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 pageserver_jwt_token: Option<String>,
352 :
353 : // All safekeepers managed by one instance of this service must have
354 : // the same public key. This JWT token will be used to authenticate
355 : // this service to the safekeepers it manages.
356 : pub safekeeper_jwt_token: Option<String>,
357 :
358 : // This JWT token will be used to authenticate this service to the control plane.
359 : pub control_plane_jwt_token: Option<String>,
360 :
361 : // This JWT token will be used to authenticate with other storage controller instances
362 : pub peer_jwt_token: Option<String>,
363 :
364 : /// Where the compute hook should send notifications of pageserver attachment locations
365 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
366 : /// assume it is running in a test environment and try to update neon_local.
367 : pub compute_hook_url: Option<String>,
368 :
369 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
370 : /// considered active. Once the grace period elapses, the next heartbeat failure will
371 : /// mark the pagseserver offline.
372 : pub max_offline_interval: Duration,
373 :
374 : /// Extended grace period within which pageserver may not respond to heartbeats.
375 : /// This extended grace period kicks in after the node has been drained for restart
376 : /// and/or upon handling the re-attach request from a node.
377 : pub max_warming_up_interval: Duration,
378 :
379 : /// How many normal-priority Reconcilers may be spawned concurrently
380 : pub reconciler_concurrency: usize,
381 :
382 : /// How many high-priority Reconcilers may be spawned concurrently
383 : pub priority_reconciler_concurrency: usize,
384 :
385 : /// How large must a shard grow in bytes before we split it?
386 : /// None disables auto-splitting.
387 : pub split_threshold: Option<u64>,
388 :
389 : // TODO: make this cfg(feature = "testing")
390 : pub neon_local_repo_dir: Option<PathBuf>,
391 :
392 : // Maximum acceptable download lag for the secondary location
393 : // while draining a node. If the secondary location is lagging
394 : // by more than the configured amount, then the secondary is not
395 : // upgraded to primary.
396 : pub max_secondary_lag_bytes: Option<u64>,
397 :
398 : pub heartbeat_interval: Duration,
399 :
400 : pub address_for_peers: Option<Uri>,
401 :
402 : pub start_as_candidate: bool,
403 :
404 : pub http_service_port: i32,
405 :
406 : pub long_reconcile_threshold: Duration,
407 :
408 : pub use_https_pageserver_api: bool,
409 : }
410 :
411 : impl From<DatabaseError> for ApiError {
412 0 : fn from(err: DatabaseError) -> ApiError {
413 0 : match err {
414 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
415 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
416 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
417 0 : ApiError::ShuttingDown
418 : }
419 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
420 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
421 : }
422 : }
423 0 : }
424 : }
425 :
426 : enum InitialShardScheduleOutcome {
427 : Scheduled(TenantCreateResponseShard),
428 : NotScheduled,
429 : ShardScheduleError(ScheduleError),
430 : }
431 :
432 : pub struct Service {
433 : inner: Arc<std::sync::RwLock<ServiceState>>,
434 : config: Config,
435 : persistence: Arc<Persistence>,
436 : compute_hook: Arc<ComputeHook>,
437 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
438 :
439 : heartbeater_ps: Heartbeater<Node, PageserverState>,
440 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
441 :
442 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
443 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
444 :
445 : // Locking on a tenant granularity (covers all shards in the tenant):
446 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
447 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
448 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
449 :
450 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
451 : // that transition it to/from Active.
452 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
453 :
454 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
455 : // and reconciliation on startup.
456 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
457 :
458 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
459 : // a human user might be waiting for.
460 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
461 :
462 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
463 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
464 : ///
465 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
466 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
467 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
468 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
469 :
470 : // Process shutdown will fire this token
471 : cancel: CancellationToken,
472 :
473 : // Child token of [`Service::cancel`] used by reconcilers
474 : reconcilers_cancel: CancellationToken,
475 :
476 : // Background tasks will hold this gate
477 : gate: Gate,
478 :
479 : // Reconcilers background tasks will hold this gate
480 : reconcilers_gate: Gate,
481 :
482 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
483 : /// passes, it isn't safe to do any actions that mutate tenants.
484 : pub(crate) startup_complete: Barrier,
485 : }
486 :
487 : impl From<ReconcileWaitError> for ApiError {
488 0 : fn from(value: ReconcileWaitError) -> Self {
489 0 : match value {
490 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
491 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
492 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
493 : }
494 0 : }
495 : }
496 :
497 : impl From<OperationError> for ApiError {
498 0 : fn from(value: OperationError) -> Self {
499 0 : match value {
500 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
501 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
502 : }
503 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
504 : }
505 0 : }
506 : }
507 :
508 : #[allow(clippy::large_enum_variant)]
509 : enum TenantCreateOrUpdate {
510 : Create(TenantCreateRequest),
511 : Update(Vec<ShardUpdate>),
512 : }
513 :
514 : struct ShardSplitParams {
515 : old_shard_count: ShardCount,
516 : new_shard_count: ShardCount,
517 : new_stripe_size: Option<ShardStripeSize>,
518 : targets: Vec<ShardSplitTarget>,
519 : policy: PlacementPolicy,
520 : config: TenantConfig,
521 : shard_ident: ShardIdentity,
522 : preferred_az_id: Option<AvailabilityZone>,
523 : }
524 :
525 : // When preparing for a shard split, we may either choose to proceed with the split,
526 : // or find that the work is already done and return NoOp.
527 : enum ShardSplitAction {
528 : Split(Box<ShardSplitParams>),
529 : NoOp(TenantShardSplitResponse),
530 : }
531 :
532 : // A parent shard which will be split
533 : struct ShardSplitTarget {
534 : parent_id: TenantShardId,
535 : node: Node,
536 : child_ids: Vec<TenantShardId>,
537 : }
538 :
539 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
540 : /// might not be available. We therefore use a queue of abort operations processed in the background.
541 : struct TenantShardSplitAbort {
542 : tenant_id: TenantId,
543 : /// The target values from the request that failed
544 : new_shard_count: ShardCount,
545 : new_stripe_size: Option<ShardStripeSize>,
546 : /// Until this abort op is complete, no other operations may be done on the tenant
547 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
548 : }
549 :
550 : #[derive(thiserror::Error, Debug)]
551 : enum TenantShardSplitAbortError {
552 : #[error(transparent)]
553 : Database(#[from] DatabaseError),
554 : #[error(transparent)]
555 : Remote(#[from] mgmt_api::Error),
556 : #[error("Unavailable")]
557 : Unavailable,
558 : }
559 :
560 : struct ShardUpdate {
561 : tenant_shard_id: TenantShardId,
562 : placement_policy: PlacementPolicy,
563 : tenant_config: TenantConfig,
564 :
565 : /// If this is None, generation is not updated.
566 : generation: Option<Generation>,
567 :
568 : /// If this is None, scheduling policy is not updated.
569 : scheduling_policy: Option<ShardSchedulingPolicy>,
570 : }
571 :
572 : enum StopReconciliationsReason {
573 : ShuttingDown,
574 : SteppingDown,
575 : }
576 :
577 : impl std::fmt::Display for StopReconciliationsReason {
578 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
579 0 : let s = match self {
580 0 : Self::ShuttingDown => "Shutting down",
581 0 : Self::SteppingDown => "Stepping down",
582 : };
583 0 : write!(writer, "{}", s)
584 0 : }
585 : }
586 :
587 : pub(crate) enum ReconcileResultRequest {
588 : ReconcileResult(ReconcileResult),
589 : Stop,
590 : }
591 :
592 : #[derive(Clone)]
593 : struct MutationLocation {
594 : node: Node,
595 : generation: Generation,
596 : }
597 :
598 : #[derive(Clone)]
599 : struct ShardMutationLocations {
600 : latest: MutationLocation,
601 : other: Vec<MutationLocation>,
602 : }
603 :
604 : #[derive(Default, Clone)]
605 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
606 :
607 : impl Service {
608 0 : pub fn get_config(&self) -> &Config {
609 0 : &self.config
610 0 : }
611 :
612 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
613 : /// view of the world, and determine which pageservers are responsive.
614 : #[instrument(skip_all)]
615 : async fn startup_reconcile(
616 : self: &Arc<Service>,
617 : current_leader: Option<ControllerPersistence>,
618 : leader_step_down_state: Option<GlobalObservedState>,
619 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
620 : Result<(), (TenantShardId, NotifyError)>,
621 : >,
622 : ) {
623 : // Startup reconciliation does I/O to other services: whether they
624 : // are responsive or not, we should aim to finish within our deadline, because:
625 : // - If we don't, a k8s readiness hook watching /ready will kill us.
626 : // - While we're waiting for startup reconciliation, we are not fully
627 : // available for end user operations like creating/deleting tenants and timelines.
628 : //
629 : // We set multiple deadlines to break up the time available between the phases of work: this is
630 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
631 : let start_at = Instant::now();
632 : let node_scan_deadline = start_at
633 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
634 : .expect("Reconcile timeout is a modest constant");
635 :
636 : let observed = if let Some(state) = leader_step_down_state {
637 : tracing::info!(
638 : "Using observed state received from leader at {}",
639 : current_leader.as_ref().unwrap().address
640 : );
641 :
642 : state
643 : } else {
644 : self.build_global_observed_state(node_scan_deadline).await
645 : };
646 :
647 : // Accumulate a list of any tenant locations that ought to be detached
648 : let mut cleanup = Vec::new();
649 :
650 : // Send initial heartbeat requests to all nodes loaded from the database
651 : let all_nodes = {
652 : let locked = self.inner.read().unwrap();
653 : locked.nodes.clone()
654 : };
655 : let (mut nodes_online, mut sks_online) =
656 : self.initial_heartbeat_round(all_nodes.keys()).await;
657 :
658 : // List of tenants for which we will attempt to notify compute of their location at startup
659 : let mut compute_notifications = Vec::new();
660 :
661 : // Populate intent and observed states for all tenants, based on reported state on pageservers
662 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
663 : let shard_count = {
664 : let mut locked = self.inner.write().unwrap();
665 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
666 :
667 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
668 : let mut new_nodes = (**nodes).clone();
669 : for (node_id, node) in new_nodes.iter_mut() {
670 : if let Some(utilization) = nodes_online.remove(node_id) {
671 : node.set_availability(NodeAvailability::Active(utilization));
672 : scheduler.node_upsert(node);
673 : }
674 : }
675 : *nodes = Arc::new(new_nodes);
676 :
677 : let mut new_sks = (**safekeepers).clone();
678 : for (node_id, node) in new_sks.iter_mut() {
679 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
680 : node.set_availability(SafekeeperState::Available {
681 : utilization,
682 : last_seen_at,
683 : });
684 : }
685 : }
686 : *safekeepers = Arc::new(new_sks);
687 :
688 : for (tenant_shard_id, observed_state) in observed.0 {
689 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
690 : for node_id in observed_state.locations.keys() {
691 : cleanup.push((tenant_shard_id, *node_id));
692 : }
693 :
694 : continue;
695 : };
696 :
697 : tenant_shard.observed = observed_state;
698 : }
699 :
700 : // Populate each tenant's intent state
701 : let mut schedule_context = ScheduleContext::default();
702 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
703 : if tenant_shard_id.shard_number == ShardNumber(0) {
704 : // Reset scheduling context each time we advance to the next Tenant
705 : schedule_context = ScheduleContext::default();
706 : }
707 :
708 : tenant_shard.intent_from_observed(scheduler);
709 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
710 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
711 : // not enough pageservers are available. The tenant may well still be available
712 : // to clients.
713 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
714 : } else {
715 : // If we're both intending and observed to be attached at a particular node, we will
716 : // emit a compute notification for this. In the case where our observed state does not
717 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
718 : if let Some(attached_at) = tenant_shard.stably_attached() {
719 : compute_notifications.push(compute_hook::ShardUpdate {
720 : tenant_shard_id: *tenant_shard_id,
721 : node_id: attached_at,
722 : stripe_size: tenant_shard.shard.stripe_size,
723 : preferred_az: tenant_shard
724 : .preferred_az()
725 0 : .map(|az| Cow::Owned(az.clone())),
726 : });
727 : }
728 : }
729 : }
730 :
731 : tenants.len()
732 : };
733 :
734 : // Before making any obeservable changes to the cluster, persist self
735 : // as leader in database and memory.
736 : let leadership = Leadership::new(
737 : self.persistence.clone(),
738 : self.config.clone(),
739 : self.cancel.child_token(),
740 : );
741 :
742 : if let Err(e) = leadership.become_leader(current_leader).await {
743 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
744 : std::process::exit(1);
745 : }
746 :
747 : self.inner.write().unwrap().become_leader();
748 :
749 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
750 : // generation_pageserver in the database.
751 :
752 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
753 : // will emit compute hook notifications when they reconcile.
754 : //
755 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
756 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
757 : // calls will be correctly ordered wrt these.
758 : //
759 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
760 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
761 : // unit and start doing I/O.
762 : tracing::info!(
763 : "Sending {} compute notifications",
764 : compute_notifications.len()
765 : );
766 : self.compute_hook.notify_background(
767 : compute_notifications,
768 : bg_compute_notify_result_tx.clone(),
769 : &self.cancel,
770 : );
771 :
772 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
773 : // which require it: under normal circumstances this should only include tenants that were in some
774 : // transient state before we restarted, or any tenants whose compute hooks failed above.
775 : tracing::info!("Checking for shards in need of reconciliation...");
776 : let reconcile_tasks = self.reconcile_all();
777 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
778 : // normal operations may proceed.
779 :
780 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
781 : // background because it does not need to complete in order to proceed with other work.
782 : if !cleanup.is_empty() {
783 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
784 : tokio::task::spawn({
785 : let cleanup_self = self.clone();
786 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
787 : });
788 : }
789 :
790 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
791 : }
792 :
793 0 : async fn initial_heartbeat_round<'a>(
794 0 : &self,
795 0 : node_ids: impl Iterator<Item = &'a NodeId>,
796 0 : ) -> (
797 0 : HashMap<NodeId, PageserverUtilization>,
798 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
799 0 : ) {
800 0 : assert!(!self.startup_complete.is_ready());
801 :
802 0 : let all_nodes = {
803 0 : let locked = self.inner.read().unwrap();
804 0 : locked.nodes.clone()
805 0 : };
806 0 :
807 0 : let mut nodes_to_heartbeat = HashMap::new();
808 0 : for node_id in node_ids {
809 0 : match all_nodes.get(node_id) {
810 0 : Some(node) => {
811 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
812 0 : }
813 : None => {
814 0 : tracing::warn!("Node {node_id} was removed during start-up");
815 : }
816 : }
817 : }
818 :
819 0 : let all_sks = {
820 0 : let locked = self.inner.read().unwrap();
821 0 : locked.safekeepers.clone()
822 0 : };
823 0 :
824 0 : tracing::info!("Sending initial heartbeats...");
825 : // Put a small, but reasonable timeout to get the initial heartbeats of the safekeepers to avoid a storage controller downtime
826 : const SK_TIMEOUT: Duration = Duration::from_secs(5);
827 0 : let (res_ps, res_sk) = tokio::join!(
828 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
829 0 : tokio::time::timeout(SK_TIMEOUT, self.heartbeater_sk.heartbeat(all_sks))
830 0 : );
831 :
832 0 : let mut online_nodes = HashMap::new();
833 0 : if let Ok(deltas) = res_ps {
834 0 : for (node_id, status) in deltas.0 {
835 0 : match status {
836 0 : PageserverState::Available { utilization, .. } => {
837 0 : online_nodes.insert(node_id, utilization);
838 0 : }
839 0 : PageserverState::Offline => {}
840 : PageserverState::WarmingUp { .. } => {
841 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
842 : }
843 : }
844 : }
845 0 : }
846 :
847 0 : let mut online_sks = HashMap::new();
848 0 : if let Ok(Ok(deltas)) = res_sk {
849 0 : for (node_id, status) in deltas.0 {
850 0 : match status {
851 : SafekeeperState::Available {
852 0 : utilization,
853 0 : last_seen_at,
854 0 : } => {
855 0 : online_sks.insert(node_id, (utilization, last_seen_at));
856 0 : }
857 0 : SafekeeperState::Offline => {}
858 : }
859 : }
860 0 : }
861 :
862 0 : (online_nodes, online_sks)
863 0 : }
864 :
865 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
866 : ///
867 : /// The result includes only nodes which responded within the deadline
868 0 : async fn scan_node_locations(
869 0 : &self,
870 0 : deadline: Instant,
871 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
872 0 : let nodes = {
873 0 : let locked = self.inner.read().unwrap();
874 0 : locked.nodes.clone()
875 0 : };
876 0 :
877 0 : let mut node_results = HashMap::new();
878 0 :
879 0 : let mut node_list_futs = FuturesUnordered::new();
880 0 :
881 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
882 0 : for node in nodes.values() {
883 0 : node_list_futs.push({
884 0 : async move {
885 0 : tracing::info!("Scanning shards on node {node}...");
886 0 : let timeout = Duration::from_secs(5);
887 0 : let response = node
888 0 : .with_client_retries(
889 0 : |client| async move { client.list_location_config().await },
890 0 : &self.config.pageserver_jwt_token,
891 0 : 1,
892 0 : 5,
893 0 : timeout,
894 0 : &self.cancel,
895 0 : )
896 0 : .await;
897 0 : (node.get_id(), response)
898 0 : }
899 0 : });
900 0 : }
901 :
902 : loop {
903 0 : let (node_id, result) = tokio::select! {
904 0 : next = node_list_futs.next() => {
905 0 : match next {
906 0 : Some(result) => result,
907 : None =>{
908 : // We got results for all our nodes
909 0 : break;
910 : }
911 :
912 : }
913 : },
914 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
915 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
916 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
917 0 : break;
918 : }
919 : };
920 :
921 0 : let Some(list_response) = result else {
922 0 : tracing::info!("Shutdown during startup_reconcile");
923 0 : break;
924 : };
925 :
926 0 : match list_response {
927 0 : Err(e) => {
928 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
929 : }
930 0 : Ok(listing) => {
931 0 : node_results.insert(node_id, listing);
932 0 : }
933 : }
934 : }
935 :
936 0 : node_results
937 0 : }
938 :
939 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
940 0 : let node_listings = self.scan_node_locations(deadline).await;
941 0 : let mut observed = GlobalObservedState::default();
942 :
943 0 : for (node_id, location_confs) in node_listings {
944 0 : tracing::info!(
945 0 : "Received {} shard statuses from pageserver {}",
946 0 : location_confs.tenant_shards.len(),
947 : node_id
948 : );
949 :
950 0 : for (tid, location_conf) in location_confs.tenant_shards {
951 0 : let entry = observed.0.entry(tid).or_default();
952 0 : entry.locations.insert(
953 0 : node_id,
954 0 : ObservedStateLocation {
955 0 : conf: location_conf,
956 0 : },
957 0 : );
958 0 : }
959 : }
960 :
961 0 : observed
962 0 : }
963 :
964 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
965 : ///
966 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
967 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
968 : /// other task trying to attach it.
969 : #[instrument(skip_all)]
970 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
971 : let nodes = self.inner.read().unwrap().nodes.clone();
972 :
973 : for (tenant_shard_id, node_id) in cleanup {
974 : // A node reported a tenant_shard_id which is unknown to us: detach it.
975 : let Some(node) = nodes.get(&node_id) else {
976 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
977 : // a location to clean up on a node that has since been removed.
978 : tracing::info!(
979 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
980 : );
981 : continue;
982 : };
983 :
984 : if self.cancel.is_cancelled() {
985 : break;
986 : }
987 :
988 : let client = PageserverClient::new(
989 : node.get_id(),
990 : node.base_url(),
991 : self.config.pageserver_jwt_token.as_deref(),
992 : );
993 : match client
994 : .location_config(
995 : tenant_shard_id,
996 : LocationConfig {
997 : mode: LocationConfigMode::Detached,
998 : generation: None,
999 : secondary_conf: None,
1000 : shard_number: tenant_shard_id.shard_number.0,
1001 : shard_count: tenant_shard_id.shard_count.literal(),
1002 : shard_stripe_size: 0,
1003 : tenant_conf: models::TenantConfig::default(),
1004 : },
1005 : None,
1006 : false,
1007 : )
1008 : .await
1009 : {
1010 : Ok(()) => {
1011 : tracing::info!(
1012 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1013 : );
1014 : }
1015 : Err(e) => {
1016 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1017 : // break anything.
1018 : tracing::error!(
1019 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1020 : );
1021 : }
1022 : }
1023 : }
1024 : }
1025 :
1026 : /// Long running background task that periodically wakes up and looks for shards that need
1027 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1028 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1029 : /// for those retries.
1030 : #[instrument(skip_all)]
1031 : async fn background_reconcile(self: &Arc<Self>) {
1032 : self.startup_complete.clone().wait().await;
1033 :
1034 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1035 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1036 : while !self.reconcilers_cancel.is_cancelled() {
1037 : tokio::select! {
1038 : _ = interval.tick() => {
1039 : let reconciles_spawned = self.reconcile_all();
1040 : if reconciles_spawned == 0 {
1041 : // Run optimizer only when we didn't find any other work to do
1042 : self.optimize_all().await;
1043 : }
1044 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1045 : // must be responsive when new projects begin ingesting and reach the threshold.
1046 : self.autosplit_tenants().await;
1047 : }
1048 : _ = self.reconcilers_cancel.cancelled() => return
1049 : }
1050 : }
1051 : }
1052 : #[instrument(skip_all)]
1053 : async fn spawn_heartbeat_driver(&self) {
1054 : self.startup_complete.clone().wait().await;
1055 :
1056 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1057 : while !self.cancel.is_cancelled() {
1058 : tokio::select! {
1059 : _ = interval.tick() => { }
1060 : _ = self.cancel.cancelled() => return
1061 : };
1062 :
1063 : let nodes = {
1064 : let locked = self.inner.read().unwrap();
1065 : locked.nodes.clone()
1066 : };
1067 :
1068 : let safekeepers = {
1069 : let locked = self.inner.read().unwrap();
1070 : locked.safekeepers.clone()
1071 : };
1072 :
1073 : const SK_TIMEOUT: Duration = Duration::from_secs(3);
1074 : let (res_ps, res_sk) = tokio::join!(
1075 : self.heartbeater_ps.heartbeat(nodes),
1076 : tokio::time::timeout(SK_TIMEOUT, self.heartbeater_sk.heartbeat(safekeepers))
1077 : );
1078 :
1079 : if let Ok(deltas) = res_ps {
1080 : let mut to_handle = Vec::default();
1081 :
1082 : for (node_id, state) in deltas.0 {
1083 : let new_availability = match state {
1084 : PageserverState::Available { utilization, .. } => {
1085 : NodeAvailability::Active(utilization)
1086 : }
1087 : PageserverState::WarmingUp { started_at } => {
1088 : NodeAvailability::WarmingUp(started_at)
1089 : }
1090 : PageserverState::Offline => {
1091 : // The node might have been placed in the WarmingUp state
1092 : // while the heartbeat round was on-going. Hence, filter out
1093 : // offline transitions for WarmingUp nodes that are still within
1094 : // their grace period.
1095 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1096 : .get_node(node_id)
1097 : .await
1098 : .as_ref()
1099 0 : .map(|n| n.get_availability())
1100 : {
1101 : let now = Instant::now();
1102 : if now - *started_at >= self.config.max_warming_up_interval {
1103 : NodeAvailability::Offline
1104 : } else {
1105 : NodeAvailability::WarmingUp(*started_at)
1106 : }
1107 : } else {
1108 : NodeAvailability::Offline
1109 : }
1110 : }
1111 : };
1112 :
1113 : let node_lock = trace_exclusive_lock(
1114 : &self.node_op_locks,
1115 : node_id,
1116 : NodeOperations::Configure,
1117 : )
1118 : .await;
1119 :
1120 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1121 :
1122 : // This is the code path for geniune availability transitions (i.e node
1123 : // goes unavailable and/or comes back online).
1124 : let res = self
1125 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1126 : .await;
1127 :
1128 : match res {
1129 : Ok(transition) => {
1130 : // Keep hold of the lock until the availability transitions
1131 : // have been handled in
1132 : // [`Service::handle_node_availability_transitions`] in order avoid
1133 : // racing with [`Service::external_node_configure`].
1134 : to_handle.push((node_id, node_lock, transition));
1135 : }
1136 : Err(ApiError::NotFound(_)) => {
1137 : // This should be rare, but legitimate since the heartbeats are done
1138 : // on a snapshot of the nodes.
1139 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1140 : }
1141 : Err(ApiError::ShuttingDown) => {
1142 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1143 : }
1144 : Err(err) => {
1145 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1146 : // becomes unavailable again, we may get an error here.
1147 : tracing::error!(
1148 : "Failed to update node state {} after heartbeat round: {}",
1149 : node_id,
1150 : err
1151 : );
1152 : }
1153 : }
1154 : }
1155 :
1156 : // We collected all the transitions above and now we handle them.
1157 : let res = self.handle_node_availability_transitions(to_handle).await;
1158 : if let Err(errs) = res {
1159 : for (node_id, err) in errs {
1160 : match err {
1161 : ApiError::NotFound(_) => {
1162 : // This should be rare, but legitimate since the heartbeats are done
1163 : // on a snapshot of the nodes.
1164 : tracing::info!(
1165 : "Node {} was not found after heartbeat round",
1166 : node_id
1167 : );
1168 : }
1169 : err => {
1170 : tracing::error!(
1171 : "Failed to handle availability transition for {} after heartbeat round: {}",
1172 : node_id,
1173 : err
1174 : );
1175 : }
1176 : }
1177 : }
1178 : }
1179 : }
1180 : if let Ok(Ok(deltas)) = res_sk {
1181 : let mut locked = self.inner.write().unwrap();
1182 : let mut safekeepers = (*locked.safekeepers).clone();
1183 : for (id, state) in deltas.0 {
1184 : let Some(sk) = safekeepers.get_mut(&id) else {
1185 : tracing::info!("Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}");
1186 : continue;
1187 : };
1188 : sk.set_availability(state);
1189 : }
1190 : locked.safekeepers = Arc::new(safekeepers);
1191 : }
1192 : }
1193 : }
1194 :
1195 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1196 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1197 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1198 : /// will indicate that reconciliation is not needed.
1199 : #[instrument(skip_all, fields(
1200 : seq=%result.sequence,
1201 : tenant_id=%result.tenant_shard_id.tenant_id,
1202 : shard_id=%result.tenant_shard_id.shard_slug(),
1203 : ))]
1204 : fn process_result(&self, result: ReconcileResult) {
1205 : let mut locked = self.inner.write().unwrap();
1206 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1207 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1208 : // A reconciliation result might race with removing a tenant: drop results for
1209 : // tenants that aren't in our map.
1210 : return;
1211 : };
1212 :
1213 : // Usually generation should only be updated via this path, so the max() isn't
1214 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1215 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1216 :
1217 : // If the reconciler signals that it failed to notify compute, set this state on
1218 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1219 : tenant.pending_compute_notification = result.pending_compute_notification;
1220 :
1221 : // Let the TenantShard know it is idle.
1222 : tenant.reconcile_complete(result.sequence);
1223 :
1224 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1225 : // make to the tenant
1226 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1227 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1228 : // make to the tenant
1229 0 : let node = nodes.get(delta.node_id())?;
1230 :
1231 0 : if node.is_available() {
1232 0 : return Some(delta);
1233 0 : }
1234 0 :
1235 0 : // In case a node became unavailable concurrently with the reconcile, observed
1236 0 : // locations on it are now uncertain. By convention, set them to None in order
1237 0 : // for them to get refreshed when the node comes back online.
1238 0 : Some(ObservedStateDelta::Upsert(Box::new((
1239 0 : node.get_id(),
1240 0 : ObservedStateLocation { conf: None },
1241 0 : ))))
1242 0 : });
1243 :
1244 : match result.result {
1245 : Ok(()) => {
1246 : tenant.apply_observed_deltas(deltas);
1247 : tenant.waiter.advance(result.sequence);
1248 : }
1249 : Err(e) => {
1250 : match e {
1251 : ReconcileError::Cancel => {
1252 : tracing::info!("Reconciler was cancelled");
1253 : }
1254 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1255 : // This might be due to the reconciler getting cancelled, or it might
1256 : // be due to the `Node` being marked offline.
1257 : tracing::info!("Reconciler cancelled during pageserver API call");
1258 : }
1259 : _ => {
1260 : tracing::warn!("Reconcile error: {}", e);
1261 : }
1262 : }
1263 :
1264 : // Ordering: populate last_error before advancing error_seq,
1265 : // so that waiters will see the correct error after waiting.
1266 : tenant.set_last_error(result.sequence, e);
1267 :
1268 : // Skip deletions on reconcile failures
1269 : let upsert_deltas =
1270 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1271 : tenant.apply_observed_deltas(upsert_deltas);
1272 : }
1273 : }
1274 :
1275 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1276 : if tenant.policy == PlacementPolicy::Detached {
1277 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1278 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1279 : // duration of the request.
1280 : let guard = self.tenant_op_locks.try_exclusive(
1281 : tenant.tenant_shard_id.tenant_id,
1282 : TenantOperations::DropDetached,
1283 : );
1284 : if let Some(guard) = guard {
1285 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1286 : }
1287 : }
1288 :
1289 : // Maybe some other work can proceed now that this job finished.
1290 : //
1291 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1292 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1293 : if self.reconciler_concurrency.available_permits() > 0 {
1294 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1295 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1296 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1297 : shard.delayed_reconcile = false;
1298 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1299 : }
1300 :
1301 : if self.reconciler_concurrency.available_permits() == 0 {
1302 : break;
1303 : }
1304 : }
1305 : }
1306 : }
1307 :
1308 0 : async fn process_results(
1309 0 : &self,
1310 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1311 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1312 0 : Result<(), (TenantShardId, NotifyError)>,
1313 0 : >,
1314 0 : ) {
1315 : loop {
1316 : // Wait for the next result, or for cancellation
1317 0 : tokio::select! {
1318 0 : r = result_rx.recv() => {
1319 0 : match r {
1320 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1321 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1322 : }
1323 : }
1324 0 : _ = async{
1325 0 : match bg_compute_hook_result_rx.recv().await {
1326 0 : Some(result) => {
1327 0 : if let Err((tenant_shard_id, notify_error)) = result {
1328 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1329 0 : let mut locked = self.inner.write().unwrap();
1330 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1331 0 : shard.pending_compute_notification = true;
1332 0 : }
1333 :
1334 0 : }
1335 : },
1336 : None => {
1337 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1338 0 : self.cancel.cancelled().await;
1339 : }
1340 : }
1341 0 : } => {},
1342 0 : _ = self.cancel.cancelled() => {
1343 0 : break;
1344 : }
1345 : };
1346 : }
1347 0 : }
1348 :
1349 0 : async fn process_aborts(
1350 0 : &self,
1351 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1352 0 : ) {
1353 : loop {
1354 : // Wait for the next result, or for cancellation
1355 0 : let op = tokio::select! {
1356 0 : r = abort_rx.recv() => {
1357 0 : match r {
1358 0 : Some(op) => {op},
1359 0 : None => {break;}
1360 : }
1361 : }
1362 0 : _ = self.cancel.cancelled() => {
1363 0 : break;
1364 : }
1365 : };
1366 :
1367 : // Retry until shutdown: we must keep this request object alive until it is properly
1368 : // processed, as it holds a lock guard that prevents other operations trying to do things
1369 : // to the tenant while it is in a weird part-split state.
