Line data Source code
1 : use hyper::Uri;
2 : use std::{
3 : borrow::Cow,
4 : cmp::Ordering,
5 : collections::{BTreeMap, HashMap, HashSet},
6 : error::Error,
7 : ops::Deref,
8 : path::PathBuf,
9 : str::FromStr,
10 : sync::Arc,
11 : time::{Duration, Instant},
12 : };
13 :
14 : use crate::{
15 : background_node_operations::{
16 : Drain, Fill, Operation, OperationError, OperationHandler, MAX_RECONCILES_PER_OPERATION,
17 : },
18 : compute_hook::NotifyError,
19 : drain_utils::{self, TenantShardDrain, TenantShardIterator},
20 : id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, TracingExclusiveGuard},
21 : leadership::Leadership,
22 : metrics,
23 : peer_client::GlobalObservedState,
24 : persistence::{
25 : AbortShardSplitStatus, ControllerPersistence, DatabaseResult, MetadataHealthPersistence,
26 : ShardGenerationState, TenantFilter,
27 : },
28 : reconciler::{ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder},
29 : scheduler::{MaySchedule, ScheduleContext, ScheduleError, ScheduleMode},
30 : tenant_shard::{
31 : MigrateAttachment, ObservedStateDelta, ReconcileNeeded, ReconcilerStatus,
32 : ScheduleOptimization, ScheduleOptimizationAction,
33 : },
34 : };
35 : use anyhow::Context;
36 : use control_plane::storage_controller::{
37 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
38 : };
39 : use diesel::result::DatabaseErrorKind;
40 : use futures::{stream::FuturesUnordered, StreamExt};
41 : use itertools::Itertools;
42 : use pageserver_api::{
43 : controller_api::{
44 : MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability, NodeRegisterRequest,
45 : NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy, ShardSchedulingPolicy,
46 : ShardsPreferredAzsRequest, ShardsPreferredAzsResponse, TenantCreateRequest,
47 : TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse,
48 : TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
49 : TenantShardMigrateRequest, TenantShardMigrateResponse,
50 : },
51 : models::{
52 : SecondaryProgress, TenantConfigRequest, TimelineArchivalConfigRequest,
53 : TopTenantShardsRequest,
54 : },
55 : };
56 : use reqwest::StatusCode;
57 : use tracing::{instrument, Instrument};
58 :
59 : use crate::pageserver_client::PageserverClient;
60 : use pageserver_api::{
61 : models::{
62 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
63 : PageserverUtilization, ShardParameters, TenantConfig, TenantLocationConfigRequest,
64 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
65 : TenantShardSplitResponse, TenantTimeTravelRequest, TimelineCreateRequest, TimelineInfo,
66 : },
67 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
68 : upcall_api::{
69 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
70 : ValidateResponse, ValidateResponseTenant,
71 : },
72 : };
73 : use pageserver_client::{mgmt_api, BlockUnblock};
74 : use tokio::sync::mpsc::error::TrySendError;
75 : use tokio_util::sync::CancellationToken;
76 : use utils::{
77 : completion::Barrier,
78 : failpoint_support,
79 : generation::Generation,
80 : http::error::ApiError,
81 : id::{NodeId, TenantId, TimelineId},
82 : sync::gate::Gate,
83 : };
84 :
85 : use crate::{
86 : compute_hook::ComputeHook,
87 : heartbeater::{Heartbeater, PageserverState},
88 : node::{AvailabilityTransition, Node},
89 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
90 : reconciler::attached_location_conf,
91 : scheduler::Scheduler,
92 : tenant_shard::{
93 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
94 : ReconcilerWaiter, TenantShard,
95 : },
96 : };
97 :
98 : pub mod chaos_injector;
99 :
100 : // For operations that should be quick, like attaching a new tenant
101 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
102 :
103 : // For operations that might be slow, like migrating a tenant with
104 : // some data in it.
105 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
106 :
107 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
108 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
109 : // input generation from future requests as authoritative.
110 : const INITIAL_GENERATION: Generation = Generation::new(0);
111 :
112 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
113 : /// up on unresponsive pageservers and proceed.
114 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
115 :
116 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
117 : /// This must be long enough to cover node restarts as well as normal operations: in future
118 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
119 :
120 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
121 : /// offline.
122 : ///
123 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
124 : /// handling of the re-attach response may take a long time and blocks heartbeats from
125 : /// being handled on the pageserver side.
126 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
127 :
128 : /// How often to send heartbeats to registered nodes?
129 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
130 :
131 : /// How long is too long for a reconciliation?
132 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
133 :
134 0 : #[derive(Clone, strum_macros::Display)]
135 : enum TenantOperations {
136 : Create,
137 : LocationConfig,
138 : ConfigSet,
139 : TimeTravelRemoteStorage,
140 : Delete,
141 : UpdatePolicy,
142 : ShardSplit,
143 : SecondaryDownload,
144 : TimelineCreate,
145 : TimelineDelete,
146 : AttachHook,
147 : TimelineArchivalConfig,
148 : TimelineDetachAncestor,
149 : TimelineGcBlockUnblock,
150 : }
151 :
152 0 : #[derive(Clone, strum_macros::Display)]
153 : enum NodeOperations {
154 : Register,
155 : Configure,
156 : Delete,
157 : }
158 :
159 : /// The leadership status for the storage controller process.
160 : /// Allowed transitions are:
161 : /// 1. Leader -> SteppedDown
162 : /// 2. Candidate -> Leader
163 : #[derive(
164 : Eq,
165 : PartialEq,
166 : Copy,
167 : Clone,
168 0 : strum_macros::Display,
169 0 : strum_macros::EnumIter,
170 : measured::FixedCardinalityLabel,
171 : )]
172 : #[strum(serialize_all = "snake_case")]
173 : pub(crate) enum LeadershipStatus {
174 : /// This is the steady state where the storage controller can produce
175 : /// side effects in the cluster.
176 : Leader,
177 : /// We've been notified to step down by another candidate. No reconciliations
178 : /// take place in this state.
179 : SteppedDown,
180 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
181 : #[allow(unused)]
182 : Candidate,
183 : }
184 :
185 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
186 :
187 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
188 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
189 : // than they're being pushed onto the queue.
190 : const MAX_DELAYED_RECONCILES: usize = 10000;
191 :
192 : // Top level state available to all HTTP handlers
193 : struct ServiceState {
194 : leadership_status: LeadershipStatus,
195 :
196 : tenants: BTreeMap<TenantShardId, TenantShard>,
197 :
198 : nodes: Arc<HashMap<NodeId, Node>>,
199 :
200 : scheduler: Scheduler,
201 :
202 : /// Ongoing background operation on the cluster if any is running.
203 : /// Note that only one such operation may run at any given time,
204 : /// hence the type choice.
205 : ongoing_operation: Option<OperationHandler>,
206 :
207 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
208 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
209 : }
210 :
211 : /// Transform an error from a pageserver into an error to return to callers of a storage
212 : /// controller API.
213 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
214 0 : match e {
215 0 : mgmt_api::Error::SendRequest(e) => {
216 0 : // Presume errors sending requests are connectivity/availability issues
217 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
218 : }
219 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
220 0 : // Presume errors receiving body are connectivity/availability issues
221 0 : ApiError::ResourceUnavailable(
222 0 : format!("{node} error receiving error body: {str}").into(),
223 0 : )
224 : }
225 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
226 0 : // Return 500 for decoding errors.
227 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
228 : }
229 0 : mgmt_api::Error::ReceiveBody(err) => {
230 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
231 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
232 0 : ApiError::ResourceUnavailable(
233 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
234 0 : )
235 : }
236 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
237 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
238 : }
239 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
240 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
241 : }
242 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
243 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
244 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
245 : // internal server errors, showing that something is wrong with the pageserver or
246 : // storage controller's auth configuration.
247 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
248 : }
249 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
250 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
251 0 : // turn ask our clients to wait + retry
252 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
253 : }
254 0 : mgmt_api::Error::ApiError(status, msg) => {
255 0 : // Presume general case of pageserver API errors is that we tried to do something
256 0 : // that can't be done right now.
257 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
258 : }
259 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
260 : }
261 0 : }
262 :
263 : impl ServiceState {
264 0 : fn new(
265 0 : nodes: HashMap<NodeId, Node>,
266 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
267 0 : scheduler: Scheduler,
268 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
269 0 : initial_leadership_status: LeadershipStatus,
270 0 : ) -> Self {
271 0 : metrics::update_leadership_status(initial_leadership_status);
272 0 :
273 0 : Self {
274 0 : leadership_status: initial_leadership_status,
275 0 : tenants,
276 0 : nodes: Arc::new(nodes),
277 0 : scheduler,
278 0 : ongoing_operation: None,
279 0 : delayed_reconcile_rx,
280 0 : }
281 0 : }
282 :
283 0 : fn parts_mut(
284 0 : &mut self,
285 0 : ) -> (
286 0 : &mut Arc<HashMap<NodeId, Node>>,
287 0 : &mut BTreeMap<TenantShardId, TenantShard>,
288 0 : &mut Scheduler,
289 0 : ) {
290 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
291 0 : }
292 :
293 0 : fn get_leadership_status(&self) -> LeadershipStatus {
294 0 : self.leadership_status
295 0 : }
296 :
297 0 : fn step_down(&mut self) {
298 0 : self.leadership_status = LeadershipStatus::SteppedDown;
299 0 : metrics::update_leadership_status(self.leadership_status);
300 0 : }
301 :
302 0 : fn become_leader(&mut self) {
303 0 : self.leadership_status = LeadershipStatus::Leader;
304 0 : metrics::update_leadership_status(self.leadership_status);
305 0 : }
306 : }
307 :
308 : #[derive(Clone)]
309 : pub struct Config {
310 : // All pageservers managed by one instance of this service must have
311 : // the same public key. This JWT token will be used to authenticate
312 : // this service to the pageservers it manages.
313 : pub jwt_token: Option<String>,
314 :
315 : // This JWT token will be used to authenticate this service to the control plane.
316 : pub control_plane_jwt_token: Option<String>,
317 :
318 : // This JWT token will be used to authenticate with other storage controller instances
319 : pub peer_jwt_token: Option<String>,
320 :
321 : /// Where the compute hook should send notifications of pageserver attachment locations
322 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
323 : /// assume it is running in a test environment and try to update neon_local.
324 : pub compute_hook_url: Option<String>,
325 :
326 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
327 : /// considered active. Once the grace period elapses, the next heartbeat failure will
328 : /// mark the pagseserver offline.
329 : pub max_offline_interval: Duration,
330 :
331 : /// Extended grace period within which pageserver may not respond to heartbeats.
332 : /// This extended grace period kicks in after the node has been drained for restart
333 : /// and/or upon handling the re-attach request from a node.
334 : pub max_warming_up_interval: Duration,
335 :
336 : /// How many Reconcilers may be spawned concurrently
337 : pub reconciler_concurrency: usize,
338 :
339 : /// How large must a shard grow in bytes before we split it?
340 : /// None disables auto-splitting.
341 : pub split_threshold: Option<u64>,
342 :
343 : // TODO: make this cfg(feature = "testing")
344 : pub neon_local_repo_dir: Option<PathBuf>,
345 :
346 : // Maximum acceptable download lag for the secondary location
347 : // while draining a node. If the secondary location is lagging
348 : // by more than the configured amount, then the secondary is not
349 : // upgraded to primary.
350 : pub max_secondary_lag_bytes: Option<u64>,
351 :
352 : pub heartbeat_interval: Duration,
353 :
354 : pub address_for_peers: Option<Uri>,
355 :
356 : pub start_as_candidate: bool,
357 :
358 : pub http_service_port: i32,
359 :
360 : pub long_reconcile_threshold: Duration,
361 : }
362 :
363 : impl From<DatabaseError> for ApiError {
364 0 : fn from(err: DatabaseError) -> ApiError {
365 0 : match err {
366 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
367 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
368 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
369 0 : ApiError::ShuttingDown
370 : }
371 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
372 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
373 : }
374 : }
375 0 : }
376 : }
377 :
378 : enum InitialShardScheduleOutcome {
379 : Scheduled(TenantCreateResponseShard),
380 : NotScheduled,
381 : ShardScheduleError(ScheduleError),
382 : }
383 :
384 : pub struct Service {
385 : inner: Arc<std::sync::RwLock<ServiceState>>,
386 : config: Config,
387 : persistence: Arc<Persistence>,
388 : compute_hook: Arc<ComputeHook>,
389 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
390 :
391 : heartbeater: Heartbeater,
392 :
393 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
394 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
395 :
396 : // Locking on a tenant granularity (covers all shards in the tenant):
397 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
398 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
399 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
400 :
401 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
402 : // that transition it to/from Active.
403 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
404 :
405 : // Limit how many Reconcilers we will spawn concurrently
406 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
407 :
408 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
409 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
410 : ///
411 : /// Note that this state logically lives inside ServiceInner, but carrying Sender here makes the code simpler
412 : /// by avoiding needing a &mut ref to something inside the ServiceInner. This could be optimized to
413 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
414 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
415 :
416 : // Process shutdown will fire this token
417 : cancel: CancellationToken,
418 :
419 : // Child token of [`Service::cancel`] used by reconcilers
420 : reconcilers_cancel: CancellationToken,
421 :
422 : // Background tasks will hold this gate
423 : gate: Gate,
424 :
425 : // Reconcilers background tasks will hold this gate
426 : reconcilers_gate: Gate,
427 :
428 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
429 : /// passes, it isn't safe to do any actions that mutate tenants.
430 : pub(crate) startup_complete: Barrier,
431 : }
432 :
433 : impl From<ReconcileWaitError> for ApiError {
434 0 : fn from(value: ReconcileWaitError) -> Self {
435 0 : match value {
436 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
437 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
438 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
439 : }
440 0 : }
441 : }
442 :
443 : impl From<OperationError> for ApiError {
444 0 : fn from(value: OperationError) -> Self {
445 0 : match value {
446 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
447 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
448 : }
449 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
450 : }
451 0 : }
452 : }
453 :
454 : #[allow(clippy::large_enum_variant)]
455 : enum TenantCreateOrUpdate {
456 : Create(TenantCreateRequest),
457 : Update(Vec<ShardUpdate>),
458 : }
459 :
460 : struct ShardSplitParams {
461 : old_shard_count: ShardCount,
462 : new_shard_count: ShardCount,
463 : new_stripe_size: Option<ShardStripeSize>,
464 : targets: Vec<ShardSplitTarget>,
465 : policy: PlacementPolicy,
466 : config: TenantConfig,
467 : shard_ident: ShardIdentity,
468 : }
469 :
470 : // When preparing for a shard split, we may either choose to proceed with the split,
471 : // or find that the work is already done and return NoOp.
472 : enum ShardSplitAction {
473 : Split(Box<ShardSplitParams>),
474 : NoOp(TenantShardSplitResponse),
475 : }
476 :
477 : // A parent shard which will be split
478 : struct ShardSplitTarget {
479 : parent_id: TenantShardId,
480 : node: Node,
481 : child_ids: Vec<TenantShardId>,
482 : }
483 :
484 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
485 : /// might not be available. We therefore use a queue of abort operations processed in the background.
486 : struct TenantShardSplitAbort {
487 : tenant_id: TenantId,
488 : /// The target values from the request that failed
489 : new_shard_count: ShardCount,
490 : new_stripe_size: Option<ShardStripeSize>,
491 : /// Until this abort op is complete, no other operations may be done on the tenant
492 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
493 : }
494 :
495 0 : #[derive(thiserror::Error, Debug)]
496 : enum TenantShardSplitAbortError {
497 : #[error(transparent)]
498 : Database(#[from] DatabaseError),
499 : #[error(transparent)]
500 : Remote(#[from] mgmt_api::Error),
501 : #[error("Unavailable")]
502 : Unavailable,
503 : }
504 :
505 : struct ShardUpdate {
506 : tenant_shard_id: TenantShardId,
507 : placement_policy: PlacementPolicy,
508 : tenant_config: TenantConfig,
509 :
510 : /// If this is None, generation is not updated.
511 : generation: Option<Generation>,
512 : }
513 :
514 : enum StopReconciliationsReason {
515 : ShuttingDown,
516 : SteppingDown,
517 : }
518 :
519 : impl std::fmt::Display for StopReconciliationsReason {
520 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
521 0 : let s = match self {
522 0 : Self::ShuttingDown => "Shutting down",
523 0 : Self::SteppingDown => "Stepping down",
524 : };
525 0 : write!(writer, "{}", s)
526 0 : }
527 : }
528 :
529 : pub(crate) enum ReconcileResultRequest {
530 : ReconcileResult(ReconcileResult),
531 : Stop,
532 : }
533 :
534 : #[derive(Clone)]
535 : struct MutationLocation {
536 : node: Node,
537 : generation: Generation,
538 : }
539 :
540 : #[derive(Clone)]
541 : struct ShardMutationLocations {
542 : latest: MutationLocation,
543 : other: Vec<MutationLocation>,
544 : }
545 :
546 : #[derive(Default, Clone)]
547 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
548 :
549 : impl Service {
550 0 : pub fn get_config(&self) -> &Config {
551 0 : &self.config
552 0 : }
553 :
554 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
555 : /// view of the world, and determine which pageservers are responsive.
