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::{
34 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder,
35 : ReconcilerPriority,
36 : },
37 : safekeeper::Safekeeper,
38 : scheduler::{MaySchedule, ScheduleContext, ScheduleError, ScheduleMode},
39 : tenant_shard::{
40 : MigrateAttachment, ObservedStateDelta, ReconcileNeeded, ReconcilerStatus,
41 : ScheduleOptimization, ScheduleOptimizationAction,
42 : },
43 : };
44 : use anyhow::Context;
45 : use control_plane::storage_controller::{
46 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
47 : };
48 : use diesel::result::DatabaseErrorKind;
49 : use futures::{stream::FuturesUnordered, StreamExt};
50 : use itertools::Itertools;
51 : use pageserver_api::{
52 : controller_api::{
53 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
54 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
55 : SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest,
56 : ShardsPreferredAzsResponse, SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse,
57 : TenantCreateResponseShard, TenantDescribeResponse, TenantDescribeResponseShard,
58 : TenantLocateResponse, TenantPolicyRequest, TenantShardMigrateRequest,
59 : TenantShardMigrateResponse,
60 : },
61 : models::{
62 : SecondaryProgress, TenantConfigPatchRequest, TenantConfigRequest,
63 : TimelineArchivalConfigRequest, TopTenantShardsRequest,
64 : },
65 : };
66 : use reqwest::StatusCode;
67 : use tracing::{instrument, Instrument};
68 :
69 : use crate::pageserver_client::PageserverClient;
70 : use http_utils::error::ApiError;
71 : use pageserver_api::{
72 : models::{
73 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
74 : PageserverUtilization, ShardParameters, TenantConfig, TenantLocationConfigRequest,
75 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
76 : TenantShardSplitResponse, TenantTimeTravelRequest, TimelineCreateRequest, TimelineInfo,
77 : },
78 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
79 : upcall_api::{
80 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
81 : ValidateResponse, ValidateResponseTenant,
82 : },
83 : };
84 : use pageserver_client::{mgmt_api, BlockUnblock};
85 : use tokio::sync::{mpsc::error::TrySendError, TryAcquireError};
86 : use tokio_util::sync::CancellationToken;
87 : use utils::{
88 : completion::Barrier,
89 : failpoint_support,
90 : generation::Generation,
91 : id::{NodeId, TenantId, TimelineId},
92 : pausable_failpoint,
93 : sync::gate::Gate,
94 : };
95 :
96 : use crate::{
97 : compute_hook::ComputeHook,
98 : heartbeater::{Heartbeater, PageserverState},
99 : node::{AvailabilityTransition, Node},
100 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
101 : reconciler::attached_location_conf,
102 : scheduler::Scheduler,
103 : tenant_shard::{
104 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
105 : ReconcilerWaiter, TenantShard,
106 : },
107 : };
108 :
109 : use context_iterator::TenantShardContextIterator;
110 :
111 : const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
112 :
113 : // For operations that should be quick, like attaching a new tenant
114 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
115 :
116 : // For operations that might be slow, like migrating a tenant with
117 : // some data in it.
118 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
119 :
120 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
121 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
122 : // input generation from future requests as authoritative.
123 : const INITIAL_GENERATION: Generation = Generation::new(0);
124 :
125 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
126 : /// up on unresponsive pageservers and proceed.
127 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
128 :
129 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
130 : /// This must be long enough to cover node restarts as well as normal operations: in future
131 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
132 :
133 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
134 : /// offline.
135 : ///
136 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
137 : /// handling of the re-attach response may take a long time and blocks heartbeats from
138 : /// being handled on the pageserver side.
139 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
140 :
141 : /// How often to send heartbeats to registered nodes?
142 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
143 :
144 : /// How long is too long for a reconciliation?
145 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
146 :
147 : #[derive(Clone, strum_macros::Display)]
148 : enum TenantOperations {
149 : Create,
150 : LocationConfig,
151 : ConfigSet,
152 : ConfigPatch,
153 : TimeTravelRemoteStorage,
154 : Delete,
155 : UpdatePolicy,
156 : ShardSplit,
157 : SecondaryDownload,
158 : TimelineCreate,
159 : TimelineDelete,
160 : AttachHook,
161 : TimelineArchivalConfig,
162 : TimelineDetachAncestor,
163 : TimelineGcBlockUnblock,
164 : DropDetached,
165 : DownloadHeatmapLayers,
166 : }
167 :
168 : #[derive(Clone, strum_macros::Display)]
169 : enum NodeOperations {
170 : Register,
171 : Configure,
172 : Delete,
173 : }
174 :
175 : /// The leadership status for the storage controller process.
176 : /// Allowed transitions are:
177 : /// 1. Leader -> SteppedDown
178 : /// 2. Candidate -> Leader
179 : #[derive(
180 : Eq,
181 : PartialEq,
182 : Copy,
183 : Clone,
184 : strum_macros::Display,
185 0 : strum_macros::EnumIter,
186 : measured::FixedCardinalityLabel,
187 : )]
188 : #[strum(serialize_all = "snake_case")]
189 : pub(crate) enum LeadershipStatus {
190 : /// This is the steady state where the storage controller can produce
191 : /// side effects in the cluster.
192 : Leader,
193 : /// We've been notified to step down by another candidate. No reconciliations
194 : /// take place in this state.
195 : SteppedDown,
196 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
197 : #[allow(unused)]
198 : Candidate,
199 : }
200 :
201 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
202 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
203 :
204 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
205 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
206 : // than they're being pushed onto the queue.
207 : const MAX_DELAYED_RECONCILES: usize = 10000;
208 :
209 : // Top level state available to all HTTP handlers
210 : struct ServiceState {
211 : leadership_status: LeadershipStatus,
212 :
213 : tenants: BTreeMap<TenantShardId, TenantShard>,
214 :
215 : nodes: Arc<HashMap<NodeId, Node>>,
216 :
217 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
218 :
219 : scheduler: Scheduler,
220 :
221 : /// Ongoing background operation on the cluster if any is running.
222 : /// Note that only one such operation may run at any given time,
223 : /// hence the type choice.
224 : ongoing_operation: Option<OperationHandler>,
225 :
226 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
227 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
228 : }
229 :
230 : /// Transform an error from a pageserver into an error to return to callers of a storage
231 : /// controller API.
232 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
233 0 : match e {
234 0 : mgmt_api::Error::SendRequest(e) => {
235 0 : // Presume errors sending requests are connectivity/availability issues
236 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
237 : }
238 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
239 0 : // Presume errors receiving body are connectivity/availability issues
240 0 : ApiError::ResourceUnavailable(
241 0 : format!("{node} error receiving error body: {str}").into(),
242 0 : )
243 : }
244 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
245 0 : // Return 500 for decoding errors.
246 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
247 : }
248 0 : mgmt_api::Error::ReceiveBody(err) => {
249 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
250 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
251 0 : ApiError::ResourceUnavailable(
252 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
253 0 : )
254 : }
255 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
256 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
257 : }
258 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
259 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
260 : }
261 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
262 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
263 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
264 : // internal server errors, showing that something is wrong with the pageserver or
265 : // storage controller's auth configuration.
266 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
267 : }
268 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
269 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
270 0 : // turn ask our clients to wait + retry
271 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
272 : }
273 0 : mgmt_api::Error::ApiError(status, msg) => {
274 0 : // Presume general case of pageserver API errors is that we tried to do something
275 0 : // that can't be done right now.
276 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
277 : }
278 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
279 : }
280 0 : }
281 :
282 : impl ServiceState {
283 0 : fn new(
284 0 : nodes: HashMap<NodeId, Node>,
285 0 : safekeepers: HashMap<NodeId, Safekeeper>,
286 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
287 0 : scheduler: Scheduler,
288 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
289 0 : initial_leadership_status: LeadershipStatus,
290 0 : ) -> Self {
291 0 : metrics::update_leadership_status(initial_leadership_status);
292 0 :
293 0 : Self {
294 0 : leadership_status: initial_leadership_status,
295 0 : tenants,
296 0 : nodes: Arc::new(nodes),
297 0 : safekeepers: Arc::new(safekeepers),
298 0 : scheduler,
299 0 : ongoing_operation: None,
300 0 : delayed_reconcile_rx,
301 0 : }
302 0 : }
303 :
304 0 : fn parts_mut(
305 0 : &mut self,
306 0 : ) -> (
307 0 : &mut Arc<HashMap<NodeId, Node>>,
308 0 : &mut BTreeMap<TenantShardId, TenantShard>,
309 0 : &mut Scheduler,
310 0 : ) {
311 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
312 0 : }
313 :
314 : #[allow(clippy::type_complexity)]
315 0 : fn parts_mut_sk(
316 0 : &mut self,
317 0 : ) -> (
318 0 : &mut Arc<HashMap<NodeId, Node>>,
319 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
320 0 : &mut BTreeMap<TenantShardId, TenantShard>,
321 0 : &mut Scheduler,
322 0 : ) {
323 0 : (
324 0 : &mut self.nodes,
325 0 : &mut self.safekeepers,
326 0 : &mut self.tenants,
327 0 : &mut self.scheduler,
328 0 : )
329 0 : }
330 :
331 0 : fn get_leadership_status(&self) -> LeadershipStatus {
332 0 : self.leadership_status
333 0 : }
334 :
335 0 : fn step_down(&mut self) {
336 0 : self.leadership_status = LeadershipStatus::SteppedDown;
337 0 : metrics::update_leadership_status(self.leadership_status);
338 0 : }
339 :
340 0 : fn become_leader(&mut self) {
341 0 : self.leadership_status = LeadershipStatus::Leader;
342 0 : metrics::update_leadership_status(self.leadership_status);
343 0 : }
344 : }
345 :
346 : #[derive(Clone)]
347 : pub struct Config {
348 : // All pageservers managed by one instance of this service must have
349 : // the same public key. This JWT token will be used to authenticate
350 : // this service to the pageservers it manages.
351 : pub jwt_token: Option<String>,
352 :
353 : // This JWT token will be used to authenticate this service to the control plane.
354 : pub control_plane_jwt_token: Option<String>,
355 :
356 : // This JWT token will be used to authenticate with other storage controller instances
357 : pub peer_jwt_token: Option<String>,
358 :
359 : /// Where the compute hook should send notifications of pageserver attachment locations
360 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
361 : /// assume it is running in a test environment and try to update neon_local.
362 : pub compute_hook_url: Option<String>,
363 :
364 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
365 : /// considered active. Once the grace period elapses, the next heartbeat failure will
366 : /// mark the pagseserver offline.
367 : pub max_offline_interval: Duration,
368 :
369 : /// Extended grace period within which pageserver may not respond to heartbeats.
370 : /// This extended grace period kicks in after the node has been drained for restart
371 : /// and/or upon handling the re-attach request from a node.
372 : pub max_warming_up_interval: Duration,
373 :
374 : /// How many normal-priority Reconcilers may be spawned concurrently
375 : pub reconciler_concurrency: usize,
376 :
377 : /// How many high-priority Reconcilers may be spawned concurrently
378 : pub priority_reconciler_concurrency: usize,
379 :
380 : /// How large must a shard grow in bytes before we split it?
381 : /// None disables auto-splitting.
382 : pub split_threshold: Option<u64>,
383 :
384 : // TODO: make this cfg(feature = "testing")
385 : pub neon_local_repo_dir: Option<PathBuf>,
386 :
387 : // Maximum acceptable download lag for the secondary location
388 : // while draining a node. If the secondary location is lagging
389 : // by more than the configured amount, then the secondary is not
390 : // upgraded to primary.
391 : pub max_secondary_lag_bytes: Option<u64>,
392 :
393 : pub heartbeat_interval: Duration,
394 :
395 : pub address_for_peers: Option<Uri>,
396 :
397 : pub start_as_candidate: bool,
398 :
399 : pub http_service_port: i32,
400 :
401 : pub long_reconcile_threshold: Duration,
402 : }
403 :
404 : impl From<DatabaseError> for ApiError {
405 0 : fn from(err: DatabaseError) -> ApiError {
406 0 : match err {
407 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
408 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
409 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
410 0 : ApiError::ShuttingDown
411 : }
412 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
413 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
414 : }
415 : }
416 0 : }
417 : }
418 :
419 : enum InitialShardScheduleOutcome {
420 : Scheduled(TenantCreateResponseShard),
421 : NotScheduled,
422 : ShardScheduleError(ScheduleError),
423 : }
424 :
425 : pub struct Service {
426 : inner: Arc<std::sync::RwLock<ServiceState>>,
427 : config: Config,
428 : persistence: Arc<Persistence>,
429 : compute_hook: Arc<ComputeHook>,
430 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
431 :
432 : heartbeater_ps: Heartbeater<Node, PageserverState>,
433 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
434 :
435 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
436 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
437 :
438 : // Locking on a tenant granularity (covers all shards in the tenant):
439 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
440 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
441 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
442 :
443 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
444 : // that transition it to/from Active.
445 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
446 :
447 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
448 : // and reconciliation on startup.
449 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
450 :
451 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
452 : // a human user might be waiting for.
453 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
454 :
455 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
456 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
457 : ///
458 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
459 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
460 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
461 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
462 :
463 : // Process shutdown will fire this token
464 : cancel: CancellationToken,
465 :
466 : // Child token of [`Service::cancel`] used by reconcilers
467 : reconcilers_cancel: CancellationToken,
468 :
469 : // Background tasks will hold this gate
470 : gate: Gate,
471 :
472 : // Reconcilers background tasks will hold this gate
473 : reconcilers_gate: Gate,
474 :
475 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
476 : /// passes, it isn't safe to do any actions that mutate tenants.
477 : pub(crate) startup_complete: Barrier,
478 : }
479 :
480 : impl From<ReconcileWaitError> for ApiError {
481 0 : fn from(value: ReconcileWaitError) -> Self {
482 0 : match value {
483 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
484 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
485 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
486 : }
487 0 : }
488 : }
489 :
490 : impl From<OperationError> for ApiError {
491 0 : fn from(value: OperationError) -> Self {
492 0 : match value {
493 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
494 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
495 : }
496 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
497 : }
498 0 : }
499 : }
500 :
501 : #[allow(clippy::large_enum_variant)]
502 : enum TenantCreateOrUpdate {
503 : Create(TenantCreateRequest),
504 : Update(Vec<ShardUpdate>),
505 : }
506 :
507 : struct ShardSplitParams {
508 : old_shard_count: ShardCount,
509 : new_shard_count: ShardCount,
510 : new_stripe_size: Option<ShardStripeSize>,
511 : targets: Vec<ShardSplitTarget>,
512 : policy: PlacementPolicy,
513 : config: TenantConfig,
514 : shard_ident: ShardIdentity,
515 : preferred_az_id: Option<AvailabilityZone>,
516 : }
517 :
518 : // When preparing for a shard split, we may either choose to proceed with the split,
519 : // or find that the work is already done and return NoOp.
520 : enum ShardSplitAction {
521 : Split(Box<ShardSplitParams>),
522 : NoOp(TenantShardSplitResponse),
523 : }
524 :
525 : // A parent shard which will be split
526 : struct ShardSplitTarget {
527 : parent_id: TenantShardId,
528 : node: Node,
529 : child_ids: Vec<TenantShardId>,
530 : }
531 :
532 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
533 : /// might not be available. We therefore use a queue of abort operations processed in the background.
534 : struct TenantShardSplitAbort {
535 : tenant_id: TenantId,
536 : /// The target values from the request that failed
537 : new_shard_count: ShardCount,
538 : new_stripe_size: Option<ShardStripeSize>,
539 : /// Until this abort op is complete, no other operations may be done on the tenant
540 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
541 : }
542 :
543 : #[derive(thiserror::Error, Debug)]
544 : enum TenantShardSplitAbortError {
545 : #[error(transparent)]
546 : Database(#[from] DatabaseError),
547 : #[error(transparent)]
548 : Remote(#[from] mgmt_api::Error),
549 : #[error("Unavailable")]
550 : Unavailable,
551 : }
552 :
553 : struct ShardUpdate {
554 : tenant_shard_id: TenantShardId,
555 : placement_policy: PlacementPolicy,
556 : tenant_config: TenantConfig,
557 :
558 : /// If this is None, generation is not updated.
559 : generation: Option<Generation>,
560 :
561 : /// If this is None, scheduling policy is not updated.
562 : scheduling_policy: Option<ShardSchedulingPolicy>,
563 : }
564 :
565 : enum StopReconciliationsReason {
566 : ShuttingDown,
567 : SteppingDown,
568 : }
569 :
570 : impl std::fmt::Display for StopReconciliationsReason {
571 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
572 0 : let s = match self {
573 0 : Self::ShuttingDown => "Shutting down",
574 0 : Self::SteppingDown => "Stepping down",
575 : };
576 0 : write!(writer, "{}", s)
577 0 : }
578 : }
579 :
580 : pub(crate) enum ReconcileResultRequest {
581 : ReconcileResult(ReconcileResult),
582 : Stop,
583 : }
584 :
585 : #[derive(Clone)]
586 : struct MutationLocation {
587 : node: Node,
588 : generation: Generation,
589 : }
590 :
591 : #[derive(Clone)]
592 : struct ShardMutationLocations {
593 : latest: MutationLocation,
594 : other: Vec<MutationLocation>,
595 : }
596 :
597 : #[derive(Default, Clone)]
598 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
599 :
600 : impl Service {
601 0 : pub fn get_config(&self) -> &Config {
602 0 : &self.config
603 0 : }
604 :
605 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
606 : /// view of the world, and determine which pageservers are responsive.
607 : #[instrument(skip_all)]
608 : async fn startup_reconcile(
609 : self: &Arc<Service>,
610 : current_leader: Option<ControllerPersistence>,
611 : leader_step_down_state: Option<GlobalObservedState>,
612 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
613 : Result<(), (TenantShardId, NotifyError)>,
614 : >,
615 : ) {
616 : // Startup reconciliation does I/O to other services: whether they
617 : // are responsive or not, we should aim to finish within our deadline, because:
618 : // - If we don't, a k8s readiness hook watching /ready will kill us.
619 : // - While we're waiting for startup reconciliation, we are not fully
620 : // available for end user operations like creating/deleting tenants and timelines.
621 : //
622 : // We set multiple deadlines to break up the time available between the phases of work: this is
623 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
624 : let start_at = Instant::now();
625 : let node_scan_deadline = start_at
626 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
627 : .expect("Reconcile timeout is a modest constant");
628 :
629 : let observed = if let Some(state) = leader_step_down_state {
630 : tracing::info!(
631 : "Using observed state received from leader at {}",
632 : current_leader.as_ref().unwrap().address
633 : );
634 :
635 : state
636 : } else {
637 : self.build_global_observed_state(node_scan_deadline).await
638 : };
639 :
640 : // Accumulate a list of any tenant locations that ought to be detached
641 : let mut cleanup = Vec::new();
642 :
643 : // Send initial heartbeat requests to all nodes loaded from the database
644 : let all_nodes = {
645 : let locked = self.inner.read().unwrap();
646 : locked.nodes.clone()
647 : };
648 : let (mut nodes_online, mut sks_online) =
649 : self.initial_heartbeat_round(all_nodes.keys()).await;
650 :
651 : // List of tenants for which we will attempt to notify compute of their location at startup
652 : let mut compute_notifications = Vec::new();
653 :
654 : // Populate intent and observed states for all tenants, based on reported state on pageservers
655 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
656 : let shard_count = {
657 : let mut locked = self.inner.write().unwrap();
658 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
659 :
660 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
661 : let mut new_nodes = (**nodes).clone();
662 : for (node_id, node) in new_nodes.iter_mut() {
663 : if let Some(utilization) = nodes_online.remove(node_id) {
664 : node.set_availability(NodeAvailability::Active(utilization));
665 : scheduler.node_upsert(node);
666 : }
667 : }
668 : *nodes = Arc::new(new_nodes);
669 :
670 : let mut new_sks = (**safekeepers).clone();
671 : for (node_id, node) in new_sks.iter_mut() {
672 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
673 : node.set_availability(SafekeeperState::Available {
674 : utilization,
675 : last_seen_at,
676 : });
677 : }
678 : }
679 : *safekeepers = Arc::new(new_sks);
680 :
681 : for (tenant_shard_id, observed_state) in observed.0 {
682 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
683 : for node_id in observed_state.locations.keys() {
684 : cleanup.push((tenant_shard_id, *node_id));
685 : }
686 :
687 : continue;
688 : };
689 :
690 : tenant_shard.observed = observed_state;
691 : }
692 :
693 : // Populate each tenant's intent state
694 : let mut schedule_context = ScheduleContext::default();
695 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
696 : if tenant_shard_id.shard_number == ShardNumber(0) {
697 : // Reset scheduling context each time we advance to the next Tenant
698 : schedule_context = ScheduleContext::default();
699 : }
700 :
701 : tenant_shard.intent_from_observed(scheduler);
702 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
703 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
704 : // not enough pageservers are available. The tenant may well still be available
705 : // to clients.
706 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
707 : } else {
708 : // If we're both intending and observed to be attached at a particular node, we will
709 : // emit a compute notification for this. In the case where our observed state does not
710 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
711 : if let Some(attached_at) = tenant_shard.stably_attached() {
712 : compute_notifications.push(compute_hook::ShardUpdate {
713 : tenant_shard_id: *tenant_shard_id,
714 : node_id: attached_at,
715 : stripe_size: tenant_shard.shard.stripe_size,
716 : preferred_az: tenant_shard
717 : .preferred_az()
718 0 : .map(|az| Cow::Owned(az.clone())),
719 : });
720 : }
721 : }
722 : }
723 :
724 : tenants.len()
725 : };
726 :
727 : // Before making any obeservable changes to the cluster, persist self
728 : // as leader in database and memory.
729 : let leadership = Leadership::new(
730 : self.persistence.clone(),
731 : self.config.clone(),
732 : self.cancel.child_token(),
733 : );
734 :
735 : if let Err(e) = leadership.become_leader(current_leader).await {
736 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
737 : std::process::exit(1);
738 : }
739 :
740 : self.inner.write().unwrap().become_leader();
741 :
742 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
743 : // generation_pageserver in the database.
744 :
745 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
746 : // will emit compute hook notifications when they reconcile.
747 : //
748 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
749 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
750 : // calls will be correctly ordered wrt these.
751 : //
752 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
753 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
754 : // unit and start doing I/O.
755 : tracing::info!(
756 : "Sending {} compute notifications",
757 : compute_notifications.len()
758 : );
759 : self.compute_hook.notify_background(
760 : compute_notifications,
761 : bg_compute_notify_result_tx.clone(),
762 : &self.cancel,
763 : );
764 :
765 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
766 : // which require it: under normal circumstances this should only include tenants that were in some
767 : // transient state before we restarted, or any tenants whose compute hooks failed above.
768 : tracing::info!("Checking for shards in need of reconciliation...");
769 : let reconcile_tasks = self.reconcile_all();
770 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
771 : // normal operations may proceed.
772 :
773 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
774 : // background because it does not need to complete in order to proceed with other work.
775 : if !cleanup.is_empty() {
776 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
777 : tokio::task::spawn({
778 : let cleanup_self = self.clone();
779 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
780 : });
781 : }
782 :
783 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
784 : }
785 :
786 0 : async fn initial_heartbeat_round<'a>(
787 0 : &self,
788 0 : node_ids: impl Iterator<Item = &'a NodeId>,
789 0 : ) -> (
790 0 : HashMap<NodeId, PageserverUtilization>,
791 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
792 0 : ) {
793 0 : assert!(!self.startup_complete.is_ready());
794 :
795 0 : let all_nodes = {
796 0 : let locked = self.inner.read().unwrap();
797 0 : locked.nodes.clone()
798 0 : };
799 0 :
800 0 : let mut nodes_to_heartbeat = HashMap::new();
801 0 : for node_id in node_ids {
802 0 : match all_nodes.get(node_id) {
803 0 : Some(node) => {
804 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
805 0 : }
806 : None => {
807 0 : tracing::warn!("Node {node_id} was removed during start-up");
808 : }
809 : }
810 : }
811 :
812 0 : let all_sks = {
813 0 : let locked = self.inner.read().unwrap();
814 0 : locked.safekeepers.clone()
815 0 : };
816 0 :
817 0 : tracing::info!("Sending initial heartbeats...");
818 0 : let res_ps = self
819 0 : .heartbeater_ps
820 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
821 0 : .await;
822 0 : let res_sk = self.heartbeater_sk.heartbeat(all_sks).await;
823 :
824 0 : let mut online_nodes = HashMap::new();
825 0 : if let Ok(deltas) = res_ps {
826 0 : for (node_id, status) in deltas.0 {
827 0 : match status {
828 0 : PageserverState::Available { utilization, .. } => {
829 0 : online_nodes.insert(node_id, utilization);
830 0 : }
831 0 : PageserverState::Offline => {}
832 : PageserverState::WarmingUp { .. } => {
833 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
834 : }
835 : }
836 : }
837 0 : }
838 :
839 0 : let mut online_sks = HashMap::new();
840 0 : if let Ok(deltas) = res_sk {
841 0 : for (node_id, status) in deltas.0 {
842 0 : match status {
843 : SafekeeperState::Available {
844 0 : utilization,
845 0 : last_seen_at,
846 0 : } => {
847 0 : online_sks.insert(node_id, (utilization, last_seen_at));
848 0 : }
849 0 : SafekeeperState::Offline => {}
850 : }
851 : }
852 0 : }
853 :
854 0 : (online_nodes, online_sks)
855 0 : }
856 :
857 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
858 : ///
859 : /// The result includes only nodes which responded within the deadline
860 0 : async fn scan_node_locations(
861 0 : &self,
862 0 : deadline: Instant,
863 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
864 0 : let nodes = {
865 0 : let locked = self.inner.read().unwrap();
866 0 : locked.nodes.clone()
867 0 : };
868 0 :
869 0 : let mut node_results = HashMap::new();
870 0 :
871 0 : let mut node_list_futs = FuturesUnordered::new();
872 0 :
873 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
874 0 : for node in nodes.values() {
875 0 : node_list_futs.push({
876 0 : async move {
877 0 : tracing::info!("Scanning shards on node {node}...");
878 0 : let timeout = Duration::from_secs(5);
879 0 : let response = node
880 0 : .with_client_retries(
881 0 : |client| async move { client.list_location_config().await },
882 0 : &self.config.jwt_token,
883 0 : 1,
884 0 : 5,
885 0 : timeout,
886 0 : &self.cancel,
887 0 : )
888 0 : .await;
889 0 : (node.get_id(), response)
890 0 : }
891 0 : });
892 0 : }
893 :
894 : loop {
895 0 : let (node_id, result) = tokio::select! {
896 0 : next = node_list_futs.next() => {
897 0 : match next {
898 0 : Some(result) => result,
899 : None =>{
900 : // We got results for all our nodes
901 0 : break;
902 : }
903 :
904 : }
905 : },
906 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
907 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
908 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
909 0 : break;
910 : }
911 : };
912 :
913 0 : let Some(list_response) = result else {
914 0 : tracing::info!("Shutdown during startup_reconcile");
915 0 : break;
916 : };
917 :
918 0 : match list_response {
919 0 : Err(e) => {
920 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
921 : }
922 0 : Ok(listing) => {
923 0 : node_results.insert(node_id, listing);
924 0 : }
925 : }
926 : }
927 :
928 0 : node_results
929 0 : }
930 :
931 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
932 0 : let node_listings = self.scan_node_locations(deadline).await;
933 0 : let mut observed = GlobalObservedState::default();
934 :
935 0 : for (node_id, location_confs) in node_listings {
936 0 : tracing::info!(
937 0 : "Received {} shard statuses from pageserver {}",
938 0 : location_confs.tenant_shards.len(),
939 : node_id
940 : );
941 :
942 0 : for (tid, location_conf) in location_confs.tenant_shards {
943 0 : let entry = observed.0.entry(tid).or_default();
944 0 : entry.locations.insert(
945 0 : node_id,
946 0 : ObservedStateLocation {
947 0 : conf: location_conf,
948 0 : },
949 0 : );
950 0 : }
951 : }
952 :
953 0 : observed
954 0 : }
955 :
956 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
957 : ///
958 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
959 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
960 : /// other task trying to attach it.
961 : #[instrument(skip_all)]
962 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
963 : let nodes = self.inner.read().unwrap().nodes.clone();
964 :
965 : for (tenant_shard_id, node_id) in cleanup {
966 : // A node reported a tenant_shard_id which is unknown to us: detach it.
967 : let Some(node) = nodes.get(&node_id) else {
968 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
969 : // a location to clean up on a node that has since been removed.
970 : tracing::info!(
971 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
972 : );
973 : continue;
974 : };
975 :
976 : if self.cancel.is_cancelled() {
977 : break;
978 : }
979 :
980 : let client = PageserverClient::new(
981 : node.get_id(),
982 : node.base_url(),
983 : self.config.jwt_token.as_deref(),
984 : );
985 : match client
986 : .location_config(
987 : tenant_shard_id,
988 : LocationConfig {
989 : mode: LocationConfigMode::Detached,
990 : generation: None,
991 : secondary_conf: None,
992 : shard_number: tenant_shard_id.shard_number.0,
993 : shard_count: tenant_shard_id.shard_count.literal(),
994 : shard_stripe_size: 0,
995 : tenant_conf: models::TenantConfig::default(),
996 : },
997 : None,
998 : false,
999 : )
1000 : .await
1001 : {
1002 : Ok(()) => {
1003 : tracing::info!(
1004 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1005 : );
1006 : }
1007 : Err(e) => {
1008 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1009 : // break anything.
1010 : tracing::error!(
1011 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1012 : );
1013 : }
1014 : }
1015 : }
1016 : }
1017 :
1018 : /// Long running background task that periodically wakes up and looks for shards that need
1019 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1020 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1021 : /// for those retries.
1022 : #[instrument(skip_all)]
1023 : async fn background_reconcile(self: &Arc<Self>) {
1024 : self.startup_complete.clone().wait().await;
1025 :
1026 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1027 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1028 : while !self.reconcilers_cancel.is_cancelled() {
1029 : tokio::select! {
1030 : _ = interval.tick() => {
1031 : let reconciles_spawned = self.reconcile_all();
1032 : if reconciles_spawned == 0 {
1033 : // Run optimizer only when we didn't find any other work to do
1034 : let optimizations = self.optimize_all().await;
1035 : if optimizations == 0 {
1036 : // Run new splits only when no optimizations are pending
1037 : self.autosplit_tenants().await;
1038 : }
1039 : }
1040 : }
1041 : _ = self.reconcilers_cancel.cancelled() => return
1042 : }
1043 : }
1044 : }
1045 : #[instrument(skip_all)]
1046 : async fn spawn_heartbeat_driver(&self) {
1047 : self.startup_complete.clone().wait().await;
1048 :
1049 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1050 : while !self.cancel.is_cancelled() {
1051 : tokio::select! {
1052 : _ = interval.tick() => { }
1053 : _ = self.cancel.cancelled() => return
1054 : };
1055 :
1056 : let nodes = {
1057 : let locked = self.inner.read().unwrap();
1058 : locked.nodes.clone()
1059 : };
1060 :
1061 : let safekeepers = {
1062 : let locked = self.inner.read().unwrap();
1063 : locked.safekeepers.clone()
1064 : };
1065 :
1066 : let res_ps = self.heartbeater_ps.heartbeat(nodes).await;
1067 : let res_sk = self.heartbeater_sk.heartbeat(safekeepers).await;
1068 : if let Ok(deltas) = res_ps {
1069 : let mut to_handle = Vec::default();
1070 :
1071 : for (node_id, state) in deltas.0 {
1072 : let new_availability = match state {
1073 : PageserverState::Available { utilization, .. } => {
1074 : NodeAvailability::Active(utilization)
1075 : }
1076 : PageserverState::WarmingUp { started_at } => {
1077 : NodeAvailability::WarmingUp(started_at)
1078 : }
1079 : PageserverState::Offline => {
1080 : // The node might have been placed in the WarmingUp state
1081 : // while the heartbeat round was on-going. Hence, filter out
1082 : // offline transitions for WarmingUp nodes that are still within
1083 : // their grace period.