1370 0 : while !self.cancel.is_cancelled() {
1371 0 : match self.abort_tenant_shard_split(&op).await {
1372 0 : Ok(_) => break,
1373 0 : Err(e) => {
1374 0 : tracing::warn!(
1375 0 : "Failed to abort shard split on {}, will retry: {e}",
1376 : op.tenant_id
1377 : );
1378 :
1379 : // If a node is unavailable, we hope that it has been properly marked Offline
1380 : // when we retry, so that the abort op will succeed. If the abort op is failing
1381 : // for some other reason, we will keep retrying forever, or until a human notices
1382 : // and does something about it (either fixing a pageserver or restarting the controller).
1383 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1384 0 : .await
1385 0 : .ok();
1386 : }
1387 : }
1388 : }
1389 : }
1390 0 : }
1391 :
1392 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1393 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1394 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1395 0 :
1396 0 : let leadership_cancel = CancellationToken::new();
1397 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1398 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1399 :
1400 : // Apply the migrations **after** the current leader has stepped down
1401 : // (or we've given up waiting for it), but **before** reading from the
1402 : // database. The only exception is reading the current leader before
1403 : // migrating.
1404 0 : persistence.migration_run().await?;
1405 :
1406 0 : tracing::info!("Loading nodes from database...");
1407 0 : let nodes = persistence
1408 0 : .list_nodes()
1409 0 : .await?
1410 0 : .into_iter()
1411 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1412 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1413 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1414 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1415 0 : metrics::METRICS_REGISTRY
1416 0 : .metrics_group
1417 0 : .storage_controller_pageserver_nodes
1418 0 : .set(nodes.len() as i64);
1419 0 :
1420 0 : tracing::info!("Loading safekeepers from database...");
1421 0 : let safekeepers = persistence
1422 0 : .list_safekeepers()
1423 0 : .await?
1424 0 : .into_iter()
1425 0 : .map(|skp| Safekeeper::from_persistence(skp, CancellationToken::new()))
1426 0 : .collect::<Vec<_>>();
1427 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1428 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1429 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1430 :
1431 0 : tracing::info!("Loading shards from database...");
1432 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1433 0 : tracing::info!(
1434 0 : "Loaded {} shards from database.",
1435 0 : tenant_shard_persistence.len()
1436 : );
1437 :
1438 : // If any shard splits were in progress, reset the database state to abort them
1439 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1440 0 : HashMap::new();
1441 0 : for tsp in &mut tenant_shard_persistence {
1442 0 : let shard = tsp.get_shard_identity()?;
1443 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1444 0 : let entry = tenant_shard_count_min_max
1445 0 : .entry(tenant_shard_id.tenant_id)
1446 0 : .or_insert_with(|| (shard.count, shard.count));
1447 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1448 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1449 0 : }
1450 :
1451 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1452 0 : if count_min != count_max {
1453 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1454 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1455 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1456 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1457 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1458 :
1459 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1460 : // identified this tenant has having mismatching min/max counts.
1461 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1462 :
1463 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1464 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1465 0 : // Set idle split state on those shards that we will retain.
1466 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1467 0 : if tsp_tenant_id == tenant_id
1468 0 : && tsp.get_shard_identity().unwrap().count == count_min
1469 0 : {
1470 0 : tsp.splitting = SplitState::Idle;
1471 0 : } else if tsp_tenant_id == tenant_id {
1472 : // Leave the splitting state on the child shards: this will be used next to
1473 : // drop them.
1474 0 : tracing::info!(
1475 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1476 : );
1477 0 : }
1478 0 : });
1479 0 :
1480 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1481 0 : tenant_shard_persistence.retain(|tsp| {
1482 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1483 0 : || tsp.splitting == SplitState::Idle
1484 0 : });
1485 0 : }
1486 : }
1487 :
1488 0 : let mut tenants = BTreeMap::new();
1489 0 :
1490 0 : let mut scheduler = Scheduler::new(nodes.values());
1491 :
1492 : #[cfg(feature = "testing")]
1493 : {
1494 : use pageserver_api::controller_api::AvailabilityZone;
1495 :
1496 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1497 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1498 : // after when pageservers start up and register.
1499 0 : let mut node_ids = HashSet::new();
1500 0 : for tsp in &tenant_shard_persistence {
1501 0 : if let Some(node_id) = tsp.generation_pageserver {
1502 0 : node_ids.insert(node_id);
1503 0 : }
1504 : }
1505 0 : for node_id in node_ids {
1506 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1507 0 : let node = Node::new(
1508 0 : NodeId(node_id as u64),
1509 0 : "".to_string(),
1510 0 : 123,
1511 0 : None,
1512 0 : "".to_string(),
1513 0 : 123,
1514 0 : AvailabilityZone("test_az".to_string()),
1515 0 : false,
1516 0 : )
1517 0 : .unwrap();
1518 0 :
1519 0 : scheduler.node_upsert(&node);
1520 : }
1521 : }
1522 0 : for tsp in tenant_shard_persistence {
1523 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1524 :
1525 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1526 : // it with what we can infer: the node for which a generation was most recently issued.
1527 0 : let mut intent = IntentState::new(
1528 0 : tsp.preferred_az_id
1529 0 : .as_ref()
1530 0 : .map(|az| AvailabilityZone(az.clone())),
1531 0 : );
1532 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1533 : {
1534 0 : if nodes.contains_key(&generation_pageserver) {
1535 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1536 0 : } else {
1537 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1538 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1539 : // on different pageservers.
1540 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1541 : }
1542 0 : }
1543 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1544 :
1545 0 : tenants.insert(tenant_shard_id, new_tenant);
1546 : }
1547 :
1548 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1549 0 :
1550 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1551 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1552 0 : tokio::sync::mpsc::channel(512);
1553 0 :
1554 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1555 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1556 0 :
1557 0 : let cancel = CancellationToken::new();
1558 0 : let reconcilers_cancel = cancel.child_token();
1559 0 :
1560 0 : let heartbeater_ps = Heartbeater::new(
1561 0 : config.pageserver_jwt_token.clone(),
1562 0 : config.max_offline_interval,
1563 0 : config.max_warming_up_interval,
1564 0 : cancel.clone(),
1565 0 : );
1566 0 :
1567 0 : let heartbeater_sk = Heartbeater::new(
1568 0 : config.safekeeper_jwt_token.clone(),
1569 0 : config.max_offline_interval,
1570 0 : config.max_warming_up_interval,
1571 0 : cancel.clone(),
1572 0 : );
1573 :
1574 0 : let initial_leadership_status = if config.start_as_candidate {
1575 0 : LeadershipStatus::Candidate
1576 : } else {
1577 0 : LeadershipStatus::Leader
1578 : };
1579 :
1580 0 : let this = Arc::new(Self {
1581 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1582 0 : nodes,
1583 0 : safekeepers,
1584 0 : tenants,
1585 0 : scheduler,
1586 0 : delayed_reconcile_rx,
1587 0 : initial_leadership_status,
1588 0 : ))),
1589 0 : config: config.clone(),
1590 0 : persistence,
1591 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1592 0 : result_tx,
1593 0 : heartbeater_ps,
1594 0 : heartbeater_sk,
1595 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1596 0 : config.reconciler_concurrency,
1597 0 : )),
1598 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1599 0 : config.priority_reconciler_concurrency,
1600 0 : )),
1601 0 : delayed_reconcile_tx,
1602 0 : abort_tx,
1603 0 : startup_complete: startup_complete.clone(),
1604 0 : cancel,
1605 0 : reconcilers_cancel,
1606 0 : gate: Gate::default(),
1607 0 : reconcilers_gate: Gate::default(),
1608 0 : tenant_op_locks: Default::default(),
1609 0 : node_op_locks: Default::default(),
1610 0 : });
1611 0 :
1612 0 : let result_task_this = this.clone();
1613 0 : tokio::task::spawn(async move {
1614 : // Block shutdown until we're done (we must respect self.cancel)
1615 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1616 0 : result_task_this
1617 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1618 0 : .await
1619 0 : }
1620 0 : });
1621 0 :
1622 0 : tokio::task::spawn({
1623 0 : let this = this.clone();
1624 0 : async move {
1625 : // Block shutdown until we're done (we must respect self.cancel)
1626 0 : if let Ok(_gate) = this.gate.enter() {
1627 0 : this.process_aborts(abort_rx).await
1628 0 : }
1629 0 : }
1630 0 : });
1631 0 :
1632 0 : tokio::task::spawn({
1633 0 : let this = this.clone();
1634 0 : async move {
1635 0 : if let Ok(_gate) = this.gate.enter() {
1636 : loop {
1637 0 : tokio::select! {
1638 0 : _ = this.cancel.cancelled() => {
1639 0 : break;
1640 : },
1641 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1642 0 : };
1643 0 : this.tenant_op_locks.housekeeping();
1644 : }
1645 0 : }
1646 0 : }
1647 0 : });
1648 0 :
1649 0 : tokio::task::spawn({
1650 0 : let this = this.clone();
1651 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1652 0 : // is done.
1653 0 : let startup_completion = startup_completion.clone();
1654 0 : async move {
1655 : // Block shutdown until we're done (we must respect self.cancel)
1656 0 : let Ok(_gate) = this.gate.enter() else {
1657 0 : return;
1658 : };
1659 :
1660 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1661 0 : .await;
1662 :
1663 0 : drop(startup_completion);
1664 0 : }
1665 0 : });
1666 0 :
1667 0 : tokio::task::spawn({
1668 0 : let this = this.clone();
1669 0 : let startup_complete = startup_complete.clone();
1670 0 : async move {
1671 0 : startup_complete.wait().await;
1672 0 : this.background_reconcile().await;
1673 0 : }
1674 0 : });
1675 0 :
1676 0 : tokio::task::spawn({
1677 0 : let this = this.clone();
1678 0 : let startup_complete = startup_complete.clone();
1679 0 : async move {
1680 0 : startup_complete.wait().await;
1681 0 : this.spawn_heartbeat_driver().await;
1682 0 : }
1683 0 : });
1684 0 :
1685 0 : Ok(this)
1686 0 : }
1687 :
1688 0 : pub(crate) async fn attach_hook(
1689 0 : &self,
1690 0 : attach_req: AttachHookRequest,
1691 0 : ) -> anyhow::Result<AttachHookResponse> {
1692 0 : let _tenant_lock = trace_exclusive_lock(
1693 0 : &self.tenant_op_locks,
1694 0 : attach_req.tenant_shard_id.tenant_id,
1695 0 : TenantOperations::AttachHook,
1696 0 : )
1697 0 : .await;
1698 :
1699 : // This is a test hook. To enable using it on tenants that were created directly with
1700 : // the pageserver API (not via this service), we will auto-create any missing tenant
1701 : // shards with default state.
1702 0 : let insert = {
1703 0 : match self
1704 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1705 0 : .await
1706 : {
1707 0 : Ok(_) => false,
1708 0 : Err(ApiError::NotFound(_)) => true,
1709 0 : Err(e) => return Err(e.into()),
1710 : }
1711 : };
1712 :
1713 0 : if insert {
1714 0 : let tsp = TenantShardPersistence {
1715 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1716 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1717 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1718 0 : shard_stripe_size: 0,
1719 0 : generation: attach_req.generation_override.or(Some(0)),
1720 0 : generation_pageserver: None,
1721 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1722 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1723 0 : splitting: SplitState::default(),
1724 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1725 0 : .unwrap(),
1726 0 : preferred_az_id: None,
1727 0 : };
1728 0 :
1729 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1730 0 : Err(e) => match e {
1731 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1732 : DatabaseErrorKind::UniqueViolation,
1733 : _,
1734 : )) => {
1735 0 : tracing::info!(
1736 0 : "Raced with another request to insert tenant {}",
1737 : attach_req.tenant_shard_id
1738 : )
1739 : }
1740 0 : _ => return Err(e.into()),
1741 : },
1742 : Ok(()) => {
1743 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1744 :
1745 0 : let mut locked = self.inner.write().unwrap();
1746 0 : locked.tenants.insert(
1747 0 : attach_req.tenant_shard_id,
1748 0 : TenantShard::new(
1749 0 : attach_req.tenant_shard_id,
1750 0 : ShardIdentity::unsharded(),
1751 0 : PlacementPolicy::Attached(0),
1752 0 : None,
1753 0 : ),
1754 0 : );
1755 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1756 : }
1757 : }
1758 0 : }
1759 :
1760 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1761 0 : let maybe_tenant_conf = {
1762 0 : let locked = self.inner.write().unwrap();
1763 0 : locked
1764 0 : .tenants
1765 0 : .get(&attach_req.tenant_shard_id)
1766 0 : .map(|t| t.config.clone())
1767 0 : };
1768 0 :
1769 0 : match maybe_tenant_conf {
1770 0 : Some(conf) => {
1771 0 : let new_generation = self
1772 0 : .persistence
1773 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1774 0 : .await?;
1775 :
1776 : // Persist the placement policy update. This is required
1777 : // when we reattaching a detached tenant.
1778 0 : self.persistence
1779 0 : .update_tenant_shard(
1780 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1781 0 : Some(PlacementPolicy::Attached(0)),
1782 0 : Some(conf),
1783 0 : None,
1784 0 : None,
1785 0 : )
1786 0 : .await?;
1787 0 : Some(new_generation)
1788 : }
1789 : None => {
1790 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1791 : }
1792 : }
1793 : } else {
1794 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1795 0 : None
1796 : };
1797 :
1798 0 : let mut locked = self.inner.write().unwrap();
1799 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1800 0 :
1801 0 : let tenant_shard = tenants
1802 0 : .get_mut(&attach_req.tenant_shard_id)
1803 0 : .expect("Checked for existence above");
1804 :
1805 0 : if let Some(new_generation) = new_generation {
1806 0 : tenant_shard.generation = Some(new_generation);
1807 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1808 0 : } else {
1809 : // This is a detach notification. We must update placement policy to avoid re-attaching
1810 : // during background scheduling/reconciliation, or during storage controller restart.
1811 0 : assert!(attach_req.node_id.is_none());
1812 0 : tenant_shard.policy = PlacementPolicy::Detached;
1813 : }
1814 :
1815 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1816 0 : tracing::info!(
1817 : tenant_id = %attach_req.tenant_shard_id,
1818 : ps_id = %attaching_pageserver,
1819 : generation = ?tenant_shard.generation,
1820 0 : "issuing",
1821 : );
1822 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1823 0 : tracing::info!(
1824 : tenant_id = %attach_req.tenant_shard_id,
1825 : %ps_id,
1826 : generation = ?tenant_shard.generation,
1827 0 : "dropping",
1828 : );
1829 : } else {
1830 0 : tracing::info!(
1831 : tenant_id = %attach_req.tenant_shard_id,
1832 0 : "no-op: tenant already has no pageserver");
1833 : }
1834 0 : tenant_shard
1835 0 : .intent
1836 0 : .set_attached(scheduler, attach_req.node_id);
1837 0 :
1838 0 : tracing::info!(
1839 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1840 0 : attach_req.tenant_shard_id,
1841 0 : tenant_shard.generation,
1842 0 : // TODO: this is an odd number of 0xf's
1843 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1844 : );
1845 :
1846 : // Trick the reconciler into not doing anything for this tenant: this helps
1847 : // tests that manually configure a tenant on the pagesrever, and then call this
1848 : // attach hook: they don't want background reconciliation to modify what they
1849 : // did to the pageserver.
1850 : #[cfg(feature = "testing")]
1851 : {
1852 0 : if let Some(node_id) = attach_req.node_id {
1853 0 : tenant_shard.observed.locations = HashMap::from([(
1854 0 : node_id,
1855 0 : ObservedStateLocation {
1856 0 : conf: Some(attached_location_conf(
1857 0 : tenant_shard.generation.unwrap(),
1858 0 : &tenant_shard.shard,
1859 0 : &tenant_shard.config,
1860 0 : &PlacementPolicy::Attached(0),
1861 0 : )),
1862 0 : },
1863 0 : )]);
1864 0 : } else {
1865 0 : tenant_shard.observed.locations.clear();
1866 0 : }
1867 : }
1868 :
1869 0 : Ok(AttachHookResponse {
1870 0 : gen: attach_req
1871 0 : .node_id
1872 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1873 0 : })
1874 0 : }
1875 :
1876 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1877 0 : let locked = self.inner.read().unwrap();
1878 0 :
1879 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1880 0 :
1881 0 : InspectResponse {
1882 0 : attachment: tenant_shard.and_then(|s| {
1883 0 : s.intent
1884 0 : .get_attached()
1885 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1886 0 : }),
1887 0 : }
1888 0 : }
1889 :
1890 : // When the availability state of a node transitions to active, we must do a full reconciliation
1891 : // of LocationConfigs on that node. This is because while a node was offline:
1892 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1893 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1894 : //
1895 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1896 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1897 : // this function.
1898 : //
1899 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1900 : // for written for a single node rather than as a batch job for all nodes.
1901 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1902 : async fn node_activate_reconcile(
1903 : &self,
1904 : mut node: Node,
1905 : _lock: &TracingExclusiveGuard<NodeOperations>,
1906 : ) -> Result<(), ApiError> {
1907 : // This Node is a mutable local copy: we will set it active so that we can use its
1908 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1909 : // later.
1910 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1911 :
1912 : let configs = match node
1913 : .with_client_retries(
1914 0 : |client| async move { client.list_location_config().await },
1915 : &self.config.pageserver_jwt_token,
1916 : 1,
1917 : 5,
1918 : SHORT_RECONCILE_TIMEOUT,
1919 : &self.cancel,
1920 : )
1921 : .await
1922 : {
1923 : None => {
1924 : // We're shutting down (the Node's cancellation token can't have fired, because
1925 : // we're the only scope that has a reference to it, and we didn't fire it).
1926 : return Err(ApiError::ShuttingDown);
1927 : }
1928 : Some(Err(e)) => {
1929 : // This node didn't succeed listing its locations: it may not proceed to active state
1930 : // as it is apparently unavailable.
1931 : return Err(ApiError::PreconditionFailed(
1932 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1933 : ));
1934 : }
1935 : Some(Ok(configs)) => configs,
1936 : };
1937 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1938 :
1939 : let mut cleanup = Vec::new();
1940 : {
1941 : let mut locked = self.inner.write().unwrap();
1942 :
1943 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1944 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1945 : cleanup.push(tenant_shard_id);
1946 : continue;
1947 : };
1948 : tenant_shard
1949 : .observed
1950 : .locations
1951 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1952 : }
1953 : }
1954 :
1955 : for tenant_shard_id in cleanup {
1956 : tracing::info!("Detaching {tenant_shard_id}");
1957 : match node
1958 : .with_client_retries(
1959 0 : |client| async move {
1960 0 : let config = LocationConfig {
1961 0 : mode: LocationConfigMode::Detached,
1962 0 : generation: None,
1963 0 : secondary_conf: None,
1964 0 : shard_number: tenant_shard_id.shard_number.0,
1965 0 : shard_count: tenant_shard_id.shard_count.literal(),
1966 0 : shard_stripe_size: 0,
1967 0 : tenant_conf: models::TenantConfig::default(),
1968 0 : };
1969 0 : client
1970 0 : .location_config(tenant_shard_id, config, None, false)
1971 0 : .await
1972 0 : },
1973 : &self.config.pageserver_jwt_token,
1974 : 1,
1975 : 5,
1976 : SHORT_RECONCILE_TIMEOUT,
1977 : &self.cancel,
1978 : )
1979 : .await
1980 : {
1981 : None => {
1982 : // We're shutting down (the Node's cancellation token can't have fired, because
1983 : // we're the only scope that has a reference to it, and we didn't fire it).
1984 : return Err(ApiError::ShuttingDown);
1985 : }
1986 : Some(Err(e)) => {
1987 : // Do not let the node proceed to Active state if it is not responsive to requests
1988 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1989 : // detach completing: we should not let this node back into the set of nodes considered
1990 : // okay for scheduling.
1991 : return Err(ApiError::Conflict(format!(
1992 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1993 : )));
1994 : }
1995 : Some(Ok(_)) => {}
1996 : };
1997 : }
1998 :
1999 : Ok(())
2000 : }
2001 :
2002 0 : pub(crate) async fn re_attach(
2003 0 : &self,
2004 0 : reattach_req: ReAttachRequest,
2005 0 : ) -> Result<ReAttachResponse, ApiError> {
2006 0 : if let Some(register_req) = reattach_req.register {
2007 0 : self.node_register(register_req).await?;
2008 0 : }
2009 :
2010 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2011 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2012 :
2013 0 : tracing::info!(
2014 : node_id=%reattach_req.node_id,
2015 0 : "Incremented {} tenant shards' generations",
2016 0 : incremented_generations.len()
2017 : );
2018 :
2019 : // Apply the updated generation to our in-memory state, and
2020 : // gather discover secondary locations.
2021 0 : let mut locked = self.inner.write().unwrap();
2022 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2023 0 :
2024 0 : let mut response = ReAttachResponse {
2025 0 : tenants: Vec::new(),
2026 0 : };
2027 :
2028 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2029 : // to call location_conf API with an old generation. Wait for cancellation to complete
2030 : // before responding to this request. Requires well implemented CancellationToken logic
2031 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2032 : // request in flight over the network: TODO handle that by making location_conf API refuse
2033 : // to go backward in generations.
2034 :
2035 : // Scan through all shards, applying updates for ones where we updated generation
2036 : // and identifying shards that intend to have a secondary location on this node.
2037 0 : for (tenant_shard_id, shard) in tenants {
2038 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2039 0 : let new_gen = *new_gen;
2040 0 : response.tenants.push(ReAttachResponseTenant {
2041 0 : id: *tenant_shard_id,
2042 0 : gen: Some(new_gen.into().unwrap()),
2043 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2044 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2045 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2046 0 : // the stale/multi states at this point.
2047 0 : mode: LocationConfigMode::AttachedSingle,
2048 0 : });
2049 0 :
2050 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2051 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2052 : // Why can we update `observed` even though we're not sure our response will be received
2053 : // by the pageserver? Because the pageserver will not proceed with startup until
2054 : // it has processed response: if it loses it, we'll see another request and increment
2055 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2056 0 : if let Some(conf) = observed.conf.as_mut() {
2057 0 : conf.generation = new_gen.into();
2058 0 : }
2059 0 : } else {
2060 0 : // This node has no observed state for the shard: perhaps it was offline
2061 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2062 0 : // will be prompted to learn the location's state before it makes changes.
2063 0 : shard
2064 0 : .observed
2065 0 : .locations
2066 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2067 0 : }
2068 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2069 0 : // Ordering: pageserver will not accept /location_config requests until it has
2070 0 : // finished processing the response from re-attach. So we can update our in-memory state
2071 0 : // now, and be confident that we are not stamping on the result of some later location config.
2072 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2073 0 : // so we might update observed state here, and then get over-written by some racing
2074 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2075 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2076 0 :
2077 0 : response.tenants.push(ReAttachResponseTenant {
2078 0 : id: *tenant_shard_id,
2079 0 : gen: None,
2080 0 : mode: LocationConfigMode::Secondary,
2081 0 : });
2082 0 :
2083 0 : // We must not update observed, because we have no guarantee that our
2084 0 : // response will be received by the pageserver. This could leave it
2085 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2086 0 : }
2087 : }
2088 :
2089 : // We consider a node Active once we have composed a re-attach response, but we
2090 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2091 : // implicitly synchronizes the LocationConfigs on the node.
2092 : //
2093 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2094 : // but those requests will not be accepted by the node until it has finished processing
2095 : // the re-attach response.
2096 : //
2097 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2098 : // in [`Persistence::re_attach`].
2099 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2100 0 : let reset_scheduling = matches!(
2101 0 : node.get_scheduling(),
2102 : NodeSchedulingPolicy::PauseForRestart
2103 : | NodeSchedulingPolicy::Draining
2104 : | NodeSchedulingPolicy::Filling
2105 : );
2106 :
2107 0 : let mut new_nodes = (**nodes).clone();
2108 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2109 0 : if reset_scheduling {
2110 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2111 0 : }
2112 :
2113 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2114 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2115 0 :
2116 0 : scheduler.node_upsert(node);
2117 0 : let new_nodes = Arc::new(new_nodes);
2118 0 : *nodes = new_nodes;
2119 : } else {
2120 0 : tracing::error!(
2121 0 : "Reattaching node {} was removed while processing the request",
2122 : reattach_req.node_id
2123 : );
2124 : }
2125 0 : }
2126 :
2127 0 : Ok(response)
2128 0 : }
2129 :
2130 0 : pub(crate) async fn validate(
2131 0 : &self,
2132 0 : validate_req: ValidateRequest,
2133 0 : ) -> Result<ValidateResponse, DatabaseError> {
2134 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2135 : // in-memory generation for a shard
2136 0 : let in_memory_result = {
2137 0 : let mut in_memory_result = Vec::new();
2138 0 : let locked = self.inner.read().unwrap();
2139 0 : for req_tenant in validate_req.tenants {
2140 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2141 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
2142 0 : tracing::info!(
2143 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2144 : req_tenant.id,
2145 : req_tenant.gen,
2146 : tenant_shard.generation
2147 : );
2148 :
2149 0 : in_memory_result.push((req_tenant.id, Generation::new(req_tenant.gen), valid));
2150 : } else {
2151 : // This is legal: for example during a shard split the pageserver may still
2152 : // have deletions in its queue from the old pre-split shard, or after deletion
2153 : // of a tenant that was busy with compaction/gc while being deleted.
2154 0 : tracing::info!(
2155 0 : "Refusing deletion validation for missing shard {}",
2156 : req_tenant.id
2157 : );
2158 : }
2159 : }
2160 :
2161 0 : in_memory_result
2162 : };
2163 :
2164 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2165 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2166 0 : let db_generations = self
2167 0 : .persistence
2168 0 : .shard_generations(in_memory_result.iter().filter_map(|i| {
2169 0 : if i.2 {
2170 0 : Some(&i.0)
2171 : } else {
2172 0 : None
2173 : }
2174 0 : }))
2175 0 : .await?;
2176 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2177 0 :
2178 0 : let mut response = ValidateResponse {
2179 0 : tenants: Vec::new(),
2180 0 : };
2181 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2182 0 : let valid = if valid {
2183 0 : let db_generation = db_generations.get(&tenant_shard_id);
2184 0 : db_generation == Some(&Some(validate_generation))
2185 : } else {
2186 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2187 : // this prevents a pageserver from cleaning up an object in S3.
2188 0 : false
2189 : };
2190 :
2191 0 : response.tenants.push(ValidateResponseTenant {
2192 0 : id: tenant_shard_id,
2193 0 : valid,
2194 0 : })
2195 : }
2196 :
2197 0 : Ok(response)
2198 0 : }
2199 :
2200 0 : pub(crate) async fn tenant_create(
2201 0 : &self,
2202 0 : create_req: TenantCreateRequest,
2203 0 : ) -> Result<TenantCreateResponse, ApiError> {
2204 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2205 :
2206 : // Exclude any concurrent attempts to create/access the same tenant ID
2207 0 : let _tenant_lock = trace_exclusive_lock(
2208 0 : &self.tenant_op_locks,
2209 0 : create_req.new_tenant_id.tenant_id,
2210 0 : TenantOperations::Create,
2211 0 : )
2212 0 : .await;
2213 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2214 :
2215 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2216 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2217 : // accept compute notifications while it is in the process of creating. Reconciliation will
2218 : // be retried in the background.
2219 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2220 0 : }
2221 0 : Ok(response)
2222 0 : }
2223 :
2224 0 : pub(crate) async fn do_tenant_create(
2225 0 : &self,
2226 0 : create_req: TenantCreateRequest,
2227 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2228 0 : let placement_policy = create_req
2229 0 : .placement_policy
2230 0 : .clone()
2231 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2232 0 : .unwrap_or(PlacementPolicy::Attached(0));
2233 :
2234 : // This service expects to handle sharding itself: it is an error to try and directly create
2235 : // a particular shard here.
2236 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2237 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2238 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2239 0 : )));
2240 : } else {
2241 0 : create_req.new_tenant_id.tenant_id
2242 0 : };
2243 0 :
2244 0 : tracing::info!(
2245 0 : "Creating tenant {}, shard_count={:?}",
2246 : create_req.new_tenant_id,
2247 : create_req.shard_parameters.count,
2248 : );
2249 :
2250 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2251 0 : .map(|i| TenantShardId {
2252 0 : tenant_id,
2253 0 : shard_number: ShardNumber(i),
2254 0 : shard_count: create_req.shard_parameters.count,
2255 0 : })
2256 0 : .collect::<Vec<_>>();
2257 :
2258 : // If the caller specifies a None generation, it means "start from default". This is different
2259 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2260 : // an incompletely-onboarded tenant.
2261 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2262 0 : tracing::info!(
2263 0 : "tenant_create: secondary mode, generation is_some={}",
2264 0 : create_req.generation.is_some()
2265 : );
2266 0 : create_req.generation.map(Generation::new)
2267 : } else {
2268 0 : tracing::info!(
2269 0 : "tenant_create: not secondary mode, generation is_some={}",
2270 0 : create_req.generation.is_some()
2271 : );
2272 0 : Some(
2273 0 : create_req
2274 0 : .generation
2275 0 : .map(Generation::new)
2276 0 : .unwrap_or(INITIAL_GENERATION),
2277 0 : )
2278 : };
2279 :
2280 0 : let preferred_az_id = {
2281 0 : let locked = self.inner.read().unwrap();
2282 : // Idempotency: take the existing value if the tenant already exists
2283 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2284 0 : shard.preferred_az().cloned()
2285 : } else {
2286 0 : locked.scheduler.get_az_for_new_tenant()
2287 : }
2288 : };
2289 :
2290 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2291 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2292 : // during the creation, rather than risking leaving orphan objects in S3.
2293 0 : let persist_tenant_shards = create_ids
2294 0 : .iter()
2295 0 : .map(|tenant_shard_id| TenantShardPersistence {
2296 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2297 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2298 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2299 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2300 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2301 0 : // The pageserver is not known until scheduling happens: we will set this column when
2302 0 : // incrementing the generation the first time we attach to a pageserver.