556 0 : #[instrument(skip_all)]
557 : async fn startup_reconcile(
558 : self: &Arc<Service>,
559 : current_leader: Option<ControllerPersistence>,
560 : leader_step_down_state: Option<GlobalObservedState>,
561 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
562 : Result<(), (TenantShardId, NotifyError)>,
563 : >,
564 : ) {
565 : // Startup reconciliation does I/O to other services: whether they
566 : // are responsive or not, we should aim to finish within our deadline, because:
567 : // - If we don't, a k8s readiness hook watching /ready will kill us.
568 : // - While we're waiting for startup reconciliation, we are not fully
569 : // available for end user operations like creating/deleting tenants and timelines.
570 : //
571 : // We set multiple deadlines to break up the time available between the phases of work: this is
572 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
573 : let start_at = Instant::now();
574 : let node_scan_deadline = start_at
575 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
576 : .expect("Reconcile timeout is a modest constant");
577 :
578 : let observed = if let Some(state) = leader_step_down_state {
579 : tracing::info!(
580 : "Using observed state received from leader at {}",
581 : current_leader.as_ref().unwrap().address
582 : );
583 :
584 : state
585 : } else {
586 : self.build_global_observed_state(node_scan_deadline).await
587 : };
588 :
589 : // Accumulate a list of any tenant locations that ought to be detached
590 : let mut cleanup = Vec::new();
591 :
592 : // Send initial heartbeat requests to all nodes loaded from the database
593 : let all_nodes = {
594 : let locked = self.inner.read().unwrap();
595 : locked.nodes.clone()
596 : };
597 : let mut nodes_online = self.initial_heartbeat_round(all_nodes.keys()).await;
598 :
599 : // List of tenants for which we will attempt to notify compute of their location at startup
600 : let mut compute_notifications = Vec::new();
601 :
602 : // Populate intent and observed states for all tenants, based on reported state on pageservers
603 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
604 : let shard_count = {
605 : let mut locked = self.inner.write().unwrap();
606 : let (nodes, tenants, scheduler) = locked.parts_mut();
607 :
608 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
609 : let mut new_nodes = (**nodes).clone();
610 : for (node_id, node) in new_nodes.iter_mut() {
611 : if let Some(utilization) = nodes_online.remove(node_id) {
612 : node.set_availability(NodeAvailability::Active(utilization));
613 : scheduler.node_upsert(node);
614 : }
615 : }
616 : *nodes = Arc::new(new_nodes);
617 :
618 : for (tenant_shard_id, observed_state) in observed.0 {
619 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
620 : for node_id in observed_state.locations.keys() {
621 : cleanup.push((tenant_shard_id, *node_id));
622 : }
623 :
624 : continue;
625 : };
626 :
627 : tenant_shard.observed = observed_state;
628 : }
629 :
630 : // Populate each tenant's intent state
631 : let mut schedule_context = ScheduleContext::default();
632 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
633 : if tenant_shard_id.shard_number == ShardNumber(0) {
634 : // Reset scheduling context each time we advance to the next Tenant
635 : schedule_context = ScheduleContext::default();
636 : }
637 :
638 : tenant_shard.intent_from_observed(scheduler);
639 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
640 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
641 : // not enough pageservers are available. The tenant may well still be available
642 : // to clients.
643 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
644 : } else {
645 : // If we're both intending and observed to be attached at a particular node, we will
646 : // emit a compute notification for this. In the case where our observed state does not
647 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
648 : if let Some(attached_at) = tenant_shard.stably_attached() {
649 : compute_notifications.push((
650 : *tenant_shard_id,
651 : attached_at,
652 : tenant_shard.shard.stripe_size,
653 : ));
654 : }
655 : }
656 : }
657 :
658 : tenants.len()
659 : };
660 :
661 : // Before making any obeservable changes to the cluster, persist self
662 : // as leader in database and memory.
663 : let leadership = Leadership::new(
664 : self.persistence.clone(),
665 : self.config.clone(),
666 : self.cancel.child_token(),
667 : );
668 :
669 : if let Err(e) = leadership.become_leader(current_leader).await {
670 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
671 : std::process::exit(1);
672 : }
673 :
674 : self.inner.write().unwrap().become_leader();
675 :
676 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
677 : // generation_pageserver in the database.
678 :
679 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
680 : // will emit compute hook notifications when they reconcile.
681 : //
682 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
683 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
684 : // calls will be correctly ordered wrt these.
685 : //
686 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
687 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
688 : // unit and start doing I/O.
689 : tracing::info!(
690 : "Sending {} compute notifications",
691 : compute_notifications.len()
692 : );
693 : self.compute_hook.notify_background(
694 : compute_notifications,
695 : bg_compute_notify_result_tx.clone(),
696 : &self.cancel,
697 : );
698 :
699 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
700 : // which require it: under normal circumstances this should only include tenants that were in some
701 : // transient state before we restarted, or any tenants whose compute hooks failed above.
702 : tracing::info!("Checking for shards in need of reconciliation...");
703 : let reconcile_tasks = self.reconcile_all();
704 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
705 : // normal operations may proceed.
706 :
707 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
708 : // background because it does not need to complete in order to proceed with other work.
709 : if !cleanup.is_empty() {
710 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
711 : tokio::task::spawn({
712 : let cleanup_self = self.clone();
713 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
714 : });
715 : }
716 :
717 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
718 : }
719 :
720 0 : async fn initial_heartbeat_round<'a>(
721 0 : &self,
722 0 : node_ids: impl Iterator<Item = &'a NodeId>,
723 0 : ) -> HashMap<NodeId, PageserverUtilization> {
724 0 : assert!(!self.startup_complete.is_ready());
725 :
726 0 : let all_nodes = {
727 0 : let locked = self.inner.read().unwrap();
728 0 : locked.nodes.clone()
729 0 : };
730 0 :
731 0 : let mut nodes_to_heartbeat = HashMap::new();
732 0 : for node_id in node_ids {
733 0 : match all_nodes.get(node_id) {
734 0 : Some(node) => {
735 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
736 0 : }
737 : None => {
738 0 : tracing::warn!("Node {node_id} was removed during start-up");
739 : }
740 : }
741 : }
742 :
743 0 : tracing::info!("Sending initial heartbeats...");
744 0 : let res = self
745 0 : .heartbeater
746 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
747 0 : .await;
748 :
749 0 : let mut online_nodes = HashMap::new();
750 0 : if let Ok(deltas) = res {
751 0 : for (node_id, status) in deltas.0 {
752 0 : match status {
753 0 : PageserverState::Available { utilization, .. } => {
754 0 : online_nodes.insert(node_id, utilization);
755 0 : }
756 0 : PageserverState::Offline => {}
757 : PageserverState::WarmingUp { .. } => {
758 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
759 : }
760 : }
761 : }
762 0 : }
763 :
764 0 : online_nodes
765 0 : }
766 :
767 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
768 : ///
769 : /// The result includes only nodes which responded within the deadline
770 0 : async fn scan_node_locations(
771 0 : &self,
772 0 : deadline: Instant,
773 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
774 0 : let nodes = {
775 0 : let locked = self.inner.read().unwrap();
776 0 : locked.nodes.clone()
777 0 : };
778 0 :
779 0 : let mut node_results = HashMap::new();
780 0 :
781 0 : let mut node_list_futs = FuturesUnordered::new();
782 0 :
783 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
784 0 : for node in nodes.values() {
785 0 : node_list_futs.push({
786 0 : async move {
787 0 : tracing::info!("Scanning shards on node {node}...");
788 0 : let timeout = Duration::from_secs(1);
789 0 : let response = node
790 0 : .with_client_retries(
791 0 : |client| async move { client.list_location_config().await },
792 0 : &self.config.jwt_token,
793 0 : 1,
794 0 : 5,
795 0 : timeout,
796 0 : &self.cancel,
797 0 : )
798 0 : .await;
799 0 : (node.get_id(), response)
800 0 : }
801 0 : });
802 0 : }
803 :
804 : loop {
805 0 : let (node_id, result) = tokio::select! {
806 0 : next = node_list_futs.next() => {
807 0 : match next {
808 0 : Some(result) => result,
809 : None =>{
810 : // We got results for all our nodes
811 0 : break;
812 : }
813 :
814 : }
815 : },
816 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
817 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
818 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
819 0 : break;
820 : }
821 : };
822 :
823 0 : let Some(list_response) = result else {
824 0 : tracing::info!("Shutdown during startup_reconcile");
825 0 : break;
826 : };
827 :
828 0 : match list_response {
829 0 : Err(e) => {
830 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
831 : }
832 0 : Ok(listing) => {
833 0 : node_results.insert(node_id, listing);
834 0 : }
835 : }
836 : }
837 :
838 0 : node_results
839 0 : }
840 :
841 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
842 0 : let node_listings = self.scan_node_locations(deadline).await;
843 0 : let mut observed = GlobalObservedState::default();
844 :
845 0 : for (node_id, location_confs) in node_listings {
846 0 : tracing::info!(
847 0 : "Received {} shard statuses from pageserver {}",
848 0 : location_confs.tenant_shards.len(),
849 : node_id
850 : );
851 :
852 0 : for (tid, location_conf) in location_confs.tenant_shards {
853 0 : let entry = observed.0.entry(tid).or_default();
854 0 : entry.locations.insert(
855 0 : node_id,
856 0 : ObservedStateLocation {
857 0 : conf: location_conf,
858 0 : },
859 0 : );
860 0 : }
861 : }
862 :
863 0 : observed
864 0 : }
865 :
866 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
867 : ///
868 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
869 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
870 : /// other task trying to attach it.
871 0 : #[instrument(skip_all)]
872 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
873 : let nodes = self.inner.read().unwrap().nodes.clone();
874 :
875 : for (tenant_shard_id, node_id) in cleanup {
876 : // A node reported a tenant_shard_id which is unknown to us: detach it.
877 : let Some(node) = nodes.get(&node_id) else {
878 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
879 : // a location to clean up on a node that has since been removed.
880 : tracing::info!(
881 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
882 : );
883 : continue;
884 : };
885 :
886 : if self.cancel.is_cancelled() {
887 : break;
888 : }
889 :
890 : let client = PageserverClient::new(
891 : node.get_id(),
892 : node.base_url(),
893 : self.config.jwt_token.as_deref(),
894 : );
895 : match client
896 : .location_config(
897 : tenant_shard_id,
898 : LocationConfig {
899 : mode: LocationConfigMode::Detached,
900 : generation: None,
901 : secondary_conf: None,
902 : shard_number: tenant_shard_id.shard_number.0,
903 : shard_count: tenant_shard_id.shard_count.literal(),
904 : shard_stripe_size: 0,
905 : tenant_conf: models::TenantConfig::default(),
906 : },
907 : None,
908 : false,
909 : )
910 : .await
911 : {
912 : Ok(()) => {
913 : tracing::info!(
914 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
915 : );
916 : }
917 : Err(e) => {
918 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
919 : // break anything.
920 : tracing::error!(
921 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
922 : );
923 : }
924 : }
925 : }
926 : }
927 :
928 : /// Long running background task that periodically wakes up and looks for shards that need
929 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
930 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
931 : /// for those retries.
932 0 : #[instrument(skip_all)]
933 : async fn background_reconcile(self: &Arc<Self>) {
934 : self.startup_complete.clone().wait().await;
935 :
936 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
937 :
938 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
939 : while !self.reconcilers_cancel.is_cancelled() {
940 : tokio::select! {
941 : _ = interval.tick() => {
942 : let reconciles_spawned = self.reconcile_all();
943 : if reconciles_spawned == 0 {
944 : // Run optimizer only when we didn't find any other work to do
945 : let optimizations = self.optimize_all().await;
946 : if optimizations == 0 {
947 : // Run new splits only when no optimizations are pending
948 : self.autosplit_tenants().await;
949 : }
950 : }
951 : }
952 : _ = self.reconcilers_cancel.cancelled() => return
953 : }
954 : }
955 : }
956 0 : #[instrument(skip_all)]
957 : async fn spawn_heartbeat_driver(&self) {
958 : self.startup_complete.clone().wait().await;
959 :
960 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
961 : while !self.cancel.is_cancelled() {
962 : tokio::select! {
963 : _ = interval.tick() => { }
964 : _ = self.cancel.cancelled() => return
965 : };
966 :
967 : let nodes = {
968 : let locked = self.inner.read().unwrap();
969 : locked.nodes.clone()
970 : };
971 :
972 : let res = self.heartbeater.heartbeat(nodes).await;
973 : if let Ok(deltas) = res {
974 : let mut to_handle = Vec::default();
975 :
976 : for (node_id, state) in deltas.0 {
977 : let new_availability = match state {
978 : PageserverState::Available { utilization, .. } => {
979 : NodeAvailability::Active(utilization)
980 : }
981 : PageserverState::WarmingUp { started_at } => {
982 : NodeAvailability::WarmingUp(started_at)
983 : }
984 : PageserverState::Offline => {
985 : // The node might have been placed in the WarmingUp state
986 : // while the heartbeat round was on-going. Hence, filter out
987 : // offline transitions for WarmingUp nodes that are still within
988 : // their grace period.
989 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
990 : .get_node(node_id)
991 : .await
992 : .as_ref()
993 0 : .map(|n| n.get_availability())
994 : {
995 : let now = Instant::now();
996 : if now - *started_at >= self.config.max_warming_up_interval {
997 : NodeAvailability::Offline
998 : } else {
999 : NodeAvailability::WarmingUp(*started_at)
1000 : }
1001 : } else {
1002 : NodeAvailability::Offline
1003 : }
1004 : }
1005 : };
1006 :
1007 : let node_lock = trace_exclusive_lock(
1008 : &self.node_op_locks,
1009 : node_id,
1010 : NodeOperations::Configure,
1011 : )
1012 : .await;
1013 :
1014 : // This is the code path for geniune availability transitions (i.e node
1015 : // goes unavailable and/or comes back online).
1016 : let res = self
1017 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1018 : .await;
1019 :
1020 : match res {
1021 : Ok(transition) => {
1022 : // Keep hold of the lock until the availability transitions
1023 : // have been handled in
1024 : // [`Service::handle_node_availability_transitions`] in order avoid
1025 : // racing with [`Service::external_node_configure`].
1026 : to_handle.push((node_id, node_lock, transition));
1027 : }
1028 : Err(ApiError::NotFound(_)) => {
1029 : // This should be rare, but legitimate since the heartbeats are done
1030 : // on a snapshot of the nodes.
1031 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1032 : }
1033 : Err(err) => {
1034 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1035 : // becomes unavailable again, we may get an error here.
1036 : tracing::error!(
1037 : "Failed to update node state {} after heartbeat round: {}",
1038 : node_id,
1039 : err
1040 : );
1041 : }
1042 : }
1043 : }
1044 :
1045 : // We collected all the transitions above and now we handle them.
1046 : let res = self.handle_node_availability_transitions(to_handle).await;
1047 : if let Err(errs) = res {
1048 : for (node_id, err) in errs {
1049 : match err {
1050 : ApiError::NotFound(_) => {
1051 : // This should be rare, but legitimate since the heartbeats are done
1052 : // on a snapshot of the nodes.
1053 : tracing::info!(
1054 : "Node {} was not found after heartbeat round",
1055 : node_id
1056 : );
1057 : }
1058 : err => {
1059 : tracing::error!(
1060 : "Failed to handle availability transition for {} after heartbeat round: {}",
1061 : node_id,
1062 : err
1063 : );
1064 : }
1065 : }
1066 : }
1067 : }
1068 : }
1069 : }
1070 : }
1071 :
1072 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1073 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1074 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1075 : /// will indicate that reconciliation is not needed.
1076 0 : #[instrument(skip_all, fields(
1077 : seq=%result.sequence,
1078 : tenant_id=%result.tenant_shard_id.tenant_id,
1079 : shard_id=%result.tenant_shard_id.shard_slug(),
1080 0 : ))]
1081 : fn process_result(&self, result: ReconcileResult) {
1082 : let mut locked = self.inner.write().unwrap();
1083 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1084 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1085 : // A reconciliation result might race with removing a tenant: drop results for
1086 : // tenants that aren't in our map.
1087 : return;
1088 : };
1089 :
1090 : // Usually generation should only be updated via this path, so the max() isn't
1091 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1092 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1093 :
1094 : // If the reconciler signals that it failed to notify compute, set this state on
1095 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1096 : tenant.pending_compute_notification = result.pending_compute_notification;
1097 :
1098 : // Let the TenantShard know it is idle.
1099 : tenant.reconcile_complete(result.sequence);
1100 :
1101 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1102 : // make to the tenant
1103 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1104 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1105 : // make to the tenant
1106 0 : let node = nodes.get(delta.node_id())?;
1107 :
1108 0 : if node.is_available() {
1109 0 : return Some(delta);
1110 0 : }
1111 0 :
1112 0 : // In case a node became unavailable concurrently with the reconcile, observed
1113 0 : // locations on it are now uncertain. By convention, set them to None in order
1114 0 : // for them to get refreshed when the node comes back online.