1084 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1085 : .get_node(node_id)
1086 : .await
1087 : .as_ref()
1088 0 : .map(|n| n.get_availability())
1089 : {
1090 : let now = Instant::now();
1091 : if now - *started_at >= self.config.max_warming_up_interval {
1092 : NodeAvailability::Offline
1093 : } else {
1094 : NodeAvailability::WarmingUp(*started_at)
1095 : }
1096 : } else {
1097 : NodeAvailability::Offline
1098 : }
1099 : }
1100 : };
1101 :
1102 : let node_lock = trace_exclusive_lock(
1103 : &self.node_op_locks,
1104 : node_id,
1105 : NodeOperations::Configure,
1106 : )
1107 : .await;
1108 :
1109 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1110 :
1111 : // This is the code path for geniune availability transitions (i.e node
1112 : // goes unavailable and/or comes back online).
1113 : let res = self
1114 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1115 : .await;
1116 :
1117 : match res {
1118 : Ok(transition) => {
1119 : // Keep hold of the lock until the availability transitions
1120 : // have been handled in
1121 : // [`Service::handle_node_availability_transitions`] in order avoid
1122 : // racing with [`Service::external_node_configure`].
1123 : to_handle.push((node_id, node_lock, transition));
1124 : }
1125 : Err(ApiError::NotFound(_)) => {
1126 : // This should be rare, but legitimate since the heartbeats are done
1127 : // on a snapshot of the nodes.
1128 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1129 : }
1130 : Err(ApiError::ShuttingDown) => {
1131 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1132 : }
1133 : Err(err) => {
1134 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1135 : // becomes unavailable again, we may get an error here.
1136 : tracing::error!(
1137 : "Failed to update node state {} after heartbeat round: {}",
1138 : node_id,
1139 : err
1140 : );
1141 : }
1142 : }
1143 : }
1144 :
1145 : // We collected all the transitions above and now we handle them.
1146 : let res = self.handle_node_availability_transitions(to_handle).await;
1147 : if let Err(errs) = res {
1148 : for (node_id, err) in errs {
1149 : match err {
1150 : ApiError::NotFound(_) => {
1151 : // This should be rare, but legitimate since the heartbeats are done
1152 : // on a snapshot of the nodes.
1153 : tracing::info!(
1154 : "Node {} was not found after heartbeat round",
1155 : node_id
1156 : );
1157 : }
1158 : err => {
1159 : tracing::error!(
1160 : "Failed to handle availability transition for {} after heartbeat round: {}",
1161 : node_id,
1162 : err
1163 : );
1164 : }
1165 : }
1166 : }
1167 : }
1168 : }
1169 : if let Ok(deltas) = res_sk {
1170 : let mut locked = self.inner.write().unwrap();
1171 : let mut safekeepers = (*locked.safekeepers).clone();
1172 : for (id, state) in deltas.0 {
1173 : let Some(sk) = safekeepers.get_mut(&id) else {
1174 : tracing::info!("Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}");
1175 : continue;
1176 : };
1177 : sk.set_availability(state);
1178 : }
1179 : locked.safekeepers = Arc::new(safekeepers);
1180 : }
1181 : }
1182 : }
1183 :
1184 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1185 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1186 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1187 : /// will indicate that reconciliation is not needed.
1188 : #[instrument(skip_all, fields(
1189 : seq=%result.sequence,
1190 : tenant_id=%result.tenant_shard_id.tenant_id,
1191 : shard_id=%result.tenant_shard_id.shard_slug(),
1192 : ))]
1193 : fn process_result(&self, result: ReconcileResult) {
1194 : let mut locked = self.inner.write().unwrap();
1195 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1196 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1197 : // A reconciliation result might race with removing a tenant: drop results for
1198 : // tenants that aren't in our map.
1199 : return;
1200 : };
1201 :
1202 : // Usually generation should only be updated via this path, so the max() isn't
1203 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1204 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1205 :
1206 : // If the reconciler signals that it failed to notify compute, set this state on
1207 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1208 : tenant.pending_compute_notification = result.pending_compute_notification;
1209 :
1210 : // Let the TenantShard know it is idle.
1211 : tenant.reconcile_complete(result.sequence);
1212 :
1213 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1214 : // make to the tenant
1215 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1216 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1217 : // make to the tenant
1218 0 : let node = nodes.get(delta.node_id())?;
1219 :
1220 0 : if node.is_available() {
1221 0 : return Some(delta);
1222 0 : }
1223 0 :
1224 0 : // In case a node became unavailable concurrently with the reconcile, observed
1225 0 : // locations on it are now uncertain. By convention, set them to None in order
1226 0 : // for them to get refreshed when the node comes back online.
1227 0 : Some(ObservedStateDelta::Upsert(Box::new((
1228 0 : node.get_id(),
1229 0 : ObservedStateLocation { conf: None },
1230 0 : ))))
1231 0 : });
1232 :
1233 : match result.result {
1234 : Ok(()) => {
1235 : tenant.apply_observed_deltas(deltas);
1236 : tenant.waiter.advance(result.sequence);
1237 : }
1238 : Err(e) => {
1239 : match e {
1240 : ReconcileError::Cancel => {
1241 : tracing::info!("Reconciler was cancelled");
1242 : }
1243 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1244 : // This might be due to the reconciler getting cancelled, or it might
1245 : // be due to the `Node` being marked offline.
1246 : tracing::info!("Reconciler cancelled during pageserver API call");
1247 : }
1248 : _ => {
1249 : tracing::warn!("Reconcile error: {}", e);
1250 : }
1251 : }
1252 :
1253 : // Ordering: populate last_error before advancing error_seq,
1254 : // so that waiters will see the correct error after waiting.
1255 : tenant.set_last_error(result.sequence, e);
1256 :
1257 : // Skip deletions on reconcile failures
1258 : let upsert_deltas =
1259 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1260 : tenant.apply_observed_deltas(upsert_deltas);
1261 : }
1262 : }
1263 :
1264 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1265 : if tenant.policy == PlacementPolicy::Detached {
1266 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1267 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1268 : // duration of the request.
1269 : let guard = self.tenant_op_locks.try_exclusive(
1270 : tenant.tenant_shard_id.tenant_id,
1271 : TenantOperations::DropDetached,
1272 : );
1273 : if let Some(guard) = guard {
1274 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1275 : }
1276 : }
1277 :
1278 : // Maybe some other work can proceed now that this job finished.
1279 : //
1280 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1281 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1282 : if self.reconciler_concurrency.available_permits() > 0 {
1283 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1284 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1285 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1286 : shard.delayed_reconcile = false;
1287 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1288 : }
1289 :
1290 : if self.reconciler_concurrency.available_permits() == 0 {
1291 : break;
1292 : }
1293 : }
1294 : }
1295 : }
1296 :
1297 0 : async fn process_results(
1298 0 : &self,
1299 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1300 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1301 0 : Result<(), (TenantShardId, NotifyError)>,
1302 0 : >,
1303 0 : ) {
1304 : loop {
1305 : // Wait for the next result, or for cancellation
1306 0 : tokio::select! {
1307 0 : r = result_rx.recv() => {
1308 0 : match r {
1309 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1310 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1311 : }
1312 : }
1313 0 : _ = async{
1314 0 : match bg_compute_hook_result_rx.recv().await {
1315 0 : Some(result) => {
1316 0 : if let Err((tenant_shard_id, notify_error)) = result {
1317 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1318 0 : let mut locked = self.inner.write().unwrap();
1319 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1320 0 : shard.pending_compute_notification = true;
1321 0 : }
1322 :
1323 0 : }
1324 : },
1325 : None => {
1326 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1327 0 : self.cancel.cancelled().await;
1328 : }
1329 : }
1330 0 : } => {},
1331 0 : _ = self.cancel.cancelled() => {
1332 0 : break;
1333 : }
1334 : };
1335 : }
1336 0 : }
1337 :
1338 0 : async fn process_aborts(
1339 0 : &self,
1340 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1341 0 : ) {
1342 : loop {
1343 : // Wait for the next result, or for cancellation
1344 0 : let op = tokio::select! {
1345 0 : r = abort_rx.recv() => {
1346 0 : match r {
1347 0 : Some(op) => {op},
1348 0 : None => {break;}
1349 : }
1350 : }
1351 0 : _ = self.cancel.cancelled() => {
1352 0 : break;
1353 : }
1354 : };
1355 :
1356 : // Retry until shutdown: we must keep this request object alive until it is properly
1357 : // processed, as it holds a lock guard that prevents other operations trying to do things
1358 : // to the tenant while it is in a weird part-split state.
1359 0 : while !self.cancel.is_cancelled() {
1360 0 : match self.abort_tenant_shard_split(&op).await {
1361 0 : Ok(_) => break,
1362 0 : Err(e) => {
1363 0 : tracing::warn!(
1364 0 : "Failed to abort shard split on {}, will retry: {e}",
1365 : op.tenant_id
1366 : );
1367 :
1368 : // If a node is unavailable, we hope that it has been properly marked Offline
1369 : // when we retry, so that the abort op will succeed. If the abort op is failing
1370 : // for some other reason, we will keep retrying forever, or until a human notices
1371 : // and does something about it (either fixing a pageserver or restarting the controller).
1372 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1373 0 : .await
1374 0 : .ok();
1375 : }
1376 : }
1377 : }
1378 : }
1379 0 : }
1380 :
1381 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1382 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1383 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1384 0 :
1385 0 : let leadership_cancel = CancellationToken::new();
1386 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1387 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1388 :
1389 : // Apply the migrations **after** the current leader has stepped down
1390 : // (or we've given up waiting for it), but **before** reading from the
1391 : // database. The only exception is reading the current leader before
1392 : // migrating.
1393 0 : persistence.migration_run().await?;
1394 :
1395 0 : tracing::info!("Loading nodes from database...");
1396 0 : let nodes = persistence
1397 0 : .list_nodes()
1398 0 : .await?
1399 0 : .into_iter()
1400 0 : .map(Node::from_persistent)
1401 0 : .collect::<Vec<_>>();
1402 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1403 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1404 0 : metrics::METRICS_REGISTRY
1405 0 : .metrics_group
1406 0 : .storage_controller_pageserver_nodes
1407 0 : .set(nodes.len() as i64);
1408 0 :
1409 0 : tracing::info!("Loading safekeepers from database...");
1410 0 : let safekeepers = persistence
1411 0 : .list_safekeepers()
1412 0 : .await?
1413 0 : .into_iter()
1414 0 : .map(|skp| Safekeeper::from_persistence(skp, CancellationToken::new()))
1415 0 : .collect::<Vec<_>>();
1416 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1417 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1418 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1419 :
1420 0 : tracing::info!("Loading shards from database...");
1421 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1422 0 : tracing::info!(
1423 0 : "Loaded {} shards from database.",
1424 0 : tenant_shard_persistence.len()
1425 : );
1426 :
1427 : // If any shard splits were in progress, reset the database state to abort them
1428 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1429 0 : HashMap::new();
1430 0 : for tsp in &mut tenant_shard_persistence {
1431 0 : let shard = tsp.get_shard_identity()?;
1432 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1433 0 : let entry = tenant_shard_count_min_max
1434 0 : .entry(tenant_shard_id.tenant_id)
1435 0 : .or_insert_with(|| (shard.count, shard.count));
1436 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1437 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1438 0 : }
1439 :
1440 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1441 0 : if count_min != count_max {
1442 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1443 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1444 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1445 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1446 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1447 :
1448 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1449 : // identified this tenant has having mismatching min/max counts.
1450 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1451 :
1452 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1453 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1454 0 : // Set idle split state on those shards that we will retain.
1455 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1456 0 : if tsp_tenant_id == tenant_id
1457 0 : && tsp.get_shard_identity().unwrap().count == count_min
1458 0 : {
1459 0 : tsp.splitting = SplitState::Idle;
1460 0 : } else if tsp_tenant_id == tenant_id {
1461 : // Leave the splitting state on the child shards: this will be used next to
1462 : // drop them.
1463 0 : tracing::info!(
1464 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1465 : );
1466 0 : }
1467 0 : });
1468 0 :
1469 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1470 0 : tenant_shard_persistence.retain(|tsp| {
1471 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1472 0 : || tsp.splitting == SplitState::Idle
1473 0 : });
1474 0 : }
1475 : }
1476 :
1477 0 : let mut tenants = BTreeMap::new();
1478 0 :
1479 0 : let mut scheduler = Scheduler::new(nodes.values());
1480 :
1481 : #[cfg(feature = "testing")]
1482 : {
1483 : use pageserver_api::controller_api::AvailabilityZone;
1484 :
1485 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1486 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1487 : // after when pageservers start up and register.
1488 0 : let mut node_ids = HashSet::new();
1489 0 : for tsp in &tenant_shard_persistence {
1490 0 : if let Some(node_id) = tsp.generation_pageserver {
1491 0 : node_ids.insert(node_id);
1492 0 : }
1493 : }
1494 0 : for node_id in node_ids {
1495 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1496 0 : let node = Node::new(
1497 0 : NodeId(node_id as u64),
1498 0 : "".to_string(),
1499 0 : 123,
1500 0 : "".to_string(),
1501 0 : 123,
1502 0 : AvailabilityZone("test_az".to_string()),
1503 0 : );
1504 0 :
1505 0 : scheduler.node_upsert(&node);
1506 : }
1507 : }
1508 0 : for tsp in tenant_shard_persistence {
1509 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1510 :
1511 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1512 : // it with what we can infer: the node for which a generation was most recently issued.
1513 0 : let mut intent = IntentState::new(
1514 0 : tsp.preferred_az_id
1515 0 : .as_ref()
1516 0 : .map(|az| AvailabilityZone(az.clone())),
1517 0 : );
1518 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1519 : {
1520 0 : if nodes.contains_key(&generation_pageserver) {
1521 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1522 0 : } else {
1523 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1524 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1525 : // on different pageservers.
1526 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1527 : }
1528 0 : }
1529 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1530 :
1531 0 : tenants.insert(tenant_shard_id, new_tenant);
1532 : }
1533 :
1534 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1535 0 :
1536 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1537 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1538 0 : tokio::sync::mpsc::channel(512);
1539 0 :
1540 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1541 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1542 0 :
1543 0 : let cancel = CancellationToken::new();
1544 0 : let reconcilers_cancel = cancel.child_token();
1545 0 :
1546 0 : let heartbeater_ps = Heartbeater::new(
1547 0 : config.jwt_token.clone(),
1548 0 : config.max_offline_interval,
1549 0 : config.max_warming_up_interval,
1550 0 : cancel.clone(),
1551 0 : );
1552 0 :
1553 0 : let heartbeater_sk = Heartbeater::new(
1554 0 : config.jwt_token.clone(),
1555 0 : config.max_offline_interval,
1556 0 : config.max_warming_up_interval,
1557 0 : cancel.clone(),
1558 0 : );
1559 :
1560 0 : let initial_leadership_status = if config.start_as_candidate {
1561 0 : LeadershipStatus::Candidate
1562 : } else {
1563 0 : LeadershipStatus::Leader
1564 : };
1565 :
1566 0 : let this = Arc::new(Self {
1567 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1568 0 : nodes,
1569 0 : safekeepers,
1570 0 : tenants,
1571 0 : scheduler,
1572 0 : delayed_reconcile_rx,
1573 0 : initial_leadership_status,
1574 0 : ))),
1575 0 : config: config.clone(),
1576 0 : persistence,
1577 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1578 0 : result_tx,
1579 0 : heartbeater_ps,
1580 0 : heartbeater_sk,
1581 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1582 0 : config.reconciler_concurrency,
1583 0 : )),
1584 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1585 0 : config.priority_reconciler_concurrency,
1586 0 : )),
1587 0 : delayed_reconcile_tx,
1588 0 : abort_tx,
1589 0 : startup_complete: startup_complete.clone(),
1590 0 : cancel,
1591 0 : reconcilers_cancel,
1592 0 : gate: Gate::default(),
1593 0 : reconcilers_gate: Gate::default(),
1594 0 : tenant_op_locks: Default::default(),
1595 0 : node_op_locks: Default::default(),
1596 0 : });
1597 0 :
1598 0 : let result_task_this = this.clone();
1599 0 : tokio::task::spawn(async move {
1600 : // Block shutdown until we're done (we must respect self.cancel)
1601 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1602 0 : result_task_this
1603 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1604 0 : .await
1605 0 : }
1606 0 : });
1607 0 :
1608 0 : tokio::task::spawn({
1609 0 : let this = this.clone();
1610 0 : async move {
1611 : // Block shutdown until we're done (we must respect self.cancel)
1612 0 : if let Ok(_gate) = this.gate.enter() {
1613 0 : this.process_aborts(abort_rx).await
1614 0 : }
1615 0 : }
1616 0 : });
1617 0 :
1618 0 : tokio::task::spawn({
1619 0 : let this = this.clone();
1620 0 : async move {
1621 0 : if let Ok(_gate) = this.gate.enter() {
1622 : loop {
1623 0 : tokio::select! {
1624 0 : _ = this.cancel.cancelled() => {
1625 0 : break;
1626 : },
1627 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1628 0 : };
1629 0 : this.tenant_op_locks.housekeeping();
1630 : }
1631 0 : }
1632 0 : }
1633 0 : });
1634 0 :
1635 0 : tokio::task::spawn({
1636 0 : let this = this.clone();
1637 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1638 0 : // is done.
1639 0 : let startup_completion = startup_completion.clone();
1640 0 : async move {
1641 : // Block shutdown until we're done (we must respect self.cancel)
1642 0 : let Ok(_gate) = this.gate.enter() else {
1643 0 : return;
1644 : };
1645 :
1646 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1647 0 : .await;
1648 :
1649 0 : drop(startup_completion);
1650 0 : }
1651 0 : });
1652 0 :
1653 0 : tokio::task::spawn({
1654 0 : let this = this.clone();
1655 0 : let startup_complete = startup_complete.clone();
1656 0 : async move {
1657 0 : startup_complete.wait().await;
1658 0 : this.background_reconcile().await;
1659 0 : }
1660 0 : });
1661 0 :
1662 0 : tokio::task::spawn({
1663 0 : let this = this.clone();
1664 0 : let startup_complete = startup_complete.clone();
1665 0 : async move {
1666 0 : startup_complete.wait().await;
1667 0 : this.spawn_heartbeat_driver().await;
1668 0 : }
1669 0 : });
1670 0 :
1671 0 : Ok(this)
1672 0 : }
1673 :
1674 0 : pub(crate) async fn attach_hook(
1675 0 : &self,
1676 0 : attach_req: AttachHookRequest,
1677 0 : ) -> anyhow::Result<AttachHookResponse> {
1678 0 : let _tenant_lock = trace_exclusive_lock(
1679 0 : &self.tenant_op_locks,
1680 0 : attach_req.tenant_shard_id.tenant_id,
1681 0 : TenantOperations::AttachHook,
1682 0 : )
1683 0 : .await;
1684 :
1685 : // This is a test hook. To enable using it on tenants that were created directly with
1686 : // the pageserver API (not via this service), we will auto-create any missing tenant
1687 : // shards with default state.
1688 0 : let insert = {
1689 0 : match self
1690 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1691 0 : .await
1692 : {
1693 0 : Ok(_) => false,
1694 0 : Err(ApiError::NotFound(_)) => true,
1695 0 : Err(e) => return Err(e.into()),
1696 : }
1697 : };
1698 :
1699 0 : if insert {
1700 0 : let tsp = TenantShardPersistence {
1701 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1702 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1703 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1704 0 : shard_stripe_size: 0,
1705 0 : generation: attach_req.generation_override.or(Some(0)),
1706 0 : generation_pageserver: None,
1707 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1708 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1709 0 : splitting: SplitState::default(),
1710 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1711 0 : .unwrap(),
1712 0 : preferred_az_id: None,
1713 0 : };
1714 0 :
1715 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1716 0 : Err(e) => match e {
1717 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1718 : DatabaseErrorKind::UniqueViolation,
1719 : _,
1720 : )) => {
1721 0 : tracing::info!(
1722 0 : "Raced with another request to insert tenant {}",
1723 : attach_req.tenant_shard_id
1724 : )
1725 : }
1726 0 : _ => return Err(e.into()),
1727 : },
1728 : Ok(()) => {
1729 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1730 :
1731 0 : let mut locked = self.inner.write().unwrap();
1732 0 : locked.tenants.insert(
1733 0 : attach_req.tenant_shard_id,
1734 0 : TenantShard::new(
1735 0 : attach_req.tenant_shard_id,
1736 0 : ShardIdentity::unsharded(),
1737 0 : PlacementPolicy::Attached(0),
1738 0 : None,
1739 0 : ),
1740 0 : );
1741 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1742 : }
1743 : }
1744 0 : }
1745 :
1746 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1747 0 : let maybe_tenant_conf = {
1748 0 : let locked = self.inner.write().unwrap();
1749 0 : locked
1750 0 : .tenants
1751 0 : .get(&attach_req.tenant_shard_id)
1752 0 : .map(|t| t.config.clone())
1753 0 : };
1754 0 :
1755 0 : match maybe_tenant_conf {
1756 0 : Some(conf) => {
1757 0 : let new_generation = self
1758 0 : .persistence
1759 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1760 0 : .await?;
1761 :
1762 : // Persist the placement policy update. This is required
1763 : // when we reattaching a detached tenant.
1764 0 : self.persistence
1765 0 : .update_tenant_shard(
1766 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1767 0 : Some(PlacementPolicy::Attached(0)),
1768 0 : Some(conf),
1769 0 : None,
1770 0 : None,
1771 0 : )
1772 0 : .await?;
1773 0 : Some(new_generation)
1774 : }
1775 : None => {
1776 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1777 : }
1778 : }
1779 : } else {
1780 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1781 0 : None
1782 : };
1783 :
1784 0 : let mut locked = self.inner.write().unwrap();
1785 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1786 0 :
1787 0 : let tenant_shard = tenants
1788 0 : .get_mut(&attach_req.tenant_shard_id)
1789 0 : .expect("Checked for existence above");
1790 :
1791 0 : if let Some(new_generation) = new_generation {
1792 0 : tenant_shard.generation = Some(new_generation);
1793 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1794 0 : } else {
1795 : // This is a detach notification. We must update placement policy to avoid re-attaching
1796 : // during background scheduling/reconciliation, or during storage controller restart.
1797 0 : assert!(attach_req.node_id.is_none());
1798 0 : tenant_shard.policy = PlacementPolicy::Detached;
1799 : }
1800 :
1801 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1802 0 : tracing::info!(
1803 : tenant_id = %attach_req.tenant_shard_id,
1804 : ps_id = %attaching_pageserver,
1805 : generation = ?tenant_shard.generation,
1806 0 : "issuing",
1807 : );
1808 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1809 0 : tracing::info!(
1810 : tenant_id = %attach_req.tenant_shard_id,
1811 : %ps_id,
1812 : generation = ?tenant_shard.generation,
1813 0 : "dropping",
1814 : );
1815 : } else {
1816 0 : tracing::info!(
1817 : tenant_id = %attach_req.tenant_shard_id,
1818 0 : "no-op: tenant already has no pageserver");
1819 : }
1820 0 : tenant_shard
1821 0 : .intent
1822 0 : .set_attached(scheduler, attach_req.node_id);
1823 0 :
1824 0 : tracing::info!(
1825 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1826 0 : attach_req.tenant_shard_id,
1827 0 : tenant_shard.generation,
1828 0 : // TODO: this is an odd number of 0xf's
1829 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1830 : );
1831 :
1832 : // Trick the reconciler into not doing anything for this tenant: this helps
1833 : // tests that manually configure a tenant on the pagesrever, and then call this
1834 : // attach hook: they don't want background reconciliation to modify what they
1835 : // did to the pageserver.
1836 : #[cfg(feature = "testing")]
1837 : {
1838 0 : if let Some(node_id) = attach_req.node_id {
1839 0 : tenant_shard.observed.locations = HashMap::from([(
1840 0 : node_id,
1841 0 : ObservedStateLocation {
1842 0 : conf: Some(attached_location_conf(
1843 0 : tenant_shard.generation.unwrap(),
1844 0 : &tenant_shard.shard,
1845 0 : &tenant_shard.config,
1846 0 : &PlacementPolicy::Attached(0),
1847 0 : )),
1848 0 : },
1849 0 : )]);
1850 0 : } else {
1851 0 : tenant_shard.observed.locations.clear();
1852 0 : }
1853 : }
1854 :
1855 0 : Ok(AttachHookResponse {
1856 0 : gen: attach_req
1857 0 : .node_id
1858 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1859 0 : })
1860 0 : }
1861 :
1862 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1863 0 : let locked = self.inner.read().unwrap();
1864 0 :
1865 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1866 0 :
1867 0 : InspectResponse {
1868 0 : attachment: tenant_shard.and_then(|s| {
1869 0 : s.intent
1870 0 : .get_attached()
1871 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1872 0 : }),
1873 0 : }
1874 0 : }
1875 :
1876 : // When the availability state of a node transitions to active, we must do a full reconciliation
1877 : // of LocationConfigs on that node. This is because while a node was offline:
1878 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1879 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1880 : //
1881 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1882 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1883 : // this function.
1884 : //
1885 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1886 : // for written for a single node rather than as a batch job for all nodes.
1887 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1888 : async fn node_activate_reconcile(
1889 : &self,
1890 : mut node: Node,
1891 : _lock: &TracingExclusiveGuard<NodeOperations>,
1892 : ) -> Result<(), ApiError> {
1893 : // This Node is a mutable local copy: we will set it active so that we can use its
1894 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1895 : // later.
1896 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1897 :
1898 : let configs = match node
1899 : .with_client_retries(
1900 0 : |client| async move { client.list_location_config().await },
1901 : &self.config.jwt_token,
1902 : 1,
1903 : 5,
1904 : SHORT_RECONCILE_TIMEOUT,
1905 : &self.cancel,
1906 : )
1907 : .await
1908 : {
1909 : None => {
1910 : // We're shutting down (the Node's cancellation token can't have fired, because
1911 : // we're the only scope that has a reference to it, and we didn't fire it).
1912 : return Err(ApiError::ShuttingDown);
1913 : }
1914 : Some(Err(e)) => {
1915 : // This node didn't succeed listing its locations: it may not proceed to active state
1916 : // as it is apparently unavailable.
1917 : return Err(ApiError::PreconditionFailed(
1918 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1919 : ));
1920 : }
1921 : Some(Ok(configs)) => configs,
1922 : };
1923 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1924 :
1925 : let mut cleanup = Vec::new();
1926 : {
1927 : let mut locked = self.inner.write().unwrap();
1928 :
1929 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1930 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1931 : cleanup.push(tenant_shard_id);
1932 : continue;
1933 : };
1934 : tenant_shard
1935 : .observed
1936 : .locations
1937 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1938 : }
1939 : }
1940 :
1941 : for tenant_shard_id in cleanup {
1942 : tracing::info!("Detaching {tenant_shard_id}");
1943 : match node
1944 : .with_client_retries(
1945 0 : |client| async move {
1946 0 : let config = LocationConfig {
1947 0 : mode: LocationConfigMode::Detached,
1948 0 : generation: None,
1949 0 : secondary_conf: None,
1950 0 : shard_number: tenant_shard_id.shard_number.0,
1951 0 : shard_count: tenant_shard_id.shard_count.literal(),
1952 0 : shard_stripe_size: 0,
1953 0 : tenant_conf: models::TenantConfig::default(),
1954 0 : };
1955 0 : client
1956 0 : .location_config(tenant_shard_id, config, None, false)
1957 0 : .await
1958 0 : },
1959 : &self.config.jwt_token,
1960 : 1,
1961 : 5,
1962 : SHORT_RECONCILE_TIMEOUT,
1963 : &self.cancel,
1964 : )
1965 : .await
1966 : {
1967 : None => {
1968 : // We're shutting down (the Node's cancellation token can't have fired, because
1969 : // we're the only scope that has a reference to it, and we didn't fire it).
1970 : return Err(ApiError::ShuttingDown);
1971 : }
1972 : Some(Err(e)) => {
1973 : // Do not let the node proceed to Active state if it is not responsive to requests
1974 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1975 : // detach completing: we should not let this node back into the set of nodes considered
1976 : // okay for scheduling.
1977 : return Err(ApiError::Conflict(format!(
1978 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1979 : )));
1980 : }
1981 : Some(Ok(_)) => {}
1982 : };
1983 : }
1984 :
1985 : Ok(())
1986 : }
1987 :
1988 0 : pub(crate) async fn re_attach(
1989 0 : &self,
1990 0 : reattach_req: ReAttachRequest,
1991 0 : ) -> Result<ReAttachResponse, ApiError> {
1992 0 : if let Some(register_req) = reattach_req.register {
1993 0 : self.node_register(register_req).await?;
1994 0 : }
1995 :
1996 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1997 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1998 :
1999 0 : tracing::info!(
2000 : node_id=%reattach_req.node_id,
2001 0 : "Incremented {} tenant shards' generations",
2002 0 : incremented_generations.len()
2003 : );
2004 :
2005 : // Apply the updated generation to our in-memory state, and
2006 : // gather discover secondary locations.
2007 0 : let mut locked = self.inner.write().unwrap();
2008 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2009 0 :
2010 0 : let mut response = ReAttachResponse {
2011 0 : tenants: Vec::new(),
2012 0 : };
2013 :
2014 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2015 : // to call location_conf API with an old generation. Wait for cancellation to complete
2016 : // before responding to this request. Requires well implemented CancellationToken logic
2017 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2018 : // request in flight over the network: TODO handle that by making location_conf API refuse
2019 : // to go backward in generations.
2020 :
2021 : // Scan through all shards, applying updates for ones where we updated generation
2022 : // and identifying shards that intend to have a secondary location on this node.
2023 0 : for (tenant_shard_id, shard) in tenants {
2024 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2025 0 : let new_gen = *new_gen;
2026 0 : response.tenants.push(ReAttachResponseTenant {
2027 0 : id: *tenant_shard_id,
2028 0 : gen: Some(new_gen.into().unwrap()),
2029 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2030 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2031 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2032 0 : // the stale/multi states at this point.
2033 0 : mode: LocationConfigMode::AttachedSingle,
2034 0 : });
2035 0 :
2036 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2037 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2038 : // Why can we update `observed` even though we're not sure our response will be received
2039 : // by the pageserver? Because the pageserver will not proceed with startup until
2040 : // it has processed response: if it loses it, we'll see another request and increment
2041 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2042 0 : if let Some(conf) = observed.conf.as_mut() {
2043 0 : conf.generation = new_gen.into();
2044 0 : }
2045 0 : } else {
2046 0 : // This node has no observed state for the shard: perhaps it was offline
2047 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2048 0 : // will be prompted to learn the location's state before it makes changes.
2049 0 : shard
2050 0 : .observed
2051 0 : .locations
2052 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2053 0 : }
2054 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2055 0 : // Ordering: pageserver will not accept /location_config requests until it has
2056 0 : // finished processing the response from re-attach. So we can update our in-memory state
2057 0 : // now, and be confident that we are not stamping on the result of some later location config.