2303 0 : generation_pageserver: None,
2304 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2305 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2306 0 : splitting: SplitState::default(),
2307 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2308 0 : .unwrap(),
2309 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2310 0 : })
2311 0 : .collect();
2312 0 :
2313 0 : match self
2314 0 : .persistence
2315 0 : .insert_tenant_shards(persist_tenant_shards)
2316 0 : .await
2317 : {
2318 0 : Ok(_) => {}
2319 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2320 : DatabaseErrorKind::UniqueViolation,
2321 : _,
2322 : ))) => {
2323 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2324 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2325 : // creation's shard count.
2326 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
2327 : }
2328 : // Any other database error is unexpected and a bug.
2329 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2330 : };
2331 :
2332 0 : let mut schedule_context = ScheduleContext::default();
2333 0 : let mut schedule_error = None;
2334 0 : let mut response_shards = Vec::new();
2335 0 : for tenant_shard_id in create_ids {
2336 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2337 :
2338 0 : let outcome = self
2339 0 : .do_initial_shard_scheduling(
2340 0 : tenant_shard_id,
2341 0 : initial_generation,
2342 0 : &create_req.shard_parameters,
2343 0 : create_req.config.clone(),
2344 0 : placement_policy.clone(),
2345 0 : preferred_az_id.as_ref(),
2346 0 : &mut schedule_context,
2347 0 : )
2348 0 : .await;
2349 :
2350 0 : match outcome {
2351 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2352 0 : InitialShardScheduleOutcome::NotScheduled => {}
2353 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2354 0 : schedule_error = Some(err);
2355 0 : }
2356 : }
2357 : }
2358 :
2359 : // If we failed to schedule shards, then they are still created in the controller,
2360 : // but we return an error to the requester to avoid a silent failure when someone
2361 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2362 : // are present in the system. We do this here rather than in the above loop, to
2363 : // avoid situations where we only create a subset of shards in the tenant.
2364 0 : if let Some(e) = schedule_error {
2365 0 : return Err(ApiError::Conflict(format!(
2366 0 : "Failed to schedule shard(s): {e}"
2367 0 : )));
2368 0 : }
2369 0 :
2370 0 : let waiters = {
2371 0 : let mut locked = self.inner.write().unwrap();
2372 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2373 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2374 0 : .tenant_creation_hint(true)
2375 0 : .build();
2376 0 : tenants
2377 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2378 0 : .filter_map(|(_shard_id, shard)| {
2379 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2380 0 : })
2381 0 : .collect::<Vec<_>>()
2382 0 : };
2383 0 :
2384 0 : Ok((
2385 0 : TenantCreateResponse {
2386 0 : shards: response_shards,
2387 0 : },
2388 0 : waiters,
2389 0 : ))
2390 0 : }
2391 :
2392 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2393 : /// case of a new tenant and a pre-existing one.
2394 : #[allow(clippy::too_many_arguments)]
2395 0 : async fn do_initial_shard_scheduling(
2396 0 : &self,
2397 0 : tenant_shard_id: TenantShardId,
2398 0 : initial_generation: Option<Generation>,
2399 0 : shard_params: &ShardParameters,
2400 0 : config: TenantConfig,
2401 0 : placement_policy: PlacementPolicy,
2402 0 : preferred_az_id: Option<&AvailabilityZone>,
2403 0 : schedule_context: &mut ScheduleContext,
2404 0 : ) -> InitialShardScheduleOutcome {
2405 0 : let mut locked = self.inner.write().unwrap();
2406 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2407 :
2408 : use std::collections::btree_map::Entry;
2409 0 : match tenants.entry(tenant_shard_id) {
2410 0 : Entry::Occupied(mut entry) => {
2411 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2412 :
2413 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2414 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2415 0 : }
2416 :
2417 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2418 0 : let generation = entry
2419 0 : .get()
2420 0 : .generation
2421 0 : .expect("Generation is set when in attached mode");
2422 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2423 0 : shard_id: tenant_shard_id,
2424 0 : node_id: *node_id,
2425 0 : generation: generation.into().unwrap(),
2426 0 : })
2427 : } else {
2428 0 : InitialShardScheduleOutcome::NotScheduled
2429 : }
2430 : }
2431 0 : Entry::Vacant(entry) => {
2432 0 : let state = entry.insert(TenantShard::new(
2433 0 : tenant_shard_id,
2434 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2435 0 : placement_policy,
2436 0 : preferred_az_id.cloned(),
2437 0 : ));
2438 0 :
2439 0 : state.generation = initial_generation;
2440 0 : state.config = config;
2441 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2442 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2443 0 : }
2444 :
2445 : // Only include shards in result if we are attaching: the purpose
2446 : // of the response is to tell the caller where the shards are attached.
2447 0 : if let Some(node_id) = state.intent.get_attached() {
2448 0 : let generation = state
2449 0 : .generation
2450 0 : .expect("Generation is set when in attached mode");
2451 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2452 0 : shard_id: tenant_shard_id,
2453 0 : node_id: *node_id,
2454 0 : generation: generation.into().unwrap(),
2455 0 : })
2456 : } else {
2457 0 : InitialShardScheduleOutcome::NotScheduled
2458 : }
2459 : }
2460 : }
2461 0 : }
2462 :
2463 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2464 : /// wait for reconciliation to complete before responding.
2465 0 : async fn await_waiters(
2466 0 : &self,
2467 0 : waiters: Vec<ReconcilerWaiter>,
2468 0 : timeout: Duration,
2469 0 : ) -> Result<(), ReconcileWaitError> {
2470 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2471 0 : for waiter in waiters {
2472 0 : let timeout = deadline.duration_since(Instant::now());
2473 0 : waiter.wait_timeout(timeout).await?;
2474 : }
2475 :
2476 0 : Ok(())
2477 0 : }
2478 :
2479 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2480 : /// in progress
2481 0 : async fn await_waiters_remainder(
2482 0 : &self,
2483 0 : waiters: Vec<ReconcilerWaiter>,
2484 0 : timeout: Duration,
2485 0 : ) -> Vec<ReconcilerWaiter> {
2486 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2487 0 : for waiter in waiters.iter() {
2488 0 : let timeout = deadline.duration_since(Instant::now());
2489 0 : let _ = waiter.wait_timeout(timeout).await;
2490 : }
2491 :
2492 0 : waiters
2493 0 : .into_iter()
2494 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2495 0 : .collect::<Vec<_>>()
2496 0 : }
2497 :
2498 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2499 : /// and transform it into either a tenant creation of a series of shard updates.
2500 : ///
2501 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2502 : /// still be returned.
2503 0 : fn tenant_location_config_prepare(
2504 0 : &self,
2505 0 : tenant_id: TenantId,
2506 0 : req: TenantLocationConfigRequest,
2507 0 : ) -> TenantCreateOrUpdate {
2508 0 : let mut updates = Vec::new();
2509 0 : let mut locked = self.inner.write().unwrap();
2510 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2511 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2512 :
2513 : // Use location config mode as an indicator of policy.
2514 0 : let placement_policy = match req.config.mode {
2515 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2516 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2517 : LocationConfigMode::AttachedMulti
2518 : | LocationConfigMode::AttachedSingle
2519 : | LocationConfigMode::AttachedStale => {
2520 0 : if nodes.len() > 1 {
2521 0 : PlacementPolicy::Attached(1)
2522 : } else {
2523 : // Convenience for dev/test: if we just have one pageserver, import
2524 : // tenants into non-HA mode so that scheduling will succeed.
2525 0 : PlacementPolicy::Attached(0)
2526 : }
2527 : }
2528 : };
2529 :
2530 : // Ordinarily we do not update scheduling policy, but when making major changes
2531 : // like detaching or demoting to secondary-only, we need to force the scheduling
2532 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2533 0 : let scheduling_policy = match req.config.mode {
2534 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2535 : // Special case: when making major changes like detaching or demoting to secondary-only,
2536 : // we need to force the scheduling mode to Active, or nothing will happen.
2537 0 : Some(ShardSchedulingPolicy::Active)
2538 : }
2539 : LocationConfigMode::AttachedMulti
2540 : | LocationConfigMode::AttachedSingle
2541 : | LocationConfigMode::AttachedStale => {
2542 : // While attached, continue to respect whatever the existing scheduling mode is.
2543 0 : None
2544 : }
2545 : };
2546 :
2547 0 : let mut create = true;
2548 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2549 : // Saw an existing shard: this is not a creation
2550 0 : create = false;
2551 :
2552 : // Shards may have initially been created by a Secondary request, where we
2553 : // would have left generation as None.
2554 : //
2555 : // We only update generation the first time we see an attached-mode request,
2556 : // and if there is no existing generation set. The caller is responsible for
2557 : // ensuring that no non-storage-controller pageserver ever uses a higher
2558 : // generation than they passed in here.
2559 : use LocationConfigMode::*;
2560 0 : let set_generation = match req.config.mode {
2561 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2562 0 : req.config.generation.map(Generation::new)
2563 : }
2564 0 : _ => None,
2565 : };
2566 :
2567 0 : updates.push(ShardUpdate {
2568 0 : tenant_shard_id: *shard_id,
2569 0 : placement_policy: placement_policy.clone(),
2570 0 : tenant_config: req.config.tenant_conf.clone(),
2571 0 : generation: set_generation,
2572 0 : scheduling_policy,
2573 0 : });
2574 : }
2575 :
2576 0 : if create {
2577 : use LocationConfigMode::*;
2578 0 : let generation = match req.config.mode {
2579 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2580 : // If a caller provided a generation in a non-attached request, ignore it
2581 : // and leave our generation as None: this enables a subsequent update to set
2582 : // the generation when setting an attached mode for the first time.
2583 0 : _ => None,
2584 : };
2585 :
2586 0 : TenantCreateOrUpdate::Create(
2587 0 : // Synthesize a creation request
2588 0 : TenantCreateRequest {
2589 0 : new_tenant_id: tenant_shard_id,
2590 0 : generation,
2591 0 : shard_parameters: ShardParameters {
2592 0 : count: tenant_shard_id.shard_count,
2593 0 : // We only import un-sharded or single-sharded tenants, so stripe
2594 0 : // size can be made up arbitrarily here.
2595 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2596 0 : },
2597 0 : placement_policy: Some(placement_policy),
2598 0 : config: req.config.tenant_conf,
2599 0 : },
2600 0 : )
2601 : } else {
2602 0 : assert!(!updates.is_empty());
2603 0 : TenantCreateOrUpdate::Update(updates)
2604 : }
2605 0 : }
2606 :
2607 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2608 : /// the tenant is present in memory. If not, load it from the database. If it is found
2609 : /// in neither location, return a NotFound error.
2610 : ///
2611 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2612 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2613 0 : async fn maybe_load_tenant(
2614 0 : &self,
2615 0 : tenant_id: TenantId,
2616 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2617 0 : ) -> Result<(), ApiError> {
2618 : // Check if the tenant is present in memory, and select an AZ to use when loading
2619 : // if we will load it.
2620 0 : let load_in_az = {
2621 0 : let locked = self.inner.read().unwrap();
2622 0 : let existing = locked
2623 0 : .tenants
2624 0 : .range(TenantShardId::tenant_range(tenant_id))
2625 0 : .next();
2626 0 :
2627 0 : // If the tenant is not present in memory, we expect to load it from database,
2628 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2629 0 : if existing.is_none() {
2630 0 : locked
2631 0 : .scheduler
2632 0 : .get_az_for_new_tenant()
2633 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2634 0 : "No AZ with nodes found to load tenant"
2635 0 : )))?
2636 : } else {
2637 : // We already have this tenant in memory
2638 0 : return Ok(());
2639 : }
2640 : };
2641 :
2642 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2643 0 : if tenant_shards.is_empty() {
2644 0 : return Err(ApiError::NotFound(
2645 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2646 0 : ));
2647 0 : }
2648 0 :
2649 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2650 0 : self.persistence
2651 0 : .set_tenant_shard_preferred_azs(
2652 0 : tenant_shards
2653 0 : .iter()
2654 0 : .map(|t| {
2655 0 : (
2656 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2657 0 : Some(load_in_az.clone()),
2658 0 : )
2659 0 : })
2660 0 : .collect(),
2661 0 : )
2662 0 : .await?;
2663 :
2664 0 : let mut locked = self.inner.write().unwrap();
2665 0 : tracing::info!(
2666 0 : "Loaded {} shards for tenant {}",
2667 0 : tenant_shards.len(),
2668 : tenant_id
2669 : );
2670 :
2671 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2672 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2673 0 : let shard =
2674 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2675 0 :
2676 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2677 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2678 0 : if shard.policy != PlacementPolicy::Detached {
2679 0 : tracing::error!(
2680 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2681 : shard.tenant_shard_id,
2682 : shard.policy
2683 : );
2684 0 : }
2685 :
2686 0 : (shard.tenant_shard_id, shard)
2687 0 : }));
2688 0 :
2689 0 : Ok(())
2690 0 : }
2691 :
2692 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2693 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2694 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2695 : ///
2696 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2697 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2698 0 : fn maybe_drop_tenant(
2699 0 : &self,
2700 0 : tenant_id: TenantId,
2701 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2702 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2703 0 : ) {
2704 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2705 0 : if tenant_shards.all(|(_id, shard)| {
2706 0 : shard.policy == PlacementPolicy::Detached
2707 0 : && shard.reconciler.is_none()
2708 0 : && shard.observed.is_empty()
2709 0 : }) {
2710 0 : let keys = locked
2711 0 : .tenants
2712 0 : .range(TenantShardId::tenant_range(tenant_id))
2713 0 : .map(|(id, _)| id)
2714 0 : .copied()
2715 0 : .collect::<Vec<_>>();
2716 0 : for key in keys {
2717 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2718 0 : locked.tenants.remove(&key);
2719 : }
2720 0 : }
2721 0 : }
2722 :
2723 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2724 : /// directly with pageservers into this service.
2725 : ///
2726 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2727 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2728 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2729 : /// tenant's source of generation numbers.
2730 : ///
2731 : /// The mode in this request coarse-grained control of tenants:
2732 : /// - Call with mode Attached* to upsert the tenant.
2733 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2734 : /// to set an existing tenant to PolicyMode::Secondary
2735 : /// - Call with mode Detached to switch to PolicyMode::Detached
2736 0 : pub(crate) async fn tenant_location_config(
2737 0 : &self,
2738 0 : tenant_shard_id: TenantShardId,
2739 0 : req: TenantLocationConfigRequest,
2740 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2741 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2742 0 : let _tenant_lock = trace_exclusive_lock(
2743 0 : &self.tenant_op_locks,
2744 0 : tenant_shard_id.tenant_id,
2745 0 : TenantOperations::LocationConfig,
2746 0 : )
2747 0 : .await;
2748 :
2749 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2750 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2751 0 : "This API is for importing single-sharded or unsharded tenants"
2752 0 : )));
2753 : } else {
2754 0 : tenant_shard_id.tenant_id
2755 0 : };
2756 0 :
2757 0 : // In case we are waking up a Detached tenant
2758 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2759 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2760 0 : // This is a creation or an update
2761 0 : }
2762 0 : Err(e) => {
2763 0 : return Err(e);
2764 : }
2765 : };
2766 :
2767 : // First check if this is a creation or an update
2768 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2769 0 :
2770 0 : let mut result = TenantLocationConfigResponse {
2771 0 : shards: Vec::new(),
2772 0 : stripe_size: None,
2773 0 : };
2774 0 : let waiters = match create_or_update {
2775 0 : TenantCreateOrUpdate::Create(create_req) => {
2776 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2777 0 : result.shards = create_resp
2778 0 : .shards
2779 0 : .into_iter()
2780 0 : .map(|s| TenantShardLocation {
2781 0 : node_id: s.node_id,
2782 0 : shard_id: s.shard_id,
2783 0 : })
2784 0 : .collect();
2785 0 : waiters
2786 : }
2787 0 : TenantCreateOrUpdate::Update(updates) => {
2788 0 : // Persist updates
2789 0 : // Ordering: write to the database before applying changes in-memory, so that
2790 0 : // we will not appear time-travel backwards on a restart.
2791 0 :
2792 0 : let mut schedule_context = ScheduleContext::default();
2793 : for ShardUpdate {
2794 0 : tenant_shard_id,
2795 0 : placement_policy,
2796 0 : tenant_config,
2797 0 : generation,
2798 0 : scheduling_policy,
2799 0 : } in &updates
2800 : {
2801 0 : self.persistence
2802 0 : .update_tenant_shard(
2803 0 : TenantFilter::Shard(*tenant_shard_id),
2804 0 : Some(placement_policy.clone()),
2805 0 : Some(tenant_config.clone()),
2806 0 : *generation,
2807 0 : *scheduling_policy,
2808 0 : )
2809 0 : .await?;
2810 : }
2811 :
2812 : // Apply updates in-memory
2813 0 : let mut waiters = Vec::new();
2814 0 : {
2815 0 : let mut locked = self.inner.write().unwrap();
2816 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2817 :
2818 : for ShardUpdate {
2819 0 : tenant_shard_id,
2820 0 : placement_policy,
2821 0 : tenant_config,
2822 0 : generation: update_generation,
2823 0 : scheduling_policy,
2824 0 : } in updates
2825 : {
2826 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2827 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2828 0 : continue;
2829 : };
2830 :
2831 : // Update stripe size
2832 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2833 0 : result.stripe_size = Some(shard.shard.stripe_size);
2834 0 : }
2835 :
2836 0 : shard.policy = placement_policy;
2837 0 : shard.config = tenant_config;
2838 0 : if let Some(generation) = update_generation {
2839 0 : shard.generation = Some(generation);
2840 0 : }
2841 :
2842 0 : if let Some(scheduling_policy) = scheduling_policy {
2843 0 : shard.set_scheduling_policy(scheduling_policy);
2844 0 : }
2845 :
2846 0 : shard.schedule(scheduler, &mut schedule_context)?;
2847 :
2848 0 : let maybe_waiter =
2849 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
2850 0 : if let Some(waiter) = maybe_waiter {
2851 0 : waiters.push(waiter);
2852 0 : }
2853 :
2854 0 : if let Some(node_id) = shard.intent.get_attached() {
2855 0 : result.shards.push(TenantShardLocation {
2856 0 : shard_id: tenant_shard_id,
2857 0 : node_id: *node_id,
2858 0 : })
2859 0 : }
2860 : }
2861 : }
2862 0 : waiters
2863 : }
2864 : };
2865 :
2866 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2867 : // Do not treat a reconcile error as fatal: we have already applied any requested
2868 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2869 : // compute notification API. In these cases, it is important that we do not
2870 : // cause the cloud control plane to retry forever on this API.
2871 0 : tracing::warn!(
2872 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2873 : );
2874 0 : }
2875 :
2876 : // Logging the full result is useful because it lets us cross-check what the cloud control
2877 : // plane's tenant_shards table should contain.
2878 0 : tracing::info!("Complete, returning {result:?}");
2879 :
2880 0 : Ok(result)
2881 0 : }
2882 :
2883 0 : pub(crate) async fn tenant_config_patch(
2884 0 : &self,
2885 0 : req: TenantConfigPatchRequest,
2886 0 : ) -> Result<(), ApiError> {
2887 0 : let _tenant_lock = trace_exclusive_lock(
2888 0 : &self.tenant_op_locks,
2889 0 : req.tenant_id,
2890 0 : TenantOperations::ConfigPatch,
2891 0 : )
2892 0 : .await;
2893 :
2894 0 : let tenant_id = req.tenant_id;
2895 0 : let patch = req.config;
2896 0 :
2897 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
2898 :
2899 0 : let base = {
2900 0 : let locked = self.inner.read().unwrap();
2901 0 : let shards = locked
2902 0 : .tenants
2903 0 : .range(TenantShardId::tenant_range(req.tenant_id));
2904 0 :
2905 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
2906 :
2907 0 : let first = match configs.peek() {
2908 0 : Some(first) => (*first).clone(),
2909 : None => {
2910 0 : return Err(ApiError::NotFound(
2911 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
2912 0 : ));
2913 : }
2914 : };
2915 :
2916 0 : if !configs.all_equal() {
2917 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
2918 : // This can't happen because we atomically update the database records
2919 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
2920 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2921 0 : "Tenant configs for {} are mismatched",
2922 0 : req.tenant_id
2923 0 : )));
2924 0 : }
2925 0 :
2926 0 : first
2927 0 : };
2928 0 :
2929 0 : let updated_config = base.apply_patch(patch);
2930 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
2931 0 : .await
2932 0 : }
2933 :
2934 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2935 : // We require an exclusive lock, because we are updating persistent and in-memory state
2936 0 : let _tenant_lock = trace_exclusive_lock(
2937 0 : &self.tenant_op_locks,
2938 0 : req.tenant_id,
2939 0 : TenantOperations::ConfigSet,
2940 0 : )
2941 0 : .await;
2942 :
2943 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
2944 :
2945 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
2946 0 : .await
2947 0 : }
2948 :
2949 0 : async fn set_tenant_config_and_reconcile(
2950 0 : &self,
2951 0 : tenant_id: TenantId,
2952 0 : config: TenantConfig,
2953 0 : ) -> Result<(), ApiError> {
2954 0 : self.persistence
2955 0 : .update_tenant_shard(
2956 0 : TenantFilter::Tenant(tenant_id),
2957 0 : None,
2958 0 : Some(config.clone()),
2959 0 : None,
2960 0 : None,
2961 0 : )
2962 0 : .await?;
2963 :
2964 0 : let waiters = {
2965 0 : let mut waiters = Vec::new();
2966 0 : let mut locked = self.inner.write().unwrap();
2967 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2968 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2969 0 : shard.config = config.clone();
2970 0 : if let Some(waiter) =
2971 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
2972 0 : {
2973 0 : waiters.push(waiter);
2974 0 : }
2975 : }
2976 0 : waiters
2977 : };
2978 :
2979 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2980 : // Treat this as success because we have stored the configuration. If e.g.
2981 : // a node was unavailable at this time, it should not stop us accepting a
2982 : // configuration change.
2983 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2984 0 : }
2985 :
2986 0 : Ok(())
2987 0 : }
2988 :
2989 0 : pub(crate) fn tenant_config_get(
2990 0 : &self,
2991 0 : tenant_id: TenantId,
2992 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2993 0 : let config = {
2994 0 : let locked = self.inner.read().unwrap();
2995 0 :
2996 0 : match locked
2997 0 : .tenants
2998 0 : .range(TenantShardId::tenant_range(tenant_id))
2999 0 : .next()
3000 : {
3001 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3002 : None => {
3003 0 : return Err(ApiError::NotFound(
3004 0 : anyhow::anyhow!("Tenant not found").into(),
3005 0 : ))
3006 : }
3007 : }
3008 : };
3009 :
3010 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3011 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3012 : // and `effective_config` in the response is meaningless, but we retain that syntax
3013 : // in order to remain compatible with the pageserver API.
3014 :
3015 0 : let response = HashMap::from([
3016 : (
3017 : "tenant_specific_overrides",
3018 0 : serde_json::to_value(&config)
3019 0 : .context("serializing tenant specific overrides")
3020 0 : .map_err(ApiError::InternalServerError)?,
3021 : ),
3022 : (
3023 0 : "effective_config",
3024 0 : serde_json::to_value(&config)
3025 0 : .context("serializing effective config")
3026 0 : .map_err(ApiError::InternalServerError)?,
3027 : ),
3028 : ]);
3029 :
3030 0 : Ok(response)
3031 0 : }
3032 :
3033 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3034 0 : &self,
3035 0 : time_travel_req: &TenantTimeTravelRequest,
3036 0 : tenant_id: TenantId,
3037 0 : timestamp: Cow<'_, str>,
3038 0 : done_if_after: Cow<'_, str>,
3039 0 : ) -> Result<(), ApiError> {
3040 0 : let _tenant_lock = trace_exclusive_lock(
3041 0 : &self.tenant_op_locks,
3042 0 : tenant_id,
3043 0 : TenantOperations::TimeTravelRemoteStorage,
3044 0 : )
3045 0 : .await;
3046 :
3047 0 : let node = {
3048 0 : let mut locked = self.inner.write().unwrap();
3049 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3050 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3051 : // but only at the start of the process, so it's really just to prevent operator
3052 : // mistakes.
3053 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3054 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3055 : {
3056 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3057 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3058 0 : )));
3059 0 : }
3060 0 : let maybe_attached = shard
3061 0 : .observed
3062 0 : .locations
3063 0 : .iter()
3064 0 : .filter_map(|(node_id, observed_location)| {
3065 0 : observed_location
3066 0 : .conf
3067 0 : .as_ref()
3068 0 : .map(|loc| (node_id, observed_location, loc.mode))
3069 0 : })
3070 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3071 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3072 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
3073 0 : }
3074 : }
3075 0 : let scheduler = &mut locked.scheduler;
3076 : // Right now we only perform the operation on a single node without parallelization
3077 : // TODO fan out the operation to multiple nodes for better performance
3078 0 : let node_id = scheduler.any_available_node()?;
3079 0 : let node = locked
3080 0 : .nodes
3081 0 : .get(&node_id)
3082 0 : .expect("Pageservers may not be deleted while lock is active");
3083 0 : node.clone()
3084 0 : };
3085 0 :
3086 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3087 0 : let mut counts = time_travel_req
3088 0 : .shard_counts
3089 0 : .iter()
3090 0 : .copied()
3091 0 : .collect::<HashSet<_>>()
3092 0 : .into_iter()
3093 0 : .collect::<Vec<_>>();
3094 0 : counts.sort_unstable();
3095 :
3096 0 : for count in counts {
3097 0 : let shard_ids = (0..count.count())
3098 0 : .map(|i| TenantShardId {
3099 0 : tenant_id,
3100 0 : shard_number: ShardNumber(i),
3101 0 : shard_count: count,
3102 0 : })
3103 0 : .collect::<Vec<_>>();
3104 0 : for tenant_shard_id in shard_ids {
3105 0 : let client = PageserverClient::new(
3106 0 : node.get_id(),
3107 0 : node.base_url(),
3108 0 : self.config.pageserver_jwt_token.as_deref(),
3109 0 : );
3110 0 :
3111 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3112 :
3113 0 : client
3114 0 : .tenant_time_travel_remote_storage(
3115 0 : tenant_shard_id,
3116 0 : ×tamp,
3117 0 : &done_if_after,
3118 0 : )
3119 0 : .await
3120 0 : .map_err(|e| {
3121 0 : ApiError::InternalServerError(anyhow::anyhow!(
3122 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3123 0 : node
3124 0 : ))
3125 0 : })?;
3126 : }
3127 : }
3128 0 : Ok(())
3129 0 : }
3130 :
3131 0 : pub(crate) async fn tenant_secondary_download(
3132 0 : &self,
3133 0 : tenant_id: TenantId,
3134 0 : wait: Option<Duration>,
3135 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3136 0 : let _tenant_lock = trace_shared_lock(
3137 0 : &self.tenant_op_locks,
3138 0 : tenant_id,
3139 0 : TenantOperations::SecondaryDownload,
3140 0 : )
3141 0 : .await;
3142 :
3143 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3144 0 : let targets = {
3145 0 : let locked = self.inner.read().unwrap();
3146 0 : let mut targets = Vec::new();
3147 :
3148 0 : for (tenant_shard_id, shard) in
3149 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3150 : {
3151 0 : for node_id in shard.intent.get_secondary() {
3152 0 : let node = locked
3153 0 : .nodes
3154 0 : .get(node_id)
3155 0 : .expect("Pageservers may not be deleted while referenced");
3156 0 :
3157 0 : targets.push((*tenant_shard_id, node.clone()));
3158 0 : }
3159 : }
3160 0 : targets
3161 0 : };
3162 0 :
3163 0 : // Issue concurrent requests to all shards' locations
3164 0 : let mut futs = FuturesUnordered::new();
3165 0 : for (tenant_shard_id, node) in targets {
3166 0 : let client = PageserverClient::new(
3167 0 : node.get_id(),
3168 0 : node.base_url(),
3169 0 : self.config.pageserver_jwt_token.as_deref(),
3170 0 : );
3171 0 : futs.push(async move {
3172 0 : let result = client
3173 0 : .tenant_secondary_download(tenant_shard_id, wait)
3174 0 : .await;
3175 0 : (result, node, tenant_shard_id)
3176 0 : })
3177 : }
3178 :
3179 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3180 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3181 : // well as more general cases like 503s, 500s, or timeouts.
3182 0 : let mut aggregate_progress = SecondaryProgress::default();
3183 0 : let mut aggregate_status: Option<StatusCode> = None;
3184 0 : let mut error: Option<mgmt_api::Error> = None;
3185 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3186 0 : match result {
3187 0 : Err(e) => {
3188 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3189 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3190 0 : // than they had hoped for.
3191 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3192 0 : error = Some(e)
3193 : }
3194 0 : Ok((status_code, progress)) => {
3195 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3196 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3197 0 : aggregate_progress.layers_total += progress.layers_total;
3198 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3199 0 : aggregate_progress.bytes_total += progress.bytes_total;
3200 0 : aggregate_progress.heatmap_mtime =
3201 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3202 0 : aggregate_status = match aggregate_status {
3203 0 : None => Some(status_code),
3204 0 : Some(StatusCode::OK) => Some(status_code),
3205 0 : Some(cur) => {
3206 0 : // Other status codes (e.g. 202) -- do not overwrite.
3207 0 : Some(cur)
3208 : }
3209 : };
3210 : }
3211 : }
3212 : }
3213 :
3214 : // If any of the shards return 202, indicate our result as 202.
3215 0 : match aggregate_status {
3216 : None => {
3217 0 : match error {
3218 0 : Some(e) => {
3219 0 : // No successes, and an error: surface it
3220 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3221 : }
3222 : None => {
3223 : // No shards found
3224 0 : Err(ApiError::NotFound(
3225 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3226 0 : ))
3227 : }
3228 : }
3229 : }
3230 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3231 : }
3232 0 : }
3233 :
3234 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
3235 0 : let _tenant_lock =
3236 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3237 :
3238 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3239 :
3240 : // Detach all shards. This also deletes local pageserver shard data.
3241 0 : let (detach_waiters, node) = {
3242 0 : let mut detach_waiters = Vec::new();
3243 0 : let mut locked = self.inner.write().unwrap();
3244 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3245 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3246 : // Update the tenant's intent to remove all attachments
3247 0 : shard.policy = PlacementPolicy::Detached;
3248 0 : shard
3249 0 : .schedule(scheduler, &mut ScheduleContext::default())
3250 0 : .expect("De-scheduling is infallible");
3251 0 : debug_assert!(shard.intent.get_attached().is_none());
3252 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3253 :
3254 0 : if let Some(waiter) =
3255 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3256 0 : {
3257 0 : detach_waiters.push(waiter);
3258 0 : }
3259 : }
3260 :
3261 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3262 : // was attached, just has to be able to see the S3 content)
3263 0 : let node_id = scheduler.any_available_node()?;
3264 0 : let node = nodes
3265 0 : .get(&node_id)
3266 0 : .expect("Pageservers may not be deleted while lock is active");
3267 0 : (detach_waiters, node.clone())
3268 0 : };
3269 0 :
3270 0 : // This reconcile wait can fail in a few ways:
3271 0 : // A there is a very long queue for the reconciler semaphore
3272 0 : // B some pageserver is failing to handle a detach promptly
3273 0 : // C some pageserver goes offline right at the moment we send it a request.