1115 0 : Some(ObservedStateDelta::Upsert(Box::new((
1116 0 : node.get_id(),
1117 0 : ObservedStateLocation { conf: None },
1118 0 : ))))
1119 0 : });
1120 :
1121 : match result.result {
1122 : Ok(()) => {
1123 : tenant.apply_observed_deltas(deltas);
1124 : tenant.waiter.advance(result.sequence);
1125 : }
1126 : Err(e) => {
1127 : match e {
1128 : ReconcileError::Cancel => {
1129 : tracing::info!("Reconciler was cancelled");
1130 : }
1131 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1132 : // This might be due to the reconciler getting cancelled, or it might
1133 : // be due to the `Node` being marked offline.
1134 : tracing::info!("Reconciler cancelled during pageserver API call");
1135 : }
1136 : _ => {
1137 : tracing::warn!("Reconcile error: {}", e);
1138 : }
1139 : }
1140 :
1141 : // Ordering: populate last_error before advancing error_seq,
1142 : // so that waiters will see the correct error after waiting.
1143 : tenant.set_last_error(result.sequence, e);
1144 :
1145 : // Skip deletions on reconcile failures
1146 : let upsert_deltas =
1147 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1148 : tenant.apply_observed_deltas(upsert_deltas);
1149 : }
1150 : }
1151 :
1152 : // Maybe some other work can proceed now that this job finished.
1153 : if self.reconciler_concurrency.available_permits() > 0 {
1154 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1155 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1156 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1157 : shard.delayed_reconcile = false;
1158 : self.maybe_reconcile_shard(shard, nodes);
1159 : }
1160 :
1161 : if self.reconciler_concurrency.available_permits() == 0 {
1162 : break;
1163 : }
1164 : }
1165 : }
1166 : }
1167 :
1168 0 : async fn process_results(
1169 0 : &self,
1170 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1171 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1172 0 : Result<(), (TenantShardId, NotifyError)>,
1173 0 : >,
1174 0 : ) {
1175 : loop {
1176 : // Wait for the next result, or for cancellation
1177 0 : tokio::select! {
1178 0 : r = result_rx.recv() => {
1179 0 : match r {
1180 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1181 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1182 : }
1183 : }
1184 0 : _ = async{
1185 0 : match bg_compute_hook_result_rx.recv().await {
1186 0 : Some(result) => {
1187 0 : if let Err((tenant_shard_id, notify_error)) = result {
1188 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1189 0 : let mut locked = self.inner.write().unwrap();
1190 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1191 0 : shard.pending_compute_notification = true;
1192 0 : }
1193 :
1194 0 : }
1195 : },
1196 : None => {
1197 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1198 0 : self.cancel.cancelled().await;
1199 : }
1200 : }
1201 0 : } => {},
1202 0 : _ = self.cancel.cancelled() => {
1203 0 : break;
1204 : }
1205 : };
1206 : }
1207 0 : }
1208 :
1209 0 : async fn process_aborts(
1210 0 : &self,
1211 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1212 0 : ) {
1213 : loop {
1214 : // Wait for the next result, or for cancellation
1215 0 : let op = tokio::select! {
1216 0 : r = abort_rx.recv() => {
1217 0 : match r {
1218 0 : Some(op) => {op},
1219 0 : None => {break;}
1220 : }
1221 : }
1222 0 : _ = self.cancel.cancelled() => {
1223 0 : break;
1224 : }
1225 : };
1226 :
1227 : // Retry until shutdown: we must keep this request object alive until it is properly
1228 : // processed, as it holds a lock guard that prevents other operations trying to do things
1229 : // to the tenant while it is in a weird part-split state.
1230 0 : while !self.cancel.is_cancelled() {
1231 0 : match self.abort_tenant_shard_split(&op).await {
1232 0 : Ok(_) => break,
1233 0 : Err(e) => {
1234 0 : tracing::warn!(
1235 0 : "Failed to abort shard split on {}, will retry: {e}",
1236 : op.tenant_id
1237 : );
1238 :
1239 : // If a node is unavailable, we hope that it has been properly marked Offline
1240 : // when we retry, so that the abort op will succeed. If the abort op is failing
1241 : // for some other reason, we will keep retrying forever, or until a human notices
1242 : // and does something about it (either fixing a pageserver or restarting the controller).
1243 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1244 0 : .await
1245 0 : .ok();
1246 : }
1247 : }
1248 : }
1249 : }
1250 0 : }
1251 :
1252 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1253 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1254 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1255 0 :
1256 0 : let leadership_cancel = CancellationToken::new();
1257 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1258 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1259 :
1260 : // Apply the migrations **after** the current leader has stepped down
1261 : // (or we've given up waiting for it), but **before** reading from the
1262 : // database. The only exception is reading the current leader before
1263 : // migrating.
1264 0 : persistence.migration_run().await?;
1265 :
1266 0 : tracing::info!("Loading nodes from database...");
1267 0 : let nodes = persistence
1268 0 : .list_nodes()
1269 0 : .await?
1270 0 : .into_iter()
1271 0 : .map(Node::from_persistent)
1272 0 : .collect::<Vec<_>>();
1273 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1274 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1275 :
1276 0 : tracing::info!("Loading shards from database...");
1277 0 : let mut tenant_shard_persistence = persistence.list_tenant_shards().await?;
1278 0 : tracing::info!(
1279 0 : "Loaded {} shards from database.",
1280 0 : tenant_shard_persistence.len()
1281 : );
1282 :
1283 : // If any shard splits were in progress, reset the database state to abort them
1284 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1285 0 : HashMap::new();
1286 0 : for tsp in &mut tenant_shard_persistence {
1287 0 : let shard = tsp.get_shard_identity()?;
1288 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1289 0 : let entry = tenant_shard_count_min_max
1290 0 : .entry(tenant_shard_id.tenant_id)
1291 0 : .or_insert_with(|| (shard.count, shard.count));
1292 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1293 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1294 0 : }
1295 :
1296 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1297 0 : if count_min != count_max {
1298 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1299 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1300 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1301 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1302 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1303 :
1304 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1305 : // identified this tenant has having mismatching min/max counts.
1306 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1307 :
1308 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1309 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1310 0 : // Set idle split state on those shards that we will retain.
1311 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1312 0 : if tsp_tenant_id == tenant_id
1313 0 : && tsp.get_shard_identity().unwrap().count == count_min
1314 0 : {
1315 0 : tsp.splitting = SplitState::Idle;
1316 0 : } else if tsp_tenant_id == tenant_id {
1317 : // Leave the splitting state on the child shards: this will be used next to
1318 : // drop them.
1319 0 : tracing::info!(
1320 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1321 : );
1322 0 : }
1323 0 : });
1324 0 :
1325 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1326 0 : tenant_shard_persistence.retain(|tsp| {
1327 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1328 0 : || tsp.splitting == SplitState::Idle
1329 0 : });
1330 0 : }
1331 : }
1332 :
1333 0 : let mut tenants = BTreeMap::new();
1334 0 :
1335 0 : let mut scheduler = Scheduler::new(nodes.values());
1336 :
1337 : #[cfg(feature = "testing")]
1338 : {
1339 : use pageserver_api::controller_api::AvailabilityZone;
1340 :
1341 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1342 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1343 : // after when pageservers start up and register.
1344 0 : let mut node_ids = HashSet::new();
1345 0 : for tsp in &tenant_shard_persistence {
1346 0 : if let Some(node_id) = tsp.generation_pageserver {
1347 0 : node_ids.insert(node_id);
1348 0 : }
1349 : }
1350 0 : for node_id in node_ids {
1351 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1352 0 : let node = Node::new(
1353 0 : NodeId(node_id as u64),
1354 0 : "".to_string(),
1355 0 : 123,
1356 0 : "".to_string(),
1357 0 : 123,
1358 0 : AvailabilityZone("test_az".to_string()),
1359 0 : );
1360 0 :
1361 0 : scheduler.node_upsert(&node);
1362 : }
1363 : }
1364 0 : for tsp in tenant_shard_persistence {
1365 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1366 :
1367 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1368 : // it with what we can infer: the node for which a generation was most recently issued.
1369 0 : let mut intent = IntentState::new();
1370 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1371 : {
1372 0 : if nodes.contains_key(&generation_pageserver) {
1373 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1374 0 : } else {
1375 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1376 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1377 : // on different pageservers.
1378 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1379 : }
1380 0 : }
1381 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1382 :
1383 0 : tenants.insert(tenant_shard_id, new_tenant);
1384 : }
1385 :
1386 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1387 0 :
1388 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1389 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1390 0 : tokio::sync::mpsc::channel(512);
1391 0 :
1392 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1393 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1394 0 :
1395 0 : let cancel = CancellationToken::new();
1396 0 : let reconcilers_cancel = cancel.child_token();
1397 0 :
1398 0 : let heartbeater = Heartbeater::new(
1399 0 : config.jwt_token.clone(),
1400 0 : config.max_offline_interval,
1401 0 : config.max_warming_up_interval,
1402 0 : cancel.clone(),
1403 0 : );
1404 :
1405 0 : let initial_leadership_status = if config.start_as_candidate {
1406 0 : LeadershipStatus::Candidate
1407 : } else {
1408 0 : LeadershipStatus::Leader
1409 : };
1410 :
1411 0 : let this = Arc::new(Self {
1412 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1413 0 : nodes,
1414 0 : tenants,
1415 0 : scheduler,
1416 0 : delayed_reconcile_rx,
1417 0 : initial_leadership_status,
1418 0 : ))),
1419 0 : config: config.clone(),
1420 0 : persistence,
1421 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1422 0 : result_tx,
1423 0 : heartbeater,
1424 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1425 0 : config.reconciler_concurrency,
1426 0 : )),
1427 0 : delayed_reconcile_tx,
1428 0 : abort_tx,
1429 0 : startup_complete: startup_complete.clone(),
1430 0 : cancel,
1431 0 : reconcilers_cancel,
1432 0 : gate: Gate::default(),
1433 0 : reconcilers_gate: Gate::default(),
1434 0 : tenant_op_locks: Default::default(),
1435 0 : node_op_locks: Default::default(),
1436 0 : });
1437 0 :
1438 0 : let result_task_this = this.clone();
1439 0 : tokio::task::spawn(async move {
1440 : // Block shutdown until we're done (we must respect self.cancel)
1441 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1442 0 : result_task_this
1443 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1444 0 : .await
1445 0 : }
1446 0 : });
1447 0 :
1448 0 : tokio::task::spawn({
1449 0 : let this = this.clone();
1450 0 : async move {
1451 : // Block shutdown until we're done (we must respect self.cancel)
1452 0 : if let Ok(_gate) = this.gate.enter() {
1453 0 : this.process_aborts(abort_rx).await
1454 0 : }
1455 0 : }
1456 0 : });
1457 0 :
1458 0 : tokio::task::spawn({
1459 0 : let this = this.clone();
1460 0 : async move {
1461 0 : if let Ok(_gate) = this.gate.enter() {
1462 : loop {
1463 0 : tokio::select! {
1464 0 : _ = this.cancel.cancelled() => {
1465 0 : break;
1466 : },
1467 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1468 0 : };
1469 0 : this.tenant_op_locks.housekeeping();
1470 : }
1471 0 : }
1472 0 : }
1473 0 : });
1474 0 :
1475 0 : tokio::task::spawn({
1476 0 : let this = this.clone();
1477 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1478 0 : // is done.
1479 0 : let startup_completion = startup_completion.clone();
1480 0 : async move {
1481 : // Block shutdown until we're done (we must respect self.cancel)
1482 0 : let Ok(_gate) = this.gate.enter() else {
1483 0 : return;
1484 : };
1485 :
1486 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1487 0 : .await;
1488 :
1489 0 : drop(startup_completion);
1490 0 : }
1491 0 : });
1492 0 :
1493 0 : tokio::task::spawn({
1494 0 : let this = this.clone();
1495 0 : let startup_complete = startup_complete.clone();
1496 0 : async move {
1497 0 : startup_complete.wait().await;
1498 0 : this.background_reconcile().await;
1499 0 : }
1500 0 : });
1501 0 :
1502 0 : tokio::task::spawn({
1503 0 : let this = this.clone();
1504 0 : let startup_complete = startup_complete.clone();
1505 0 : async move {
1506 0 : startup_complete.wait().await;
1507 0 : this.spawn_heartbeat_driver().await;
1508 0 : }
1509 0 : });
1510 0 :
1511 0 : Ok(this)
1512 0 : }
1513 :
1514 0 : pub(crate) async fn attach_hook(
1515 0 : &self,
1516 0 : attach_req: AttachHookRequest,
1517 0 : ) -> anyhow::Result<AttachHookResponse> {
1518 0 : let _tenant_lock = trace_exclusive_lock(
1519 0 : &self.tenant_op_locks,
1520 0 : attach_req.tenant_shard_id.tenant_id,
1521 0 : TenantOperations::AttachHook,
1522 0 : )
1523 0 : .await;
1524 :
1525 : // This is a test hook. To enable using it on tenants that were created directly with
1526 : // the pageserver API (not via this service), we will auto-create any missing tenant
1527 : // shards with default state.
1528 0 : let insert = {
1529 0 : let locked = self.inner.write().unwrap();
1530 0 : !locked.tenants.contains_key(&attach_req.tenant_shard_id)
1531 0 : };
1532 0 :
1533 0 : if insert {
1534 0 : let tsp = TenantShardPersistence {
1535 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1536 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1537 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1538 0 : shard_stripe_size: 0,
1539 0 : generation: attach_req.generation_override.or(Some(0)),
1540 0 : generation_pageserver: None,
1541 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1542 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1543 0 : splitting: SplitState::default(),
1544 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1545 0 : .unwrap(),
1546 0 : preferred_az_id: None,
1547 0 : };
1548 0 :
1549 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1550 0 : Err(e) => match e {
1551 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1552 : DatabaseErrorKind::UniqueViolation,
1553 : _,
1554 : )) => {
1555 0 : tracing::info!(
1556 0 : "Raced with another request to insert tenant {}",
1557 : attach_req.tenant_shard_id
1558 : )
1559 : }
1560 0 : _ => return Err(e.into()),
1561 : },
1562 : Ok(()) => {
1563 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1564 :
1565 0 : let mut locked = self.inner.write().unwrap();
1566 0 : locked.tenants.insert(
1567 0 : attach_req.tenant_shard_id,
1568 0 : TenantShard::new(
1569 0 : attach_req.tenant_shard_id,
1570 0 : ShardIdentity::unsharded(),
1571 0 : PlacementPolicy::Attached(0),
1572 0 : ),
1573 0 : );
1574 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1575 : }
1576 : }
1577 0 : }
1578 :
1579 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1580 0 : let maybe_tenant_conf = {
1581 0 : let locked = self.inner.write().unwrap();
1582 0 : locked
1583 0 : .tenants
1584 0 : .get(&attach_req.tenant_shard_id)
1585 0 : .map(|t| t.config.clone())
1586 0 : };
1587 0 :
1588 0 : match maybe_tenant_conf {
1589 0 : Some(conf) => {
1590 0 : let new_generation = self
1591 0 : .persistence
1592 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1593 0 : .await?;
1594 :
1595 : // Persist the placement policy update. This is required
1596 : // when we reattaching a detached tenant.
1597 0 : self.persistence
1598 0 : .update_tenant_shard(
1599 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1600 0 : Some(PlacementPolicy::Attached(0)),
1601 0 : Some(conf),
1602 0 : None,
1603 0 : None,
1604 0 : )
1605 0 : .await?;
1606 0 : Some(new_generation)
1607 : }
1608 : None => {
1609 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1610 : }
1611 : }
1612 : } else {
1613 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1614 0 : None
1615 : };
1616 :
1617 0 : let mut locked = self.inner.write().unwrap();
1618 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1619 0 :
1620 0 : let tenant_shard = tenants
1621 0 : .get_mut(&attach_req.tenant_shard_id)
1622 0 : .expect("Checked for existence above");
1623 :
1624 0 : if let Some(new_generation) = new_generation {
1625 0 : tenant_shard.generation = Some(new_generation);
1626 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1627 0 : } else {
1628 : // This is a detach notification. We must update placement policy to avoid re-attaching
1629 : // during background scheduling/reconciliation, or during storage controller restart.
1630 0 : assert!(attach_req.node_id.is_none());
1631 0 : tenant_shard.policy = PlacementPolicy::Detached;
1632 : }
1633 :
1634 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1635 0 : tracing::info!(
1636 : tenant_id = %attach_req.tenant_shard_id,
1637 : ps_id = %attaching_pageserver,
1638 : generation = ?tenant_shard.generation,
1639 0 : "issuing",
1640 : );
1641 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1642 0 : tracing::info!(
1643 : tenant_id = %attach_req.tenant_shard_id,
1644 : %ps_id,
1645 : generation = ?tenant_shard.generation,
1646 0 : "dropping",
1647 : );
1648 : } else {
1649 0 : tracing::info!(
1650 : tenant_id = %attach_req.tenant_shard_id,
1651 0 : "no-op: tenant already has no pageserver");
1652 : }
1653 0 : tenant_shard
1654 0 : .intent
1655 0 : .set_attached(scheduler, attach_req.node_id);
1656 0 :
1657 0 : tracing::info!(
1658 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1659 0 : attach_req.tenant_shard_id,
1660 0 : tenant_shard.generation,
1661 0 : // TODO: this is an odd number of 0xf's
1662 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1663 : );
1664 :
1665 : // Trick the reconciler into not doing anything for this tenant: this helps
1666 : // tests that manually configure a tenant on the pagesrever, and then call this
1667 : // attach hook: they don't want background reconciliation to modify what they
1668 : // did to the pageserver.