2058 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2059 0 : // so we might update observed state here, and then get over-written by some racing
2060 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2061 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2062 0 :
2063 0 : response.tenants.push(ReAttachResponseTenant {
2064 0 : id: *tenant_shard_id,
2065 0 : gen: None,
2066 0 : mode: LocationConfigMode::Secondary,
2067 0 : });
2068 0 :
2069 0 : // We must not update observed, because we have no guarantee that our
2070 0 : // response will be received by the pageserver. This could leave it
2071 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2072 0 : }
2073 : }
2074 :
2075 : // We consider a node Active once we have composed a re-attach response, but we
2076 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2077 : // implicitly synchronizes the LocationConfigs on the node.
2078 : //
2079 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2080 : // but those requests will not be accepted by the node until it has finished processing
2081 : // the re-attach response.
2082 : //
2083 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2084 : // in [`Persistence::re_attach`].
2085 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2086 0 : let reset_scheduling = matches!(
2087 0 : node.get_scheduling(),
2088 : NodeSchedulingPolicy::PauseForRestart
2089 : | NodeSchedulingPolicy::Draining
2090 : | NodeSchedulingPolicy::Filling
2091 : );
2092 :
2093 0 : let mut new_nodes = (**nodes).clone();
2094 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2095 0 : if reset_scheduling {
2096 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2097 0 : }
2098 :
2099 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2100 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2101 0 :
2102 0 : scheduler.node_upsert(node);
2103 0 : let new_nodes = Arc::new(new_nodes);
2104 0 : *nodes = new_nodes;
2105 : } else {
2106 0 : tracing::error!(
2107 0 : "Reattaching node {} was removed while processing the request",
2108 : reattach_req.node_id
2109 : );
2110 : }
2111 0 : }
2112 :
2113 0 : Ok(response)
2114 0 : }
2115 :
2116 0 : pub(crate) async fn validate(
2117 0 : &self,
2118 0 : validate_req: ValidateRequest,
2119 0 : ) -> Result<ValidateResponse, DatabaseError> {
2120 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2121 : // in-memory generation for a shard
2122 0 : let in_memory_result = {
2123 0 : let mut in_memory_result = Vec::new();
2124 0 : let locked = self.inner.read().unwrap();
2125 0 : for req_tenant in validate_req.tenants {
2126 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2127 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
2128 0 : tracing::info!(
2129 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2130 : req_tenant.id,
2131 : req_tenant.gen,
2132 : tenant_shard.generation
2133 : );
2134 :
2135 0 : in_memory_result.push((req_tenant.id, Generation::new(req_tenant.gen), valid));
2136 : } else {
2137 : // This is legal: for example during a shard split the pageserver may still
2138 : // have deletions in its queue from the old pre-split shard, or after deletion
2139 : // of a tenant that was busy with compaction/gc while being deleted.
2140 0 : tracing::info!(
2141 0 : "Refusing deletion validation for missing shard {}",
2142 : req_tenant.id
2143 : );
2144 : }
2145 : }
2146 :
2147 0 : in_memory_result
2148 : };
2149 :
2150 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2151 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2152 0 : let db_generations = self
2153 0 : .persistence
2154 0 : .shard_generations(in_memory_result.iter().filter_map(|i| {
2155 0 : if i.2 {
2156 0 : Some(&i.0)
2157 : } else {
2158 0 : None
2159 : }
2160 0 : }))
2161 0 : .await?;
2162 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2163 0 :
2164 0 : let mut response = ValidateResponse {
2165 0 : tenants: Vec::new(),
2166 0 : };
2167 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2168 0 : let valid = if valid {
2169 0 : let db_generation = db_generations.get(&tenant_shard_id);
2170 0 : db_generation == Some(&Some(validate_generation))
2171 : } else {
2172 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2173 : // this prevents a pageserver from cleaning up an object in S3.
2174 0 : false
2175 : };
2176 :
2177 0 : response.tenants.push(ValidateResponseTenant {
2178 0 : id: tenant_shard_id,
2179 0 : valid,
2180 0 : })
2181 : }
2182 :
2183 0 : Ok(response)
2184 0 : }
2185 :
2186 0 : pub(crate) async fn tenant_create(
2187 0 : &self,
2188 0 : create_req: TenantCreateRequest,
2189 0 : ) -> Result<TenantCreateResponse, ApiError> {
2190 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2191 :
2192 : // Exclude any concurrent attempts to create/access the same tenant ID
2193 0 : let _tenant_lock = trace_exclusive_lock(
2194 0 : &self.tenant_op_locks,
2195 0 : create_req.new_tenant_id.tenant_id,
2196 0 : TenantOperations::Create,
2197 0 : )
2198 0 : .await;
2199 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2200 :
2201 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2202 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2203 : // accept compute notifications while it is in the process of creating. Reconciliation will
2204 : // be retried in the background.
2205 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2206 0 : }
2207 0 : Ok(response)
2208 0 : }
2209 :
2210 0 : pub(crate) async fn do_tenant_create(
2211 0 : &self,
2212 0 : create_req: TenantCreateRequest,
2213 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2214 0 : let placement_policy = create_req
2215 0 : .placement_policy
2216 0 : .clone()
2217 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2218 0 : .unwrap_or(PlacementPolicy::Attached(0));
2219 :
2220 : // This service expects to handle sharding itself: it is an error to try and directly create
2221 : // a particular shard here.
2222 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2223 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2224 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2225 0 : )));
2226 : } else {
2227 0 : create_req.new_tenant_id.tenant_id
2228 0 : };
2229 0 :
2230 0 : tracing::info!(
2231 0 : "Creating tenant {}, shard_count={:?}",
2232 : create_req.new_tenant_id,
2233 : create_req.shard_parameters.count,
2234 : );
2235 :
2236 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2237 0 : .map(|i| TenantShardId {
2238 0 : tenant_id,
2239 0 : shard_number: ShardNumber(i),
2240 0 : shard_count: create_req.shard_parameters.count,
2241 0 : })
2242 0 : .collect::<Vec<_>>();
2243 :
2244 : // If the caller specifies a None generation, it means "start from default". This is different
2245 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2246 : // an incompletely-onboarded tenant.
2247 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2248 0 : tracing::info!(
2249 0 : "tenant_create: secondary mode, generation is_some={}",
2250 0 : create_req.generation.is_some()
2251 : );
2252 0 : create_req.generation.map(Generation::new)
2253 : } else {
2254 0 : tracing::info!(
2255 0 : "tenant_create: not secondary mode, generation is_some={}",
2256 0 : create_req.generation.is_some()
2257 : );
2258 0 : Some(
2259 0 : create_req
2260 0 : .generation
2261 0 : .map(Generation::new)
2262 0 : .unwrap_or(INITIAL_GENERATION),
2263 0 : )
2264 : };
2265 :
2266 0 : let preferred_az_id = {
2267 0 : let locked = self.inner.read().unwrap();
2268 : // Idempotency: take the existing value if the tenant already exists
2269 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2270 0 : shard.preferred_az().cloned()
2271 : } else {
2272 0 : locked.scheduler.get_az_for_new_tenant()
2273 : }
2274 : };
2275 :
2276 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2277 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2278 : // during the creation, rather than risking leaving orphan objects in S3.
2279 0 : let persist_tenant_shards = create_ids
2280 0 : .iter()
2281 0 : .map(|tenant_shard_id| TenantShardPersistence {
2282 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2283 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2284 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2285 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2286 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2287 0 : // The pageserver is not known until scheduling happens: we will set this column when
2288 0 : // incrementing the generation the first time we attach to a pageserver.
2289 0 : generation_pageserver: None,
2290 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2291 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2292 0 : splitting: SplitState::default(),
2293 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2294 0 : .unwrap(),
2295 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2296 0 : })
2297 0 : .collect();
2298 0 :
2299 0 : match self
2300 0 : .persistence
2301 0 : .insert_tenant_shards(persist_tenant_shards)
2302 0 : .await
2303 : {
2304 0 : Ok(_) => {}
2305 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2306 : DatabaseErrorKind::UniqueViolation,
2307 : _,
2308 : ))) => {
2309 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2310 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2311 : // creation's shard count.
2312 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
2313 : }
2314 : // Any other database error is unexpected and a bug.
2315 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2316 : };
2317 :
2318 0 : let mut schedule_context = ScheduleContext::default();
2319 0 : let mut schedule_error = None;
2320 0 : let mut response_shards = Vec::new();
2321 0 : for tenant_shard_id in create_ids {
2322 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2323 :
2324 0 : let outcome = self
2325 0 : .do_initial_shard_scheduling(
2326 0 : tenant_shard_id,
2327 0 : initial_generation,
2328 0 : &create_req.shard_parameters,
2329 0 : create_req.config.clone(),
2330 0 : placement_policy.clone(),
2331 0 : preferred_az_id.as_ref(),
2332 0 : &mut schedule_context,
2333 0 : )
2334 0 : .await;
2335 :
2336 0 : match outcome {
2337 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2338 0 : InitialShardScheduleOutcome::NotScheduled => {}
2339 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2340 0 : schedule_error = Some(err);
2341 0 : }
2342 : }
2343 : }
2344 :
2345 : // If we failed to schedule shards, then they are still created in the controller,
2346 : // but we return an error to the requester to avoid a silent failure when someone
2347 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2348 : // are present in the system. We do this here rather than in the above loop, to
2349 : // avoid situations where we only create a subset of shards in the tenant.
2350 0 : if let Some(e) = schedule_error {
2351 0 : return Err(ApiError::Conflict(format!(
2352 0 : "Failed to schedule shard(s): {e}"
2353 0 : )));
2354 0 : }
2355 0 :
2356 0 : let waiters = {
2357 0 : let mut locked = self.inner.write().unwrap();
2358 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2359 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2360 0 : .tenant_creation_hint(true)
2361 0 : .build();
2362 0 : tenants
2363 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2364 0 : .filter_map(|(_shard_id, shard)| {
2365 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2366 0 : })
2367 0 : .collect::<Vec<_>>()
2368 0 : };
2369 0 :
2370 0 : Ok((
2371 0 : TenantCreateResponse {
2372 0 : shards: response_shards,
2373 0 : },
2374 0 : waiters,
2375 0 : ))
2376 0 : }
2377 :
2378 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2379 : /// case of a new tenant and a pre-existing one.
2380 : #[allow(clippy::too_many_arguments)]
2381 0 : async fn do_initial_shard_scheduling(
2382 0 : &self,
2383 0 : tenant_shard_id: TenantShardId,
2384 0 : initial_generation: Option<Generation>,
2385 0 : shard_params: &ShardParameters,
2386 0 : config: TenantConfig,
2387 0 : placement_policy: PlacementPolicy,
2388 0 : preferred_az_id: Option<&AvailabilityZone>,
2389 0 : schedule_context: &mut ScheduleContext,
2390 0 : ) -> InitialShardScheduleOutcome {
2391 0 : let mut locked = self.inner.write().unwrap();
2392 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2393 :
2394 : use std::collections::btree_map::Entry;
2395 0 : match tenants.entry(tenant_shard_id) {
2396 0 : Entry::Occupied(mut entry) => {
2397 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2398 :
2399 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2400 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2401 0 : }
2402 :
2403 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2404 0 : let generation = entry
2405 0 : .get()
2406 0 : .generation
2407 0 : .expect("Generation is set when in attached mode");
2408 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2409 0 : shard_id: tenant_shard_id,
2410 0 : node_id: *node_id,
2411 0 : generation: generation.into().unwrap(),
2412 0 : })
2413 : } else {
2414 0 : InitialShardScheduleOutcome::NotScheduled
2415 : }
2416 : }
2417 0 : Entry::Vacant(entry) => {
2418 0 : let state = entry.insert(TenantShard::new(
2419 0 : tenant_shard_id,
2420 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2421 0 : placement_policy,
2422 0 : preferred_az_id.cloned(),
2423 0 : ));
2424 0 :
2425 0 : state.generation = initial_generation;
2426 0 : state.config = config;
2427 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2428 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2429 0 : }
2430 :
2431 : // Only include shards in result if we are attaching: the purpose
2432 : // of the response is to tell the caller where the shards are attached.
2433 0 : if let Some(node_id) = state.intent.get_attached() {
2434 0 : let generation = state
2435 0 : .generation
2436 0 : .expect("Generation is set when in attached mode");
2437 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2438 0 : shard_id: tenant_shard_id,
2439 0 : node_id: *node_id,
2440 0 : generation: generation.into().unwrap(),
2441 0 : })
2442 : } else {
2443 0 : InitialShardScheduleOutcome::NotScheduled
2444 : }
2445 : }
2446 : }
2447 0 : }
2448 :
2449 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2450 : /// wait for reconciliation to complete before responding.
2451 0 : async fn await_waiters(
2452 0 : &self,
2453 0 : waiters: Vec<ReconcilerWaiter>,
2454 0 : timeout: Duration,
2455 0 : ) -> Result<(), ReconcileWaitError> {
2456 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2457 0 : for waiter in waiters {
2458 0 : let timeout = deadline.duration_since(Instant::now());
2459 0 : waiter.wait_timeout(timeout).await?;
2460 : }
2461 :
2462 0 : Ok(())
2463 0 : }
2464 :
2465 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2466 : /// in progress
2467 0 : async fn await_waiters_remainder(
2468 0 : &self,
2469 0 : waiters: Vec<ReconcilerWaiter>,
2470 0 : timeout: Duration,
2471 0 : ) -> Vec<ReconcilerWaiter> {
2472 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2473 0 : for waiter in waiters.iter() {
2474 0 : let timeout = deadline.duration_since(Instant::now());
2475 0 : let _ = waiter.wait_timeout(timeout).await;
2476 : }
2477 :
2478 0 : waiters
2479 0 : .into_iter()
2480 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2481 0 : .collect::<Vec<_>>()
2482 0 : }
2483 :
2484 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2485 : /// and transform it into either a tenant creation of a series of shard updates.
2486 : ///
2487 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2488 : /// still be returned.
2489 0 : fn tenant_location_config_prepare(
2490 0 : &self,
2491 0 : tenant_id: TenantId,
2492 0 : req: TenantLocationConfigRequest,
2493 0 : ) -> TenantCreateOrUpdate {
2494 0 : let mut updates = Vec::new();
2495 0 : let mut locked = self.inner.write().unwrap();
2496 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2497 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2498 :
2499 : // Use location config mode as an indicator of policy.
2500 0 : let placement_policy = match req.config.mode {
2501 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2502 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2503 : LocationConfigMode::AttachedMulti
2504 : | LocationConfigMode::AttachedSingle
2505 : | LocationConfigMode::AttachedStale => {
2506 0 : if nodes.len() > 1 {
2507 0 : PlacementPolicy::Attached(1)
2508 : } else {
2509 : // Convenience for dev/test: if we just have one pageserver, import
2510 : // tenants into non-HA mode so that scheduling will succeed.
2511 0 : PlacementPolicy::Attached(0)
2512 : }
2513 : }
2514 : };
2515 :
2516 : // Ordinarily we do not update scheduling policy, but when making major changes
2517 : // like detaching or demoting to secondary-only, we need to force the scheduling
2518 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2519 0 : let scheduling_policy = match req.config.mode {
2520 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2521 : // Special case: when making major changes like detaching or demoting to secondary-only,
2522 : // we need to force the scheduling mode to Active, or nothing will happen.
2523 0 : Some(ShardSchedulingPolicy::Active)
2524 : }
2525 : LocationConfigMode::AttachedMulti
2526 : | LocationConfigMode::AttachedSingle
2527 : | LocationConfigMode::AttachedStale => {
2528 : // While attached, continue to respect whatever the existing scheduling mode is.
2529 0 : None
2530 : }
2531 : };
2532 :
2533 0 : let mut create = true;
2534 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2535 : // Saw an existing shard: this is not a creation
2536 0 : create = false;
2537 :
2538 : // Shards may have initially been created by a Secondary request, where we
2539 : // would have left generation as None.
2540 : //
2541 : // We only update generation the first time we see an attached-mode request,
2542 : // and if there is no existing generation set. The caller is responsible for
2543 : // ensuring that no non-storage-controller pageserver ever uses a higher
2544 : // generation than they passed in here.
2545 : use LocationConfigMode::*;
2546 0 : let set_generation = match req.config.mode {
2547 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2548 0 : req.config.generation.map(Generation::new)
2549 : }
2550 0 : _ => None,
2551 : };
2552 :
2553 0 : updates.push(ShardUpdate {
2554 0 : tenant_shard_id: *shard_id,
2555 0 : placement_policy: placement_policy.clone(),
2556 0 : tenant_config: req.config.tenant_conf.clone(),
2557 0 : generation: set_generation,
2558 0 : scheduling_policy,
2559 0 : });
2560 : }
2561 :
2562 0 : if create {
2563 : use LocationConfigMode::*;
2564 0 : let generation = match req.config.mode {
2565 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2566 : // If a caller provided a generation in a non-attached request, ignore it
2567 : // and leave our generation as None: this enables a subsequent update to set
2568 : // the generation when setting an attached mode for the first time.
2569 0 : _ => None,
2570 : };
2571 :
2572 0 : TenantCreateOrUpdate::Create(
2573 0 : // Synthesize a creation request
2574 0 : TenantCreateRequest {
2575 0 : new_tenant_id: tenant_shard_id,
2576 0 : generation,
2577 0 : shard_parameters: ShardParameters {
2578 0 : count: tenant_shard_id.shard_count,
2579 0 : // We only import un-sharded or single-sharded tenants, so stripe
2580 0 : // size can be made up arbitrarily here.
2581 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2582 0 : },
2583 0 : placement_policy: Some(placement_policy),
2584 0 : config: req.config.tenant_conf,
2585 0 : },
2586 0 : )
2587 : } else {
2588 0 : assert!(!updates.is_empty());
2589 0 : TenantCreateOrUpdate::Update(updates)
2590 : }
2591 0 : }
2592 :
2593 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2594 : /// the tenant is present in memory. If not, load it from the database. If it is found
2595 : /// in neither location, return a NotFound error.
2596 : ///
2597 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2598 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2599 0 : async fn maybe_load_tenant(
2600 0 : &self,
2601 0 : tenant_id: TenantId,
2602 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2603 0 : ) -> Result<(), ApiError> {
2604 : // Check if the tenant is present in memory, and select an AZ to use when loading
2605 : // if we will load it.
2606 0 : let load_in_az = {
2607 0 : let locked = self.inner.read().unwrap();
2608 0 : let existing = locked
2609 0 : .tenants
2610 0 : .range(TenantShardId::tenant_range(tenant_id))
2611 0 : .next();
2612 0 :
2613 0 : // If the tenant is not present in memory, we expect to load it from database,
2614 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2615 0 : if existing.is_none() {
2616 0 : locked
2617 0 : .scheduler
2618 0 : .get_az_for_new_tenant()
2619 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2620 0 : "No AZ with nodes found to load tenant"
2621 0 : )))?
2622 : } else {
2623 : // We already have this tenant in memory
2624 0 : return Ok(());
2625 : }
2626 : };
2627 :
2628 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2629 0 : if tenant_shards.is_empty() {
2630 0 : return Err(ApiError::NotFound(
2631 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2632 0 : ));
2633 0 : }
2634 0 :
2635 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2636 0 : self.persistence
2637 0 : .set_tenant_shard_preferred_azs(
2638 0 : tenant_shards
2639 0 : .iter()
2640 0 : .map(|t| {
2641 0 : (
2642 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2643 0 : Some(load_in_az.clone()),
2644 0 : )
2645 0 : })
2646 0 : .collect(),
2647 0 : )
2648 0 : .await?;
2649 :
2650 0 : let mut locked = self.inner.write().unwrap();
2651 0 : tracing::info!(
2652 0 : "Loaded {} shards for tenant {}",
2653 0 : tenant_shards.len(),
2654 : tenant_id
2655 : );
2656 :
2657 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2658 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2659 0 : let shard =
2660 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2661 0 :
2662 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2663 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2664 0 : if shard.policy != PlacementPolicy::Detached {
2665 0 : tracing::error!(
2666 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2667 : shard.tenant_shard_id,
2668 : shard.policy
2669 : );
2670 0 : }
2671 :
2672 0 : (shard.tenant_shard_id, shard)
2673 0 : }));
2674 0 :
2675 0 : Ok(())
2676 0 : }
2677 :
2678 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2679 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2680 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2681 : ///
2682 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2683 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2684 0 : fn maybe_drop_tenant(
2685 0 : &self,
2686 0 : tenant_id: TenantId,
2687 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2688 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2689 0 : ) {
2690 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2691 0 : if tenant_shards.all(|(_id, shard)| {
2692 0 : shard.policy == PlacementPolicy::Detached
2693 0 : && shard.reconciler.is_none()
2694 0 : && shard.observed.is_empty()
2695 0 : }) {
2696 0 : let keys = locked
2697 0 : .tenants
2698 0 : .range(TenantShardId::tenant_range(tenant_id))
2699 0 : .map(|(id, _)| id)
2700 0 : .copied()
2701 0 : .collect::<Vec<_>>();
2702 0 : for key in keys {
2703 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2704 0 : locked.tenants.remove(&key);
2705 : }
2706 0 : }
2707 0 : }
2708 :
2709 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2710 : /// directly with pageservers into this service.
2711 : ///
2712 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2713 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2714 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2715 : /// tenant's source of generation numbers.
2716 : ///
2717 : /// The mode in this request coarse-grained control of tenants:
2718 : /// - Call with mode Attached* to upsert the tenant.
2719 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2720 : /// to set an existing tenant to PolicyMode::Secondary
2721 : /// - Call with mode Detached to switch to PolicyMode::Detached
2722 0 : pub(crate) async fn tenant_location_config(
2723 0 : &self,
2724 0 : tenant_shard_id: TenantShardId,
2725 0 : req: TenantLocationConfigRequest,
2726 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2727 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2728 0 : let _tenant_lock = trace_exclusive_lock(
2729 0 : &self.tenant_op_locks,
2730 0 : tenant_shard_id.tenant_id,
2731 0 : TenantOperations::LocationConfig,
2732 0 : )
2733 0 : .await;
2734 :
2735 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2736 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2737 0 : "This API is for importing single-sharded or unsharded tenants"
2738 0 : )));
2739 : } else {
2740 0 : tenant_shard_id.tenant_id
2741 0 : };
2742 0 :
2743 0 : // In case we are waking up a Detached tenant
2744 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2745 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2746 0 : // This is a creation or an update
2747 0 : }
2748 0 : Err(e) => {
2749 0 : return Err(e);
2750 : }
2751 : };
2752 :
2753 : // First check if this is a creation or an update
2754 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2755 0 :
2756 0 : let mut result = TenantLocationConfigResponse {
2757 0 : shards: Vec::new(),
2758 0 : stripe_size: None,
2759 0 : };
2760 0 : let waiters = match create_or_update {
2761 0 : TenantCreateOrUpdate::Create(create_req) => {
2762 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2763 0 : result.shards = create_resp
2764 0 : .shards
2765 0 : .into_iter()
2766 0 : .map(|s| TenantShardLocation {
2767 0 : node_id: s.node_id,
2768 0 : shard_id: s.shard_id,
2769 0 : })
2770 0 : .collect();
2771 0 : waiters
2772 : }
2773 0 : TenantCreateOrUpdate::Update(updates) => {
2774 0 : // Persist updates
2775 0 : // Ordering: write to the database before applying changes in-memory, so that
2776 0 : // we will not appear time-travel backwards on a restart.
2777 0 :
2778 0 : let mut schedule_context = ScheduleContext::default();
2779 : for ShardUpdate {
2780 0 : tenant_shard_id,
2781 0 : placement_policy,
2782 0 : tenant_config,
2783 0 : generation,
2784 0 : scheduling_policy,
2785 0 : } in &updates
2786 : {
2787 0 : self.persistence
2788 0 : .update_tenant_shard(
2789 0 : TenantFilter::Shard(*tenant_shard_id),
2790 0 : Some(placement_policy.clone()),
2791 0 : Some(tenant_config.clone()),
2792 0 : *generation,
2793 0 : *scheduling_policy,
2794 0 : )
2795 0 : .await?;
2796 : }
2797 :
2798 : // Apply updates in-memory
2799 0 : let mut waiters = Vec::new();
2800 0 : {
2801 0 : let mut locked = self.inner.write().unwrap();
2802 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2803 :
2804 : for ShardUpdate {
2805 0 : tenant_shard_id,
2806 0 : placement_policy,
2807 0 : tenant_config,
2808 0 : generation: update_generation,
2809 0 : scheduling_policy,
2810 0 : } in updates
2811 : {
2812 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2813 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2814 0 : continue;
2815 : };
2816 :
2817 : // Update stripe size
2818 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2819 0 : result.stripe_size = Some(shard.shard.stripe_size);
2820 0 : }
2821 :
2822 0 : shard.policy = placement_policy;
2823 0 : shard.config = tenant_config;
2824 0 : if let Some(generation) = update_generation {
2825 0 : shard.generation = Some(generation);
2826 0 : }
2827 :
2828 0 : if let Some(scheduling_policy) = scheduling_policy {
2829 0 : shard.set_scheduling_policy(scheduling_policy);
2830 0 : }
2831 :
2832 0 : shard.schedule(scheduler, &mut schedule_context)?;
2833 :
2834 0 : let maybe_waiter =
2835 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
2836 0 : if let Some(waiter) = maybe_waiter {
2837 0 : waiters.push(waiter);
2838 0 : }
2839 :
2840 0 : if let Some(node_id) = shard.intent.get_attached() {
2841 0 : result.shards.push(TenantShardLocation {
2842 0 : shard_id: tenant_shard_id,
2843 0 : node_id: *node_id,
2844 0 : })
2845 0 : }
2846 : }
2847 : }
2848 0 : waiters
2849 : }
2850 : };
2851 :
2852 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2853 : // Do not treat a reconcile error as fatal: we have already applied any requested
2854 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2855 : // compute notification API. In these cases, it is important that we do not
2856 : // cause the cloud control plane to retry forever on this API.
2857 0 : tracing::warn!(
2858 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2859 : );
2860 0 : }
2861 :
2862 : // Logging the full result is useful because it lets us cross-check what the cloud control
2863 : // plane's tenant_shards table should contain.
2864 0 : tracing::info!("Complete, returning {result:?}");
2865 :
2866 0 : Ok(result)
2867 0 : }
2868 :
2869 0 : pub(crate) async fn tenant_config_patch(
2870 0 : &self,
2871 0 : req: TenantConfigPatchRequest,
2872 0 : ) -> Result<(), ApiError> {
2873 0 : let _tenant_lock = trace_exclusive_lock(
2874 0 : &self.tenant_op_locks,
2875 0 : req.tenant_id,
2876 0 : TenantOperations::ConfigPatch,
2877 0 : )
2878 0 : .await;
2879 :
2880 0 : let tenant_id = req.tenant_id;
2881 0 : let patch = req.config;
2882 0 :
2883 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
2884 :
2885 0 : let base = {
2886 0 : let locked = self.inner.read().unwrap();
2887 0 : let shards = locked
2888 0 : .tenants
2889 0 : .range(TenantShardId::tenant_range(req.tenant_id));
2890 0 :
2891 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
2892 :
2893 0 : let first = match configs.peek() {
2894 0 : Some(first) => (*first).clone(),
2895 : None => {
2896 0 : return Err(ApiError::NotFound(
2897 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
2898 0 : ));
2899 : }
2900 : };
2901 :
2902 0 : if !configs.all_equal() {
2903 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
2904 : // This can't happen because we atomically update the database records
2905 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
2906 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2907 0 : "Tenant configs for {} are mismatched",
2908 0 : req.tenant_id
2909 0 : )));
2910 0 : }
2911 0 :
2912 0 : first
2913 0 : };
2914 0 :
2915 0 : let updated_config = base.apply_patch(patch);
2916 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
2917 0 : .await
2918 0 : }
2919 :
2920 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2921 : // We require an exclusive lock, because we are updating persistent and in-memory state
2922 0 : let _tenant_lock = trace_exclusive_lock(
2923 0 : &self.tenant_op_locks,
2924 0 : req.tenant_id,
2925 0 : TenantOperations::ConfigSet,
2926 0 : )
2927 0 : .await;
2928 :
2929 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
2930 :
2931 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
2932 0 : .await
2933 0 : }
2934 :
2935 0 : async fn set_tenant_config_and_reconcile(
2936 0 : &self,
2937 0 : tenant_id: TenantId,
2938 0 : config: TenantConfig,
2939 0 : ) -> Result<(), ApiError> {
2940 0 : self.persistence
2941 0 : .update_tenant_shard(
2942 0 : TenantFilter::Tenant(tenant_id),
2943 0 : None,
2944 0 : Some(config.clone()),
2945 0 : None,
2946 0 : None,
2947 0 : )
2948 0 : .await?;
2949 :
2950 0 : let waiters = {
2951 0 : let mut waiters = Vec::new();
2952 0 : let mut locked = self.inner.write().unwrap();
2953 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2954 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2955 0 : shard.config = config.clone();
2956 0 : if let Some(waiter) =
2957 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
2958 0 : {
2959 0 : waiters.push(waiter);
2960 0 : }
2961 : }
2962 0 : waiters
2963 : };
2964 :
2965 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2966 : // Treat this as success because we have stored the configuration. If e.g.
2967 : // a node was unavailable at this time, it should not stop us accepting a
2968 : // configuration change.
2969 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2970 0 : }
2971 :
2972 0 : Ok(())
2973 0 : }
2974 :
2975 0 : pub(crate) fn tenant_config_get(
2976 0 : &self,
2977 0 : tenant_id: TenantId,
2978 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2979 0 : let config = {
2980 0 : let locked = self.inner.read().unwrap();
2981 0 :
2982 0 : match locked
2983 0 : .tenants
2984 0 : .range(TenantShardId::tenant_range(tenant_id))
2985 0 : .next()
2986 : {
2987 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2988 : None => {
2989 0 : return Err(ApiError::NotFound(
2990 0 : anyhow::anyhow!("Tenant not found").into(),
2991 0 : ))
2992 : }
2993 : }
2994 : };
2995 :
2996 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2997 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2998 : // and `effective_config` in the response is meaningless, but we retain that syntax
2999 : // in order to remain compatible with the pageserver API.
3000 :
3001 0 : let response = HashMap::from([
3002 : (
3003 : "tenant_specific_overrides",
3004 0 : serde_json::to_value(&config)
3005 0 : .context("serializing tenant specific overrides")
3006 0 : .map_err(ApiError::InternalServerError)?,
3007 : ),
3008 : (
3009 0 : "effective_config",
3010 0 : serde_json::to_value(&config)
3011 0 : .context("serializing effective config")
3012 0 : .map_err(ApiError::InternalServerError)?,
3013 : ),
3014 : ]);
3015 :
3016 0 : Ok(response)
3017 0 : }
3018 :
3019 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3020 0 : &self,
3021 0 : time_travel_req: &TenantTimeTravelRequest,
3022 0 : tenant_id: TenantId,
3023 0 : timestamp: Cow<'_, str>,
3024 0 : done_if_after: Cow<'_, str>,
3025 0 : ) -> Result<(), ApiError> {
3026 0 : let _tenant_lock = trace_exclusive_lock(
3027 0 : &self.tenant_op_locks,
3028 0 : tenant_id,
3029 0 : TenantOperations::TimeTravelRemoteStorage,
3030 0 : )
3031 0 : .await;
3032 :
3033 0 : let node = {
3034 0 : let mut locked = self.inner.write().unwrap();
3035 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3036 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3037 : // but only at the start of the process, so it's really just to prevent operator
3038 : // mistakes.
