets {
3130 0 : let client = PageserverClient::new(
3131 0 : node.get_id(),
3132 0 : node.base_url(),
3133 0 : self.config.jwt_token.as_deref(),
3134 0 : );
3135 0 : futs.push(async move {
3136 0 : let result = client
3137 0 : .tenant_secondary_download(tenant_shard_id, wait)
3138 0 : .await;
3139 0 : (result, node, tenant_shard_id)
3140 0 : })
3141 : }
3142 :
3143 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3144 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3145 : // well as more general cases like 503s, 500s, or timeouts.
3146 0 : let mut aggregate_progress = SecondaryProgress::default();
3147 0 : let mut aggregate_status: Option<StatusCode> = None;
3148 0 : let mut error: Option<mgmt_api::Error> = None;
3149 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3150 0 : match result {
3151 0 : Err(e) => {
3152 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3153 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3154 0 : // than they had hoped for.
3155 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3156 0 : error = Some(e)
3157 : }
3158 0 : Ok((status_code, progress)) => {
3159 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3160 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3161 0 : aggregate_progress.layers_total += progress.layers_total;
3162 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3163 0 : aggregate_progress.bytes_total += progress.bytes_total;
3164 0 : aggregate_progress.heatmap_mtime =
3165 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3166 0 : aggregate_status = match aggregate_status {
3167 0 : None => Some(status_code),
3168 0 : Some(StatusCode::OK) => Some(status_code),
3169 0 : Some(cur) => {
3170 0 : // Other status codes (e.g. 202) -- do not overwrite.
3171 0 : Some(cur)
3172 : }
3173 : };
3174 : }
3175 : }
3176 : }
3177 :
3178 : // If any of the shards return 202, indicate our result as 202.
3179 0 : match aggregate_status {
3180 : None => {
3181 0 : match error {
3182 0 : Some(e) => {
3183 0 : // No successes, and an error: surface it
3184 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3185 : }
3186 : None => {
3187 : // No shards found
3188 0 : Err(ApiError::NotFound(
3189 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3190 0 : ))
3191 : }
3192 : }
3193 : }
3194 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3195 : }
3196 0 : }
3197 :
3198 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
3199 0 : let _tenant_lock =
3200 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3201 :
3202 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3203 :
3204 : // Detach all shards. This also deletes local pageserver shard data.
3205 0 : let (detach_waiters, node) = {
3206 0 : let mut detach_waiters = Vec::new();
3207 0 : let mut locked = self.inner.write().unwrap();
3208 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3209 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3210 : // Update the tenant's intent to remove all attachments
3211 0 : shard.policy = PlacementPolicy::Detached;
3212 0 : shard
3213 0 : .schedule(scheduler, &mut ScheduleContext::default())
3214 0 : .expect("De-scheduling is infallible");
3215 0 : debug_assert!(shard.intent.get_attached().is_none());
3216 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3217 :
3218 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
3219 0 : detach_waiters.push(waiter);
3220 0 : }
3221 : }
3222 :
3223 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3224 : // was attached, just has to be able to see the S3 content)
3225 0 : let node_id = scheduler.any_available_node()?;
3226 0 : let node = nodes
3227 0 : .get(&node_id)
3228 0 : .expect("Pageservers may not be deleted while lock is active");
3229 0 : (detach_waiters, node.clone())
3230 0 : };
3231 0 :
3232 0 : // This reconcile wait can fail in a few ways:
3233 0 : // A there is a very long queue for the reconciler semaphore
3234 0 : // B some pageserver is failing to handle a detach promptly
3235 0 : // C some pageserver goes offline right at the moment we send it a request.
3236 0 : //
3237 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3238 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3239 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3240 0 : // deleting the underlying data).
3241 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3242 0 : .await?;
3243 :
3244 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3245 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3246 : // the tenant ID prefix, including all shards (even possibly stale ones).
3247 0 : match node
3248 0 : .with_client_retries(
3249 0 : |client| async move {
3250 0 : client
3251 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3252 0 : .await
3253 0 : },
3254 0 : &self.config.jwt_token,
3255 0 : 1,
3256 0 : 3,
3257 0 : RECONCILE_TIMEOUT,
3258 0 : &self.cancel,
3259 0 : )
3260 0 : .await
3261 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3262 : {
3263 0 : Ok(_) => {}
3264 : Err(mgmt_api::Error::Cancelled) => {
3265 0 : return Err(ApiError::ShuttingDown);
3266 : }
3267 0 : Err(e) => {
3268 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3269 0 : // at large is unavailable.
3270 0 : tracing::error!("Error deleting via node {node}: {e}");
3271 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3272 : }
3273 : }
3274 :
3275 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3276 : // our in-memory state and database state.
3277 :
3278 : // Ordering: we delete persistent state first: if we then
3279 : // crash, we will drop the in-memory state.
3280 :
3281 : // Drop persistent state.
3282 0 : self.persistence.delete_tenant(tenant_id).await?;
3283 :
3284 : // Drop in-memory state
3285 : {
3286 0 : let mut locked = self.inner.write().unwrap();
3287 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3288 :
3289 : // Dereference Scheduler from shards before dropping them
3290 0 : for (_tenant_shard_id, shard) in
3291 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3292 0 : {
3293 0 : shard.intent.clear(scheduler);
3294 0 : }
3295 :
3296 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3297 0 : tracing::info!(
3298 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3299 0 : locked.tenants.len()
3300 : );
3301 : };
3302 :
3303 : // Success is represented as 404, to imitate the existing pageserver deletion API
3304 0 : Ok(StatusCode::NOT_FOUND)
3305 0 : }
3306 :
3307 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3308 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3309 : /// the tenant's policies (configuration) within the storage controller
3310 0 : pub(crate) async fn tenant_update_policy(
3311 0 : &self,
3312 0 : tenant_id: TenantId,
3313 0 : req: TenantPolicyRequest,
3314 0 : ) -> Result<(), ApiError> {
3315 : // We require an exclusive lock, because we are updating persistent and in-memory state
3316 0 : let _tenant_lock = trace_exclusive_lock(
3317 0 : &self.tenant_op_locks,
3318 0 : tenant_id,
3319 0 : TenantOperations::UpdatePolicy,
3320 0 : )
3321 0 : .await;
3322 :
3323 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3324 :
3325 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3326 :
3327 : let TenantPolicyRequest {
3328 0 : placement,
3329 0 : mut scheduling,
3330 0 : } = req;
3331 :
3332 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3333 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3334 : // this to take effect.
3335 0 : if scheduling.is_none() {
3336 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3337 0 : }
3338 0 : }
3339 :
3340 0 : self.persistence
3341 0 : .update_tenant_shard(
3342 0 : TenantFilter::Tenant(tenant_id),
3343 0 : placement.clone(),
3344 0 : None,
3345 0 : None,
3346 0 : scheduling,
3347 0 : )
3348 0 : .await?;
3349 :
3350 0 : let mut schedule_context = ScheduleContext::default();
3351 0 : let mut locked = self.inner.write().unwrap();
3352 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3353 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3354 0 : if let Some(placement) = &placement {
3355 0 : shard.policy = placement.clone();
3356 0 :
3357 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3358 0 : "Updated placement policy to {placement:?}");
3359 0 : }
3360 :
3361 0 : if let Some(scheduling) = &scheduling {
3362 0 : shard.set_scheduling_policy(*scheduling);
3363 0 :
3364 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3365 0 : "Updated scheduling policy to {scheduling:?}");
3366 0 : }
3367 :
3368 : // In case scheduling is being switched back on, try it now.
3369 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3370 0 : self.maybe_reconcile_shard(shard, nodes);
3371 : }
3372 :
3373 0 : Ok(())
3374 0 : }
3375 :
3376 0 : pub(crate) async fn tenant_timeline_create(
3377 0 : &self,
3378 0 : tenant_id: TenantId,
3379 0 : mut create_req: TimelineCreateRequest,
3380 0 : ) -> Result<TimelineInfo, ApiError> {
3381 0 : tracing::info!(
3382 0 : "Creating timeline {}/{}",
3383 : tenant_id,
3384 : create_req.new_timeline_id,
3385 : );
3386 :
3387 0 : let _tenant_lock = trace_shared_lock(
3388 0 : &self.tenant_op_locks,
3389 0 : tenant_id,
3390 0 : TenantOperations::TimelineCreate,
3391 0 : )
3392 0 : .await;
3393 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3394 :
3395 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3396 0 : if targets.0.is_empty() {
3397 0 : return Err(ApiError::NotFound(
3398 0 : anyhow::anyhow!("Tenant not found").into(),
3399 0 : ));
3400 0 : };
3401 0 :
3402 0 : let (shard_zero_tid, shard_zero_locations) =
3403 0 : targets.0.pop_first().expect("Must have at least one shard");
3404 0 : assert!(shard_zero_tid.is_shard_zero());
3405 :
3406 0 : async fn create_one(
3407 0 : tenant_shard_id: TenantShardId,
3408 0 : locations: ShardMutationLocations,
3409 0 : jwt: Option<String>,
3410 0 : create_req: TimelineCreateRequest,
3411 0 : ) -> Result<TimelineInfo, ApiError> {
3412 0 : let latest = locations.latest.node;
3413 0 :
3414 0 : tracing::info!(
3415 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3416 : tenant_shard_id,
3417 : create_req.new_timeline_id,
3418 : locations.latest.generation
3419 : );
3420 :
3421 0 : let client =
3422 0 : PageserverClient::new(latest.get_id(), latest.base_url(), jwt.as_deref());
3423 :
3424 0 : let timeline_info = client
3425 0 : .timeline_create(tenant_shard_id, &create_req)
3426 0 : .await
3427 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3428 :
3429 : // We propagate timeline creations to all attached locations such that a compute
3430 : // for the new timeline is able to start regardless of the current state of the
3431 : // tenant shard reconciliation.
3432 0 : for location in locations.other {
3433 0 : tracing::info!(
3434 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3435 : tenant_shard_id,
3436 : create_req.new_timeline_id,
3437 : location.node,
3438 : location.generation
3439 : );
3440 :
3441 0 : let client = PageserverClient::new(
3442 0 : location.node.get_id(),
3443 0 : location.node.base_url(),
3444 0 : jwt.as_deref(),
3445 0 : );
3446 :
3447 0 : let res = client
3448 0 : .timeline_create(tenant_shard_id, &create_req)
3449 0 : .await;
3450 :
3451 0 : if let Err(e) = res {
3452 0 : match e {
3453 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3454 0 : // Tenant might have been detached from the stale location,
3455 0 : // so ignore 404s.
3456 0 : },
3457 : _ => {
3458 0 : return Err(passthrough_api_error(&location.node, e));
3459 : }
3460 : }
3461 0 : }
3462 : }
3463 :
3464 0 : Ok(timeline_info)
3465 0 : }
3466 :
3467 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3468 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3469 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3470 0 : let timeline_info = create_one(
3471 0 : shard_zero_tid,
3472 0 : shard_zero_locations,
3473 0 : self.config.jwt_token.clone(),
3474 0 : create_req.clone(),
3475 0 : )
3476 0 : .await?;
3477 :
3478 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3479 0 : match &mut create_req.mode {
3480 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3481 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3482 0 : },
3483 0 : _ => {}
3484 : }
3485 :
3486 : // Create timeline on remaining shards with number >0
3487 0 : if !targets.0.is_empty() {
3488 : // If we had multiple shards, issue requests for the remainder now.
3489 0 : let jwt = &self.config.jwt_token;
3490 0 : self.tenant_for_shards(
3491 0 : targets
3492 0 : .0
3493 0 : .iter()
3494 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3495 0 : .collect(),
3496 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3497 0 : let create_req = create_req.clone();
3498 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3499 0 : Box::pin(create_one(
3500 0 : tenant_shard_id,
3501 0 : mutation_locations,
3502 0 : jwt.clone(),
3503 0 : create_req,
3504 0 : ))
3505 0 : },
3506 0 : )
3507 0 : .await?;
3508 0 : }
3509 :
3510 0 : Ok(timeline_info)
3511 0 : })
3512 0 : .await?
3513 0 : }
3514 :
3515 0 : pub(crate) async fn tenant_timeline_archival_config(
3516 0 : &self,
3517 0 : tenant_id: TenantId,
3518 0 : timeline_id: TimelineId,
3519 0 : req: TimelineArchivalConfigRequest,
3520 0 : ) -> Result<(), ApiError> {
3521 0 : tracing::info!(
3522 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3523 : req.state
3524 : );
3525 :
3526 0 : let _tenant_lock = trace_shared_lock(
3527 0 : &self.tenant_op_locks,
3528 0 : tenant_id,
3529 0 : TenantOperations::TimelineArchivalConfig,
3530 0 : )
3531 0 : .await;
3532 :
3533 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3534 0 : if targets.0.is_empty() {
3535 0 : return Err(ApiError::NotFound(
3536 0 : anyhow::anyhow!("Tenant not found").into(),
3537 0 : ));
3538 0 : }
3539 0 : async fn config_one(
3540 0 : tenant_shard_id: TenantShardId,
3541 0 : timeline_id: TimelineId,
3542 0 : node: Node,
3543 0 : jwt: Option<String>,
3544 0 : req: TimelineArchivalConfigRequest,
3545 0 : ) -> Result<(), ApiError> {
3546 0 : tracing::info!(
3547 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3548 : );
3549 :
3550 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3551 0 :
3552 0 : client
3553 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3554 0 : .await
3555 0 : .map_err(|e| match e {
3556 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3557 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3558 : }
3559 0 : _ => passthrough_api_error(&node, e),
3560 0 : })
3561 0 : }
3562 :
3563 : // no shard needs to go first/last; the operation should be idempotent
3564 : // TODO: it would be great to ensure that all shards return the same error
3565 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3566 0 : let results = self
3567 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3568 0 : futures::FutureExt::boxed(config_one(
3569 0 : tenant_shard_id,
3570 0 : timeline_id,
3571 0 : node,
3572 0 : self.config.jwt_token.clone(),
3573 0 : req.clone(),
3574 0 : ))
3575 0 : })
3576 0 : .await?;
3577 0 : assert!(!results.is_empty(), "must have at least one result");
3578 :
3579 0 : Ok(())
3580 0 : }).await?
3581 0 : }
3582 :
3583 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3584 0 : &self,
3585 0 : tenant_id: TenantId,
3586 0 : timeline_id: TimelineId,
3587 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3588 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3589 :
3590 0 : let _tenant_lock = trace_shared_lock(
3591 0 : &self.tenant_op_locks,
3592 0 : tenant_id,
3593 0 : TenantOperations::TimelineDetachAncestor,
3594 0 : )
3595 0 : .await;
3596 :
3597 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3598 0 : if targets.0.is_empty() {
3599 0 : return Err(ApiError::NotFound(
3600 0 : anyhow::anyhow!("Tenant not found").into(),
3601 0 : ));
3602 0 : }
3603 :
3604 0 : async fn detach_one(
3605 0 : tenant_shard_id: TenantShardId,
3606 0 : timeline_id: TimelineId,
3607 0 : node: Node,
3608 0 : jwt: Option<String>,
3609 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3610 0 : tracing::info!(
3611 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3612 : );
3613 :
3614 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3615 0 :
3616 0 : client
3617 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3618 0 : .await
3619 0 : .map_err(|e| {
3620 : use mgmt_api::Error;
3621 :
3622 0 : match e {
3623 : // no ancestor (ever)
3624 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3625 0 : "{node}: {}",
3626 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3627 0 : )),
3628 : // too many ancestors
3629 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3630 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3631 : }
3632 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3633 0 : // avoid turning these into conflicts to remain compatible with
3634 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3635 0 : // detach
3636 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3637 : }
3638 : // rest can be mapped as usual
3639 0 : other => passthrough_api_error(&node, other),
3640 : }
3641 0 : })
3642 0 : .map(|res| (tenant_shard_id.shard_number, res))
3643 0 : }
3644 :
3645 : // no shard needs to go first/last; the operation should be idempotent
3646 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3647 0 : let mut results = self
3648 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3649 0 : futures::FutureExt::boxed(detach_one(
3650 0 : tenant_shard_id,
3651 0 : timeline_id,
3652 0 : node,
3653 0 : self.config.jwt_token.clone(),
3654 0 : ))
3655 0 : })
3656 0 : .await?;
3657 :
3658 0 : let any = results.pop().expect("we must have at least one response");
3659 0 :
3660 0 : let mismatching = results
3661 0 : .iter()
3662 0 : .filter(|(_, res)| res != &any.1)
3663 0 : .collect::<Vec<_>>();
3664 0 : if !mismatching.is_empty() {
3665 : // this can be hit by races which should not happen because operation lock on cplane
3666 0 : let matching = results.len() - mismatching.len();
3667 0 : tracing::error!(
3668 : matching,
3669 : compared_against=?any,
3670 : ?mismatching,
3671 0 : "shards returned different results"
3672 : );
3673 :
3674 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3675 0 : }
3676 0 :
3677 0 : Ok(any.1)
3678 0 : }).await?