1669 : #[cfg(feature = "testing")]
1670 : {
1671 0 : if let Some(node_id) = attach_req.node_id {
1672 0 : tenant_shard.observed.locations = HashMap::from([(
1673 0 : node_id,
1674 0 : ObservedStateLocation {
1675 0 : conf: Some(attached_location_conf(
1676 0 : tenant_shard.generation.unwrap(),
1677 0 : &tenant_shard.shard,
1678 0 : &tenant_shard.config,
1679 0 : &PlacementPolicy::Attached(0),
1680 0 : )),
1681 0 : },
1682 0 : )]);
1683 0 : } else {
1684 0 : tenant_shard.observed.locations.clear();
1685 0 : }
1686 : }
1687 :
1688 0 : Ok(AttachHookResponse {
1689 0 : gen: attach_req
1690 0 : .node_id
1691 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1692 0 : })
1693 0 : }
1694 :
1695 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1696 0 : let locked = self.inner.read().unwrap();
1697 0 :
1698 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1699 0 :
1700 0 : InspectResponse {
1701 0 : attachment: tenant_shard.and_then(|s| {
1702 0 : s.intent
1703 0 : .get_attached()
1704 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1705 0 : }),
1706 0 : }
1707 0 : }
1708 :
1709 : // When the availability state of a node transitions to active, we must do a full reconciliation
1710 : // of LocationConfigs on that node. This is because while a node was offline:
1711 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1712 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1713 : //
1714 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1715 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1716 : // this function.
1717 : //
1718 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1719 : // for written for a single node rather than as a batch job for all nodes.
1720 0 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1721 : async fn node_activate_reconcile(
1722 : &self,
1723 : mut node: Node,
1724 : _lock: &TracingExclusiveGuard<NodeOperations>,
1725 : ) -> Result<(), ApiError> {
1726 : // This Node is a mutable local copy: we will set it active so that we can use its
1727 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1728 : // later.
1729 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1730 :
1731 : let configs = match node
1732 : .with_client_retries(
1733 0 : |client| async move { client.list_location_config().await },
1734 : &self.config.jwt_token,
1735 : 1,
1736 : 5,
1737 : SHORT_RECONCILE_TIMEOUT,
1738 : &self.cancel,
1739 : )
1740 : .await
1741 : {
1742 : None => {
1743 : // We're shutting down (the Node's cancellation token can't have fired, because
1744 : // we're the only scope that has a reference to it, and we didn't fire it).
1745 : return Err(ApiError::ShuttingDown);
1746 : }
1747 : Some(Err(e)) => {
1748 : // This node didn't succeed listing its locations: it may not proceed to active state
1749 : // as it is apparently unavailable.
1750 : return Err(ApiError::PreconditionFailed(
1751 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1752 : ));
1753 : }
1754 : Some(Ok(configs)) => configs,
1755 : };
1756 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1757 :
1758 : let mut cleanup = Vec::new();
1759 : {
1760 : let mut locked = self.inner.write().unwrap();
1761 :
1762 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1763 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1764 : cleanup.push(tenant_shard_id);
1765 : continue;
1766 : };
1767 : tenant_shard
1768 : .observed
1769 : .locations
1770 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1771 : }
1772 : }
1773 :
1774 : for tenant_shard_id in cleanup {
1775 : tracing::info!("Detaching {tenant_shard_id}");
1776 : match node
1777 : .with_client_retries(
1778 0 : |client| async move {
1779 0 : let config = LocationConfig {
1780 0 : mode: LocationConfigMode::Detached,
1781 0 : generation: None,
1782 0 : secondary_conf: None,
1783 0 : shard_number: tenant_shard_id.shard_number.0,
1784 0 : shard_count: tenant_shard_id.shard_count.literal(),
1785 0 : shard_stripe_size: 0,
1786 0 : tenant_conf: models::TenantConfig::default(),
1787 0 : };
1788 0 : client
1789 0 : .location_config(tenant_shard_id, config, None, false)
1790 0 : .await
1791 0 : },
1792 : &self.config.jwt_token,
1793 : 1,
1794 : 5,
1795 : SHORT_RECONCILE_TIMEOUT,
1796 : &self.cancel,
1797 : )
1798 : .await
1799 : {
1800 : None => {
1801 : // We're shutting down (the Node's cancellation token can't have fired, because
1802 : // we're the only scope that has a reference to it, and we didn't fire it).
1803 : return Err(ApiError::ShuttingDown);
1804 : }
1805 : Some(Err(e)) => {
1806 : // Do not let the node proceed to Active state if it is not responsive to requests
1807 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1808 : // detach completing: we should not let this node back into the set of nodes considered
1809 : // okay for scheduling.
1810 : return Err(ApiError::Conflict(format!(
1811 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1812 : )));
1813 : }
1814 : Some(Ok(_)) => {}
1815 : };
1816 : }
1817 :
1818 : Ok(())
1819 : }
1820 :
1821 0 : pub(crate) async fn re_attach(
1822 0 : &self,
1823 0 : reattach_req: ReAttachRequest,
1824 0 : ) -> Result<ReAttachResponse, ApiError> {
1825 0 : if let Some(register_req) = reattach_req.register {
1826 0 : self.node_register(register_req).await?;
1827 0 : }
1828 :
1829 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1830 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1831 :
1832 0 : tracing::info!(
1833 : node_id=%reattach_req.node_id,
1834 0 : "Incremented {} tenant shards' generations",
1835 0 : incremented_generations.len()
1836 : );
1837 :
1838 : // Apply the updated generation to our in-memory state, and
1839 : // gather discover secondary locations.
1840 0 : let mut locked = self.inner.write().unwrap();
1841 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1842 0 :
1843 0 : let mut response = ReAttachResponse {
1844 0 : tenants: Vec::new(),
1845 0 : };
1846 :
1847 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
1848 : // to call location_conf API with an old generation. Wait for cancellation to complete
1849 : // before responding to this request. Requires well implemented CancellationToken logic
1850 : // all the way to where we call location_conf. Even then, there can still be a location_conf
1851 : // request in flight over the network: TODO handle that by making location_conf API refuse
1852 : // to go backward in generations.
1853 :
1854 : // Scan through all shards, applying updates for ones where we updated generation
1855 : // and identifying shards that intend to have a secondary location on this node.
1856 0 : for (tenant_shard_id, shard) in tenants {
1857 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
1858 0 : let new_gen = *new_gen;
1859 0 : response.tenants.push(ReAttachResponseTenant {
1860 0 : id: *tenant_shard_id,
1861 0 : gen: Some(new_gen.into().unwrap()),
1862 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
1863 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
1864 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
1865 0 : // the stale/multi states at this point.
1866 0 : mode: LocationConfigMode::AttachedSingle,
1867 0 : });
1868 0 :
1869 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
1870 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
1871 : // Why can we update `observed` even though we're not sure our response will be received
1872 : // by the pageserver? Because the pageserver will not proceed with startup until
1873 : // it has processed response: if it loses it, we'll see another request and increment
1874 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
1875 0 : if let Some(conf) = observed.conf.as_mut() {
1876 0 : conf.generation = new_gen.into();
1877 0 : }
1878 0 : } else {
1879 0 : // This node has no observed state for the shard: perhaps it was offline
1880 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
1881 0 : // will be prompted to learn the location's state before it makes changes.
1882 0 : shard
1883 0 : .observed
1884 0 : .locations
1885 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
1886 0 : }
1887 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
1888 0 : // Ordering: pageserver will not accept /location_config requests until it has
1889 0 : // finished processing the response from re-attach. So we can update our in-memory state
1890 0 : // now, and be confident that we are not stamping on the result of some later location config.
1891 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
1892 0 : // so we might update observed state here, and then get over-written by some racing
1893 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
1894 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
1895 0 :
1896 0 : response.tenants.push(ReAttachResponseTenant {
1897 0 : id: *tenant_shard_id,
1898 0 : gen: None,
1899 0 : mode: LocationConfigMode::Secondary,
1900 0 : });
1901 0 :
1902 0 : // We must not update observed, because we have no guarantee that our
1903 0 : // response will be received by the pageserver. This could leave it
1904 0 : // falsely dirty, but the resulting reconcile should be idempotent.
1905 0 : }
1906 : }
1907 :
1908 : // We consider a node Active once we have composed a re-attach response, but we
1909 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
1910 : // implicitly synchronizes the LocationConfigs on the node.
1911 : //
1912 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
1913 : // but those requests will not be accepted by the node until it has finished processing
1914 : // the re-attach response.
1915 : //
1916 : // Additionally, reset the nodes scheduling policy to match the conditional update done
1917 : // in [`Persistence::re_attach`].
1918 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
1919 0 : let reset_scheduling = matches!(
1920 0 : node.get_scheduling(),
1921 : NodeSchedulingPolicy::PauseForRestart
1922 : | NodeSchedulingPolicy::Draining
1923 : | NodeSchedulingPolicy::Filling
1924 : );
1925 :
1926 0 : let mut new_nodes = (**nodes).clone();
1927 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
1928 0 : if reset_scheduling {
1929 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
1930 0 : }
1931 :
1932 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
1933 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
1934 0 :
1935 0 : scheduler.node_upsert(node);
1936 0 : let new_nodes = Arc::new(new_nodes);
1937 0 : *nodes = new_nodes;
1938 : } else {
1939 0 : tracing::error!(
1940 0 : "Reattaching node {} was removed while processing the request",
1941 : reattach_req.node_id
1942 : );
1943 : }
1944 0 : }
1945 :
1946 0 : Ok(response)
1947 0 : }
1948 :
1949 0 : pub(crate) async fn validate(
1950 0 : &self,
1951 0 : validate_req: ValidateRequest,
1952 0 : ) -> Result<ValidateResponse, DatabaseError> {
1953 : // Fast in-memory check: we may reject validation on anything that doesn't match our
1954 : // in-memory generation for a shard
1955 0 : let in_memory_result = {
1956 0 : let mut in_memory_result = Vec::new();
1957 0 : let locked = self.inner.read().unwrap();
1958 0 : for req_tenant in validate_req.tenants {
1959 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
1960 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
1961 0 : tracing::info!(
1962 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
1963 : req_tenant.id,
1964 : req_tenant.gen,
1965 : tenant_shard.generation
1966 : );
1967 :
1968 0 : in_memory_result.push((req_tenant.id, Generation::new(req_tenant.gen), valid));
1969 : } else {
1970 : // This is legal: for example during a shard split the pageserver may still
1971 : // have deletions in its queue from the old pre-split shard, or after deletion
1972 : // of a tenant that was busy with compaction/gc while being deleted.
1973 0 : tracing::info!(
1974 0 : "Refusing deletion validation for missing shard {}",
1975 : req_tenant.id
1976 : );
1977 : }
1978 : }
1979 :
1980 0 : in_memory_result
1981 : };
1982 :
1983 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
1984 : // in case of controller split-brain, where some other controller process might have incremented the generation.
1985 0 : let db_generations = self
1986 0 : .persistence
1987 0 : .shard_generations(in_memory_result.iter().filter_map(|i| {
1988 0 : if i.2 {
1989 0 : Some(&i.0)
1990 : } else {
1991 0 : None
1992 : }
1993 0 : }))
1994 0 : .await?;
1995 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
1996 0 :
1997 0 : let mut response = ValidateResponse {
1998 0 : tenants: Vec::new(),
1999 0 : };
2000 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2001 0 : let valid = if valid {
2002 0 : let db_generation = db_generations.get(&tenant_shard_id);
2003 0 : db_generation == Some(&Some(validate_generation))
2004 : } else {
2005 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2006 : // this prevents a pageserver from cleaning up an object in S3.
2007 0 : false
2008 : };
2009 :
2010 0 : response.tenants.push(ValidateResponseTenant {
2011 0 : id: tenant_shard_id,
2012 0 : valid,
2013 0 : })
2014 : }
2015 :
2016 0 : Ok(response)
2017 0 : }
2018 :
2019 0 : pub(crate) async fn tenant_create(
2020 0 : &self,
2021 0 : create_req: TenantCreateRequest,
2022 0 : ) -> Result<TenantCreateResponse, ApiError> {
2023 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2024 :
2025 : // Exclude any concurrent attempts to create/access the same tenant ID
2026 0 : let _tenant_lock = trace_exclusive_lock(
2027 0 : &self.tenant_op_locks,
2028 0 : create_req.new_tenant_id.tenant_id,
2029 0 : TenantOperations::Create,
2030 0 : )
2031 0 : .await;
2032 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2033 :
2034 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2035 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2036 : // accept compute notifications while it is in the process of creating. Reconciliation will
2037 : // be retried in the background.
2038 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2039 0 : }
2040 0 : Ok(response)
2041 0 : }
2042 :
2043 0 : pub(crate) async fn do_tenant_create(
2044 0 : &self,
2045 0 : create_req: TenantCreateRequest,
2046 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2047 0 : let placement_policy = create_req
2048 0 : .placement_policy
2049 0 : .clone()
2050 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2051 0 : .unwrap_or(PlacementPolicy::Attached(0));
2052 :
2053 : // This service expects to handle sharding itself: it is an error to try and directly create
2054 : // a particular shard here.
2055 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2056 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2057 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2058 0 : )));
2059 : } else {
2060 0 : create_req.new_tenant_id.tenant_id
2061 0 : };
2062 0 :
2063 0 : tracing::info!(
2064 0 : "Creating tenant {}, shard_count={:?}",
2065 : create_req.new_tenant_id,
2066 : create_req.shard_parameters.count,
2067 : );
2068 :
2069 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2070 0 : .map(|i| TenantShardId {
2071 0 : tenant_id,
2072 0 : shard_number: ShardNumber(i),
2073 0 : shard_count: create_req.shard_parameters.count,
2074 0 : })
2075 0 : .collect::<Vec<_>>();
2076 :
2077 : // If the caller specifies a None generation, it means "start from default". This is different
2078 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2079 : // an incompletely-onboarded tenant.
2080 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2081 0 : tracing::info!(
2082 0 : "tenant_create: secondary mode, generation is_some={}",
2083 0 : create_req.generation.is_some()
2084 : );
2085 0 : create_req.generation.map(Generation::new)
2086 : } else {
2087 0 : tracing::info!(
2088 0 : "tenant_create: not secondary mode, generation is_some={}",
2089 0 : create_req.generation.is_some()
2090 : );
2091 0 : Some(
2092 0 : create_req
2093 0 : .generation
2094 0 : .map(Generation::new)
2095 0 : .unwrap_or(INITIAL_GENERATION),
2096 0 : )
2097 : };
2098 :
2099 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2100 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2101 : // during the creation, rather than risking leaving orphan objects in S3.
2102 0 : let persist_tenant_shards = create_ids
2103 0 : .iter()
2104 0 : .map(|tenant_shard_id| TenantShardPersistence {
2105 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2106 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2107 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2108 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2109 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2110 0 : // The pageserver is not known until scheduling happens: we will set this column when
2111 0 : // incrementing the generation the first time we attach to a pageserver.
2112 0 : generation_pageserver: None,
2113 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2114 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2115 0 : splitting: SplitState::default(),
2116 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2117 0 : .unwrap(),
2118 0 : preferred_az_id: None,
2119 0 : })
2120 0 : .collect();
2121 0 :
2122 0 : match self
2123 0 : .persistence
2124 0 : .insert_tenant_shards(persist_tenant_shards)
2125 0 : .await
2126 : {
2127 0 : Ok(_) => {}
2128 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2129 : DatabaseErrorKind::UniqueViolation,
2130 : _,
2131 : ))) => {
2132 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2133 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2134 : // creation's shard count.
2135 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
2136 : }
2137 : // Any other database error is unexpected and a bug.
2138 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2139 : };
2140 :
2141 0 : let mut schedule_context = ScheduleContext::default();
2142 0 : let mut schedule_error = None;
2143 0 : let mut response_shards = Vec::new();
2144 0 : for tenant_shard_id in create_ids {
2145 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2146 :
2147 0 : let outcome = self
2148 0 : .do_initial_shard_scheduling(
2149 0 : tenant_shard_id,
2150 0 : initial_generation,
2151 0 : &create_req.shard_parameters,
2152 0 : create_req.config.clone(),
2153 0 : placement_policy.clone(),
2154 0 : &mut schedule_context,
2155 0 : )
2156 0 : .await;
2157 :
2158 0 : match outcome {
2159 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2160 0 : InitialShardScheduleOutcome::NotScheduled => {}
2161 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2162 0 : schedule_error = Some(err);
2163 0 : }
2164 : }
2165 : }
2166 :
2167 0 : let preferred_azs = {
2168 0 : let locked = self.inner.read().unwrap();
2169 0 : response_shards
2170 0 : .iter()
2171 0 : .filter_map(|resp| {
2172 0 : let az_id = locked
2173 0 : .nodes
2174 0 : .get(&resp.node_id)
2175 0 : .map(|n| n.get_availability_zone_id().clone())?;
2176 :
2177 0 : Some((resp.shard_id, az_id))
2178 0 : })
2179 0 : .collect::<Vec<_>>()
2180 : };
2181 :
2182 : // Note that we persist the preferred AZ for the new shards separately.