3039 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3040 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3041 : {
3042 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3043 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3044 0 : )));
3045 0 : }
3046 0 : let maybe_attached = shard
3047 0 : .observed
3048 0 : .locations
3049 0 : .iter()
3050 0 : .filter_map(|(node_id, observed_location)| {
3051 0 : observed_location
3052 0 : .conf
3053 0 : .as_ref()
3054 0 : .map(|loc| (node_id, observed_location, loc.mode))
3055 0 : })
3056 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3057 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3058 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
3059 0 : }
3060 : }
3061 0 : let scheduler = &mut locked.scheduler;
3062 : // Right now we only perform the operation on a single node without parallelization
3063 : // TODO fan out the operation to multiple nodes for better performance
3064 0 : let node_id = scheduler.any_available_node()?;
3065 0 : let node = locked
3066 0 : .nodes
3067 0 : .get(&node_id)
3068 0 : .expect("Pageservers may not be deleted while lock is active");
3069 0 : node.clone()
3070 0 : };
3071 0 :
3072 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3073 0 : let mut counts = time_travel_req
3074 0 : .shard_counts
3075 0 : .iter()
3076 0 : .copied()
3077 0 : .collect::<HashSet<_>>()
3078 0 : .into_iter()
3079 0 : .collect::<Vec<_>>();
3080 0 : counts.sort_unstable();
3081 :
3082 0 : for count in counts {
3083 0 : let shard_ids = (0..count.count())
3084 0 : .map(|i| TenantShardId {
3085 0 : tenant_id,
3086 0 : shard_number: ShardNumber(i),
3087 0 : shard_count: count,
3088 0 : })
3089 0 : .collect::<Vec<_>>();
3090 0 : for tenant_shard_id in shard_ids {
3091 0 : let client = PageserverClient::new(
3092 0 : node.get_id(),
3093 0 : node.base_url(),
3094 0 : self.config.jwt_token.as_deref(),
3095 0 : );
3096 0 :
3097 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3098 :
3099 0 : client
3100 0 : .tenant_time_travel_remote_storage(
3101 0 : tenant_shard_id,
3102 0 : ×tamp,
3103 0 : &done_if_after,
3104 0 : )
3105 0 : .await
3106 0 : .map_err(|e| {
3107 0 : ApiError::InternalServerError(anyhow::anyhow!(
3108 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3109 0 : node
3110 0 : ))
3111 0 : })?;
3112 : }
3113 : }
3114 0 : Ok(())
3115 0 : }
3116 :
3117 0 : pub(crate) async fn tenant_secondary_download(
3118 0 : &self,
3119 0 : tenant_id: TenantId,
3120 0 : wait: Option<Duration>,
3121 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3122 0 : let _tenant_lock = trace_shared_lock(
3123 0 : &self.tenant_op_locks,
3124 0 : tenant_id,
3125 0 : TenantOperations::SecondaryDownload,
3126 0 : )
3127 0 : .await;
3128 :
3129 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3130 0 : let targets = {
3131 0 : let locked = self.inner.read().unwrap();
3132 0 : let mut targets = Vec::new();
3133 :
3134 0 : for (tenant_shard_id, shard) in
3135 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3136 : {
3137 0 : for node_id in shard.intent.get_secondary() {
3138 0 : let node = locked
3139 0 : .nodes
3140 0 : .get(node_id)
3141 0 : .expect("Pageservers may not be deleted while referenced");
3142 0 :
3143 0 : targets.push((*tenant_shard_id, node.clone()));
3144 0 : }
3145 : }
3146 0 : targets
3147 0 : };
3148 0 :
3149 0 : // Issue concurrent requests to all shards' locations
3150 0 : let mut futs = FuturesUnordered::new();
3151 0 : for (tenant_shard_id, node) in targets {
3152 0 : let client = PageserverClient::new(
3153 0 : node.get_id(),
3154 0 : node.base_url(),
3155 0 : self.config.jwt_token.as_deref(),
3156 0 : );
3157 0 : futs.push(async move {
3158 0 : let result = client
3159 0 : .tenant_secondary_download(tenant_shard_id, wait)
3160 0 : .await;
3161 0 : (result, node, tenant_shard_id)
3162 0 : })
3163 : }
3164 :
3165 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3166 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3167 : // well as more general cases like 503s, 500s, or timeouts.
3168 0 : let mut aggregate_progress = SecondaryProgress::default();
3169 0 : let mut aggregate_status: Option<StatusCode> = None;
3170 0 : let mut error: Option<mgmt_api::Error> = None;
3171 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3172 0 : match result {
3173 0 : Err(e) => {
3174 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3175 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3176 0 : // than they had hoped for.
3177 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3178 0 : error = Some(e)
3179 : }
3180 0 : Ok((status_code, progress)) => {
3181 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3182 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3183 0 : aggregate_progress.layers_total += progress.layers_total;
3184 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3185 0 : aggregate_progress.bytes_total += progress.bytes_total;
3186 0 : aggregate_progress.heatmap_mtime =
3187 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3188 0 : aggregate_status = match aggregate_status {
3189 0 : None => Some(status_code),
3190 0 : Some(StatusCode::OK) => Some(status_code),
3191 0 : Some(cur) => {
3192 0 : // Other status codes (e.g. 202) -- do not overwrite.
3193 0 : Some(cur)
3194 : }
3195 : };
3196 : }
3197 : }
3198 : }
3199 :
3200 : // If any of the shards return 202, indicate our result as 202.
3201 0 : match aggregate_status {
3202 : None => {
3203 0 : match error {
3204 0 : Some(e) => {
3205 0 : // No successes, and an error: surface it
3206 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3207 : }
3208 : None => {
3209 : // No shards found
3210 0 : Err(ApiError::NotFound(
3211 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3212 0 : ))
3213 : }
3214 : }
3215 : }
3216 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3217 : }
3218 0 : }
3219 :
3220 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
3221 0 : let _tenant_lock =
3222 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3223 :
3224 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3225 :
3226 : // Detach all shards. This also deletes local pageserver shard data.
3227 0 : let (detach_waiters, node) = {
3228 0 : let mut detach_waiters = Vec::new();
3229 0 : let mut locked = self.inner.write().unwrap();
3230 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3231 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3232 : // Update the tenant's intent to remove all attachments
3233 0 : shard.policy = PlacementPolicy::Detached;
3234 0 : shard
3235 0 : .schedule(scheduler, &mut ScheduleContext::default())
3236 0 : .expect("De-scheduling is infallible");
3237 0 : debug_assert!(shard.intent.get_attached().is_none());
3238 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3239 :
3240 0 : if let Some(waiter) =
3241 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3242 0 : {
3243 0 : detach_waiters.push(waiter);
3244 0 : }
3245 : }
3246 :
3247 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3248 : // was attached, just has to be able to see the S3 content)
3249 0 : let node_id = scheduler.any_available_node()?;
3250 0 : let node = nodes
3251 0 : .get(&node_id)
3252 0 : .expect("Pageservers may not be deleted while lock is active");
3253 0 : (detach_waiters, node.clone())
3254 0 : };
3255 0 :
3256 0 : // This reconcile wait can fail in a few ways:
3257 0 : // A there is a very long queue for the reconciler semaphore
3258 0 : // B some pageserver is failing to handle a detach promptly
3259 0 : // C some pageserver goes offline right at the moment we send it a request.
3260 0 : //
3261 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3262 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3263 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3264 0 : // deleting the underlying data).
3265 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3266 0 : .await?;
3267 :
3268 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3269 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3270 : // the tenant ID prefix, including all shards (even possibly stale ones).
3271 0 : match node
3272 0 : .with_client_retries(
3273 0 : |client| async move {
3274 0 : client
3275 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3276 0 : .await
3277 0 : },
3278 0 : &self.config.jwt_token,
3279 0 : 1,
3280 0 : 3,
3281 0 : RECONCILE_TIMEOUT,
3282 0 : &self.cancel,
3283 0 : )
3284 0 : .await
3285 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3286 : {
3287 0 : Ok(_) => {}
3288 : Err(mgmt_api::Error::Cancelled) => {
3289 0 : return Err(ApiError::ShuttingDown);
3290 : }
3291 0 : Err(e) => {
3292 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3293 0 : // at large is unavailable.
3294 0 : tracing::error!("Error deleting via node {node}: {e}");
3295 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3296 : }
3297 : }
3298 :
3299 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3300 : // our in-memory state and database state.
3301 :
3302 : // Ordering: we delete persistent state first: if we then
3303 : // crash, we will drop the in-memory state.
3304 :
3305 : // Drop persistent state.
3306 0 : self.persistence.delete_tenant(tenant_id).await?;
3307 :
3308 : // Drop in-memory state
3309 : {
3310 0 : let mut locked = self.inner.write().unwrap();
3311 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3312 :
3313 : // Dereference Scheduler from shards before dropping them
3314 0 : for (_tenant_shard_id, shard) in
3315 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3316 0 : {
3317 0 : shard.intent.clear(scheduler);
3318 0 : }
3319 :
3320 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3321 0 : tracing::info!(
3322 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3323 0 : locked.tenants.len()
3324 : );
3325 : };
3326 :
3327 : // Success is represented as 404, to imitate the existing pageserver deletion API
3328 0 : Ok(StatusCode::NOT_FOUND)
3329 0 : }
3330 :
3331 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3332 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3333 : /// the tenant's policies (configuration) within the storage controller
3334 0 : pub(crate) async fn tenant_update_policy(
3335 0 : &self,
3336 0 : tenant_id: TenantId,
3337 0 : req: TenantPolicyRequest,
3338 0 : ) -> Result<(), ApiError> {
3339 : // We require an exclusive lock, because we are updating persistent and in-memory state
3340 0 : let _tenant_lock = trace_exclusive_lock(
3341 0 : &self.tenant_op_locks,
3342 0 : tenant_id,
3343 0 : TenantOperations::UpdatePolicy,
3344 0 : )
3345 0 : .await;
3346 :
3347 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3348 :
3349 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3350 :
3351 : let TenantPolicyRequest {
3352 0 : placement,
3353 0 : mut scheduling,
3354 0 : } = req;
3355 :
3356 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3357 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3358 : // this to take effect.
3359 0 : if scheduling.is_none() {
3360 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3361 0 : }
3362 0 : }
3363 :
3364 0 : self.persistence
3365 0 : .update_tenant_shard(
3366 0 : TenantFilter::Tenant(tenant_id),
3367 0 : placement.clone(),
3368 0 : None,
3369 0 : None,
3370 0 : scheduling,
3371 0 : )
3372 0 : .await?;
3373 :
3374 0 : let mut schedule_context = ScheduleContext::default();
3375 0 : let mut locked = self.inner.write().unwrap();
3376 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3377 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3378 0 : if let Some(placement) = &placement {
3379 0 : shard.policy = placement.clone();
3380 0 :
3381 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3382 0 : "Updated placement policy to {placement:?}");
3383 0 : }
3384 :
3385 0 : if let Some(scheduling) = &scheduling {
3386 0 : shard.set_scheduling_policy(*scheduling);
3387 0 :
3388 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3389 0 : "Updated scheduling policy to {scheduling:?}");
3390 0 : }
3391 :
3392 : // In case scheduling is being switched back on, try it now.
3393 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3394 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3395 : }
3396 :
3397 0 : Ok(())
3398 0 : }
3399 :
3400 0 : pub(crate) async fn tenant_timeline_create(
3401 0 : &self,
3402 0 : tenant_id: TenantId,
3403 0 : mut create_req: TimelineCreateRequest,
3404 0 : ) -> Result<TimelineInfo, ApiError> {
3405 0 : tracing::info!(
3406 0 : "Creating timeline {}/{}",
3407 : tenant_id,
3408 : create_req.new_timeline_id,
3409 : );
3410 :
3411 0 : let _tenant_lock = trace_shared_lock(
3412 0 : &self.tenant_op_locks,
3413 0 : tenant_id,
3414 0 : TenantOperations::TimelineCreate,
3415 0 : )
3416 0 : .await;
3417 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3418 :
3419 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3420 0 : if targets.0.is_empty() {
3421 0 : return Err(ApiError::NotFound(
3422 0 : anyhow::anyhow!("Tenant not found").into(),
3423 0 : ));
3424 0 : };
3425 0 :
3426 0 : let (shard_zero_tid, shard_zero_locations) =
3427 0 : targets.0.pop_first().expect("Must have at least one shard");
3428 0 : assert!(shard_zero_tid.is_shard_zero());
3429 :
3430 0 : async fn create_one(
3431 0 : tenant_shard_id: TenantShardId,
3432 0 : locations: ShardMutationLocations,
3433 0 : jwt: Option<String>,
3434 0 : create_req: TimelineCreateRequest,
3435 0 : ) -> Result<TimelineInfo, ApiError> {
3436 0 : let latest = locations.latest.node;
3437 0 :
3438 0 : tracing::info!(
3439 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3440 : tenant_shard_id,
3441 : create_req.new_timeline_id,
3442 : locations.latest.generation
3443 : );
3444 :
3445 0 : let client =
3446 0 : PageserverClient::new(latest.get_id(), latest.base_url(), jwt.as_deref());
3447 :
3448 0 : let timeline_info = client
3449 0 : .timeline_create(tenant_shard_id, &create_req)
3450 0 : .await
3451 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3452 :
3453 : // We propagate timeline creations to all attached locations such that a compute
3454 : // for the new timeline is able to start regardless of the current state of the
3455 : // tenant shard reconciliation.
3456 0 : for location in locations.other {
3457 0 : tracing::info!(
3458 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3459 : tenant_shard_id,
3460 : create_req.new_timeline_id,
3461 : location.node,
3462 : location.generation
3463 : );
3464 :
3465 0 : let client = PageserverClient::new(
3466 0 : location.node.get_id(),
3467 0 : location.node.base_url(),
3468 0 : jwt.as_deref(),
3469 0 : );
3470 :
3471 0 : let res = client
3472 0 : .timeline_create(tenant_shard_id, &create_req)
3473 0 : .await;
3474 :
3475 0 : if let Err(e) = res {
3476 0 : match e {
3477 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3478 0 : // Tenant might have been detached from the stale location,
3479 0 : // so ignore 404s.
3480 0 : },
3481 : _ => {
3482 0 : return Err(passthrough_api_error(&location.node, e));
3483 : }
3484 : }
3485 0 : }
3486 : }
3487 :
3488 0 : Ok(timeline_info)
3489 0 : }
3490 :
3491 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3492 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3493 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3494 0 : let timeline_info = create_one(
3495 0 : shard_zero_tid,
3496 0 : shard_zero_locations,
3497 0 : self.config.jwt_token.clone(),
3498 0 : create_req.clone(),
3499 0 : )
3500 0 : .await?;
3501 :
3502 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3503 0 : match &mut create_req.mode {
3504 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3505 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3506 0 : },
3507 0 : _ => {}
3508 : }
3509 :
3510 : // Create timeline on remaining shards with number >0
3511 0 : if !targets.0.is_empty() {
3512 : // If we had multiple shards, issue requests for the remainder now.
3513 0 : let jwt = &self.config.jwt_token;
3514 0 : self.tenant_for_shards(
3515 0 : targets
3516 0 : .0
3517 0 : .iter()
3518 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3519 0 : .collect(),
3520 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3521 0 : let create_req = create_req.clone();
3522 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3523 0 : Box::pin(create_one(
3524 0 : tenant_shard_id,
3525 0 : mutation_locations,
3526 0 : jwt.clone(),
3527 0 : create_req,
3528 0 : ))
3529 0 : },
3530 0 : )
3531 0 : .await?;
3532 0 : }
3533 :
3534 0 : Ok(timeline_info)
3535 0 : })
3536 0 : .await?
3537 0 : }
3538 :
3539 0 : pub(crate) async fn tenant_timeline_archival_config(
3540 0 : &self,
3541 0 : tenant_id: TenantId,
3542 0 : timeline_id: TimelineId,
3543 0 : req: TimelineArchivalConfigRequest,
3544 0 : ) -> Result<(), ApiError> {
3545 0 : tracing::info!(
3546 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3547 : req.state
3548 : );
3549 :
3550 0 : let _tenant_lock = trace_shared_lock(
3551 0 : &self.tenant_op_locks,
3552 0 : tenant_id,
3553 0 : TenantOperations::TimelineArchivalConfig,
3554 0 : )
3555 0 : .await;
3556 :
3557 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3558 0 : if targets.0.is_empty() {
3559 0 : return Err(ApiError::NotFound(
3560 0 : anyhow::anyhow!("Tenant not found").into(),
3561 0 : ));
3562 0 : }
3563 0 : async fn config_one(
3564 0 : tenant_shard_id: TenantShardId,
3565 0 : timeline_id: TimelineId,
3566 0 : node: Node,
3567 0 : jwt: Option<String>,
3568 0 : req: TimelineArchivalConfigRequest,
3569 0 : ) -> Result<(), ApiError> {
3570 0 : tracing::info!(
3571 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3572 : );
3573 :
3574 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3575 0 :
3576 0 : client
3577 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3578 0 : .await
3579 0 : .map_err(|e| match e {
3580 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3581 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3582 : }
3583 0 : _ => passthrough_api_error(&node, e),
3584 0 : })
3585 0 : }
3586 :
3587 : // no shard needs to go first/last; the operation should be idempotent
3588 : // TODO: it would be great to ensure that all shards return the same error
3589 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3590 0 : let results = self
3591 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3592 0 : futures::FutureExt::boxed(config_one(
3593 0 : tenant_shard_id,
3594 0 : timeline_id,
3595 0 : node,
3596 0 : self.config.jwt_token.clone(),
3597 0 : req.clone(),
3598 0 : ))
3599 0 : })
3600 0 : .await?;
3601 0 : assert!(!results.is_empty(), "must have at least one result");
3602 :
3603 0 : Ok(())
3604 0 : }).await?
3605 0 : }
3606 :
3607 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3608 0 : &self,
3609 0 : tenant_id: TenantId,
3610 0 : timeline_id: TimelineId,
3611 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3612 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3613 :
3614 0 : let _tenant_lock = trace_shared_lock(
3615 0 : &self.tenant_op_locks,
3616 0 : tenant_id,
3617 0 : TenantOperations::TimelineDetachAncestor,
3618 0 : )
3619 0 : .await;
3620 :
3621 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3622 0 : if targets.0.is_empty() {
3623 0 : return Err(ApiError::NotFound(
3624 0 : anyhow::anyhow!("Tenant not found").into(),
3625 0 : ));
3626 0 : }
3627 :
3628 0 : async fn detach_one(
3629 0 : tenant_shard_id: TenantShardId,
3630 0 : timeline_id: TimelineId,
3631 0 : node: Node,
3632 0 : jwt: Option<String>,
3633 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3634 0 : tracing::info!(
3635 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3636 : );
3637 :
3638 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3639 0 :
3640 0 : client
3641 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3642 0 : .await
3643 0 : .map_err(|e| {
3644 : use mgmt_api::Error;
3645 :
3646 0 : match e {
3647 : // no ancestor (ever)
3648 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3649 0 : "{node}: {}",
3650 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3651 0 : )),
3652 : // too many ancestors
3653 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3654 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3655 : }
3656 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3657 0 : // avoid turning these into conflicts to remain compatible with
3658 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3659 0 : // detach
3660 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3661 : }
3662 : // rest can be mapped as usual
3663 0 : other => passthrough_api_error(&node, other),
3664 : }
3665 0 : })
3666 0 : .map(|res| (tenant_shard_id.shard_number, res))
3667 0 : }
3668 :
3669 : // no shard needs to go first/last; the operation should be idempotent
3670 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3671 0 : let mut results = self
3672 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3673 0 : futures::FutureExt::boxed(detach_one(
3674 0 : tenant_shard_id,
3675 0 : timeline_id,
3676 0 : node,
3677 0 : self.config.jwt_token.clone(),
3678 0 : ))
3679 0 : })
3680 0 : .await?;
3681 :
3682 0 : let any = results.pop().expect("we must have at least one response");
3683 0 :
3684 0 : let mismatching = results
3685 0 : .iter()
3686 0 : .filter(|(_, res)| res != &any.1)
3687 0 : .collect::<Vec<_>>();
3688 0 : if !mismatching.is_empty() {
3689 : // this can be hit by races which should not happen because operation lock on cplane
3690 0 : let matching = results.len() - mismatching.len();
3691 0 : tracing::error!(
3692 : matching,
3693 : compared_against=?any,
3694 : ?mismatching,
3695 0 : "shards returned different results"
3696 : );
3697 :
3698 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3699 0 : }
3700 0 :
3701 0 : Ok(any.1)
3702 0 : }).await?
3703 0 : }
3704 :
3705 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3706 0 : &self,
3707 0 : tenant_id: TenantId,
3708 0 : timeline_id: TimelineId,
3709 0 : dir: BlockUnblock,
3710 0 : ) -> Result<(), ApiError> {
3711 0 : let _tenant_lock = trace_shared_lock(
3712 0 : &self.tenant_op_locks,
3713 0 : tenant_id,
3714 0 : TenantOperations::TimelineGcBlockUnblock,
3715 0 : )
3716 0 : .await;
3717 :
3718 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3719 0 : if targets.0.is_empty() {
3720 0 : return Err(ApiError::NotFound(
3721 0 : anyhow::anyhow!("Tenant not found").into(),
3722 0 : ));
3723 0 : }
3724 :
3725 0 : async fn do_one(
3726 0 : tenant_shard_id: TenantShardId,
3727 0 : timeline_id: TimelineId,
3728 0 : node: Node,
3729 0 : jwt: Option<String>,
3730 0 : dir: BlockUnblock,
3731 0 : ) -> Result<(), ApiError> {
3732 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3733 0 :
3734 0 : client
3735 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3736 0 : .await
3737 0 : .map_err(|e| passthrough_api_error(&node, e))
3738 0 : }
3739 :
3740 : // no shard needs to go first/last; the operation should be idempotent
3741 0 : let locations = targets
3742 0 : .0
3743 0 : .iter()
3744 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3745 0 : .collect();
3746 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
3747 0 : futures::FutureExt::boxed(do_one(
3748 0 : tenant_shard_id,
3749 0 : timeline_id,
3750 0 : node,
3751 0 : self.config.jwt_token.clone(),
3752 0 : dir,
3753 0 : ))
3754 0 : })
3755 0 : .await
3756 0 : })
3757 0 : .await??;
3758 0 : Ok(())
3759 0 : }
3760 :
3761 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
3762 0 : &self,
3763 0 : tenant_shard_id: TenantShardId,
3764 0 : timeline_id: TimelineId,
3765 0 : concurrency: Option<usize>,
3766 0 : ) -> Result<(), ApiError> {
3767 0 : let _tenant_lock = trace_shared_lock(
3768 0 : &self.tenant_op_locks,
3769 0 : tenant_shard_id.tenant_id,
3770 0 : TenantOperations::DownloadHeatmapLayers,
3771 0 : )
3772 0 : .await;
3773 :
3774 0 : let targets = {
3775 0 : let locked = self.inner.read().unwrap();
3776 0 : let mut targets = Vec::new();
3777 :
3778 : // If the request got an unsharded tenant id, then apply
3779 : // the operation to all shards. Otherwise, apply it to a specific shard.
3780 0 : let shards_range = if tenant_shard_id.is_unsharded() {
3781 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
3782 : } else {
3783 0 : tenant_shard_id.range()
3784 : };
3785 :
3786 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
3787 0 : if let Some(node_id) = shard.intent.get_attached() {
3788 0 : let node = locked
3789 0 : .nodes
3790 0 : .get(node_id)
3791 0 : .expect("Pageservers may not be deleted while referenced");
3792 0 :
3793 0 : targets.push((*tenant_shard_id, node.clone()));
3794 0 : }
3795 : }
3796 0 : targets
3797 0 : };
3798 0 :
3799 0 : self.tenant_for_shards_api(
3800 0 : targets,
3801 0 : |tenant_shard_id, client| async move {
3802 0 : client
3803 0 : .timeline_download_heatmap_layers(tenant_shard_id, timeline_id, concurrency)
3804 0 : .await
3805 0 : },
3806 0 : 1,
3807 0 : 1,
3808 0 : SHORT_RECONCILE_TIMEOUT,
3809 0 : &self.cancel,
3810 0 : )
3811 0 : .await;
3812 :
3813 0 : Ok(())
3814 0 : }
3815 :
3816 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3817 : ///
3818 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3819 0 : async fn tenant_for_shards<F, R>(
3820 0 : &self,
3821 0 : locations: Vec<(TenantShardId, Node)>,
3822 0 : mut req_fn: F,
3823 0 : ) -> Result<Vec<R>, ApiError>
3824 0 : where
3825 0 : F: FnMut(
3826 0 : TenantShardId,
3827 0 : Node,
3828 0 : )
3829 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3830 0 : {
3831 0 : let mut futs = FuturesUnordered::new();
3832 0 : let mut results = Vec::with_capacity(locations.len());
3833 :
3834 0 : for (tenant_shard_id, node) in locations {
3835 0 : futs.push(req_fn(tenant_shard_id, node));
3836 0 : }
3837 :
3838 0 : while let Some(r) = futs.next().await {
3839 0 : results.push(r?);
3840 : }
3841 :
3842 0 : Ok(results)
3843 0 : }
3844 :
3845 : /// Concurrently invoke a pageserver API call on many shards at once
3846 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3847 0 : &self,
3848 0 : locations: Vec<(TenantShardId, Node)>,
3849 0 : op: O,
3850 0 : warn_threshold: u32,
3851 0 : max_retries: u32,
3852 0 : timeout: Duration,
3853 0 : cancel: &CancellationToken,
3854 0 : ) -> Vec<mgmt_api::Result<T>>
3855 0 : where
3856 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3857 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3858 0 : {
3859 0 : let mut futs = FuturesUnordered::new();
3860 0 : let mut results = Vec::with_capacity(locations.len());
3861 :
3862 0 : for (tenant_shard_id, node) in locations {
3863 0 : futs.push(async move {
3864 0 : node.with_client_retries(
3865 0 : |client| op(tenant_shard_id, client),
3866 0 : &self.config.jwt_token,
3867 0 : warn_threshold,
3868 0 : max_retries,
3869 0 : timeout,
3870 0 : cancel,
3871 0 : )
3872 0 : .await
3873 0 : });
3874 0 : }
3875 :
3876 0 : while let Some(r) = futs.next().await {
3877 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3878 0 : results.push(r);
3879 0 : }
3880 :
3881 0 : results
3882 0 : }
3883 :
3884 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3885 : /// when creating and deleting timelines:
3886 : /// - Makes sure shards are attached somewhere if they weren't already
3887 : /// - Looks up the shards and the nodes where they were most recently attached
3888 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3889 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3890 0 : async fn tenant_remote_mutation<R, O, F>(
3891 0 : &self,
3892 0 : tenant_id: TenantId,
3893 0 : op: O,
3894 0 : ) -> Result<R, ApiError>
3895 0 : where
3896 0 : O: FnOnce(TenantMutationLocations) -> F,
3897 0 : F: std::future::Future<Output = R>,
3898 0 : {
3899 0 : let mutation_locations = {
3900 0 : let mut locations = TenantMutationLocations::default();
3901 :
3902 : // Load the currently attached pageservers for the latest generation of each shard. This can
3903 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3904 : // will still be the latest when we're done: we will check generations again at the end of
3905 : // this function to handle that.
3906 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
3907 :
3908 0 : if generations
3909 0 : .iter()
3910 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3911 : {
3912 0 : let shard_generations = generations
3913 0 : .into_iter()
3914 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
3915 0 : .collect::<HashMap<_, _>>();
3916 0 :
3917 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3918 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)
3919 0 : let locked = self.inner.read().unwrap();
3920 0 : for (shard_id, shard) in
3921 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3922 : {
3923 0 : match shard.policy {
3924 : PlacementPolicy::Attached(_) => {
3925 : // This shard is meant to be attached: the caller is not wrong to try and
3926 : // use this function, but we can't service the request right now.
3927 0 : let Some(generation) = shard_generations.get(shard_id) else {
3928 : // This can only happen if there is a split brain controller modifying the database. This should
3929 : // never happen when testing, and if it happens in production we can only log the issue.
3930 0 : debug_assert!(false);
3931 0 : tracing::error!("Shard {shard_id} not found in generation state! Is another rogue controller running?");
3932 0 : continue;
3933 : };
3934 0 : let (generation, generation_pageserver) = generation;
3935 0 : if let Some(generation) = generation {
3936 0 : if generation_pageserver.is_none() {
3937 : // This is legitimate only in a very narrow window where the shard was only just configured into
3938 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
3939 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
3940 0 : tracing::warn!("Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?");
3941 0 : }
3942 : } else {
3943 : // This should never happen: a shard with no generation is only permitted when it was created in some state
3944 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
3945 0 : debug_assert!(false);
3946 0 : tracing::error!("Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!");
3947 0 : continue;
3948 : }
3949 : }
3950 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3951 0 : return Err(ApiError::Conflict(format!(
3952 0 : "Shard {shard_id} tenant has policy {:?}",
3953 0 : shard.policy
3954 0 : )));
3955 : }
3956 : }
3957 : }
3958 :
3959 0 : return Err(ApiError::ResourceUnavailable(
3960 0 : "One or more shards in tenant is not yet attached".into(),
3961 0 : ));
3962 0 : }
3963 0 :
3964 0 : let locked = self.inner.read().unwrap();
3965 : for ShardGenerationState {
3966 0 : tenant_shard_id,
3967 0 : generation,
3968 0 : generation_pageserver,
3969 0 : } in generations
3970 : {
3971 0 : let node_id = generation_pageserver.expect("We checked for None above");
3972 0 : let node = locked
3973 0 : .nodes
3974 0 : .get(&node_id)
3975 0 : .ok_or(ApiError::Conflict(format!(
3976 0 : "Raced with removal of node {node_id}"
3977 0 : )))?;
3978 0 : let generation = generation.expect("Checked above");
3979 0 :
3980 0 : let tenant = locked.tenants.get(&tenant_shard_id);
3981 :
3982 : // TODO(vlad): Abstract the logic that finds stale attached locations
3983 : // from observed state into a [`Service`] method.
3984 0 : let other_locations = match tenant {
3985 0 : Some(tenant) => {
3986 0 : let mut other = tenant.attached_locations();
3987 0 : let latest_location_index =
3988 0 : other.iter().position(|&l| l == (node.get_id(), generation));
3989 0 : if let Some(idx) = latest_location_index {
3990 0 : other.remove(idx);
3991 0 : }
3992 :
3993 0 : other
3994 : }
3995 0 : None => Vec::default(),
3996 : };
3997 :
3998 0 : let location = ShardMutationLocations {
3999 0 : latest: MutationLocation {
4000 0 : node: node.clone(),
4001 0 : generation,
4002 0 : },
4003 0 : other: other_locations
4004 0 : .into_iter()
4005 0 : .filter_map(|(node_id, generation)| {
4006 0 : let node = locked.nodes.get(&node_id)?;
4007 :
4008 0 : Some(MutationLocation {
4009 0 : node: node.clone(),
4010 0 : generation,
4011 0 : })
4012 0 : })
4013 0 : .collect(),
4014 0 : };
4015 0 : locations.0.insert(tenant_shard_id, location);
4016 0 : }
4017 :
4018 0 : locations
4019 : };
4020 :
4021 0 : let result = op(mutation_locations.clone()).await;
4022 :
4023 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
4024 : // our remote operation executed on the latest generation and is therefore persistent.
4025 : {
4026 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
4027 0 : if latest_generations
4028 0 : .into_iter()
4029 0 : .map(
4030 0 : |ShardGenerationState {
4031 : tenant_shard_id,
4032 : generation,
4033 : generation_pageserver: _,
4034 0 : }| (tenant_shard_id, generation),
4035 0 : )
4036 0 : .collect::<Vec<_>>()
4037 0 : != mutation_locations
4038 0 : .0
4039 0 : .into_iter()
4040 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
4041 0 : .collect::<Vec<_>>()
4042 : {
4043 : // We raced with something that incremented the generation, and therefore cannot be
4044 : // confident that our actions are persistent (they might have hit an old generation).