3679 0 : }
3680 :
3681 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3682 0 : &self,
3683 0 : tenant_id: TenantId,
3684 0 : timeline_id: TimelineId,
3685 0 : dir: BlockUnblock,
3686 0 : ) -> Result<(), ApiError> {
3687 0 : let _tenant_lock = trace_shared_lock(
3688 0 : &self.tenant_op_locks,
3689 0 : tenant_id,
3690 0 : TenantOperations::TimelineGcBlockUnblock,
3691 0 : )
3692 0 : .await;
3693 :
3694 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3695 0 : if targets.0.is_empty() {
3696 0 : return Err(ApiError::NotFound(
3697 0 : anyhow::anyhow!("Tenant not found").into(),
3698 0 : ));
3699 0 : }
3700 :
3701 0 : async fn do_one(
3702 0 : tenant_shard_id: TenantShardId,
3703 0 : timeline_id: TimelineId,
3704 0 : node: Node,
3705 0 : jwt: Option<String>,
3706 0 : dir: BlockUnblock,
3707 0 : ) -> Result<(), ApiError> {
3708 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3709 0 :
3710 0 : client
3711 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3712 0 : .await
3713 0 : .map_err(|e| passthrough_api_error(&node, e))
3714 0 : }
3715 :
3716 : // no shard needs to go first/last; the operation should be idempotent
3717 0 : let locations = targets
3718 0 : .0
3719 0 : .iter()
3720 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3721 0 : .collect();
3722 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
3723 0 : futures::FutureExt::boxed(do_one(
3724 0 : tenant_shard_id,
3725 0 : timeline_id,
3726 0 : node,
3727 0 : self.config.jwt_token.clone(),
3728 0 : dir,
3729 0 : ))
3730 0 : })
3731 0 : .await
3732 0 : })
3733 0 : .await??;
3734 0 : Ok(())
3735 0 : }
3736 :
3737 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3738 : ///
3739 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3740 0 : async fn tenant_for_shards<F, R>(
3741 0 : &self,
3742 0 : locations: Vec<(TenantShardId, Node)>,
3743 0 : mut req_fn: F,
3744 0 : ) -> Result<Vec<R>, ApiError>
3745 0 : where
3746 0 : F: FnMut(
3747 0 : TenantShardId,
3748 0 : Node,
3749 0 : )
3750 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3751 0 : {
3752 0 : let mut futs = FuturesUnordered::new();
3753 0 : let mut results = Vec::with_capacity(locations.len());
3754 :
3755 0 : for (tenant_shard_id, node) in locations {
3756 0 : futs.push(req_fn(tenant_shard_id, node));
3757 0 : }
3758 :
3759 0 : while let Some(r) = futs.next().await {
3760 0 : results.push(r?);
3761 : }
3762 :
3763 0 : Ok(results)
3764 0 : }
3765 :
3766 : /// Concurrently invoke a pageserver API call on many shards at once
3767 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3768 0 : &self,
3769 0 : locations: Vec<(TenantShardId, Node)>,
3770 0 : op: O,
3771 0 : warn_threshold: u32,
3772 0 : max_retries: u32,
3773 0 : timeout: Duration,
3774 0 : cancel: &CancellationToken,
3775 0 : ) -> Vec<mgmt_api::Result<T>>
3776 0 : where
3777 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3778 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3779 0 : {
3780 0 : let mut futs = FuturesUnordered::new();
3781 0 : let mut results = Vec::with_capacity(locations.len());
3782 :
3783 0 : for (tenant_shard_id, node) in locations {
3784 0 : futs.push(async move {
3785 0 : node.with_client_retries(
3786 0 : |client| op(tenant_shard_id, client),
3787 0 : &self.config.jwt_token,
3788 0 : warn_threshold,
3789 0 : max_retries,
3790 0 : timeout,
3791 0 : cancel,
3792 0 : )
3793 0 : .await
3794 0 : });
3795 0 : }
3796 :
3797 0 : while let Some(r) = futs.next().await {
3798 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3799 0 : results.push(r);
3800 0 : }
3801 :
3802 0 : results
3803 0 : }
3804 :
3805 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3806 : /// when creating and deleting timelines:
3807 : /// - Makes sure shards are attached somewhere if they weren't already
3808 : /// - Looks up the shards and the nodes where they were most recently attached
3809 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3810 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3811 0 : async fn tenant_remote_mutation<R, O, F>(
3812 0 : &self,
3813 0 : tenant_id: TenantId,
3814 0 : op: O,
3815 0 : ) -> Result<R, ApiError>
3816 0 : where
3817 0 : O: FnOnce(TenantMutationLocations) -> F,
3818 0 : F: std::future::Future<Output = R>,
3819 0 : {
3820 0 : let mutation_locations = {
3821 0 : let mut locations = TenantMutationLocations::default();
3822 :
3823 : // Load the currently attached pageservers for the latest generation of each shard. This can
3824 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3825 : // will still be the latest when we're done: we will check generations again at the end of
3826 : // this function to handle that.
3827 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
3828 :
3829 0 : if generations
3830 0 : .iter()
3831 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3832 : {
3833 0 : let shard_generations = generations
3834 0 : .into_iter()
3835 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
3836 0 : .collect::<HashMap<_, _>>();
3837 0 :
3838 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3839 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)
3840 0 : let locked = self.inner.read().unwrap();
3841 0 : for (shard_id, shard) in
3842 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3843 : {
3844 0 : match shard.policy {
3845 : PlacementPolicy::Attached(_) => {
3846 : // This shard is meant to be attached: the caller is not wrong to try and
3847 : // use this function, but we can't service the request right now.
3848 0 : let Some(generation) = shard_generations.get(shard_id) else {
3849 : // This can only happen if there is a split brain controller modifying the database. This should
3850 : // never happen when testing, and if it happens in production we can only log the issue.
3851 0 : debug_assert!(false);
3852 0 : tracing::error!("Shard {shard_id} not found in generation state! Is another rogue controller running?");
3853 0 : continue;
3854 : };
3855 0 : let (generation, generation_pageserver) = generation;
3856 0 : if let Some(generation) = generation {
3857 0 : if generation_pageserver.is_none() {
3858 : // This is legitimate only in a very narrow window where the shard was only just configured into
3859 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
3860 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
3861 0 : tracing::warn!("Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?");
3862 0 : }
3863 : } else {
3864 : // This should never happen: a shard with no generation is only permitted when it was created in some state
3865 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
3866 0 : debug_assert!(false);
3867 0 : tracing::error!("Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!");
3868 0 : continue;
3869 : }
3870 : }
3871 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3872 0 : return Err(ApiError::Conflict(format!(
3873 0 : "Shard {shard_id} tenant has policy {:?}",
3874 0 : shard.policy
3875 0 : )));
3876 : }
3877 : }
3878 : }
3879 :
3880 0 : return Err(ApiError::ResourceUnavailable(
3881 0 : "One or more shards in tenant is not yet attached".into(),
3882 0 : ));
3883 0 : }
3884 0 :
3885 0 : let locked = self.inner.read().unwrap();
3886 : for ShardGenerationState {
3887 0 : tenant_shard_id,
3888 0 : generation,
3889 0 : generation_pageserver,
3890 0 : } in generations
3891 : {
3892 0 : let node_id = generation_pageserver.expect("We checked for None above");
3893 0 : let node = locked
3894 0 : .nodes
3895 0 : .get(&node_id)
3896 0 : .ok_or(ApiError::Conflict(format!(
3897 0 : "Raced with removal of node {node_id}"
3898 0 : )))?;
3899 0 : let generation = generation.expect("Checked above");
3900 0 :
3901 0 : let tenant = locked.tenants.get(&tenant_shard_id);
3902 :
3903 : // TODO(vlad): Abstract the logic that finds stale attached locations
3904 : // from observed state into a [`Service`] method.
3905 0 : let other_locations = match tenant {
3906 0 : Some(tenant) => {
3907 0 : let mut other = tenant.attached_locations();
3908 0 : let latest_location_index =
3909 0 : other.iter().position(|&l| l == (node.get_id(), generation));
3910 0 : if let Some(idx) = latest_location_index {
3911 0 : other.remove(idx);
3912 0 : }
3913 :
3914 0 : other
3915 : }
3916 0 : None => Vec::default(),
3917 : };
3918 :
3919 0 : let location = ShardMutationLocations {
3920 0 : latest: MutationLocation {
3921 0 : node: node.clone(),
3922 0 : generation,
3923 0 : },
3924 0 : other: other_locations
3925 0 : .into_iter()
3926 0 : .filter_map(|(node_id, generation)| {
3927 0 : let node = locked.nodes.get(&node_id)?;
3928 :
3929 0 : Some(MutationLocation {
3930 0 : node: node.clone(),
3931 0 : generation,
3932 0 : })
3933 0 : })
3934 0 : .collect(),
3935 0 : };
3936 0 : locations.0.insert(tenant_shard_id, location);
3937 0 : }
3938 :
3939 0 : locations
3940 : };
3941 :
3942 0 : let result = op(mutation_locations.clone()).await;
3943 :
3944 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
3945 : // our remote operation executed on the latest generation and is therefore persistent.
3946 : {
3947 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
3948 0 : if latest_generations
3949 0 : .into_iter()
3950 0 : .map(
3951 0 : |ShardGenerationState {
3952 : tenant_shard_id,
3953 : generation,
3954 : generation_pageserver: _,
3955 0 : }| (tenant_shard_id, generation),
3956 0 : )
3957 0 : .collect::<Vec<_>>()
3958 0 : != mutation_locations
3959 0 : .0
3960 0 : .into_iter()
3961 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
3962 0 : .collect::<Vec<_>>()
3963 : {
3964 : // We raced with something that incremented the generation, and therefore cannot be
3965 : // confident that our actions are persistent (they might have hit an old generation).
3966 : //
3967 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
3968 0 : return Err(ApiError::ResourceUnavailable(
3969 0 : "Tenant attachment changed, please retry".into(),
3970 0 : ));
3971 0 : }
3972 0 : }
3973 0 :
3974 0 : Ok(result)
3975 0 : }
3976 :
3977 0 : pub(crate) async fn tenant_timeline_delete(
3978 0 : &self,
3979 0 : tenant_id: TenantId,
3980 0 : timeline_id: TimelineId,
3981 0 : ) -> Result<StatusCode, ApiError> {
3982 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
3983 0 : let _tenant_lock = trace_shared_lock(
3984 0 : &self.tenant_op_locks,
3985 0 : tenant_id,
3986 0 : TenantOperations::TimelineDelete,
3987 0 : )
3988 0 : .await;
3989 :
3990 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3991 0 : if targets.0.is_empty() {
3992 0 : return Err(ApiError::NotFound(
3993 0 : anyhow::anyhow!("Tenant not found").into(),
3994 0 : ));
3995 0 : }
3996 0 :
3997 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
3998 0 : assert!(shard_zero_tid.is_shard_zero());
3999 :
4000 0 : async fn delete_one(
4001 0 : tenant_shard_id: TenantShardId,
4002 0 : timeline_id: TimelineId,
4003 0 : node: Node,
4004 0 : jwt: Option<String>,
4005 0 : ) -> Result<StatusCode, ApiError> {
4006 0 : tracing::info!(
4007 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4008 : );
4009 :
4010 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
4011 0 : let res = client
4012 0 : .timeline_delete(tenant_shard_id, timeline_id)
4013 0 : .await;
4014 :
4015 0 : match res {
4016 0 : Ok(ok) => Ok(ok),
4017 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
4018 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
4019 0 : Err(e) => {
4020 0 : Err(
4021 0 : ApiError::InternalServerError(anyhow::anyhow!(
4022 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
4023 0 : ))
4024 0 : )
4025 : }
4026 : }
4027 0 : }
4028 :
4029 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4030 0 : let statuses = self
4031 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
4032 0 : Box::pin(delete_one(
4033 0 : tenant_shard_id,
4034 0 : timeline_id,
4035 0 : node,
4036 0 : self.config.jwt_token.clone(),
4037 0 : ))
4038 0 : })
4039 0 : .await?;
4040 :
4041 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
4042 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
4043 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
4044 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
4045 0 : return Ok(StatusCode::CONFLICT);
4046 0 : }
4047 0 :
4048 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
4049 0 : return Ok(StatusCode::ACCEPTED);
4050 0 : }
4051 :
4052 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
4053 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
4054 0 : let shard_zero_status = delete_one(
4055 0 : shard_zero_tid,
4056 0 : timeline_id,
4057 0 : shard_zero_locations.latest.node,
4058 0 : self.config.jwt_token.clone(),
4059 0 : )
4060 0 : .await?;
4061 0 : Ok(shard_zero_status)
4062 0 : }).await?
4063 0 : }
4064 :
4065 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
4066 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
4067 0 : pub(crate) async fn tenant_shard0_node(
4068 0 : &self,
4069 0 : tenant_id: TenantId,
4070 0 : ) -> Result<(Node, TenantShardId), ApiError> {
4071 0 : // Look up in-memory state and maybe use the node from there.
4072 0 : {
4073 0 : let locked = self.inner.read().unwrap();
4074 0 : let Some((tenant_shard_id, shard)) = locked
4075 0 : .tenants
4076 0 : .range(TenantShardId::tenant_range(tenant_id))
4077 0 : .next()
4078 : else {
4079 0 : return Err(ApiError::NotFound(
4080 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
4081 0 : ));
4082 : };
4083 :
4084 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
4085 0 : tracing::warn!(
4086 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
4087 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
4088 : shard.policy
4089 : );
4090 0 : return Err(ApiError::Conflict(
4091 0 : "Cannot call timeline API on non-attached tenant".to_string(),
4092 0 : ));
4093 : };
4094 :
4095 0 : if shard.reconciler.is_none() {
4096 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
4097 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
4098 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
4099 : // This should never happen
4100 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4101 0 : "Shard refers to nonexistent node"
4102 0 : )));
4103 : };
4104 0 : return Ok((node.clone(), *tenant_shard_id));
4105 0 : }
4106 : };
4107 :
4108 : // Look up the latest attached pageserver location from the database
4109 : // generation state: this will reflect the progress of any ongoing migration.
4110 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4111 : // the case where they call through to the pageserver and get a 404.
4112 0 : let db_result = self.persistence.tenant_generations(tenant_id).await?;
4113 : let Some(ShardGenerationState {
4114 0 : tenant_shard_id,
4115 0 : generation: _,
4116 0 : generation_pageserver: Some(node_id),
4117 0 : }) = db_result.first()
4118 : else {
4119 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4120 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4121 : // or a conflict response (case where tenant was detached in background)
4122 0 : return Err(ApiError::ResourceUnavailable(
4123 0 : "Shard {} not found in database, or is not attached".into(),
4124 0 : ));
4125 : };
4126 0 : let locked = self.inner.read().unwrap();
4127 0 : let Some(node) = locked.nodes.get(node_id) else {
4128 : // This should never happen
4129 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4130 0 : "Shard refers to nonexistent node"
4131 0 : )));
4132 : };
4133 :
4134 0 : Ok((node.clone(), *tenant_shard_id))
4135 0 : }
4136 :
4137 0 : pub(crate) fn tenant_locate(
4138 0 : &self,
4139 0 : tenant_id: TenantId,
4140 0 : ) -> Result<TenantLocateResponse, ApiError> {
4141 0 : let locked = self.inner.read().unwrap();
4142 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4143 :
4144 0 : let mut result = Vec::new();
4145 0 : let mut shard_params: Option<ShardParameters> = None;
4146 :
4147 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4148 : {
4149 0 : let node_id =
4150 0 : shard
4151 0 : .intent
4152 0 : .get_attached()
4153 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4154 0 : "Cannot locate a tenant that is not attached"
4155 0 : )))?;
4156 :
4157 0 : let node = locked
4158 0 : .nodes
4159 0 : .get(&node_id)
4160 0 : .expect("Pageservers may not be deleted while referenced");
4161 0 :
4162 0 : result.push(node.shard_location(*tenant_shard_id));
4163 0 :
4164 0 : match &shard_params {
4165 0 : None => {
4166 0 : shard_params = Some(ShardParameters {
4167 0 : stripe_size: shard.shard.stripe_size,
4168 0 : count: shard.shard.count,
4169 0 : });
4170 0 : }
4171 0 : Some(params) => {
4172 0 : if params.stripe_size != shard.shard.stripe_size {
4173 : // This should never happen. We enforce at runtime because it's simpler than
4174 : // adding an extra per-tenant data structure to store the things that should be the same
4175 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4176 0 : "Inconsistent shard stripe size parameters!"
4177 0 : )));
4178 0 : }
4179 : }
4180 : }
4181 : }
4182 :
4183 0 : if result.is_empty() {
4184 0 : return Err(ApiError::NotFound(
4185 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4186 0 : ));
4187 0 : }
4188 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4189 0 : tracing::info!(
4190 0 : "Located tenant {} with params {:?} on shards {}",
4191 0 : tenant_id,
4192 0 : shard_params,
4193 0 : result
4194 0 : .iter()
4195 0 : .map(|s| format!("{:?}", s))
4196 0 : .collect::<Vec<_>>()
4197 0 : .join(",")
4198 : );
4199 :
4200 0 : Ok(TenantLocateResponse {
4201 0 : shards: result,
4202 0 : shard_params,
4203 0 : })
4204 0 : }
4205 :
4206 : /// Returns None if the input iterator of shards does not include a shard with number=0
4207 0 : fn tenant_describe_impl<'a>(
4208 0 : &self,
4209 0 : shards: impl Iterator<Item = &'a TenantShard>,
4210 0 : ) -> Option<TenantDescribeResponse> {
4211 0 : let mut shard_zero = None;
4212 0 : let mut describe_shards = Vec::new();
4213 :
4214 0 : for shard in shards {
4215 0 : if shard.tenant_shard_id.is_shard_zero() {
4216 0 : shard_zero = Some(shard);
4217 0 : }
4218 :
4219 0 : describe_shards.push(TenantDescribeResponseShard {
4220 0 : tenant_shard_id: shard.tenant_shard_id,
4221 0 : node_attached: *shard.intent.get_attached(),
4222 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4223 0 : last_error: shard
4224 0 : .last_error
4225 0 : .lock()
4226 0 : .unwrap()
4227 0 : .as_ref()
4228 0 : .map(|e| format!("{e}"))
4229 0 : .unwrap_or("".to_string())
4230 0 : .clone(),
4231 0 : is_reconciling: shard.reconciler.is_some(),
4232 0 : is_pending_compute_notification: shard.pending_compute_notification,
4233 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4234 0 : scheduling_policy: *shard.get_scheduling_policy(),
4235 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4236 : })
4237 : }
4238 :
4239 0 : let shard_zero = shard_zero?;
4240 :
4241 0 : Some(TenantDescribeResponse {
4242 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
4243 0 : shards: describe_shards,
4244 0 : stripe_size: shard_zero.shard.stripe_size,
4245 0 : policy: shard_zero.policy.clone(),
4246 0 : config: shard_zero.config.clone(),
4247 0 : })
4248 0 : }
4249 :
4250 0 : pub(crate) fn tenant_describe(
4251 0 : &self,
4252 0 : tenant_id: TenantId,
4253 0 : ) -> Result<TenantDescribeResponse, ApiError> {
4254 0 : let locked = self.inner.read().unwrap();
4255 0 :
4256 0 : self.tenant_describe_impl(
4257 0 : locked
4258 0 : .tenants
4259 0 : .range(TenantShardId::tenant_range(tenant_id))
4260 0 : .map(|(_k, v)| v),
4261 0 : )
4262 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
4263 0 : }
4264 :
4265 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
4266 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
4267 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
4268 : /// in our external API.
4269 0 : pub(crate) fn tenant_list(
4270 0 : &self,
4271 0 : limit: Option<usize>,
4272 0 : start_after: Option<TenantId>,
4273 0 : ) -> Vec<TenantDescribeResponse> {
4274 0 : let locked = self.inner.read().unwrap();
4275 :
4276 : // Apply start_from parameter
4277 0 : let shard_range = match start_after {
4278 0 : None => locked.tenants.range(..),
4279 0 : Some(tenant_id) => locked.tenants.range(
4280 0 : TenantShardId {
4281 0 : tenant_id,
4282 0 : shard_number: ShardNumber(u8::MAX),
4283 0 : shard_count: ShardCount(u8::MAX),
4284 0 : }..,
4285 0 : ),
4286 : };
4287 :
4288 0 : let mut result = Vec::new();
4289 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
4290 0 : result.push(
4291 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
4292 0 : .expect("Groups are always non-empty"),
4293 0 : );
4294 :
4295 : // Enforce `limit` parameter
4296 0 : if let Some(limit) = limit {
4297 0 : if result.len() >= limit {
4298 0 : break;
4299 0 : }
4300 0 : }
4301 : }
4302 :
4303 0 : result
4304 0 : }
4305 :
4306 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
4307 : async fn abort_tenant_shard_split(
4308 : &self,
4309 : op: &TenantShardSplitAbort,
4310 : ) -> Result<(), TenantShardSplitAbortError> {
4311 : // Cleaning up a split:
4312 : // - Parent shards are not destroyed during a split, just detached.
4313 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
4314 : // just the children attached, or both.
4315 : //
4316 : // Therefore our work to do is to:
4317 : // 1. Clean up storage controller's internal state to just refer to parents, no children
4318 : // 2. Call out to pageservers to ensure that children are detached
4319 : // 3. Call out to pageservers to ensure that parents are attached.