2183 : // In theory, we could "peek" the scheduler to determine where the shard will
2184 : // land, but the subsequent "real" call into the scheduler might select a different
2185 : // node. Hence, we do this awkward update to keep things consistent.
2186 0 : let updated = self
2187 0 : .persistence
2188 0 : .set_tenant_shard_preferred_azs(preferred_azs)
2189 0 : .await
2190 0 : .map_err(|err| {
2191 0 : ApiError::InternalServerError(anyhow::anyhow!(
2192 0 : "Failed to persist preferred az ids: {err}"
2193 0 : ))
2194 0 : })?;
2195 :
2196 : {
2197 0 : let mut locked = self.inner.write().unwrap();
2198 0 : for (tid, az_id) in updated {
2199 0 : if let Some(shard) = locked.tenants.get_mut(&tid) {
2200 0 : shard.set_preferred_az(az_id);
2201 0 : }
2202 : }
2203 : }
2204 :
2205 : // If we failed to schedule shards, then they are still created in the controller,
2206 : // but we return an error to the requester to avoid a silent failure when someone
2207 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2208 : // are present in the system. We do this here rather than in the above loop, to
2209 : // avoid situations where we only create a subset of shards in the tenant.
2210 0 : if let Some(e) = schedule_error {
2211 0 : return Err(ApiError::Conflict(format!(
2212 0 : "Failed to schedule shard(s): {e}"
2213 0 : )));
2214 0 : }
2215 0 :
2216 0 : let waiters = {
2217 0 : let mut locked = self.inner.write().unwrap();
2218 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2219 0 : tenants
2220 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2221 0 : .filter_map(|(_shard_id, shard)| self.maybe_reconcile_shard(shard, nodes))
2222 0 : .collect::<Vec<_>>()
2223 0 : };
2224 0 :
2225 0 : Ok((
2226 0 : TenantCreateResponse {
2227 0 : shards: response_shards,
2228 0 : },
2229 0 : waiters,
2230 0 : ))
2231 0 : }
2232 :
2233 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2234 : /// case of a new tenant and a pre-existing one.
2235 0 : async fn do_initial_shard_scheduling(
2236 0 : &self,
2237 0 : tenant_shard_id: TenantShardId,
2238 0 : initial_generation: Option<Generation>,
2239 0 : shard_params: &ShardParameters,
2240 0 : config: TenantConfig,
2241 0 : placement_policy: PlacementPolicy,
2242 0 : schedule_context: &mut ScheduleContext,
2243 0 : ) -> InitialShardScheduleOutcome {
2244 0 : let mut locked = self.inner.write().unwrap();
2245 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2246 :
2247 : use std::collections::btree_map::Entry;
2248 0 : match tenants.entry(tenant_shard_id) {
2249 0 : Entry::Occupied(mut entry) => {
2250 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2251 :
2252 : // TODO: schedule() should take an anti-affinity expression that pushes
2253 : // attached and secondary locations (independently) away frorm those
2254 : // pageservers also holding a shard for this tenant.
2255 :
2256 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2257 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2258 0 : }
2259 :
2260 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2261 0 : let generation = entry
2262 0 : .get()
2263 0 : .generation
2264 0 : .expect("Generation is set when in attached mode");
2265 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2266 0 : shard_id: tenant_shard_id,
2267 0 : node_id: *node_id,
2268 0 : generation: generation.into().unwrap(),
2269 0 : })
2270 : } else {
2271 0 : InitialShardScheduleOutcome::NotScheduled
2272 : }
2273 : }
2274 0 : Entry::Vacant(entry) => {
2275 0 : let state = entry.insert(TenantShard::new(
2276 0 : tenant_shard_id,
2277 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2278 0 : placement_policy,
2279 0 : ));
2280 0 :
2281 0 : state.generation = initial_generation;
2282 0 : state.config = config;
2283 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2284 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2285 0 : }
2286 :
2287 : // Only include shards in result if we are attaching: the purpose
2288 : // of the response is to tell the caller where the shards are attached.
2289 0 : if let Some(node_id) = state.intent.get_attached() {
2290 0 : let generation = state
2291 0 : .generation
2292 0 : .expect("Generation is set when in attached mode");
2293 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2294 0 : shard_id: tenant_shard_id,
2295 0 : node_id: *node_id,
2296 0 : generation: generation.into().unwrap(),
2297 0 : })
2298 : } else {
2299 0 : InitialShardScheduleOutcome::NotScheduled
2300 : }
2301 : }
2302 : }
2303 0 : }
2304 :
2305 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2306 : /// wait for reconciliation to complete before responding.
2307 0 : async fn await_waiters(
2308 0 : &self,
2309 0 : waiters: Vec<ReconcilerWaiter>,
2310 0 : timeout: Duration,
2311 0 : ) -> Result<(), ReconcileWaitError> {
2312 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2313 0 : for waiter in waiters {
2314 0 : let timeout = deadline.duration_since(Instant::now());
2315 0 : waiter.wait_timeout(timeout).await?;
2316 : }
2317 :
2318 0 : Ok(())
2319 0 : }
2320 :
2321 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2322 : /// in progress
2323 0 : async fn await_waiters_remainder(
2324 0 : &self,
2325 0 : waiters: Vec<ReconcilerWaiter>,
2326 0 : timeout: Duration,
2327 0 : ) -> Vec<ReconcilerWaiter> {
2328 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2329 0 : for waiter in waiters.iter() {
2330 0 : let timeout = deadline.duration_since(Instant::now());
2331 0 : let _ = waiter.wait_timeout(timeout).await;
2332 : }
2333 :
2334 0 : waiters
2335 0 : .into_iter()
2336 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2337 0 : .collect::<Vec<_>>()
2338 0 : }
2339 :
2340 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2341 : /// and transform it into either a tenant creation of a series of shard updates.
2342 : ///
2343 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2344 : /// still be returned.
2345 0 : fn tenant_location_config_prepare(
2346 0 : &self,
2347 0 : tenant_id: TenantId,
2348 0 : req: TenantLocationConfigRequest,
2349 0 : ) -> TenantCreateOrUpdate {
2350 0 : let mut updates = Vec::new();
2351 0 : let mut locked = self.inner.write().unwrap();
2352 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2353 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2354 :
2355 : // Use location config mode as an indicator of policy.
2356 0 : let placement_policy = match req.config.mode {
2357 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2358 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2359 : LocationConfigMode::AttachedMulti
2360 : | LocationConfigMode::AttachedSingle
2361 : | LocationConfigMode::AttachedStale => {
2362 0 : if nodes.len() > 1 {
2363 0 : PlacementPolicy::Attached(1)
2364 : } else {
2365 : // Convenience for dev/test: if we just have one pageserver, import
2366 : // tenants into non-HA mode so that scheduling will succeed.
2367 0 : PlacementPolicy::Attached(0)
2368 : }
2369 : }
2370 : };
2371 :
2372 0 : let mut create = true;
2373 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2374 : // Saw an existing shard: this is not a creation
2375 0 : create = false;
2376 :
2377 : // Shards may have initially been created by a Secondary request, where we
2378 : // would have left generation as None.
2379 : //
2380 : // We only update generation the first time we see an attached-mode request,
2381 : // and if there is no existing generation set. The caller is responsible for
2382 : // ensuring that no non-storage-controller pageserver ever uses a higher
2383 : // generation than they passed in here.
2384 : use LocationConfigMode::*;
2385 0 : let set_generation = match req.config.mode {
2386 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2387 0 : req.config.generation.map(Generation::new)
2388 : }
2389 0 : _ => None,
2390 : };
2391 :
2392 0 : updates.push(ShardUpdate {
2393 0 : tenant_shard_id: *shard_id,
2394 0 : placement_policy: placement_policy.clone(),
2395 0 : tenant_config: req.config.tenant_conf.clone(),
2396 0 : generation: set_generation,
2397 0 : });
2398 : }
2399 :
2400 0 : if create {
2401 : use LocationConfigMode::*;
2402 0 : let generation = match req.config.mode {
2403 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2404 : // If a caller provided a generation in a non-attached request, ignore it
2405 : // and leave our generation as None: this enables a subsequent update to set
2406 : // the generation when setting an attached mode for the first time.
2407 0 : _ => None,
2408 : };
2409 :
2410 0 : TenantCreateOrUpdate::Create(
2411 0 : // Synthesize a creation request
2412 0 : TenantCreateRequest {
2413 0 : new_tenant_id: tenant_shard_id,
2414 0 : generation,
2415 0 : shard_parameters: ShardParameters {
2416 0 : count: tenant_shard_id.shard_count,
2417 0 : // We only import un-sharded or single-sharded tenants, so stripe
2418 0 : // size can be made up arbitrarily here.
2419 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2420 0 : },
2421 0 : placement_policy: Some(placement_policy),
2422 0 : config: req.config.tenant_conf,
2423 0 : },
2424 0 : )
2425 : } else {
2426 0 : assert!(!updates.is_empty());
2427 0 : TenantCreateOrUpdate::Update(updates)
2428 : }
2429 0 : }
2430 :
2431 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2432 : /// directly with pageservers into this service.
2433 : ///
2434 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2435 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2436 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2437 : /// tenant's source of generation numbers.
2438 : ///
2439 : /// The mode in this request coarse-grained control of tenants:
2440 : /// - Call with mode Attached* to upsert the tenant.
2441 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2442 : /// to set an existing tenant to PolicyMode::Secondary
2443 : /// - Call with mode Detached to switch to PolicyMode::Detached
2444 0 : pub(crate) async fn tenant_location_config(
2445 0 : &self,
2446 0 : tenant_shard_id: TenantShardId,
2447 0 : req: TenantLocationConfigRequest,
2448 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2449 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2450 0 : let _tenant_lock = trace_exclusive_lock(
2451 0 : &self.tenant_op_locks,
2452 0 : tenant_shard_id.tenant_id,
2453 0 : TenantOperations::LocationConfig,
2454 0 : )
2455 0 : .await;
2456 :
2457 0 : if !tenant_shard_id.is_unsharded() {
2458 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2459 0 : "This API is for importing single-sharded or unsharded tenants"
2460 0 : )));
2461 0 : }
2462 0 :
2463 0 : // First check if this is a creation or an update
2464 0 : let create_or_update = self.tenant_location_config_prepare(tenant_shard_id.tenant_id, req);
2465 0 :
2466 0 : let mut result = TenantLocationConfigResponse {
2467 0 : shards: Vec::new(),
2468 0 : stripe_size: None,
2469 0 : };
2470 0 : let waiters = match create_or_update {
2471 0 : TenantCreateOrUpdate::Create(create_req) => {
2472 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2473 0 : result.shards = create_resp
2474 0 : .shards
2475 0 : .into_iter()
2476 0 : .map(|s| TenantShardLocation {
2477 0 : node_id: s.node_id,
2478 0 : shard_id: s.shard_id,
2479 0 : })
2480 0 : .collect();
2481 0 : waiters
2482 : }
2483 0 : TenantCreateOrUpdate::Update(updates) => {
2484 0 : // Persist updates
2485 0 : // Ordering: write to the database before applying changes in-memory, so that
2486 0 : // we will not appear time-travel backwards on a restart.
2487 0 : let mut schedule_context = ScheduleContext::default();
2488 : for ShardUpdate {
2489 0 : tenant_shard_id,
2490 0 : placement_policy,
2491 0 : tenant_config,
2492 0 : generation,
2493 0 : } in &updates
2494 : {
2495 0 : self.persistence
2496 0 : .update_tenant_shard(
2497 0 : TenantFilter::Shard(*tenant_shard_id),
2498 0 : Some(placement_policy.clone()),
2499 0 : Some(tenant_config.clone()),
2500 0 : *generation,
2501 0 : None,
2502 0 : )
2503 0 : .await?;
2504 : }
2505 :
2506 : // Apply updates in-memory
2507 0 : let mut waiters = Vec::new();
2508 0 : {
2509 0 : let mut locked = self.inner.write().unwrap();
2510 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2511 :
2512 : for ShardUpdate {
2513 0 : tenant_shard_id,
2514 0 : placement_policy,
2515 0 : tenant_config,
2516 0 : generation: update_generation,
2517 0 : } in updates
2518 : {
2519 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2520 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2521 0 : continue;
2522 : };
2523 :
2524 : // Update stripe size
2525 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2526 0 : result.stripe_size = Some(shard.shard.stripe_size);
2527 0 : }
2528 :
2529 0 : shard.policy = placement_policy;
2530 0 : shard.config = tenant_config;
2531 0 : if let Some(generation) = update_generation {
2532 0 : shard.generation = Some(generation);
2533 0 : }
2534 :
2535 0 : shard.schedule(scheduler, &mut schedule_context)?;
2536 :
2537 0 : let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
2538 0 : if let Some(waiter) = maybe_waiter {
2539 0 : waiters.push(waiter);
2540 0 : }
2541 :
2542 0 : if let Some(node_id) = shard.intent.get_attached() {
2543 0 : result.shards.push(TenantShardLocation {
2544 0 : shard_id: tenant_shard_id,
2545 0 : node_id: *node_id,
2546 0 : })
2547 0 : }
2548 : }
2549 : }
2550 0 : waiters
2551 : }
2552 : };
2553 :
2554 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2555 : // Do not treat a reconcile error as fatal: we have already applied any requested
2556 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2557 : // compute notification API. In these cases, it is important that we do not
2558 : // cause the cloud control plane to retry forever on this API.
2559 0 : tracing::warn!(
2560 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2561 : );
2562 0 : }
2563 :
2564 : // Logging the full result is useful because it lets us cross-check what the cloud control
2565 : // plane's tenant_shards table should contain.
2566 0 : tracing::info!("Complete, returning {result:?}");
2567 :
2568 0 : Ok(result)
2569 0 : }
2570 :
2571 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2572 : // We require an exclusive lock, because we are updating persistent and in-memory state
2573 0 : let _tenant_lock = trace_exclusive_lock(
2574 0 : &self.tenant_op_locks,
2575 0 : req.tenant_id,
2576 0 : TenantOperations::ConfigSet,
2577 0 : )
2578 0 : .await;
2579 :
2580 0 : let tenant_id = req.tenant_id;
2581 0 : let config = req.config;
2582 0 :
2583 0 : self.persistence
2584 0 : .update_tenant_shard(
2585 0 : TenantFilter::Tenant(req.tenant_id),
2586 0 : None,
2587 0 : Some(config.clone()),
2588 0 : None,
2589 0 : None,
2590 0 : )
2591 0 : .await?;
2592 :
2593 0 : let waiters = {
2594 0 : let mut waiters = Vec::new();
2595 0 : let mut locked = self.inner.write().unwrap();
2596 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2597 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2598 0 : shard.config = config.clone();
2599 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2600 0 : waiters.push(waiter);
2601 0 : }
2602 : }
2603 0 : waiters
2604 : };
2605 :
2606 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2607 : // Treat this as success because we have stored the configuration. If e.g.
2608 : // a node was unavailable at this time, it should not stop us accepting a
2609 : // configuration change.
2610 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2611 0 : }
2612 :
2613 0 : Ok(())
2614 0 : }
2615 :
2616 0 : pub(crate) fn tenant_config_get(
2617 0 : &self,
2618 0 : tenant_id: TenantId,
2619 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2620 0 : let config = {
2621 0 : let locked = self.inner.read().unwrap();
2622 0 :
2623 0 : match locked
2624 0 : .tenants
2625 0 : .range(TenantShardId::tenant_range(tenant_id))
2626 0 : .next()
2627 : {
2628 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2629 : None => {
2630 0 : return Err(ApiError::NotFound(
2631 0 : anyhow::anyhow!("Tenant not found").into(),
2632 0 : ))
2633 : }
2634 : }
2635 : };
2636 :
2637 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2638 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2639 : // and `effective_config` in the response is meaningless, but we retain that syntax
2640 : // in order to remain compatible with the pageserver API.
2641 :
2642 0 : let response = HashMap::from([
2643 : (
2644 : "tenant_specific_overrides",
2645 0 : serde_json::to_value(&config)
2646 0 : .context("serializing tenant specific overrides")
2647 0 : .map_err(ApiError::InternalServerError)?,
2648 : ),
2649 : (
2650 0 : "effective_config",
2651 0 : serde_json::to_value(&config)
2652 0 : .context("serializing effective config")
2653 0 : .map_err(ApiError::InternalServerError)?,
2654 : ),
2655 : ]);
2656 :
2657 0 : Ok(response)
2658 0 : }
2659 :
2660 0 : pub(crate) async fn tenant_time_travel_remote_storage(
2661 0 : &self,
2662 0 : time_travel_req: &TenantTimeTravelRequest,
2663 0 : tenant_id: TenantId,
2664 0 : timestamp: Cow<'_, str>,
2665 0 : done_if_after: Cow<'_, str>,
2666 0 : ) -> Result<(), ApiError> {
2667 0 : let _tenant_lock = trace_exclusive_lock(
2668 0 : &self.tenant_op_locks,
2669 0 : tenant_id,
2670 0 : TenantOperations::TimeTravelRemoteStorage,
2671 0 : )
2672 0 : .await;
2673 :
2674 0 : let node = {
2675 0 : let mut locked = self.inner.write().unwrap();
2676 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
2677 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
2678 : // but only at the start of the process, so it's really just to prevent operator
2679 : // mistakes.