3274 0 : //
3275 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3276 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3277 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3278 0 : // deleting the underlying data).
3279 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3280 0 : .await?;
3281 :
3282 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3283 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3284 : // the tenant ID prefix, including all shards (even possibly stale ones).
3285 0 : match node
3286 0 : .with_client_retries(
3287 0 : |client| async move {
3288 0 : client
3289 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3290 0 : .await
3291 0 : },
3292 0 : &self.config.pageserver_jwt_token,
3293 0 : 1,
3294 0 : 3,
3295 0 : RECONCILE_TIMEOUT,
3296 0 : &self.cancel,
3297 0 : )
3298 0 : .await
3299 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3300 : {
3301 0 : Ok(_) => {}
3302 : Err(mgmt_api::Error::Cancelled) => {
3303 0 : return Err(ApiError::ShuttingDown);
3304 : }
3305 0 : Err(e) => {
3306 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3307 0 : // at large is unavailable.
3308 0 : tracing::error!("Error deleting via node {node}: {e}");
3309 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3310 : }
3311 : }
3312 :
3313 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3314 : // our in-memory state and database state.
3315 :
3316 : // Ordering: we delete persistent state first: if we then
3317 : // crash, we will drop the in-memory state.
3318 :
3319 : // Drop persistent state.
3320 0 : self.persistence.delete_tenant(tenant_id).await?;
3321 :
3322 : // Drop in-memory state
3323 : {
3324 0 : let mut locked = self.inner.write().unwrap();
3325 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3326 :
3327 : // Dereference Scheduler from shards before dropping them
3328 0 : for (_tenant_shard_id, shard) in
3329 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3330 0 : {
3331 0 : shard.intent.clear(scheduler);
3332 0 : }
3333 :
3334 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3335 0 : tracing::info!(
3336 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3337 0 : locked.tenants.len()
3338 : );
3339 : };
3340 :
3341 : // Success is represented as 404, to imitate the existing pageserver deletion API
3342 0 : Ok(StatusCode::NOT_FOUND)
3343 0 : }
3344 :
3345 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3346 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3347 : /// the tenant's policies (configuration) within the storage controller
3348 0 : pub(crate) async fn tenant_update_policy(
3349 0 : &self,
3350 0 : tenant_id: TenantId,
3351 0 : req: TenantPolicyRequest,
3352 0 : ) -> Result<(), ApiError> {
3353 : // We require an exclusive lock, because we are updating persistent and in-memory state
3354 0 : let _tenant_lock = trace_exclusive_lock(
3355 0 : &self.tenant_op_locks,
3356 0 : tenant_id,
3357 0 : TenantOperations::UpdatePolicy,
3358 0 : )
3359 0 : .await;
3360 :
3361 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3362 :
3363 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3364 :
3365 : let TenantPolicyRequest {
3366 0 : placement,
3367 0 : mut scheduling,
3368 0 : } = req;
3369 :
3370 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3371 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3372 : // this to take effect.
3373 0 : if scheduling.is_none() {
3374 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3375 0 : }
3376 0 : }
3377 :
3378 0 : self.persistence
3379 0 : .update_tenant_shard(
3380 0 : TenantFilter::Tenant(tenant_id),
3381 0 : placement.clone(),
3382 0 : None,
3383 0 : None,
3384 0 : scheduling,
3385 0 : )
3386 0 : .await?;
3387 :
3388 0 : let mut schedule_context = ScheduleContext::default();
3389 0 : let mut locked = self.inner.write().unwrap();
3390 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3391 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3392 0 : if let Some(placement) = &placement {
3393 0 : shard.policy = placement.clone();
3394 0 :
3395 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3396 0 : "Updated placement policy to {placement:?}");
3397 0 : }
3398 :
3399 0 : if let Some(scheduling) = &scheduling {
3400 0 : shard.set_scheduling_policy(*scheduling);
3401 0 :
3402 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3403 0 : "Updated scheduling policy to {scheduling:?}");
3404 0 : }
3405 :
3406 : // In case scheduling is being switched back on, try it now.
3407 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3408 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3409 : }
3410 :
3411 0 : Ok(())
3412 0 : }
3413 :
3414 0 : pub(crate) async fn tenant_timeline_create(
3415 0 : &self,
3416 0 : tenant_id: TenantId,
3417 0 : mut create_req: TimelineCreateRequest,
3418 0 : ) -> Result<TimelineInfo, ApiError> {
3419 0 : tracing::info!(
3420 0 : "Creating timeline {}/{}",
3421 : tenant_id,
3422 : create_req.new_timeline_id,
3423 : );
3424 :
3425 0 : let _tenant_lock = trace_shared_lock(
3426 0 : &self.tenant_op_locks,
3427 0 : tenant_id,
3428 0 : TenantOperations::TimelineCreate,
3429 0 : )
3430 0 : .await;
3431 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3432 :
3433 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3434 0 : if targets.0.is_empty() {
3435 0 : return Err(ApiError::NotFound(
3436 0 : anyhow::anyhow!("Tenant not found").into(),
3437 0 : ));
3438 0 : };
3439 0 :
3440 0 : let (shard_zero_tid, shard_zero_locations) =
3441 0 : targets.0.pop_first().expect("Must have at least one shard");
3442 0 : assert!(shard_zero_tid.is_shard_zero());
3443 :
3444 0 : async fn create_one(
3445 0 : tenant_shard_id: TenantShardId,
3446 0 : locations: ShardMutationLocations,
3447 0 : jwt: Option<String>,
3448 0 : create_req: TimelineCreateRequest,
3449 0 : ) -> Result<TimelineInfo, ApiError> {
3450 0 : let latest = locations.latest.node;
3451 0 :
3452 0 : tracing::info!(
3453 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3454 : tenant_shard_id,
3455 : create_req.new_timeline_id,
3456 : locations.latest.generation
3457 : );
3458 :
3459 0 : let client =
3460 0 : PageserverClient::new(latest.get_id(), latest.base_url(), jwt.as_deref());
3461 :
3462 0 : let timeline_info = client
3463 0 : .timeline_create(tenant_shard_id, &create_req)
3464 0 : .await
3465 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3466 :
3467 : // We propagate timeline creations to all attached locations such that a compute
3468 : // for the new timeline is able to start regardless of the current state of the
3469 : // tenant shard reconciliation.
3470 0 : for location in locations.other {
3471 0 : tracing::info!(
3472 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3473 : tenant_shard_id,
3474 : create_req.new_timeline_id,
3475 : location.node,
3476 : location.generation
3477 : );
3478 :
3479 0 : let client = PageserverClient::new(
3480 0 : location.node.get_id(),
3481 0 : location.node.base_url(),
3482 0 : jwt.as_deref(),
3483 0 : );
3484 :
3485 0 : let res = client
3486 0 : .timeline_create(tenant_shard_id, &create_req)
3487 0 : .await;
3488 :
3489 0 : if let Err(e) = res {
3490 0 : match e {
3491 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3492 0 : // Tenant might have been detached from the stale location,
3493 0 : // so ignore 404s.
3494 0 : },
3495 : _ => {
3496 0 : return Err(passthrough_api_error(&location.node, e));
3497 : }
3498 : }
3499 0 : }
3500 : }
3501 :
3502 0 : Ok(timeline_info)
3503 0 : }
3504 :
3505 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3506 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3507 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3508 0 : let timeline_info = create_one(
3509 0 : shard_zero_tid,
3510 0 : shard_zero_locations,
3511 0 : self.config.pageserver_jwt_token.clone(),
3512 0 : create_req.clone(),
3513 0 : )
3514 0 : .await?;
3515 :
3516 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3517 0 : match &mut create_req.mode {
3518 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3519 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3520 0 : },
3521 0 : _ => {}
3522 : }
3523 :
3524 : // Create timeline on remaining shards with number >0
3525 0 : if !targets.0.is_empty() {
3526 : // If we had multiple shards, issue requests for the remainder now.
3527 0 : let jwt = &self.config.pageserver_jwt_token;
3528 0 : self.tenant_for_shards(
3529 0 : targets
3530 0 : .0
3531 0 : .iter()
3532 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3533 0 : .collect(),
3534 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3535 0 : let create_req = create_req.clone();
3536 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3537 0 : Box::pin(create_one(
3538 0 : tenant_shard_id,
3539 0 : mutation_locations,
3540 0 : jwt.clone(),
3541 0 : create_req,
3542 0 : ))
3543 0 : },
3544 0 : )
3545 0 : .await?;
3546 0 : }
3547 :
3548 0 : Ok(timeline_info)
3549 0 : })
3550 0 : .await?
3551 0 : }
3552 :
3553 0 : pub(crate) async fn tenant_timeline_archival_config(
3554 0 : &self,
3555 0 : tenant_id: TenantId,
3556 0 : timeline_id: TimelineId,
3557 0 : req: TimelineArchivalConfigRequest,
3558 0 : ) -> Result<(), ApiError> {
3559 0 : tracing::info!(
3560 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3561 : req.state
3562 : );
3563 :
3564 0 : let _tenant_lock = trace_shared_lock(
3565 0 : &self.tenant_op_locks,
3566 0 : tenant_id,
3567 0 : TenantOperations::TimelineArchivalConfig,
3568 0 : )
3569 0 : .await;
3570 :
3571 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3572 0 : if targets.0.is_empty() {
3573 0 : return Err(ApiError::NotFound(
3574 0 : anyhow::anyhow!("Tenant not found").into(),
3575 0 : ));
3576 0 : }
3577 0 : async fn config_one(
3578 0 : tenant_shard_id: TenantShardId,
3579 0 : timeline_id: TimelineId,
3580 0 : node: Node,
3581 0 : jwt: Option<String>,
3582 0 : req: TimelineArchivalConfigRequest,
3583 0 : ) -> Result<(), ApiError> {
3584 0 : tracing::info!(
3585 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3586 : );
3587 :
3588 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3589 0 :
3590 0 : client
3591 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3592 0 : .await
3593 0 : .map_err(|e| match e {
3594 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3595 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3596 : }
3597 0 : _ => passthrough_api_error(&node, e),
3598 0 : })
3599 0 : }
3600 :
3601 : // no shard needs to go first/last; the operation should be idempotent
3602 : // TODO: it would be great to ensure that all shards return the same error
3603 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3604 0 : let results = self
3605 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3606 0 : futures::FutureExt::boxed(config_one(
3607 0 : tenant_shard_id,
3608 0 : timeline_id,
3609 0 : node,
3610 0 : self.config.pageserver_jwt_token.clone(),
3611 0 : req.clone(),
3612 0 : ))
3613 0 : })
3614 0 : .await?;
3615 0 : assert!(!results.is_empty(), "must have at least one result");
3616 :
3617 0 : Ok(())
3618 0 : }).await?
3619 0 : }
3620 :
3621 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3622 0 : &self,
3623 0 : tenant_id: TenantId,
3624 0 : timeline_id: TimelineId,
3625 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3626 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3627 :
3628 0 : let _tenant_lock = trace_shared_lock(
3629 0 : &self.tenant_op_locks,
3630 0 : tenant_id,
3631 0 : TenantOperations::TimelineDetachAncestor,
3632 0 : )
3633 0 : .await;
3634 :
3635 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3636 0 : if targets.0.is_empty() {
3637 0 : return Err(ApiError::NotFound(
3638 0 : anyhow::anyhow!("Tenant not found").into(),
3639 0 : ));
3640 0 : }
3641 :
3642 0 : async fn detach_one(
3643 0 : tenant_shard_id: TenantShardId,
3644 0 : timeline_id: TimelineId,
3645 0 : node: Node,
3646 0 : jwt: Option<String>,
3647 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3648 0 : tracing::info!(
3649 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3650 : );
3651 :
3652 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3653 0 :
3654 0 : client
3655 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3656 0 : .await
3657 0 : .map_err(|e| {
3658 : use mgmt_api::Error;
3659 :
3660 0 : match e {
3661 : // no ancestor (ever)
3662 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3663 0 : "{node}: {}",
3664 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3665 0 : )),
3666 : // too many ancestors
3667 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3668 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3669 : }
3670 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3671 0 : // avoid turning these into conflicts to remain compatible with
3672 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3673 0 : // detach
3674 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3675 : }
3676 : // rest can be mapped as usual
3677 0 : other => passthrough_api_error(&node, other),
3678 : }
3679 0 : })
3680 0 : .map(|res| (tenant_shard_id.shard_number, res))
3681 0 : }
3682 :
3683 : // no shard needs to go first/last; the operation should be idempotent
3684 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3685 0 : let mut results = self
3686 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3687 0 : futures::FutureExt::boxed(detach_one(
3688 0 : tenant_shard_id,
3689 0 : timeline_id,
3690 0 : node,
3691 0 : self.config.pageserver_jwt_token.clone(),
3692 0 : ))
3693 0 : })
3694 0 : .await?;
3695 :
3696 0 : let any = results.pop().expect("we must have at least one response");
3697 0 :
3698 0 : let mismatching = results
3699 0 : .iter()
3700 0 : .filter(|(_, res)| res != &any.1)
3701 0 : .collect::<Vec<_>>();
3702 0 : if !mismatching.is_empty() {
3703 : // this can be hit by races which should not happen because operation lock on cplane
3704 0 : let matching = results.len() - mismatching.len();
3705 0 : tracing::error!(
3706 : matching,
3707 : compared_against=?any,
3708 : ?mismatching,
3709 0 : "shards returned different results"
3710 : );
3711 :
3712 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3713 0 : }
3714 0 :
3715 0 : Ok(any.1)
3716 0 : }).await?
3717 0 : }
3718 :
3719 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3720 0 : &self,
3721 0 : tenant_id: TenantId,
3722 0 : timeline_id: TimelineId,
3723 0 : dir: BlockUnblock,
3724 0 : ) -> Result<(), ApiError> {
3725 0 : let _tenant_lock = trace_shared_lock(
3726 0 : &self.tenant_op_locks,
3727 0 : tenant_id,
3728 0 : TenantOperations::TimelineGcBlockUnblock,
3729 0 : )
3730 0 : .await;
3731 :
3732 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3733 0 : if targets.0.is_empty() {
3734 0 : return Err(ApiError::NotFound(
3735 0 : anyhow::anyhow!("Tenant not found").into(),
3736 0 : ));
3737 0 : }
3738 :
3739 0 : async fn do_one(
3740 0 : tenant_shard_id: TenantShardId,
3741 0 : timeline_id: TimelineId,
3742 0 : node: Node,
3743 0 : jwt: Option<String>,
3744 0 : dir: BlockUnblock,
3745 0 : ) -> Result<(), ApiError> {
3746 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3747 0 :
3748 0 : client
3749 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3750 0 : .await
3751 0 : .map_err(|e| passthrough_api_error(&node, e))
3752 0 : }
3753 :
3754 : // no shard needs to go first/last; the operation should be idempotent
3755 0 : let locations = targets
3756 0 : .0
3757 0 : .iter()
3758 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3759 0 : .collect();
3760 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
3761 0 : futures::FutureExt::boxed(do_one(
3762 0 : tenant_shard_id,
3763 0 : timeline_id,
3764 0 : node,
3765 0 : self.config.pageserver_jwt_token.clone(),
3766 0 : dir,
3767 0 : ))
3768 0 : })
3769 0 : .await
3770 0 : })
3771 0 : .await??;
3772 0 : Ok(())
3773 0 : }
3774 :
3775 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
3776 0 : &self,
3777 0 : tenant_shard_id: TenantShardId,
3778 0 : timeline_id: TimelineId,
3779 0 : concurrency: Option<usize>,
3780 0 : ) -> Result<(), ApiError> {
3781 0 : let _tenant_lock = trace_shared_lock(
3782 0 : &self.tenant_op_locks,
3783 0 : tenant_shard_id.tenant_id,
3784 0 : TenantOperations::DownloadHeatmapLayers,
3785 0 : )
3786 0 : .await;
3787 :
3788 0 : let targets = {
3789 0 : let locked = self.inner.read().unwrap();
3790 0 : let mut targets = Vec::new();
3791 :
3792 : // If the request got an unsharded tenant id, then apply
3793 : // the operation to all shards. Otherwise, apply it to a specific shard.
3794 0 : let shards_range = if tenant_shard_id.is_unsharded() {
3795 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
3796 : } else {
3797 0 : tenant_shard_id.range()
3798 : };
3799 :
3800 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
3801 0 : if let Some(node_id) = shard.intent.get_attached() {
3802 0 : let node = locked
3803 0 : .nodes
3804 0 : .get(node_id)
3805 0 : .expect("Pageservers may not be deleted while referenced");
3806 0 :
3807 0 : targets.push((*tenant_shard_id, node.clone()));
3808 0 : }
3809 : }
3810 0 : targets
3811 0 : };
3812 0 :
3813 0 : self.tenant_for_shards_api(
3814 0 : targets,
3815 0 : |tenant_shard_id, client| async move {
3816 0 : client
3817 0 : .timeline_download_heatmap_layers(tenant_shard_id, timeline_id, concurrency)
3818 0 : .await
3819 0 : },
3820 0 : 1,
3821 0 : 1,
3822 0 : SHORT_RECONCILE_TIMEOUT,
3823 0 : &self.cancel,
3824 0 : )
3825 0 : .await;
3826 :
3827 0 : Ok(())
3828 0 : }
3829 :
3830 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3831 : ///
3832 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3833 0 : async fn tenant_for_shards<F, R>(
3834 0 : &self,
3835 0 : locations: Vec<(TenantShardId, Node)>,
3836 0 : mut req_fn: F,
3837 0 : ) -> Result<Vec<R>, ApiError>
3838 0 : where
3839 0 : F: FnMut(
3840 0 : TenantShardId,
3841 0 : Node,
3842 0 : )
3843 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3844 0 : {
3845 0 : let mut futs = FuturesUnordered::new();
3846 0 : let mut results = Vec::with_capacity(locations.len());
3847 :
3848 0 : for (tenant_shard_id, node) in locations {
3849 0 : futs.push(req_fn(tenant_shard_id, node));
3850 0 : }
3851 :
3852 0 : while let Some(r) = futs.next().await {
3853 0 : results.push(r?);
3854 : }
3855 :
3856 0 : Ok(results)
3857 0 : }
3858 :
3859 : /// Concurrently invoke a pageserver API call on many shards at once
3860 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3861 0 : &self,
3862 0 : locations: Vec<(TenantShardId, Node)>,
3863 0 : op: O,
3864 0 : warn_threshold: u32,
3865 0 : max_retries: u32,
3866 0 : timeout: Duration,
3867 0 : cancel: &CancellationToken,
3868 0 : ) -> Vec<mgmt_api::Result<T>>
3869 0 : where
3870 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3871 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3872 0 : {
3873 0 : let mut futs = FuturesUnordered::new();
3874 0 : let mut results = Vec::with_capacity(locations.len());
3875 :
3876 0 : for (tenant_shard_id, node) in locations {
3877 0 : futs.push(async move {
3878 0 : node.with_client_retries(
3879 0 : |client| op(tenant_shard_id, client),
3880 0 : &self.config.pageserver_jwt_token,
3881 0 : warn_threshold,
3882 0 : max_retries,
3883 0 : timeout,
3884 0 : cancel,
3885 0 : )
3886 0 : .await
3887 0 : });
3888 0 : }
3889 :
3890 0 : while let Some(r) = futs.next().await {
3891 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3892 0 : results.push(r);
3893 0 : }
3894 :
3895 0 : results
3896 0 : }
3897 :
3898 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3899 : /// when creating and deleting timelines:
3900 : /// - Makes sure shards are attached somewhere if they weren't already
3901 : /// - Looks up the shards and the nodes where they were most recently attached
3902 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3903 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3904 0 : async fn tenant_remote_mutation<R, O, F>(
3905 0 : &self,
3906 0 : tenant_id: TenantId,
3907 0 : op: O,
3908 0 : ) -> Result<R, ApiError>
3909 0 : where
3910 0 : O: FnOnce(TenantMutationLocations) -> F,
3911 0 : F: std::future::Future<Output = R>,
3912 0 : {
3913 0 : let mutation_locations = {
3914 0 : let mut locations = TenantMutationLocations::default();
3915 :
3916 : // Load the currently attached pageservers for the latest generation of each shard. This can
3917 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3918 : // will still be the latest when we're done: we will check generations again at the end of
3919 : // this function to handle that.
3920 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
3921 :
3922 0 : if generations
3923 0 : .iter()
3924 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3925 : {
3926 0 : let shard_generations = generations
3927 0 : .into_iter()
3928 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
3929 0 : .collect::<HashMap<_, _>>();
3930 0 :
3931 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3932 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)
3933 0 : let locked = self.inner.read().unwrap();
3934 0 : for (shard_id, shard) in
3935 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3936 : {
3937 0 : match shard.policy {
3938 : PlacementPolicy::Attached(_) => {
3939 : // This shard is meant to be attached: the caller is not wrong to try and
3940 : // use this function, but we can't service the request right now.
3941 0 : let Some(generation) = shard_generations.get(shard_id) else {
3942 : // This can only happen if there is a split brain controller modifying the database. This should
3943 : // never happen when testing, and if it happens in production we can only log the issue.
3944 0 : debug_assert!(false);
3945 0 : tracing::error!("Shard {shard_id} not found in generation state! Is another rogue controller running?");
3946 0 : continue;
3947 : };
3948 0 : let (generation, generation_pageserver) = generation;
3949 0 : if let Some(generation) = generation {
3950 0 : if generation_pageserver.is_none() {
3951 : // This is legitimate only in a very narrow window where the shard was only just configured into
3952 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
3953 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
3954 0 : tracing::warn!("Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?");
3955 0 : }
3956 : } else {
3957 : // This should never happen: a shard with no generation is only permitted when it was created in some state
3958 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
3959 0 : debug_assert!(false);
3960 0 : tracing::error!("Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!");
3961 0 : continue;
3962 : }
3963 : }
3964 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3965 0 : return Err(ApiError::Conflict(format!(
3966 0 : "Shard {shard_id} tenant has policy {:?}",
3967 0 : shard.policy
3968 0 : )));
3969 : }
3970 : }
3971 : }
3972 :
3973 0 : return Err(ApiError::ResourceUnavailable(
3974 0 : "One or more shards in tenant is not yet attached".into(),
3975 0 : ));
3976 0 : }
3977 0 :
3978 0 : let locked = self.inner.read().unwrap();
3979 : for ShardGenerationState {
3980 0 : tenant_shard_id,
3981 0 : generation,
3982 0 : generation_pageserver,
3983 0 : } in generations
3984 : {
3985 0 : let node_id = generation_pageserver.expect("We checked for None above");
3986 0 : let node = locked
3987 0 : .nodes
3988 0 : .get(&node_id)
3989 0 : .ok_or(ApiError::Conflict(format!(
3990 0 : "Raced with removal of node {node_id}"
3991 0 : )))?;
3992 0 : let generation = generation.expect("Checked above");
3993 0 :
3994 0 : let tenant = locked.tenants.get(&tenant_shard_id);
3995 :
3996 : // TODO(vlad): Abstract the logic that finds stale attached locations
3997 : // from observed state into a [`Service`] method.
3998 0 : let other_locations = match tenant {
3999 0 : Some(tenant) => {
4000 0 : let mut other = tenant.attached_locations();
4001 0 : let latest_location_index =
4002 0 : other.iter().position(|&l| l == (node.get_id(), generation));
4003 0 : if let Some(idx) = latest_location_index {
4004 0 : other.remove(idx);
4005 0 : }
4006 :
4007 0 : other
4008 : }
4009 0 : None => Vec::default(),
4010 : };
4011 :
4012 0 : let location = ShardMutationLocations {
4013 0 : latest: MutationLocation {
4014 0 : node: node.clone(),
4015 0 : generation,
4016 0 : },
4017 0 : other: other_locations
4018 0 : .into_iter()
4019 0 : .filter_map(|(node_id, generation)| {
4020 0 : let node = locked.nodes.get(&node_id)?;
4021 :
4022 0 : Some(MutationLocation {
4023 0 : node: node.clone(),
4024 0 : generation,
4025 0 : })
4026 0 : })
4027 0 : .collect(),
4028 0 : };
4029 0 : locations.0.insert(tenant_shard_id, location);
4030 0 : }
4031 :
4032 0 : locations
4033 : };
4034 :
4035 0 : let result = op(mutation_locations.clone()).await;
4036 :
4037 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
4038 : // our remote operation executed on the latest generation and is therefore persistent.
4039 : {
4040 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
4041 0 : if latest_generations
4042 0 : .into_iter()
4043 0 : .map(
4044 0 : |ShardGenerationState {
4045 : tenant_shard_id,
4046 : generation,
4047 : generation_pageserver: _,
4048 0 : }| (tenant_shard_id, generation),
4049 0 : )
4050 0 : .collect::<Vec<_>>()
4051 0 : != mutation_locations
4052 0 : .0
4053 0 : .into_iter()
4054 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
4055 0 : .collect::<Vec<_>>()
4056 : {
4057 : // We raced with something that incremented the generation, and therefore cannot be
4058 : // confident that our actions are persistent (they might have hit an old generation).
4059 : //
4060 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
4061 0 : return Err(ApiError::ResourceUnavailable(
4062 0 : "Tenant attachment changed, please retry".into(),
4063 0 : ));
4064 0 : }
4065 0 : }
4066 0 :
4067 0 : Ok(result)
4068 0 : }
4069 :
4070 0 : pub(crate) async fn tenant_timeline_delete(
4071 0 : &self,
4072 0 : tenant_id: TenantId,
4073 0 : timeline_id: TimelineId,
4074 0 : ) -> Result<StatusCode, ApiError> {
4075 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
4076 0 : let _tenant_lock = trace_shared_lock(
4077 0 : &self.tenant_op_locks,
4078 0 : tenant_id,
4079 0 : TenantOperations::TimelineDelete,
4080 0 : )
4081 0 : .await;
4082 :
4083 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
4084 0 : if targets.0.is_empty() {
4085 0 : return Err(ApiError::NotFound(
4086 0 : anyhow::anyhow!("Tenant not found").into(),
4087 0 : ));
4088 0 : }
4089 0 :
4090 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
4091 0 : assert!(shard_zero_tid.is_shard_zero());
4092 :
4093 0 : async fn delete_one(
4094 0 : tenant_shard_id: TenantShardId,
4095 0 : timeline_id: TimelineId,
4096 0 : node: Node,
4097 0 : jwt: Option<String>,
4098 0 : ) -> Result<StatusCode, ApiError> {
4099 0 : tracing::info!(
4100 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4101 : );
4102 :
4103 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
4104 0 : let res = client
4105 0 : .timeline_delete(tenant_shard_id, timeline_id)
4106 0 : .await;
4107 :
4108 0 : match res {
4109 0 : Ok(ok) => Ok(ok),
4110 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
4111 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
4112 0 : Err(e) => {
4113 0 : Err(
4114 0 : ApiError::InternalServerError(anyhow::anyhow!(
4115 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
4116 0 : ))
4117 0 : )
4118 : }
4119 : }
4120 0 : }
4121 :
4122 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4123 0 : let statuses = self
4124 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
4125 0 : Box::pin(delete_one(
4126 0 : tenant_shard_id,
4127 0 : timeline_id,
4128 0 : node,
4129 0 : self.config.pageserver_jwt_token.clone(),
4130 0 : ))
4131 0 : })
4132 0 : .await?;
4133 :
4134 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
4135 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
4136 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
4137 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
4138 0 : return Ok(StatusCode::CONFLICT);
4139 0 : }
4140 0 :
4141 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
4142 0 : return Ok(StatusCode::ACCEPTED);
4143 0 : }
4144 :
4145 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
4146 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
4147 0 : let shard_zero_status = delete_one(
4148 0 : shard_zero_tid,
4149 0 : timeline_id,
4150 0 : shard_zero_locations.latest.node,
4151 0 : self.config.pageserver_jwt_token.clone(),
4152 0 : )
4153 0 : .await?;
4154 0 : Ok(shard_zero_status)
4155 0 : }).await?
4156 0 : }
4157 :
4158 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
4159 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
4160 0 : pub(crate) async fn tenant_shard0_node(
4161 0 : &self,
4162 0 : tenant_id: TenantId,
4163 0 : ) -> Result<(Node, TenantShardId), ApiError> {
4164 0 : // Look up in-memory state and maybe use the node from there.
4165 0 : {
4166 0 : let locked = self.inner.read().unwrap();
4167 0 : let Some((tenant_shard_id, shard)) = locked
4168 0 : .tenants
4169 0 : .range(TenantShardId::tenant_range(tenant_id))
4170 0 : .next()
4171 : else {
4172 0 : return Err(ApiError::NotFound(
4173 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
4174 0 : ));
4175 : };
4176 :
4177 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
4178 0 : tracing::warn!(
4179 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
4180 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
4181 : shard.policy
4182 : );
4183 0 : return Err(ApiError::Conflict(
4184 0 : "Cannot call timeline API on non-attached tenant".to_string(),
4185 0 : ));
4186 : };
4187 :
4188 0 : if shard.reconciler.is_none() {
4189 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
4190 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
4191 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
4192 : // This should never happen
4193 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4194 0 : "Shard refers to nonexistent node"
4195 0 : )));
4196 : };
4197 0 : return Ok((node.clone(), *tenant_shard_id));
4198 0 : }
4199 : };
4200 :
4201 : // Look up the latest attached pageserver location from the database
4202 : // generation state: this will reflect the progress of any ongoing migration.
4203 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4204 : // the case where they call through to the pageserver and get a 404.