4320 : //
4321 : // Crash safety:
4322 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
4323 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
4324 : // and detach them.
4325 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
4326 : // from our database, then we will re-enter this cleanup routine on startup.
4327 :
4328 : let TenantShardSplitAbort {
4329 : tenant_id,
4330 : new_shard_count,
4331 : new_stripe_size,
4332 : ..
4333 : } = op;
4334 :
4335 : // First abort persistent state, if any exists.
4336 : match self
4337 : .persistence
4338 : .abort_shard_split(*tenant_id, *new_shard_count)
4339 : .await?
4340 : {
4341 : AbortShardSplitStatus::Aborted => {
4342 : // Proceed to roll back any child shards created on pageservers
4343 : }
4344 : AbortShardSplitStatus::Complete => {
4345 : // The split completed (we might hit that path if e.g. our database transaction
4346 : // to write the completion landed in the database, but we dropped connection
4347 : // before seeing the result).
4348 : //
4349 : // We must update in-memory state to reflect the successful split.
4350 : self.tenant_shard_split_commit_inmem(
4351 : *tenant_id,
4352 : *new_shard_count,
4353 : *new_stripe_size,
4354 : );
4355 : return Ok(());
4356 : }
4357 : }
4358 :
4359 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
4360 : let detach_locations: Vec<(Node, TenantShardId)> = {
4361 : let mut detach_locations = Vec::new();
4362 : let mut locked = self.inner.write().unwrap();
4363 : let (nodes, tenants, scheduler) = locked.parts_mut();
4364 :
4365 : for (tenant_shard_id, shard) in
4366 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
4367 : {
4368 : if shard.shard.count == op.new_shard_count {
4369 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
4370 : // is infallible, so if we got an error we shouldn't have got that far.
4371 : tracing::warn!(
4372 : "During split abort, child shard {tenant_shard_id} found in-memory"
4373 : );
4374 : continue;
4375 : }
4376 :
4377 : // Add the children of this shard to this list of things to detach
4378 : if let Some(node_id) = shard.intent.get_attached() {
4379 : for child_id in tenant_shard_id.split(*new_shard_count) {
4380 : detach_locations.push((
4381 : nodes
4382 : .get(node_id)
4383 : .expect("Intent references nonexistent node")
4384 : .clone(),
4385 : child_id,
4386 : ));
4387 : }
4388 : } else {
4389 : tracing::warn!(
4390 : "During split abort, shard {tenant_shard_id} has no attached location"
4391 : );
4392 : }
4393 :
4394 : tracing::info!("Restoring parent shard {tenant_shard_id}");
4395 :
4396 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
4397 : // if the original attachment location is offline.
4398 : if let Some(node_id) = shard.intent.get_attached() {
4399 : if !nodes.get(node_id).unwrap().is_available() {
4400 : tracing::info!("Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}");
4401 : shard.intent.demote_attached(scheduler, *node_id);
4402 : }
4403 : }
4404 : for node_id in shard.intent.get_secondary().clone() {
4405 : if !nodes.get(&node_id).unwrap().is_available() {
4406 : tracing::info!("Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}");
4407 : shard.intent.remove_secondary(scheduler, node_id);
4408 : }
4409 : }
4410 :
4411 : shard.splitting = SplitState::Idle;
4412 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
4413 : // If this shard can't be scheduled now (perhaps due to offline nodes or
4414 : // capacity issues), that must not prevent us rolling back a split. In this
4415 : // case it should be eventually scheduled in the background.
4416 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
4417 : }
4418 :
4419 : self.maybe_reconcile_shard(shard, nodes);
4420 : }
4421 :
4422 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
4423 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
4424 :
4425 : detach_locations
4426 : };
4427 :
4428 : for (node, child_id) in detach_locations {
4429 : if !node.is_available() {
4430 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
4431 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
4432 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
4433 : // them from the node.
4434 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
4435 : continue;
4436 : }
4437 :
4438 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
4439 : // a 503 and retry, up to our limit.
4440 : tracing::info!("Detaching {child_id} on {node}...");
4441 : match node
4442 : .with_client_retries(
4443 0 : |client| async move {
4444 0 : let config = LocationConfig {
4445 0 : mode: LocationConfigMode::Detached,
4446 0 : generation: None,
4447 0 : secondary_conf: None,
4448 0 : shard_number: child_id.shard_number.0,
4449 0 : shard_count: child_id.shard_count.literal(),
4450 0 : // Stripe size and tenant config don't matter when detaching
4451 0 : shard_stripe_size: 0,
4452 0 : tenant_conf: TenantConfig::default(),
4453 0 : };
4454 0 :
4455 0 : client.location_config(child_id, config, None, false).await
4456 0 : },
4457 : &self.config.jwt_token,
4458 : 1,
4459 : 10,
4460 : Duration::from_secs(5),
4461 : &self.cancel,
4462 : )
4463 : .await
4464 : {
4465 : Some(Ok(_)) => {}
4466 : Some(Err(e)) => {
4467 : // We failed to communicate with the remote node. This is problematic: we may be
4468 : // leaving it with a rogue child shard.
4469 : tracing::warn!(
4470 : "Failed to detach child {child_id} from node {node} during abort"
4471 : );
4472 : return Err(e.into());
4473 : }
4474 : None => {
4475 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
4476 : // clean up on restart. The node going offline requires a retry.
4477 : return Err(TenantShardSplitAbortError::Unavailable);
4478 : }
4479 : };
4480 : }
4481 :
4482 : tracing::info!("Successfully aborted split");
4483 : Ok(())
4484 : }
4485 :
4486 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
4487 : /// of the tenant map to reflect the child shards that exist after the split.
4488 0 : fn tenant_shard_split_commit_inmem(
4489 0 : &self,
4490 0 : tenant_id: TenantId,
4491 0 : new_shard_count: ShardCount,
4492 0 : new_stripe_size: Option<ShardStripeSize>,
4493 0 : ) -> (
4494 0 : TenantShardSplitResponse,
4495 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
4496 0 : Vec<ReconcilerWaiter>,
4497 0 : ) {
4498 0 : let mut response = TenantShardSplitResponse {
4499 0 : new_shards: Vec::new(),
4500 0 : };
4501 0 : let mut child_locations = Vec::new();
4502 0 : let mut waiters = Vec::new();
4503 0 :
4504 0 : {
4505 0 : let mut locked = self.inner.write().unwrap();
4506 0 :
4507 0 : let parent_ids = locked
4508 0 : .tenants
4509 0 : .range(TenantShardId::tenant_range(tenant_id))
4510 0 : .map(|(shard_id, _)| *shard_id)
4511 0 : .collect::<Vec<_>>();
4512 0 :
4513 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4514 0 : for parent_id in parent_ids {
4515 0 : let child_ids = parent_id.split(new_shard_count);
4516 :
4517 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
4518 0 : let mut old_state = tenants
4519 0 : .remove(&parent_id)
4520 0 : .expect("It was present, we just split it");
4521 0 :
4522 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
4523 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
4524 0 : // nothing else can clear this.
4525 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
4526 :
4527 0 : let old_attached = old_state.intent.get_attached().unwrap();
4528 0 : old_state.intent.clear(scheduler);
4529 0 : let generation = old_state.generation.expect("Shard must have been attached");
4530 0 : (
4531 0 : old_attached,
4532 0 : generation,
4533 0 : old_state.policy.clone(),
4534 0 : old_state.shard,
4535 0 : old_state.config.clone(),
4536 0 : old_state.preferred_az().cloned(),
4537 0 : )
4538 0 : };
4539 0 :
4540 0 : let mut schedule_context = ScheduleContext::default();
4541 0 : for child in child_ids {
4542 0 : let mut child_shard = parent_ident;
4543 0 : child_shard.number = child.shard_number;
4544 0 : child_shard.count = child.shard_count;
4545 0 : if let Some(stripe_size) = new_stripe_size {
4546 0 : child_shard.stripe_size = stripe_size;
4547 0 : }
4548 :
4549 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
4550 0 : child_observed.insert(
4551 0 : pageserver,
4552 0 : ObservedStateLocation {
4553 0 : conf: Some(attached_location_conf(
4554 0 : generation,
4555 0 : &child_shard,
4556 0 : &config,
4557 0 : &policy,
4558 0 : )),
4559 0 : },
4560 0 : );
4561 0 :
4562 0 : let mut child_state =
4563 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
4564 0 : child_state.intent =
4565 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
4566 0 : child_state.observed = ObservedState {
4567 0 : locations: child_observed,
4568 0 : };
4569 0 : child_state.generation = Some(generation);
4570 0 : child_state.config = config.clone();
4571 0 :
4572 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
4573 0 : // as at this point in the split process we have succeeded and this part is infallible:
4574 0 : // we will never need to do any special recovery from this state.
4575 0 :
4576 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
4577 :
4578 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
4579 : // This is not fatal, because we've implicitly already got an attached
4580 : // location for the child shard. Failure here just means we couldn't
4581 : // find a secondary (e.g. because cluster is overloaded).
4582 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
4583 0 : }
4584 : // In the background, attach secondary locations for the new shards
4585 0 : if let Some(waiter) = self.maybe_reconcile_shard(&mut child_state, nodes) {
4586 0 : waiters.push(waiter);
4587 0 : }
4588 :
4589 0 : tenants.insert(child, child_state);
4590 0 : response.new_shards.push(child);
4591 : }
4592 : }
4593 0 : (response, child_locations, waiters)
4594 0 : }
4595 0 : }
4596 :
4597 0 : async fn tenant_shard_split_start_secondaries(
4598 0 : &self,
4599 0 : tenant_id: TenantId,
4600 0 : waiters: Vec<ReconcilerWaiter>,
4601 0 : ) {
4602 : // Wait for initial reconcile of child shards, this creates the secondary locations
4603 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
4604 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
4605 : // their secondaries couldn't be attached.
4606 0 : tracing::warn!("Failed to reconcile after split: {e}");
4607 0 : return;
4608 0 : }
4609 :
4610 : // Take the state lock to discover the attached & secondary intents for all shards
4611 0 : let (attached, secondary) = {
4612 0 : let locked = self.inner.read().unwrap();
4613 0 : let mut attached = Vec::new();
4614 0 : let mut secondary = Vec::new();
4615 :
4616 0 : for (tenant_shard_id, shard) in
4617 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4618 : {
4619 0 : let Some(node_id) = shard.intent.get_attached() else {
4620 : // Unexpected. Race with a PlacementPolicy change?
4621 0 : tracing::warn!(
4622 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
4623 : );
4624 0 : continue;
4625 : };
4626 :
4627 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
4628 : // No secondary location. Nothing for us to do.
4629 0 : continue;
4630 : };
4631 :
4632 0 : let attached_node = locked
4633 0 : .nodes
4634 0 : .get(node_id)
4635 0 : .expect("Pageservers may not be deleted while referenced");
4636 0 :
4637 0 : let secondary_node = locked
4638 0 : .nodes
4639 0 : .get(secondary_node_id)
4640 0 : .expect("Pageservers may not be deleted while referenced");
4641 0 :
4642 0 : attached.push((*tenant_shard_id, attached_node.clone()));
4643 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
4644 : }
4645 0 : (attached, secondary)
4646 0 : };
4647 0 :
4648 0 : if secondary.is_empty() {
4649 : // No secondary locations; nothing for us to do
4650 0 : return;
4651 0 : }
4652 :
4653 0 : for result in self
4654 0 : .tenant_for_shards_api(
4655 0 : attached,
4656 0 : |tenant_shard_id, client| async move {
4657 0 : client.tenant_heatmap_upload(tenant_shard_id).await
4658 0 : },
4659 0 : 1,
4660 0 : 1,
4661 0 : SHORT_RECONCILE_TIMEOUT,
4662 0 : &self.cancel,
4663 0 : )
4664 0 : .await
4665 : {
4666 0 : if let Err(e) = result {
4667 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
4668 0 : return;
4669 0 : }
4670 : }
4671 :
4672 0 : for result in self
4673 0 : .tenant_for_shards_api(
4674 0 : secondary,
4675 0 : |tenant_shard_id, client| async move {
4676 0 : client
4677 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
4678 0 : .await
4679 0 : },
4680 0 : 1,
4681 0 : 1,
4682 0 : SHORT_RECONCILE_TIMEOUT,
4683 0 : &self.cancel,
4684 0 : )
4685 0 : .await
4686 : {
4687 0 : if let Err(e) = result {
4688 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
4689 0 : return;
4690 0 : }
4691 : }
4692 0 : }
4693 :
4694 0 : pub(crate) async fn tenant_shard_split(
4695 0 : &self,
4696 0 : tenant_id: TenantId,
4697 0 : split_req: TenantShardSplitRequest,
4698 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
4699 : // TODO: return 503 if we get stuck waiting for this lock
4700 : // (issue https://github.com/neondatabase/neon/issues/7108)
4701 0 : let _tenant_lock = trace_exclusive_lock(
4702 0 : &self.tenant_op_locks,
4703 0 : tenant_id,
4704 0 : TenantOperations::ShardSplit,
4705 0 : )
4706 0 : .await;
4707 :
4708 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
4709 0 : let new_stripe_size = split_req.new_stripe_size;
4710 :
4711 : // Validate the request and construct parameters. This phase is fallible, but does not require
4712 : // rollback on errors, as it does no I/O and mutates no state.
4713 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
4714 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
4715 0 : ShardSplitAction::Split(params) => params,
4716 : };
4717 :
4718 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
4719 : // we must roll back.
4720 0 : let r = self
4721 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
4722 0 : .await;
4723 :
4724 0 : let (response, waiters) = match r {
4725 0 : Ok(r) => r,
4726 0 : Err(e) => {
4727 0 : // Split might be part-done, we must do work to abort it.
4728 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
4729 0 : self.abort_tx
4730 0 : .send(TenantShardSplitAbort {
4731 0 : tenant_id,
4732 0 : new_shard_count,
4733 0 : new_stripe_size,
4734 0 : _tenant_lock,
4735 0 : })
4736 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
4737 0 : .ok();
4738 0 : return Err(e);
4739 : }
4740 : };
4741 :
4742 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
4743 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
4744 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
4745 : // in [`Self::optimize_all`]
4746 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
4747 0 : .await;
4748 0 : Ok(response)
4749 0 : }
4750 :
4751 0 : fn prepare_tenant_shard_split(
4752 0 : &self,
4753 0 : tenant_id: TenantId,
4754 0 : split_req: TenantShardSplitRequest,
4755 0 : ) -> Result<ShardSplitAction, ApiError> {
4756 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
4757 0 : anyhow::anyhow!("failpoint")
4758 0 : )));
4759 :
4760 0 : let mut policy = None;
4761 0 : let mut config = None;
4762 0 : let mut shard_ident = None;
4763 0 : let mut preferred_az_id = None;
4764 : // Validate input, and calculate which shards we will create
4765 0 : let (old_shard_count, targets) =
4766 : {
4767 0 : let locked = self.inner.read().unwrap();
4768 0 :
4769 0 : let pageservers = locked.nodes.clone();
4770 0 :
4771 0 : let mut targets = Vec::new();
4772 0 :
4773 0 : // In case this is a retry, count how many already-split shards we found
4774 0 : let mut children_found = Vec::new();
4775 0 : let mut old_shard_count = None;
4776 :
4777 0 : for (tenant_shard_id, shard) in
4778 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4779 : {
4780 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
4781 : Ordering::Equal => {
4782 : // Already split this
4783 0 : children_found.push(*tenant_shard_id);
4784 0 : continue;
4785 : }
4786 : Ordering::Greater => {
4787 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4788 0 : "Requested count {} but already have shards at count {}",
4789 0 : split_req.new_shard_count,
4790 0 : shard.shard.count.count()
4791 0 : )));
4792 : }
4793 0 : Ordering::Less => {
4794 0 : // Fall through: this shard has lower count than requested,
4795 0 : // is a candidate for splitting.
4796 0 : }
4797 0 : }
4798 0 :
4799 0 : match old_shard_count {
4800 0 : None => old_shard_count = Some(shard.shard.count),
4801 0 : Some(old_shard_count) => {
4802 0 : if old_shard_count != shard.shard.count {
4803 : // We may hit this case if a caller asked for two splits to
4804 : // different sizes, before the first one is complete.
4805 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
4806 : // of shard_count=1 and shard_count=2 shards in the map.
4807 0 : return Err(ApiError::Conflict(
4808 0 : "Cannot split, currently mid-split".to_string(),
4809 0 : ));
4810 0 : }
4811 : }
4812 : }
4813 0 : if policy.is_none() {
4814 0 : policy = Some(shard.policy.clone());
4815 0 : }
4816 0 : if shard_ident.is_none() {
4817 0 : shard_ident = Some(shard.shard);
4818 0 : }
4819 0 : if config.is_none() {
4820 0 : config = Some(shard.config.clone());
4821 0 : }
4822 0 : if preferred_az_id.is_none() {
4823 0 : preferred_az_id = shard.preferred_az().cloned();
4824 0 : }
4825 :
4826 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4827 0 : tracing::info!(
4828 0 : "Tenant shard {} already has shard count {}",
4829 : tenant_shard_id,
4830 : split_req.new_shard_count
4831 : );
4832 0 : continue;
4833 0 : }
4834 :
4835 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4836 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4837 0 : ))?;
4838 :
4839 0 : let node = pageservers
4840 0 : .get(&node_id)
4841 0 : .expect("Pageservers may not be deleted while referenced");
4842 0 :
4843 0 : targets.push(ShardSplitTarget {
4844 0 : parent_id: *tenant_shard_id,
4845 0 : node: node.clone(),
4846 0 : child_ids: tenant_shard_id
4847 0 : .split(ShardCount::new(split_req.new_shard_count)),
4848 0 : });
4849 : }
4850 :
4851 0 : if targets.is_empty() {
4852 0 : if children_found.len() == split_req.new_shard_count as usize {
4853 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4854 0 : new_shards: children_found,
4855 0 : }));
4856 : } else {
4857 : // No shards found to split, and no existing children found: the
4858 : // tenant doesn't exist at all.
4859 0 : return Err(ApiError::NotFound(
4860 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4861 0 : ));
4862 : }
4863 0 : }
4864 0 :
4865 0 : (old_shard_count, targets)
4866 0 : };
4867 0 :
4868 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
4869 0 : let old_shard_count = old_shard_count.unwrap();
4870 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
4871 : // This ShardIdentity will be used as the template for all children, so this implicitly
4872 : // applies the new stripe size to the children.