4045 : //
4046 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
4047 0 : return Err(ApiError::ResourceUnavailable(
4048 0 : "Tenant attachment changed, please retry".into(),
4049 0 : ));
4050 0 : }
4051 0 : }
4052 0 :
4053 0 : Ok(result)
4054 0 : }
4055 :
4056 0 : pub(crate) async fn tenant_timeline_delete(
4057 0 : &self,
4058 0 : tenant_id: TenantId,
4059 0 : timeline_id: TimelineId,
4060 0 : ) -> Result<StatusCode, ApiError> {
4061 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
4062 0 : let _tenant_lock = trace_shared_lock(
4063 0 : &self.tenant_op_locks,
4064 0 : tenant_id,
4065 0 : TenantOperations::TimelineDelete,
4066 0 : )
4067 0 : .await;
4068 :
4069 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
4070 0 : if targets.0.is_empty() {
4071 0 : return Err(ApiError::NotFound(
4072 0 : anyhow::anyhow!("Tenant not found").into(),
4073 0 : ));
4074 0 : }
4075 0 :
4076 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
4077 0 : assert!(shard_zero_tid.is_shard_zero());
4078 :
4079 0 : async fn delete_one(
4080 0 : tenant_shard_id: TenantShardId,
4081 0 : timeline_id: TimelineId,
4082 0 : node: Node,
4083 0 : jwt: Option<String>,
4084 0 : ) -> Result<StatusCode, ApiError> {
4085 0 : tracing::info!(
4086 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4087 : );
4088 :
4089 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
4090 0 : let res = client
4091 0 : .timeline_delete(tenant_shard_id, timeline_id)
4092 0 : .await;
4093 :
4094 0 : match res {
4095 0 : Ok(ok) => Ok(ok),
4096 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
4097 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
4098 0 : Err(e) => {
4099 0 : Err(
4100 0 : ApiError::InternalServerError(anyhow::anyhow!(
4101 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
4102 0 : ))
4103 0 : )
4104 : }
4105 : }
4106 0 : }
4107 :
4108 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4109 0 : let statuses = self
4110 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
4111 0 : Box::pin(delete_one(
4112 0 : tenant_shard_id,
4113 0 : timeline_id,
4114 0 : node,
4115 0 : self.config.jwt_token.clone(),
4116 0 : ))
4117 0 : })
4118 0 : .await?;
4119 :
4120 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
4121 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
4122 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
4123 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
4124 0 : return Ok(StatusCode::CONFLICT);
4125 0 : }
4126 0 :
4127 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
4128 0 : return Ok(StatusCode::ACCEPTED);
4129 0 : }
4130 :
4131 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
4132 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
4133 0 : let shard_zero_status = delete_one(
4134 0 : shard_zero_tid,
4135 0 : timeline_id,
4136 0 : shard_zero_locations.latest.node,
4137 0 : self.config.jwt_token.clone(),
4138 0 : )
4139 0 : .await?;
4140 0 : Ok(shard_zero_status)
4141 0 : }).await?
4142 0 : }
4143 :
4144 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
4145 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
4146 0 : pub(crate) async fn tenant_shard0_node(
4147 0 : &self,
4148 0 : tenant_id: TenantId,
4149 0 : ) -> Result<(Node, TenantShardId), ApiError> {
4150 0 : // Look up in-memory state and maybe use the node from there.
4151 0 : {
4152 0 : let locked = self.inner.read().unwrap();
4153 0 : let Some((tenant_shard_id, shard)) = locked
4154 0 : .tenants
4155 0 : .range(TenantShardId::tenant_range(tenant_id))
4156 0 : .next()
4157 : else {
4158 0 : return Err(ApiError::NotFound(
4159 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
4160 0 : ));
4161 : };
4162 :
4163 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
4164 0 : tracing::warn!(
4165 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
4166 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
4167 : shard.policy
4168 : );
4169 0 : return Err(ApiError::Conflict(
4170 0 : "Cannot call timeline API on non-attached tenant".to_string(),
4171 0 : ));
4172 : };
4173 :
4174 0 : if shard.reconciler.is_none() {
4175 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
4176 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
4177 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
4178 : // This should never happen
4179 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4180 0 : "Shard refers to nonexistent node"
4181 0 : )));
4182 : };
4183 0 : return Ok((node.clone(), *tenant_shard_id));
4184 0 : }
4185 : };
4186 :
4187 : // Look up the latest attached pageserver location from the database
4188 : // generation state: this will reflect the progress of any ongoing migration.
4189 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4190 : // the case where they call through to the pageserver and get a 404.
4191 0 : let db_result = self.persistence.tenant_generations(tenant_id).await?;
4192 : let Some(ShardGenerationState {
4193 0 : tenant_shard_id,
4194 0 : generation: _,
4195 0 : generation_pageserver: Some(node_id),
4196 0 : }) = db_result.first()
4197 : else {
4198 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4199 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4200 : // or a conflict response (case where tenant was detached in background)
4201 0 : return Err(ApiError::ResourceUnavailable(
4202 0 : "Shard {} not found in database, or is not attached".into(),
4203 0 : ));
4204 : };
4205 0 : let locked = self.inner.read().unwrap();
4206 0 : let Some(node) = locked.nodes.get(node_id) else {
4207 : // This should never happen
4208 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4209 0 : "Shard refers to nonexistent node"
4210 0 : )));
4211 : };
4212 :
4213 0 : Ok((node.clone(), *tenant_shard_id))
4214 0 : }
4215 :
4216 0 : pub(crate) fn tenant_locate(
4217 0 : &self,
4218 0 : tenant_id: TenantId,
4219 0 : ) -> Result<TenantLocateResponse, ApiError> {
4220 0 : let locked = self.inner.read().unwrap();
4221 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4222 :
4223 0 : let mut result = Vec::new();
4224 0 : let mut shard_params: Option<ShardParameters> = None;
4225 :
4226 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4227 : {
4228 0 : let node_id =
4229 0 : shard
4230 0 : .intent
4231 0 : .get_attached()
4232 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4233 0 : "Cannot locate a tenant that is not attached"
4234 0 : )))?;
4235 :
4236 0 : let node = locked
4237 0 : .nodes
4238 0 : .get(&node_id)
4239 0 : .expect("Pageservers may not be deleted while referenced");
4240 0 :
4241 0 : result.push(node.shard_location(*tenant_shard_id));
4242 0 :
4243 0 : match &shard_params {
4244 0 : None => {
4245 0 : shard_params = Some(ShardParameters {
4246 0 : stripe_size: shard.shard.stripe_size,
4247 0 : count: shard.shard.count,
4248 0 : });
4249 0 : }
4250 0 : Some(params) => {
4251 0 : if params.stripe_size != shard.shard.stripe_size {
4252 : // This should never happen. We enforce at runtime because it's simpler than
4253 : // adding an extra per-tenant data structure to store the things that should be the same
4254 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4255 0 : "Inconsistent shard stripe size parameters!"
4256 0 : )));
4257 0 : }
4258 : }
4259 : }
4260 : }
4261 :
4262 0 : if result.is_empty() {
4263 0 : return Err(ApiError::NotFound(
4264 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4265 0 : ));
4266 0 : }
4267 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4268 0 : tracing::info!(
4269 0 : "Located tenant {} with params {:?} on shards {}",
4270 0 : tenant_id,
4271 0 : shard_params,
4272 0 : result
4273 0 : .iter()
4274 0 : .map(|s| format!("{:?}", s))
4275 0 : .collect::<Vec<_>>()
4276 0 : .join(",")
4277 : );
4278 :
4279 0 : Ok(TenantLocateResponse {
4280 0 : shards: result,
4281 0 : shard_params,
4282 0 : })
4283 0 : }
4284 :
4285 : /// Returns None if the input iterator of shards does not include a shard with number=0
4286 0 : fn tenant_describe_impl<'a>(
4287 0 : &self,
4288 0 : shards: impl Iterator<Item = &'a TenantShard>,
4289 0 : ) -> Option<TenantDescribeResponse> {
4290 0 : let mut shard_zero = None;
4291 0 : let mut describe_shards = Vec::new();
4292 :
4293 0 : for shard in shards {
4294 0 : if shard.tenant_shard_id.is_shard_zero() {
4295 0 : shard_zero = Some(shard);
4296 0 : }
4297 :
4298 0 : describe_shards.push(TenantDescribeResponseShard {
4299 0 : tenant_shard_id: shard.tenant_shard_id,
4300 0 : node_attached: *shard.intent.get_attached(),
4301 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4302 0 : last_error: shard
4303 0 : .last_error
4304 0 : .lock()
4305 0 : .unwrap()
4306 0 : .as_ref()
4307 0 : .map(|e| format!("{e}"))
4308 0 : .unwrap_or("".to_string())
4309 0 : .clone(),
4310 0 : is_reconciling: shard.reconciler.is_some(),
4311 0 : is_pending_compute_notification: shard.pending_compute_notification,
4312 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4313 0 : scheduling_policy: *shard.get_scheduling_policy(),
4314 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4315 : })
4316 : }
4317 :
4318 0 : let shard_zero = shard_zero?;
4319 :
4320 0 : Some(TenantDescribeResponse {
4321 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
4322 0 : shards: describe_shards,
4323 0 : stripe_size: shard_zero.shard.stripe_size,
4324 0 : policy: shard_zero.policy.clone(),
4325 0 : config: shard_zero.config.clone(),
4326 0 : })
4327 0 : }
4328 :
4329 0 : pub(crate) fn tenant_describe(
4330 0 : &self,
4331 0 : tenant_id: TenantId,
4332 0 : ) -> Result<TenantDescribeResponse, ApiError> {
4333 0 : let locked = self.inner.read().unwrap();
4334 0 :
4335 0 : self.tenant_describe_impl(
4336 0 : locked
4337 0 : .tenants
4338 0 : .range(TenantShardId::tenant_range(tenant_id))
4339 0 : .map(|(_k, v)| v),
4340 0 : )
4341 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
4342 0 : }
4343 :
4344 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
4345 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
4346 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
4347 : /// in our external API.
4348 0 : pub(crate) fn tenant_list(
4349 0 : &self,
4350 0 : limit: Option<usize>,
4351 0 : start_after: Option<TenantId>,
4352 0 : ) -> Vec<TenantDescribeResponse> {
4353 0 : let locked = self.inner.read().unwrap();
4354 :
4355 : // Apply start_from parameter
4356 0 : let shard_range = match start_after {
4357 0 : None => locked.tenants.range(..),
4358 0 : Some(tenant_id) => locked.tenants.range(
4359 0 : TenantShardId {
4360 0 : tenant_id,
4361 0 : shard_number: ShardNumber(u8::MAX),
4362 0 : shard_count: ShardCount(u8::MAX),
4363 0 : }..,
4364 0 : ),
4365 : };
4366 :
4367 0 : let mut result = Vec::new();
4368 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
4369 0 : result.push(
4370 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
4371 0 : .expect("Groups are always non-empty"),
4372 0 : );
4373 :
4374 : // Enforce `limit` parameter
4375 0 : if let Some(limit) = limit {
4376 0 : if result.len() >= limit {
4377 0 : break;
4378 0 : }
4379 0 : }
4380 : }
4381 :
4382 0 : result
4383 0 : }
4384 :
4385 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
4386 : async fn abort_tenant_shard_split(
4387 : &self,
4388 : op: &TenantShardSplitAbort,
4389 : ) -> Result<(), TenantShardSplitAbortError> {
4390 : // Cleaning up a split:
4391 : // - Parent shards are not destroyed during a split, just detached.
4392 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
4393 : // just the children attached, or both.
4394 : //
4395 : // Therefore our work to do is to:
4396 : // 1. Clean up storage controller's internal state to just refer to parents, no children
4397 : // 2. Call out to pageservers to ensure that children are detached
4398 : // 3. Call out to pageservers to ensure that parents are attached.
4399 : //
4400 : // Crash safety:
4401 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
4402 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
4403 : // and detach them.
4404 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
4405 : // from our database, then we will re-enter this cleanup routine on startup.
4406 :
4407 : let TenantShardSplitAbort {
4408 : tenant_id,
4409 : new_shard_count,
4410 : new_stripe_size,
4411 : ..
4412 : } = op;
4413 :
4414 : // First abort persistent state, if any exists.
4415 : match self
4416 : .persistence
4417 : .abort_shard_split(*tenant_id, *new_shard_count)
4418 : .await?
4419 : {
4420 : AbortShardSplitStatus::Aborted => {
4421 : // Proceed to roll back any child shards created on pageservers
4422 : }
4423 : AbortShardSplitStatus::Complete => {
4424 : // The split completed (we might hit that path if e.g. our database transaction
4425 : // to write the completion landed in the database, but we dropped connection
4426 : // before seeing the result).
4427 : //
4428 : // We must update in-memory state to reflect the successful split.
4429 : self.tenant_shard_split_commit_inmem(
4430 : *tenant_id,
4431 : *new_shard_count,
4432 : *new_stripe_size,
4433 : );
4434 : return Ok(());
4435 : }
4436 : }
4437 :
4438 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
4439 : let detach_locations: Vec<(Node, TenantShardId)> = {
4440 : let mut detach_locations = Vec::new();
4441 : let mut locked = self.inner.write().unwrap();
4442 : let (nodes, tenants, scheduler) = locked.parts_mut();
4443 :
4444 : for (tenant_shard_id, shard) in
4445 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
4446 : {
4447 : if shard.shard.count == op.new_shard_count {
4448 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
4449 : // is infallible, so if we got an error we shouldn't have got that far.
4450 : tracing::warn!(
4451 : "During split abort, child shard {tenant_shard_id} found in-memory"
4452 : );
4453 : continue;
4454 : }
4455 :
4456 : // Add the children of this shard to this list of things to detach
4457 : if let Some(node_id) = shard.intent.get_attached() {
4458 : for child_id in tenant_shard_id.split(*new_shard_count) {
4459 : detach_locations.push((
4460 : nodes
4461 : .get(node_id)
4462 : .expect("Intent references nonexistent node")
4463 : .clone(),
4464 : child_id,
4465 : ));
4466 : }
4467 : } else {
4468 : tracing::warn!(
4469 : "During split abort, shard {tenant_shard_id} has no attached location"
4470 : );
4471 : }
4472 :
4473 : tracing::info!("Restoring parent shard {tenant_shard_id}");
4474 :
4475 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
4476 : // if the original attachment location is offline.
4477 : if let Some(node_id) = shard.intent.get_attached() {
4478 : if !nodes.get(node_id).unwrap().is_available() {
4479 : tracing::info!("Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}");
4480 : shard.intent.demote_attached(scheduler, *node_id);
4481 : }
4482 : }
4483 : for node_id in shard.intent.get_secondary().clone() {
4484 : if !nodes.get(&node_id).unwrap().is_available() {
4485 : tracing::info!("Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}");
4486 : shard.intent.remove_secondary(scheduler, node_id);
4487 : }
4488 : }
4489 :
4490 : shard.splitting = SplitState::Idle;
4491 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
4492 : // If this shard can't be scheduled now (perhaps due to offline nodes or
4493 : // capacity issues), that must not prevent us rolling back a split. In this
4494 : // case it should be eventually scheduled in the background.
4495 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
4496 : }
4497 :
4498 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
4499 : }
4500 :
4501 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
4502 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
4503 :
4504 : detach_locations
4505 : };
4506 :
4507 : for (node, child_id) in detach_locations {
4508 : if !node.is_available() {
4509 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
4510 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
4511 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
4512 : // them from the node.
4513 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
4514 : continue;
4515 : }
4516 :
4517 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
4518 : // a 503 and retry, up to our limit.
4519 : tracing::info!("Detaching {child_id} on {node}...");
4520 : match node
4521 : .with_client_retries(
4522 0 : |client| async move {
4523 0 : let config = LocationConfig {
4524 0 : mode: LocationConfigMode::Detached,
4525 0 : generation: None,
4526 0 : secondary_conf: None,
4527 0 : shard_number: child_id.shard_number.0,
4528 0 : shard_count: child_id.shard_count.literal(),
4529 0 : // Stripe size and tenant config don't matter when detaching
4530 0 : shard_stripe_size: 0,
4531 0 : tenant_conf: TenantConfig::default(),
4532 0 : };
4533 0 :
4534 0 : client.location_config(child_id, config, None, false).await
4535 0 : },
4536 : &self.config.jwt_token,
4537 : 1,
4538 : 10,
4539 : Duration::from_secs(5),
4540 : &self.cancel,
4541 : )
4542 : .await
4543 : {
4544 : Some(Ok(_)) => {}
4545 : Some(Err(e)) => {
4546 : // We failed to communicate with the remote node. This is problematic: we may be
4547 : // leaving it with a rogue child shard.
4548 : tracing::warn!(
4549 : "Failed to detach child {child_id} from node {node} during abort"
4550 : );
4551 : return Err(e.into());
4552 : }
4553 : None => {
4554 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
4555 : // clean up on restart. The node going offline requires a retry.
4556 : return Err(TenantShardSplitAbortError::Unavailable);
4557 : }
4558 : };
4559 : }
4560 :
4561 : tracing::info!("Successfully aborted split");
4562 : Ok(())
4563 : }
4564 :
4565 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
4566 : /// of the tenant map to reflect the child shards that exist after the split.
4567 0 : fn tenant_shard_split_commit_inmem(
4568 0 : &self,
4569 0 : tenant_id: TenantId,
4570 0 : new_shard_count: ShardCount,
4571 0 : new_stripe_size: Option<ShardStripeSize>,
4572 0 : ) -> (
4573 0 : TenantShardSplitResponse,
4574 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
4575 0 : Vec<ReconcilerWaiter>,
4576 0 : ) {
4577 0 : let mut response = TenantShardSplitResponse {
4578 0 : new_shards: Vec::new(),
4579 0 : };
4580 0 : let mut child_locations = Vec::new();
4581 0 : let mut waiters = Vec::new();
4582 0 :
4583 0 : {
4584 0 : let mut locked = self.inner.write().unwrap();
4585 0 :
4586 0 : let parent_ids = locked
4587 0 : .tenants
4588 0 : .range(TenantShardId::tenant_range(tenant_id))
4589 0 : .map(|(shard_id, _)| *shard_id)
4590 0 : .collect::<Vec<_>>();
4591 0 :
4592 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4593 0 : for parent_id in parent_ids {
4594 0 : let child_ids = parent_id.split(new_shard_count);
4595 :
4596 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
4597 0 : let mut old_state = tenants
4598 0 : .remove(&parent_id)
4599 0 : .expect("It was present, we just split it");
4600 0 :
4601 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
4602 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
4603 0 : // nothing else can clear this.
4604 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
4605 :
4606 0 : let old_attached = old_state.intent.get_attached().unwrap();
4607 0 : old_state.intent.clear(scheduler);
4608 0 : let generation = old_state.generation.expect("Shard must have been attached");
4609 0 : (
4610 0 : old_attached,
4611 0 : generation,
4612 0 : old_state.policy.clone(),
4613 0 : old_state.shard,
4614 0 : old_state.config.clone(),
4615 0 : old_state.preferred_az().cloned(),
4616 0 : )
4617 0 : };
4618 0 :
4619 0 : let mut schedule_context = ScheduleContext::default();
4620 0 : for child in child_ids {
4621 0 : let mut child_shard = parent_ident;
4622 0 : child_shard.number = child.shard_number;
4623 0 : child_shard.count = child.shard_count;
4624 0 : if let Some(stripe_size) = new_stripe_size {
4625 0 : child_shard.stripe_size = stripe_size;
4626 0 : }
4627 :
4628 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
4629 0 : child_observed.insert(
4630 0 : pageserver,
4631 0 : ObservedStateLocation {
4632 0 : conf: Some(attached_location_conf(
4633 0 : generation,
4634 0 : &child_shard,
4635 0 : &config,
4636 0 : &policy,
4637 0 : )),
4638 0 : },
4639 0 : );
4640 0 :
4641 0 : let mut child_state =
4642 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
4643 0 : child_state.intent =
4644 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
4645 0 : child_state.observed = ObservedState {
4646 0 : locations: child_observed,
4647 0 : };
4648 0 : child_state.generation = Some(generation);
4649 0 : child_state.config = config.clone();
4650 0 :
4651 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
4652 0 : // as at this point in the split process we have succeeded and this part is infallible:
4653 0 : // we will never need to do any special recovery from this state.
4654 0 :
4655 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
4656 :
4657 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
4658 : // This is not fatal, because we've implicitly already got an attached
4659 : // location for the child shard. Failure here just means we couldn't
4660 : // find a secondary (e.g. because cluster is overloaded).
4661 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
4662 0 : }
4663 : // In the background, attach secondary locations for the new shards
4664 0 : if let Some(waiter) = self.maybe_reconcile_shard(
4665 0 : &mut child_state,
4666 0 : nodes,
4667 0 : ReconcilerPriority::High,
4668 0 : ) {
4669 0 : waiters.push(waiter);
4670 0 : }
4671 :
4672 0 : tenants.insert(child, child_state);
4673 0 : response.new_shards.push(child);
4674 : }
4675 : }
4676 0 : (response, child_locations, waiters)
4677 0 : }
4678 0 : }
4679 :
4680 0 : async fn tenant_shard_split_start_secondaries(
4681 0 : &self,
4682 0 : tenant_id: TenantId,
4683 0 : waiters: Vec<ReconcilerWaiter>,
4684 0 : ) {
4685 : // Wait for initial reconcile of child shards, this creates the secondary locations
4686 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
4687 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
4688 : // their secondaries couldn't be attached.
4689 0 : tracing::warn!("Failed to reconcile after split: {e}");
4690 0 : return;
4691 0 : }
4692 :
4693 : // Take the state lock to discover the attached & secondary intents for all shards
4694 0 : let (attached, secondary) = {
4695 0 : let locked = self.inner.read().unwrap();
4696 0 : let mut attached = Vec::new();
4697 0 : let mut secondary = Vec::new();
4698 :
4699 0 : for (tenant_shard_id, shard) in
4700 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4701 : {
4702 0 : let Some(node_id) = shard.intent.get_attached() else {
4703 : // Unexpected. Race with a PlacementPolicy change?
4704 0 : tracing::warn!(
4705 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
4706 : );
4707 0 : continue;
4708 : };
4709 :
4710 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
4711 : // No secondary location. Nothing for us to do.
4712 0 : continue;
4713 : };
4714 :
4715 0 : let attached_node = locked
4716 0 : .nodes
4717 0 : .get(node_id)
4718 0 : .expect("Pageservers may not be deleted while referenced");
4719 0 :
4720 0 : let secondary_node = locked
4721 0 : .nodes
4722 0 : .get(secondary_node_id)
4723 0 : .expect("Pageservers may not be deleted while referenced");
4724 0 :
4725 0 : attached.push((*tenant_shard_id, attached_node.clone()));
4726 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
4727 : }
4728 0 : (attached, secondary)
4729 0 : };
4730 0 :
4731 0 : if secondary.is_empty() {
4732 : // No secondary locations; nothing for us to do
4733 0 : return;
4734 0 : }
4735 :
4736 0 : for result in self
4737 0 : .tenant_for_shards_api(
4738 0 : attached,
4739 0 : |tenant_shard_id, client| async move {
4740 0 : client.tenant_heatmap_upload(tenant_shard_id).await
4741 0 : },
4742 0 : 1,
4743 0 : 1,
4744 0 : SHORT_RECONCILE_TIMEOUT,
4745 0 : &self.cancel,
4746 0 : )
4747 0 : .await
4748 : {
4749 0 : if let Err(e) = result {
4750 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
4751 0 : return;
4752 0 : }
4753 : }
4754 :
4755 0 : for result in self
4756 0 : .tenant_for_shards_api(
4757 0 : secondary,
4758 0 : |tenant_shard_id, client| async move {
4759 0 : client
4760 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
4761 0 : .await
4762 0 : },
4763 0 : 1,
4764 0 : 1,
4765 0 : SHORT_RECONCILE_TIMEOUT,
4766 0 : &self.cancel,
4767 0 : )
4768 0 : .await
4769 : {
4770 0 : if let Err(e) = result {
4771 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
4772 0 : return;
4773 0 : }
4774 : }
4775 0 : }
4776 :
4777 0 : pub(crate) async fn tenant_shard_split(
4778 0 : &self,
4779 0 : tenant_id: TenantId,
4780 0 : split_req: TenantShardSplitRequest,
4781 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
4782 : // TODO: return 503 if we get stuck waiting for this lock
4783 : // (issue https://github.com/neondatabase/neon/issues/7108)
4784 0 : let _tenant_lock = trace_exclusive_lock(
4785 0 : &self.tenant_op_locks,
4786 0 : tenant_id,
4787 0 : TenantOperations::ShardSplit,
4788 0 : )
4789 0 : .await;
4790 :
4791 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
4792 0 : let new_stripe_size = split_req.new_stripe_size;
4793 :
4794 : // Validate the request and construct parameters. This phase is fallible, but does not require
4795 : // rollback on errors, as it does no I/O and mutates no state.
4796 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
4797 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
4798 0 : ShardSplitAction::Split(params) => params,
4799 : };
4800 :
4801 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
4802 : // we must roll back.
4803 0 : let r = self
4804 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
4805 0 : .await;
4806 :
4807 0 : let (response, waiters) = match r {
4808 0 : Ok(r) => r,
4809 0 : Err(e) => {
4810 0 : // Split might be part-done, we must do work to abort it.
4811 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
4812 0 : self.abort_tx
4813 0 : .send(TenantShardSplitAbort {
4814 0 : tenant_id,
4815 0 : new_shard_count,
4816 0 : new_stripe_size,
4817 0 : _tenant_lock,
4818 0 : })
4819 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
4820 0 : .ok();
4821 0 : return Err(e);
4822 : }
4823 : };
4824 :
4825 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
4826 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
4827 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
4828 : // in [`Self::optimize_all`]
4829 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
4830 0 : .await;
4831 0 : Ok(response)
4832 0 : }
4833 :
4834 0 : fn prepare_tenant_shard_split(
4835 0 : &self,
4836 0 : tenant_id: TenantId,
4837 0 : split_req: TenantShardSplitRequest,
4838 0 : ) -> Result<ShardSplitAction, ApiError> {
4839 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
4840 0 : anyhow::anyhow!("failpoint")
4841 0 : )));
4842 :
4843 0 : let mut policy = None;
4844 0 : let mut config = None;
4845 0 : let mut shard_ident = None;
4846 0 : let mut preferred_az_id = None;
4847 : // Validate input, and calculate which shards we will create
4848 0 : let (old_shard_count, targets) =
4849 : {
4850 0 : let locked = self.inner.read().unwrap();
4851 0 :
4852 0 : let pageservers = locked.nodes.clone();
4853 0 :
4854 0 : let mut targets = Vec::new();
4855 0 :
4856 0 : // In case this is a retry, count how many already-split shards we found
4857 0 : let mut children_found = Vec::new();
4858 0 : let mut old_shard_count = None;
4859 :
4860 0 : for (tenant_shard_id, shard) in
4861 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4862 : {
4863 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
4864 : Ordering::Equal => {
4865 : // Already split this
4866 0 : children_found.push(*tenant_shard_id);
4867 0 : continue;
4868 : }
4869 : Ordering::Greater => {
4870 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4871 0 : "Requested count {} but already have shards at count {}",
4872 0 : split_req.new_shard_count,
4873 0 : shard.shard.count.count()
4874 0 : )));
4875 : }
4876 0 : Ordering::Less => {
4877 0 : // Fall through: this shard has lower count than requested,
4878 0 : // is a candidate for splitting.
4879 0 : }
4880 0 : }
4881 0 :
4882 0 : match old_shard_count {
4883 0 : None => old_shard_count = Some(shard.shard.count),
4884 0 : Some(old_shard_count) => {
4885 0 : if old_shard_count != shard.shard.count {
4886 : // We may hit this case if a caller asked for two splits to
4887 : // different sizes, before the first one is complete.
4888 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
4889 : // of shard_count=1 and shard_count=2 shards in the map.
4890 0 : return Err(ApiError::Conflict(
4891 0 : "Cannot split, currently mid-split".to_string(),
4892 0 : ));
4893 0 : }
4894 : }
4895 : }
4896 0 : if policy.is_none() {
4897 0 : policy = Some(shard.policy.clone());
4898 0 : }
4899 0 : if shard_ident.is_none() {
4900 0 : shard_ident = Some(shard.shard);
4901 0 : }
4902 0 : if config.is_none() {
4903 0 : config = Some(shard.config.clone());
4904 0 : }
4905 0 : if preferred_az_id.is_none() {
4906 0 : preferred_az_id = shard.preferred_az().cloned();
4907 0 : }
4908 :
4909 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4910 0 : tracing::info!(
4911 0 : "Tenant shard {} already has shard count {}",
4912 : tenant_shard_id,
4913 : split_req.new_shard_count
4914 : );
4915 0 : continue;
4916 0 : }
4917 :
4918 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4919 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4920 0 : ))?;
4921 :
4922 0 : let node = pageservers
4923 0 : .get(&node_id)
4924 0 : .expect("Pageservers may not be deleted while referenced");
4925 0 :
4926 0 : targets.push(ShardSplitTarget {
4927 0 : parent_id: *tenant_shard_id,
4928 0 : node: node.clone(),
4929 0 : child_ids: tenant_shard_id
4930 0 : .split(ShardCount::new(split_req.new_shard_count)),
4931 0 : });
4932 : }
4933 :
4934 0 : if targets.is_empty() {
4935 0 : if children_found.len() == split_req.new_shard_count as usize {
4936 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4937 0 : new_shards: children_found,
4938 0 : }));
4939 : } else {
4940 : // No shards found to split, and no existing children found: the
4941 : // tenant doesn't exist at all.
4942 0 : return Err(ApiError::NotFound(
4943 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4944 0 : ));
4945 : }
4946 0 : }
4947 0 :
4948 0 : (old_shard_count, targets)
4949 0 : };
4950 0 :
4951 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
4952 0 : let old_shard_count = old_shard_count.unwrap();
4953 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
4954 : // This ShardIdentity will be used as the template for all children, so this implicitly
4955 : // applies the new stripe size to the children.
4956 0 : let mut shard_ident = shard_ident.unwrap();
4957 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
4958 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)));
4959 0 : }
4960 0 :
4961 0 : shard_ident.stripe_size = new_stripe_size;
4962 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
4963 0 : shard_ident
4964 : } else {
4965 0 : shard_ident.unwrap()
4966 : };
4967 0 : let policy = policy.unwrap();
4968 0 : let config = config.unwrap();
4969 0 :
4970 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
4971 0 : old_shard_count,
4972 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
4973 0 : new_stripe_size: split_req.new_stripe_size,
4974 0 : targets,
4975 0 : policy,
4976 0 : config,
4977 0 : shard_ident,
4978 0 : preferred_az_id,
4979 0 : })))
4980 0 : }
4981 :
4982 0 : async fn do_tenant_shard_split(
4983 0 : &self,
4984 0 : tenant_id: TenantId,
4985 0 : params: Box<ShardSplitParams>,
4986 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
4987 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
4988 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
4989 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
4990 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
4991 0 : // (https://github.com/neondatabase/neon/issues/6676)
4992 0 :
4993 0 : let ShardSplitParams {
4994 0 : old_shard_count,
4995 0 : new_shard_count,
4996 0 : new_stripe_size,
4997 0 : mut targets,
4998 0 : policy,
4999 0 : config,
5000 0 : shard_ident,
5001 0 : preferred_az_id,
5002 0 : } = *params;
5003 :
5004 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
5005 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
5006 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
5007 : // at the time of split.