2680 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
2681 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
2682 : {
2683 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2684 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
2685 0 : )));
2686 0 : }
2687 0 : let maybe_attached = shard
2688 0 : .observed
2689 0 : .locations
2690 0 : .iter()
2691 0 : .filter_map(|(node_id, observed_location)| {
2692 0 : observed_location
2693 0 : .conf
2694 0 : .as_ref()
2695 0 : .map(|loc| (node_id, observed_location, loc.mode))
2696 0 : })
2697 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
2698 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
2699 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
2700 0 : }
2701 : }
2702 0 : let scheduler = &mut locked.scheduler;
2703 : // Right now we only perform the operation on a single node without parallelization
2704 : // TODO fan out the operation to multiple nodes for better performance
2705 0 : let node_id = scheduler.any_available_node()?;
2706 0 : let node = locked
2707 0 : .nodes
2708 0 : .get(&node_id)
2709 0 : .expect("Pageservers may not be deleted while lock is active");
2710 0 : node.clone()
2711 0 : };
2712 0 :
2713 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
2714 0 : let mut counts = time_travel_req
2715 0 : .shard_counts
2716 0 : .iter()
2717 0 : .copied()
2718 0 : .collect::<HashSet<_>>()
2719 0 : .into_iter()
2720 0 : .collect::<Vec<_>>();
2721 0 : counts.sort_unstable();
2722 :
2723 0 : for count in counts {
2724 0 : let shard_ids = (0..count.count())
2725 0 : .map(|i| TenantShardId {
2726 0 : tenant_id,
2727 0 : shard_number: ShardNumber(i),
2728 0 : shard_count: count,
2729 0 : })
2730 0 : .collect::<Vec<_>>();
2731 0 : for tenant_shard_id in shard_ids {
2732 0 : let client = PageserverClient::new(
2733 0 : node.get_id(),
2734 0 : node.base_url(),
2735 0 : self.config.jwt_token.as_deref(),
2736 0 : );
2737 0 :
2738 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
2739 :
2740 0 : client
2741 0 : .tenant_time_travel_remote_storage(
2742 0 : tenant_shard_id,
2743 0 : ×tamp,
2744 0 : &done_if_after,
2745 0 : )
2746 0 : .await
2747 0 : .map_err(|e| {
2748 0 : ApiError::InternalServerError(anyhow::anyhow!(
2749 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
2750 0 : node
2751 0 : ))
2752 0 : })?;
2753 : }
2754 : }
2755 0 : Ok(())
2756 0 : }
2757 :
2758 0 : pub(crate) async fn tenant_secondary_download(
2759 0 : &self,
2760 0 : tenant_id: TenantId,
2761 0 : wait: Option<Duration>,
2762 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
2763 0 : let _tenant_lock = trace_shared_lock(
2764 0 : &self.tenant_op_locks,
2765 0 : tenant_id,
2766 0 : TenantOperations::SecondaryDownload,
2767 0 : )
2768 0 : .await;
2769 :
2770 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
2771 0 : let targets = {
2772 0 : let locked = self.inner.read().unwrap();
2773 0 : let mut targets = Vec::new();
2774 :
2775 0 : for (tenant_shard_id, shard) in
2776 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2777 : {
2778 0 : for node_id in shard.intent.get_secondary() {
2779 0 : let node = locked
2780 0 : .nodes
2781 0 : .get(node_id)
2782 0 : .expect("Pageservers may not be deleted while referenced");
2783 0 :
2784 0 : targets.push((*tenant_shard_id, node.clone()));
2785 0 : }
2786 : }
2787 0 : targets
2788 0 : };
2789 0 :
2790 0 : // Issue concurrent requests to all shards' locations
2791 0 : let mut futs = FuturesUnordered::new();
2792 0 : for (tenant_shard_id, node) in targets {
2793 0 : let client = PageserverClient::new(
2794 0 : node.get_id(),
2795 0 : node.base_url(),
2796 0 : self.config.jwt_token.as_deref(),
2797 0 : );
2798 0 : futs.push(async move {
2799 0 : let result = client
2800 0 : .tenant_secondary_download(tenant_shard_id, wait)
2801 0 : .await;
2802 0 : (result, node, tenant_shard_id)
2803 0 : })
2804 : }
2805 :
2806 : // Handle any errors returned by pageservers. This includes cases like this request racing with
2807 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
2808 : // well as more general cases like 503s, 500s, or timeouts.
2809 0 : let mut aggregate_progress = SecondaryProgress::default();
2810 0 : let mut aggregate_status: Option<StatusCode> = None;
2811 0 : let mut error: Option<mgmt_api::Error> = None;
2812 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
2813 0 : match result {
2814 0 : Err(e) => {
2815 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
2816 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
2817 0 : // than they had hoped for.
2818 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
2819 0 : error = Some(e)
2820 : }
2821 0 : Ok((status_code, progress)) => {
2822 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
2823 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
2824 0 : aggregate_progress.layers_total += progress.layers_total;
2825 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
2826 0 : aggregate_progress.bytes_total += progress.bytes_total;
2827 0 : aggregate_progress.heatmap_mtime =
2828 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
2829 0 : aggregate_status = match aggregate_status {
2830 0 : None => Some(status_code),
2831 0 : Some(StatusCode::OK) => Some(status_code),
2832 0 : Some(cur) => {
2833 0 : // Other status codes (e.g. 202) -- do not overwrite.
2834 0 : Some(cur)
2835 : }
2836 : };
2837 : }
2838 : }
2839 : }
2840 :
2841 : // If any of the shards return 202, indicate our result as 202.
2842 0 : match aggregate_status {
2843 : None => {
2844 0 : match error {
2845 0 : Some(e) => {
2846 0 : // No successes, and an error: surface it
2847 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
2848 : }
2849 : None => {
2850 : // No shards found
2851 0 : Err(ApiError::NotFound(
2852 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2853 0 : ))
2854 : }
2855 : }
2856 : }
2857 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
2858 : }
2859 0 : }
2860 :
2861 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
2862 0 : let _tenant_lock =
2863 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
2864 :
2865 : // Detach all shards. This also deletes local pageserver shard data.
2866 0 : let (detach_waiters, node) = {
2867 0 : let mut detach_waiters = Vec::new();
2868 0 : let mut locked = self.inner.write().unwrap();
2869 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2870 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2871 : // Update the tenant's intent to remove all attachments
2872 0 : shard.policy = PlacementPolicy::Detached;
2873 0 : shard
2874 0 : .schedule(scheduler, &mut ScheduleContext::default())
2875 0 : .expect("De-scheduling is infallible");
2876 0 : debug_assert!(shard.intent.get_attached().is_none());
2877 0 : debug_assert!(shard.intent.get_secondary().is_empty());
2878 :
2879 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2880 0 : detach_waiters.push(waiter);
2881 0 : }
2882 : }
2883 :
2884 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
2885 : // was attached, just has to be able to see the S3 content)
2886 0 : let node_id = scheduler.any_available_node()?;
2887 0 : let node = nodes
2888 0 : .get(&node_id)
2889 0 : .expect("Pageservers may not be deleted while lock is active");
2890 0 : (detach_waiters, node.clone())
2891 0 : };
2892 0 :
2893 0 : // This reconcile wait can fail in a few ways:
2894 0 : // A there is a very long queue for the reconciler semaphore
2895 0 : // B some pageserver is failing to handle a detach promptly
2896 0 : // C some pageserver goes offline right at the moment we send it a request.
2897 0 : //
2898 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
2899 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
2900 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
2901 0 : // deleting the underlying data).
2902 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
2903 0 : .await?;
2904 :
2905 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
2906 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
2907 : // the tenant ID prefix, including all shards (even possibly stale ones).
2908 0 : match node
2909 0 : .with_client_retries(
2910 0 : |client| async move {
2911 0 : client
2912 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
2913 0 : .await
2914 0 : },
2915 0 : &self.config.jwt_token,
2916 0 : 1,
2917 0 : 3,
2918 0 : RECONCILE_TIMEOUT,
2919 0 : &self.cancel,
2920 0 : )
2921 0 : .await
2922 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
2923 : {
2924 0 : Ok(_) => {}
2925 : Err(mgmt_api::Error::Cancelled) => {
2926 0 : return Err(ApiError::ShuttingDown);
2927 : }
2928 0 : Err(e) => {
2929 0 : // This is unexpected: remote deletion should be infallible, unless the object store
2930 0 : // at large is unavailable.
2931 0 : tracing::error!("Error deleting via node {node}: {e}");
2932 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
2933 : }
2934 : }
2935 :
2936 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
2937 : // our in-memory state and database state.
2938 :
2939 : // Ordering: we delete persistent state first: if we then
2940 : // crash, we will drop the in-memory state.
2941 :
2942 : // Drop persistent state.
2943 0 : self.persistence.delete_tenant(tenant_id).await?;
2944 :
2945 : // Drop in-memory state
2946 : {
2947 0 : let mut locked = self.inner.write().unwrap();
2948 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2949 :
2950 : // Dereference Scheduler from shards before dropping them
2951 0 : for (_tenant_shard_id, shard) in
2952 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2953 0 : {
2954 0 : shard.intent.clear(scheduler);
2955 0 : }
2956 :
2957 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
2958 0 : tracing::info!(
2959 0 : "Deleted tenant {tenant_id}, now have {} tenants",
2960 0 : locked.tenants.len()
2961 : );
2962 : };
2963 :
2964 : // Success is represented as 404, to imitate the existing pageserver deletion API
2965 0 : Ok(StatusCode::NOT_FOUND)
2966 0 : }
2967 :
2968 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
2969 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
2970 : /// the tenant's policies (configuration) within the storage controller
2971 0 : pub(crate) async fn tenant_update_policy(
2972 0 : &self,
2973 0 : tenant_id: TenantId,
2974 0 : req: TenantPolicyRequest,
2975 0 : ) -> Result<(), ApiError> {
2976 : // We require an exclusive lock, because we are updating persistent and in-memory state
2977 0 : let _tenant_lock = trace_exclusive_lock(
2978 0 : &self.tenant_op_locks,
2979 0 : tenant_id,
2980 0 : TenantOperations::UpdatePolicy,
2981 0 : )
2982 0 : .await;
2983 :
2984 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
2985 :
2986 : let TenantPolicyRequest {
2987 0 : placement,
2988 0 : scheduling,
2989 0 : } = req;
2990 0 :
2991 0 : self.persistence
2992 0 : .update_tenant_shard(
2993 0 : TenantFilter::Tenant(tenant_id),
2994 0 : placement.clone(),
2995 0 : None,
2996 0 : None,
2997 0 : scheduling,
2998 0 : )
2999 0 : .await?;
3000 :
3001 0 : let mut schedule_context = ScheduleContext::default();
3002 0 : let mut locked = self.inner.write().unwrap();
3003 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3004 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3005 0 : if let Some(placement) = &placement {
3006 0 : shard.policy = placement.clone();
3007 0 :
3008 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3009 0 : "Updated placement policy to {placement:?}");
3010 0 : }
3011 :
3012 0 : if let Some(scheduling) = &scheduling {
3013 0 : shard.set_scheduling_policy(*scheduling);
3014 0 :
3015 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3016 0 : "Updated scheduling policy to {scheduling:?}");
3017 0 : }
3018 :
3019 : // In case scheduling is being switched back on, try it now.
3020 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3021 0 : self.maybe_reconcile_shard(shard, nodes);
3022 : }
3023 :
3024 0 : Ok(())
3025 0 : }
3026 :
3027 0 : pub(crate) async fn tenant_timeline_create(
3028 0 : &self,
3029 0 : tenant_id: TenantId,
3030 0 : mut create_req: TimelineCreateRequest,
3031 0 : ) -> Result<TimelineInfo, ApiError> {
3032 0 : tracing::info!(
3033 0 : "Creating timeline {}/{}",
3034 : tenant_id,
3035 : create_req.new_timeline_id,
3036 : );
3037 :
3038 0 : let _tenant_lock = trace_shared_lock(
3039 0 : &self.tenant_op_locks,
3040 0 : tenant_id,
3041 0 : TenantOperations::TimelineCreate,
3042 0 : )
3043 0 : .await;
3044 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3045 :
3046 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3047 0 : if targets.0.is_empty() {
3048 0 : return Err(ApiError::NotFound(
3049 0 : anyhow::anyhow!("Tenant not found").into(),
3050 0 : ));
3051 0 : };
3052 0 :
3053 0 : let (shard_zero_tid, shard_zero_locations) =
3054 0 : targets.0.pop_first().expect("Must have at least one shard");
3055 0 : assert!(shard_zero_tid.is_shard_zero());
3056 :
3057 0 : async fn create_one(
3058 0 : tenant_shard_id: TenantShardId,
3059 0 : locations: ShardMutationLocations,
3060 0 : jwt: Option<String>,
3061 0 : create_req: TimelineCreateRequest,
3062 0 : ) -> Result<TimelineInfo, ApiError> {
3063 0 : let latest = locations.latest.node;
3064 0 :
3065 0 : tracing::info!(
3066 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3067 : tenant_shard_id,
3068 : create_req.new_timeline_id,
3069 : locations.latest.generation
3070 : );
3071 :
3072 0 : let client =
3073 0 : PageserverClient::new(latest.get_id(), latest.base_url(), jwt.as_deref());
3074 :
3075 0 : let timeline_info = client
3076 0 : .timeline_create(tenant_shard_id, &create_req)
3077 0 : .await
3078 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3079 :
3080 : // We propagate timeline creations to all attached locations such that a compute
3081 : // for the new timeline is able to start regardless of the current state of the
3082 : // tenant shard reconciliation.
3083 0 : for location in locations.other {
3084 0 : tracing::info!(
3085 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3086 : tenant_shard_id,
3087 : create_req.new_timeline_id,
3088 : location.node,
3089 : location.generation
3090 : );
3091 :
3092 0 : let client = PageserverClient::new(
3093 0 : location.node.get_id(),
3094 0 : location.node.base_url(),
3095 0 : jwt.as_deref(),
3096 0 : );
3097 :
3098 0 : let res = client
3099 0 : .timeline_create(tenant_shard_id, &create_req)
3100 0 : .await;
3101 :
3102 0 : if let Err(e) = res {
3103 0 : match e {
3104 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3105 0 : // Tenant might have been detached from the stale location,
3106 0 : // so ignore 404s.
3107 0 : },
3108 : _ => {
3109 0 : return Err(passthrough_api_error(&location.node, e));
3110 : }
3111 : }
3112 0 : }
3113 : }
3114 :
3115 0 : Ok(timeline_info)
3116 0 : }
3117 :
3118 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3119 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3120 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3121 0 : let timeline_info = create_one(
3122 0 : shard_zero_tid,
3123 0 : shard_zero_locations,
3124 0 : self.config.jwt_token.clone(),
3125 0 : create_req.clone(),
3126 0 : )
3127 0 : .await?;
3128 :
3129 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3130 0 : if create_req.ancestor_timeline_id.is_some() && create_req.ancestor_start_lsn.is_none()
3131 0 : {
3132 0 : create_req.ancestor_start_lsn = timeline_info.ancestor_lsn;
3133 0 : }
3134 :
3135 : // Create timeline on remaining shards with number >0
3136 0 : if !targets.0.is_empty() {
3137 : // If we had multiple shards, issue requests for the remainder now.
3138 0 : let jwt = &self.config.jwt_token;
3139 0 : self.tenant_for_shards(
3140 0 : targets
3141 0 : .0
3142 0 : .iter()
3143 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3144 0 : .collect(),
3145 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3146 0 : let create_req = create_req.clone();
3147 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3148 0 : Box::pin(create_one(
3149 0 : tenant_shard_id,
3150 0 : mutation_locations,
3151 0 : jwt.clone(),
3152 0 : create_req,
3153 0 : ))
3154 0 : },
3155 0 : )
3156 0 : .await?;
3157 0 : }
3158 :
3159 0 : Ok(timeline_info)
3160 0 : })
3161 0 : .await?