4205 0 : let db_result = self.persistence.tenant_generations(tenant_id).await?;
4206 : let Some(ShardGenerationState {
4207 0 : tenant_shard_id,
4208 0 : generation: _,
4209 0 : generation_pageserver: Some(node_id),
4210 0 : }) = db_result.first()
4211 : else {
4212 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4213 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4214 : // or a conflict response (case where tenant was detached in background)
4215 0 : return Err(ApiError::ResourceUnavailable(
4216 0 : "Shard {} not found in database, or is not attached".into(),
4217 0 : ));
4218 : };
4219 0 : let locked = self.inner.read().unwrap();
4220 0 : let Some(node) = locked.nodes.get(node_id) else {
4221 : // This should never happen
4222 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4223 0 : "Shard refers to nonexistent node"
4224 0 : )));
4225 : };
4226 :
4227 0 : Ok((node.clone(), *tenant_shard_id))
4228 0 : }
4229 :
4230 0 : pub(crate) fn tenant_locate(
4231 0 : &self,
4232 0 : tenant_id: TenantId,
4233 0 : ) -> Result<TenantLocateResponse, ApiError> {
4234 0 : let locked = self.inner.read().unwrap();
4235 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4236 :
4237 0 : let mut result = Vec::new();
4238 0 : let mut shard_params: Option<ShardParameters> = None;
4239 :
4240 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4241 : {
4242 0 : let node_id =
4243 0 : shard
4244 0 : .intent
4245 0 : .get_attached()
4246 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4247 0 : "Cannot locate a tenant that is not attached"
4248 0 : )))?;
4249 :
4250 0 : let node = locked
4251 0 : .nodes
4252 0 : .get(&node_id)
4253 0 : .expect("Pageservers may not be deleted while referenced");
4254 0 :
4255 0 : result.push(node.shard_location(*tenant_shard_id));
4256 0 :
4257 0 : match &shard_params {
4258 0 : None => {
4259 0 : shard_params = Some(ShardParameters {
4260 0 : stripe_size: shard.shard.stripe_size,
4261 0 : count: shard.shard.count,
4262 0 : });
4263 0 : }
4264 0 : Some(params) => {
4265 0 : if params.stripe_size != shard.shard.stripe_size {
4266 : // This should never happen. We enforce at runtime because it's simpler than
4267 : // adding an extra per-tenant data structure to store the things that should be the same
4268 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4269 0 : "Inconsistent shard stripe size parameters!"
4270 0 : )));
4271 0 : }
4272 : }
4273 : }
4274 : }
4275 :
4276 0 : if result.is_empty() {
4277 0 : return Err(ApiError::NotFound(
4278 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4279 0 : ));
4280 0 : }
4281 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4282 0 : tracing::info!(
4283 0 : "Located tenant {} with params {:?} on shards {}",
4284 0 : tenant_id,
4285 0 : shard_params,
4286 0 : result
4287 0 : .iter()
4288 0 : .map(|s| format!("{:?}", s))
4289 0 : .collect::<Vec<_>>()
4290 0 : .join(",")
4291 : );
4292 :
4293 0 : Ok(TenantLocateResponse {
4294 0 : shards: result,
4295 0 : shard_params,
4296 0 : })
4297 0 : }
4298 :
4299 : /// Returns None if the input iterator of shards does not include a shard with number=0
4300 0 : fn tenant_describe_impl<'a>(
4301 0 : &self,
4302 0 : shards: impl Iterator<Item = &'a TenantShard>,
4303 0 : ) -> Option<TenantDescribeResponse> {
4304 0 : let mut shard_zero = None;
4305 0 : let mut describe_shards = Vec::new();
4306 :
4307 0 : for shard in shards {
4308 0 : if shard.tenant_shard_id.is_shard_zero() {
4309 0 : shard_zero = Some(shard);
4310 0 : }
4311 :
4312 0 : describe_shards.push(TenantDescribeResponseShard {
4313 0 : tenant_shard_id: shard.tenant_shard_id,
4314 0 : node_attached: *shard.intent.get_attached(),
4315 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4316 0 : last_error: shard
4317 0 : .last_error
4318 0 : .lock()
4319 0 : .unwrap()
4320 0 : .as_ref()
4321 0 : .map(|e| format!("{e}"))
4322 0 : .unwrap_or("".to_string())
4323 0 : .clone(),
4324 0 : is_reconciling: shard.reconciler.is_some(),
4325 0 : is_pending_compute_notification: shard.pending_compute_notification,
4326 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4327 0 : scheduling_policy: *shard.get_scheduling_policy(),
4328 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4329 : })
4330 : }
4331 :
4332 0 : let shard_zero = shard_zero?;
4333 :
4334 0 : Some(TenantDescribeResponse {
4335 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
4336 0 : shards: describe_shards,
4337 0 : stripe_size: shard_zero.shard.stripe_size,
4338 0 : policy: shard_zero.policy.clone(),
4339 0 : config: shard_zero.config.clone(),
4340 0 : })
4341 0 : }
4342 :
4343 0 : pub(crate) fn tenant_describe(
4344 0 : &self,
4345 0 : tenant_id: TenantId,
4346 0 : ) -> Result<TenantDescribeResponse, ApiError> {
4347 0 : let locked = self.inner.read().unwrap();
4348 0 :
4349 0 : self.tenant_describe_impl(
4350 0 : locked
4351 0 : .tenants
4352 0 : .range(TenantShardId::tenant_range(tenant_id))
4353 0 : .map(|(_k, v)| v),
4354 0 : )
4355 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
4356 0 : }
4357 :
4358 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
4359 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
4360 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
4361 : /// in our external API.
4362 0 : pub(crate) fn tenant_list(
4363 0 : &self,
4364 0 : limit: Option<usize>,
4365 0 : start_after: Option<TenantId>,
4366 0 : ) -> Vec<TenantDescribeResponse> {
4367 0 : let locked = self.inner.read().unwrap();
4368 :
4369 : // Apply start_from parameter
4370 0 : let shard_range = match start_after {
4371 0 : None => locked.tenants.range(..),
4372 0 : Some(tenant_id) => locked.tenants.range(
4373 0 : TenantShardId {
4374 0 : tenant_id,
4375 0 : shard_number: ShardNumber(u8::MAX),
4376 0 : shard_count: ShardCount(u8::MAX),
4377 0 : }..,
4378 0 : ),
4379 : };
4380 :
4381 0 : let mut result = Vec::new();
4382 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
4383 0 : result.push(
4384 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
4385 0 : .expect("Groups are always non-empty"),
4386 0 : );
4387 :
4388 : // Enforce `limit` parameter
4389 0 : if let Some(limit) = limit {
4390 0 : if result.len() >= limit {
4391 0 : break;
4392 0 : }
4393 0 : }
4394 : }
4395 :
4396 0 : result
4397 0 : }
4398 :
4399 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
4400 : async fn abort_tenant_shard_split(
4401 : &self,
4402 : op: &TenantShardSplitAbort,
4403 : ) -> Result<(), TenantShardSplitAbortError> {
4404 : // Cleaning up a split:
4405 : // - Parent shards are not destroyed during a split, just detached.
4406 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
4407 : // just the children attached, or both.
4408 : //
4409 : // Therefore our work to do is to:
4410 : // 1. Clean up storage controller's internal state to just refer to parents, no children
4411 : // 2. Call out to pageservers to ensure that children are detached
4412 : // 3. Call out to pageservers to ensure that parents are attached.
4413 : //
4414 : // Crash safety:
4415 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
4416 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
4417 : // and detach them.
4418 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
4419 : // from our database, then we will re-enter this cleanup routine on startup.
4420 :
4421 : let TenantShardSplitAbort {
4422 : tenant_id,
4423 : new_shard_count,
4424 : new_stripe_size,
4425 : ..
4426 : } = op;
4427 :
4428 : // First abort persistent state, if any exists.
4429 : match self
4430 : .persistence
4431 : .abort_shard_split(*tenant_id, *new_shard_count)
4432 : .await?
4433 : {
4434 : AbortShardSplitStatus::Aborted => {
4435 : // Proceed to roll back any child shards created on pageservers
4436 : }
4437 : AbortShardSplitStatus::Complete => {
4438 : // The split completed (we might hit that path if e.g. our database transaction
4439 : // to write the completion landed in the database, but we dropped connection
4440 : // before seeing the result).
4441 : //
4442 : // We must update in-memory state to reflect the successful split.
4443 : self.tenant_shard_split_commit_inmem(
4444 : *tenant_id,
4445 : *new_shard_count,
4446 : *new_stripe_size,
4447 : );
4448 : return Ok(());
4449 : }
4450 : }
4451 :
4452 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
4453 : let detach_locations: Vec<(Node, TenantShardId)> = {
4454 : let mut detach_locations = Vec::new();
4455 : let mut locked = self.inner.write().unwrap();
4456 : let (nodes, tenants, scheduler) = locked.parts_mut();
4457 :
4458 : for (tenant_shard_id, shard) in
4459 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
4460 : {
4461 : if shard.shard.count == op.new_shard_count {
4462 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
4463 : // is infallible, so if we got an error we shouldn't have got that far.
4464 : tracing::warn!(
4465 : "During split abort, child shard {tenant_shard_id} found in-memory"
4466 : );
4467 : continue;
4468 : }
4469 :
4470 : // Add the children of this shard to this list of things to detach
4471 : if let Some(node_id) = shard.intent.get_attached() {
4472 : for child_id in tenant_shard_id.split(*new_shard_count) {
4473 : detach_locations.push((
4474 : nodes
4475 : .get(node_id)
4476 : .expect("Intent references nonexistent node")
4477 : .clone(),
4478 : child_id,
4479 : ));
4480 : }
4481 : } else {
4482 : tracing::warn!(
4483 : "During split abort, shard {tenant_shard_id} has no attached location"
4484 : );
4485 : }
4486 :
4487 : tracing::info!("Restoring parent shard {tenant_shard_id}");
4488 :
4489 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
4490 : // if the original attachment location is offline.
4491 : if let Some(node_id) = shard.intent.get_attached() {
4492 : if !nodes.get(node_id).unwrap().is_available() {
4493 : tracing::info!("Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}");
4494 : shard.intent.demote_attached(scheduler, *node_id);
4495 : }
4496 : }
4497 : for node_id in shard.intent.get_secondary().clone() {
4498 : if !nodes.get(&node_id).unwrap().is_available() {
4499 : tracing::info!("Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}");
4500 : shard.intent.remove_secondary(scheduler, node_id);
4501 : }
4502 : }
4503 :
4504 : shard.splitting = SplitState::Idle;
4505 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
4506 : // If this shard can't be scheduled now (perhaps due to offline nodes or
4507 : // capacity issues), that must not prevent us rolling back a split. In this
4508 : // case it should be eventually scheduled in the background.
4509 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
4510 : }
4511 :
4512 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
4513 : }
4514 :
4515 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
4516 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
4517 :
4518 : detach_locations
4519 : };
4520 :
4521 : for (node, child_id) in detach_locations {
4522 : if !node.is_available() {
4523 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
4524 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
4525 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
4526 : // them from the node.
4527 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
4528 : continue;
4529 : }
4530 :
4531 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
4532 : // a 503 and retry, up to our limit.
4533 : tracing::info!("Detaching {child_id} on {node}...");
4534 : match node
4535 : .with_client_retries(
4536 0 : |client| async move {
4537 0 : let config = LocationConfig {
4538 0 : mode: LocationConfigMode::Detached,
4539 0 : generation: None,
4540 0 : secondary_conf: None,
4541 0 : shard_number: child_id.shard_number.0,
4542 0 : shard_count: child_id.shard_count.literal(),
4543 0 : // Stripe size and tenant config don't matter when detaching
4544 0 : shard_stripe_size: 0,
4545 0 : tenant_conf: TenantConfig::default(),
4546 0 : };
4547 0 :
4548 0 : client.location_config(child_id, config, None, false).await
4549 0 : },
4550 : &self.config.pageserver_jwt_token,
4551 : 1,
4552 : 10,
4553 : Duration::from_secs(5),
4554 : &self.cancel,
4555 : )
4556 : .await
4557 : {
4558 : Some(Ok(_)) => {}
4559 : Some(Err(e)) => {
4560 : // We failed to communicate with the remote node. This is problematic: we may be
4561 : // leaving it with a rogue child shard.
4562 : tracing::warn!(
4563 : "Failed to detach child {child_id} from node {node} during abort"
4564 : );
4565 : return Err(e.into());
4566 : }
4567 : None => {
4568 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
4569 : // clean up on restart. The node going offline requires a retry.
4570 : return Err(TenantShardSplitAbortError::Unavailable);
4571 : }
4572 : };
4573 : }
4574 :
4575 : tracing::info!("Successfully aborted split");
4576 : Ok(())
4577 : }
4578 :
4579 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
4580 : /// of the tenant map to reflect the child shards that exist after the split.
4581 0 : fn tenant_shard_split_commit_inmem(
4582 0 : &self,
4583 0 : tenant_id: TenantId,
4584 0 : new_shard_count: ShardCount,
4585 0 : new_stripe_size: Option<ShardStripeSize>,
4586 0 : ) -> (
4587 0 : TenantShardSplitResponse,
4588 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
4589 0 : Vec<ReconcilerWaiter>,
4590 0 : ) {
4591 0 : let mut response = TenantShardSplitResponse {
4592 0 : new_shards: Vec::new(),
4593 0 : };
4594 0 : let mut child_locations = Vec::new();
4595 0 : let mut waiters = Vec::new();
4596 0 :
4597 0 : {
4598 0 : let mut locked = self.inner.write().unwrap();
4599 0 :
4600 0 : let parent_ids = locked
4601 0 : .tenants
4602 0 : .range(TenantShardId::tenant_range(tenant_id))
4603 0 : .map(|(shard_id, _)| *shard_id)
4604 0 : .collect::<Vec<_>>();
4605 0 :
4606 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4607 0 : for parent_id in parent_ids {
4608 0 : let child_ids = parent_id.split(new_shard_count);
4609 :
4610 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
4611 0 : let mut old_state = tenants
4612 0 : .remove(&parent_id)
4613 0 : .expect("It was present, we just split it");
4614 0 :
4615 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
4616 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
4617 0 : // nothing else can clear this.
4618 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
4619 :
4620 0 : let old_attached = old_state.intent.get_attached().unwrap();
4621 0 : old_state.intent.clear(scheduler);
4622 0 : let generation = old_state.generation.expect("Shard must have been attached");
4623 0 : (
4624 0 : old_attached,
4625 0 : generation,
4626 0 : old_state.policy.clone(),
4627 0 : old_state.shard,
4628 0 : old_state.config.clone(),
4629 0 : old_state.preferred_az().cloned(),
4630 0 : )
4631 0 : };
4632 0 :
4633 0 : let mut schedule_context = ScheduleContext::default();
4634 0 : for child in child_ids {
4635 0 : let mut child_shard = parent_ident;
4636 0 : child_shard.number = child.shard_number;
4637 0 : child_shard.count = child.shard_count;
4638 0 : if let Some(stripe_size) = new_stripe_size {
4639 0 : child_shard.stripe_size = stripe_size;
4640 0 : }
4641 :
4642 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
4643 0 : child_observed.insert(
4644 0 : pageserver,
4645 0 : ObservedStateLocation {
4646 0 : conf: Some(attached_location_conf(
4647 0 : generation,
4648 0 : &child_shard,
4649 0 : &config,
4650 0 : &policy,
4651 0 : )),
4652 0 : },
4653 0 : );
4654 0 :
4655 0 : let mut child_state =
4656 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
4657 0 : child_state.intent =
4658 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
4659 0 : child_state.observed = ObservedState {
4660 0 : locations: child_observed,
4661 0 : };
4662 0 : child_state.generation = Some(generation);
4663 0 : child_state.config = config.clone();
4664 0 :
4665 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
4666 0 : // as at this point in the split process we have succeeded and this part is infallible:
4667 0 : // we will never need to do any special recovery from this state.
4668 0 :
4669 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
4670 :
4671 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
4672 : // This is not fatal, because we've implicitly already got an attached
4673 : // location for the child shard. Failure here just means we couldn't
4674 : // find a secondary (e.g. because cluster is overloaded).
4675 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
4676 0 : }
4677 : // In the background, attach secondary locations for the new shards
4678 0 : if let Some(waiter) = self.maybe_reconcile_shard(
4679 0 : &mut child_state,
4680 0 : nodes,
4681 0 : ReconcilerPriority::High,
4682 0 : ) {
4683 0 : waiters.push(waiter);
4684 0 : }
4685 :
4686 0 : tenants.insert(child, child_state);
4687 0 : response.new_shards.push(child);
4688 : }
4689 : }
4690 0 : (response, child_locations, waiters)
4691 0 : }
4692 0 : }
4693 :
4694 0 : async fn tenant_shard_split_start_secondaries(
4695 0 : &self,
4696 0 : tenant_id: TenantId,
4697 0 : waiters: Vec<ReconcilerWaiter>,
4698 0 : ) {
4699 : // Wait for initial reconcile of child shards, this creates the secondary locations
4700 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
4701 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
4702 : // their secondaries couldn't be attached.
4703 0 : tracing::warn!("Failed to reconcile after split: {e}");
4704 0 : return;
4705 0 : }
4706 :
4707 : // Take the state lock to discover the attached & secondary intents for all shards
4708 0 : let (attached, secondary) = {
4709 0 : let locked = self.inner.read().unwrap();
4710 0 : let mut attached = Vec::new();
4711 0 : let mut secondary = Vec::new();
4712 :
4713 0 : for (tenant_shard_id, shard) in
4714 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4715 : {
4716 0 : let Some(node_id) = shard.intent.get_attached() else {
4717 : // Unexpected. Race with a PlacementPolicy change?
4718 0 : tracing::warn!(
4719 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
4720 : );
4721 0 : continue;
4722 : };
4723 :
4724 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
4725 : // No secondary location. Nothing for us to do.
4726 0 : continue;
4727 : };
4728 :
4729 0 : let attached_node = locked
4730 0 : .nodes
4731 0 : .get(node_id)
4732 0 : .expect("Pageservers may not be deleted while referenced");
4733 0 :
4734 0 : let secondary_node = locked
4735 0 : .nodes
4736 0 : .get(secondary_node_id)
4737 0 : .expect("Pageservers may not be deleted while referenced");
4738 0 :
4739 0 : attached.push((*tenant_shard_id, attached_node.clone()));
4740 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
4741 : }
4742 0 : (attached, secondary)
4743 0 : };
4744 0 :
4745 0 : if secondary.is_empty() {
4746 : // No secondary locations; nothing for us to do
4747 0 : return;
4748 0 : }
4749 :
4750 0 : for result in self
4751 0 : .tenant_for_shards_api(
4752 0 : attached,
4753 0 : |tenant_shard_id, client| async move {
4754 0 : client.tenant_heatmap_upload(tenant_shard_id).await
4755 0 : },
4756 0 : 1,
4757 0 : 1,
4758 0 : SHORT_RECONCILE_TIMEOUT,
4759 0 : &self.cancel,
4760 0 : )
4761 0 : .await
4762 : {
4763 0 : if let Err(e) = result {
4764 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
4765 0 : return;
4766 0 : }
4767 : }
4768 :
4769 0 : for result in self
4770 0 : .tenant_for_shards_api(
4771 0 : secondary,
4772 0 : |tenant_shard_id, client| async move {
4773 0 : client
4774 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
4775 0 : .await
4776 0 : },
4777 0 : 1,
4778 0 : 1,
4779 0 : SHORT_RECONCILE_TIMEOUT,
4780 0 : &self.cancel,
4781 0 : )
4782 0 : .await
4783 : {
4784 0 : if let Err(e) = result {
4785 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
4786 0 : return;
4787 0 : }
4788 : }
4789 0 : }
4790 :
4791 0 : pub(crate) async fn tenant_shard_split(
4792 0 : &self,
4793 0 : tenant_id: TenantId,
4794 0 : split_req: TenantShardSplitRequest,
4795 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
4796 : // TODO: return 503 if we get stuck waiting for this lock
4797 : // (issue https://github.com/neondatabase/neon/issues/7108)
4798 0 : let _tenant_lock = trace_exclusive_lock(
4799 0 : &self.tenant_op_locks,
4800 0 : tenant_id,
4801 0 : TenantOperations::ShardSplit,
4802 0 : )
4803 0 : .await;
4804 :
4805 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
4806 0 : let new_stripe_size = split_req.new_stripe_size;
4807 :
4808 : // Validate the request and construct parameters. This phase is fallible, but does not require
4809 : // rollback on errors, as it does no I/O and mutates no state.
4810 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
4811 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
4812 0 : ShardSplitAction::Split(params) => params,
4813 : };
4814 :
4815 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
4816 : // we must roll back.
4817 0 : let r = self
4818 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
4819 0 : .await;
4820 :
4821 0 : let (response, waiters) = match r {
4822 0 : Ok(r) => r,
4823 0 : Err(e) => {
4824 0 : // Split might be part-done, we must do work to abort it.
4825 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
4826 0 : self.abort_tx
4827 0 : .send(TenantShardSplitAbort {
4828 0 : tenant_id,
4829 0 : new_shard_count,
4830 0 : new_stripe_size,
4831 0 : _tenant_lock,
4832 0 : })
4833 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
4834 0 : .ok();
4835 0 : return Err(e);
4836 : }
4837 : };
4838 :
4839 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
4840 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
4841 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
4842 : // in [`Self::optimize_all`]
4843 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
4844 0 : .await;
4845 0 : Ok(response)
4846 0 : }
4847 :
4848 0 : fn prepare_tenant_shard_split(
4849 0 : &self,
4850 0 : tenant_id: TenantId,
4851 0 : split_req: TenantShardSplitRequest,
4852 0 : ) -> Result<ShardSplitAction, ApiError> {
4853 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
4854 0 : anyhow::anyhow!("failpoint")
4855 0 : )));
4856 :
4857 0 : let mut policy = None;
4858 0 : let mut config = None;
4859 0 : let mut shard_ident = None;
4860 0 : let mut preferred_az_id = None;
4861 : // Validate input, and calculate which shards we will create
4862 0 : let (old_shard_count, targets) =
4863 : {
4864 0 : let locked = self.inner.read().unwrap();
4865 0 :
4866 0 : let pageservers = locked.nodes.clone();
4867 0 :
4868 0 : let mut targets = Vec::new();
4869 0 :
4870 0 : // In case this is a retry, count how many already-split shards we found
4871 0 : let mut children_found = Vec::new();
4872 0 : let mut old_shard_count = None;
4873 :
4874 0 : for (tenant_shard_id, shard) in
4875 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4876 : {
4877 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
4878 : Ordering::Equal => {
4879 : // Already split this
4880 0 : children_found.push(*tenant_shard_id);
4881 0 : continue;
4882 : }
4883 : Ordering::Greater => {
4884 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4885 0 : "Requested count {} but already have shards at count {}",
4886 0 : split_req.new_shard_count,
4887 0 : shard.shard.count.count()
4888 0 : )));
4889 : }
4890 0 : Ordering::Less => {
4891 0 : // Fall through: this shard has lower count than requested,
4892 0 : // is a candidate for splitting.
4893 0 : }
4894 0 : }
4895 0 :
4896 0 : match old_shard_count {
4897 0 : None => old_shard_count = Some(shard.shard.count),
4898 0 : Some(old_shard_count) => {
4899 0 : if old_shard_count != shard.shard.count {
4900 : // We may hit this case if a caller asked for two splits to
4901 : // different sizes, before the first one is complete.
4902 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
4903 : // of shard_count=1 and shard_count=2 shards in the map.
4904 0 : return Err(ApiError::Conflict(
4905 0 : "Cannot split, currently mid-split".to_string(),
4906 0 : ));
4907 0 : }
4908 : }
4909 : }
4910 0 : if policy.is_none() {
4911 0 : policy = Some(shard.policy.clone());
4912 0 : }
4913 0 : if shard_ident.is_none() {
4914 0 : shard_ident = Some(shard.shard);
4915 0 : }
4916 0 : if config.is_none() {
4917 0 : config = Some(shard.config.clone());
4918 0 : }
4919 0 : if preferred_az_id.is_none() {
4920 0 : preferred_az_id = shard.preferred_az().cloned();
4921 0 : }
4922 :
4923 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4924 0 : tracing::info!(
4925 0 : "Tenant shard {} already has shard count {}",
4926 : tenant_shard_id,
4927 : split_req.new_shard_count
4928 : );
4929 0 : continue;
4930 0 : }
4931 :
4932 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4933 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4934 0 : ))?;
4935 :
4936 0 : let node = pageservers
4937 0 : .get(&node_id)
4938 0 : .expect("Pageservers may not be deleted while referenced");
4939 0 :
4940 0 : targets.push(ShardSplitTarget {
4941 0 : parent_id: *tenant_shard_id,
4942 0 : node: node.clone(),
4943 0 : child_ids: tenant_shard_id
4944 0 : .split(ShardCount::new(split_req.new_shard_count)),
4945 0 : });
4946 : }
4947 :
4948 0 : if targets.is_empty() {
4949 0 : if children_found.len() == split_req.new_shard_count as usize {
4950 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4951 0 : new_shards: children_found,
4952 0 : }));
4953 : } else {
4954 : // No shards found to split, and no existing children found: the
4955 : // tenant doesn't exist at all.
4956 0 : return Err(ApiError::NotFound(
4957 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4958 0 : ));
4959 : }
4960 0 : }
4961 0 :
4962 0 : (old_shard_count, targets)
4963 0 : };
4964 0 :
4965 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
4966 0 : let old_shard_count = old_shard_count.unwrap();
4967 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
4968 : // This ShardIdentity will be used as the template for all children, so this implicitly
4969 : // applies the new stripe size to the children.
4970 0 : let mut shard_ident = shard_ident.unwrap();
4971 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
4972 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)));
4973 0 : }
4974 0 :
4975 0 : shard_ident.stripe_size = new_stripe_size;
4976 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
4977 0 : shard_ident
4978 : } else {
4979 0 : shard_ident.unwrap()
4980 : };
4981 0 : let policy = policy.unwrap();
4982 0 : let config = config.unwrap();
4983 0 :
4984 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
4985 0 : old_shard_count,
4986 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
4987 0 : new_stripe_size: split_req.new_stripe_size,
4988 0 : targets,
4989 0 : policy,
4990 0 : config,
4991 0 : shard_ident,
4992 0 : preferred_az_id,
4993 0 : })))
4994 0 : }
4995 :
4996 0 : async fn do_tenant_shard_split(
4997 0 : &self,
4998 0 : tenant_id: TenantId,
4999 0 : params: Box<ShardSplitParams>,
5000 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
5001 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
5002 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
5003 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
5004 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
5005 0 : // (https://github.com/neondatabase/neon/issues/6676)
5006 0 :
5007 0 : let ShardSplitParams {
5008 0 : old_shard_count,
5009 0 : new_shard_count,
5010 0 : new_stripe_size,
5011 0 : mut targets,
5012 0 : policy,
5013 0 : config,
5014 0 : shard_ident,
5015 0 : preferred_az_id,
5016 0 : } = *params;
5017 :
5018 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
5019 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
5020 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
5021 : // at the time of split.
5022 0 : let waiters = {
5023 0 : let mut locked = self.inner.write().unwrap();
5024 0 : let mut waiters = Vec::new();
5025 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5026 0 : for target in &mut targets {
5027 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
5028 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5029 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5030 0 : "Shard {} not found",
5031 0 : target.parent_id
5032 0 : )));
5033 : };
5034 :
5035 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
5036 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5037 0 : return Err(ApiError::Conflict(format!(
5038 0 : "Shard {} unexpectedly rescheduled during split",
5039 0 : target.parent_id
5040 0 : )));
5041 0 : }
5042 0 :
5043 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
5044 0 : shard.intent.clear_secondary(scheduler);
5045 :
5046 : // Run Reconciler to execute detach fo secondary locations.
5047 0 : if let Some(waiter) =
5048 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5049 0 : {
5050 0 : waiters.push(waiter);
5051 0 : }
5052 : }
5053 0 : waiters
5054 0 : };
5055 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
5056 :
5057 : // Before creating any new child shards in memory or on the pageservers, persist them: this
5058 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
5059 : // acts as the protection against two concurrent attempts to split: one of them will get a database
5060 : // error trying to insert the child shards.
5061 0 : let mut child_tsps = Vec::new();
5062 0 : for target in &targets {
5063 0 : let mut this_child_tsps = Vec::new();
5064 0 : for child in &target.child_ids {
5065 0 : let mut child_shard = shard_ident;
5066 0 : child_shard.number = child.shard_number;
5067 0 : child_shard.count = child.shard_count;
5068 0 :
5069 0 : tracing::info!(
5070 0 : "Create child shard persistence with stripe size {}",
5071 : shard_ident.stripe_size.0
5072 : );
5073 :
5074 0 : this_child_tsps.push(TenantShardPersistence {
5075 0 : tenant_id: child.tenant_id.to_string(),
5076 0 : shard_number: child.shard_number.0 as i32,
5077 0 : shard_count: child.shard_count.literal() as i32,
5078 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
5079 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
5080 0 : // populate the correct generation as part of its transaction, to protect us
5081 0 : // against racing with changes in the state of the parent.
5082 0 : generation: None,
5083 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
5084 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
5085 0 : config: serde_json::to_string(&config).unwrap(),
5086 0 : splitting: SplitState::Splitting,
5087 0 :
5088 0 : // Scheduling policies and preferred AZ do not carry through to children
5089 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
5090 0 : .unwrap(),
5091 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
5092 0 : });
5093 0 : }
5094 :
5095 0 : child_tsps.push((target.parent_id, this_child_tsps));
5096 : }
5097 :
5098 0 : if let Err(e) = self
5099 0 : .persistence
5100 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
5101 0 : .await
5102 : {
5103 0 : match e {
5104 : DatabaseError::Query(diesel::result::Error::DatabaseError(
5105 : DatabaseErrorKind::UniqueViolation,
5106 : _,
5107 : )) => {
5108 : // Inserting a child shard violated a unique constraint: we raced with another call to
5109 : // this function
5110 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
5111 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
5112 : }
5113 0 : _ => return Err(ApiError::InternalServerError(e.into())),
5114 : }
5115 0 : }
5116 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
5117 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5118 0 : ));
5119 :
5120 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
5121 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
5122 : // is not set in memory, then it was not persisted.
5123 : {
5124 0 : let mut locked = self.inner.write().unwrap();
5125 0 : for target in &targets {
5126 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
5127 0 : parent_shard.splitting = SplitState::Splitting;
5128 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
5129 0 : // split operation.
5130 0 : parent_shard
5131 0 : .observed
5132 0 : .locations
5133 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
5134 0 : }
5135 : }
5136 : }
5137 :
5138 : // TODO: issue split calls concurrently (this only matters once we're splitting
5139 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
5140 :
5141 0 : for target in &targets {
5142 : let ShardSplitTarget {
5143 0 : parent_id,
5144 0 : node,
5145 0 : child_ids,
5146 0 : } = target;
5147 0 : let client = PageserverClient::new(
5148 0 : node.get_id(),
5149 0 : node.base_url(),
5150 0 : self.config.pageserver_jwt_token.as_deref(),
5151 0 : );
5152 0 : let response = client
5153 0 : .tenant_shard_split(
5154 0 : *parent_id,
5155 0 : TenantShardSplitRequest {
5156 0 : new_shard_count: new_shard_count.literal(),
5157 0 : new_stripe_size,
5158 0 : },
5159 0 : )
5160 0 : .await
5161 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
5162 :
5163 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
5164 0 : "failpoint".to_string()
5165 0 : )));
5166 :
5167 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
5168 :
5169 0 : tracing::info!(
5170 0 : "Split {} into {}",
5171 0 : parent_id,
5172 0 : response
5173 0 : .new_shards
5174 0 : .iter()
5175 0 : .map(|s| format!("{:?}", s))
5176 0 : .collect::<Vec<_>>()
5177 0 : .join(",")
5178 : );
5179 :
5180 0 : if &response.new_shards != child_ids {
5181 : // This should never happen: the pageserver should agree with us on how shard splits work.