4873 0 : let mut shard_ident = shard_ident.unwrap();
4874 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
4875 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards", shard_ident.stripe_size)));
4876 0 : }
4877 0 :
4878 0 : shard_ident.stripe_size = new_stripe_size;
4879 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
4880 0 : shard_ident
4881 : } else {
4882 0 : shard_ident.unwrap()
4883 : };
4884 0 : let policy = policy.unwrap();
4885 0 : let config = config.unwrap();
4886 0 :
4887 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
4888 0 : old_shard_count,
4889 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
4890 0 : new_stripe_size: split_req.new_stripe_size,
4891 0 : targets,
4892 0 : policy,
4893 0 : config,
4894 0 : shard_ident,
4895 0 : preferred_az_id,
4896 0 : })))
4897 0 : }
4898 :
4899 0 : async fn do_tenant_shard_split(
4900 0 : &self,
4901 0 : tenant_id: TenantId,
4902 0 : params: Box<ShardSplitParams>,
4903 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
4904 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
4905 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
4906 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
4907 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
4908 0 : // (https://github.com/neondatabase/neon/issues/6676)
4909 0 :
4910 0 : let ShardSplitParams {
4911 0 : old_shard_count,
4912 0 : new_shard_count,
4913 0 : new_stripe_size,
4914 0 : mut targets,
4915 0 : policy,
4916 0 : config,
4917 0 : shard_ident,
4918 0 : preferred_az_id,
4919 0 : } = *params;
4920 :
4921 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
4922 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
4923 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
4924 : // at the time of split.
4925 0 : let waiters = {
4926 0 : let mut locked = self.inner.write().unwrap();
4927 0 : let mut waiters = Vec::new();
4928 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4929 0 : for target in &mut targets {
4930 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
4931 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4932 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4933 0 : "Shard {} not found",
4934 0 : target.parent_id
4935 0 : )));
4936 : };
4937 :
4938 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
4939 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4940 0 : return Err(ApiError::Conflict(format!(
4941 0 : "Shard {} unexpectedly rescheduled during split",
4942 0 : target.parent_id
4943 0 : )));
4944 0 : }
4945 0 :
4946 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
4947 0 : shard.intent.clear_secondary(scheduler);
4948 :
4949 : // Run Reconciler to execute detach fo secondary locations.
4950 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
4951 0 : waiters.push(waiter);
4952 0 : }
4953 : }
4954 0 : waiters
4955 0 : };
4956 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
4957 :
4958 : // Before creating any new child shards in memory or on the pageservers, persist them: this
4959 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
4960 : // acts as the protection against two concurrent attempts to split: one of them will get a database
4961 : // error trying to insert the child shards.
4962 0 : let mut child_tsps = Vec::new();
4963 0 : for target in &targets {
4964 0 : let mut this_child_tsps = Vec::new();
4965 0 : for child in &target.child_ids {
4966 0 : let mut child_shard = shard_ident;
4967 0 : child_shard.number = child.shard_number;
4968 0 : child_shard.count = child.shard_count;
4969 0 :
4970 0 : tracing::info!(
4971 0 : "Create child shard persistence with stripe size {}",
4972 : shard_ident.stripe_size.0
4973 : );
4974 :
4975 0 : this_child_tsps.push(TenantShardPersistence {
4976 0 : tenant_id: child.tenant_id.to_string(),
4977 0 : shard_number: child.shard_number.0 as i32,
4978 0 : shard_count: child.shard_count.literal() as i32,
4979 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
4980 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
4981 0 : // populate the correct generation as part of its transaction, to protect us
4982 0 : // against racing with changes in the state of the parent.
4983 0 : generation: None,
4984 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
4985 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
4986 0 : config: serde_json::to_string(&config).unwrap(),
4987 0 : splitting: SplitState::Splitting,
4988 0 :
4989 0 : // Scheduling policies and preferred AZ do not carry through to children
4990 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
4991 0 : .unwrap(),
4992 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
4993 0 : });
4994 0 : }
4995 :
4996 0 : child_tsps.push((target.parent_id, this_child_tsps));
4997 : }
4998 :
4999 0 : if let Err(e) = self
5000 0 : .persistence
5001 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
5002 0 : .await
5003 : {
5004 0 : match e {
5005 : DatabaseError::Query(diesel::result::Error::DatabaseError(
5006 : DatabaseErrorKind::UniqueViolation,
5007 : _,
5008 : )) => {
5009 : // Inserting a child shard violated a unique constraint: we raced with another call to
5010 : // this function
5011 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
5012 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
5013 : }
5014 0 : _ => return Err(ApiError::InternalServerError(e.into())),
5015 : }
5016 0 : }
5017 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
5018 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5019 0 : ));
5020 :
5021 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
5022 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
5023 : // is not set in memory, then it was not persisted.
5024 : {
5025 0 : let mut locked = self.inner.write().unwrap();
5026 0 : for target in &targets {
5027 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
5028 0 : parent_shard.splitting = SplitState::Splitting;
5029 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
5030 0 : // split operation.
5031 0 : parent_shard
5032 0 : .observed
5033 0 : .locations
5034 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
5035 0 : }
5036 : }
5037 : }
5038 :
5039 : // TODO: issue split calls concurrently (this only matters once we're splitting
5040 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
5041 :
5042 0 : for target in &targets {
5043 : let ShardSplitTarget {
5044 0 : parent_id,
5045 0 : node,
5046 0 : child_ids,
5047 0 : } = target;
5048 0 : let client = PageserverClient::new(
5049 0 : node.get_id(),
5050 0 : node.base_url(),
5051 0 : self.config.jwt_token.as_deref(),
5052 0 : );
5053 0 : let response = client
5054 0 : .tenant_shard_split(
5055 0 : *parent_id,
5056 0 : TenantShardSplitRequest {
5057 0 : new_shard_count: new_shard_count.literal(),
5058 0 : new_stripe_size,
5059 0 : },
5060 0 : )
5061 0 : .await
5062 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
5063 :
5064 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
5065 0 : "failpoint".to_string()
5066 0 : )));
5067 :
5068 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
5069 :
5070 0 : tracing::info!(
5071 0 : "Split {} into {}",
5072 0 : parent_id,
5073 0 : response
5074 0 : .new_shards
5075 0 : .iter()
5076 0 : .map(|s| format!("{:?}", s))
5077 0 : .collect::<Vec<_>>()
5078 0 : .join(",")
5079 : );
5080 :
5081 0 : if &response.new_shards != child_ids {
5082 : // This should never happen: the pageserver should agree with us on how shard splits work.
5083 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5084 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
5085 0 : parent_id,
5086 0 : response.new_shards,
5087 0 : child_ids
5088 0 : )));
5089 0 : }
5090 : }
5091 :
5092 : // TODO: if the pageserver restarted concurrently with our split API call,
5093 : // the actual generation of the child shard might differ from the generation
5094 : // we expect it to have. In order for our in-database generation to end up
5095 : // correct, we should carry the child generation back in the response and apply it here
5096 : // in complete_shard_split (and apply the correct generation in memory)
5097 : // (or, we can carry generation in the request and reject the request if
5098 : // it doesn't match, but that requires more retry logic on this side)
5099 :
5100 0 : self.persistence
5101 0 : .complete_shard_split(tenant_id, old_shard_count)
5102 0 : .await?;
5103 :
5104 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5105 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5106 0 : ));
5107 :
5108 : // Replace all the shards we just split with their children: this phase is infallible.
5109 0 : let (response, child_locations, waiters) =
5110 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5111 0 :
5112 0 : // Send compute notifications for all the new shards
5113 0 : let mut failed_notifications = Vec::new();
5114 0 : for (child_id, child_ps, stripe_size) in child_locations {
5115 0 : if let Err(e) = self
5116 0 : .compute_hook
5117 0 : .notify(
5118 0 : compute_hook::ShardUpdate {
5119 0 : tenant_shard_id: child_id,
5120 0 : node_id: child_ps,
5121 0 : stripe_size,
5122 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5123 0 : },
5124 0 : &self.cancel,
5125 0 : )
5126 0 : .await
5127 : {
5128 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5129 : child_id, child_ps);
5130 0 : failed_notifications.push(child_id);
5131 0 : }
5132 : }
5133 :
5134 : // If we failed any compute notifications, make a note to retry later.
5135 0 : if !failed_notifications.is_empty() {
5136 0 : let mut locked = self.inner.write().unwrap();
5137 0 : for failed in failed_notifications {
5138 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5139 0 : shard.pending_compute_notification = true;
5140 0 : }
5141 : }
5142 0 : }
5143 :
5144 0 : Ok((response, waiters))
5145 0 : }
5146 :
5147 0 : pub(crate) async fn tenant_shard_migrate(
5148 0 : &self,
5149 0 : tenant_shard_id: TenantShardId,
5150 0 : migrate_req: TenantShardMigrateRequest,
5151 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5152 0 : let waiter = {
5153 0 : let mut locked = self.inner.write().unwrap();
5154 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5155 :
5156 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5157 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5158 0 : "Node {} not found",
5159 0 : migrate_req.node_id
5160 0 : )));
5161 : };
5162 :
5163 0 : if !node.is_available() {
5164 : // Warn but proceed: the caller may intend to manually adjust the placement of
5165 : // a shard even if the node is down, e.g. if intervening during an incident.
5166 0 : tracing::warn!("Migrating to unavailable node {node}");
5167 0 : }
5168 :
5169 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5170 0 : return Err(ApiError::NotFound(
5171 0 : anyhow::anyhow!("Tenant shard not found").into(),
5172 0 : ));
5173 : };
5174 :
5175 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5176 : // No-op case: we will still proceed to wait for reconciliation in case it is
5177 : // incomplete from an earlier update to the intent.
5178 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
5179 : } else {
5180 0 : let old_attached = *shard.intent.get_attached();
5181 0 :
5182 0 : match shard.policy {
5183 0 : PlacementPolicy::Attached(n) => {
5184 0 : // If our new attached node was a secondary, it no longer should be.
5185 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
5186 0 :
5187 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
5188 :
5189 : // If we were already attached to something, demote that to a secondary
5190 0 : if let Some(old_attached) = old_attached {
5191 0 : if n > 0 {
5192 : // Remove other secondaries to make room for the location we'll demote
5193 0 : while shard.intent.get_secondary().len() >= n {
5194 0 : shard.intent.pop_secondary(scheduler);
5195 0 : }
5196 :
5197 0 : shard.intent.push_secondary(scheduler, old_attached);
5198 0 : }
5199 0 : }
5200 : }
5201 0 : PlacementPolicy::Secondary => {
5202 0 : shard.intent.clear(scheduler);
5203 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5204 0 : }
5205 : PlacementPolicy::Detached => {
5206 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5207 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
5208 0 : )))
5209 : }
5210 : }
5211 :
5212 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
5213 0 : shard.sequence = shard.sequence.next();
5214 : }
5215 :
5216 0 : let reconciler_config = match migrate_req.migration_config {
5217 0 : Some(cfg) => (&cfg).into(),
5218 0 : None => ReconcilerConfig::default(),
5219 : };
5220 :
5221 0 : self.maybe_configured_reconcile_shard(shard, nodes, reconciler_config)
5222 : };
5223 :
5224 0 : if let Some(waiter) = waiter {
5225 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5226 : } else {
5227 0 : tracing::info!("Migration is a no-op");
5228 : }
5229 :
5230 0 : Ok(TenantShardMigrateResponse {})
5231 0 : }
5232 :
5233 0 : pub(crate) async fn tenant_shard_migrate_secondary(
5234 0 : &self,
5235 0 : tenant_shard_id: TenantShardId,
5236 0 : migrate_req: TenantShardMigrateRequest,
5237 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5238 0 : let waiter = {
5239 0 : let mut locked = self.inner.write().unwrap();
5240 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5241 :
5242 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5243 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5244 0 : "Node {} not found",
5245 0 : migrate_req.node_id
5246 0 : )));
5247 : };
5248 :
5249 0 : if !node.is_available() {
5250 : // Warn but proceed: the caller may intend to manually adjust the placement of
5251 : // a shard even if the node is down, e.g. if intervening during an incident.
5252 0 : tracing::warn!("Migrating to unavailable node {node}");
5253 0 : }
5254 :
5255 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5256 0 : return Err(ApiError::NotFound(
5257 0 : anyhow::anyhow!("Tenant shard not found").into(),
5258 0 : ));
5259 : };
5260 :
5261 0 : if shard.intent.get_secondary().len() == 1
5262 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
5263 : {
5264 0 : tracing::info!(
5265 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
5266 : shard.intent
5267 : );
5268 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5269 0 : tracing::info!("Migrating secondary to {node}: already attached where we were asked to create a secondary");
5270 : } else {
5271 0 : let old_secondaries = shard.intent.get_secondary().clone();
5272 0 : for secondary in old_secondaries {
5273 0 : shard.intent.remove_secondary(scheduler, secondary);
5274 0 : }
5275 :
5276 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5277 0 : shard.sequence = shard.sequence.next();
5278 0 : tracing::info!(
5279 0 : "Migrating secondary to {node}: new intent {:?}",
5280 : shard.intent
5281 : );
5282 : }
5283 :
5284 0 : self.maybe_reconcile_shard(shard, nodes)
5285 : };
5286 :
5287 0 : if let Some(waiter) = waiter {
5288 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5289 : } else {
5290 0 : tracing::info!("Migration is a no-op");
5291 : }
5292 :
5293 0 : Ok(TenantShardMigrateResponse {})
5294 0 : }
5295 :
5296 : /// 'cancel' in this context means cancel any ongoing reconcile
5297 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
5298 0 : &self,
5299 0 : tenant_shard_id: TenantShardId,
5300 0 : ) -> Result<(), ApiError> {
5301 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
5302 0 : let waiter = {
5303 0 : let locked = self.inner.write().unwrap();
5304 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5305 0 : return Err(ApiError::NotFound(
5306 0 : anyhow::anyhow!("Tenant shard not found").into(),
5307 0 : ));
5308 : };
5309 :
5310 0 : let waiter = shard.get_waiter();
5311 0 : match waiter {
5312 : None => {
5313 0 : tracing::info!("Shard does not have an ongoing Reconciler");
5314 0 : return Ok(());
5315 : }
5316 0 : Some(waiter) => {
5317 0 : tracing::info!("Cancelling Reconciler");
5318 0 : shard.cancel_reconciler();
5319 0 : waiter
5320 0 : }
5321 0 : }
5322 0 : };
5323 0 :
5324 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
5325 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
5326 0 : // the Reconciler is misbehaving and not respecting the cancellation token
5327 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
5328 0 : .await?;
5329 :
5330 0 : Ok(())
5331 0 : }
5332 :
5333 : /// This is for debug/support only: we simply drop all state for a tenant, without
5334 : /// detaching or deleting it on pageservers.
5335 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
5336 0 : self.persistence.delete_tenant(tenant_id).await?;
5337 :
5338 0 : let mut locked = self.inner.write().unwrap();
5339 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
5340 0 : let mut shards = Vec::new();
5341 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
5342 0 : shards.push(*tenant_shard_id);
5343 0 : }
5344 :
5345 0 : for shard_id in shards {
5346 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
5347 0 : shard.intent.clear(scheduler);
5348 0 : }
5349 : }
5350 :
5351 0 : Ok(())
5352 0 : }
5353 :
5354 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
5355 : /// tenant with a very high generation number so that it will see the existing data.
5356 0 : pub(crate) async fn tenant_import(
5357 0 : &self,
5358 0 : tenant_id: TenantId,
5359 0 : ) -> Result<TenantCreateResponse, ApiError> {
5360 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
5361 0 : let maybe_node = {
5362 0 : self.inner
5363 0 : .read()
5364 0 : .unwrap()
5365 0 : .nodes
5366 0 : .values()
5367 0 : .find(|n| n.is_available())
5368 0 : .cloned()
5369 : };
5370 0 : let Some(node) = maybe_node else {
5371 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
5372 : };
5373 :
5374 0 : let client = PageserverClient::new(
5375 0 : node.get_id(),
5376 0 : node.base_url(),
5377 0 : self.config.jwt_token.as_deref(),
5378 0 : );
5379 :
5380 0 : let scan_result = client
5381 0 : .tenant_scan_remote_storage(tenant_id)
5382 0 : .await
5383 0 : .map_err(|e| passthrough_api_error(&node, e))?;
5384 :
5385 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
5386 0 : let Some(shard_count) = scan_result
5387 0 : .shards
5388 0 : .iter()
5389 0 : .map(|s| s.tenant_shard_id.shard_count)
5390 0 : .max()
5391 : else {
5392 0 : return Err(ApiError::NotFound(
5393 0 : anyhow::anyhow!("No shards found").into(),
5394 0 : ));
5395 : };
5396 :
5397 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
5398 : // to
5399 0 : let generation = scan_result
5400 0 : .shards
5401 0 : .iter()
5402 0 : .map(|s| s.generation)
5403 0 : .max()
5404 0 : .expect("We already validated >0 shards");
5405 0 :
5406 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
5407 0 : // only work if they were using the default stripe size.
5408 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
5409 :
5410 0 : let (response, waiters) = self
5411 0 : .do_tenant_create(TenantCreateRequest {
5412 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
5413 0 : generation,
5414 0 :
5415 0 : shard_parameters: ShardParameters {
5416 0 : count: shard_count,
5417 0 : stripe_size,
5418 0 : },
5419 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
5420 0 : config: TenantConfig::default(),
5421 0 : })
5422 0 : .await?;
5423 :
5424 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
5425 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
5426 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
5427 : // reconcile, as reconciliation includes notifying compute.
5428 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
5429 0 : }
5430 :
5431 0 : Ok(response)
5432 0 : }
5433 :
5434 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
5435 : /// we don't have to make TenantShard clonable in the return path.
5436 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5437 0 : let serialized = {
5438 0 : let locked = self.inner.read().unwrap();
5439 0 : let result = locked.tenants.values().collect::<Vec<_>>();
5440 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
5441 : };
5442 :
5443 0 : hyper::Response::builder()
5444 0 : .status(hyper::StatusCode::OK)
5445 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5446 0 : .body(hyper::Body::from(serialized))
5447 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5448 0 : }
5449 :
5450 : /// Check the consistency of in-memory state vs. persistent state, and check that the
5451 : /// scheduler's statistics are up to date.
5452 : ///
5453 : /// These consistency checks expect an **idle** system. If changes are going on while
5454 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
5455 : /// checks, but not suitable for running continuously in the background in the field.
5456 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
5457 0 : let (mut expect_nodes, mut expect_shards) = {
5458 0 : let locked = self.inner.read().unwrap();
5459 0 :
5460 0 : locked
5461 0 : .scheduler
5462 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
5463 0 : .context("Scheduler checks")
5464 0 : .map_err(ApiError::InternalServerError)?;
5465 :
5466 0 : let expect_nodes = locked
5467 0 : .nodes
5468 0 : .values()
5469 0 : .map(|n| n.to_persistent())
5470 0 : .collect::<Vec<_>>();
5471 0 :
5472 0 : let expect_shards = locked
5473 0 : .tenants
5474 0 : .values()
5475 0 : .map(|t| t.to_persistent())
5476 0 : .collect::<Vec<_>>();
5477 :
5478 : // This method can only validate the state of an idle system: if a reconcile is in
5479 : // progress, fail out early to avoid giving false errors on state that won't match
5480 : // between database and memory under a ReconcileResult is processed.
