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