3162 0 : }
3163 :
3164 0 : pub(crate) async fn tenant_timeline_archival_config(
3165 0 : &self,
3166 0 : tenant_id: TenantId,
3167 0 : timeline_id: TimelineId,
3168 0 : req: TimelineArchivalConfigRequest,
3169 0 : ) -> Result<(), ApiError> {
3170 0 : tracing::info!(
3171 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3172 : req.state
3173 : );
3174 :
3175 0 : let _tenant_lock = trace_shared_lock(
3176 0 : &self.tenant_op_locks,
3177 0 : tenant_id,
3178 0 : TenantOperations::TimelineArchivalConfig,
3179 0 : )
3180 0 : .await;
3181 :
3182 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3183 0 : if targets.0.is_empty() {
3184 0 : return Err(ApiError::NotFound(
3185 0 : anyhow::anyhow!("Tenant not found").into(),
3186 0 : ));
3187 0 : }
3188 0 : async fn config_one(
3189 0 : tenant_shard_id: TenantShardId,
3190 0 : timeline_id: TimelineId,
3191 0 : node: Node,
3192 0 : jwt: Option<String>,
3193 0 : req: TimelineArchivalConfigRequest,
3194 0 : ) -> Result<(), ApiError> {
3195 0 : tracing::info!(
3196 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3197 : );
3198 :
3199 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3200 0 :
3201 0 : client
3202 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3203 0 : .await
3204 0 : .map_err(|e| match e {
3205 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3206 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3207 : }
3208 0 : _ => passthrough_api_error(&node, e),
3209 0 : })
3210 0 : }
3211 :
3212 : // no shard needs to go first/last; the operation should be idempotent
3213 : // TODO: it would be great to ensure that all shards return the same error
3214 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3215 0 : let results = self
3216 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3217 0 : futures::FutureExt::boxed(config_one(
3218 0 : tenant_shard_id,
3219 0 : timeline_id,
3220 0 : node,
3221 0 : self.config.jwt_token.clone(),
3222 0 : req.clone(),
3223 0 : ))
3224 0 : })
3225 0 : .await?;
3226 0 : assert!(!results.is_empty(), "must have at least one result");
3227 :
3228 0 : Ok(())
3229 0 : }).await?
3230 0 : }
3231 :
3232 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3233 0 : &self,
3234 0 : tenant_id: TenantId,
3235 0 : timeline_id: TimelineId,
3236 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3237 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3238 :
3239 0 : let _tenant_lock = trace_shared_lock(
3240 0 : &self.tenant_op_locks,
3241 0 : tenant_id,
3242 0 : TenantOperations::TimelineDetachAncestor,
3243 0 : )
3244 0 : .await;
3245 :
3246 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3247 0 : if targets.0.is_empty() {
3248 0 : return Err(ApiError::NotFound(
3249 0 : anyhow::anyhow!("Tenant not found").into(),
3250 0 : ));
3251 0 : }
3252 :
3253 0 : async fn detach_one(
3254 0 : tenant_shard_id: TenantShardId,
3255 0 : timeline_id: TimelineId,
3256 0 : node: Node,
3257 0 : jwt: Option<String>,
3258 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3259 0 : tracing::info!(
3260 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3261 : );
3262 :
3263 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3264 0 :
3265 0 : client
3266 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3267 0 : .await
3268 0 : .map_err(|e| {
3269 : use mgmt_api::Error;
3270 :
3271 0 : match e {
3272 : // no ancestor (ever)
3273 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3274 0 : "{node}: {}",
3275 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3276 0 : )),
3277 : // too many ancestors
3278 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3279 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3280 : }
3281 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3282 0 : // avoid turning these into conflicts to remain compatible with
3283 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3284 0 : // detach
3285 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3286 : }
3287 : // rest can be mapped as usual
3288 0 : other => passthrough_api_error(&node, other),
3289 : }
3290 0 : })
3291 0 : .map(|res| (tenant_shard_id.shard_number, res))
3292 0 : }
3293 :
3294 : // no shard needs to go first/last; the operation should be idempotent
3295 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3296 0 : let mut results = self
3297 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3298 0 : futures::FutureExt::boxed(detach_one(
3299 0 : tenant_shard_id,
3300 0 : timeline_id,
3301 0 : node,
3302 0 : self.config.jwt_token.clone(),
3303 0 : ))
3304 0 : })
3305 0 : .await?;
3306 :
3307 0 : let any = results.pop().expect("we must have at least one response");
3308 0 :
3309 0 : let mismatching = results
3310 0 : .iter()
3311 0 : .filter(|(_, res)| res != &any.1)
3312 0 : .collect::<Vec<_>>();
3313 0 : if !mismatching.is_empty() {
3314 : // this can be hit by races which should not happen because operation lock on cplane
3315 0 : let matching = results.len() - mismatching.len();
3316 0 : tracing::error!(
3317 : matching,
3318 : compared_against=?any,
3319 : ?mismatching,
3320 0 : "shards returned different results"
3321 : );
3322 :
3323 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3324 0 : }
3325 0 :
3326 0 : Ok(any.1)
3327 0 : }).await?
3328 0 : }
3329 :
3330 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3331 0 : &self,
3332 0 : tenant_id: TenantId,
3333 0 : timeline_id: TimelineId,
3334 0 : dir: BlockUnblock,
3335 0 : ) -> Result<(), ApiError> {
3336 0 : let _tenant_lock = trace_shared_lock(
3337 0 : &self.tenant_op_locks,
3338 0 : tenant_id,
3339 0 : TenantOperations::TimelineGcBlockUnblock,
3340 0 : )
3341 0 : .await;
3342 :
3343 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3344 0 : if targets.0.is_empty() {
3345 0 : return Err(ApiError::NotFound(
3346 0 : anyhow::anyhow!("Tenant not found").into(),
3347 0 : ));
3348 0 : }
3349 :
3350 0 : async fn do_one(
3351 0 : tenant_shard_id: TenantShardId,
3352 0 : timeline_id: TimelineId,
3353 0 : node: Node,
3354 0 : jwt: Option<String>,
3355 0 : dir: BlockUnblock,
3356 0 : ) -> Result<(), ApiError> {
3357 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3358 0 :
3359 0 : client
3360 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3361 0 : .await
3362 0 : .map_err(|e| passthrough_api_error(&node, e))
3363 0 : }
3364 :
3365 : // no shard needs to go first/last; the operation should be idempotent
3366 0 : let locations = targets
3367 0 : .0
3368 0 : .iter()
3369 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3370 0 : .collect();
3371 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
3372 0 : futures::FutureExt::boxed(do_one(
3373 0 : tenant_shard_id,
3374 0 : timeline_id,
3375 0 : node,
3376 0 : self.config.jwt_token.clone(),
3377 0 : dir,
3378 0 : ))
3379 0 : })
3380 0 : .await
3381 0 : })
3382 0 : .await??;
3383 0 : Ok(())
3384 0 : }
3385 :
3386 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3387 : ///
3388 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3389 0 : async fn tenant_for_shards<F, R>(
3390 0 : &self,
3391 0 : locations: Vec<(TenantShardId, Node)>,
3392 0 : mut req_fn: F,
3393 0 : ) -> Result<Vec<R>, ApiError>
3394 0 : where
3395 0 : F: FnMut(
3396 0 : TenantShardId,
3397 0 : Node,
3398 0 : )
3399 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3400 0 : {
3401 0 : let mut futs = FuturesUnordered::new();
3402 0 : let mut results = Vec::with_capacity(locations.len());
3403 :
3404 0 : for (tenant_shard_id, node) in locations {
3405 0 : futs.push(req_fn(tenant_shard_id, node));
3406 0 : }
3407 :
3408 0 : while let Some(r) = futs.next().await {
3409 0 : results.push(r?);
3410 : }
3411 :
3412 0 : Ok(results)
3413 0 : }
3414 :
3415 : /// Concurrently invoke a pageserver API call on many shards at once
3416 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3417 0 : &self,
3418 0 : locations: Vec<(TenantShardId, Node)>,
3419 0 : op: O,
3420 0 : warn_threshold: u32,
3421 0 : max_retries: u32,
3422 0 : timeout: Duration,
3423 0 : cancel: &CancellationToken,
3424 0 : ) -> Vec<mgmt_api::Result<T>>
3425 0 : where
3426 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3427 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3428 0 : {
3429 0 : let mut futs = FuturesUnordered::new();
3430 0 : let mut results = Vec::with_capacity(locations.len());
3431 :
3432 0 : for (tenant_shard_id, node) in locations {
3433 0 : futs.push(async move {
3434 0 : node.with_client_retries(
3435 0 : |client| op(tenant_shard_id, client),
3436 0 : &self.config.jwt_token,
3437 0 : warn_threshold,
3438 0 : max_retries,
3439 0 : timeout,
3440 0 : cancel,
3441 0 : )
3442 0 : .await
3443 0 : });
3444 0 : }
3445 :
3446 0 : while let Some(r) = futs.next().await {
3447 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3448 0 : results.push(r);
3449 0 : }
3450 :
3451 0 : results
3452 0 : }
3453 :
3454 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3455 : /// when creating and deleting timelines:
3456 : /// - Makes sure shards are attached somewhere if they weren't already
3457 : /// - Looks up the shards and the nodes where they were most recently attached
3458 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3459 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3460 0 : async fn tenant_remote_mutation<R, O, F>(
3461 0 : &self,
3462 0 : tenant_id: TenantId,
3463 0 : op: O,
3464 0 : ) -> Result<R, ApiError>
3465 0 : where
3466 0 : O: FnOnce(TenantMutationLocations) -> F,
3467 0 : F: std::future::Future<Output = R>,
3468 0 : {
3469 0 : let mutation_locations = {
3470 0 : let mut locations = TenantMutationLocations::default();
3471 :
3472 : // Load the currently attached pageservers for the latest generation of each shard. This can
3473 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3474 : // will still be the latest when we're done: we will check generations again at the end of
3475 : // this function to handle that.
3476 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
3477 :
3478 0 : if generations
3479 0 : .iter()
3480 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3481 : {
3482 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3483 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
3484 0 : let locked = self.inner.read().unwrap();
3485 0 : for (shard_id, shard) in
3486 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3487 : {
3488 0 : match shard.policy {
3489 0 : PlacementPolicy::Attached(_) => {
3490 0 : // This shard is meant to be attached: the caller is not wrong to try and
3491 0 : // use this function, but we can't service the request right now.
3492 0 : }
3493 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3494 0 : return Err(ApiError::Conflict(format!(
3495 0 : "Shard {shard_id} tenant has policy {:?}",
3496 0 : shard.policy
3497 0 : )));
3498 : }
3499 : }
3500 : }
3501 :
3502 0 : return Err(ApiError::ResourceUnavailable(
3503 0 : "One or more shards in tenant is not yet attached".into(),
3504 0 : ));
3505 0 : }
3506 0 :
3507 0 : let locked = self.inner.read().unwrap();
3508 : for ShardGenerationState {
3509 0 : tenant_shard_id,
3510 0 : generation,
3511 0 : generation_pageserver,
3512 0 : } in generations
3513 : {
3514 0 : let node_id = generation_pageserver.expect("We checked for None above");
3515 0 : let node = locked
3516 0 : .nodes
3517 0 : .get(&node_id)
3518 0 : .ok_or(ApiError::Conflict(format!(
3519 0 : "Raced with removal of node {node_id}"
3520 0 : )))?;
3521 0 : let generation = generation.expect("Checked above");
3522 0 :
3523 0 : let tenant = locked.tenants.get(&tenant_shard_id);
3524 :
3525 : // TODO(vlad): Abstract the logic that finds stale attached locations
3526 : // from observed state into a [`Service`] method.
3527 0 : let other_locations = match tenant {
3528 0 : Some(tenant) => {
3529 0 : let mut other = tenant.attached_locations();
3530 0 : let latest_location_index =
3531 0 : other.iter().position(|&l| l == (node.get_id(), generation));
3532 0 : if let Some(idx) = latest_location_index {
3533 0 : other.remove(idx);
3534 0 : }
3535 :
3536 0 : other
3537 : }
3538 0 : None => Vec::default(),
3539 : };
3540 :
3541 0 : let location = ShardMutationLocations {
3542 0 : latest: MutationLocation {
3543 0 : node: node.clone(),
3544 0 : generation,
3545 0 : },
3546 0 : other: other_locations
3547 0 : .into_iter()
3548 0 : .filter_map(|(node_id, generation)| {
3549 0 : let node = locked.nodes.get(&node_id)?;
3550 :
3551 0 : Some(MutationLocation {
3552 0 : node: node.clone(),
3553 0 : generation,
3554 0 : })
3555 0 : })
3556 0 : .collect(),
3557 0 : };
3558 0 : locations.0.insert(tenant_shard_id, location);
3559 0 : }
3560 :
3561 0 : locations
3562 : };
3563 :
3564 0 : let result = op(mutation_locations.clone()).await;
3565 :
3566 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
3567 : // our remote operation executed on the latest generation and is therefore persistent.
3568 : {
3569 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
3570 0 : if latest_generations
3571 0 : .into_iter()
3572 0 : .map(
3573 0 : |ShardGenerationState {
3574 : tenant_shard_id,
3575 : generation,
3576 : generation_pageserver: _,
3577 0 : }| (tenant_shard_id, generation),
3578 0 : )
3579 0 : .collect::<Vec<_>>()
3580 0 : != mutation_locations
3581 0 : .0
3582 0 : .into_iter()
3583 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
3584 0 : .collect::<Vec<_>>()
3585 : {
3586 : // We raced with something that incremented the generation, and therefore cannot be
3587 : // confident that our actions are persistent (they might have hit an old generation).
3588 : //
3589 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
3590 0 : return Err(ApiError::ResourceUnavailable(
3591 0 : "Tenant attachment changed, please retry".into(),
3592 0 : ));
3593 0 : }
3594 0 : }
3595 0 :
3596 0 : Ok(result)
3597 0 : }
3598 :
3599 0 : pub(crate) async fn tenant_timeline_delete(
3600 0 : &self,
3601 0 : tenant_id: TenantId,
3602 0 : timeline_id: TimelineId,
3603 0 : ) -> Result<StatusCode, ApiError> {
3604 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
3605 0 : let _tenant_lock = trace_shared_lock(
3606 0 : &self.tenant_op_locks,
3607 0 : tenant_id,
3608 0 : TenantOperations::TimelineDelete,
3609 0 : )
3610 0 : .await;
3611 :
3612 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3613 0 : if targets.0.is_empty() {
3614 0 : return Err(ApiError::NotFound(
3615 0 : anyhow::anyhow!("Tenant not found").into(),
3616 0 : ));
3617 0 : }
3618 0 :
3619 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
3620 0 : assert!(shard_zero_tid.is_shard_zero());
3621 :
3622 0 : async fn delete_one(
3623 0 : tenant_shard_id: TenantShardId,
3624 0 : timeline_id: TimelineId,
3625 0 : node: Node,
3626 0 : jwt: Option<String>,
3627 0 : ) -> Result<StatusCode, ApiError> {
3628 0 : tracing::info!(
3629 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3630 : );
3631 :
3632 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3633 0 : let res = client
3634 0 : .timeline_delete(tenant_shard_id, timeline_id)
3635 0 : .await;
3636 :
3637 0 : match res {
3638 0 : Ok(ok) => Ok(ok),
3639 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
3640 0 : Err(e) => {
3641 0 : Err(
3642 0 : ApiError::InternalServerError(anyhow::anyhow!(
3643 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
3644 0 : ))
3645 0 : )
3646 : }
3647 : }
3648 0 : }
3649 :
3650 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3651 0 : let statuses = self
3652 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
3653 0 : Box::pin(delete_one(
3654 0 : tenant_shard_id,
3655 0 : timeline_id,
3656 0 : node,
3657 0 : self.config.jwt_token.clone(),
3658 0 : ))
3659 0 : })
3660 0 : .await?;
3661 :
3662 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
3663 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
3664 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
3665 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
3666 0 : return Ok(StatusCode::CONFLICT);
3667 0 : }
3668 0 :
3669 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
3670 0 : return Ok(StatusCode::ACCEPTED);
3671 0 : }
3672 :
3673 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
3674 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
3675 0 : let shard_zero_status = delete_one(
3676 0 : shard_zero_tid,
3677 0 : timeline_id,
3678 0 : shard_zero_locations.latest.node,
3679 0 : self.config.jwt_token.clone(),
3680 0 : )
3681 0 : .await?;
3682 0 : Ok(shard_zero_status)
3683 0 : }).await?
3684 0 : }
3685 :
3686 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
3687 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
3688 0 : pub(crate) async fn tenant_shard0_node(
3689 0 : &self,
3690 0 : tenant_id: TenantId,
3691 0 : ) -> Result<(Node, TenantShardId), ApiError> {
3692 0 : // Look up in-memory state and maybe use the node from there.
3693 0 : {
3694 0 : let locked = self.inner.read().unwrap();
3695 0 : let Some((tenant_shard_id, shard)) = locked
3696 0 : .tenants
3697 0 : .range(TenantShardId::tenant_range(tenant_id))
3698 0 : .next()
3699 : else {
3700 0 : return Err(ApiError::NotFound(
3701 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
3702 0 : ));
3703 : };
3704 :
3705 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
3706 0 : tracing::warn!(
3707 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
3708 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
3709 : shard.policy
3710 : );
3711 0 : return Err(ApiError::Conflict(
3712 0 : "Cannot call timeline API on non-attached tenant".to_string(),
3713 0 : ));
3714 : };
3715 :
3716 0 : if shard.reconciler.is_none() {
3717 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
3718 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
3719 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
3720 : // This should never happen
3721 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3722 0 : "Shard refers to nonexistent node"
3723 0 : )));
3724 : };
3725 0 : return Ok((node.clone(), *tenant_shard_id));
3726 0 : }
3727 : };
3728 :
3729 : // Look up the latest attached pageserver location from the database
3730 : // generation state: this will reflect the progress of any ongoing migration.