5182 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5183 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
5184 0 : parent_id,
5185 0 : response.new_shards,
5186 0 : child_ids
5187 0 : )));
5188 0 : }
5189 : }
5190 :
5191 : // TODO: if the pageserver restarted concurrently with our split API call,
5192 : // the actual generation of the child shard might differ from the generation
5193 : // we expect it to have. In order for our in-database generation to end up
5194 : // correct, we should carry the child generation back in the response and apply it here
5195 : // in complete_shard_split (and apply the correct generation in memory)
5196 : // (or, we can carry generation in the request and reject the request if
5197 : // it doesn't match, but that requires more retry logic on this side)
5198 :
5199 0 : self.persistence
5200 0 : .complete_shard_split(tenant_id, old_shard_count)
5201 0 : .await?;
5202 :
5203 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5204 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5205 0 : ));
5206 :
5207 : // Replace all the shards we just split with their children: this phase is infallible.
5208 0 : let (response, child_locations, waiters) =
5209 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5210 0 :
5211 0 : // Send compute notifications for all the new shards
5212 0 : let mut failed_notifications = Vec::new();
5213 0 : for (child_id, child_ps, stripe_size) in child_locations {
5214 0 : if let Err(e) = self
5215 0 : .compute_hook
5216 0 : .notify(
5217 0 : compute_hook::ShardUpdate {
5218 0 : tenant_shard_id: child_id,
5219 0 : node_id: child_ps,
5220 0 : stripe_size,
5221 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5222 0 : },
5223 0 : &self.cancel,
5224 0 : )
5225 0 : .await
5226 : {
5227 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5228 : child_id, child_ps);
5229 0 : failed_notifications.push(child_id);
5230 0 : }
5231 : }
5232 :
5233 : // If we failed any compute notifications, make a note to retry later.
5234 0 : if !failed_notifications.is_empty() {
5235 0 : let mut locked = self.inner.write().unwrap();
5236 0 : for failed in failed_notifications {
5237 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5238 0 : shard.pending_compute_notification = true;
5239 0 : }
5240 : }
5241 0 : }
5242 :
5243 0 : Ok((response, waiters))
5244 0 : }
5245 :
5246 0 : pub(crate) async fn tenant_shard_migrate(
5247 0 : &self,
5248 0 : tenant_shard_id: TenantShardId,
5249 0 : migrate_req: TenantShardMigrateRequest,
5250 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5251 0 : let waiter = {
5252 0 : let mut locked = self.inner.write().unwrap();
5253 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5254 :
5255 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5256 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5257 0 : "Node {} not found",
5258 0 : migrate_req.node_id
5259 0 : )));
5260 : };
5261 :
5262 0 : if !node.is_available() {
5263 : // Warn but proceed: the caller may intend to manually adjust the placement of
5264 : // a shard even if the node is down, e.g. if intervening during an incident.
5265 0 : tracing::warn!("Migrating to unavailable node {node}");
5266 0 : }
5267 :
5268 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5269 0 : return Err(ApiError::NotFound(
5270 0 : anyhow::anyhow!("Tenant shard not found").into(),
5271 0 : ));
5272 : };
5273 :
5274 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5275 : // No-op case: we will still proceed to wait for reconciliation in case it is
5276 : // incomplete from an earlier update to the intent.
5277 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
5278 : } else {
5279 0 : let old_attached = *shard.intent.get_attached();
5280 0 :
5281 0 : match shard.policy {
5282 0 : PlacementPolicy::Attached(n) => {
5283 0 : // If our new attached node was a secondary, it no longer should be.
5284 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
5285 0 :
5286 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
5287 :
5288 : // If we were already attached to something, demote that to a secondary
5289 0 : if let Some(old_attached) = old_attached {
5290 0 : if n > 0 {
5291 : // Remove other secondaries to make room for the location we'll demote
5292 0 : while shard.intent.get_secondary().len() >= n {
5293 0 : shard.intent.pop_secondary(scheduler);
5294 0 : }
5295 :
5296 0 : shard.intent.push_secondary(scheduler, old_attached);
5297 0 : }
5298 0 : }
5299 : }
5300 0 : PlacementPolicy::Secondary => {
5301 0 : shard.intent.clear(scheduler);
5302 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5303 0 : }
5304 : PlacementPolicy::Detached => {
5305 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5306 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
5307 0 : )))
5308 : }
5309 : }
5310 :
5311 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
5312 0 : shard.sequence = shard.sequence.next();
5313 : }
5314 :
5315 0 : let reconciler_config = match migrate_req.migration_config {
5316 0 : Some(cfg) => (&cfg).into(),
5317 0 : None => ReconcilerConfig::new(ReconcilerPriority::High),
5318 : };
5319 :
5320 0 : self.maybe_configured_reconcile_shard(shard, nodes, reconciler_config)
5321 : };
5322 :
5323 0 : if let Some(waiter) = waiter {
5324 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5325 : } else {
5326 0 : tracing::info!("Migration is a no-op");
5327 : }
5328 :
5329 0 : Ok(TenantShardMigrateResponse {})
5330 0 : }
5331 :
5332 0 : pub(crate) async fn tenant_shard_migrate_secondary(
5333 0 : &self,
5334 0 : tenant_shard_id: TenantShardId,
5335 0 : migrate_req: TenantShardMigrateRequest,
5336 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5337 0 : let waiter = {
5338 0 : let mut locked = self.inner.write().unwrap();
5339 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5340 :
5341 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5342 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5343 0 : "Node {} not found",
5344 0 : migrate_req.node_id
5345 0 : )));
5346 : };
5347 :
5348 0 : if !node.is_available() {
5349 : // Warn but proceed: the caller may intend to manually adjust the placement of
5350 : // a shard even if the node is down, e.g. if intervening during an incident.
5351 0 : tracing::warn!("Migrating to unavailable node {node}");
5352 0 : }
5353 :
5354 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5355 0 : return Err(ApiError::NotFound(
5356 0 : anyhow::anyhow!("Tenant shard not found").into(),
5357 0 : ));
5358 : };
5359 :
5360 0 : if shard.intent.get_secondary().len() == 1
5361 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
5362 : {
5363 0 : tracing::info!(
5364 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
5365 : shard.intent
5366 : );
5367 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5368 0 : tracing::info!("Migrating secondary to {node}: already attached where we were asked to create a secondary");
5369 : } else {
5370 0 : let old_secondaries = shard.intent.get_secondary().clone();
5371 0 : for secondary in old_secondaries {
5372 0 : shard.intent.remove_secondary(scheduler, secondary);
5373 0 : }
5374 :
5375 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5376 0 : shard.sequence = shard.sequence.next();
5377 0 : tracing::info!(
5378 0 : "Migrating secondary to {node}: new intent {:?}",
5379 : shard.intent
5380 : );
5381 : }
5382 :
5383 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5384 : };
5385 :
5386 0 : if let Some(waiter) = waiter {
5387 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5388 : } else {
5389 0 : tracing::info!("Migration is a no-op");
5390 : }
5391 :
5392 0 : Ok(TenantShardMigrateResponse {})
5393 0 : }
5394 :
5395 : /// 'cancel' in this context means cancel any ongoing reconcile
5396 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
5397 0 : &self,
5398 0 : tenant_shard_id: TenantShardId,
5399 0 : ) -> Result<(), ApiError> {
5400 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
5401 0 : let waiter = {
5402 0 : let locked = self.inner.write().unwrap();
5403 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5404 0 : return Err(ApiError::NotFound(
5405 0 : anyhow::anyhow!("Tenant shard not found").into(),
5406 0 : ));
5407 : };
5408 :
5409 0 : let waiter = shard.get_waiter();
5410 0 : match waiter {
5411 : None => {
5412 0 : tracing::info!("Shard does not have an ongoing Reconciler");
5413 0 : return Ok(());
5414 : }
5415 0 : Some(waiter) => {
5416 0 : tracing::info!("Cancelling Reconciler");
5417 0 : shard.cancel_reconciler();
5418 0 : waiter
5419 0 : }
5420 0 : }
5421 0 : };
5422 0 :
5423 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
5424 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
5425 0 : // the Reconciler is misbehaving and not respecting the cancellation token
5426 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
5427 0 : .await?;
5428 :
5429 0 : Ok(())
5430 0 : }
5431 :
5432 : /// This is for debug/support only: we simply drop all state for a tenant, without
5433 : /// detaching or deleting it on pageservers.
5434 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
5435 0 : self.persistence.delete_tenant(tenant_id).await?;
5436 :
5437 0 : let mut locked = self.inner.write().unwrap();
5438 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
5439 0 : let mut shards = Vec::new();
5440 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
5441 0 : shards.push(*tenant_shard_id);
5442 0 : }
5443 :
5444 0 : for shard_id in shards {
5445 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
5446 0 : shard.intent.clear(scheduler);
5447 0 : }
5448 : }
5449 :
5450 0 : Ok(())
5451 0 : }
5452 :
5453 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
5454 : /// tenant with a very high generation number so that it will see the existing data.
5455 0 : pub(crate) async fn tenant_import(
5456 0 : &self,
5457 0 : tenant_id: TenantId,
5458 0 : ) -> Result<TenantCreateResponse, ApiError> {
5459 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
5460 0 : let maybe_node = {
5461 0 : self.inner
5462 0 : .read()
5463 0 : .unwrap()
5464 0 : .nodes
5465 0 : .values()
5466 0 : .find(|n| n.is_available())
5467 0 : .cloned()
5468 : };
5469 0 : let Some(node) = maybe_node else {
5470 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
5471 : };
5472 :
5473 0 : let client = PageserverClient::new(
5474 0 : node.get_id(),
5475 0 : node.base_url(),
5476 0 : self.config.pageserver_jwt_token.as_deref(),
5477 0 : );
5478 :
5479 0 : let scan_result = client
5480 0 : .tenant_scan_remote_storage(tenant_id)
5481 0 : .await
5482 0 : .map_err(|e| passthrough_api_error(&node, e))?;
5483 :
5484 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
5485 0 : let Some(shard_count) = scan_result
5486 0 : .shards
5487 0 : .iter()
5488 0 : .map(|s| s.tenant_shard_id.shard_count)
5489 0 : .max()
5490 : else {
5491 0 : return Err(ApiError::NotFound(
5492 0 : anyhow::anyhow!("No shards found").into(),
5493 0 : ));
5494 : };
5495 :
5496 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
5497 : // to
5498 0 : let generation = scan_result
5499 0 : .shards
5500 0 : .iter()
5501 0 : .map(|s| s.generation)
5502 0 : .max()
5503 0 : .expect("We already validated >0 shards");
5504 0 :
5505 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
5506 0 : // only work if they were using the default stripe size.
5507 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
5508 :
5509 0 : let (response, waiters) = self
5510 0 : .do_tenant_create(TenantCreateRequest {
5511 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
5512 0 : generation,
5513 0 :
5514 0 : shard_parameters: ShardParameters {
5515 0 : count: shard_count,
5516 0 : stripe_size,
5517 0 : },
5518 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
5519 0 : config: TenantConfig::default(),
5520 0 : })
5521 0 : .await?;
5522 :
5523 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
5524 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
5525 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
5526 : // reconcile, as reconciliation includes notifying compute.
5527 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
5528 0 : }
5529 :
5530 0 : Ok(response)
5531 0 : }
5532 :
5533 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
5534 : /// we don't have to make TenantShard clonable in the return path.
5535 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5536 0 : let serialized = {
5537 0 : let locked = self.inner.read().unwrap();
5538 0 : let result = locked.tenants.values().collect::<Vec<_>>();
5539 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
5540 : };
5541 :
5542 0 : hyper::Response::builder()
5543 0 : .status(hyper::StatusCode::OK)
5544 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5545 0 : .body(hyper::Body::from(serialized))
5546 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5547 0 : }
5548 :
5549 : /// Check the consistency of in-memory state vs. persistent state, and check that the
5550 : /// scheduler's statistics are up to date.
5551 : ///
5552 : /// These consistency checks expect an **idle** system. If changes are going on while
5553 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
5554 : /// checks, but not suitable for running continuously in the background in the field.
5555 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
5556 0 : let (mut expect_nodes, mut expect_shards) = {
5557 0 : let locked = self.inner.read().unwrap();
5558 0 :
5559 0 : locked
5560 0 : .scheduler
5561 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
5562 0 : .context("Scheduler checks")
5563 0 : .map_err(ApiError::InternalServerError)?;
5564 :
5565 0 : let expect_nodes = locked
5566 0 : .nodes
5567 0 : .values()
5568 0 : .map(|n| n.to_persistent())
5569 0 : .collect::<Vec<_>>();
5570 0 :
5571 0 : let expect_shards = locked
5572 0 : .tenants
5573 0 : .values()
5574 0 : .map(|t| t.to_persistent())
5575 0 : .collect::<Vec<_>>();
5576 :
5577 : // This method can only validate the state of an idle system: if a reconcile is in
5578 : // progress, fail out early to avoid giving false errors on state that won't match
5579 : // between database and memory under a ReconcileResult is processed.
5580 0 : for t in locked.tenants.values() {
5581 0 : if t.reconciler.is_some() {
5582 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5583 0 : "Shard {} reconciliation in progress",
5584 0 : t.tenant_shard_id
5585 0 : )));
5586 0 : }
5587 : }
5588 :
5589 0 : (expect_nodes, expect_shards)
5590 : };
5591 :
5592 0 : let mut nodes = self.persistence.list_nodes().await?;
5593 0 : expect_nodes.sort_by_key(|n| n.node_id);
5594 0 : nodes.sort_by_key(|n| n.node_id);
5595 :
5596 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
5597 0 : let node_result = if nodes != expect_nodes {
5598 0 : tracing::error!("Consistency check failed on nodes.");
5599 0 : tracing::error!(
5600 0 : "Nodes in memory: {}",
5601 0 : serde_json::to_string(&expect_nodes)
5602 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5603 : );
5604 0 : tracing::error!(
5605 0 : "Nodes in database: {}",
5606 0 : serde_json::to_string(&nodes)
5607 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5608 : );
5609 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
5610 0 : "Node consistency failure"
5611 0 : )))
5612 : } else {
5613 0 : Ok(())
5614 : };
5615 :
5616 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
5617 0 : persistent_shards
5618 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5619 0 :
5620 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5621 :
5622 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
5623 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
5624 : // fields are added, before doing a comparison.
5625 0 : for tsp in &mut persistent_shards {
5626 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
5627 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
5628 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
5629 : }
5630 :
5631 0 : if persistent_shards != expect_shards {
5632 0 : tracing::error!("Consistency check failed on shards.");
5633 :
5634 0 : tracing::error!(
5635 0 : "Shards in memory: {}",
5636 0 : serde_json::to_string(&expect_shards)
5637 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5638 : );
5639 0 : tracing::error!(
5640 0 : "Shards in database: {}",
5641 0 : serde_json::to_string(&persistent_shards)
5642 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5643 : );
5644 :
5645 : // The total dump log lines above are useful in testing but in the field grafana will
5646 : // usually just drop them because they're so large. So we also do some explicit logging
5647 : // of just the diffs.
5648 0 : let persistent_shards = persistent_shards
5649 0 : .into_iter()
5650 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5651 0 : .collect::<HashMap<_, _>>();
5652 0 : let expect_shards = expect_shards
5653 0 : .into_iter()
5654 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5655 0 : .collect::<HashMap<_, _>>();
5656 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
5657 0 : match expect_shards.get(tenant_shard_id) {
5658 : None => {
5659 0 : tracing::error!(
5660 0 : "Shard {} found in database but not in memory",
5661 : tenant_shard_id
5662 : );
5663 : }
5664 0 : Some(expect_tsp) => {
5665 0 : if expect_tsp != persistent_tsp {
5666 0 : tracing::error!(
5667 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
5668 0 : tenant_shard_id,
5669 0 : serde_json::to_string(expect_tsp).unwrap(),
5670 0 : serde_json::to_string(&persistent_tsp).unwrap()
5671 : );
5672 0 : }
5673 : }
5674 : }
5675 : }
5676 :
5677 : // Having already logged any differences, log any shards that simply aren't present in the database
5678 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
5679 0 : if !persistent_shards.contains_key(tenant_shard_id) {
5680 0 : tracing::error!(
5681 0 : "Shard {} found in memory but not in database: {}",
5682 0 : tenant_shard_id,
5683 0 : serde_json::to_string(memory_tsp)
5684 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5685 : );
5686 0 : }
5687 : }
5688 :
5689 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5690 0 : "Shard consistency failure"
5691 0 : )));
5692 0 : }
5693 0 :
5694 0 : node_result
5695 0 : }
5696 :
5697 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
5698 : /// we don't have to make TenantShard clonable in the return path.
5699 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5700 0 : let serialized = {
5701 0 : let locked = self.inner.read().unwrap();
5702 0 : serde_json::to_string(&locked.scheduler)
5703 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5704 : };
5705 :
5706 0 : hyper::Response::builder()
5707 0 : .status(hyper::StatusCode::OK)
5708 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5709 0 : .body(hyper::Body::from(serialized))
5710 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5711 0 : }
5712 :
5713 : /// This is for debug/support only: we simply drop all state for a tenant, without
5714 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
5715 : /// tenants that were on this node.
5716 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
5717 0 : self.persistence.delete_node(node_id).await?;
5718 :
5719 0 : let mut locked = self.inner.write().unwrap();
5720 :
5721 0 : for shard in locked.tenants.values_mut() {
5722 0 : shard.deref_node(node_id);
5723 0 : shard.observed.locations.remove(&node_id);
5724 0 : }
5725 :
5726 0 : let mut nodes = (*locked.nodes).clone();
5727 0 : nodes.remove(&node_id);
5728 0 : locked.nodes = Arc::new(nodes);
5729 0 : metrics::METRICS_REGISTRY
5730 0 : .metrics_group
5731 0 : .storage_controller_pageserver_nodes
5732 0 : .set(locked.nodes.len() as i64);
5733 0 :
5734 0 : locked.scheduler.node_remove(node_id);
5735 0 :
5736 0 : Ok(())
5737 0 : }
5738 :
5739 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
5740 : /// that we don't leave any bad state behind in the storage controller, but unclean
5741 : /// in the sense that we are not carefully draining the node.
5742 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
5743 0 : let _node_lock =
5744 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
5745 :
5746 : // 1. Atomically update in-memory state:
5747 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
5748 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
5749 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
5750 : // re-insert references to this node into the ObservedState of shards
5751 : // - drop the node from the scheduler
5752 : {
5753 0 : let mut locked = self.inner.write().unwrap();
5754 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5755 0 :
5756 0 : {
5757 0 : let mut nodes_mut = (*nodes).deref().clone();
5758 0 : match nodes_mut.get_mut(&node_id) {
5759 0 : Some(node) => {
5760 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
5761 0 : // if we crash it would not be desirable to leave the node paused after a restart.
5762 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
5763 0 : }
5764 : None => {
5765 0 : tracing::info!(
5766 0 : "Node not found: presuming this is a retry and returning success"
5767 : );
5768 0 : return Ok(());
5769 : }
5770 : }
5771 :
5772 0 : *nodes = Arc::new(nodes_mut);
5773 : }
5774 :
5775 0 : for (_tenant_id, mut schedule_context, shards) in
5776 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
5777 : {
5778 0 : for shard in shards {
5779 0 : if shard.deref_node(node_id) {
5780 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
5781 : // TODO: implement force flag to remove a node even if we can't reschedule
5782 : // a tenant
5783 0 : tracing::error!(
5784 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
5785 : shard.tenant_shard_id
5786 : );
5787 0 : return Err(e.into());
5788 : } else {
5789 0 : tracing::info!(
5790 0 : "Rescheduled shard {} away from node during deletion",
5791 : shard.tenant_shard_id
5792 : )
5793 : }
5794 :
5795 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
5796 0 : }
5797 :
5798 : // Here we remove an existing observed location for the node we're removing, and it will
5799 : // not be re-added by a reconciler's completion because we filter out removed nodes in
5800 : // process_result.
5801 : //
5802 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
5803 : // means any reconciles we spawned will know about the node we're deleting, enabling them
5804 : // to do live migrations if it's still online.
5805 0 : shard.observed.locations.remove(&node_id);
5806 : }
5807 : }
5808 :
5809 0 : scheduler.node_remove(node_id);
5810 0 :
5811 0 : {
5812 0 : let mut nodes_mut = (**nodes).clone();
5813 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
5814 0 : // Ensure that any reconciler holding an Arc<> to this node will
5815 0 : // drop out when trying to RPC to it (setting Offline state sets the
5816 0 : // cancellation token on the Node object).
5817 0 : removed_node.set_availability(NodeAvailability::Offline);
5818 0 : }
5819 0 : *nodes = Arc::new(nodes_mut);
5820 0 : metrics::METRICS_REGISTRY
5821 0 : .metrics_group
5822 0 : .storage_controller_pageserver_nodes
5823 0 : .set(nodes.len() as i64);
5824 0 : }
5825 0 : }
5826 0 :
5827 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
5828 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
5829 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
5830 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
5831 0 : // that exists.
5832 0 :
5833 0 : // 2. Actually delete the node from the database and from in-memory state
5834 0 : tracing::info!("Deleting node from database");
5835 0 : self.persistence.delete_node(node_id).await?;
5836 :
5837 0 : Ok(())
5838 0 : }
5839 :
5840 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
5841 0 : let nodes = {
5842 0 : self.inner
5843 0 : .read()
5844 0 : .unwrap()
5845 0 : .nodes
5846 0 : .values()
5847 0 : .cloned()
5848 0 : .collect::<Vec<_>>()
5849 0 : };
5850 0 :
5851 0 : Ok(nodes)
5852 0 : }
5853 :
5854 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
5855 0 : self.inner
5856 0 : .read()
5857 0 : .unwrap()
5858 0 : .nodes
5859 0 : .get(&node_id)
5860 0 : .cloned()
5861 0 : .ok_or(ApiError::NotFound(
5862 0 : format!("Node {node_id} not registered").into(),
5863 0 : ))
5864 0 : }
5865 :
5866 0 : pub(crate) async fn get_node_shards(
5867 0 : &self,
5868 0 : node_id: NodeId,
5869 0 : ) -> Result<NodeShardResponse, ApiError> {
5870 0 : let locked = self.inner.read().unwrap();
5871 0 : let mut shards = Vec::new();
5872 0 : for (tid, tenant) in locked.tenants.iter() {
5873 0 : let is_intended_secondary = match (
5874 0 : tenant.intent.get_attached() == &Some(node_id),
5875 0 : tenant.intent.get_secondary().contains(&node_id),
5876 0 : ) {
5877 : (true, true) => {
5878 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5879 0 : "{} attached as primary+secondary on the same node",
5880 0 : tid
5881 0 : )))
5882 : }
5883 0 : (true, false) => Some(false),
5884 0 : (false, true) => Some(true),
5885 0 : (false, false) => None,
5886 : };
5887 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
5888 0 : tenant.observed.locations.get(&node_id)
5889 : {
5890 0 : Some(conf.secondary_conf.is_some())
5891 : } else {
5892 0 : None
5893 : };
5894 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
5895 0 : shards.push(NodeShard {
5896 0 : tenant_shard_id: *tid,
5897 0 : is_intended_secondary,
5898 0 : is_observed_secondary,
5899 0 : });
5900 0 : }
5901 : }
5902 0 : Ok(NodeShardResponse { node_id, shards })
5903 0 : }
5904 :
5905 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
5906 0 : self.persistence.get_leader().await
5907 0 : }
5908 :
5909 0 : pub(crate) async fn node_register(
5910 0 : &self,
5911 0 : register_req: NodeRegisterRequest,
5912 0 : ) -> Result<(), ApiError> {
5913 0 : let _node_lock = trace_exclusive_lock(
5914 0 : &self.node_op_locks,
5915 0 : register_req.node_id,
5916 0 : NodeOperations::Register,
5917 0 : )
5918 0 : .await;
5919 :
5920 : #[derive(PartialEq)]
5921 : enum RegistrationStatus {
5922 : UpToDate,
5923 : NeedUpdate,
5924 : Mismatched,
5925 : New,
5926 : }
5927 :
5928 0 : let registration_status = {
5929 0 : let locked = self.inner.read().unwrap();
5930 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
5931 0 : if node.registration_match(®ister_req) {
5932 0 : if node.need_update(®ister_req) {
5933 0 : RegistrationStatus::NeedUpdate
5934 : } else {
5935 0 : RegistrationStatus::UpToDate
5936 : }
5937 : } else {
5938 0 : RegistrationStatus::Mismatched
5939 : }
5940 : } else {
5941 0 : RegistrationStatus::New
5942 : }
5943 : };
5944 :
5945 0 : match registration_status {
5946 : RegistrationStatus::UpToDate => {
5947 0 : tracing::info!(
5948 0 : "Node {} re-registered with matching address and is up to date",
5949 : register_req.node_id
5950 : );
5951 :
5952 0 : return Ok(());
5953 : }
5954 : RegistrationStatus::Mismatched => {
5955 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
5956 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
5957 : // a fixed address through the lifetime of a node.
5958 0 : tracing::warn!(
5959 0 : "Node {} tried to register with different address",
5960 : register_req.node_id
5961 : );
5962 0 : return Err(ApiError::Conflict(
5963 0 : "Node is already registered with different address".to_string(),
5964 0 : ));
5965 : }
5966 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
5967 0 : // fallthrough
5968 0 : }
5969 0 : }
5970 0 :
5971 0 : // We do not require that a node is actually online when registered (it will start life
5972 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
5973 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
5974 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
5975 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
5976 0 : if tokio::net::lookup_host(format!(
5977 0 : "{}:{}",
5978 0 : register_req.listen_http_addr, register_req.listen_http_port
5979 0 : ))
5980 0 : .await
5981 0 : .is_err()
5982 : {
5983 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
5984 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
5985 : // we return a soft 503 error, to encourage callers to retry past transient issues.
5986 0 : return Err(ApiError::ResourceUnavailable(
5987 0 : format!(
5988 0 : "Node {} tried to register with unknown DNS name '{}'",
5989 0 : register_req.node_id, register_req.listen_http_addr
5990 0 : )
5991 0 : .into(),
5992 0 : ));
5993 0 : }
5994 0 :
5995 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
5996 0 : return Err(ApiError::PreconditionFailed(
5997 0 : format!(
5998 0 : "Node {} has no https port, but use_https is enabled",
5999 0 : register_req.node_id
6000 0 : )
6001 0 : .into(),
6002 0 : ));
6003 0 : }
6004 0 :
6005 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
6006 0 : // This ensures that before we use it for anything or expose it via any external
6007 0 : // API, it is guaranteed to be available after a restart.
6008 0 : let new_node = Node::new(
6009 0 : register_req.node_id,
6010 0 : register_req.listen_http_addr,
6011 0 : register_req.listen_http_port,
6012 0 : register_req.listen_https_port,
6013 0 : register_req.listen_pg_addr,
6014 0 : register_req.listen_pg_port,
6015 0 : register_req.availability_zone_id.clone(),
6016 0 : self.config.use_https_pageserver_api,
6017 0 : );
6018 0 : let new_node = match new_node {
6019 0 : Ok(new_node) => new_node,
6020 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
6021 : };
6022 :
6023 0 : match registration_status {
6024 0 : RegistrationStatus::New => self.persistence.insert_node(&new_node).await?,
6025 : RegistrationStatus::NeedUpdate => {
6026 0 : self.persistence
6027 0 : .update_node_on_registration(
6028 0 : register_req.node_id,
6029 0 : register_req.listen_https_port,
6030 0 : )
6031 0 : .await?
6032 : }
6033 0 : _ => unreachable!("Other statuses have been processed earlier"),
6034 : }
6035 :
6036 0 : let mut locked = self.inner.write().unwrap();
6037 0 : let mut new_nodes = (*locked.nodes).clone();
6038 0 :
6039 0 : locked.scheduler.node_upsert(&new_node);
6040 0 : new_nodes.insert(register_req.node_id, new_node);
6041 0 :
6042 0 : locked.nodes = Arc::new(new_nodes);
6043 0 :
6044 0 : metrics::METRICS_REGISTRY
6045 0 : .metrics_group
6046 0 : .storage_controller_pageserver_nodes
6047 0 : .set(locked.nodes.len() as i64);
6048 0 :
6049 0 : match registration_status {
6050 : RegistrationStatus::New => {
6051 0 : tracing::info!(
6052 0 : "Registered pageserver {} ({}), now have {} pageservers",
6053 0 : register_req.node_id,
6054 0 : register_req.availability_zone_id,
6055 0 : locked.nodes.len()
6056 : );
6057 : }
6058 : RegistrationStatus::NeedUpdate => {
6059 0 : tracing::info!(
6060 0 : "Re-registered and updated node {} ({})",
6061 : register_req.node_id,
6062 : register_req.availability_zone_id,
6063 : );
6064 : }
6065 0 : _ => unreachable!("Other statuses have been processed earlier"),
6066 : }
6067 0 : Ok(())
6068 0 : }
6069 :
6070 : /// Configure in-memory and persistent state of a node as requested
6071 : ///
6072 : /// Note that this function does not trigger any immediate side effects in response
6073 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
6074 0 : async fn node_state_configure(
6075 0 : &self,
6076 0 : node_id: NodeId,
6077 0 : availability: Option<NodeAvailability>,
6078 0 : scheduling: Option<NodeSchedulingPolicy>,
6079 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
6080 0 : ) -> Result<AvailabilityTransition, ApiError> {
6081 0 : if let Some(scheduling) = scheduling {
6082 : // Scheduling is a persistent part of Node: we must write updates to the database before
6083 : // applying them in memory
6084 0 : self.persistence
6085 0 : .update_node_scheduling_policy(node_id, scheduling)
6086 0 : .await?;
6087 0 : }
6088 :
6089 : // If we're activating a node, then before setting it active we must reconcile any shard locations
6090 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
6091 : // by calling [`Self::node_activate_reconcile`]
6092 : //
6093 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
6094 : // nothing else can mutate its availability while we run.