5481 0 : for t in locked.tenants.values() {
5482 0 : if t.reconciler.is_some() {
5483 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5484 0 : "Shard {} reconciliation in progress",
5485 0 : t.tenant_shard_id
5486 0 : )));
5487 0 : }
5488 : }
5489 :
5490 0 : (expect_nodes, expect_shards)
5491 : };
5492 :
5493 0 : let mut nodes = self.persistence.list_nodes().await?;
5494 0 : expect_nodes.sort_by_key(|n| n.node_id);
5495 0 : nodes.sort_by_key(|n| n.node_id);
5496 :
5497 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
5498 0 : let node_result = if nodes != expect_nodes {
5499 0 : tracing::error!("Consistency check failed on nodes.");
5500 0 : tracing::error!(
5501 0 : "Nodes in memory: {}",
5502 0 : serde_json::to_string(&expect_nodes)
5503 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5504 : );
5505 0 : tracing::error!(
5506 0 : "Nodes in database: {}",
5507 0 : serde_json::to_string(&nodes)
5508 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5509 : );
5510 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
5511 0 : "Node consistency failure"
5512 0 : )))
5513 : } else {
5514 0 : Ok(())
5515 : };
5516 :
5517 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
5518 0 : persistent_shards
5519 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5520 0 :
5521 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5522 :
5523 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
5524 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
5525 : // fields are added, before doing a comparison.
5526 0 : for tsp in &mut persistent_shards {
5527 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
5528 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
5529 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
5530 : }
5531 :
5532 0 : if persistent_shards != expect_shards {
5533 0 : tracing::error!("Consistency check failed on shards.");
5534 :
5535 0 : tracing::error!(
5536 0 : "Shards in memory: {}",
5537 0 : serde_json::to_string(&expect_shards)
5538 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5539 : );
5540 0 : tracing::error!(
5541 0 : "Shards in database: {}",
5542 0 : serde_json::to_string(&persistent_shards)
5543 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5544 : );
5545 :
5546 : // The total dump log lines above are useful in testing but in the field grafana will
5547 : // usually just drop them because they're so large. So we also do some explicit logging
5548 : // of just the diffs.
5549 0 : let persistent_shards = persistent_shards
5550 0 : .into_iter()
5551 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5552 0 : .collect::<HashMap<_, _>>();
5553 0 : let expect_shards = expect_shards
5554 0 : .into_iter()
5555 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5556 0 : .collect::<HashMap<_, _>>();
5557 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
5558 0 : match expect_shards.get(tenant_shard_id) {
5559 : None => {
5560 0 : tracing::error!(
5561 0 : "Shard {} found in database but not in memory",
5562 : tenant_shard_id
5563 : );
5564 : }
5565 0 : Some(expect_tsp) => {
5566 0 : if expect_tsp != persistent_tsp {
5567 0 : tracing::error!(
5568 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
5569 0 : tenant_shard_id,
5570 0 : serde_json::to_string(expect_tsp).unwrap(),
5571 0 : serde_json::to_string(&persistent_tsp).unwrap()
5572 : );
5573 0 : }
5574 : }
5575 : }
5576 : }
5577 :
5578 : // Having already logged any differences, log any shards that simply aren't present in the database
5579 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
5580 0 : if !persistent_shards.contains_key(tenant_shard_id) {
5581 0 : tracing::error!(
5582 0 : "Shard {} found in memory but not in database: {}",
5583 0 : tenant_shard_id,
5584 0 : serde_json::to_string(memory_tsp)
5585 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5586 : );
5587 0 : }
5588 : }
5589 :
5590 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5591 0 : "Shard consistency failure"
5592 0 : )));
5593 0 : }
5594 0 :
5595 0 : node_result
5596 0 : }
5597 :
5598 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
5599 : /// we don't have to make TenantShard clonable in the return path.
5600 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5601 0 : let serialized = {
5602 0 : let locked = self.inner.read().unwrap();
5603 0 : serde_json::to_string(&locked.scheduler)
5604 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5605 : };
5606 :
5607 0 : hyper::Response::builder()
5608 0 : .status(hyper::StatusCode::OK)
5609 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5610 0 : .body(hyper::Body::from(serialized))
5611 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5612 0 : }
5613 :
5614 : /// This is for debug/support only: we simply drop all state for a tenant, without
5615 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
5616 : /// tenants that were on this node.
5617 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
5618 0 : self.persistence.delete_node(node_id).await?;
5619 :
5620 0 : let mut locked = self.inner.write().unwrap();
5621 :
5622 0 : for shard in locked.tenants.values_mut() {
5623 0 : shard.deref_node(node_id);
5624 0 : shard.observed.locations.remove(&node_id);
5625 0 : }
5626 :
5627 0 : let mut nodes = (*locked.nodes).clone();
5628 0 : nodes.remove(&node_id);
5629 0 : locked.nodes = Arc::new(nodes);
5630 0 : metrics::METRICS_REGISTRY
5631 0 : .metrics_group
5632 0 : .storage_controller_pageserver_nodes
5633 0 : .set(locked.nodes.len() as i64);
5634 0 :
5635 0 : locked.scheduler.node_remove(node_id);
5636 0 :
5637 0 : Ok(())
5638 0 : }
5639 :
5640 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
5641 : /// that we don't leave any bad state behind in the storage controller, but unclean
5642 : /// in the sense that we are not carefully draining the node.
5643 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
5644 0 : let _node_lock =
5645 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
5646 :
5647 : // 1. Atomically update in-memory state:
5648 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
5649 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
5650 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
5651 : // re-insert references to this node into the ObservedState of shards
5652 : // - drop the node from the scheduler
5653 : {
5654 0 : let mut locked = self.inner.write().unwrap();
5655 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5656 0 :
5657 0 : {
5658 0 : let mut nodes_mut = (*nodes).deref().clone();
5659 0 : match nodes_mut.get_mut(&node_id) {
5660 0 : Some(node) => {
5661 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
5662 0 : // if we crash it would not be desirable to leave the node paused after a restart.
5663 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
5664 0 : }
5665 : None => {
5666 0 : tracing::info!(
5667 0 : "Node not found: presuming this is a retry and returning success"
5668 : );
5669 0 : return Ok(());
5670 : }
5671 : }
5672 :
5673 0 : *nodes = Arc::new(nodes_mut);
5674 : }
5675 :
5676 0 : for (_tenant_id, mut schedule_context, shards) in
5677 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
5678 : {
5679 0 : for shard in shards {
5680 0 : if shard.deref_node(node_id) {
5681 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
5682 : // TODO: implement force flag to remove a node even if we can't reschedule
5683 : // a tenant
5684 0 : tracing::error!(
5685 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
5686 : shard.tenant_shard_id
5687 : );
5688 0 : return Err(e.into());
5689 : } else {
5690 0 : tracing::info!(
5691 0 : "Rescheduled shard {} away from node during deletion",
5692 : shard.tenant_shard_id
5693 : )
5694 : }
5695 :
5696 0 : self.maybe_reconcile_shard(shard, nodes);
5697 0 : }
5698 :
5699 : // Here we remove an existing observed location for the node we're removing, and it will
5700 : // not be re-added by a reconciler's completion because we filter out removed nodes in
5701 : // process_result.
5702 : //
5703 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
5704 : // means any reconciles we spawned will know about the node we're deleting, enabling them
5705 : // to do live migrations if it's still online.
5706 0 : shard.observed.locations.remove(&node_id);
5707 : }
5708 : }
5709 :
5710 0 : scheduler.node_remove(node_id);
5711 0 :
5712 0 : {
5713 0 : let mut nodes_mut = (**nodes).clone();
5714 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
5715 0 : // Ensure that any reconciler holding an Arc<> to this node will
5716 0 : // drop out when trying to RPC to it (setting Offline state sets the
5717 0 : // cancellation token on the Node object).
5718 0 : removed_node.set_availability(NodeAvailability::Offline);
5719 0 : }
5720 0 : *nodes = Arc::new(nodes_mut);
5721 0 : metrics::METRICS_REGISTRY
5722 0 : .metrics_group
5723 0 : .storage_controller_pageserver_nodes
5724 0 : .set(nodes.len() as i64);
5725 0 : }
5726 0 : }
5727 0 :
5728 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
5729 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
5730 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
5731 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
5732 0 : // that exists.
5733 0 :
5734 0 : // 2. Actually delete the node from the database and from in-memory state
5735 0 : tracing::info!("Deleting node from database");
5736 0 : self.persistence.delete_node(node_id).await?;
5737 :
5738 0 : Ok(())
5739 0 : }
5740 :
5741 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
5742 0 : let nodes = {
5743 0 : self.inner
5744 0 : .read()
5745 0 : .unwrap()
5746 0 : .nodes
5747 0 : .values()
5748 0 : .cloned()
5749 0 : .collect::<Vec<_>>()
5750 0 : };
5751 0 :
5752 0 : Ok(nodes)
5753 0 : }
5754 :
5755 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
5756 0 : self.inner
5757 0 : .read()
5758 0 : .unwrap()
5759 0 : .nodes
5760 0 : .get(&node_id)
5761 0 : .cloned()
5762 0 : .ok_or(ApiError::NotFound(
5763 0 : format!("Node {node_id} not registered").into(),
5764 0 : ))
5765 0 : }
5766 :
5767 0 : pub(crate) async fn get_node_shards(
5768 0 : &self,
5769 0 : node_id: NodeId,
5770 0 : ) -> Result<NodeShardResponse, ApiError> {
5771 0 : let locked = self.inner.read().unwrap();
5772 0 : let mut shards = Vec::new();
5773 0 : for (tid, tenant) in locked.tenants.iter() {
5774 0 : let is_intended_secondary = match (
5775 0 : tenant.intent.get_attached() == &Some(node_id),
5776 0 : tenant.intent.get_secondary().contains(&node_id),
5777 0 : ) {
5778 : (true, true) => {
5779 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5780 0 : "{} attached as primary+secondary on the same node",
5781 0 : tid
5782 0 : )))
5783 : }
5784 0 : (true, false) => Some(false),
5785 0 : (false, true) => Some(true),
5786 0 : (false, false) => None,
5787 : };
5788 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
5789 0 : tenant.observed.locations.get(&node_id)
5790 : {
5791 0 : Some(conf.secondary_conf.is_some())
5792 : } else {
5793 0 : None
5794 : };
5795 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
5796 0 : shards.push(NodeShard {
5797 0 : tenant_shard_id: *tid,
5798 0 : is_intended_secondary,
5799 0 : is_observed_secondary,
5800 0 : });
5801 0 : }
5802 : }
5803 0 : Ok(NodeShardResponse { node_id, shards })
5804 0 : }
5805 :
5806 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
5807 0 : self.persistence.get_leader().await
5808 0 : }
5809 :
5810 0 : pub(crate) async fn node_register(
5811 0 : &self,
5812 0 : register_req: NodeRegisterRequest,
5813 0 : ) -> Result<(), ApiError> {
5814 0 : let _node_lock = trace_exclusive_lock(
5815 0 : &self.node_op_locks,
5816 0 : register_req.node_id,
5817 0 : NodeOperations::Register,
5818 0 : )
5819 0 : .await;
5820 :
5821 : enum RegistrationStatus {
5822 : Matched,
5823 : Mismatched,
5824 : New,
5825 : }
5826 :
5827 0 : let registration_status = {
5828 0 : let locked = self.inner.read().unwrap();
5829 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
5830 0 : if node.registration_match(®ister_req) {
5831 0 : RegistrationStatus::Matched
5832 : } else {
5833 0 : RegistrationStatus::Mismatched
5834 : }
5835 : } else {
5836 0 : RegistrationStatus::New
5837 : }
5838 : };
5839 :
5840 0 : match registration_status {
5841 : RegistrationStatus::Matched => {
5842 0 : tracing::info!(
5843 0 : "Node {} re-registered with matching address",
5844 : register_req.node_id
5845 : );
5846 :
5847 0 : return Ok(());
5848 : }
5849 : RegistrationStatus::Mismatched => {
5850 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
5851 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
5852 : // a fixed address through the lifetime of a node.
5853 0 : tracing::warn!(
5854 0 : "Node {} tried to register with different address",
5855 : register_req.node_id
5856 : );
5857 0 : return Err(ApiError::Conflict(
5858 0 : "Node is already registered with different address".to_string(),
5859 0 : ));
5860 : }
5861 0 : RegistrationStatus::New => {
5862 0 : // fallthrough
5863 0 : }
5864 0 : }
5865 0 :
5866 0 : // We do not require that a node is actually online when registered (it will start life
5867 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
5868 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
5869 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
5870 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
5871 0 : if tokio::net::lookup_host(format!(
5872 0 : "{}:{}",
5873 0 : register_req.listen_http_addr, register_req.listen_http_port
5874 0 : ))
5875 0 : .await
5876 0 : .is_err()
5877 : {
5878 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
5879 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
5880 : // we return a soft 503 error, to encourage callers to retry past transient issues.
5881 0 : return Err(ApiError::ResourceUnavailable(
5882 0 : format!(
5883 0 : "Node {} tried to register with unknown DNS name '{}'",
5884 0 : register_req.node_id, register_req.listen_http_addr
5885 0 : )
5886 0 : .into(),
5887 0 : ));
5888 0 : }
5889 0 :
5890 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
5891 0 : // This ensures that before we use it for anything or expose it via any external
5892 0 : // API, it is guaranteed to be available after a restart.
5893 0 : let new_node = Node::new(
5894 0 : register_req.node_id,
5895 0 : register_req.listen_http_addr,
5896 0 : register_req.listen_http_port,
5897 0 : register_req.listen_pg_addr,
5898 0 : register_req.listen_pg_port,
5899 0 : register_req.availability_zone_id.clone(),
5900 0 : );
5901 0 :
5902 0 : // TODO: idempotency if the node already exists in the database
5903 0 : self.persistence.insert_node(&new_node).await?;
5904 :
5905 0 : let mut locked = self.inner.write().unwrap();
5906 0 : let mut new_nodes = (*locked.nodes).clone();
5907 0 :
5908 0 : locked.scheduler.node_upsert(&new_node);
5909 0 : new_nodes.insert(register_req.node_id, new_node);
5910 0 :
5911 0 : locked.nodes = Arc::new(new_nodes);
5912 0 :
5913 0 : metrics::METRICS_REGISTRY
5914 0 : .metrics_group
5915 0 : .storage_controller_pageserver_nodes
5916 0 : .set(locked.nodes.len() as i64);
5917 0 :
5918 0 : tracing::info!(
5919 0 : "Registered pageserver {} ({}), now have {} pageservers",
5920 0 : register_req.node_id,
5921 0 : register_req.availability_zone_id,
5922 0 : locked.nodes.len()
5923 : );
5924 0 : Ok(())
5925 0 : }
5926 :
5927 : /// Configure in-memory and persistent state of a node as requested
5928 : ///
5929 : /// Note that this function does not trigger any immediate side effects in response
5930 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
5931 0 : async fn node_state_configure(
5932 0 : &self,
5933 0 : node_id: NodeId,
5934 0 : availability: Option<NodeAvailability>,
5935 0 : scheduling: Option<NodeSchedulingPolicy>,
5936 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
5937 0 : ) -> Result<AvailabilityTransition, ApiError> {
5938 0 : if let Some(scheduling) = scheduling {
5939 : // Scheduling is a persistent part of Node: we must write updates to the database before
5940 : // applying them in memory
5941 0 : self.persistence.update_node(node_id, scheduling).await?;
5942 0 : }
5943 :
5944 : // If we're activating a node, then before setting it active we must reconcile any shard locations
5945 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
5946 : // by calling [`Self::node_activate_reconcile`]
5947 : //
5948 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
5949 : // nothing else can mutate its availability while we run.
5950 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
5951 0 : let (activate_node, availability_transition) = {
5952 0 : let locked = self.inner.read().unwrap();
5953 0 : let Some(node) = locked.nodes.get(&node_id) else {
5954 0 : return Err(ApiError::NotFound(
5955 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5956 0 : ));
5957 : };
5958 :
5959 0 : (
5960 0 : node.clone(),
5961 0 : node.get_availability_transition(input_availability),
5962 0 : )
5963 : };
5964 :
5965 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
5966 0 : self.node_activate_reconcile(activate_node, node_lock)
5967 0 : .await?;
5968 0 : }
5969 0 : availability_transition
5970 : } else {
5971 0 : AvailabilityTransition::Unchanged
5972 : };
5973 :
5974 : // Apply changes from the request to our in-memory state for the Node
5975 0 : let mut locked = self.inner.write().unwrap();
5976 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
5977 0 :
5978 0 : let mut new_nodes = (**nodes).clone();
5979 :
5980 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
5981 0 : return Err(ApiError::NotFound(
5982 0 : anyhow::anyhow!("Node not registered").into(),
5983 0 : ));
5984 : };
5985 :
5986 0 : if let Some(availability) = availability {
5987 0 : node.set_availability(availability);
5988 0 : }
5989 :
5990 0 : if let Some(scheduling) = scheduling {
5991 0 : node.set_scheduling(scheduling);
5992 0 : }
5993 :
5994 : // Update the scheduler, in case the elegibility of the node for new shards has changed
5995 0 : scheduler.node_upsert(node);
5996 0 :
5997 0 : let new_nodes = Arc::new(new_nodes);
5998 0 : locked.nodes = new_nodes;
5999 0 :
6000 0 : Ok(availability_transition)
6001 0 : }
6002 :
6003 : /// Handle availability transition of one node
6004 : ///
6005 : /// Note that you should first call [`Self::node_state_configure`] to update
6006 : /// the in-memory state referencing that node. If you need to handle more than one transition
6007 : /// consider using [`Self::handle_node_availability_transitions`].
6008 0 : async fn handle_node_availability_transition(
6009 0 : &self,
6010 0 : node_id: NodeId,
6011 0 : transition: AvailabilityTransition,
6012 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
6013 0 : ) -> Result<(), ApiError> {
6014 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
6015 0 : match transition {
6016 : AvailabilityTransition::ToOffline => {
6017 0 : tracing::info!("Node {} transition to offline", node_id);
6018 :
6019 0 : let mut locked = self.inner.write().unwrap();
6020 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6021 0 :
6022 0 : let mut tenants_affected: usize = 0;
6023 :
6024 0 : for (_tenant_id, mut schedule_context, shards) in
6025 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6026 : {
6027 0 : for tenant_shard in shards {
6028 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
6029 0 : if let Some(observed_loc) =
6030 0 : tenant_shard.observed.locations.get_mut(&node_id)
6031 0 : {
6032 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
6033 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
6034 0 : observed_loc.conf = None;
6035 0 : }
6036 :
6037 0 : if nodes.len() == 1 {
6038 : // Special case for single-node cluster: there is no point trying to reschedule
6039 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
6040 : // failures to schedule them.
6041 0 : continue;
6042 0 : }
6043 0 :
6044 0 : if !nodes
6045 0 : .values()
6046 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6047 : {
6048 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
6049 : // trying to reschedule since there's nowhere else to go. Without this
6050 : // branch we incorrectly detach tenants in response to node unavailability.
6051 0 : continue;
6052 0 : }
6053 0 :
6054 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
6055 0 : tenant_shard.sequence = tenant_shard.sequence.next();
6056 0 :
6057 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
6058 0 : Err(e) => {
6059 0 : // It is possible that some tenants will become unschedulable when too many pageservers
6060 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
6061 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
6062 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
6063 : }
6064 : Ok(()) => {
6065 0 : if self.maybe_reconcile_shard(tenant_shard, nodes).is_some() {
6066 0 : tenants_affected += 1;
6067 0 : };
6068 : }
6069 : }
6070 0 : }
6071 : }
6072 : }
6073 0 : tracing::info!(
6074 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
6075 : tenants_affected,
6076 : node_id
6077 : )
6078 : }
6079 : AvailabilityTransition::ToActive => {
6080 0 : tracing::info!("Node {} transition to active", node_id);
6081 :
6082 0 : let mut locked = self.inner.write().unwrap();
6083 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
6084 :
6085 : // When a node comes back online, we must reconcile any tenant that has a None observed
6086 : // location on the node.