5008 0 : let waiters = {
5009 0 : let mut locked = self.inner.write().unwrap();
5010 0 : let mut waiters = Vec::new();
5011 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5012 0 : for target in &mut targets {
5013 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
5014 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5015 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5016 0 : "Shard {} not found",
5017 0 : target.parent_id
5018 0 : )));
5019 : };
5020 :
5021 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
5022 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5023 0 : return Err(ApiError::Conflict(format!(
5024 0 : "Shard {} unexpectedly rescheduled during split",
5025 0 : target.parent_id
5026 0 : )));
5027 0 : }
5028 0 :
5029 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
5030 0 : shard.intent.clear_secondary(scheduler);
5031 :
5032 : // Run Reconciler to execute detach fo secondary locations.
5033 0 : if let Some(waiter) =
5034 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5035 0 : {
5036 0 : waiters.push(waiter);
5037 0 : }
5038 : }
5039 0 : waiters
5040 0 : };
5041 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
5042 :
5043 : // Before creating any new child shards in memory or on the pageservers, persist them: this
5044 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
5045 : // acts as the protection against two concurrent attempts to split: one of them will get a database
5046 : // error trying to insert the child shards.
5047 0 : let mut child_tsps = Vec::new();
5048 0 : for target in &targets {
5049 0 : let mut this_child_tsps = Vec::new();
5050 0 : for child in &target.child_ids {
5051 0 : let mut child_shard = shard_ident;
5052 0 : child_shard.number = child.shard_number;
5053 0 : child_shard.count = child.shard_count;
5054 0 :
5055 0 : tracing::info!(
5056 0 : "Create child shard persistence with stripe size {}",
5057 : shard_ident.stripe_size.0
5058 : );
5059 :
5060 0 : this_child_tsps.push(TenantShardPersistence {
5061 0 : tenant_id: child.tenant_id.to_string(),
5062 0 : shard_number: child.shard_number.0 as i32,
5063 0 : shard_count: child.shard_count.literal() as i32,
5064 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
5065 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
5066 0 : // populate the correct generation as part of its transaction, to protect us
5067 0 : // against racing with changes in the state of the parent.
5068 0 : generation: None,
5069 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
5070 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
5071 0 : config: serde_json::to_string(&config).unwrap(),
5072 0 : splitting: SplitState::Splitting,
5073 0 :
5074 0 : // Scheduling policies and preferred AZ do not carry through to children
5075 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
5076 0 : .unwrap(),
5077 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
5078 0 : });
5079 0 : }
5080 :
5081 0 : child_tsps.push((target.parent_id, this_child_tsps));
5082 : }
5083 :
5084 0 : if let Err(e) = self
5085 0 : .persistence
5086 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
5087 0 : .await
5088 : {
5089 0 : match e {
5090 : DatabaseError::Query(diesel::result::Error::DatabaseError(
5091 : DatabaseErrorKind::UniqueViolation,
5092 : _,
5093 : )) => {
5094 : // Inserting a child shard violated a unique constraint: we raced with another call to
5095 : // this function
5096 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
5097 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
5098 : }
5099 0 : _ => return Err(ApiError::InternalServerError(e.into())),
5100 : }
5101 0 : }
5102 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
5103 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5104 0 : ));
5105 :
5106 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
5107 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
5108 : // is not set in memory, then it was not persisted.
5109 : {
5110 0 : let mut locked = self.inner.write().unwrap();
5111 0 : for target in &targets {
5112 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
5113 0 : parent_shard.splitting = SplitState::Splitting;
5114 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
5115 0 : // split operation.
5116 0 : parent_shard
5117 0 : .observed
5118 0 : .locations
5119 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
5120 0 : }
5121 : }
5122 : }
5123 :
5124 : // TODO: issue split calls concurrently (this only matters once we're splitting
5125 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
5126 :
5127 0 : for target in &targets {
5128 : let ShardSplitTarget {
5129 0 : parent_id,
5130 0 : node,
5131 0 : child_ids,
5132 0 : } = target;
5133 0 : let client = PageserverClient::new(
5134 0 : node.get_id(),
5135 0 : node.base_url(),
5136 0 : self.config.jwt_token.as_deref(),
5137 0 : );
5138 0 : let response = client
5139 0 : .tenant_shard_split(
5140 0 : *parent_id,
5141 0 : TenantShardSplitRequest {
5142 0 : new_shard_count: new_shard_count.literal(),
5143 0 : new_stripe_size,
5144 0 : },
5145 0 : )
5146 0 : .await
5147 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
5148 :
5149 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
5150 0 : "failpoint".to_string()
5151 0 : )));
5152 :
5153 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
5154 :
5155 0 : tracing::info!(
5156 0 : "Split {} into {}",
5157 0 : parent_id,
5158 0 : response
5159 0 : .new_shards
5160 0 : .iter()
5161 0 : .map(|s| format!("{:?}", s))
5162 0 : .collect::<Vec<_>>()
5163 0 : .join(",")
5164 : );
5165 :
5166 0 : if &response.new_shards != child_ids {
5167 : // This should never happen: the pageserver should agree with us on how shard splits work.
5168 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5169 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
5170 0 : parent_id,
5171 0 : response.new_shards,
5172 0 : child_ids
5173 0 : )));
5174 0 : }
5175 : }
5176 :
5177 : // TODO: if the pageserver restarted concurrently with our split API call,
5178 : // the actual generation of the child shard might differ from the generation
5179 : // we expect it to have. In order for our in-database generation to end up
5180 : // correct, we should carry the child generation back in the response and apply it here
5181 : // in complete_shard_split (and apply the correct generation in memory)
5182 : // (or, we can carry generation in the request and reject the request if
5183 : // it doesn't match, but that requires more retry logic on this side)
5184 :
5185 0 : self.persistence
5186 0 : .complete_shard_split(tenant_id, old_shard_count)
5187 0 : .await?;
5188 :
5189 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5190 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5191 0 : ));
5192 :
5193 : // Replace all the shards we just split with their children: this phase is infallible.
5194 0 : let (response, child_locations, waiters) =
5195 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5196 0 :
5197 0 : // Send compute notifications for all the new shards
5198 0 : let mut failed_notifications = Vec::new();
5199 0 : for (child_id, child_ps, stripe_size) in child_locations {
5200 0 : if let Err(e) = self
5201 0 : .compute_hook
5202 0 : .notify(
5203 0 : compute_hook::ShardUpdate {
5204 0 : tenant_shard_id: child_id,
5205 0 : node_id: child_ps,
5206 0 : stripe_size,
5207 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5208 0 : },
5209 0 : &self.cancel,
5210 0 : )
5211 0 : .await
5212 : {
5213 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5214 : child_id, child_ps);
5215 0 : failed_notifications.push(child_id);
5216 0 : }
5217 : }
5218 :
5219 : // If we failed any compute notifications, make a note to retry later.
5220 0 : if !failed_notifications.is_empty() {
5221 0 : let mut locked = self.inner.write().unwrap();
5222 0 : for failed in failed_notifications {
5223 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5224 0 : shard.pending_compute_notification = true;
5225 0 : }
5226 : }
5227 0 : }
5228 :
5229 0 : Ok((response, waiters))
5230 0 : }
5231 :
5232 0 : pub(crate) async fn tenant_shard_migrate(
5233 0 : &self,
5234 0 : tenant_shard_id: TenantShardId,
5235 0 : migrate_req: TenantShardMigrateRequest,
5236 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5237 0 : let waiter = {
5238 0 : let mut locked = self.inner.write().unwrap();
5239 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5240 :
5241 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5242 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5243 0 : "Node {} not found",
5244 0 : migrate_req.node_id
5245 0 : )));
5246 : };
5247 :
5248 0 : if !node.is_available() {
5249 : // Warn but proceed: the caller may intend to manually adjust the placement of
5250 : // a shard even if the node is down, e.g. if intervening during an incident.
5251 0 : tracing::warn!("Migrating to unavailable node {node}");
5252 0 : }
5253 :
5254 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5255 0 : return Err(ApiError::NotFound(
5256 0 : anyhow::anyhow!("Tenant shard not found").into(),
5257 0 : ));
5258 : };
5259 :
5260 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5261 : // No-op case: we will still proceed to wait for reconciliation in case it is
5262 : // incomplete from an earlier update to the intent.
5263 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
5264 : } else {
5265 0 : let old_attached = *shard.intent.get_attached();
5266 0 :
5267 0 : match shard.policy {
5268 0 : PlacementPolicy::Attached(n) => {
5269 0 : // If our new attached node was a secondary, it no longer should be.
5270 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
5271 0 :
5272 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
5273 :
5274 : // If we were already attached to something, demote that to a secondary
5275 0 : if let Some(old_attached) = old_attached {
5276 0 : if n > 0 {
5277 : // Remove other secondaries to make room for the location we'll demote
5278 0 : while shard.intent.get_secondary().len() >= n {
5279 0 : shard.intent.pop_secondary(scheduler);
5280 0 : }
5281 :
5282 0 : shard.intent.push_secondary(scheduler, old_attached);
5283 0 : }
5284 0 : }
5285 : }
5286 0 : PlacementPolicy::Secondary => {
5287 0 : shard.intent.clear(scheduler);
5288 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5289 0 : }
5290 : PlacementPolicy::Detached => {
5291 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5292 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
5293 0 : )))
5294 : }
5295 : }
5296 :
5297 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
5298 0 : shard.sequence = shard.sequence.next();
5299 : }
5300 :
5301 0 : let reconciler_config = match migrate_req.migration_config {
5302 0 : Some(cfg) => (&cfg).into(),
5303 0 : None => ReconcilerConfig::new(ReconcilerPriority::High),
5304 : };
5305 :
5306 0 : self.maybe_configured_reconcile_shard(shard, nodes, reconciler_config)
5307 : };
5308 :
5309 0 : if let Some(waiter) = waiter {
5310 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5311 : } else {
5312 0 : tracing::info!("Migration is a no-op");
5313 : }
5314 :
5315 0 : Ok(TenantShardMigrateResponse {})
5316 0 : }
5317 :
5318 0 : pub(crate) async fn tenant_shard_migrate_secondary(
5319 0 : &self,
5320 0 : tenant_shard_id: TenantShardId,
5321 0 : migrate_req: TenantShardMigrateRequest,
5322 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5323 0 : let waiter = {
5324 0 : let mut locked = self.inner.write().unwrap();
5325 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5326 :
5327 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5328 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5329 0 : "Node {} not found",
5330 0 : migrate_req.node_id
5331 0 : )));
5332 : };
5333 :
5334 0 : if !node.is_available() {
5335 : // Warn but proceed: the caller may intend to manually adjust the placement of
5336 : // a shard even if the node is down, e.g. if intervening during an incident.
5337 0 : tracing::warn!("Migrating to unavailable node {node}");
5338 0 : }
5339 :
5340 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5341 0 : return Err(ApiError::NotFound(
5342 0 : anyhow::anyhow!("Tenant shard not found").into(),
5343 0 : ));
5344 : };
5345 :
5346 0 : if shard.intent.get_secondary().len() == 1
5347 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
5348 : {
5349 0 : tracing::info!(
5350 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
5351 : shard.intent
5352 : );
5353 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5354 0 : tracing::info!("Migrating secondary to {node}: already attached where we were asked to create a secondary");
5355 : } else {
5356 0 : let old_secondaries = shard.intent.get_secondary().clone();
5357 0 : for secondary in old_secondaries {
5358 0 : shard.intent.remove_secondary(scheduler, secondary);
5359 0 : }
5360 :
5361 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5362 0 : shard.sequence = shard.sequence.next();
5363 0 : tracing::info!(
5364 0 : "Migrating secondary to {node}: new intent {:?}",
5365 : shard.intent
5366 : );
5367 : }
5368 :
5369 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5370 : };
5371 :
5372 0 : if let Some(waiter) = waiter {
5373 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5374 : } else {
5375 0 : tracing::info!("Migration is a no-op");
5376 : }
5377 :
5378 0 : Ok(TenantShardMigrateResponse {})
5379 0 : }
5380 :
5381 : /// 'cancel' in this context means cancel any ongoing reconcile
5382 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
5383 0 : &self,
5384 0 : tenant_shard_id: TenantShardId,
5385 0 : ) -> Result<(), ApiError> {
5386 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
5387 0 : let waiter = {
5388 0 : let locked = self.inner.write().unwrap();
5389 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5390 0 : return Err(ApiError::NotFound(
5391 0 : anyhow::anyhow!("Tenant shard not found").into(),
5392 0 : ));
5393 : };
5394 :
5395 0 : let waiter = shard.get_waiter();
5396 0 : match waiter {
5397 : None => {
5398 0 : tracing::info!("Shard does not have an ongoing Reconciler");
5399 0 : return Ok(());
5400 : }
5401 0 : Some(waiter) => {
5402 0 : tracing::info!("Cancelling Reconciler");
5403 0 : shard.cancel_reconciler();
5404 0 : waiter
5405 0 : }
5406 0 : }
5407 0 : };
5408 0 :
5409 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
5410 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
5411 0 : // the Reconciler is misbehaving and not respecting the cancellation token
5412 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
5413 0 : .await?;
5414 :
5415 0 : Ok(())
5416 0 : }
5417 :
5418 : /// This is for debug/support only: we simply drop all state for a tenant, without
5419 : /// detaching or deleting it on pageservers.
5420 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
5421 0 : self.persistence.delete_tenant(tenant_id).await?;
5422 :
5423 0 : let mut locked = self.inner.write().unwrap();
5424 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
5425 0 : let mut shards = Vec::new();
5426 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
5427 0 : shards.push(*tenant_shard_id);
5428 0 : }
5429 :
5430 0 : for shard_id in shards {
5431 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
5432 0 : shard.intent.clear(scheduler);
5433 0 : }
5434 : }
5435 :
5436 0 : Ok(())
5437 0 : }
5438 :
5439 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
5440 : /// tenant with a very high generation number so that it will see the existing data.
5441 0 : pub(crate) async fn tenant_import(
5442 0 : &self,
5443 0 : tenant_id: TenantId,
5444 0 : ) -> Result<TenantCreateResponse, ApiError> {
5445 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
5446 0 : let maybe_node = {
5447 0 : self.inner
5448 0 : .read()
5449 0 : .unwrap()
5450 0 : .nodes
5451 0 : .values()
5452 0 : .find(|n| n.is_available())
5453 0 : .cloned()
5454 : };
5455 0 : let Some(node) = maybe_node else {
5456 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
5457 : };
5458 :
5459 0 : let client = PageserverClient::new(
5460 0 : node.get_id(),
5461 0 : node.base_url(),
5462 0 : self.config.jwt_token.as_deref(),
5463 0 : );
5464 :
5465 0 : let scan_result = client
5466 0 : .tenant_scan_remote_storage(tenant_id)
5467 0 : .await
5468 0 : .map_err(|e| passthrough_api_error(&node, e))?;
5469 :
5470 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
5471 0 : let Some(shard_count) = scan_result
5472 0 : .shards
5473 0 : .iter()
5474 0 : .map(|s| s.tenant_shard_id.shard_count)
5475 0 : .max()
5476 : else {
5477 0 : return Err(ApiError::NotFound(
5478 0 : anyhow::anyhow!("No shards found").into(),
5479 0 : ));
5480 : };
5481 :
5482 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
5483 : // to
5484 0 : let generation = scan_result
5485 0 : .shards
5486 0 : .iter()
5487 0 : .map(|s| s.generation)
5488 0 : .max()
5489 0 : .expect("We already validated >0 shards");
5490 0 :
5491 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
5492 0 : // only work if they were using the default stripe size.
5493 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
5494 :
5495 0 : let (response, waiters) = self
5496 0 : .do_tenant_create(TenantCreateRequest {
5497 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
5498 0 : generation,
5499 0 :
5500 0 : shard_parameters: ShardParameters {
5501 0 : count: shard_count,
5502 0 : stripe_size,
5503 0 : },
5504 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
5505 0 : config: TenantConfig::default(),
5506 0 : })
5507 0 : .await?;
5508 :
5509 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
5510 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
5511 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
5512 : // reconcile, as reconciliation includes notifying compute.
5513 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
5514 0 : }
5515 :
5516 0 : Ok(response)
5517 0 : }
5518 :
5519 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
5520 : /// we don't have to make TenantShard clonable in the return path.
5521 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5522 0 : let serialized = {
5523 0 : let locked = self.inner.read().unwrap();
5524 0 : let result = locked.tenants.values().collect::<Vec<_>>();
5525 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
5526 : };
5527 :
5528 0 : hyper::Response::builder()
5529 0 : .status(hyper::StatusCode::OK)
5530 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5531 0 : .body(hyper::Body::from(serialized))
5532 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5533 0 : }
5534 :
5535 : /// Check the consistency of in-memory state vs. persistent state, and check that the
5536 : /// scheduler's statistics are up to date.
5537 : ///
5538 : /// These consistency checks expect an **idle** system. If changes are going on while
5539 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
5540 : /// checks, but not suitable for running continuously in the background in the field.
5541 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
5542 0 : let (mut expect_nodes, mut expect_shards) = {
5543 0 : let locked = self.inner.read().unwrap();
5544 0 :
5545 0 : locked
5546 0 : .scheduler
5547 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
5548 0 : .context("Scheduler checks")
5549 0 : .map_err(ApiError::InternalServerError)?;
5550 :
5551 0 : let expect_nodes = locked
5552 0 : .nodes
5553 0 : .values()
5554 0 : .map(|n| n.to_persistent())
5555 0 : .collect::<Vec<_>>();
5556 0 :
5557 0 : let expect_shards = locked
5558 0 : .tenants
5559 0 : .values()
5560 0 : .map(|t| t.to_persistent())
5561 0 : .collect::<Vec<_>>();
5562 :
5563 : // This method can only validate the state of an idle system: if a reconcile is in
5564 : // progress, fail out early to avoid giving false errors on state that won't match
5565 : // between database and memory under a ReconcileResult is processed.
5566 0 : for t in locked.tenants.values() {
5567 0 : if t.reconciler.is_some() {
5568 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5569 0 : "Shard {} reconciliation in progress",
5570 0 : t.tenant_shard_id
5571 0 : )));
5572 0 : }
5573 : }
5574 :
5575 0 : (expect_nodes, expect_shards)
5576 : };
5577 :
5578 0 : let mut nodes = self.persistence.list_nodes().await?;
5579 0 : expect_nodes.sort_by_key(|n| n.node_id);
5580 0 : nodes.sort_by_key(|n| n.node_id);
5581 :
5582 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
5583 0 : let node_result = if nodes != expect_nodes {
5584 0 : tracing::error!("Consistency check failed on nodes.");
5585 0 : tracing::error!(
5586 0 : "Nodes in memory: {}",
5587 0 : serde_json::to_string(&expect_nodes)
5588 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5589 : );
5590 0 : tracing::error!(
5591 0 : "Nodes in database: {}",
5592 0 : serde_json::to_string(&nodes)
5593 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5594 : );
5595 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
5596 0 : "Node consistency failure"
5597 0 : )))
5598 : } else {
5599 0 : Ok(())
5600 : };
5601 :
5602 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
5603 0 : persistent_shards
5604 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5605 0 :
5606 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5607 :
5608 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
5609 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
5610 : // fields are added, before doing a comparison.
5611 0 : for tsp in &mut persistent_shards {
5612 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
5613 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
5614 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
5615 : }
5616 :
5617 0 : if persistent_shards != expect_shards {
5618 0 : tracing::error!("Consistency check failed on shards.");
5619 :
5620 0 : tracing::error!(
5621 0 : "Shards in memory: {}",
5622 0 : serde_json::to_string(&expect_shards)
5623 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5624 : );
5625 0 : tracing::error!(
5626 0 : "Shards in database: {}",
5627 0 : serde_json::to_string(&persistent_shards)
5628 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5629 : );
5630 :
5631 : // The total dump log lines above are useful in testing but in the field grafana will
5632 : // usually just drop them because they're so large. So we also do some explicit logging
5633 : // of just the diffs.
5634 0 : let persistent_shards = persistent_shards
5635 0 : .into_iter()
5636 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5637 0 : .collect::<HashMap<_, _>>();
5638 0 : let expect_shards = expect_shards
5639 0 : .into_iter()
5640 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5641 0 : .collect::<HashMap<_, _>>();
5642 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
5643 0 : match expect_shards.get(tenant_shard_id) {
5644 : None => {
5645 0 : tracing::error!(
5646 0 : "Shard {} found in database but not in memory",
5647 : tenant_shard_id
5648 : );
5649 : }
5650 0 : Some(expect_tsp) => {
5651 0 : if expect_tsp != persistent_tsp {
5652 0 : tracing::error!(
5653 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
5654 0 : tenant_shard_id,
5655 0 : serde_json::to_string(expect_tsp).unwrap(),
5656 0 : serde_json::to_string(&persistent_tsp).unwrap()
5657 : );
5658 0 : }
5659 : }
5660 : }
5661 : }
5662 :
5663 : // Having already logged any differences, log any shards that simply aren't present in the database
5664 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
5665 0 : if !persistent_shards.contains_key(tenant_shard_id) {
5666 0 : tracing::error!(
5667 0 : "Shard {} found in memory but not in database: {}",
5668 0 : tenant_shard_id,
5669 0 : serde_json::to_string(memory_tsp)
5670 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5671 : );
5672 0 : }
5673 : }
5674 :
5675 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5676 0 : "Shard consistency failure"
5677 0 : )));
5678 0 : }
5679 0 :
5680 0 : node_result
5681 0 : }
5682 :
5683 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
5684 : /// we don't have to make TenantShard clonable in the return path.
5685 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5686 0 : let serialized = {
5687 0 : let locked = self.inner.read().unwrap();
5688 0 : serde_json::to_string(&locked.scheduler)
5689 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5690 : };
5691 :
5692 0 : hyper::Response::builder()
5693 0 : .status(hyper::StatusCode::OK)
5694 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5695 0 : .body(hyper::Body::from(serialized))
5696 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5697 0 : }
5698 :
5699 : /// This is for debug/support only: we simply drop all state for a tenant, without
5700 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
5701 : /// tenants that were on this node.
5702 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
5703 0 : self.persistence.delete_node(node_id).await?;
5704 :
5705 0 : let mut locked = self.inner.write().unwrap();
5706 :
5707 0 : for shard in locked.tenants.values_mut() {
5708 0 : shard.deref_node(node_id);
5709 0 : shard.observed.locations.remove(&node_id);
5710 0 : }
5711 :
5712 0 : let mut nodes = (*locked.nodes).clone();
5713 0 : nodes.remove(&node_id);
5714 0 : locked.nodes = Arc::new(nodes);
5715 0 : metrics::METRICS_REGISTRY
5716 0 : .metrics_group
5717 0 : .storage_controller_pageserver_nodes
5718 0 : .set(locked.nodes.len() as i64);
5719 0 :
5720 0 : locked.scheduler.node_remove(node_id);
5721 0 :
5722 0 : Ok(())
5723 0 : }
5724 :
5725 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
5726 : /// that we don't leave any bad state behind in the storage controller, but unclean
5727 : /// in the sense that we are not carefully draining the node.
5728 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
5729 0 : let _node_lock =
5730 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
5731 :
5732 : // 1. Atomically update in-memory state:
5733 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
5734 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
5735 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
5736 : // re-insert references to this node into the ObservedState of shards
5737 : // - drop the node from the scheduler
5738 : {
5739 0 : let mut locked = self.inner.write().unwrap();
5740 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5741 0 :
5742 0 : {
5743 0 : let mut nodes_mut = (*nodes).deref().clone();
5744 0 : match nodes_mut.get_mut(&node_id) {
5745 0 : Some(node) => {
5746 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
5747 0 : // if we crash it would not be desirable to leave the node paused after a restart.
5748 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
5749 0 : }
5750 : None => {
5751 0 : tracing::info!(
5752 0 : "Node not found: presuming this is a retry and returning success"
5753 : );
5754 0 : return Ok(());
5755 : }
5756 : }
5757 :
5758 0 : *nodes = Arc::new(nodes_mut);
5759 : }
5760 :
5761 0 : for (_tenant_id, mut schedule_context, shards) in
5762 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
5763 : {
5764 0 : for shard in shards {
5765 0 : if shard.deref_node(node_id) {
5766 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
5767 : // TODO: implement force flag to remove a node even if we can't reschedule
5768 : // a tenant
5769 0 : tracing::error!(
5770 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
5771 : shard.tenant_shard_id
5772 : );
5773 0 : return Err(e.into());
5774 : } else {
5775 0 : tracing::info!(
5776 0 : "Rescheduled shard {} away from node during deletion",
5777 : shard.tenant_shard_id
5778 : )
5779 : }
5780 :
5781 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
5782 0 : }
5783 :
5784 : // Here we remove an existing observed location for the node we're removing, and it will
5785 : // not be re-added by a reconciler's completion because we filter out removed nodes in
5786 : // process_result.
5787 : //
5788 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
5789 : // means any reconciles we spawned will know about the node we're deleting, enabling them
5790 : // to do live migrations if it's still online.
5791 0 : shard.observed.locations.remove(&node_id);
5792 : }
5793 : }
5794 :
5795 0 : scheduler.node_remove(node_id);
5796 0 :
5797 0 : {
5798 0 : let mut nodes_mut = (**nodes).clone();
5799 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
5800 0 : // Ensure that any reconciler holding an Arc<> to this node will
5801 0 : // drop out when trying to RPC to it (setting Offline state sets the
5802 0 : // cancellation token on the Node object).
5803 0 : removed_node.set_availability(NodeAvailability::Offline);
5804 0 : }
5805 0 : *nodes = Arc::new(nodes_mut);
5806 0 : metrics::METRICS_REGISTRY
5807 0 : .metrics_group
5808 0 : .storage_controller_pageserver_nodes
5809 0 : .set(nodes.len() as i64);
5810 0 : }
5811 0 : }
5812 0 :
5813 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
5814 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
5815 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
5816 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
5817 0 : // that exists.
5818 0 :
5819 0 : // 2. Actually delete the node from the database and from in-memory state
5820 0 : tracing::info!("Deleting node from database");
5821 0 : self.persistence.delete_node(node_id).await?;
5822 :
5823 0 : Ok(())
5824 0 : }
5825 :
5826 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
5827 0 : let nodes = {
5828 0 : self.inner
5829 0 : .read()
5830 0 : .unwrap()
5831 0 : .nodes
5832 0 : .values()
5833 0 : .cloned()
5834 0 : .collect::<Vec<_>>()
5835 0 : };
5836 0 :
5837 0 : Ok(nodes)
5838 0 : }
5839 :
5840 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
5841 0 : self.inner
5842 0 : .read()
5843 0 : .unwrap()
5844 0 : .nodes
5845 0 : .get(&node_id)
5846 0 : .cloned()
5847 0 : .ok_or(ApiError::NotFound(
5848 0 : format!("Node {node_id} not registered").into(),
5849 0 : ))
5850 0 : }
5851 :
5852 0 : pub(crate) async fn get_node_shards(
5853 0 : &self,
5854 0 : node_id: NodeId,
5855 0 : ) -> Result<NodeShardResponse, ApiError> {
5856 0 : let locked = self.inner.read().unwrap();
5857 0 : let mut shards = Vec::new();
5858 0 : for (tid, tenant) in locked.tenants.iter() {
5859 0 : let is_intended_secondary = match (
5860 0 : tenant.intent.get_attached() == &Some(node_id),
5861 0 : tenant.intent.get_secondary().contains(&node_id),
5862 0 : ) {
5863 : (true, true) => {
5864 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5865 0 : "{} attached as primary+secondary on the same node",
5866 0 : tid
5867 0 : )))
5868 : }
5869 0 : (true, false) => Some(false),
5870 0 : (false, true) => Some(true),
5871 0 : (false, false) => None,
5872 : };
5873 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
5874 0 : tenant.observed.locations.get(&node_id)
5875 : {
5876 0 : Some(conf.secondary_conf.is_some())
5877 : } else {
5878 0 : None
5879 : };
5880 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
5881 0 : shards.push(NodeShard {
5882 0 : tenant_shard_id: *tid,
5883 0 : is_intended_secondary,
5884 0 : is_observed_secondary,
5885 0 : });
5886 0 : }
5887 : }
5888 0 : Ok(NodeShardResponse { node_id, shards })
5889 0 : }
5890 :
5891 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
5892 0 : self.persistence.get_leader().await
5893 0 : }
5894 :
5895 0 : pub(crate) async fn node_register(
5896 0 : &self,
5897 0 : register_req: NodeRegisterRequest,
5898 0 : ) -> Result<(), ApiError> {
5899 0 : let _node_lock = trace_exclusive_lock(
5900 0 : &self.node_op_locks,
5901 0 : register_req.node_id,
5902 0 : NodeOperations::Register,
5903 0 : )
5904 0 : .await;
5905 :
5906 : enum RegistrationStatus {
5907 : Matched,
5908 : Mismatched,
5909 : New,
5910 : }
5911 :
5912 0 : let registration_status = {
5913 0 : let locked = self.inner.read().unwrap();
5914 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
5915 0 : if node.registration_match(®ister_req) {
5916 0 : RegistrationStatus::Matched
5917 : } else {
5918 0 : RegistrationStatus::Mismatched
5919 : }
5920 : } else {
5921 0 : RegistrationStatus::New
5922 : }
5923 : };
5924 :
5925 0 : match registration_status {
5926 : RegistrationStatus::Matched => {
5927 0 : tracing::info!(
5928 0 : "Node {} re-registered with matching address",
5929 : register_req.node_id
5930 : );
5931 :
5932 0 : return Ok(());
5933 : }
5934 : RegistrationStatus::Mismatched => {
5935 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
5936 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
5937 : // a fixed address through the lifetime of a node.
5938 0 : tracing::warn!(
5939 0 : "Node {} tried to register with different address",
5940 : register_req.node_id
5941 : );
5942 0 : return Err(ApiError::Conflict(
5943 0 : "Node is already registered with different address".to_string(),
5944 0 : ));
5945 : }
5946 0 : RegistrationStatus::New => {
5947 0 : // fallthrough
5948 0 : }
5949 0 : }
5950 0 :
5951 0 : // We do not require that a node is actually online when registered (it will start life
5952 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
5953 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
5954 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
5955 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
5956 0 : if tokio::net::lookup_host(format!(
5957 0 : "{}:{}",
5958 0 : register_req.listen_http_addr, register_req.listen_http_port
5959 0 : ))
5960 0 : .await
5961 0 : .is_err()
5962 : {
5963 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
5964 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
5965 : // we return a soft 503 error, to encourage callers to retry past transient issues.
5966 0 : return Err(ApiError::ResourceUnavailable(
5967 0 : format!(
5968 0 : "Node {} tried to register with unknown DNS name '{}'",
5969 0 : register_req.node_id, register_req.listen_http_addr
5970 0 : )
5971 0 : .into(),
5972 0 : ));
5973 0 : }
5974 0 :
5975 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
5976 0 : // This ensures that before we use it for anything or expose it via any external
5977 0 : // API, it is guaranteed to be available after a restart.