6095 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
6096 0 : let (activate_node, availability_transition) = {
6097 0 : let locked = self.inner.read().unwrap();
6098 0 : let Some(node) = locked.nodes.get(&node_id) else {
6099 0 : return Err(ApiError::NotFound(
6100 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6101 0 : ));
6102 : };
6103 :
6104 0 : (
6105 0 : node.clone(),
6106 0 : node.get_availability_transition(input_availability),
6107 0 : )
6108 : };
6109 :
6110 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
6111 0 : self.node_activate_reconcile(activate_node, node_lock)
6112 0 : .await?;
6113 0 : }
6114 0 : availability_transition
6115 : } else {
6116 0 : AvailabilityTransition::Unchanged
6117 : };
6118 :
6119 : // Apply changes from the request to our in-memory state for the Node
6120 0 : let mut locked = self.inner.write().unwrap();
6121 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
6122 0 :
6123 0 : let mut new_nodes = (**nodes).clone();
6124 :
6125 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
6126 0 : return Err(ApiError::NotFound(
6127 0 : anyhow::anyhow!("Node not registered").into(),
6128 0 : ));
6129 : };
6130 :
6131 0 : if let Some(availability) = availability {
6132 0 : node.set_availability(availability);
6133 0 : }
6134 :
6135 0 : if let Some(scheduling) = scheduling {
6136 0 : node.set_scheduling(scheduling);
6137 0 : }
6138 :
6139 : // Update the scheduler, in case the elegibility of the node for new shards has changed
6140 0 : scheduler.node_upsert(node);
6141 0 :
6142 0 : let new_nodes = Arc::new(new_nodes);
6143 0 : locked.nodes = new_nodes;
6144 0 :
6145 0 : Ok(availability_transition)
6146 0 : }
6147 :
6148 : /// Handle availability transition of one node
6149 : ///
6150 : /// Note that you should first call [`Self::node_state_configure`] to update
6151 : /// the in-memory state referencing that node. If you need to handle more than one transition
6152 : /// consider using [`Self::handle_node_availability_transitions`].
6153 0 : async fn handle_node_availability_transition(
6154 0 : &self,
6155 0 : node_id: NodeId,
6156 0 : transition: AvailabilityTransition,
6157 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
6158 0 : ) -> Result<(), ApiError> {
6159 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
6160 0 : match transition {
6161 : AvailabilityTransition::ToOffline => {
6162 0 : tracing::info!("Node {} transition to offline", node_id);
6163 :
6164 0 : let mut locked = self.inner.write().unwrap();
6165 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6166 0 :
6167 0 : let mut tenants_affected: usize = 0;
6168 :
6169 0 : for (_tenant_id, mut schedule_context, shards) in
6170 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6171 : {
6172 0 : for tenant_shard in shards {
6173 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
6174 0 : if let Some(observed_loc) =
6175 0 : tenant_shard.observed.locations.get_mut(&node_id)
6176 0 : {
6177 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
6178 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
6179 0 : observed_loc.conf = None;
6180 0 : }
6181 :
6182 0 : if nodes.len() == 1 {
6183 : // Special case for single-node cluster: there is no point trying to reschedule
6184 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
6185 : // failures to schedule them.
6186 0 : continue;
6187 0 : }
6188 0 :
6189 0 : if !nodes
6190 0 : .values()
6191 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6192 : {
6193 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
6194 : // trying to reschedule since there's nowhere else to go. Without this
6195 : // branch we incorrectly detach tenants in response to node unavailability.
6196 0 : continue;
6197 0 : }
6198 0 :
6199 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
6200 0 : tenant_shard.sequence = tenant_shard.sequence.next();
6201 0 :
6202 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
6203 0 : Err(e) => {
6204 0 : // It is possible that some tenants will become unschedulable when too many pageservers
6205 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
6206 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
6207 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
6208 : }
6209 : Ok(()) => {
6210 0 : if self
6211 0 : .maybe_reconcile_shard(
6212 0 : tenant_shard,
6213 0 : nodes,
6214 0 : ReconcilerPriority::Normal,
6215 0 : )
6216 0 : .is_some()
6217 0 : {
6218 0 : tenants_affected += 1;
6219 0 : };
6220 : }
6221 : }
6222 0 : }
6223 : }
6224 : }
6225 0 : tracing::info!(
6226 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
6227 : tenants_affected,
6228 : node_id
6229 : )
6230 : }
6231 : AvailabilityTransition::ToActive => {
6232 0 : tracing::info!("Node {} transition to active", node_id);
6233 :
6234 0 : let mut locked = self.inner.write().unwrap();
6235 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
6236 :
6237 : // When a node comes back online, we must reconcile any tenant that has a None observed
6238 : // location on the node.
6239 0 : for tenant_shard in tenants.values_mut() {
6240 : // If a reconciliation is already in progress, rely on the previous scheduling
6241 : // decision and skip triggering a new reconciliation.
6242 0 : if tenant_shard.reconciler.is_some() {
6243 0 : continue;
6244 0 : }
6245 :
6246 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
6247 0 : if observed_loc.conf.is_none() {
6248 0 : self.maybe_reconcile_shard(
6249 0 : tenant_shard,
6250 0 : nodes,
6251 0 : ReconcilerPriority::Normal,
6252 0 : );
6253 0 : }
6254 0 : }
6255 : }
6256 :
6257 : // TODO: in the background, we should balance work back onto this pageserver
6258 : }
6259 : // No action required for the intermediate unavailable state.
6260 : // When we transition into active or offline from the unavailable state,
6261 : // the correct handling above will kick in.
6262 : AvailabilityTransition::ToWarmingUpFromActive => {
6263 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
6264 : }
6265 : AvailabilityTransition::ToWarmingUpFromOffline => {
6266 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
6267 : }
6268 : AvailabilityTransition::Unchanged => {
6269 0 : tracing::debug!("Node {} no availability change during config", node_id);
6270 : }
6271 : }
6272 :
6273 0 : Ok(())
6274 0 : }
6275 :
6276 : /// Handle availability transition for multiple nodes
6277 : ///
6278 : /// Note that you should first call [`Self::node_state_configure`] for
6279 : /// all nodes being handled here for the handling to use fresh in-memory state.
6280 0 : async fn handle_node_availability_transitions(
6281 0 : &self,
6282 0 : transitions: Vec<(
6283 0 : NodeId,
6284 0 : TracingExclusiveGuard<NodeOperations>,
6285 0 : AvailabilityTransition,
6286 0 : )>,
6287 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
6288 0 : let mut errors = Vec::default();
6289 0 : for (node_id, node_lock, transition) in transitions {
6290 0 : let res = self
6291 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
6292 0 : .await;
6293 0 : if let Err(err) = res {
6294 0 : errors.push((node_id, err));
6295 0 : }
6296 : }
6297 :
6298 0 : if errors.is_empty() {
6299 0 : Ok(())
6300 : } else {
6301 0 : Err(errors)
6302 : }
6303 0 : }
6304 :
6305 0 : pub(crate) async fn node_configure(
6306 0 : &self,
6307 0 : node_id: NodeId,
6308 0 : availability: Option<NodeAvailability>,
6309 0 : scheduling: Option<NodeSchedulingPolicy>,
6310 0 : ) -> Result<(), ApiError> {
6311 0 : let node_lock =
6312 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
6313 :
6314 0 : let transition = self
6315 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
6316 0 : .await?;
6317 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
6318 0 : .await
6319 0 : }
6320 :
6321 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
6322 : /// operation for HTTP api.
6323 0 : pub(crate) async fn external_node_configure(
6324 0 : &self,
6325 0 : node_id: NodeId,
6326 0 : availability: Option<NodeAvailability>,
6327 0 : scheduling: Option<NodeSchedulingPolicy>,
6328 0 : ) -> Result<(), ApiError> {
6329 0 : {
6330 0 : let locked = self.inner.read().unwrap();
6331 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
6332 0 : return Err(ApiError::PreconditionFailed(
6333 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
6334 0 : ));
6335 0 : }
6336 0 : }
6337 0 :
6338 0 : self.node_configure(node_id, availability, scheduling).await
6339 0 : }
6340 :
6341 0 : pub(crate) async fn start_node_drain(
6342 0 : self: &Arc<Self>,
6343 0 : node_id: NodeId,
6344 0 : ) -> Result<(), ApiError> {
6345 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
6346 0 : let locked = self.inner.read().unwrap();
6347 0 : let nodes = &locked.nodes;
6348 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6349 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6350 0 : ))?;
6351 0 : let schedulable_nodes_count = nodes
6352 0 : .iter()
6353 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6354 0 : .count();
6355 0 :
6356 0 : (
6357 0 : locked
6358 0 : .ongoing_operation
6359 0 : .as_ref()
6360 0 : .map(|ongoing| ongoing.operation),
6361 0 : node.is_available(),
6362 0 : node.get_scheduling(),
6363 0 : schedulable_nodes_count,
6364 0 : )
6365 0 : };
6366 :
6367 0 : if let Some(ongoing) = ongoing_op {
6368 0 : return Err(ApiError::PreconditionFailed(
6369 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6370 0 : ));
6371 0 : }
6372 0 :
6373 0 : if !node_available {
6374 0 : return Err(ApiError::ResourceUnavailable(
6375 0 : format!("Node {node_id} is currently unavailable").into(),
6376 0 : ));
6377 0 : }
6378 0 :
6379 0 : if schedulable_nodes_count == 0 {
6380 0 : return Err(ApiError::PreconditionFailed(
6381 0 : "No other schedulable nodes to drain to".into(),
6382 0 : ));
6383 0 : }
6384 0 :
6385 0 : match node_policy {
6386 : NodeSchedulingPolicy::Active => {
6387 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
6388 0 : .await?;
6389 :
6390 0 : let cancel = self.cancel.child_token();
6391 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6392 :
6393 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6394 0 : operation: Operation::Drain(Drain { node_id }),
6395 0 : cancel: cancel.clone(),
6396 0 : });
6397 :
6398 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
6399 :
6400 0 : tokio::task::spawn({
6401 0 : let service = self.clone();
6402 0 : let cancel = cancel.clone();
6403 0 : async move {
6404 0 : let _gate_guard = gate_guard;
6405 0 :
6406 0 : scopeguard::defer! {
6407 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6408 0 :
6409 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
6410 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
6411 0 : } else {
6412 0 : panic!("We always remove the same operation")
6413 0 : }
6414 0 : }
6415 0 :
6416 0 : tracing::info!("Drain background operation starting");
6417 0 : let res = service.drain_node(node_id, cancel).await;
6418 0 : match res {
6419 : Ok(()) => {
6420 0 : tracing::info!("Drain background operation completed successfully");
6421 : }
6422 : Err(OperationError::Cancelled) => {
6423 0 : tracing::info!("Drain background operation was cancelled");
6424 : }
6425 0 : Err(err) => {
6426 0 : tracing::error!("Drain background operation encountered: {err}")
6427 : }
6428 : }
6429 0 : }
6430 0 : }.instrument(span));
6431 0 : }
6432 : NodeSchedulingPolicy::Draining => {
6433 0 : return Err(ApiError::Conflict(format!(
6434 0 : "Node {node_id} has drain in progress"
6435 0 : )));
6436 : }
6437 0 : policy => {
6438 0 : return Err(ApiError::PreconditionFailed(
6439 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
6440 0 : ));
6441 : }
6442 : }
6443 :
6444 0 : Ok(())
6445 0 : }
6446 :
6447 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
6448 0 : let node_available = {
6449 0 : let locked = self.inner.read().unwrap();
6450 0 : let nodes = &locked.nodes;
6451 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6452 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6453 0 : ))?;
6454 :
6455 0 : node.is_available()
6456 0 : };
6457 0 :
6458 0 : if !node_available {
6459 0 : return Err(ApiError::ResourceUnavailable(
6460 0 : format!("Node {node_id} is currently unavailable").into(),
6461 0 : ));
6462 0 : }
6463 :
6464 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6465 0 : if let Operation::Drain(drain) = op_handler.operation {
6466 0 : if drain.node_id == node_id {
6467 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6468 0 : op_handler.cancel.cancel();
6469 0 : return Ok(());
6470 0 : }
6471 0 : }
6472 0 : }
6473 :
6474 0 : Err(ApiError::PreconditionFailed(
6475 0 : format!("Node {node_id} has no drain in progress").into(),
6476 0 : ))
6477 0 : }
6478 :
6479 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
6480 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
6481 0 : let locked = self.inner.read().unwrap();
6482 0 : let nodes = &locked.nodes;
6483 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6484 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6485 0 : ))?;
6486 :
6487 0 : (
6488 0 : locked
6489 0 : .ongoing_operation
6490 0 : .as_ref()
6491 0 : .map(|ongoing| ongoing.operation),
6492 0 : node.is_available(),
6493 0 : node.get_scheduling(),
6494 0 : nodes.len(),
6495 0 : )
6496 0 : };
6497 :
6498 0 : if let Some(ongoing) = ongoing_op {
6499 0 : return Err(ApiError::PreconditionFailed(
6500 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6501 0 : ));
6502 0 : }
6503 0 :
6504 0 : if !node_available {
6505 0 : return Err(ApiError::ResourceUnavailable(
6506 0 : format!("Node {node_id} is currently unavailable").into(),
6507 0 : ));
6508 0 : }
6509 0 :
6510 0 : if total_nodes_count <= 1 {
6511 0 : return Err(ApiError::PreconditionFailed(
6512 0 : "No other nodes to fill from".into(),
6513 0 : ));
6514 0 : }
6515 0 :
6516 0 : match node_policy {
6517 : NodeSchedulingPolicy::Active => {
6518 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
6519 0 : .await?;
6520 :
6521 0 : let cancel = self.cancel.child_token();
6522 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6523 :
6524 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6525 0 : operation: Operation::Fill(Fill { node_id }),
6526 0 : cancel: cancel.clone(),
6527 0 : });
6528 :
6529 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
6530 :
6531 0 : tokio::task::spawn({
6532 0 : let service = self.clone();
6533 0 : let cancel = cancel.clone();
6534 0 : async move {
6535 0 : let _gate_guard = gate_guard;
6536 0 :
6537 0 : scopeguard::defer! {
6538 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6539 0 :
6540 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
6541 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
6542 0 : } else {
6543 0 : panic!("We always remove the same operation")
6544 0 : }
6545 0 : }
6546 0 :
6547 0 : tracing::info!("Fill background operation starting");
6548 0 : let res = service.fill_node(node_id, cancel).await;
6549 0 : match res {
6550 : Ok(()) => {
6551 0 : tracing::info!("Fill background operation completed successfully");
6552 : }
6553 : Err(OperationError::Cancelled) => {
6554 0 : tracing::info!("Fill background operation was cancelled");
6555 : }
6556 0 : Err(err) => {
6557 0 : tracing::error!("Fill background operation encountered: {err}")
6558 : }
6559 : }
6560 0 : }
6561 0 : }.instrument(span));
6562 0 : }
6563 : NodeSchedulingPolicy::Filling => {
6564 0 : return Err(ApiError::Conflict(format!(
6565 0 : "Node {node_id} has fill in progress"
6566 0 : )));
6567 : }
6568 0 : policy => {
6569 0 : return Err(ApiError::PreconditionFailed(
6570 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
6571 0 : ));
6572 : }
6573 : }
6574 :
6575 0 : Ok(())
6576 0 : }
6577 :
6578 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
6579 0 : let node_available = {
6580 0 : let locked = self.inner.read().unwrap();
6581 0 : let nodes = &locked.nodes;
6582 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6583 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6584 0 : ))?;
6585 :
6586 0 : node.is_available()
6587 0 : };
6588 0 :
6589 0 : if !node_available {
6590 0 : return Err(ApiError::ResourceUnavailable(
6591 0 : format!("Node {node_id} is currently unavailable").into(),
6592 0 : ));
6593 0 : }
6594 :
6595 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6596 0 : if let Operation::Fill(fill) = op_handler.operation {
6597 0 : if fill.node_id == node_id {
6598 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6599 0 : op_handler.cancel.cancel();
6600 0 : return Ok(());
6601 0 : }
6602 0 : }
6603 0 : }
6604 :
6605 0 : Err(ApiError::PreconditionFailed(
6606 0 : format!("Node {node_id} has no fill in progress").into(),
6607 0 : ))
6608 0 : }
6609 :
6610 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
6611 : /// configuration
6612 0 : fn maybe_reconcile_shard(
6613 0 : &self,
6614 0 : shard: &mut TenantShard,
6615 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6616 0 : priority: ReconcilerPriority,
6617 0 : ) -> Option<ReconcilerWaiter> {
6618 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
6619 0 : }
6620 :
6621 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
6622 0 : fn get_reconciler_units(
6623 0 : &self,
6624 0 : priority: ReconcilerPriority,
6625 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
6626 0 : let units = match priority {
6627 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
6628 : ReconcilerPriority::High => {
6629 0 : match self
6630 0 : .priority_reconciler_concurrency
6631 0 : .clone()
6632 0 : .try_acquire_owned()
6633 : {
6634 0 : Ok(u) => Ok(u),
6635 : Err(TryAcquireError::NoPermits) => {
6636 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
6637 : // the normal priority semaphore.
6638 0 : self.reconciler_concurrency.clone().try_acquire_owned()
6639 : }
6640 0 : Err(e) => Err(e),
6641 : }
6642 : }
6643 : };
6644 :
6645 0 : units.map(ReconcileUnits::new)
6646 0 : }
6647 :
6648 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
6649 0 : fn maybe_configured_reconcile_shard(
6650 0 : &self,
6651 0 : shard: &mut TenantShard,
6652 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6653 0 : reconciler_config: ReconcilerConfig,
6654 0 : ) -> Option<ReconcilerWaiter> {
6655 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
6656 0 :
6657 0 : match reconcile_needed {
6658 0 : ReconcileNeeded::No => return None,
6659 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
6660 0 : ReconcileNeeded::Yes => {
6661 0 : // Fall through to try and acquire units for spawning reconciler
6662 0 : }
6663 : };
6664 :
6665 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
6666 0 : Ok(u) => u,
6667 : Err(_) => {
6668 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
6669 0 : "Concurrency limited: enqueued for reconcile later");
6670 0 : if !shard.delayed_reconcile {
6671 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
6672 0 : Err(TrySendError::Closed(_)) => {
6673 0 : // Weird mid-shutdown case?
6674 0 : }
6675 : Err(TrySendError::Full(_)) => {
6676 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
6677 0 : tracing::warn!(
6678 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
6679 : );
6680 : }
6681 0 : Ok(()) => {
6682 0 : shard.delayed_reconcile = true;
6683 0 : }
6684 : }
6685 0 : }
6686 :
6687 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
6688 : // number to advance. When this function is eventually called again and succeeds in getting units,
6689 : // it will spawn a reconciler that makes this waiter complete.
6690 0 : return Some(shard.future_reconcile_waiter());
6691 : }
6692 : };
6693 :
6694 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
6695 : // Gate closed: we're shutting down, drop out.
6696 0 : return None;
6697 : };
6698 :
6699 0 : shard.spawn_reconciler(
6700 0 : &self.result_tx,
6701 0 : nodes,
6702 0 : &self.compute_hook,
6703 0 : reconciler_config,
6704 0 : &self.config,
6705 0 : &self.persistence,
6706 0 : units,
6707 0 : gate_guard,
6708 0 : &self.reconcilers_cancel,
6709 0 : )
6710 0 : }
6711 :
6712 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
6713 : /// Additionally, reschedule tenants that require it.
6714 : ///
6715 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
6716 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
6717 : /// available. A return value of 0 indicates that everything is fully reconciled already.
6718 0 : fn reconcile_all(&self) -> usize {
6719 0 : let mut locked = self.inner.write().unwrap();
6720 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6721 0 : let pageservers = nodes.clone();
6722 0 :
6723 0 : // This function is an efficient place to update lazy statistics, since we are walking
6724 0 : // all tenants.
6725 0 : let mut pending_reconciles = 0;
6726 0 : let mut az_violations = 0;
6727 0 :
6728 0 : // If we find any tenants to drop from memory, stash them to offload after
6729 0 : // we're done traversing the map of tenants.
6730 0 : let mut drop_detached_tenants = Vec::new();
6731 0 :
6732 0 : let mut reconciles_spawned = 0;
6733 0 : for shard in tenants.values_mut() {
6734 : // Accumulate scheduling statistics
6735 0 : if let (Some(attached), Some(preferred)) =
6736 0 : (shard.intent.get_attached(), shard.preferred_az())
6737 : {
6738 0 : let node_az = nodes
6739 0 : .get(attached)
6740 0 : .expect("Nodes exist if referenced")
6741 0 : .get_availability_zone_id();
6742 0 : if node_az != preferred {
6743 0 : az_violations += 1;
6744 0 : }
6745 0 : }
6746 :
6747 : // Skip checking if this shard is already enqueued for reconciliation
6748 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
6749 : // If there is something delayed, then return a nonzero count so that
6750 : // callers like reconcile_all_now do not incorrectly get the impression
6751 : // that the system is in a quiescent state.
6752 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
6753 0 : pending_reconciles += 1;
6754 0 : continue;
6755 0 : }
6756 0 :
6757 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
6758 0 : // dirty, spawn another rone
6759 0 : if self
6760 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
6761 0 : .is_some()
6762 0 : {
6763 0 : reconciles_spawned += 1;
6764 0 : } else if shard.delayed_reconcile {
6765 0 : // Shard wanted to reconcile but for some reason couldn't.
6766 0 : pending_reconciles += 1;
6767 0 : }
6768 :
6769 : // If this tenant is detached, try dropping it from memory. This is usually done
6770 : // proactively in [`Self::process_results`], but we do it here to handle the edge
6771 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
6772 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
6773 0 : && shard.policy == PlacementPolicy::Detached
6774 : {
6775 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
6776 0 : shard.tenant_shard_id.tenant_id,
6777 0 : TenantOperations::DropDetached,
6778 0 : ) {
6779 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
6780 0 : }
6781 0 : }
6782 : }
6783 :
6784 : // Some metrics are calculated from SchedulerNode state, update these periodically
6785 0 : scheduler.update_metrics();
6786 :
6787 : // Process any deferred tenant drops
6788 0 : for (tenant_id, guard) in drop_detached_tenants {
6789 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
6790 0 : }
6791 :
6792 0 : metrics::METRICS_REGISTRY
6793 0 : .metrics_group
6794 0 : .storage_controller_schedule_az_violation
6795 0 : .set(az_violations as i64);
6796 0 :
6797 0 : metrics::METRICS_REGISTRY
6798 0 : .metrics_group
6799 0 : .storage_controller_pending_reconciles
6800 0 : .set(pending_reconciles as i64);
6801 0 :
6802 0 : reconciles_spawned
6803 0 : }
6804 :
6805 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
6806 : /// could be scheduled somewhere better:
6807 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
6808 : /// * e.g. after a node fails then recovers, to move some work back to it
6809 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
6810 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
6811 : /// we did the split, but are probably better placed elsewhere.
6812 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
6813 : /// * e.g. after a shard split, some locations will be on the same node (where the split
6814 : /// happened), and will probably be better placed elsewhere.
6815 : ///
6816 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
6817 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
6818 : /// according to those same soft constraints.
6819 0 : async fn optimize_all(&self) -> usize {
6820 : // Limit on how many shards' optmizations each call to this function will execute. Combined
6821 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
6822 : // trickle of optimizations in the background, rather than executing a large number in parallel
6823 : // when a change occurs.
6824 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
6825 :
6826 : // Synchronous prepare: scan shards for possible scheduling optimizations
6827 0 : let candidate_work = self.optimize_all_plan();
6828 0 : let candidate_work_len = candidate_work.len();
6829 :
6830 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
6831 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
6832 :
6833 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
6834 0 :
6835 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
6836 0 : let mut reconciles_spawned = 0;
6837 0 : let mut optimizations_applied = 0;
6838 0 : let mut locked = self.inner.write().unwrap();
6839 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6840 0 : for (tenant_shard_id, optimization) in validated_work {
6841 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6842 : // Shard was dropped between planning and execution;
6843 0 : continue;
6844 : };
6845 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
6846 0 : if shard.apply_optimization(scheduler, optimization) {
6847 0 : optimizations_applied += 1;
6848 0 : if self
6849 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
6850 0 : .is_some()
6851 0 : {
6852 0 : reconciles_spawned += 1;
6853 0 : }
6854 0 : }
6855 :
6856 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
6857 0 : break;
6858 0 : }
6859 : }
6860 :
6861 0 : if was_work_filtered {
6862 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
6863 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
6864 0 : // as these validations start passing.
6865 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
6866 0 : }
6867 :
6868 0 : reconciles_spawned
6869 0 : }
6870 :
6871 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
6872 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
6873 : // this higher than the execution limit gives us a chance to execute some work even if the first
6874 : // few optimizations we find are not ready.
6875 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
6876 :
6877 0 : let mut work = Vec::new();
6878 0 : let mut locked = self.inner.write().unwrap();
6879 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6880 :
6881 : // We are going to plan a bunch of optimisations before applying any of them, so the
6882 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
6883 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
6884 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
6885 : // to ignore the utilisation component of the score.
6886 :
6887 0 : for (_tenant_id, schedule_context, shards) in
6888 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
6889 : {
6890 0 : for shard in shards {
6891 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
6892 0 : break;
6893 0 : }
6894 0 : match shard.get_scheduling_policy() {
6895 0 : ShardSchedulingPolicy::Active => {
6896 0 : // Ok to do optimization
6897 0 : }
6898 : ShardSchedulingPolicy::Essential
6899 : | ShardSchedulingPolicy::Pause
6900 : | ShardSchedulingPolicy::Stop => {
6901 : // Policy prevents optimizing this shard.
6902 0 : continue;
6903 : }
6904 : }
6905 :
6906 0 : if !matches!(shard.splitting, SplitState::Idle)
6907 0 : || matches!(shard.policy, PlacementPolicy::Detached)
6908 0 : || shard.reconciler.is_some()
6909 : {
6910 : // Do not start any optimizations while another change to the tenant is ongoing: this
6911 : // is not necessary for correctness, but simplifies operations and implicitly throttles
6912 : // optimization changes to happen in a "trickle" over time.
6913 0 : continue;
6914 0 : }
6915 0 :
6916 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
6917 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
6918 0 : if cfg!(feature = "testing") {
6919 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
6920 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
6921 : // panic in prod if we hit this, or spend cycles on it in prod.
6922 0 : assert!(shard
6923 0 : .optimize_attachment(scheduler, &schedule_context)
6924 0 : .is_none());
6925 0 : assert!(shard
6926 0 : .optimize_secondary(scheduler, &schedule_context)
6927 0 : .is_none());
6928 0 : }
6929 0 : continue;
6930 0 : }
6931 :
6932 0 : if let Some(optimization) =
6933 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
6934 : // its primary location based on soft constraints, cut it over.
6935 0 : shard.optimize_attachment(scheduler, &schedule_context)
6936 : {
6937 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
6938 0 : work.push((shard.tenant_shard_id, optimization));
6939 0 : break;
6940 0 : } else if let Some(optimization) =
6941 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
6942 : // better placed on another node, based on ScheduleContext, then adjust it. This
6943 : // covers cases like after a shard split, where we might have too many shards
6944 : // in the same tenant with secondary locations on the node where they originally split.
6945 0 : shard.optimize_secondary(scheduler, &schedule_context)
6946 : {
6947 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
6948 0 : work.push((shard.tenant_shard_id, optimization));
6949 0 : break;
6950 0 : }
6951 : }
6952 : }
6953 :
6954 0 : work
6955 0 : }
6956 :
6957 0 : async fn optimize_all_validate(
6958 0 : &self,
6959 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
6960 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
6961 0 : // Take a clone of the node map to use outside the lock in async validation phase
6962 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
6963 0 :
6964 0 : let mut want_secondary_status = Vec::new();
6965 0 :
6966 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
6967 0 : // check that the state of locations is acceptable to run the optimization, such as
6968 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
6969 0 : // in a live migration.
6970 0 : let mut validated_work = Vec::new();
6971 0 : for (tenant_shard_id, optimization) in candidate_work {
6972 0 : match optimization.action {
6973 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
6974 : old_attached_node_id: _,
6975 0 : new_attached_node_id,
6976 0 : }) => {
6977 0 : match validation_nodes.get(&new_attached_node_id) {
6978 0 : None => {
6979 0 : // Node was dropped between planning and validation
6980 0 : }
6981 0 : Some(node) => {
6982 0 : if !node.is_available() {
6983 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
6984 0 : } else {
6985 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
6986 0 : // remote API requests concurrently.
6987 0 : want_secondary_status.push((
6988 0 : tenant_shard_id,
6989 0 : node.clone(),
6990 0 : optimization,
6991 0 : ));
6992 0 : }
6993 : }
6994 : }
6995 : }
6996 : ScheduleOptimizationAction::ReplaceSecondary(_)
6997 : | ScheduleOptimizationAction::CreateSecondary(_)
6998 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
6999 : // No extra checks needed to manage secondaries: this does not interrupt client access
7000 0 : validated_work.push((tenant_shard_id, optimization))
7001 : }
7002 : };
7003 : }
7004 :
7005 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
7006 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
7007 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
7008 0 : let results = self
7009 0 : .tenant_for_shards_api(
7010 0 : want_secondary_status
7011 0 : .iter()
7012 0 : .map(|i| (i.0, i.1.clone()))
7013 0 : .collect(),
7014 0 : |tenant_shard_id, client| async move {
7015 0 : client.tenant_secondary_status(tenant_shard_id).await
7016 0 : },
7017 0 : 1,
7018 0 : 1,
7019 0 : SHORT_RECONCILE_TIMEOUT,
7020 0 : &self.cancel,
7021 0 : )
7022 0 : .await;
7023 :
7024 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
7025 0 : want_secondary_status.into_iter().zip(results.into_iter())
7026 : {
7027 0 : match secondary_status {
7028 0 : Err(e) => {
7029 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
7030 : }
7031 0 : Ok(progress) => {
7032 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
7033 : // them in an optimization
7034 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
7035 :
7036 0 : if progress.heatmap_mtime.is_none()
7037 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
7038 0 : && progress.bytes_downloaded != progress.bytes_total
7039 0 : || progress.bytes_total - progress.bytes_downloaded
7040 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
7041 : {
7042 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
7043 :
7044 : #[cfg(feature = "testing")]
7045 0 : if progress.heatmap_mtime.is_none() {
7046 : // No heatmap might mean the attached location has never uploaded one, or that
7047 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
7048 : // but fairly common in tests.
7049 0 : self.kick_secondary_download(tenant_shard_id).await;
7050 0 : }
7051 : } else {
7052 : // Location looks ready: proceed
7053 0 : tracing::info!(
7054 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
7055 : );
7056 0 : validated_work.push((tenant_shard_id, optimization))
7057 : }
7058 : }
7059 : }
7060 : }
7061 :
7062 0 : validated_work
7063 0 : }
7064 :
7065 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
7066 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
7067 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
7068 : /// we have this helper to move things along faster.