6087 0 : for tenant_shard in tenants.values_mut() {
6088 : // If a reconciliation is already in progress, rely on the previous scheduling
6089 : // decision and skip triggering a new reconciliation.
6090 0 : if tenant_shard.reconciler.is_some() {
6091 0 : continue;
6092 0 : }
6093 :
6094 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
6095 0 : if observed_loc.conf.is_none() {
6096 0 : self.maybe_reconcile_shard(tenant_shard, nodes);
6097 0 : }
6098 0 : }
6099 : }
6100 :
6101 : // TODO: in the background, we should balance work back onto this pageserver
6102 : }
6103 : // No action required for the intermediate unavailable state.
6104 : // When we transition into active or offline from the unavailable state,
6105 : // the correct handling above will kick in.
6106 : AvailabilityTransition::ToWarmingUpFromActive => {
6107 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
6108 : }
6109 : AvailabilityTransition::ToWarmingUpFromOffline => {
6110 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
6111 : }
6112 : AvailabilityTransition::Unchanged => {
6113 0 : tracing::debug!("Node {} no availability change during config", node_id);
6114 : }
6115 : }
6116 :
6117 0 : Ok(())
6118 0 : }
6119 :
6120 : /// Handle availability transition for multiple nodes
6121 : ///
6122 : /// Note that you should first call [`Self::node_state_configure`] for
6123 : /// all nodes being handled here for the handling to use fresh in-memory state.
6124 0 : async fn handle_node_availability_transitions(
6125 0 : &self,
6126 0 : transitions: Vec<(
6127 0 : NodeId,
6128 0 : TracingExclusiveGuard<NodeOperations>,
6129 0 : AvailabilityTransition,
6130 0 : )>,
6131 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
6132 0 : let mut errors = Vec::default();
6133 0 : for (node_id, node_lock, transition) in transitions {
6134 0 : let res = self
6135 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
6136 0 : .await;
6137 0 : if let Err(err) = res {
6138 0 : errors.push((node_id, err));
6139 0 : }
6140 : }
6141 :
6142 0 : if errors.is_empty() {
6143 0 : Ok(())
6144 : } else {
6145 0 : Err(errors)
6146 : }
6147 0 : }
6148 :
6149 0 : pub(crate) async fn node_configure(
6150 0 : &self,
6151 0 : node_id: NodeId,
6152 0 : availability: Option<NodeAvailability>,
6153 0 : scheduling: Option<NodeSchedulingPolicy>,
6154 0 : ) -> Result<(), ApiError> {
6155 0 : let node_lock =
6156 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
6157 :
6158 0 : let transition = self
6159 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
6160 0 : .await?;
6161 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
6162 0 : .await
6163 0 : }
6164 :
6165 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
6166 : /// operation for HTTP api.
6167 0 : pub(crate) async fn external_node_configure(
6168 0 : &self,
6169 0 : node_id: NodeId,
6170 0 : availability: Option<NodeAvailability>,
6171 0 : scheduling: Option<NodeSchedulingPolicy>,
6172 0 : ) -> Result<(), ApiError> {
6173 0 : {
6174 0 : let locked = self.inner.read().unwrap();
6175 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
6176 0 : return Err(ApiError::PreconditionFailed(
6177 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
6178 0 : ));
6179 0 : }
6180 0 : }
6181 0 :
6182 0 : self.node_configure(node_id, availability, scheduling).await
6183 0 : }
6184 :
6185 0 : pub(crate) async fn start_node_drain(
6186 0 : self: &Arc<Self>,
6187 0 : node_id: NodeId,
6188 0 : ) -> Result<(), ApiError> {
6189 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
6190 0 : let locked = self.inner.read().unwrap();
6191 0 : let nodes = &locked.nodes;
6192 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6193 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6194 0 : ))?;
6195 0 : let schedulable_nodes_count = nodes
6196 0 : .iter()
6197 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6198 0 : .count();
6199 0 :
6200 0 : (
6201 0 : locked
6202 0 : .ongoing_operation
6203 0 : .as_ref()
6204 0 : .map(|ongoing| ongoing.operation),
6205 0 : node.is_available(),
6206 0 : node.get_scheduling(),
6207 0 : schedulable_nodes_count,
6208 0 : )
6209 0 : };
6210 :
6211 0 : if let Some(ongoing) = ongoing_op {
6212 0 : return Err(ApiError::PreconditionFailed(
6213 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6214 0 : ));
6215 0 : }
6216 0 :
6217 0 : if !node_available {
6218 0 : return Err(ApiError::ResourceUnavailable(
6219 0 : format!("Node {node_id} is currently unavailable").into(),
6220 0 : ));
6221 0 : }
6222 0 :
6223 0 : if schedulable_nodes_count == 0 {
6224 0 : return Err(ApiError::PreconditionFailed(
6225 0 : "No other schedulable nodes to drain to".into(),
6226 0 : ));
6227 0 : }
6228 0 :
6229 0 : match node_policy {
6230 : NodeSchedulingPolicy::Active => {
6231 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
6232 0 : .await?;
6233 :
6234 0 : let cancel = self.cancel.child_token();
6235 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6236 :
6237 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6238 0 : operation: Operation::Drain(Drain { node_id }),
6239 0 : cancel: cancel.clone(),
6240 0 : });
6241 :
6242 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
6243 :
6244 0 : tokio::task::spawn({
6245 0 : let service = self.clone();
6246 0 : let cancel = cancel.clone();
6247 0 : async move {
6248 0 : let _gate_guard = gate_guard;
6249 0 :
6250 0 : scopeguard::defer! {
6251 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6252 0 :
6253 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
6254 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
6255 0 : } else {
6256 0 : panic!("We always remove the same operation")
6257 0 : }
6258 0 : }
6259 0 :
6260 0 : tracing::info!("Drain background operation starting");
6261 0 : let res = service.drain_node(node_id, cancel).await;
6262 0 : match res {
6263 : Ok(()) => {
6264 0 : tracing::info!("Drain background operation completed successfully");
6265 : }
6266 : Err(OperationError::Cancelled) => {
6267 0 : tracing::info!("Drain background operation was cancelled");
6268 : }
6269 0 : Err(err) => {
6270 0 : tracing::error!("Drain background operation encountered: {err}")
6271 : }
6272 : }
6273 0 : }
6274 0 : }.instrument(span));
6275 0 : }
6276 : NodeSchedulingPolicy::Draining => {
6277 0 : return Err(ApiError::Conflict(format!(
6278 0 : "Node {node_id} has drain in progress"
6279 0 : )));
6280 : }
6281 0 : policy => {
6282 0 : return Err(ApiError::PreconditionFailed(
6283 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
6284 0 : ));
6285 : }
6286 : }
6287 :
6288 0 : Ok(())
6289 0 : }
6290 :
6291 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
6292 0 : let node_available = {
6293 0 : let locked = self.inner.read().unwrap();
6294 0 : let nodes = &locked.nodes;
6295 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6296 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6297 0 : ))?;
6298 :
6299 0 : node.is_available()
6300 0 : };
6301 0 :
6302 0 : if !node_available {
6303 0 : return Err(ApiError::ResourceUnavailable(
6304 0 : format!("Node {node_id} is currently unavailable").into(),
6305 0 : ));
6306 0 : }
6307 :
6308 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6309 0 : if let Operation::Drain(drain) = op_handler.operation {
6310 0 : if drain.node_id == node_id {
6311 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6312 0 : op_handler.cancel.cancel();
6313 0 : return Ok(());
6314 0 : }
6315 0 : }
6316 0 : }
6317 :
6318 0 : Err(ApiError::PreconditionFailed(
6319 0 : format!("Node {node_id} has no drain in progress").into(),
6320 0 : ))
6321 0 : }
6322 :
6323 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
6324 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
6325 0 : let locked = self.inner.read().unwrap();
6326 0 : let nodes = &locked.nodes;
6327 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6328 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6329 0 : ))?;
6330 :
6331 0 : (
6332 0 : locked
6333 0 : .ongoing_operation
6334 0 : .as_ref()
6335 0 : .map(|ongoing| ongoing.operation),
6336 0 : node.is_available(),
6337 0 : node.get_scheduling(),
6338 0 : nodes.len(),
6339 0 : )
6340 0 : };
6341 :
6342 0 : if let Some(ongoing) = ongoing_op {
6343 0 : return Err(ApiError::PreconditionFailed(
6344 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6345 0 : ));
6346 0 : }
6347 0 :
6348 0 : if !node_available {
6349 0 : return Err(ApiError::ResourceUnavailable(
6350 0 : format!("Node {node_id} is currently unavailable").into(),
6351 0 : ));
6352 0 : }
6353 0 :
6354 0 : if total_nodes_count <= 1 {
6355 0 : return Err(ApiError::PreconditionFailed(
6356 0 : "No other nodes to fill from".into(),
6357 0 : ));
6358 0 : }
6359 0 :
6360 0 : match node_policy {
6361 : NodeSchedulingPolicy::Active => {
6362 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
6363 0 : .await?;
6364 :
6365 0 : let cancel = self.cancel.child_token();
6366 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6367 :
6368 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6369 0 : operation: Operation::Fill(Fill { node_id }),
6370 0 : cancel: cancel.clone(),
6371 0 : });
6372 :
6373 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
6374 :
6375 0 : tokio::task::spawn({
6376 0 : let service = self.clone();
6377 0 : let cancel = cancel.clone();
6378 0 : async move {
6379 0 : let _gate_guard = gate_guard;
6380 0 :
6381 0 : scopeguard::defer! {
6382 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6383 0 :
6384 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
6385 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
6386 0 : } else {
6387 0 : panic!("We always remove the same operation")
6388 0 : }
6389 0 : }
6390 0 :
6391 0 : tracing::info!("Fill background operation starting");
6392 0 : let res = service.fill_node(node_id, cancel).await;
6393 0 : match res {
6394 : Ok(()) => {
6395 0 : tracing::info!("Fill background operation completed successfully");
6396 : }
6397 : Err(OperationError::Cancelled) => {
6398 0 : tracing::info!("Fill background operation was cancelled");
6399 : }
6400 0 : Err(err) => {
6401 0 : tracing::error!("Fill background operation encountered: {err}")
6402 : }
6403 : }
6404 0 : }
6405 0 : }.instrument(span));
6406 0 : }
6407 : NodeSchedulingPolicy::Filling => {
6408 0 : return Err(ApiError::Conflict(format!(
6409 0 : "Node {node_id} has fill in progress"
6410 0 : )));
6411 : }
6412 0 : policy => {
6413 0 : return Err(ApiError::PreconditionFailed(
6414 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
6415 0 : ));
6416 : }
6417 : }
6418 :
6419 0 : Ok(())
6420 0 : }
6421 :
6422 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
6423 0 : let node_available = {
6424 0 : let locked = self.inner.read().unwrap();
6425 0 : let nodes = &locked.nodes;
6426 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6427 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6428 0 : ))?;
6429 :
6430 0 : node.is_available()
6431 0 : };
6432 0 :
6433 0 : if !node_available {
6434 0 : return Err(ApiError::ResourceUnavailable(
6435 0 : format!("Node {node_id} is currently unavailable").into(),
6436 0 : ));
6437 0 : }
6438 :
6439 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6440 0 : if let Operation::Fill(fill) = op_handler.operation {
6441 0 : if fill.node_id == node_id {
6442 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6443 0 : op_handler.cancel.cancel();
6444 0 : return Ok(());
6445 0 : }
6446 0 : }
6447 0 : }
6448 :
6449 0 : Err(ApiError::PreconditionFailed(
6450 0 : format!("Node {node_id} has no fill in progress").into(),
6451 0 : ))
6452 0 : }
6453 :
6454 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
6455 : /// configuration
6456 0 : fn maybe_reconcile_shard(
6457 0 : &self,
6458 0 : shard: &mut TenantShard,
6459 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6460 0 : ) -> Option<ReconcilerWaiter> {
6461 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::default())
6462 0 : }
6463 :
6464 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
6465 0 : fn maybe_configured_reconcile_shard(
6466 0 : &self,
6467 0 : shard: &mut TenantShard,
6468 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6469 0 : reconciler_config: ReconcilerConfig,
6470 0 : ) -> Option<ReconcilerWaiter> {
6471 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
6472 0 :
6473 0 : match reconcile_needed {
6474 0 : ReconcileNeeded::No => return None,
6475 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
6476 0 : ReconcileNeeded::Yes => {
6477 0 : // Fall through to try and acquire units for spawning reconciler
6478 0 : }
6479 : };
6480 :
6481 0 : let units = match self.reconciler_concurrency.clone().try_acquire_owned() {
6482 0 : Ok(u) => ReconcileUnits::new(u),
6483 : Err(_) => {
6484 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
6485 0 : "Concurrency limited: enqueued for reconcile later");
6486 0 : if !shard.delayed_reconcile {
6487 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
6488 0 : Err(TrySendError::Closed(_)) => {
6489 0 : // Weird mid-shutdown case?
6490 0 : }
6491 : Err(TrySendError::Full(_)) => {
6492 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
6493 0 : tracing::warn!(
6494 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
6495 : );
6496 : }
6497 0 : Ok(()) => {
6498 0 : shard.delayed_reconcile = true;
6499 0 : }
6500 : }
6501 0 : }
6502 :
6503 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
6504 : // number to advance. When this function is eventually called again and succeeds in getting units,
6505 : // it will spawn a reconciler that makes this waiter complete.
6506 0 : return Some(shard.future_reconcile_waiter());
6507 : }
6508 : };
6509 :
6510 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
6511 : // Gate closed: we're shutting down, drop out.
6512 0 : return None;
6513 : };
6514 :
6515 0 : shard.spawn_reconciler(
6516 0 : &self.result_tx,
6517 0 : nodes,
6518 0 : &self.compute_hook,
6519 0 : reconciler_config,
6520 0 : &self.config,
6521 0 : &self.persistence,
6522 0 : units,
6523 0 : gate_guard,
6524 0 : &self.reconcilers_cancel,
6525 0 : )
6526 0 : }
6527 :
6528 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
6529 : /// Additionally, reschedule tenants that require it.
6530 : ///
6531 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
6532 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
6533 : /// available. A return value of 0 indicates that everything is fully reconciled already.
6534 0 : fn reconcile_all(&self) -> usize {
6535 0 : let mut locked = self.inner.write().unwrap();
6536 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6537 0 : let pageservers = nodes.clone();
6538 0 :
6539 0 : // This function is an efficient place to update lazy statistics, since we are walking
6540 0 : // all tenants.
6541 0 : let mut pending_reconciles = 0;
6542 0 : let mut az_violations = 0;
6543 0 :
6544 0 : // If we find any tenants to drop from memory, stash them to offload after
6545 0 : // we're done traversing the map of tenants.
6546 0 : let mut drop_detached_tenants = Vec::new();
6547 0 :
6548 0 : let mut reconciles_spawned = 0;
6549 0 : for shard in tenants.values_mut() {
6550 : // Accumulate scheduling statistics
6551 0 : if let (Some(attached), Some(preferred)) =
6552 0 : (shard.intent.get_attached(), shard.preferred_az())
6553 : {
6554 0 : let node_az = nodes
6555 0 : .get(attached)
6556 0 : .expect("Nodes exist if referenced")
6557 0 : .get_availability_zone_id();
6558 0 : if node_az != preferred {
6559 0 : az_violations += 1;
6560 0 : }
6561 0 : }
6562 :
6563 : // Skip checking if this shard is already enqueued for reconciliation
6564 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
6565 : // If there is something delayed, then return a nonzero count so that
6566 : // callers like reconcile_all_now do not incorrectly get the impression
6567 : // that the system is in a quiescent state.
6568 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
6569 0 : pending_reconciles += 1;
6570 0 : continue;
6571 0 : }
6572 0 :
6573 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
6574 0 : // dirty, spawn another rone
6575 0 : if self.maybe_reconcile_shard(shard, &pageservers).is_some() {
6576 0 : reconciles_spawned += 1;
6577 0 : } else if shard.delayed_reconcile {
6578 0 : // Shard wanted to reconcile but for some reason couldn't.
6579 0 : pending_reconciles += 1;
6580 0 : }
6581 :
6582 : // If this tenant is detached, try dropping it from memory. This is usually done
6583 : // proactively in [`Self::process_results`], but we do it here to handle the edge
6584 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
6585 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
6586 0 : && shard.policy == PlacementPolicy::Detached
6587 : {
6588 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
6589 0 : shard.tenant_shard_id.tenant_id,
6590 0 : TenantOperations::DropDetached,
6591 0 : ) {
6592 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
6593 0 : }
6594 0 : }
6595 : }
6596 :
6597 : // Some metrics are calculated from SchedulerNode state, update these periodically
6598 0 : scheduler.update_metrics();
6599 :
6600 : // Process any deferred tenant drops
6601 0 : for (tenant_id, guard) in drop_detached_tenants {
6602 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
6603 0 : }
6604 :
6605 0 : metrics::METRICS_REGISTRY
6606 0 : .metrics_group
6607 0 : .storage_controller_schedule_az_violation
6608 0 : .set(az_violations as i64);
6609 0 :
6610 0 : metrics::METRICS_REGISTRY
6611 0 : .metrics_group
6612 0 : .storage_controller_pending_reconciles
6613 0 : .set(pending_reconciles as i64);
6614 0 :
6615 0 : reconciles_spawned
6616 0 : }
6617 :
6618 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
6619 : /// could be scheduled somewhere better:
6620 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
6621 : /// * e.g. after a node fails then recovers, to move some work back to it
6622 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
6623 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
6624 : /// we did the split, but are probably better placed elsewhere.
6625 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
6626 : /// * e.g. after a shard split, some locations will be on the same node (where the split
6627 : /// happened), and will probably be better placed elsewhere.
6628 : ///
6629 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
6630 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
6631 : /// according to those same soft constraints.
6632 0 : async fn optimize_all(&self) -> usize {
6633 : // Limit on how many shards' optmizations each call to this function will execute. Combined
6634 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
6635 : // trickle of optimizations in the background, rather than executing a large number in parallel
6636 : // when a change occurs.
6637 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
6638 :
6639 : // Synchronous prepare: scan shards for possible scheduling optimizations
6640 0 : let candidate_work = self.optimize_all_plan();
6641 0 : let candidate_work_len = candidate_work.len();
6642 :
6643 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
6644 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
6645 :
6646 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
6647 0 :
6648 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
6649 0 : let mut reconciles_spawned = 0;
6650 0 : let mut optimizations_applied = 0;
6651 0 : let mut locked = self.inner.write().unwrap();
6652 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6653 0 : for (tenant_shard_id, optimization) in validated_work {
6654 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6655 : // Shard was dropped between planning and execution;
6656 0 : continue;
6657 : };
6658 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
6659 0 : if shard.apply_optimization(scheduler, optimization) {
6660 0 : optimizations_applied += 1;
6661 0 : if self.maybe_reconcile_shard(shard, nodes).is_some() {
6662 0 : reconciles_spawned += 1;
6663 0 : }
6664 0 : }
6665 :
6666 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
6667 0 : break;
6668 0 : }
6669 : }
6670 :
6671 0 : if was_work_filtered {
6672 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
6673 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
6674 0 : // as these validations start passing.