5978 0 : let new_node = Node::new(
5979 0 : register_req.node_id,
5980 0 : register_req.listen_http_addr,
5981 0 : register_req.listen_http_port,
5982 0 : register_req.listen_pg_addr,
5983 0 : register_req.listen_pg_port,
5984 0 : register_req.availability_zone_id.clone(),
5985 0 : );
5986 0 :
5987 0 : // TODO: idempotency if the node already exists in the database
5988 0 : self.persistence.insert_node(&new_node).await?;
5989 :
5990 0 : let mut locked = self.inner.write().unwrap();
5991 0 : let mut new_nodes = (*locked.nodes).clone();
5992 0 :
5993 0 : locked.scheduler.node_upsert(&new_node);
5994 0 : new_nodes.insert(register_req.node_id, new_node);
5995 0 :
5996 0 : locked.nodes = Arc::new(new_nodes);
5997 0 :
5998 0 : metrics::METRICS_REGISTRY
5999 0 : .metrics_group
6000 0 : .storage_controller_pageserver_nodes
6001 0 : .set(locked.nodes.len() as i64);
6002 0 :
6003 0 : tracing::info!(
6004 0 : "Registered pageserver {} ({}), now have {} pageservers",
6005 0 : register_req.node_id,
6006 0 : register_req.availability_zone_id,
6007 0 : locked.nodes.len()
6008 : );
6009 0 : Ok(())
6010 0 : }
6011 :
6012 : /// Configure in-memory and persistent state of a node as requested
6013 : ///
6014 : /// Note that this function does not trigger any immediate side effects in response
6015 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
6016 0 : async fn node_state_configure(
6017 0 : &self,
6018 0 : node_id: NodeId,
6019 0 : availability: Option<NodeAvailability>,
6020 0 : scheduling: Option<NodeSchedulingPolicy>,
6021 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
6022 0 : ) -> Result<AvailabilityTransition, ApiError> {
6023 0 : if let Some(scheduling) = scheduling {
6024 : // Scheduling is a persistent part of Node: we must write updates to the database before
6025 : // applying them in memory
6026 0 : self.persistence.update_node(node_id, scheduling).await?;
6027 0 : }
6028 :
6029 : // If we're activating a node, then before setting it active we must reconcile any shard locations
6030 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
6031 : // by calling [`Self::node_activate_reconcile`]
6032 : //
6033 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
6034 : // nothing else can mutate its availability while we run.
6035 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
6036 0 : let (activate_node, availability_transition) = {
6037 0 : let locked = self.inner.read().unwrap();
6038 0 : let Some(node) = locked.nodes.get(&node_id) else {
6039 0 : return Err(ApiError::NotFound(
6040 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6041 0 : ));
6042 : };
6043 :
6044 0 : (
6045 0 : node.clone(),
6046 0 : node.get_availability_transition(input_availability),
6047 0 : )
6048 : };
6049 :
6050 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
6051 0 : self.node_activate_reconcile(activate_node, node_lock)
6052 0 : .await?;
6053 0 : }
6054 0 : availability_transition
6055 : } else {
6056 0 : AvailabilityTransition::Unchanged
6057 : };
6058 :
6059 : // Apply changes from the request to our in-memory state for the Node
6060 0 : let mut locked = self.inner.write().unwrap();
6061 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
6062 0 :
6063 0 : let mut new_nodes = (**nodes).clone();
6064 :
6065 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
6066 0 : return Err(ApiError::NotFound(
6067 0 : anyhow::anyhow!("Node not registered").into(),
6068 0 : ));
6069 : };
6070 :
6071 0 : if let Some(availability) = availability {
6072 0 : node.set_availability(availability);
6073 0 : }
6074 :
6075 0 : if let Some(scheduling) = scheduling {
6076 0 : node.set_scheduling(scheduling);
6077 0 : }
6078 :
6079 : // Update the scheduler, in case the elegibility of the node for new shards has changed
6080 0 : scheduler.node_upsert(node);
6081 0 :
6082 0 : let new_nodes = Arc::new(new_nodes);
6083 0 : locked.nodes = new_nodes;
6084 0 :
6085 0 : Ok(availability_transition)
6086 0 : }
6087 :
6088 : /// Handle availability transition of one node
6089 : ///
6090 : /// Note that you should first call [`Self::node_state_configure`] to update
6091 : /// the in-memory state referencing that node. If you need to handle more than one transition
6092 : /// consider using [`Self::handle_node_availability_transitions`].
6093 0 : async fn handle_node_availability_transition(
6094 0 : &self,
6095 0 : node_id: NodeId,
6096 0 : transition: AvailabilityTransition,
6097 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
6098 0 : ) -> Result<(), ApiError> {
6099 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
6100 0 : match transition {
6101 : AvailabilityTransition::ToOffline => {
6102 0 : tracing::info!("Node {} transition to offline", node_id);
6103 :
6104 0 : let mut locked = self.inner.write().unwrap();
6105 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6106 0 :
6107 0 : let mut tenants_affected: usize = 0;
6108 :
6109 0 : for (_tenant_id, mut schedule_context, shards) in
6110 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6111 : {
6112 0 : for tenant_shard in shards {
6113 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
6114 0 : if let Some(observed_loc) =
6115 0 : tenant_shard.observed.locations.get_mut(&node_id)
6116 0 : {
6117 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
6118 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
6119 0 : observed_loc.conf = None;
6120 0 : }
6121 :
6122 0 : if nodes.len() == 1 {
6123 : // Special case for single-node cluster: there is no point trying to reschedule
6124 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
6125 : // failures to schedule them.
6126 0 : continue;
6127 0 : }
6128 0 :
6129 0 : if !nodes
6130 0 : .values()
6131 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6132 : {
6133 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
6134 : // trying to reschedule since there's nowhere else to go. Without this
6135 : // branch we incorrectly detach tenants in response to node unavailability.
6136 0 : continue;
6137 0 : }
6138 0 :
6139 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
6140 0 : tenant_shard.sequence = tenant_shard.sequence.next();
6141 0 :
6142 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
6143 0 : Err(e) => {
6144 0 : // It is possible that some tenants will become unschedulable when too many pageservers
6145 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
6146 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
6147 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
6148 : }
6149 : Ok(()) => {
6150 0 : if self
6151 0 : .maybe_reconcile_shard(
6152 0 : tenant_shard,
6153 0 : nodes,
6154 0 : ReconcilerPriority::Normal,
6155 0 : )
6156 0 : .is_some()
6157 0 : {
6158 0 : tenants_affected += 1;
6159 0 : };
6160 : }
6161 : }
6162 0 : }
6163 : }
6164 : }
6165 0 : tracing::info!(
6166 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
6167 : tenants_affected,
6168 : node_id
6169 : )
6170 : }
6171 : AvailabilityTransition::ToActive => {
6172 0 : tracing::info!("Node {} transition to active", node_id);
6173 :
6174 0 : let mut locked = self.inner.write().unwrap();
6175 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
6176 :
6177 : // When a node comes back online, we must reconcile any tenant that has a None observed
6178 : // location on the node.
6179 0 : for tenant_shard in tenants.values_mut() {
6180 : // If a reconciliation is already in progress, rely on the previous scheduling
6181 : // decision and skip triggering a new reconciliation.
6182 0 : if tenant_shard.reconciler.is_some() {
6183 0 : continue;
6184 0 : }
6185 :
6186 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
6187 0 : if observed_loc.conf.is_none() {
6188 0 : self.maybe_reconcile_shard(
6189 0 : tenant_shard,
6190 0 : nodes,
6191 0 : ReconcilerPriority::Normal,
6192 0 : );
6193 0 : }
6194 0 : }
6195 : }
6196 :
6197 : // TODO: in the background, we should balance work back onto this pageserver
6198 : }
6199 : // No action required for the intermediate unavailable state.
6200 : // When we transition into active or offline from the unavailable state,
6201 : // the correct handling above will kick in.
6202 : AvailabilityTransition::ToWarmingUpFromActive => {
6203 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
6204 : }
6205 : AvailabilityTransition::ToWarmingUpFromOffline => {
6206 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
6207 : }
6208 : AvailabilityTransition::Unchanged => {
6209 0 : tracing::debug!("Node {} no availability change during config", node_id);
6210 : }
6211 : }
6212 :
6213 0 : Ok(())
6214 0 : }
6215 :
6216 : /// Handle availability transition for multiple nodes
6217 : ///
6218 : /// Note that you should first call [`Self::node_state_configure`] for
6219 : /// all nodes being handled here for the handling to use fresh in-memory state.
6220 0 : async fn handle_node_availability_transitions(
6221 0 : &self,
6222 0 : transitions: Vec<(
6223 0 : NodeId,
6224 0 : TracingExclusiveGuard<NodeOperations>,
6225 0 : AvailabilityTransition,
6226 0 : )>,
6227 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
6228 0 : let mut errors = Vec::default();
6229 0 : for (node_id, node_lock, transition) in transitions {
6230 0 : let res = self
6231 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
6232 0 : .await;
6233 0 : if let Err(err) = res {
6234 0 : errors.push((node_id, err));
6235 0 : }
6236 : }
6237 :
6238 0 : if errors.is_empty() {
6239 0 : Ok(())
6240 : } else {
6241 0 : Err(errors)
6242 : }
6243 0 : }
6244 :
6245 0 : pub(crate) async fn node_configure(
6246 0 : &self,
6247 0 : node_id: NodeId,
6248 0 : availability: Option<NodeAvailability>,
6249 0 : scheduling: Option<NodeSchedulingPolicy>,
6250 0 : ) -> Result<(), ApiError> {
6251 0 : let node_lock =
6252 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
6253 :
6254 0 : let transition = self
6255 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
6256 0 : .await?;
6257 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
6258 0 : .await
6259 0 : }
6260 :
6261 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
6262 : /// operation for HTTP api.
6263 0 : pub(crate) async fn external_node_configure(
6264 0 : &self,
6265 0 : node_id: NodeId,
6266 0 : availability: Option<NodeAvailability>,
6267 0 : scheduling: Option<NodeSchedulingPolicy>,
6268 0 : ) -> Result<(), ApiError> {
6269 0 : {
6270 0 : let locked = self.inner.read().unwrap();
6271 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
6272 0 : return Err(ApiError::PreconditionFailed(
6273 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
6274 0 : ));
6275 0 : }
6276 0 : }
6277 0 :
6278 0 : self.node_configure(node_id, availability, scheduling).await
6279 0 : }
6280 :
6281 0 : pub(crate) async fn start_node_drain(
6282 0 : self: &Arc<Self>,
6283 0 : node_id: NodeId,
6284 0 : ) -> Result<(), ApiError> {
6285 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
6286 0 : let locked = self.inner.read().unwrap();
6287 0 : let nodes = &locked.nodes;
6288 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6289 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6290 0 : ))?;
6291 0 : let schedulable_nodes_count = nodes
6292 0 : .iter()
6293 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6294 0 : .count();
6295 0 :
6296 0 : (
6297 0 : locked
6298 0 : .ongoing_operation
6299 0 : .as_ref()
6300 0 : .map(|ongoing| ongoing.operation),
6301 0 : node.is_available(),
6302 0 : node.get_scheduling(),
6303 0 : schedulable_nodes_count,
6304 0 : )
6305 0 : };
6306 :
6307 0 : if let Some(ongoing) = ongoing_op {
6308 0 : return Err(ApiError::PreconditionFailed(
6309 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6310 0 : ));
6311 0 : }
6312 0 :
6313 0 : if !node_available {
6314 0 : return Err(ApiError::ResourceUnavailable(
6315 0 : format!("Node {node_id} is currently unavailable").into(),
6316 0 : ));
6317 0 : }
6318 0 :
6319 0 : if schedulable_nodes_count == 0 {
6320 0 : return Err(ApiError::PreconditionFailed(
6321 0 : "No other schedulable nodes to drain to".into(),
6322 0 : ));
6323 0 : }
6324 0 :
6325 0 : match node_policy {
6326 : NodeSchedulingPolicy::Active => {
6327 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
6328 0 : .await?;
6329 :
6330 0 : let cancel = self.cancel.child_token();
6331 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6332 :
6333 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6334 0 : operation: Operation::Drain(Drain { node_id }),
6335 0 : cancel: cancel.clone(),
6336 0 : });
6337 :
6338 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
6339 :
6340 0 : tokio::task::spawn({
6341 0 : let service = self.clone();
6342 0 : let cancel = cancel.clone();
6343 0 : async move {
6344 0 : let _gate_guard = gate_guard;
6345 0 :
6346 0 : scopeguard::defer! {
6347 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6348 0 :
6349 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
6350 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
6351 0 : } else {
6352 0 : panic!("We always remove the same operation")
6353 0 : }
6354 0 : }
6355 0 :
6356 0 : tracing::info!("Drain background operation starting");
6357 0 : let res = service.drain_node(node_id, cancel).await;
6358 0 : match res {
6359 : Ok(()) => {
6360 0 : tracing::info!("Drain background operation completed successfully");
6361 : }
6362 : Err(OperationError::Cancelled) => {
6363 0 : tracing::info!("Drain background operation was cancelled");
6364 : }
6365 0 : Err(err) => {
6366 0 : tracing::error!("Drain background operation encountered: {err}")
6367 : }
6368 : }
6369 0 : }
6370 0 : }.instrument(span));
6371 0 : }
6372 : NodeSchedulingPolicy::Draining => {
6373 0 : return Err(ApiError::Conflict(format!(
6374 0 : "Node {node_id} has drain in progress"
6375 0 : )));
6376 : }
6377 0 : policy => {
6378 0 : return Err(ApiError::PreconditionFailed(
6379 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
6380 0 : ));
6381 : }
6382 : }
6383 :
6384 0 : Ok(())
6385 0 : }
6386 :
6387 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
6388 0 : let node_available = {
6389 0 : let locked = self.inner.read().unwrap();
6390 0 : let nodes = &locked.nodes;
6391 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6392 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6393 0 : ))?;
6394 :
6395 0 : node.is_available()
6396 0 : };
6397 0 :
6398 0 : if !node_available {
6399 0 : return Err(ApiError::ResourceUnavailable(
6400 0 : format!("Node {node_id} is currently unavailable").into(),
6401 0 : ));
6402 0 : }
6403 :
6404 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6405 0 : if let Operation::Drain(drain) = op_handler.operation {
6406 0 : if drain.node_id == node_id {
6407 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6408 0 : op_handler.cancel.cancel();
6409 0 : return Ok(());
6410 0 : }
6411 0 : }
6412 0 : }
6413 :
6414 0 : Err(ApiError::PreconditionFailed(
6415 0 : format!("Node {node_id} has no drain in progress").into(),
6416 0 : ))
6417 0 : }
6418 :
6419 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
6420 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
6421 0 : let locked = self.inner.read().unwrap();
6422 0 : let nodes = &locked.nodes;
6423 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6424 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6425 0 : ))?;
6426 :
6427 0 : (
6428 0 : locked
6429 0 : .ongoing_operation
6430 0 : .as_ref()
6431 0 : .map(|ongoing| ongoing.operation),
6432 0 : node.is_available(),
6433 0 : node.get_scheduling(),
6434 0 : nodes.len(),
6435 0 : )
6436 0 : };
6437 :
6438 0 : if let Some(ongoing) = ongoing_op {
6439 0 : return Err(ApiError::PreconditionFailed(
6440 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6441 0 : ));
6442 0 : }
6443 0 :
6444 0 : if !node_available {
6445 0 : return Err(ApiError::ResourceUnavailable(
6446 0 : format!("Node {node_id} is currently unavailable").into(),
6447 0 : ));
6448 0 : }
6449 0 :
6450 0 : if total_nodes_count <= 1 {
6451 0 : return Err(ApiError::PreconditionFailed(
6452 0 : "No other nodes to fill from".into(),
6453 0 : ));
6454 0 : }
6455 0 :
6456 0 : match node_policy {
6457 : NodeSchedulingPolicy::Active => {
6458 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
6459 0 : .await?;
6460 :
6461 0 : let cancel = self.cancel.child_token();
6462 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6463 :
6464 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6465 0 : operation: Operation::Fill(Fill { node_id }),
6466 0 : cancel: cancel.clone(),
6467 0 : });
6468 :
6469 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
6470 :
6471 0 : tokio::task::spawn({
6472 0 : let service = self.clone();
6473 0 : let cancel = cancel.clone();
6474 0 : async move {
6475 0 : let _gate_guard = gate_guard;
6476 0 :
6477 0 : scopeguard::defer! {
6478 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6479 0 :
6480 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
6481 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
6482 0 : } else {
6483 0 : panic!("We always remove the same operation")
6484 0 : }
6485 0 : }
6486 0 :
6487 0 : tracing::info!("Fill background operation starting");
6488 0 : let res = service.fill_node(node_id, cancel).await;
6489 0 : match res {
6490 : Ok(()) => {
6491 0 : tracing::info!("Fill background operation completed successfully");
6492 : }
6493 : Err(OperationError::Cancelled) => {
6494 0 : tracing::info!("Fill background operation was cancelled");
6495 : }
6496 0 : Err(err) => {
6497 0 : tracing::error!("Fill background operation encountered: {err}")
6498 : }
6499 : }
6500 0 : }
6501 0 : }.instrument(span));
6502 0 : }
6503 : NodeSchedulingPolicy::Filling => {
6504 0 : return Err(ApiError::Conflict(format!(
6505 0 : "Node {node_id} has fill in progress"
6506 0 : )));
6507 : }
6508 0 : policy => {
6509 0 : return Err(ApiError::PreconditionFailed(
6510 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
6511 0 : ));
6512 : }
6513 : }
6514 :
6515 0 : Ok(())
6516 0 : }
6517 :
6518 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
6519 0 : let node_available = {
6520 0 : let locked = self.inner.read().unwrap();
6521 0 : let nodes = &locked.nodes;
6522 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6523 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6524 0 : ))?;
6525 :
6526 0 : node.is_available()
6527 0 : };
6528 0 :
6529 0 : if !node_available {
6530 0 : return Err(ApiError::ResourceUnavailable(
6531 0 : format!("Node {node_id} is currently unavailable").into(),
6532 0 : ));
6533 0 : }
6534 :
6535 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6536 0 : if let Operation::Fill(fill) = op_handler.operation {
6537 0 : if fill.node_id == node_id {
6538 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6539 0 : op_handler.cancel.cancel();
6540 0 : return Ok(());
6541 0 : }
6542 0 : }
6543 0 : }
6544 :
6545 0 : Err(ApiError::PreconditionFailed(
6546 0 : format!("Node {node_id} has no fill in progress").into(),
6547 0 : ))
6548 0 : }
6549 :
6550 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
6551 : /// configuration
6552 0 : fn maybe_reconcile_shard(
6553 0 : &self,
6554 0 : shard: &mut TenantShard,
6555 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6556 0 : priority: ReconcilerPriority,
6557 0 : ) -> Option<ReconcilerWaiter> {
6558 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
6559 0 : }
6560 :
6561 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
6562 0 : fn get_reconciler_units(
6563 0 : &self,
6564 0 : priority: ReconcilerPriority,
6565 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
6566 0 : let units = match priority {
6567 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
6568 : ReconcilerPriority::High => {
6569 0 : match self
6570 0 : .priority_reconciler_concurrency
6571 0 : .clone()
6572 0 : .try_acquire_owned()
6573 : {
6574 0 : Ok(u) => Ok(u),
6575 : Err(TryAcquireError::NoPermits) => {
6576 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
6577 : // the normal priority semaphore.
6578 0 : self.reconciler_concurrency.clone().try_acquire_owned()
6579 : }
6580 0 : Err(e) => Err(e),
6581 : }
6582 : }
6583 : };
6584 :
6585 0 : units.map(ReconcileUnits::new)
6586 0 : }
6587 :
6588 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
6589 0 : fn maybe_configured_reconcile_shard(
6590 0 : &self,
6591 0 : shard: &mut TenantShard,
6592 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6593 0 : reconciler_config: ReconcilerConfig,
6594 0 : ) -> Option<ReconcilerWaiter> {
6595 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
6596 0 :
6597 0 : match reconcile_needed {
6598 0 : ReconcileNeeded::No => return None,
6599 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
6600 0 : ReconcileNeeded::Yes => {
6601 0 : // Fall through to try and acquire units for spawning reconciler
6602 0 : }
6603 : };
6604 :
6605 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
6606 0 : Ok(u) => u,
6607 : Err(_) => {
6608 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
6609 0 : "Concurrency limited: enqueued for reconcile later");
6610 0 : if !shard.delayed_reconcile {
6611 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
6612 0 : Err(TrySendError::Closed(_)) => {
6613 0 : // Weird mid-shutdown case?
6614 0 : }
6615 : Err(TrySendError::Full(_)) => {
6616 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
6617 0 : tracing::warn!(
6618 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
6619 : );
6620 : }
6621 0 : Ok(()) => {
6622 0 : shard.delayed_reconcile = true;
6623 0 : }
6624 : }
6625 0 : }
6626 :
6627 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
6628 : // number to advance. When this function is eventually called again and succeeds in getting units,
6629 : // it will spawn a reconciler that makes this waiter complete.
6630 0 : return Some(shard.future_reconcile_waiter());
6631 : }
6632 : };
6633 :
6634 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
6635 : // Gate closed: we're shutting down, drop out.
6636 0 : return None;
6637 : };
6638 :
6639 0 : shard.spawn_reconciler(
6640 0 : &self.result_tx,
6641 0 : nodes,
6642 0 : &self.compute_hook,
6643 0 : reconciler_config,
6644 0 : &self.config,
6645 0 : &self.persistence,
6646 0 : units,
6647 0 : gate_guard,
6648 0 : &self.reconcilers_cancel,
6649 0 : )
6650 0 : }
6651 :
6652 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
6653 : /// Additionally, reschedule tenants that require it.
6654 : ///
6655 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
6656 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
6657 : /// available. A return value of 0 indicates that everything is fully reconciled already.
6658 0 : fn reconcile_all(&self) -> usize {
6659 0 : let mut locked = self.inner.write().unwrap();
6660 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6661 0 : let pageservers = nodes.clone();
6662 0 :
6663 0 : // This function is an efficient place to update lazy statistics, since we are walking
6664 0 : // all tenants.
6665 0 : let mut pending_reconciles = 0;
6666 0 : let mut az_violations = 0;
6667 0 :
6668 0 : // If we find any tenants to drop from memory, stash them to offload after
6669 0 : // we're done traversing the map of tenants.
6670 0 : let mut drop_detached_tenants = Vec::new();
6671 0 :
6672 0 : let mut reconciles_spawned = 0;
6673 0 : for shard in tenants.values_mut() {
6674 : // Accumulate scheduling statistics
6675 0 : if let (Some(attached), Some(preferred)) =
6676 0 : (shard.intent.get_attached(), shard.preferred_az())
6677 : {
6678 0 : let node_az = nodes
6679 0 : .get(attached)
6680 0 : .expect("Nodes exist if referenced")
6681 0 : .get_availability_zone_id();
6682 0 : if node_az != preferred {
6683 0 : az_violations += 1;
6684 0 : }
6685 0 : }
6686 :
6687 : // Skip checking if this shard is already enqueued for reconciliation
6688 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
6689 : // If there is something delayed, then return a nonzero count so that
6690 : // callers like reconcile_all_now do not incorrectly get the impression
6691 : // that the system is in a quiescent state.
6692 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
6693 0 : pending_reconciles += 1;
6694 0 : continue;
6695 0 : }
6696 0 :
6697 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
6698 0 : // dirty, spawn another rone
6699 0 : if self
6700 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
6701 0 : .is_some()
6702 0 : {
6703 0 : reconciles_spawned += 1;
6704 0 : } else if shard.delayed_reconcile {
6705 0 : // Shard wanted to reconcile but for some reason couldn't.
6706 0 : pending_reconciles += 1;
6707 0 : }
6708 :
6709 : // If this tenant is detached, try dropping it from memory. This is usually done
6710 : // proactively in [`Self::process_results`], but we do it here to handle the edge
6711 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
6712 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
6713 0 : && shard.policy == PlacementPolicy::Detached
6714 : {
6715 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
6716 0 : shard.tenant_shard_id.tenant_id,
6717 0 : TenantOperations::DropDetached,
6718 0 : ) {
6719 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
6720 0 : }
6721 0 : }
6722 : }
6723 :
6724 : // Some metrics are calculated from SchedulerNode state, update these periodically
6725 0 : scheduler.update_metrics();
6726 :
6727 : // Process any deferred tenant drops
6728 0 : for (tenant_id, guard) in drop_detached_tenants {
6729 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
6730 0 : }
6731 :
6732 0 : metrics::METRICS_REGISTRY
6733 0 : .metrics_group
6734 0 : .storage_controller_schedule_az_violation
6735 0 : .set(az_violations as i64);
6736 0 :
6737 0 : metrics::METRICS_REGISTRY
6738 0 : .metrics_group
6739 0 : .storage_controller_pending_reconciles
6740 0 : .set(pending_reconciles as i64);
6741 0 :
6742 0 : reconciles_spawned
6743 0 : }
6744 :
6745 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
6746 : /// could be scheduled somewhere better:
6747 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
6748 : /// * e.g. after a node fails then recovers, to move some work back to it
6749 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
6750 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
6751 : /// we did the split, but are probably better placed elsewhere.
6752 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
6753 : /// * e.g. after a shard split, some locations will be on the same node (where the split
6754 : /// happened), and will probably be better placed elsewhere.
6755 : ///
6756 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
6757 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
6758 : /// according to those same soft constraints.
6759 0 : async fn optimize_all(&self) -> usize {
6760 : // Limit on how many shards' optmizations each call to this function will execute. Combined
6761 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
6762 : // trickle of optimizations in the background, rather than executing a large number in parallel
6763 : // when a change occurs.
6764 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
6765 :
6766 : // Synchronous prepare: scan shards for possible scheduling optimizations
6767 0 : let candidate_work = self.optimize_all_plan();
6768 0 : let candidate_work_len = candidate_work.len();
6769 :
6770 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
6771 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
6772 :
6773 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
6774 0 :
6775 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
6776 0 : let mut reconciles_spawned = 0;
6777 0 : let mut optimizations_applied = 0;
6778 0 : let mut locked = self.inner.write().unwrap();
6779 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6780 0 : for (tenant_shard_id, optimization) in validated_work {
6781 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6782 : // Shard was dropped between planning and execution;
6783 0 : continue;
6784 : };
6785 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
6786 0 : if shard.apply_optimization(scheduler, optimization) {
6787 0 : optimizations_applied += 1;
6788 0 : if self
6789 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
6790 0 : .is_some()
6791 0 : {
6792 0 : reconciles_spawned += 1;
6793 0 : }
6794 0 : }
6795 :
6796 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
6797 0 : break;
6798 0 : }
6799 : }
6800 :
6801 0 : if was_work_filtered {
6802 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
6803 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
6804 0 : // as these validations start passing.
6805 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
6806 0 : }
6807 :
6808 0 : reconciles_spawned
6809 0 : }
6810 :
6811 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
6812 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
6813 : // this higher than the execution limit gives us a chance to execute some work even if the first
6814 : // few optimizations we find are not ready.
6815 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
6816 :
6817 0 : let mut work = Vec::new();
6818 0 : let mut locked = self.inner.write().unwrap();
6819 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6820 :
6821 : // We are going to plan a bunch of optimisations before applying any of them, so the
6822 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
6823 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
6824 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
6825 : // to ignore the utilisation component of the score.
6826 :
6827 0 : for (_tenant_id, schedule_context, shards) in
6828 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
6829 : {
6830 0 : for shard in shards {
6831 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
6832 0 : break;
6833 0 : }
6834 0 : match shard.get_scheduling_policy() {
6835 0 : ShardSchedulingPolicy::Active => {
6836 0 : // Ok to do optimization
6837 0 : }
6838 : ShardSchedulingPolicy::Essential
6839 : | ShardSchedulingPolicy::Pause
6840 : | ShardSchedulingPolicy::Stop => {
6841 : // Policy prevents optimizing this shard.
6842 0 : continue;
6843 : }
6844 : }
6845 :
6846 0 : if !matches!(shard.splitting, SplitState::Idle)
6847 0 : || matches!(shard.policy, PlacementPolicy::Detached)
6848 0 : || shard.reconciler.is_some()
6849 : {
6850 : // Do not start any optimizations while another change to the tenant is ongoing: this
6851 : // is not necessary for correctness, but simplifies operations and implicitly throttles
6852 : // optimization changes to happen in a "trickle" over time.
6853 0 : continue;
6854 0 : }
6855 0 :
6856 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
6857 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
6858 0 : if cfg!(feature = "testing") {
6859 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
6860 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
6861 : // panic in prod if we hit this, or spend cycles on it in prod.
6862 0 : assert!(shard
6863 0 : .optimize_attachment(scheduler, &schedule_context)
6864 0 : .is_none());
6865 0 : assert!(shard
6866 0 : .optimize_secondary(scheduler, &schedule_context)
6867 0 : .is_none());
6868 0 : }
6869 0 : continue;
6870 0 : }
6871 :
6872 0 : if let Some(optimization) =
6873 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
6874 : // its primary location based on soft constraints, cut it over.
6875 0 : shard.optimize_attachment(scheduler, &schedule_context)
6876 : {
6877 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
6878 0 : work.push((shard.tenant_shard_id, optimization));
6879 0 : break;
6880 0 : } else if let Some(optimization) =
6881 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
6882 : // better placed on another node, based on ScheduleContext, then adjust it. This
6883 : // covers cases like after a shard split, where we might have too many shards
6884 : // in the same tenant with secondary locations on the node where they originally split.
6885 0 : shard.optimize_secondary(scheduler, &schedule_context)
6886 : {
6887 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
6888 0 : work.push((shard.tenant_shard_id, optimization));
6889 0 : break;
6890 0 : }
6891 : }
6892 : }
6893 :
6894 0 : work
6895 0 : }
6896 :
6897 0 : async fn optimize_all_validate(
6898 0 : &self,
6899 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
6900 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
6901 0 : // Take a clone of the node map to use outside the lock in async validation phase
6902 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
6903 0 :
6904 0 : let mut want_secondary_status = Vec::new();
6905 0 :
6906 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
6907 0 : // check that the state of locations is acceptable to run the optimization, such as
6908 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
6909 0 : // in a live migration.
6910 0 : let mut validated_work = Vec::new();
6911 0 : for (tenant_shard_id, optimization) in candidate_work {
6912 0 : match optimization.action {
6913 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
6914 : old_attached_node_id: _,
6915 0 : new_attached_node_id,
6916 0 : }) => {
6917 0 : match validation_nodes.get(&new_attached_node_id) {
6918 0 : None => {
6919 0 : // Node was dropped between planning and validation
6920 0 : }
6921 0 : Some(node) => {
6922 0 : if !node.is_available() {
6923 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
6924 0 : } else {
6925 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
6926 0 : // remote API requests concurrently.
6927 0 : want_secondary_status.push((
6928 0 : tenant_shard_id,
6929 0 : node.clone(),
6930 0 : optimization,
6931 0 : ));
6932 0 : }
6933 : }
6934 : }
6935 : }
6936 : ScheduleOptimizationAction::ReplaceSecondary(_)
6937 : | ScheduleOptimizationAction::CreateSecondary(_)
6938 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
6939 : // No extra checks needed to manage secondaries: this does not interrupt client access
6940 0 : validated_work.push((tenant_shard_id, optimization))
6941 : }
6942 : };
6943 : }
6944 :
6945 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
6946 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
6947 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
6948 0 : let results = self
6949 0 : .tenant_for_shards_api(
6950 0 : want_secondary_status
6951 0 : .iter()
6952 0 : .map(|i| (i.0, i.1.clone()))
6953 0 : .collect(),
6954 0 : |tenant_shard_id, client| async move {
6955 0 : client.tenant_secondary_status(tenant_shard_id).await
6956 0 : },
6957 0 : 1,
6958 0 : 1,
6959 0 : SHORT_RECONCILE_TIMEOUT,
6960 0 : &self.cancel,
6961 0 : )
6962 0 : .await;
6963 :
6964 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
6965 0 : want_secondary_status.into_iter().zip(results.into_iter())
6966 : {
6967 0 : match secondary_status {
6968 0 : Err(e) => {
6969 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
6970 : }
6971 0 : Ok(progress) => {
6972 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
6973 : // them in an optimization
6974 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
6975 :
6976 0 : if progress.heatmap_mtime.is_none()
6977 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
6978 0 : && progress.bytes_downloaded != progress.bytes_total
6979 0 : || progress.bytes_total - progress.bytes_downloaded
6980 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
6981 : {
6982 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
6983 :
6984 : #[cfg(feature = "testing")]
6985 0 : if progress.heatmap_mtime.is_none() {
6986 : // No heatmap might mean the attached location has never uploaded one, or that
6987 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
6988 : // but fairly common in tests.