7069 : #[cfg(feature = "testing")]
7070 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
7071 0 : let (attached_node, secondaries) = {
7072 0 : let locked = self.inner.read().unwrap();
7073 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
7074 0 : tracing::warn!(
7075 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
7076 : );
7077 0 : return;
7078 : };
7079 :
7080 0 : let Some(attached) = shard.intent.get_attached() else {
7081 0 : tracing::warn!(
7082 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
7083 : );
7084 0 : return;
7085 : };
7086 :
7087 0 : let secondaries = shard
7088 0 : .intent
7089 0 : .get_secondary()
7090 0 : .iter()
7091 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
7092 0 : .collect::<Vec<_>>();
7093 0 :
7094 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
7095 0 : };
7096 0 :
7097 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
7098 0 : // a 'kick' to let scheduling optimisation run more promptly.
7099 0 : match attached_node
7100 0 : .with_client_retries(
7101 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
7102 0 : &self.config.pageserver_jwt_token,
7103 0 : 3,
7104 0 : 10,
7105 0 : SHORT_RECONCILE_TIMEOUT,
7106 0 : &self.cancel,
7107 0 : )
7108 0 : .await
7109 : {
7110 0 : Some(Err(e)) => {
7111 0 : tracing::info!(
7112 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
7113 : );
7114 : }
7115 : None => {
7116 0 : tracing::info!(
7117 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
7118 : );
7119 : }
7120 : Some(Ok(_)) => {
7121 0 : tracing::info!(
7122 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
7123 : );
7124 : }
7125 : }
7126 :
7127 0 : for secondary_node in secondaries {
7128 0 : match secondary_node
7129 0 : .with_client_retries(
7130 0 : |client| async move {
7131 0 : client
7132 0 : .tenant_secondary_download(
7133 0 : tenant_shard_id,
7134 0 : Some(Duration::from_secs(1)),
7135 0 : )
7136 0 : .await
7137 0 : },
7138 0 : &self.config.pageserver_jwt_token,
7139 0 : 3,
7140 0 : 10,
7141 0 : SHORT_RECONCILE_TIMEOUT,
7142 0 : &self.cancel,
7143 0 : )
7144 0 : .await
7145 : {
7146 0 : Some(Err(e)) => {
7147 0 : tracing::info!(
7148 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
7149 : );
7150 : }
7151 : None => {
7152 0 : tracing::info!("Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}");
7153 : }
7154 0 : Some(Ok(progress)) => {
7155 0 : tracing::info!("Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}");
7156 : }
7157 : }
7158 : }
7159 0 : }
7160 :
7161 : /// Look for shards which are oversized and in need of splitting
7162 0 : async fn autosplit_tenants(self: &Arc<Self>) {
7163 0 : let Some(split_threshold) = self.config.split_threshold else {
7164 : // Auto-splitting is disabled
7165 0 : return;
7166 : };
7167 :
7168 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7169 :
7170 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
7171 :
7172 0 : let mut top_n = Vec::new();
7173 0 :
7174 0 : // Call into each node to look for big tenants
7175 0 : let top_n_request = TopTenantShardsRequest {
7176 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
7177 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
7178 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
7179 0 : // disk space impact of one shard.
7180 0 : order_by: models::TenantSorting::MaxLogicalSize,
7181 0 : limit: 10,
7182 0 : where_shards_lt: Some(SPLIT_TO_MAX),
7183 0 : where_gt: Some(split_threshold),
7184 0 : };
7185 0 : for node in nodes.values() {
7186 0 : let request_ref = &top_n_request;
7187 0 : match node
7188 0 : .with_client_retries(
7189 0 : |client| async move {
7190 0 : let request = request_ref.clone();
7191 0 : client.top_tenant_shards(request.clone()).await
7192 0 : },
7193 0 : &self.config.pageserver_jwt_token,
7194 0 : 3,
7195 0 : 3,
7196 0 : Duration::from_secs(5),
7197 0 : &self.cancel,
7198 0 : )
7199 0 : .await
7200 : {
7201 0 : Some(Ok(node_top_n)) => {
7202 0 : top_n.extend(node_top_n.shards.into_iter());
7203 0 : }
7204 : Some(Err(mgmt_api::Error::Cancelled)) => {
7205 0 : continue;
7206 : }
7207 0 : Some(Err(e)) => {
7208 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
7209 0 : continue;
7210 : }
7211 : None => {
7212 : // Node is shutting down
7213 0 : continue;
7214 : }
7215 : };
7216 : }
7217 :
7218 : // Pick the biggest tenant to split first
7219 0 : top_n.sort_by_key(|i| i.resident_size);
7220 0 :
7221 0 : // Filter out tenants in a prohibiting scheduling mode
7222 0 : {
7223 0 : let locked = self.inner.read().unwrap();
7224 0 : top_n.retain(|i| {
7225 0 : if let Some(shard) = locked.tenants.get(&i.id) {
7226 0 : matches!(shard.get_scheduling_policy(), ShardSchedulingPolicy::Active)
7227 : } else {
7228 0 : false
7229 : }
7230 0 : });
7231 0 : }
7232 :
7233 0 : let Some(split_candidate) = top_n.into_iter().next() else {
7234 0 : tracing::debug!("No split-elegible shards found");
7235 0 : return;
7236 : };
7237 :
7238 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
7239 : // want to block the background reconcile loop on this.
7240 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
7241 :
7242 0 : let this = self.clone();
7243 0 : tokio::spawn(
7244 0 : async move {
7245 0 : match this
7246 0 : .tenant_shard_split(
7247 0 : split_candidate.id.tenant_id,
7248 0 : TenantShardSplitRequest {
7249 0 : // Always split to the max number of shards: this avoids stepping through
7250 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
7251 0 : // each time as a tenant grows, and is not too expensive because our max shard
7252 0 : // count is relatively low anyway.
7253 0 : // This policy will be adjusted in future once we support higher shard count.
7254 0 : new_shard_count: SPLIT_TO_MAX.literal(),
7255 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
7256 0 : },
7257 0 : )
7258 0 : .await
7259 : {
7260 : Ok(_) => {
7261 0 : tracing::info!("Successful auto-split");
7262 : }
7263 0 : Err(e) => {
7264 0 : tracing::error!("Auto-split failed: {e}");
7265 : }
7266 : }
7267 0 : }
7268 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
7269 : );
7270 0 : }
7271 :
7272 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
7273 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
7274 : /// put the system into a quiescent state where future background reconciliations won't do anything.
7275 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
7276 0 : let reconciles_spawned = self.reconcile_all();
7277 0 : let reconciles_spawned = if reconciles_spawned == 0 {
7278 : // Only optimize when we are otherwise idle
7279 0 : self.optimize_all().await
7280 : } else {
7281 0 : reconciles_spawned
7282 : };
7283 :
7284 0 : let waiters = {
7285 0 : let mut waiters = Vec::new();
7286 0 : let locked = self.inner.read().unwrap();
7287 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
7288 0 : if let Some(waiter) = shard.get_waiter() {
7289 0 : waiters.push(waiter);
7290 0 : }
7291 : }
7292 0 : waiters
7293 0 : };
7294 0 :
7295 0 : let waiter_count = waiters.len();
7296 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
7297 0 : Ok(()) => {}
7298 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
7299 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
7300 0 : {
7301 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
7302 0 : // because some other change superceded it. We will return a nonzero number,
7303 0 : // so the caller knows they might have to call again to quiesce the system.
7304 0 : }
7305 0 : Err(e) => {
7306 0 : return Err(e);
7307 : }
7308 : };
7309 :
7310 0 : tracing::info!(
7311 0 : "{} reconciles in reconcile_all, {} waiters",
7312 : reconciles_spawned,
7313 : waiter_count
7314 : );
7315 :
7316 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
7317 0 : }
7318 :
7319 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
7320 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
7321 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
7322 0 : self.reconcilers_cancel.cancel();
7323 0 : self.reconcilers_gate.close().await;
7324 :
7325 : // Signal the background loop in [`Service::process_results`] to exit once
7326 : // it has proccessed the results from all the reconciles we cancelled earlier.
7327 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
7328 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
7329 0 : self.result_tx.closed().await;
7330 0 : }
7331 :
7332 0 : pub async fn shutdown(&self) {
7333 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
7334 0 : .await;
7335 :
7336 : // Background tasks hold gate guards: this notifies them of the cancellation and
7337 : // waits for them all to complete.
7338 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
7339 0 : self.cancel.cancel();
7340 0 : self.gate.close().await;
7341 0 : }
7342 :
7343 : /// Spot check the download lag for a secondary location of a shard.
7344 : /// Should be used as a heuristic, since it's not always precise: the
7345 : /// secondary might have not downloaded the new heat map yet and, hence,
7346 : /// is not aware of the lag.
7347 : ///
7348 : /// Returns:
7349 : /// * Ok(None) if the lag could not be determined from the status,
7350 : /// * Ok(Some(_)) if the lag could be determind
7351 : /// * Err on failures to query the pageserver.
7352 0 : async fn secondary_lag(
7353 0 : &self,
7354 0 : secondary: &NodeId,
7355 0 : tenant_shard_id: TenantShardId,
7356 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
7357 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7358 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
7359 0 : StatusCode::NOT_FOUND,
7360 0 : format!("Node with id {} not found", secondary),
7361 0 : ))?;
7362 :
7363 0 : match node
7364 0 : .with_client_retries(
7365 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
7366 0 : &self.config.pageserver_jwt_token,
7367 0 : 1,
7368 0 : 3,
7369 0 : Duration::from_millis(250),
7370 0 : &self.cancel,
7371 0 : )
7372 0 : .await
7373 : {
7374 0 : Some(Ok(status)) => match status.heatmap_mtime {
7375 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
7376 0 : None => Ok(None),
7377 : },
7378 0 : Some(Err(e)) => Err(e),
7379 0 : None => Err(mgmt_api::Error::Cancelled),
7380 : }
7381 0 : }
7382 :
7383 : /// Drain a node by moving the shards attached to it as primaries.
7384 : /// This is a long running operation and it should run as a separate Tokio task.
7385 0 : pub(crate) async fn drain_node(
7386 0 : self: &Arc<Self>,
7387 0 : node_id: NodeId,
7388 0 : cancel: CancellationToken,
7389 0 : ) -> Result<(), OperationError> {
7390 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
7391 0 : let max_secondary_lag_bytes = self
7392 0 : .config
7393 0 : .max_secondary_lag_bytes
7394 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
7395 :
7396 : // By default, live migrations are generous about the wait time for getting
7397 : // the secondary location up to speed. When draining, give up earlier in order
7398 : // to not stall the operation when a cold secondary is encountered.
7399 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7400 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7401 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
7402 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7403 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7404 0 : .build();
7405 0 :
7406 0 : let mut waiters = Vec::new();
7407 0 :
7408 0 : let mut tid_iter = TenantShardIterator::new({
7409 0 : let service = self.clone();
7410 0 : move |last_inspected_shard: Option<TenantShardId>| {
7411 0 : let locked = &service.inner.read().unwrap();
7412 0 : let tenants = &locked.tenants;
7413 0 : let entry = match last_inspected_shard {
7414 0 : Some(skip_past) => {
7415 0 : // Skip to the last seen tenant shard id
7416 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
7417 0 :
7418 0 : // Skip past the last seen
7419 0 : cursor.nth(1)
7420 : }
7421 0 : None => tenants.first_key_value(),
7422 : };
7423 :
7424 0 : entry.map(|(tid, _)| tid).copied()
7425 0 : }
7426 0 : });
7427 :
7428 0 : while !tid_iter.finished() {
7429 0 : if cancel.is_cancelled() {
7430 0 : match self
7431 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7432 0 : .await
7433 : {
7434 0 : Ok(()) => return Err(OperationError::Cancelled),
7435 0 : Err(err) => {
7436 0 : return Err(OperationError::FinalizeError(
7437 0 : format!(
7438 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7439 0 : node_id, err
7440 0 : )
7441 0 : .into(),
7442 0 : ));
7443 : }
7444 : }
7445 0 : }
7446 0 :
7447 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
7448 :
7449 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7450 0 : let tid = match tid_iter.next() {
7451 0 : Some(tid) => tid,
7452 : None => {
7453 0 : break;
7454 : }
7455 : };
7456 :
7457 0 : let tid_drain = TenantShardDrain {
7458 0 : drained_node: node_id,
7459 0 : tenant_shard_id: tid,
7460 0 : };
7461 :
7462 0 : let dest_node_id = {
7463 0 : let locked = self.inner.read().unwrap();
7464 0 :
7465 0 : match tid_drain
7466 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
7467 : {
7468 0 : Some(node_id) => node_id,
7469 : None => {
7470 0 : continue;
7471 : }
7472 : }
7473 : };
7474 :
7475 0 : match self.secondary_lag(&dest_node_id, tid).await {
7476 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
7477 0 : // The secondary is reasonably up to date.
7478 0 : // Migrate to it
7479 0 : }
7480 0 : Ok(Some(lag)) => {
7481 0 : tracing::info!(
7482 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7483 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
7484 : );
7485 0 : continue;
7486 : }
7487 : Ok(None) => {
7488 0 : tracing::info!(
7489 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7490 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
7491 : );
7492 0 : continue;
7493 : }
7494 0 : Err(err) => {
7495 0 : tracing::warn!(
7496 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7497 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
7498 : );
7499 0 : continue;
7500 : }
7501 : }
7502 :
7503 : {
7504 0 : let mut locked = self.inner.write().unwrap();
7505 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7506 0 : let rescheduled = tid_drain.reschedule_to_secondary(
7507 0 : dest_node_id,
7508 0 : tenants,
7509 0 : scheduler,
7510 0 : nodes,
7511 0 : )?;
7512 :
7513 0 : if let Some(tenant_shard) = rescheduled {
7514 0 : let waiter = self.maybe_configured_reconcile_shard(
7515 0 : tenant_shard,
7516 0 : nodes,
7517 0 : reconciler_config,
7518 0 : );
7519 0 : if let Some(some) = waiter {
7520 0 : waiters.push(some);
7521 0 : }
7522 0 : }
7523 : }
7524 : }
7525 :
7526 0 : waiters = self
7527 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7528 0 : .await;
7529 :
7530 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
7531 : }
7532 :
7533 0 : while !waiters.is_empty() {
7534 0 : if cancel.is_cancelled() {
7535 0 : match self
7536 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7537 0 : .await
7538 : {
7539 0 : Ok(()) => return Err(OperationError::Cancelled),
7540 0 : Err(err) => {
7541 0 : return Err(OperationError::FinalizeError(
7542 0 : format!(
7543 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7544 0 : node_id, err
7545 0 : )
7546 0 : .into(),
7547 0 : ));
7548 : }
7549 : }
7550 0 : }
7551 0 :
7552 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
7553 :
7554 0 : waiters = self
7555 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7556 0 : .await;
7557 : }
7558 :
7559 : // At this point we have done the best we could to drain shards from this node.
7560 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
7561 : // to complete the drain.
7562 0 : if let Err(err) = self
7563 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
7564 0 : .await
7565 : {
7566 : // This is not fatal. Anything that is polling the node scheduling policy to detect
7567 : // the end of the drain operations will hang, but all such places should enforce an
7568 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
7569 : // by the counterpart fill operation.
7570 0 : return Err(OperationError::FinalizeError(
7571 0 : format!(
7572 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
7573 0 : )
7574 0 : .into(),
7575 0 : ));
7576 0 : }
7577 0 :
7578 0 : Ok(())
7579 0 : }
7580 :
7581 : /// Create a node fill plan (pick secondaries to promote), based on:
7582 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
7583 : /// outside their home AZ, should be migrated back here.
7584 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
7585 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
7586 : /// those shards have a home AZ that doesn't match the node we're filling.
7587 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
7588 0 : let mut locked = self.inner.write().unwrap();
7589 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7590 0 :
7591 0 : let node_az = nodes
7592 0 : .get(&node_id)
7593 0 : .expect("Node must exist")
7594 0 : .get_availability_zone_id()
7595 0 : .clone();
7596 0 :
7597 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
7598 0 : let mut plan = Vec::new();
7599 0 :
7600 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
7601 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
7602 0 :
7603 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
7604 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
7605 0 : let respect_azs = nodes
7606 0 : .values()
7607 0 : .map(|n| n.get_availability_zone_id())
7608 0 : .unique()
7609 0 : .count()
7610 0 : > 1;
7611 :
7612 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
7613 : // requires it.
7614 0 : for (tsid, tenant_shard) in tenants {
7615 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7616 : // Shard doesn't have a secondary on this node, ignore it.
7617 0 : continue;
7618 0 : }
7619 0 :
7620 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
7621 0 : // shards which belong on this node, not to promote shards whose scheduling preference
7622 0 : // would be on their currently attached node. So will avoid promoting shards whose
7623 0 : // home AZ doesn't match the AZ of the node we're filling.
7624 0 : match tenant_shard.preferred_az() {
7625 0 : _ if !respect_azs => {
7626 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7627 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7628 0 : }
7629 : }
7630 : None => {
7631 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
7632 : // will only do so if our target shard count requires it.
7633 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7634 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7635 0 : }
7636 : }
7637 0 : Some(az) if az == &node_az => {
7638 : // This shard's home AZ is equal to the node we're filling: it should
7639 : // be moved back to this node as part of filling, unless its currently
7640 : // attached location is also in its home AZ.
7641 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7642 0 : if nodes
7643 0 : .get(primary)
7644 0 : .expect("referenced node must exist")
7645 0 : .get_availability_zone_id()
7646 0 : != tenant_shard
7647 0 : .preferred_az()
7648 0 : .expect("tenant must have an AZ preference")
7649 : {
7650 0 : plan.push(*tsid)
7651 0 : }
7652 : } else {
7653 0 : plan.push(*tsid)
7654 : }
7655 : }
7656 0 : Some(_) => {
7657 0 : // This shard's home AZ is somewhere other than the node we're filling,
7658 0 : // it may not be moved back to this node as part of filling. Ignore it
7659 0 : }
7660 : }
7661 : }
7662 :
7663 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
7664 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
7665 0 :
7666 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
7667 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
7668 0 :
7669 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
7670 :
7671 0 : for (node_id, attached) in nodes_by_load {
7672 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
7673 0 : if !available {
7674 0 : continue;
7675 0 : }
7676 0 :
7677 0 : if plan.len() >= fill_requirement
7678 0 : || free_tids_by_node.is_empty()
7679 0 : || attached <= expected_attached
7680 : {
7681 0 : break;
7682 0 : }
7683 0 :
7684 0 : let can_take = attached - expected_attached;
7685 0 : let needed = fill_requirement - plan.len();
7686 0 : let mut take = std::cmp::min(can_take, needed);
7687 0 :
7688 0 : let mut remove_node = false;
7689 0 : while take > 0 {
7690 0 : match free_tids_by_node.get_mut(&node_id) {
7691 0 : Some(tids) => match tids.pop() {
7692 0 : Some(tid) => {
7693 0 : let max_promote_for_tenant = std::cmp::max(
7694 0 : tid.shard_count.count() as usize / locked.nodes.len(),
7695 0 : 1,
7696 0 : );
7697 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
7698 0 : if *promoted < max_promote_for_tenant {
7699 0 : plan.push(tid);
7700 0 : *promoted += 1;
7701 0 : take -= 1;
7702 0 : }
7703 : }
7704 : None => {
7705 0 : remove_node = true;
7706 0 : break;
7707 : }
7708 : },
7709 : None => {
7710 0 : break;
7711 : }
7712 : }
7713 : }
7714 :
7715 0 : if remove_node {
7716 0 : free_tids_by_node.remove(&node_id);
7717 0 : }
7718 : }
7719 :
7720 0 : plan
7721 0 : }
7722 :
7723 : /// Fill a node by promoting its secondaries until the cluster is balanced
7724 : /// with regards to attached shard counts. Note that this operation only
7725 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
7726 : /// This is a long running operation and it should run as a separate Tokio task.
7727 0 : pub(crate) async fn fill_node(
7728 0 : &self,
7729 0 : node_id: NodeId,
7730 0 : cancel: CancellationToken,
7731 0 : ) -> Result<(), OperationError> {
7732 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7733 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7734 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
7735 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7736 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7737 0 : .build();
7738 0 :
7739 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
7740 0 : let mut waiters = Vec::new();
7741 :
7742 : // Execute the plan we've composed above. Before aplying each move from the plan,
7743 : // we validate to ensure that it has not gone stale in the meantime.
7744 0 : while !tids_to_promote.is_empty() {
7745 0 : if cancel.is_cancelled() {
7746 0 : match self
7747 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7748 0 : .await
7749 : {
7750 0 : Ok(()) => return Err(OperationError::Cancelled),
7751 0 : Err(err) => {
7752 0 : return Err(OperationError::FinalizeError(
7753 0 : format!(
7754 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7755 0 : node_id, err
7756 0 : )
7757 0 : .into(),
7758 0 : ));
7759 : }
7760 : }
7761 0 : }
7762 0 :
7763 0 : {
7764 0 : let mut locked = self.inner.write().unwrap();
7765 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7766 :
7767 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
7768 0 : format!("node {node_id} was removed").into(),
7769 0 : ))?;
7770 :
7771 0 : let current_policy = node.get_scheduling();
7772 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
7773 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
7774 : // about it
7775 0 : return Err(OperationError::NodeStateChanged(
7776 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
7777 0 : ));
7778 0 : }
7779 :
7780 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7781 0 : if let Some(tid) = tids_to_promote.pop() {
7782 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
7783 : // If the node being filled is not a secondary anymore,
7784 : // skip the promotion.
7785 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7786 0 : continue;
7787 0 : }
7788 0 :
7789 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
7790 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
7791 0 : Err(e) => {
7792 0 : tracing::warn!(
7793 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7794 0 : "Scheduling error when filling pageserver {} : {e}", node_id
7795 : );
7796 : }
7797 : Ok(()) => {
7798 0 : tracing::info!(
7799 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7800 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
7801 : node_id,
7802 : previously_attached_to,
7803 : node_id
7804 : );
7805 :
7806 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
7807 0 : tenant_shard,
7808 0 : nodes,
7809 0 : reconciler_config,
7810 0 : ) {
7811 0 : waiters.push(waiter);
7812 0 : }
7813 : }
7814 : }
7815 0 : }
7816 : } else {
7817 0 : break;
7818 : }
7819 : }
7820 : }
7821 :
7822 0 : waiters = self
7823 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7824 0 : .await;
7825 : }
7826 :
7827 0 : while !waiters.is_empty() {
7828 0 : if cancel.is_cancelled() {
7829 0 : match self
7830 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7831 0 : .await
7832 : {
7833 0 : Ok(()) => return Err(OperationError::Cancelled),
7834 0 : Err(err) => {
7835 0 : return Err(OperationError::FinalizeError(
7836 0 : format!(
7837 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7838 0 : node_id, err
7839 0 : )
7840 0 : .into(),
7841 0 : ));
7842 : }
7843 : }
7844 0 : }
7845 0 :
7846 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
7847 :
7848 0 : waiters = self
7849 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7850 0 : .await;
7851 : }
7852 :
7853 0 : if let Err(err) = self
7854 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7855 0 : .await
7856 : {
7857 : // This isn't a huge issue since the filling process starts upon request. However, it
7858 : // will prevent the next drain from starting. The only case in which this can fail
7859 : // is database unavailability. Such a case will require manual intervention.
7860 0 : return Err(OperationError::FinalizeError(
7861 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
7862 0 : .into(),
7863 0 : ));
7864 0 : }
7865 0 :
7866 0 : Ok(())
7867 0 : }
7868 :
7869 : /// Updates scrubber metadata health check results.
7870 0 : pub(crate) async fn metadata_health_update(
7871 0 : &self,
7872 0 : update_req: MetadataHealthUpdateRequest,
7873 0 : ) -> Result<(), ApiError> {
7874 0 : let now = chrono::offset::Utc::now();
7875 0 : let (healthy_records, unhealthy_records) = {
7876 0 : let locked = self.inner.read().unwrap();
7877 0 : let healthy_records = update_req
7878 0 : .healthy_tenant_shards
7879 0 : .into_iter()
7880 0 : // Retain only health records associated with tenant shards managed by storage controller.
7881 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7882 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
7883 0 : .collect();
7884 0 : let unhealthy_records = update_req
7885 0 : .unhealthy_tenant_shards
7886 0 : .into_iter()
7887 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7888 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
7889 0 : .collect();
7890 0 :
7891 0 : (healthy_records, unhealthy_records)
7892 0 : };
7893 0 :
7894 0 : self.persistence
7895 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
7896 0 : .await?;
7897 0 : Ok(())
7898 0 : }
7899 :
7900 : /// Lists the tenant shards that has unhealthy metadata status.
7901 0 : pub(crate) async fn metadata_health_list_unhealthy(
7902 0 : &self,
7903 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
7904 0 : let result = self
7905 0 : .persistence
7906 0 : .list_unhealthy_metadata_health_records()
7907 0 : .await?
7908 0 : .iter()
7909 0 : .map(|p| p.get_tenant_shard_id().unwrap())
7910 0 : .collect();
7911 0 :
7912 0 : Ok(result)
7913 0 : }
7914 :
7915 : /// Lists the tenant shards that have not been scrubbed for some duration.
7916 0 : pub(crate) async fn metadata_health_list_outdated(
7917 0 : &self,
7918 0 : not_scrubbed_for: Duration,
7919 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
7920 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
7921 0 : let result = self
7922 0 : .persistence
7923 0 : .list_outdated_metadata_health_records(earlier)
7924 0 : .await?
7925 0 : .into_iter()
7926 0 : .map(|record| record.into())
7927 0 : .collect();
7928 0 : Ok(result)
7929 0 : }
7930 :
7931 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
7932 0 : self.inner.read().unwrap().get_leadership_status()
7933 0 : }
7934 :
7935 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
7936 0 : tracing::info!("Received step down request from peer");
7937 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
7938 :
7939 0 : self.inner.write().unwrap().step_down();
7940 0 : // TODO: would it make sense to have a time-out for this?
7941 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
7942 0 : .await;
7943 :
7944 0 : let mut global_observed = GlobalObservedState::default();
7945 0 : let locked = self.inner.read().unwrap();
7946 0 : for (tid, tenant_shard) in locked.tenants.iter() {
7947 0 : global_observed
7948 0 : .0
7949 0 : .insert(*tid, tenant_shard.observed.clone());
7950 0 : }
7951 :
7952 0 : global_observed
7953 0 : }
7954 :
7955 0 : pub(crate) async fn safekeepers_list(
7956 0 : &self,
7957 0 : ) -> Result<Vec<SafekeeperDescribeResponse>, DatabaseError> {
7958 0 : let locked = self.inner.read().unwrap();
7959 0 : let mut list = locked
7960 0 : .safekeepers
7961 0 : .iter()
7962 0 : .map(|sk| sk.1.describe_response())
7963 0 : .collect::<Result<Vec<_>, _>>()?;
7964 0 : list.sort_by_key(|v| v.id);
7965 0 : Ok(list)
7966 0 : }
7967 :
7968 0 : pub(crate) async fn get_safekeeper(
7969 0 : &self,
7970 0 : id: i64,
7971 0 : ) -> Result<SafekeeperDescribeResponse, DatabaseError> {
7972 0 : let locked = self.inner.read().unwrap();
7973 0 : let sk = locked
7974 0 : .safekeepers
7975 0 : .get(&NodeId(id as u64))
7976 0 : .ok_or(diesel::result::Error::NotFound)?;
7977 0 : sk.describe_response()
7978 0 : }
7979 :
7980 0 : pub(crate) async fn upsert_safekeeper(
7981 0 : &self,
7982 0 : record: crate::persistence::SafekeeperUpsert,
7983 0 : ) -> Result<(), DatabaseError> {
7984 0 : let node_id = NodeId(record.id as u64);
7985 0 : self.persistence.safekeeper_upsert(record.clone()).await?;
7986 : {
7987 0 : let mut locked = self.inner.write().unwrap();
7988 0 : let mut safekeepers = (*locked.safekeepers).clone();
7989 0 : match safekeepers.entry(node_id) {
7990 0 : std::collections::hash_map::Entry::Occupied(mut entry) => {
7991 0 : entry.get_mut().update_from_record(record);
7992 0 : }
7993 0 : std::collections::hash_map::Entry::Vacant(entry) => {
7994 0 : entry.insert(Safekeeper::from_persistence(
7995 0 : crate::persistence::SafekeeperPersistence::from_upsert(
7996 0 : record,
7997 0 : SkSchedulingPolicy::Pause,
7998 0 : ),
7999 0 : CancellationToken::new(),
8000 0 : ));
8001 0 : }
8002 : }
8003 0 : locked.safekeepers = Arc::new(safekeepers);
8004 0 : }
8005 0 : Ok(())
8006 0 : }
8007 :
8008 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
8009 0 : &self,
8010 0 : id: i64,
8011 0 : scheduling_policy: SkSchedulingPolicy,
8012 0 : ) -> Result<(), DatabaseError> {
8013 0 : self.persistence
8014 0 : .set_safekeeper_scheduling_policy(id, scheduling_policy)
8015 0 : .await?;
8016 0 : let node_id = NodeId(id as u64);
8017 0 : // After the change has been persisted successfully, update the in-memory state
8018 0 : {
8019 0 : let mut locked = self.inner.write().unwrap();
8020 0 : let mut safekeepers = (*locked.safekeepers).clone();
8021 0 : let sk = safekeepers
8022 0 : .get_mut(&node_id)
8023 0 : .ok_or(DatabaseError::Logical("Not found".to_string()))?;
8024 0 : sk.set_scheduling_policy(scheduling_policy);
8025 0 :
8026 0 : locked.safekeepers = Arc::new(safekeepers);
8027 0 : }
8028 0 : Ok(())
8029 0 : }
8030 :
8031 0 : pub(crate) async fn update_shards_preferred_azs(
8032 0 : &self,
8033 0 : req: ShardsPreferredAzsRequest,
8034 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
8035 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
8036 0 : let updated = self
8037 0 : .persistence
8038 0 : .set_tenant_shard_preferred_azs(preferred_azs)
8039 0 : .await
8040 0 : .map_err(|err| {
8041 0 : ApiError::InternalServerError(anyhow::anyhow!(
8042 0 : "Failed to persist preferred AZs: {err}"
8043 0 : ))
8044 0 : })?;
8045 :
8046 0 : let mut updated_in_mem_and_db = Vec::default();
8047 0 :
8048 0 : let mut locked = self.inner.write().unwrap();
8049 0 : for (tid, az_id) in updated {
8050 0 : let shard = locked.tenants.get_mut(&tid);
8051 0 : if let Some(shard) = shard {
8052 0 : shard.set_preferred_az(az_id);
8053 0 : updated_in_mem_and_db.push(tid);
8054 0 : }
8055 : }
8056 :
8057 0 : Ok(ShardsPreferredAzsResponse {
8058 0 : updated: updated_in_mem_and_db,
8059 0 : })
8060 0 : }
8061 : }
|