6675 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
6676 0 : }
6677 :
6678 0 : reconciles_spawned
6679 0 : }
6680 :
6681 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
6682 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
6683 : // this higher than the execution limit gives us a chance to execute some work even if the first
6684 : // few optimizations we find are not ready.
6685 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
6686 :
6687 0 : let mut work = Vec::new();
6688 0 : let mut locked = self.inner.write().unwrap();
6689 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6690 :
6691 : // We are going to plan a bunch of optimisations before applying any of them, so the
6692 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
6693 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
6694 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
6695 : // to ignore the utilisation component of the score.
6696 :
6697 0 : for (_tenant_id, schedule_context, shards) in
6698 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
6699 : {
6700 0 : for shard in shards {
6701 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
6702 0 : break;
6703 0 : }
6704 0 : match shard.get_scheduling_policy() {
6705 0 : ShardSchedulingPolicy::Active => {
6706 0 : // Ok to do optimization
6707 0 : }
6708 : ShardSchedulingPolicy::Essential
6709 : | ShardSchedulingPolicy::Pause
6710 : | ShardSchedulingPolicy::Stop => {
6711 : // Policy prevents optimizing this shard.
6712 0 : continue;
6713 : }
6714 : }
6715 :
6716 0 : if !matches!(shard.splitting, SplitState::Idle)
6717 0 : || matches!(shard.policy, PlacementPolicy::Detached)
6718 0 : || shard.reconciler.is_some()
6719 : {
6720 : // Do not start any optimizations while another change to the tenant is ongoing: this
6721 : // is not necessary for correctness, but simplifies operations and implicitly throttles
6722 : // optimization changes to happen in a "trickle" over time.
6723 0 : continue;
6724 0 : }
6725 0 :
6726 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
6727 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
6728 0 : if cfg!(feature = "testing") {
6729 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
6730 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
6731 : // panic in prod if we hit this, or spend cycles on it in prod.
6732 0 : assert!(shard
6733 0 : .optimize_attachment(scheduler, &schedule_context)
6734 0 : .is_none());
6735 0 : assert!(shard
6736 0 : .optimize_secondary(scheduler, &schedule_context)
6737 0 : .is_none());
6738 0 : }
6739 0 : continue;
6740 0 : }
6741 :
6742 0 : if let Some(optimization) =
6743 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
6744 : // its primary location based on soft constraints, cut it over.
6745 0 : shard.optimize_attachment(scheduler, &schedule_context)
6746 : {
6747 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
6748 0 : work.push((shard.tenant_shard_id, optimization));
6749 0 : break;
6750 0 : } else if let Some(optimization) =
6751 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
6752 : // better placed on another node, based on ScheduleContext, then adjust it. This
6753 : // covers cases like after a shard split, where we might have too many shards
6754 : // in the same tenant with secondary locations on the node where they originally split.
6755 0 : shard.optimize_secondary(scheduler, &schedule_context)
6756 : {
6757 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
6758 0 : work.push((shard.tenant_shard_id, optimization));
6759 0 : break;
6760 0 : }
6761 : }
6762 : }
6763 :
6764 0 : work
6765 0 : }
6766 :
6767 0 : async fn optimize_all_validate(
6768 0 : &self,
6769 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
6770 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
6771 0 : // Take a clone of the node map to use outside the lock in async validation phase
6772 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
6773 0 :
6774 0 : let mut want_secondary_status = Vec::new();
6775 0 :
6776 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
6777 0 : // check that the state of locations is acceptable to run the optimization, such as
6778 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
6779 0 : // in a live migration.
6780 0 : let mut validated_work = Vec::new();
6781 0 : for (tenant_shard_id, optimization) in candidate_work {
6782 0 : match optimization.action {
6783 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
6784 : old_attached_node_id: _,
6785 0 : new_attached_node_id,
6786 0 : }) => {
6787 0 : match validation_nodes.get(&new_attached_node_id) {
6788 0 : None => {
6789 0 : // Node was dropped between planning and validation
6790 0 : }
6791 0 : Some(node) => {
6792 0 : if !node.is_available() {
6793 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
6794 0 : } else {
6795 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
6796 0 : // remote API requests concurrently.
6797 0 : want_secondary_status.push((
6798 0 : tenant_shard_id,
6799 0 : node.clone(),
6800 0 : optimization,
6801 0 : ));
6802 0 : }
6803 : }
6804 : }
6805 : }
6806 : ScheduleOptimizationAction::ReplaceSecondary(_)
6807 : | ScheduleOptimizationAction::CreateSecondary(_)
6808 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
6809 : // No extra checks needed to manage secondaries: this does not interrupt client access
6810 0 : validated_work.push((tenant_shard_id, optimization))
6811 : }
6812 : };
6813 : }
6814 :
6815 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
6816 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
6817 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
6818 0 : let results = self
6819 0 : .tenant_for_shards_api(
6820 0 : want_secondary_status
6821 0 : .iter()
6822 0 : .map(|i| (i.0, i.1.clone()))
6823 0 : .collect(),
6824 0 : |tenant_shard_id, client| async move {
6825 0 : client.tenant_secondary_status(tenant_shard_id).await
6826 0 : },
6827 0 : 1,
6828 0 : 1,
6829 0 : SHORT_RECONCILE_TIMEOUT,
6830 0 : &self.cancel,
6831 0 : )
6832 0 : .await;
6833 :
6834 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
6835 0 : want_secondary_status.into_iter().zip(results.into_iter())
6836 : {
6837 0 : match secondary_status {
6838 0 : Err(e) => {
6839 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
6840 : }
6841 0 : Ok(progress) => {
6842 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
6843 : // them in an optimization
6844 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
6845 :
6846 0 : if progress.heatmap_mtime.is_none()
6847 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
6848 0 : && progress.bytes_downloaded != progress.bytes_total
6849 0 : || progress.bytes_total - progress.bytes_downloaded
6850 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
6851 : {
6852 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
6853 :
6854 : #[cfg(feature = "testing")]
6855 0 : if progress.heatmap_mtime.is_none() {
6856 : // No heatmap might mean the attached location has never uploaded one, or that
6857 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
6858 : // but fairly common in tests.
6859 0 : self.kick_secondary_download(tenant_shard_id).await;
6860 0 : }
6861 : } else {
6862 : // Location looks ready: proceed
6863 0 : tracing::info!(
6864 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
6865 : );
6866 0 : validated_work.push((tenant_shard_id, optimization))
6867 : }
6868 : }
6869 : }
6870 : }
6871 :
6872 0 : validated_work
6873 0 : }
6874 :
6875 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
6876 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
6877 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
6878 : /// we have this helper to move things along faster.
6879 : #[cfg(feature = "testing")]
6880 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
6881 0 : let (attached_node, secondaries) = {
6882 0 : let locked = self.inner.read().unwrap();
6883 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6884 0 : tracing::warn!(
6885 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
6886 : );
6887 0 : return;
6888 : };
6889 :
6890 0 : let Some(attached) = shard.intent.get_attached() else {
6891 0 : tracing::warn!(
6892 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
6893 : );
6894 0 : return;
6895 : };
6896 :
6897 0 : let secondaries = shard
6898 0 : .intent
6899 0 : .get_secondary()
6900 0 : .iter()
6901 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
6902 0 : .collect::<Vec<_>>();
6903 0 :
6904 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
6905 0 : };
6906 0 :
6907 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
6908 0 : // a 'kick' to let scheduling optimisation run more promptly.
6909 0 : match attached_node
6910 0 : .with_client_retries(
6911 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
6912 0 : &self.config.jwt_token,
6913 0 : 3,
6914 0 : 10,
6915 0 : SHORT_RECONCILE_TIMEOUT,
6916 0 : &self.cancel,
6917 0 : )
6918 0 : .await
6919 : {
6920 0 : Some(Err(e)) => {
6921 0 : tracing::info!(
6922 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
6923 : );
6924 : }
6925 : None => {
6926 0 : tracing::info!(
6927 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
6928 : );
6929 : }
6930 : Some(Ok(_)) => {
6931 0 : tracing::info!(
6932 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
6933 : );
6934 : }
6935 : }
6936 :
6937 0 : for secondary_node in secondaries {
6938 0 : match secondary_node
6939 0 : .with_client_retries(
6940 0 : |client| async move {
6941 0 : client
6942 0 : .tenant_secondary_download(
6943 0 : tenant_shard_id,
6944 0 : Some(Duration::from_secs(1)),
6945 0 : )
6946 0 : .await
6947 0 : },
6948 0 : &self.config.jwt_token,
6949 0 : 3,
6950 0 : 10,
6951 0 : SHORT_RECONCILE_TIMEOUT,
6952 0 : &self.cancel,
6953 0 : )
6954 0 : .await
6955 : {
6956 0 : Some(Err(e)) => {
6957 0 : tracing::info!(
6958 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
6959 : );
6960 : }
6961 : None => {
6962 0 : tracing::info!("Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}");
6963 : }
6964 0 : Some(Ok(progress)) => {
6965 0 : tracing::info!("Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}");
6966 : }
6967 : }
6968 : }
6969 0 : }
6970 :
6971 : /// Look for shards which are oversized and in need of splitting
6972 0 : async fn autosplit_tenants(self: &Arc<Self>) {
6973 0 : let Some(split_threshold) = self.config.split_threshold else {
6974 : // Auto-splitting is disabled
6975 0 : return;
6976 : };
6977 :
6978 0 : let nodes = self.inner.read().unwrap().nodes.clone();
6979 :
6980 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
6981 :
6982 0 : let mut top_n = Vec::new();
6983 0 :
6984 0 : // Call into each node to look for big tenants
6985 0 : let top_n_request = TopTenantShardsRequest {
6986 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
6987 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
6988 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
6989 0 : // disk space impact of one shard.
6990 0 : order_by: models::TenantSorting::MaxLogicalSize,
6991 0 : limit: 10,
6992 0 : where_shards_lt: Some(SPLIT_TO_MAX),
6993 0 : where_gt: Some(split_threshold),
6994 0 : };
6995 0 : for node in nodes.values() {
6996 0 : let request_ref = &top_n_request;
6997 0 : match node
6998 0 : .with_client_retries(
6999 0 : |client| async move {
7000 0 : let request = request_ref.clone();
7001 0 : client.top_tenant_shards(request.clone()).await
7002 0 : },
7003 0 : &self.config.jwt_token,
7004 0 : 3,
7005 0 : 3,
7006 0 : Duration::from_secs(5),
7007 0 : &self.cancel,
7008 0 : )
7009 0 : .await
7010 : {
7011 0 : Some(Ok(node_top_n)) => {
7012 0 : top_n.extend(node_top_n.shards.into_iter());
7013 0 : }
7014 : Some(Err(mgmt_api::Error::Cancelled)) => {
7015 0 : continue;
7016 : }
7017 0 : Some(Err(e)) => {
7018 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
7019 0 : continue;
7020 : }
7021 : None => {
7022 : // Node is shutting down
7023 0 : continue;
7024 : }
7025 : };
7026 : }
7027 :
7028 : // Pick the biggest tenant to split first
7029 0 : top_n.sort_by_key(|i| i.resident_size);
7030 0 :
7031 0 : // Filter out tenants in a prohibiting scheduling mode
7032 0 : {
7033 0 : let locked = self.inner.read().unwrap();
7034 0 : top_n.retain(|i| {
7035 0 : if let Some(shard) = locked.tenants.get(&i.id) {
7036 0 : matches!(shard.get_scheduling_policy(), ShardSchedulingPolicy::Active)
7037 : } else {
7038 0 : false
7039 : }
7040 0 : });
7041 0 : }
7042 :
7043 0 : let Some(split_candidate) = top_n.into_iter().next() else {
7044 0 : tracing::debug!("No split-elegible shards found");
7045 0 : return;
7046 : };
7047 :
7048 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
7049 : // want to block the background reconcile loop on this.
7050 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
7051 :
7052 0 : let this = self.clone();
7053 0 : tokio::spawn(
7054 0 : async move {
7055 0 : match this
7056 0 : .tenant_shard_split(
7057 0 : split_candidate.id.tenant_id,
7058 0 : TenantShardSplitRequest {
7059 0 : // Always split to the max number of shards: this avoids stepping through
7060 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
7061 0 : // each time as a tenant grows, and is not too expensive because our max shard
7062 0 : // count is relatively low anyway.
7063 0 : // This policy will be adjusted in future once we support higher shard count.
7064 0 : new_shard_count: SPLIT_TO_MAX.literal(),
7065 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
7066 0 : },
7067 0 : )
7068 0 : .await
7069 : {
7070 : Ok(_) => {
7071 0 : tracing::info!("Successful auto-split");
7072 : }
7073 0 : Err(e) => {
7074 0 : tracing::error!("Auto-split failed: {e}");
7075 : }
7076 : }
7077 0 : }
7078 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
7079 : );
7080 0 : }
7081 :
7082 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
7083 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
7084 : /// put the system into a quiescent state where future background reconciliations won't do anything.
7085 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
7086 0 : let reconciles_spawned = self.reconcile_all();
7087 0 : let reconciles_spawned = if reconciles_spawned == 0 {
7088 : // Only optimize when we are otherwise idle
7089 0 : self.optimize_all().await
7090 : } else {
7091 0 : reconciles_spawned
7092 : };
7093 :
7094 0 : let waiters = {
7095 0 : let mut waiters = Vec::new();
7096 0 : let locked = self.inner.read().unwrap();
7097 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
7098 0 : if let Some(waiter) = shard.get_waiter() {
7099 0 : waiters.push(waiter);
7100 0 : }
7101 : }
7102 0 : waiters
7103 0 : };
7104 0 :
7105 0 : let waiter_count = waiters.len();
7106 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
7107 0 : Ok(()) => {}
7108 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
7109 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
7110 0 : {
7111 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
7112 0 : // because some other change superceded it. We will return a nonzero number,
7113 0 : // so the caller knows they might have to call again to quiesce the system.
7114 0 : }
7115 0 : Err(e) => {
7116 0 : return Err(e);
7117 : }
7118 : };
7119 :
7120 0 : tracing::info!(
7121 0 : "{} reconciles in reconcile_all, {} waiters",
7122 : reconciles_spawned,
7123 : waiter_count
7124 : );
7125 :
7126 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
7127 0 : }
7128 :
7129 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
7130 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
7131 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
7132 0 : self.reconcilers_cancel.cancel();
7133 0 : self.reconcilers_gate.close().await;
7134 :
7135 : // Signal the background loop in [`Service::process_results`] to exit once
7136 : // it has proccessed the results from all the reconciles we cancelled earlier.
7137 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
7138 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
7139 0 : self.result_tx.closed().await;
7140 0 : }
7141 :
7142 0 : pub async fn shutdown(&self) {
7143 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
7144 0 : .await;
7145 :
7146 : // Background tasks hold gate guards: this notifies them of the cancellation and
7147 : // waits for them all to complete.
7148 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
7149 0 : self.cancel.cancel();
7150 0 : self.gate.close().await;
7151 0 : }
7152 :
7153 : /// Spot check the download lag for a secondary location of a shard.
7154 : /// Should be used as a heuristic, since it's not always precise: the
7155 : /// secondary might have not downloaded the new heat map yet and, hence,
7156 : /// is not aware of the lag.
7157 : ///
7158 : /// Returns:
7159 : /// * Ok(None) if the lag could not be determined from the status,
7160 : /// * Ok(Some(_)) if the lag could be determind
7161 : /// * Err on failures to query the pageserver.
7162 0 : async fn secondary_lag(
7163 0 : &self,
7164 0 : secondary: &NodeId,
7165 0 : tenant_shard_id: TenantShardId,
7166 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
7167 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7168 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
7169 0 : StatusCode::NOT_FOUND,
7170 0 : format!("Node with id {} not found", secondary),
7171 0 : ))?;
7172 :
7173 0 : match node
7174 0 : .with_client_retries(
7175 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
7176 0 : &self.config.jwt_token,
7177 0 : 1,
7178 0 : 3,
7179 0 : Duration::from_millis(250),
7180 0 : &self.cancel,
7181 0 : )
7182 0 : .await
7183 : {
7184 0 : Some(Ok(status)) => match status.heatmap_mtime {
7185 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
7186 0 : None => Ok(None),
7187 : },
7188 0 : Some(Err(e)) => Err(e),
7189 0 : None => Err(mgmt_api::Error::Cancelled),
7190 : }
7191 0 : }
7192 :
7193 : /// Drain a node by moving the shards attached to it as primaries.
7194 : /// This is a long running operation and it should run as a separate Tokio task.
7195 0 : pub(crate) async fn drain_node(
7196 0 : self: &Arc<Self>,
7197 0 : node_id: NodeId,
7198 0 : cancel: CancellationToken,
7199 0 : ) -> Result<(), OperationError> {
7200 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
7201 0 : let max_secondary_lag_bytes = self
7202 0 : .config
7203 0 : .max_secondary_lag_bytes
7204 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
7205 :
7206 : // By default, live migrations are generous about the wait time for getting
7207 : // the secondary location up to speed. When draining, give up earlier in order
7208 : // to not stall the operation when a cold secondary is encountered.
7209 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7210 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7211 0 : let reconciler_config = ReconcilerConfigBuilder::new()
7212 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7213 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7214 0 : .build();
7215 0 :
7216 0 : let mut waiters = Vec::new();
7217 0 :
7218 0 : let mut tid_iter = TenantShardIterator::new({
7219 0 : let service = self.clone();
7220 0 : move |last_inspected_shard: Option<TenantShardId>| {
7221 0 : let locked = &service.inner.read().unwrap();
7222 0 : let tenants = &locked.tenants;
7223 0 : let entry = match last_inspected_shard {
7224 0 : Some(skip_past) => {
7225 0 : // Skip to the last seen tenant shard id
7226 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
7227 0 :
7228 0 : // Skip past the last seen
7229 0 : cursor.nth(1)
7230 : }
7231 0 : None => tenants.first_key_value(),
7232 : };
7233 :
7234 0 : entry.map(|(tid, _)| tid).copied()
7235 0 : }
7236 0 : });
7237 :
7238 0 : while !tid_iter.finished() {
7239 0 : if cancel.is_cancelled() {
7240 0 : match self
7241 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7242 0 : .await
7243 : {
7244 0 : Ok(()) => return Err(OperationError::Cancelled),
7245 0 : Err(err) => {
7246 0 : return Err(OperationError::FinalizeError(
7247 0 : format!(
7248 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7249 0 : node_id, err
7250 0 : )
7251 0 : .into(),
7252 0 : ));
7253 : }
7254 : }
7255 0 : }
7256 0 :
7257 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
7258 :
7259 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7260 0 : let tid = match tid_iter.next() {
7261 0 : Some(tid) => tid,
7262 : None => {
7263 0 : break;
7264 : }
7265 : };
7266 :
7267 0 : let tid_drain = TenantShardDrain {
7268 0 : drained_node: node_id,
7269 0 : tenant_shard_id: tid,
7270 0 : };
7271 :
7272 0 : let dest_node_id = {
7273 0 : let locked = self.inner.read().unwrap();
7274 0 :
7275 0 : match tid_drain
7276 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
7277 : {
7278 0 : Some(node_id) => node_id,
7279 : None => {
7280 0 : continue;
7281 : }
7282 : }
7283 : };
7284 :
7285 0 : match self.secondary_lag(&dest_node_id, tid).await {
7286 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
7287 0 : // The secondary is reasonably up to date.