6989 0 : self.kick_secondary_download(tenant_shard_id).await;
6990 0 : }
6991 : } else {
6992 : // Location looks ready: proceed
6993 0 : tracing::info!(
6994 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
6995 : );
6996 0 : validated_work.push((tenant_shard_id, optimization))
6997 : }
6998 : }
6999 : }
7000 : }
7001 :
7002 0 : validated_work
7003 0 : }
7004 :
7005 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
7006 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
7007 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
7008 : /// we have this helper to move things along faster.
7009 : #[cfg(feature = "testing")]
7010 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
7011 0 : let (attached_node, secondaries) = {
7012 0 : let locked = self.inner.read().unwrap();
7013 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
7014 0 : tracing::warn!(
7015 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
7016 : );
7017 0 : return;
7018 : };
7019 :
7020 0 : let Some(attached) = shard.intent.get_attached() else {
7021 0 : tracing::warn!(
7022 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
7023 : );
7024 0 : return;
7025 : };
7026 :
7027 0 : let secondaries = shard
7028 0 : .intent
7029 0 : .get_secondary()
7030 0 : .iter()
7031 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
7032 0 : .collect::<Vec<_>>();
7033 0 :
7034 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
7035 0 : };
7036 0 :
7037 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
7038 0 : // a 'kick' to let scheduling optimisation run more promptly.
7039 0 : match attached_node
7040 0 : .with_client_retries(
7041 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
7042 0 : &self.config.jwt_token,
7043 0 : 3,
7044 0 : 10,
7045 0 : SHORT_RECONCILE_TIMEOUT,
7046 0 : &self.cancel,
7047 0 : )
7048 0 : .await
7049 : {
7050 0 : Some(Err(e)) => {
7051 0 : tracing::info!(
7052 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
7053 : );
7054 : }
7055 : None => {
7056 0 : tracing::info!(
7057 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
7058 : );
7059 : }
7060 : Some(Ok(_)) => {
7061 0 : tracing::info!(
7062 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
7063 : );
7064 : }
7065 : }
7066 :
7067 0 : for secondary_node in secondaries {
7068 0 : match secondary_node
7069 0 : .with_client_retries(
7070 0 : |client| async move {
7071 0 : client
7072 0 : .tenant_secondary_download(
7073 0 : tenant_shard_id,
7074 0 : Some(Duration::from_secs(1)),
7075 0 : )
7076 0 : .await
7077 0 : },
7078 0 : &self.config.jwt_token,
7079 0 : 3,
7080 0 : 10,
7081 0 : SHORT_RECONCILE_TIMEOUT,
7082 0 : &self.cancel,
7083 0 : )
7084 0 : .await
7085 : {
7086 0 : Some(Err(e)) => {
7087 0 : tracing::info!(
7088 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
7089 : );
7090 : }
7091 : None => {
7092 0 : tracing::info!("Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}");
7093 : }
7094 0 : Some(Ok(progress)) => {
7095 0 : tracing::info!("Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}");
7096 : }
7097 : }
7098 : }
7099 0 : }
7100 :
7101 : /// Look for shards which are oversized and in need of splitting
7102 0 : async fn autosplit_tenants(self: &Arc<Self>) {
7103 0 : let Some(split_threshold) = self.config.split_threshold else {
7104 : // Auto-splitting is disabled
7105 0 : return;
7106 : };
7107 :
7108 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7109 :
7110 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
7111 :
7112 0 : let mut top_n = Vec::new();
7113 0 :
7114 0 : // Call into each node to look for big tenants
7115 0 : let top_n_request = TopTenantShardsRequest {
7116 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
7117 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
7118 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
7119 0 : // disk space impact of one shard.
7120 0 : order_by: models::TenantSorting::MaxLogicalSize,
7121 0 : limit: 10,
7122 0 : where_shards_lt: Some(SPLIT_TO_MAX),
7123 0 : where_gt: Some(split_threshold),
7124 0 : };
7125 0 : for node in nodes.values() {
7126 0 : let request_ref = &top_n_request;
7127 0 : match node
7128 0 : .with_client_retries(
7129 0 : |client| async move {
7130 0 : let request = request_ref.clone();
7131 0 : client.top_tenant_shards(request.clone()).await
7132 0 : },
7133 0 : &self.config.jwt_token,
7134 0 : 3,
7135 0 : 3,
7136 0 : Duration::from_secs(5),
7137 0 : &self.cancel,
7138 0 : )
7139 0 : .await
7140 : {
7141 0 : Some(Ok(node_top_n)) => {
7142 0 : top_n.extend(node_top_n.shards.into_iter());
7143 0 : }
7144 : Some(Err(mgmt_api::Error::Cancelled)) => {
7145 0 : continue;
7146 : }
7147 0 : Some(Err(e)) => {
7148 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
7149 0 : continue;
7150 : }
7151 : None => {
7152 : // Node is shutting down
7153 0 : continue;
7154 : }
7155 : };
7156 : }
7157 :
7158 : // Pick the biggest tenant to split first
7159 0 : top_n.sort_by_key(|i| i.resident_size);
7160 0 :
7161 0 : // Filter out tenants in a prohibiting scheduling mode
7162 0 : {
7163 0 : let locked = self.inner.read().unwrap();
7164 0 : top_n.retain(|i| {
7165 0 : if let Some(shard) = locked.tenants.get(&i.id) {
7166 0 : matches!(shard.get_scheduling_policy(), ShardSchedulingPolicy::Active)
7167 : } else {
7168 0 : false
7169 : }
7170 0 : });
7171 0 : }
7172 :
7173 0 : let Some(split_candidate) = top_n.into_iter().next() else {
7174 0 : tracing::debug!("No split-elegible shards found");
7175 0 : return;
7176 : };
7177 :
7178 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
7179 : // want to block the background reconcile loop on this.
7180 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
7181 :
7182 0 : let this = self.clone();
7183 0 : tokio::spawn(
7184 0 : async move {
7185 0 : match this
7186 0 : .tenant_shard_split(
7187 0 : split_candidate.id.tenant_id,
7188 0 : TenantShardSplitRequest {
7189 0 : // Always split to the max number of shards: this avoids stepping through
7190 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
7191 0 : // each time as a tenant grows, and is not too expensive because our max shard
7192 0 : // count is relatively low anyway.
7193 0 : // This policy will be adjusted in future once we support higher shard count.
7194 0 : new_shard_count: SPLIT_TO_MAX.literal(),
7195 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
7196 0 : },
7197 0 : )
7198 0 : .await
7199 : {
7200 : Ok(_) => {
7201 0 : tracing::info!("Successful auto-split");
7202 : }
7203 0 : Err(e) => {
7204 0 : tracing::error!("Auto-split failed: {e}");
7205 : }
7206 : }
7207 0 : }
7208 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
7209 : );
7210 0 : }
7211 :
7212 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
7213 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
7214 : /// put the system into a quiescent state where future background reconciliations won't do anything.
7215 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
7216 0 : let reconciles_spawned = self.reconcile_all();
7217 0 : let reconciles_spawned = if reconciles_spawned == 0 {
7218 : // Only optimize when we are otherwise idle
7219 0 : self.optimize_all().await
7220 : } else {
7221 0 : reconciles_spawned
7222 : };
7223 :
7224 0 : let waiters = {
7225 0 : let mut waiters = Vec::new();
7226 0 : let locked = self.inner.read().unwrap();
7227 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
7228 0 : if let Some(waiter) = shard.get_waiter() {
7229 0 : waiters.push(waiter);
7230 0 : }
7231 : }
7232 0 : waiters
7233 0 : };
7234 0 :
7235 0 : let waiter_count = waiters.len();
7236 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
7237 0 : Ok(()) => {}
7238 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
7239 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
7240 0 : {
7241 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
7242 0 : // because some other change superceded it. We will return a nonzero number,
7243 0 : // so the caller knows they might have to call again to quiesce the system.
7244 0 : }
7245 0 : Err(e) => {
7246 0 : return Err(e);
7247 : }
7248 : };
7249 :
7250 0 : tracing::info!(
7251 0 : "{} reconciles in reconcile_all, {} waiters",
7252 : reconciles_spawned,
7253 : waiter_count
7254 : );
7255 :
7256 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
7257 0 : }
7258 :
7259 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
7260 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
7261 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
7262 0 : self.reconcilers_cancel.cancel();
7263 0 : self.reconcilers_gate.close().await;
7264 :
7265 : // Signal the background loop in [`Service::process_results`] to exit once
7266 : // it has proccessed the results from all the reconciles we cancelled earlier.
7267 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
7268 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
7269 0 : self.result_tx.closed().await;
7270 0 : }
7271 :
7272 0 : pub async fn shutdown(&self) {
7273 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
7274 0 : .await;
7275 :
7276 : // Background tasks hold gate guards: this notifies them of the cancellation and
7277 : // waits for them all to complete.
7278 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
7279 0 : self.cancel.cancel();
7280 0 : self.gate.close().await;
7281 0 : }
7282 :
7283 : /// Spot check the download lag for a secondary location of a shard.
7284 : /// Should be used as a heuristic, since it's not always precise: the
7285 : /// secondary might have not downloaded the new heat map yet and, hence,
7286 : /// is not aware of the lag.
7287 : ///
7288 : /// Returns:
7289 : /// * Ok(None) if the lag could not be determined from the status,
7290 : /// * Ok(Some(_)) if the lag could be determind
7291 : /// * Err on failures to query the pageserver.
7292 0 : async fn secondary_lag(
7293 0 : &self,
7294 0 : secondary: &NodeId,
7295 0 : tenant_shard_id: TenantShardId,
7296 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
7297 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7298 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
7299 0 : StatusCode::NOT_FOUND,
7300 0 : format!("Node with id {} not found", secondary),
7301 0 : ))?;
7302 :
7303 0 : match node
7304 0 : .with_client_retries(
7305 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
7306 0 : &self.config.jwt_token,
7307 0 : 1,
7308 0 : 3,
7309 0 : Duration::from_millis(250),
7310 0 : &self.cancel,
7311 0 : )
7312 0 : .await
7313 : {
7314 0 : Some(Ok(status)) => match status.heatmap_mtime {
7315 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
7316 0 : None => Ok(None),
7317 : },
7318 0 : Some(Err(e)) => Err(e),
7319 0 : None => Err(mgmt_api::Error::Cancelled),
7320 : }
7321 0 : }
7322 :
7323 : /// Drain a node by moving the shards attached to it as primaries.
7324 : /// This is a long running operation and it should run as a separate Tokio task.
7325 0 : pub(crate) async fn drain_node(
7326 0 : self: &Arc<Self>,
7327 0 : node_id: NodeId,
7328 0 : cancel: CancellationToken,
7329 0 : ) -> Result<(), OperationError> {
7330 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
7331 0 : let max_secondary_lag_bytes = self
7332 0 : .config
7333 0 : .max_secondary_lag_bytes
7334 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
7335 :
7336 : // By default, live migrations are generous about the wait time for getting
7337 : // the secondary location up to speed. When draining, give up earlier in order
7338 : // to not stall the operation when a cold secondary is encountered.
7339 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7340 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7341 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
7342 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7343 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7344 0 : .build();
7345 0 :
7346 0 : let mut waiters = Vec::new();
7347 0 :
7348 0 : let mut tid_iter = TenantShardIterator::new({
7349 0 : let service = self.clone();
7350 0 : move |last_inspected_shard: Option<TenantShardId>| {
7351 0 : let locked = &service.inner.read().unwrap();
7352 0 : let tenants = &locked.tenants;
7353 0 : let entry = match last_inspected_shard {
7354 0 : Some(skip_past) => {
7355 0 : // Skip to the last seen tenant shard id
7356 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
7357 0 :
7358 0 : // Skip past the last seen
7359 0 : cursor.nth(1)
7360 : }
7361 0 : None => tenants.first_key_value(),
7362 : };
7363 :
7364 0 : entry.map(|(tid, _)| tid).copied()
7365 0 : }
7366 0 : });
7367 :
7368 0 : while !tid_iter.finished() {
7369 0 : if cancel.is_cancelled() {
7370 0 : match self
7371 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7372 0 : .await
7373 : {
7374 0 : Ok(()) => return Err(OperationError::Cancelled),
7375 0 : Err(err) => {
7376 0 : return Err(OperationError::FinalizeError(
7377 0 : format!(
7378 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7379 0 : node_id, err
7380 0 : )
7381 0 : .into(),
7382 0 : ));
7383 : }
7384 : }
7385 0 : }
7386 0 :
7387 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
7388 :
7389 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7390 0 : let tid = match tid_iter.next() {
7391 0 : Some(tid) => tid,
7392 : None => {
7393 0 : break;
7394 : }
7395 : };
7396 :
7397 0 : let tid_drain = TenantShardDrain {
7398 0 : drained_node: node_id,
7399 0 : tenant_shard_id: tid,
7400 0 : };
7401 :
7402 0 : let dest_node_id = {
7403 0 : let locked = self.inner.read().unwrap();
7404 0 :
7405 0 : match tid_drain
7406 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
7407 : {
7408 0 : Some(node_id) => node_id,
7409 : None => {
7410 0 : continue;
7411 : }
7412 : }
7413 : };
7414 :
7415 0 : match self.secondary_lag(&dest_node_id, tid).await {
7416 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
7417 0 : // The secondary is reasonably up to date.
7418 0 : // Migrate to it
7419 0 : }
7420 0 : Ok(Some(lag)) => {
7421 0 : tracing::info!(
7422 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7423 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
7424 : );
7425 0 : continue;
7426 : }
7427 : Ok(None) => {
7428 0 : tracing::info!(
7429 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7430 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
7431 : );
7432 0 : continue;
7433 : }
7434 0 : Err(err) => {
7435 0 : tracing::warn!(
7436 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7437 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
7438 : );
7439 0 : continue;
7440 : }
7441 : }
7442 :
7443 : {
7444 0 : let mut locked = self.inner.write().unwrap();
7445 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7446 0 : let rescheduled = tid_drain.reschedule_to_secondary(
7447 0 : dest_node_id,
7448 0 : tenants,
7449 0 : scheduler,
7450 0 : nodes,
7451 0 : )?;
7452 :
7453 0 : if let Some(tenant_shard) = rescheduled {
7454 0 : let waiter = self.maybe_configured_reconcile_shard(
7455 0 : tenant_shard,
7456 0 : nodes,
7457 0 : reconciler_config,
7458 0 : );
7459 0 : if let Some(some) = waiter {
7460 0 : waiters.push(some);
7461 0 : }
7462 0 : }
7463 : }
7464 : }
7465 :
7466 0 : waiters = self
7467 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7468 0 : .await;
7469 :
7470 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
7471 : }
7472 :
7473 0 : while !waiters.is_empty() {
7474 0 : if cancel.is_cancelled() {
7475 0 : match self
7476 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7477 0 : .await
7478 : {
7479 0 : Ok(()) => return Err(OperationError::Cancelled),
7480 0 : Err(err) => {
7481 0 : return Err(OperationError::FinalizeError(
7482 0 : format!(
7483 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7484 0 : node_id, err
7485 0 : )
7486 0 : .into(),
7487 0 : ));
7488 : }
7489 : }
7490 0 : }
7491 0 :
7492 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
7493 :
7494 0 : waiters = self
7495 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7496 0 : .await;
7497 : }
7498 :
7499 : // At this point we have done the best we could to drain shards from this node.
7500 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
7501 : // to complete the drain.
7502 0 : if let Err(err) = self
7503 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
7504 0 : .await
7505 : {
7506 : // This is not fatal. Anything that is polling the node scheduling policy to detect
7507 : // the end of the drain operations will hang, but all such places should enforce an
7508 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
7509 : // by the counterpart fill operation.
7510 0 : return Err(OperationError::FinalizeError(
7511 0 : format!(
7512 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
7513 0 : )
7514 0 : .into(),
7515 0 : ));
7516 0 : }
7517 0 :
7518 0 : Ok(())
7519 0 : }
7520 :
7521 : /// Create a node fill plan (pick secondaries to promote), based on:
7522 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
7523 : /// outside their home AZ, should be migrated back here.
7524 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
7525 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
7526 : /// those shards have a home AZ that doesn't match the node we're filling.
7527 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
7528 0 : let mut locked = self.inner.write().unwrap();
7529 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7530 0 :
7531 0 : let node_az = nodes
7532 0 : .get(&node_id)
7533 0 : .expect("Node must exist")
7534 0 : .get_availability_zone_id()
7535 0 : .clone();
7536 0 :
7537 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
7538 0 : let mut plan = Vec::new();
7539 0 :
7540 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
7541 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
7542 0 :
7543 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
7544 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
7545 0 : let respect_azs = nodes
7546 0 : .values()
7547 0 : .map(|n| n.get_availability_zone_id())
7548 0 : .unique()
7549 0 : .count()
7550 0 : > 1;
7551 :
7552 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
7553 : // requires it.
7554 0 : for (tsid, tenant_shard) in tenants {
7555 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7556 : // Shard doesn't have a secondary on this node, ignore it.
7557 0 : continue;
7558 0 : }
7559 0 :
7560 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
7561 0 : // shards which belong on this node, not to promote shards whose scheduling preference
7562 0 : // would be on their currently attached node. So will avoid promoting shards whose
7563 0 : // home AZ doesn't match the AZ of the node we're filling.
7564 0 : match tenant_shard.preferred_az() {
7565 0 : _ if !respect_azs => {
7566 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7567 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7568 0 : }
7569 : }
7570 : None => {
7571 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
7572 : // will only do so if our target shard count requires it.
7573 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7574 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7575 0 : }
7576 : }
7577 0 : Some(az) if az == &node_az => {
7578 : // This shard's home AZ is equal to the node we're filling: it should
7579 : // be moved back to this node as part of filling, unless its currently
7580 : // attached location is also in its home AZ.
7581 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7582 0 : if nodes
7583 0 : .get(primary)
7584 0 : .expect("referenced node must exist")
7585 0 : .get_availability_zone_id()
7586 0 : != tenant_shard
7587 0 : .preferred_az()
7588 0 : .expect("tenant must have an AZ preference")
7589 : {
7590 0 : plan.push(*tsid)
7591 0 : }
7592 : } else {
7593 0 : plan.push(*tsid)
7594 : }
7595 : }
7596 0 : Some(_) => {
7597 0 : // This shard's home AZ is somewhere other than the node we're filling,
7598 0 : // it may not be moved back to this node as part of filling. Ignore it
7599 0 : }
7600 : }
7601 : }
7602 :
7603 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
7604 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
7605 0 :
7606 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
7607 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
7608 0 :
7609 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
7610 :
7611 0 : for (node_id, attached) in nodes_by_load {
7612 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
7613 0 : if !available {
7614 0 : continue;
7615 0 : }
7616 0 :
7617 0 : if plan.len() >= fill_requirement
7618 0 : || free_tids_by_node.is_empty()
7619 0 : || attached <= expected_attached
7620 : {
7621 0 : break;
7622 0 : }
7623 0 :
7624 0 : let can_take = attached - expected_attached;
7625 0 : let needed = fill_requirement - plan.len();
7626 0 : let mut take = std::cmp::min(can_take, needed);
7627 0 :
7628 0 : let mut remove_node = false;
7629 0 : while take > 0 {
7630 0 : match free_tids_by_node.get_mut(&node_id) {
7631 0 : Some(tids) => match tids.pop() {
7632 0 : Some(tid) => {
7633 0 : let max_promote_for_tenant = std::cmp::max(
7634 0 : tid.shard_count.count() as usize / locked.nodes.len(),
7635 0 : 1,
7636 0 : );
7637 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
7638 0 : if *promoted < max_promote_for_tenant {
7639 0 : plan.push(tid);
7640 0 : *promoted += 1;
7641 0 : take -= 1;
7642 0 : }
7643 : }
7644 : None => {
7645 0 : remove_node = true;
7646 0 : break;
7647 : }
7648 : },
7649 : None => {
7650 0 : break;
7651 : }
7652 : }
7653 : }
7654 :
7655 0 : if remove_node {
7656 0 : free_tids_by_node.remove(&node_id);
7657 0 : }
7658 : }
7659 :
7660 0 : plan
7661 0 : }
7662 :
7663 : /// Fill a node by promoting its secondaries until the cluster is balanced
7664 : /// with regards to attached shard counts. Note that this operation only
7665 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
7666 : /// This is a long running operation and it should run as a separate Tokio task.
7667 0 : pub(crate) async fn fill_node(
7668 0 : &self,
7669 0 : node_id: NodeId,
7670 0 : cancel: CancellationToken,
7671 0 : ) -> Result<(), OperationError> {
7672 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7673 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7674 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
7675 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7676 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7677 0 : .build();
7678 0 :
7679 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
7680 0 : let mut waiters = Vec::new();
7681 :
7682 : // Execute the plan we've composed above. Before aplying each move from the plan,
7683 : // we validate to ensure that it has not gone stale in the meantime.
7684 0 : while !tids_to_promote.is_empty() {
7685 0 : if cancel.is_cancelled() {
7686 0 : match self
7687 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7688 0 : .await
7689 : {
7690 0 : Ok(()) => return Err(OperationError::Cancelled),
7691 0 : Err(err) => {
7692 0 : return Err(OperationError::FinalizeError(
7693 0 : format!(
7694 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7695 0 : node_id, err
7696 0 : )
7697 0 : .into(),
7698 0 : ));
7699 : }
7700 : }
7701 0 : }
7702 0 :
7703 0 : {
7704 0 : let mut locked = self.inner.write().unwrap();
7705 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7706 :
7707 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
7708 0 : format!("node {node_id} was removed").into(),
7709 0 : ))?;
7710 :
7711 0 : let current_policy = node.get_scheduling();
7712 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
7713 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
7714 : // about it
7715 0 : return Err(OperationError::NodeStateChanged(
7716 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
7717 0 : ));
7718 0 : }
7719 :
7720 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7721 0 : if let Some(tid) = tids_to_promote.pop() {
7722 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
7723 : // If the node being filled is not a secondary anymore,
7724 : // skip the promotion.
7725 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7726 0 : continue;
7727 0 : }
7728 0 :
7729 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
7730 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
7731 0 : Err(e) => {
7732 0 : tracing::warn!(
7733 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7734 0 : "Scheduling error when filling pageserver {} : {e}", node_id
7735 : );
7736 : }
7737 : Ok(()) => {
7738 0 : tracing::info!(
7739 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7740 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
7741 : node_id,
7742 : previously_attached_to,
7743 : node_id
7744 : );
7745 :
7746 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
7747 0 : tenant_shard,
7748 0 : nodes,
7749 0 : reconciler_config,
7750 0 : ) {
7751 0 : waiters.push(waiter);
7752 0 : }
7753 : }
7754 : }
7755 0 : }
7756 : } else {
7757 0 : break;
7758 : }
7759 : }
7760 : }
7761 :
7762 0 : waiters = self
7763 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7764 0 : .await;
7765 : }
7766 :
7767 0 : while !waiters.is_empty() {
7768 0 : if cancel.is_cancelled() {
7769 0 : match self
7770 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7771 0 : .await
7772 : {
7773 0 : Ok(()) => return Err(OperationError::Cancelled),
7774 0 : Err(err) => {
7775 0 : return Err(OperationError::FinalizeError(
7776 0 : format!(
7777 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7778 0 : node_id, err
7779 0 : )
7780 0 : .into(),
7781 0 : ));
7782 : }
7783 : }
7784 0 : }
7785 0 :
7786 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
7787 :
7788 0 : waiters = self
7789 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7790 0 : .await;
7791 : }
7792 :
7793 0 : if let Err(err) = self
7794 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7795 0 : .await
7796 : {
7797 : // This isn't a huge issue since the filling process starts upon request. However, it
7798 : // will prevent the next drain from starting. The only case in which this can fail
7799 : // is database unavailability. Such a case will require manual intervention.
7800 0 : return Err(OperationError::FinalizeError(
7801 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
7802 0 : .into(),
7803 0 : ));
7804 0 : }
7805 0 :
7806 0 : Ok(())
7807 0 : }
7808 :
7809 : /// Updates scrubber metadata health check results.
7810 0 : pub(crate) async fn metadata_health_update(
7811 0 : &self,
7812 0 : update_req: MetadataHealthUpdateRequest,
7813 0 : ) -> Result<(), ApiError> {
7814 0 : let now = chrono::offset::Utc::now();
7815 0 : let (healthy_records, unhealthy_records) = {
7816 0 : let locked = self.inner.read().unwrap();
7817 0 : let healthy_records = update_req
7818 0 : .healthy_tenant_shards
7819 0 : .into_iter()
7820 0 : // Retain only health records associated with tenant shards managed by storage controller.
7821 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7822 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
7823 0 : .collect();
7824 0 : let unhealthy_records = update_req
7825 0 : .unhealthy_tenant_shards
7826 0 : .into_iter()
7827 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7828 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
7829 0 : .collect();
7830 0 :
7831 0 : (healthy_records, unhealthy_records)
7832 0 : };
7833 0 :
7834 0 : self.persistence
7835 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
7836 0 : .await?;
7837 0 : Ok(())
7838 0 : }
7839 :
7840 : /// Lists the tenant shards that has unhealthy metadata status.
7841 0 : pub(crate) async fn metadata_health_list_unhealthy(
7842 0 : &self,
7843 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
7844 0 : let result = self
7845 0 : .persistence
7846 0 : .list_unhealthy_metadata_health_records()
7847 0 : .await?
7848 0 : .iter()
7849 0 : .map(|p| p.get_tenant_shard_id().unwrap())
7850 0 : .collect();
7851 0 :
7852 0 : Ok(result)
7853 0 : }
7854 :
7855 : /// Lists the tenant shards that have not been scrubbed for some duration.
7856 0 : pub(crate) async fn metadata_health_list_outdated(
7857 0 : &self,
7858 0 : not_scrubbed_for: Duration,
7859 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
7860 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
7861 0 : let result = self
7862 0 : .persistence
7863 0 : .list_outdated_metadata_health_records(earlier)
7864 0 : .await?
7865 0 : .into_iter()
7866 0 : .map(|record| record.into())
7867 0 : .collect();
7868 0 : Ok(result)
7869 0 : }
7870 :
7871 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
7872 0 : self.inner.read().unwrap().get_leadership_status()
7873 0 : }
7874 :
7875 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
7876 0 : tracing::info!("Received step down request from peer");
7877 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
7878 :
7879 0 : self.inner.write().unwrap().step_down();
7880 0 : // TODO: would it make sense to have a time-out for this?
7881 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
7882 0 : .await;
7883 :
7884 0 : let mut global_observed = GlobalObservedState::default();
7885 0 : let locked = self.inner.read().unwrap();
7886 0 : for (tid, tenant_shard) in locked.tenants.iter() {
7887 0 : global_observed
7888 0 : .0
7889 0 : .insert(*tid, tenant_shard.observed.clone());
7890 0 : }
7891 :
7892 0 : global_observed
7893 0 : }
7894 :
7895 0 : pub(crate) async fn safekeepers_list(
7896 0 : &self,
7897 0 : ) -> Result<Vec<SafekeeperDescribeResponse>, DatabaseError> {
7898 0 : let locked = self.inner.read().unwrap();
7899 0 : let mut list = locked
7900 0 : .safekeepers
7901 0 : .iter()
7902 0 : .map(|sk| sk.1.describe_response())
7903 0 : .collect::<Result<Vec<_>, _>>()?;
7904 0 : list.sort_by_key(|v| v.id);
7905 0 : Ok(list)
7906 0 : }
7907 :
7908 0 : pub(crate) async fn get_safekeeper(
7909 0 : &self,
7910 0 : id: i64,
7911 0 : ) -> Result<SafekeeperDescribeResponse, DatabaseError> {
7912 0 : let locked = self.inner.read().unwrap();
7913 0 : let sk = locked
7914 0 : .safekeepers
7915 0 : .get(&NodeId(id as u64))
7916 0 : .ok_or(diesel::result::Error::NotFound)?;
7917 0 : sk.describe_response()
7918 0 : }
7919 :
7920 0 : pub(crate) async fn upsert_safekeeper(
7921 0 : &self,
7922 0 : record: crate::persistence::SafekeeperUpsert,
7923 0 : ) -> Result<(), DatabaseError> {
7924 0 : let node_id = NodeId(record.id as u64);
7925 0 : self.persistence.safekeeper_upsert(record.clone()).await?;
7926 : {
7927 0 : let mut locked = self.inner.write().unwrap();
7928 0 : let mut safekeepers = (*locked.safekeepers).clone();
7929 0 : match safekeepers.entry(node_id) {
7930 0 : std::collections::hash_map::Entry::Occupied(mut entry) => {
7931 0 : entry.get_mut().update_from_record(record);
7932 0 : }
7933 0 : std::collections::hash_map::Entry::Vacant(entry) => {
7934 0 : entry.insert(Safekeeper::from_persistence(
7935 0 : crate::persistence::SafekeeperPersistence::from_upsert(
7936 0 : record,
7937 0 : SkSchedulingPolicy::Pause,
7938 0 : ),
7939 0 : CancellationToken::new(),
7940 0 : ));
7941 0 : }
7942 : }
7943 0 : locked.safekeepers = Arc::new(safekeepers);
7944 0 : }
7945 0 : Ok(())
7946 0 : }
7947 :
7948 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
7949 0 : &self,
7950 0 : id: i64,
7951 0 : scheduling_policy: SkSchedulingPolicy,
7952 0 : ) -> Result<(), DatabaseError> {
7953 0 : self.persistence
7954 0 : .set_safekeeper_scheduling_policy(id, scheduling_policy)
7955 0 : .await?;
7956 0 : let node_id = NodeId(id as u64);
7957 0 : // After the change has been persisted successfully, update the in-memory state
7958 0 : {
7959 0 : let mut locked = self.inner.write().unwrap();
7960 0 : let mut safekeepers = (*locked.safekeepers).clone();
7961 0 : let sk = safekeepers
7962 0 : .get_mut(&node_id)
7963 0 : .ok_or(DatabaseError::Logical("Not found".to_string()))?;
7964 0 : sk.skp.scheduling_policy = String::from(scheduling_policy);
7965 0 :
7966 0 : locked.safekeepers = Arc::new(safekeepers);
7967 0 : }
7968 0 : Ok(())
7969 0 : }
7970 :
7971 0 : pub(crate) async fn update_shards_preferred_azs(
7972 0 : &self,
7973 0 : req: ShardsPreferredAzsRequest,
7974 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
7975 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
7976 0 : let updated = self
7977 0 : .persistence
7978 0 : .set_tenant_shard_preferred_azs(preferred_azs)
7979 0 : .await
7980 0 : .map_err(|err| {
7981 0 : ApiError::InternalServerError(anyhow::anyhow!(
7982 0 : "Failed to persist preferred AZs: {err}"
7983 0 : ))
7984 0 : })?;
7985 :
7986 0 : let mut updated_in_mem_and_db = Vec::default();
7987 0 :
7988 0 : let mut locked = self.inner.write().unwrap();
7989 0 : for (tid, az_id) in updated {
7990 0 : let shard = locked.tenants.get_mut(&tid);
7991 0 : if let Some(shard) = shard {
7992 0 : shard.set_preferred_az(az_id);
7993 0 : updated_in_mem_and_db.push(tid);
7994 0 : }
7995 : }
7996 :
7997 0 : Ok(ShardsPreferredAzsResponse {
7998 0 : updated: updated_in_mem_and_db,
7999 0 : })
8000 0 : }
8001 : }
|