7288 0 : // Migrate to it
7289 0 : }
7290 0 : Ok(Some(lag)) => {
7291 0 : tracing::info!(
7292 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7293 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
7294 : );
7295 0 : continue;
7296 : }
7297 : Ok(None) => {
7298 0 : tracing::info!(
7299 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7300 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
7301 : );
7302 0 : continue;
7303 : }
7304 0 : Err(err) => {
7305 0 : tracing::warn!(
7306 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7307 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
7308 : );
7309 0 : continue;
7310 : }
7311 : }
7312 :
7313 : {
7314 0 : let mut locked = self.inner.write().unwrap();
7315 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7316 0 : let rescheduled = tid_drain.reschedule_to_secondary(
7317 0 : dest_node_id,
7318 0 : tenants,
7319 0 : scheduler,
7320 0 : nodes,
7321 0 : )?;
7322 :
7323 0 : if let Some(tenant_shard) = rescheduled {
7324 0 : let waiter = self.maybe_configured_reconcile_shard(
7325 0 : tenant_shard,
7326 0 : nodes,
7327 0 : reconciler_config,
7328 0 : );
7329 0 : if let Some(some) = waiter {
7330 0 : waiters.push(some);
7331 0 : }
7332 0 : }
7333 : }
7334 : }
7335 :
7336 0 : waiters = self
7337 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7338 0 : .await;
7339 :
7340 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
7341 : }
7342 :
7343 0 : while !waiters.is_empty() {
7344 0 : if cancel.is_cancelled() {
7345 0 : match self
7346 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7347 0 : .await
7348 : {
7349 0 : Ok(()) => return Err(OperationError::Cancelled),
7350 0 : Err(err) => {
7351 0 : return Err(OperationError::FinalizeError(
7352 0 : format!(
7353 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7354 0 : node_id, err
7355 0 : )
7356 0 : .into(),
7357 0 : ));
7358 : }
7359 : }
7360 0 : }
7361 0 :
7362 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
7363 :
7364 0 : waiters = self
7365 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7366 0 : .await;
7367 : }
7368 :
7369 : // At this point we have done the best we could to drain shards from this node.
7370 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
7371 : // to complete the drain.
7372 0 : if let Err(err) = self
7373 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
7374 0 : .await
7375 : {
7376 : // This is not fatal. Anything that is polling the node scheduling policy to detect
7377 : // the end of the drain operations will hang, but all such places should enforce an
7378 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
7379 : // by the counterpart fill operation.
7380 0 : return Err(OperationError::FinalizeError(
7381 0 : format!(
7382 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
7383 0 : )
7384 0 : .into(),
7385 0 : ));
7386 0 : }
7387 0 :
7388 0 : Ok(())
7389 0 : }
7390 :
7391 : /// Create a node fill plan (pick secondaries to promote), based on:
7392 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
7393 : /// outside their home AZ, should be migrated back here.
7394 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
7395 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
7396 : /// those shards have a home AZ that doesn't match the node we're filling.
7397 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
7398 0 : let mut locked = self.inner.write().unwrap();
7399 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7400 0 :
7401 0 : let node_az = nodes
7402 0 : .get(&node_id)
7403 0 : .expect("Node must exist")
7404 0 : .get_availability_zone_id()
7405 0 : .clone();
7406 0 :
7407 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
7408 0 : let mut plan = Vec::new();
7409 0 :
7410 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
7411 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
7412 0 :
7413 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
7414 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
7415 0 : let respect_azs = nodes
7416 0 : .values()
7417 0 : .map(|n| n.get_availability_zone_id())
7418 0 : .unique()
7419 0 : .count()
7420 0 : > 1;
7421 :
7422 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
7423 : // requires it.
7424 0 : for (tsid, tenant_shard) in tenants {
7425 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7426 : // Shard doesn't have a secondary on this node, ignore it.
7427 0 : continue;
7428 0 : }
7429 0 :
7430 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
7431 0 : // shards which belong on this node, not to promote shards whose scheduling preference
7432 0 : // would be on their currently attached node. So will avoid promoting shards whose
7433 0 : // home AZ doesn't match the AZ of the node we're filling.
7434 0 : match tenant_shard.preferred_az() {
7435 0 : _ if !respect_azs => {
7436 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7437 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7438 0 : }
7439 : }
7440 : None => {
7441 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
7442 : // will only do so if our target shard count requires it.
7443 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7444 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7445 0 : }
7446 : }
7447 0 : Some(az) if az == &node_az => {
7448 : // This shard's home AZ is equal to the node we're filling: it should
7449 : // be moved back to this node as part of filling, unless its currently
7450 : // attached location is also in its home AZ.
7451 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7452 0 : if nodes
7453 0 : .get(primary)
7454 0 : .expect("referenced node must exist")
7455 0 : .get_availability_zone_id()
7456 0 : != tenant_shard
7457 0 : .preferred_az()
7458 0 : .expect("tenant must have an AZ preference")
7459 : {
7460 0 : plan.push(*tsid)
7461 0 : }
7462 : } else {
7463 0 : plan.push(*tsid)
7464 : }
7465 : }
7466 0 : Some(_) => {
7467 0 : // This shard's home AZ is somewhere other than the node we're filling,
7468 0 : // it may not be moved back to this node as part of filling. Ignore it
7469 0 : }
7470 : }
7471 : }
7472 :
7473 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
7474 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
7475 0 :
7476 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
7477 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
7478 0 :
7479 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
7480 :
7481 0 : for (node_id, attached) in nodes_by_load {
7482 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
7483 0 : if !available {
7484 0 : continue;
7485 0 : }
7486 0 :
7487 0 : if plan.len() >= fill_requirement
7488 0 : || free_tids_by_node.is_empty()
7489 0 : || attached <= expected_attached
7490 : {
7491 0 : break;
7492 0 : }
7493 0 :
7494 0 : let can_take = attached - expected_attached;
7495 0 : let needed = fill_requirement - plan.len();
7496 0 : let mut take = std::cmp::min(can_take, needed);
7497 0 :
7498 0 : let mut remove_node = false;
7499 0 : while take > 0 {
7500 0 : match free_tids_by_node.get_mut(&node_id) {
7501 0 : Some(tids) => match tids.pop() {
7502 0 : Some(tid) => {
7503 0 : let max_promote_for_tenant = std::cmp::max(
7504 0 : tid.shard_count.count() as usize / locked.nodes.len(),
7505 0 : 1,
7506 0 : );
7507 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
7508 0 : if *promoted < max_promote_for_tenant {
7509 0 : plan.push(tid);
7510 0 : *promoted += 1;
7511 0 : take -= 1;
7512 0 : }
7513 : }
7514 : None => {
7515 0 : remove_node = true;
7516 0 : break;
7517 : }
7518 : },
7519 : None => {
7520 0 : break;
7521 : }
7522 : }
7523 : }
7524 :
7525 0 : if remove_node {
7526 0 : free_tids_by_node.remove(&node_id);
7527 0 : }
7528 : }
7529 :
7530 0 : plan
7531 0 : }
7532 :
7533 : /// Fill a node by promoting its secondaries until the cluster is balanced
7534 : /// with regards to attached shard counts. Note that this operation only
7535 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
7536 : /// This is a long running operation and it should run as a separate Tokio task.
7537 0 : pub(crate) async fn fill_node(
7538 0 : &self,
7539 0 : node_id: NodeId,
7540 0 : cancel: CancellationToken,
7541 0 : ) -> Result<(), OperationError> {
7542 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7543 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7544 0 : let reconciler_config = ReconcilerConfigBuilder::new()
7545 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7546 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7547 0 : .build();
7548 0 :
7549 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
7550 0 : let mut waiters = Vec::new();
7551 :
7552 : // Execute the plan we've composed above. Before aplying each move from the plan,
7553 : // we validate to ensure that it has not gone stale in the meantime.
7554 0 : while !tids_to_promote.is_empty() {
7555 0 : if cancel.is_cancelled() {
7556 0 : match self
7557 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7558 0 : .await
7559 : {
7560 0 : Ok(()) => return Err(OperationError::Cancelled),
7561 0 : Err(err) => {
7562 0 : return Err(OperationError::FinalizeError(
7563 0 : format!(
7564 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7565 0 : node_id, err
7566 0 : )
7567 0 : .into(),
7568 0 : ));
7569 : }
7570 : }
7571 0 : }
7572 0 :
7573 0 : {
7574 0 : let mut locked = self.inner.write().unwrap();
7575 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7576 :
7577 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
7578 0 : format!("node {node_id} was removed").into(),
7579 0 : ))?;
7580 :
7581 0 : let current_policy = node.get_scheduling();
7582 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
7583 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
7584 : // about it
7585 0 : return Err(OperationError::NodeStateChanged(
7586 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
7587 0 : ));
7588 0 : }
7589 :
7590 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7591 0 : if let Some(tid) = tids_to_promote.pop() {
7592 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
7593 : // If the node being filled is not a secondary anymore,
7594 : // skip the promotion.
7595 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7596 0 : continue;
7597 0 : }
7598 0 :
7599 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
7600 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
7601 0 : Err(e) => {
7602 0 : tracing::warn!(
7603 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7604 0 : "Scheduling error when filling pageserver {} : {e}", node_id
7605 : );
7606 : }
7607 : Ok(()) => {
7608 0 : tracing::info!(
7609 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7610 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
7611 : node_id,
7612 : previously_attached_to,
7613 : node_id
7614 : );
7615 :
7616 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
7617 0 : tenant_shard,
7618 0 : nodes,
7619 0 : reconciler_config,
7620 0 : ) {
7621 0 : waiters.push(waiter);
7622 0 : }
7623 : }
7624 : }
7625 0 : }
7626 : } else {
7627 0 : break;
7628 : }
7629 : }
7630 : }
7631 :
7632 0 : waiters = self
7633 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7634 0 : .await;
7635 : }
7636 :
7637 0 : while !waiters.is_empty() {
7638 0 : if cancel.is_cancelled() {
7639 0 : match self
7640 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7641 0 : .await
7642 : {
7643 0 : Ok(()) => return Err(OperationError::Cancelled),
7644 0 : Err(err) => {
7645 0 : return Err(OperationError::FinalizeError(
7646 0 : format!(
7647 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7648 0 : node_id, err
7649 0 : )
7650 0 : .into(),
7651 0 : ));
7652 : }
7653 : }
7654 0 : }
7655 0 :
7656 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
7657 :
7658 0 : waiters = self
7659 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7660 0 : .await;
7661 : }
7662 :
7663 0 : if let Err(err) = self
7664 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7665 0 : .await
7666 : {
7667 : // This isn't a huge issue since the filling process starts upon request. However, it
7668 : // will prevent the next drain from starting. The only case in which this can fail
7669 : // is database unavailability. Such a case will require manual intervention.
7670 0 : return Err(OperationError::FinalizeError(
7671 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
7672 0 : .into(),
7673 0 : ));
7674 0 : }
7675 0 :
7676 0 : Ok(())
7677 0 : }
7678 :
7679 : /// Updates scrubber metadata health check results.
7680 0 : pub(crate) async fn metadata_health_update(
7681 0 : &self,
7682 0 : update_req: MetadataHealthUpdateRequest,
7683 0 : ) -> Result<(), ApiError> {
7684 0 : let now = chrono::offset::Utc::now();
7685 0 : let (healthy_records, unhealthy_records) = {
7686 0 : let locked = self.inner.read().unwrap();
7687 0 : let healthy_records = update_req
7688 0 : .healthy_tenant_shards
7689 0 : .into_iter()
7690 0 : // Retain only health records associated with tenant shards managed by storage controller.
7691 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7692 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
7693 0 : .collect();
7694 0 : let unhealthy_records = update_req
7695 0 : .unhealthy_tenant_shards
7696 0 : .into_iter()
7697 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7698 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
7699 0 : .collect();
7700 0 :
7701 0 : (healthy_records, unhealthy_records)
7702 0 : };
7703 0 :
7704 0 : self.persistence
7705 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
7706 0 : .await?;
7707 0 : Ok(())
7708 0 : }
7709 :
7710 : /// Lists the tenant shards that has unhealthy metadata status.
7711 0 : pub(crate) async fn metadata_health_list_unhealthy(
7712 0 : &self,
7713 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
7714 0 : let result = self
7715 0 : .persistence
7716 0 : .list_unhealthy_metadata_health_records()
7717 0 : .await?
7718 0 : .iter()
7719 0 : .map(|p| p.get_tenant_shard_id().unwrap())
7720 0 : .collect();
7721 0 :
7722 0 : Ok(result)
7723 0 : }
7724 :
7725 : /// Lists the tenant shards that have not been scrubbed for some duration.
7726 0 : pub(crate) async fn metadata_health_list_outdated(
7727 0 : &self,
7728 0 : not_scrubbed_for: Duration,
7729 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
7730 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
7731 0 : let result = self
7732 0 : .persistence
7733 0 : .list_outdated_metadata_health_records(earlier)
7734 0 : .await?
7735 0 : .into_iter()
7736 0 : .map(|record| record.into())
7737 0 : .collect();
7738 0 : Ok(result)
7739 0 : }
7740 :
7741 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
7742 0 : self.inner.read().unwrap().get_leadership_status()
7743 0 : }
7744 :
7745 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
7746 0 : tracing::info!("Received step down request from peer");
7747 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
7748 :
7749 0 : self.inner.write().unwrap().step_down();
7750 0 : // TODO: would it make sense to have a time-out for this?
7751 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
7752 0 : .await;
7753 :
7754 0 : let mut global_observed = GlobalObservedState::default();
7755 0 : let locked = self.inner.read().unwrap();
7756 0 : for (tid, tenant_shard) in locked.tenants.iter() {
7757 0 : global_observed
7758 0 : .0
7759 0 : .insert(*tid, tenant_shard.observed.clone());
7760 0 : }
7761 :
7762 0 : global_observed
7763 0 : }
7764 :
7765 0 : pub(crate) async fn safekeepers_list(
7766 0 : &self,
7767 0 : ) -> Result<Vec<SafekeeperDescribeResponse>, DatabaseError> {
7768 0 : let locked = self.inner.read().unwrap();
7769 0 : let mut list = locked
7770 0 : .safekeepers
7771 0 : .iter()
7772 0 : .map(|sk| sk.1.describe_response())
7773 0 : .collect::<Result<Vec<_>, _>>()?;
7774 0 : list.sort_by_key(|v| v.id);
7775 0 : Ok(list)
7776 0 : }
7777 :
7778 0 : pub(crate) async fn get_safekeeper(
7779 0 : &self,
7780 0 : id: i64,
7781 0 : ) -> Result<SafekeeperDescribeResponse, DatabaseError> {
7782 0 : let locked = self.inner.read().unwrap();
7783 0 : let sk = locked
7784 0 : .safekeepers
7785 0 : .get(&NodeId(id as u64))
7786 0 : .ok_or(diesel::result::Error::NotFound)?;
7787 0 : sk.describe_response()
7788 0 : }
7789 :
7790 0 : pub(crate) async fn upsert_safekeeper(
7791 0 : &self,
7792 0 : record: crate::persistence::SafekeeperUpsert,
7793 0 : ) -> Result<(), DatabaseError> {
7794 0 : let node_id = NodeId(record.id as u64);
7795 0 : self.persistence.safekeeper_upsert(record.clone()).await?;
7796 : {
7797 0 : let mut locked = self.inner.write().unwrap();
7798 0 : let mut safekeepers = (*locked.safekeepers).clone();
7799 0 : match safekeepers.entry(node_id) {
7800 0 : std::collections::hash_map::Entry::Occupied(mut entry) => {
7801 0 : entry.get_mut().update_from_record(record);
7802 0 : }
7803 0 : std::collections::hash_map::Entry::Vacant(entry) => {
7804 0 : entry.insert(Safekeeper::from_persistence(
7805 0 : crate::persistence::SafekeeperPersistence::from_upsert(
7806 0 : record,
7807 0 : SkSchedulingPolicy::Pause,
7808 0 : ),
7809 0 : CancellationToken::new(),
7810 0 : ));
7811 0 : }
7812 : }
7813 0 : locked.safekeepers = Arc::new(safekeepers);
7814 0 : }
7815 0 : Ok(())
7816 0 : }
7817 :
7818 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
7819 0 : &self,
7820 0 : id: i64,
7821 0 : scheduling_policy: SkSchedulingPolicy,
7822 0 : ) -> Result<(), DatabaseError> {
7823 0 : self.persistence
7824 0 : .set_safekeeper_scheduling_policy(id, scheduling_policy)
7825 0 : .await?;
7826 0 : let node_id = NodeId(id as u64);
7827 0 : // After the change has been persisted successfully, update the in-memory state
7828 0 : {
7829 0 : let mut locked = self.inner.write().unwrap();
7830 0 : let mut safekeepers = (*locked.safekeepers).clone();
7831 0 : let sk = safekeepers
7832 0 : .get_mut(&node_id)
7833 0 : .ok_or(DatabaseError::Logical("Not found".to_string()))?;
7834 0 : sk.skp.scheduling_policy = String::from(scheduling_policy);
7835 0 :
7836 0 : locked.safekeepers = Arc::new(safekeepers);
7837 0 : }
7838 0 : Ok(())
7839 0 : }
7840 :
7841 0 : pub(crate) async fn update_shards_preferred_azs(
7842 0 : &self,
7843 0 : req: ShardsPreferredAzsRequest,
7844 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
7845 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
7846 0 : let updated = self
7847 0 : .persistence
7848 0 : .set_tenant_shard_preferred_azs(preferred_azs)
7849 0 : .await
7850 0 : .map_err(|err| {
7851 0 : ApiError::InternalServerError(anyhow::anyhow!(
7852 0 : "Failed to persist preferred AZs: {err}"
7853 0 : ))
7854 0 : })?;
7855 :
7856 0 : let mut updated_in_mem_and_db = Vec::default();
7857 0 :
7858 0 : let mut locked = self.inner.write().unwrap();
7859 0 : for (tid, az_id) in updated {
7860 0 : let shard = locked.tenants.get_mut(&tid);
7861 0 : if let Some(shard) = shard {
7862 0 : shard.set_preferred_az(az_id);
7863 0 : updated_in_mem_and_db.push(tid);
7864 0 : }
7865 : }
7866 :
7867 0 : Ok(ShardsPreferredAzsResponse {
7868 0 : updated: updated_in_mem_and_db,
7869 0 : })
7870 0 : }
7871 : }
|