3731 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
3732 : // the case where they call through to the pageserver and get a 404.
3733 0 : let db_result = self.persistence.tenant_generations(tenant_id).await?;
3734 : let Some(ShardGenerationState {
3735 0 : tenant_shard_id,
3736 0 : generation: _,
3737 0 : generation_pageserver: Some(node_id),
3738 0 : }) = db_result.first()
3739 : else {
3740 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
3741 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
3742 : // or a conflict response (case where tenant was detached in background)
3743 0 : return Err(ApiError::ResourceUnavailable(
3744 0 : "Shard {} not found in database, or is not attached".into(),
3745 0 : ));
3746 : };
3747 0 : let locked = self.inner.read().unwrap();
3748 0 : let Some(node) = locked.nodes.get(node_id) else {
3749 : // This should never happen
3750 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3751 0 : "Shard refers to nonexistent node"
3752 0 : )));
3753 : };
3754 :
3755 0 : Ok((node.clone(), *tenant_shard_id))
3756 0 : }
3757 :
3758 0 : pub(crate) fn tenant_locate(
3759 0 : &self,
3760 0 : tenant_id: TenantId,
3761 0 : ) -> Result<TenantLocateResponse, ApiError> {
3762 0 : let locked = self.inner.read().unwrap();
3763 0 : tracing::info!("Locating shards for tenant {tenant_id}");
3764 :
3765 0 : let mut result = Vec::new();
3766 0 : let mut shard_params: Option<ShardParameters> = None;
3767 :
3768 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3769 : {
3770 0 : let node_id =
3771 0 : shard
3772 0 : .intent
3773 0 : .get_attached()
3774 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
3775 0 : "Cannot locate a tenant that is not attached"
3776 0 : )))?;
3777 :
3778 0 : let node = locked
3779 0 : .nodes
3780 0 : .get(&node_id)
3781 0 : .expect("Pageservers may not be deleted while referenced");
3782 0 :
3783 0 : result.push(node.shard_location(*tenant_shard_id));
3784 0 :
3785 0 : match &shard_params {
3786 0 : None => {
3787 0 : shard_params = Some(ShardParameters {
3788 0 : stripe_size: shard.shard.stripe_size,
3789 0 : count: shard.shard.count,
3790 0 : });
3791 0 : }
3792 0 : Some(params) => {
3793 0 : if params.stripe_size != shard.shard.stripe_size {
3794 : // This should never happen. We enforce at runtime because it's simpler than
3795 : // adding an extra per-tenant data structure to store the things that should be the same
3796 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3797 0 : "Inconsistent shard stripe size parameters!"
3798 0 : )));
3799 0 : }
3800 : }
3801 : }
3802 : }
3803 :
3804 0 : if result.is_empty() {
3805 0 : return Err(ApiError::NotFound(
3806 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
3807 0 : ));
3808 0 : }
3809 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
3810 0 : tracing::info!(
3811 0 : "Located tenant {} with params {:?} on shards {}",
3812 0 : tenant_id,
3813 0 : shard_params,
3814 0 : result
3815 0 : .iter()
3816 0 : .map(|s| format!("{:?}", s))
3817 0 : .collect::<Vec<_>>()
3818 0 : .join(",")
3819 : );
3820 :
3821 0 : Ok(TenantLocateResponse {
3822 0 : shards: result,
3823 0 : shard_params,
3824 0 : })
3825 0 : }
3826 :
3827 : /// Returns None if the input iterator of shards does not include a shard with number=0
3828 0 : fn tenant_describe_impl<'a>(
3829 0 : &self,
3830 0 : shards: impl Iterator<Item = &'a TenantShard>,
3831 0 : ) -> Option<TenantDescribeResponse> {
3832 0 : let mut shard_zero = None;
3833 0 : let mut describe_shards = Vec::new();
3834 :
3835 0 : for shard in shards {
3836 0 : if shard.tenant_shard_id.is_shard_zero() {
3837 0 : shard_zero = Some(shard);
3838 0 : }
3839 :
3840 0 : describe_shards.push(TenantDescribeResponseShard {
3841 0 : tenant_shard_id: shard.tenant_shard_id,
3842 0 : node_attached: *shard.intent.get_attached(),
3843 0 : node_secondary: shard.intent.get_secondary().to_vec(),
3844 0 : last_error: shard
3845 0 : .last_error
3846 0 : .lock()
3847 0 : .unwrap()
3848 0 : .as_ref()
3849 0 : .map(|e| format!("{e}"))
3850 0 : .unwrap_or("".to_string())
3851 0 : .clone(),
3852 0 : is_reconciling: shard.reconciler.is_some(),
3853 0 : is_pending_compute_notification: shard.pending_compute_notification,
3854 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
3855 0 : scheduling_policy: *shard.get_scheduling_policy(),
3856 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
3857 : })
3858 : }
3859 :
3860 0 : let shard_zero = shard_zero?;
3861 :
3862 0 : Some(TenantDescribeResponse {
3863 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
3864 0 : shards: describe_shards,
3865 0 : stripe_size: shard_zero.shard.stripe_size,
3866 0 : policy: shard_zero.policy.clone(),
3867 0 : config: shard_zero.config.clone(),
3868 0 : })
3869 0 : }
3870 :
3871 0 : pub(crate) fn tenant_describe(
3872 0 : &self,
3873 0 : tenant_id: TenantId,
3874 0 : ) -> Result<TenantDescribeResponse, ApiError> {
3875 0 : let locked = self.inner.read().unwrap();
3876 0 :
3877 0 : self.tenant_describe_impl(
3878 0 : locked
3879 0 : .tenants
3880 0 : .range(TenantShardId::tenant_range(tenant_id))
3881 0 : .map(|(_k, v)| v),
3882 0 : )
3883 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
3884 0 : }
3885 :
3886 0 : pub(crate) fn tenant_list(&self) -> Vec<TenantDescribeResponse> {
3887 0 : let locked = self.inner.read().unwrap();
3888 0 :
3889 0 : let mut result = Vec::new();
3890 0 : for (_tenant_id, tenant_shards) in
3891 0 : &locked.tenants.iter().group_by(|(id, _shard)| id.tenant_id)
3892 0 : {
3893 0 : result.push(
3894 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
3895 0 : .expect("Groups are always non-empty"),
3896 0 : );
3897 0 : }
3898 :
3899 0 : result
3900 0 : }
3901 :
3902 0 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
3903 : async fn abort_tenant_shard_split(
3904 : &self,
3905 : op: &TenantShardSplitAbort,
3906 : ) -> Result<(), TenantShardSplitAbortError> {
3907 : // Cleaning up a split:
3908 : // - Parent shards are not destroyed during a split, just detached.
3909 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
3910 : // just the children attached, or both.
3911 : //
3912 : // Therefore our work to do is to:
3913 : // 1. Clean up storage controller's internal state to just refer to parents, no children
3914 : // 2. Call out to pageservers to ensure that children are detached
3915 : // 3. Call out to pageservers to ensure that parents are attached.
3916 : //
3917 : // Crash safety:
3918 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
3919 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
3920 : // and detach them.
3921 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
3922 : // from our database, then we will re-enter this cleanup routine on startup.
3923 :
3924 : let TenantShardSplitAbort {
3925 : tenant_id,
3926 : new_shard_count,
3927 : new_stripe_size,
3928 : ..
3929 : } = op;
3930 :
3931 : // First abort persistent state, if any exists.
3932 : match self
3933 : .persistence
3934 : .abort_shard_split(*tenant_id, *new_shard_count)
3935 : .await?
3936 : {
3937 : AbortShardSplitStatus::Aborted => {
3938 : // Proceed to roll back any child shards created on pageservers
3939 : }
3940 : AbortShardSplitStatus::Complete => {
3941 : // The split completed (we might hit that path if e.g. our database transaction
3942 : // to write the completion landed in the database, but we dropped connection
3943 : // before seeing the result).
3944 : //
3945 : // We must update in-memory state to reflect the successful split.
3946 : self.tenant_shard_split_commit_inmem(
3947 : *tenant_id,
3948 : *new_shard_count,
3949 : *new_stripe_size,
3950 : );
3951 : return Ok(());
3952 : }
3953 : }
3954 :
3955 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
3956 : let detach_locations: Vec<(Node, TenantShardId)> = {
3957 : let mut detach_locations = Vec::new();
3958 : let mut locked = self.inner.write().unwrap();
3959 : let (nodes, tenants, scheduler) = locked.parts_mut();
3960 :
3961 : for (tenant_shard_id, shard) in
3962 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
3963 : {
3964 : if shard.shard.count == op.new_shard_count {
3965 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
3966 : // is infallible, so if we got an error we shouldn't have got that far.
3967 : tracing::warn!(
3968 : "During split abort, child shard {tenant_shard_id} found in-memory"
3969 : );
3970 : continue;
3971 : }
3972 :
3973 : // Add the children of this shard to this list of things to detach
3974 : if let Some(node_id) = shard.intent.get_attached() {
3975 : for child_id in tenant_shard_id.split(*new_shard_count) {
3976 : detach_locations.push((
3977 : nodes
3978 : .get(node_id)
3979 : .expect("Intent references nonexistent node")
3980 : .clone(),
3981 : child_id,
3982 : ));
3983 : }
3984 : } else {
3985 : tracing::warn!(
3986 : "During split abort, shard {tenant_shard_id} has no attached location"
3987 : );
3988 : }
3989 :
3990 : tracing::info!("Restoring parent shard {tenant_shard_id}");
3991 : shard.splitting = SplitState::Idle;
3992 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
3993 : // If this shard can't be scheduled now (perhaps due to offline nodes or
3994 : // capacity issues), that must not prevent us rolling back a split. In this
3995 : // case it should be eventually scheduled in the background.
3996 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
3997 : }
3998 :
3999 : self.maybe_reconcile_shard(shard, nodes);
4000 : }
4001 :
4002 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
4003 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
4004 :
4005 : detach_locations
4006 : };
4007 :
4008 : for (node, child_id) in detach_locations {
4009 : if !node.is_available() {
4010 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
4011 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
4012 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
4013 : // them from the node.
4014 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
4015 : continue;
4016 : }
4017 :
4018 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
4019 : // a 503 and retry, up to our limit.
4020 : tracing::info!("Detaching {child_id} on {node}...");
4021 : match node
4022 : .with_client_retries(
4023 0 : |client| async move {
4024 0 : let config = LocationConfig {
4025 0 : mode: LocationConfigMode::Detached,
4026 0 : generation: None,
4027 0 : secondary_conf: None,
4028 0 : shard_number: child_id.shard_number.0,
4029 0 : shard_count: child_id.shard_count.literal(),
4030 0 : // Stripe size and tenant config don't matter when detaching
4031 0 : shard_stripe_size: 0,
4032 0 : tenant_conf: TenantConfig::default(),
4033 0 : };
4034 0 :
4035 0 : client.location_config(child_id, config, None, false).await
4036 0 : },
4037 : &self.config.jwt_token,
4038 : 1,
4039 : 10,
4040 : Duration::from_secs(5),
4041 : &self.cancel,
4042 : )
4043 : .await
4044 : {
4045 : Some(Ok(_)) => {}
4046 : Some(Err(e)) => {
4047 : // We failed to communicate with the remote node. This is problematic: we may be
4048 : // leaving it with a rogue child shard.
4049 : tracing::warn!(
4050 : "Failed to detach child {child_id} from node {node} during abort"
4051 : );
4052 : return Err(e.into());
4053 : }
4054 : None => {
4055 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
4056 : // clean up on restart. The node going offline requires a retry.
4057 : return Err(TenantShardSplitAbortError::Unavailable);
4058 : }
4059 : };
4060 : }
4061 :
4062 : tracing::info!("Successfully aborted split");
4063 : Ok(())
4064 : }
4065 :
4066 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
4067 : /// of the tenant map to reflect the child shards that exist after the split.
4068 0 : fn tenant_shard_split_commit_inmem(
4069 0 : &self,
4070 0 : tenant_id: TenantId,
4071 0 : new_shard_count: ShardCount,
4072 0 : new_stripe_size: Option<ShardStripeSize>,
4073 0 : ) -> (
4074 0 : TenantShardSplitResponse,
4075 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
4076 0 : Vec<ReconcilerWaiter>,
4077 0 : ) {
4078 0 : let mut response = TenantShardSplitResponse {
4079 0 : new_shards: Vec::new(),
4080 0 : };
4081 0 : let mut child_locations = Vec::new();
4082 0 : let mut waiters = Vec::new();
4083 0 :
4084 0 : {
4085 0 : let mut locked = self.inner.write().unwrap();
4086 0 :
4087 0 : let parent_ids = locked
4088 0 : .tenants
4089 0 : .range(TenantShardId::tenant_range(tenant_id))
4090 0 : .map(|(shard_id, _)| *shard_id)
4091 0 : .collect::<Vec<_>>();
4092 0 :
4093 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4094 0 : for parent_id in parent_ids {
4095 0 : let child_ids = parent_id.split(new_shard_count);
4096 :
4097 0 : let (pageserver, generation, policy, parent_ident, config) = {
4098 0 : let mut old_state = tenants
4099 0 : .remove(&parent_id)
4100 0 : .expect("It was present, we just split it");
4101 0 :
4102 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
4103 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
4104 0 : // nothing else can clear this.
4105 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
4106 :
4107 0 : let old_attached = old_state.intent.get_attached().unwrap();
4108 0 : old_state.intent.clear(scheduler);
4109 0 : let generation = old_state.generation.expect("Shard must have been attached");
4110 0 : (
4111 0 : old_attached,
4112 0 : generation,
4113 0 : old_state.policy,
4114 0 : old_state.shard,
4115 0 : old_state.config,
4116 0 : )
4117 0 : };
4118 0 :
4119 0 : let mut schedule_context = ScheduleContext::default();
4120 0 : for child in child_ids {
4121 0 : let mut child_shard = parent_ident;
4122 0 : child_shard.number = child.shard_number;
4123 0 : child_shard.count = child.shard_count;
4124 0 : if let Some(stripe_size) = new_stripe_size {
4125 0 : child_shard.stripe_size = stripe_size;
4126 0 : }
4127 :
4128 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
4129 0 : child_observed.insert(
4130 0 : pageserver,
4131 0 : ObservedStateLocation {
4132 0 : conf: Some(attached_location_conf(
4133 0 : generation,
4134 0 : &child_shard,
4135 0 : &config,
4136 0 : &policy,
4137 0 : )),
4138 0 : },
4139 0 : );
4140 0 :
4141 0 : let mut child_state = TenantShard::new(child, child_shard, policy.clone());
4142 0 : child_state.intent = IntentState::single(scheduler, Some(pageserver));
4143 0 : child_state.observed = ObservedState {
4144 0 : locations: child_observed,
4145 0 : };
4146 0 : child_state.generation = Some(generation);
4147 0 : child_state.config = config.clone();
4148 0 :
4149 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
4150 0 : // as at this point in the split process we have succeeded and this part is infallible:
4151 0 : // we will never need to do any special recovery from this state.
4152 0 :
4153 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
4154 :
4155 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
4156 : // This is not fatal, because we've implicitly already got an attached
4157 : // location for the child shard. Failure here just means we couldn't
4158 : // find a secondary (e.g. because cluster is overloaded).
4159 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
4160 0 : }
4161 : // In the background, attach secondary locations for the new shards
4162 0 : if let Some(waiter) = self.maybe_reconcile_shard(&mut child_state, nodes) {
4163 0 : waiters.push(waiter);
4164 0 : }
4165 :
4166 0 : tenants.insert(child, child_state);
4167 0 : response.new_shards.push(child);
4168 : }
4169 : }
4170 0 : (response, child_locations, waiters)
4171 0 : }
4172 0 : }
4173 :
4174 0 : async fn tenant_shard_split_start_secondaries(
4175 0 : &self,
4176 0 : tenant_id: TenantId,
4177 0 : waiters: Vec<ReconcilerWaiter>,
4178 0 : ) {
4179 : // Wait for initial reconcile of child shards, this creates the secondary locations
4180 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
4181 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
4182 : // their secondaries couldn't be attached.
4183 0 : tracing::warn!("Failed to reconcile after split: {e}");
4184 0 : return;
4185 0 : }
4186 :
4187 : // Take the state lock to discover the attached & secondary intents for all shards
4188 0 : let (attached, secondary) = {
4189 0 : let locked = self.inner.read().unwrap();
4190 0 : let mut attached = Vec::new();
4191 0 : let mut secondary = Vec::new();
4192 :
4193 0 : for (tenant_shard_id, shard) in
4194 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4195 : {
4196 0 : let Some(node_id) = shard.intent.get_attached() else {
4197 : // Unexpected. Race with a PlacementPolicy change?
4198 0 : tracing::warn!(
4199 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
4200 : );
4201 0 : continue;
4202 : };
4203 :
4204 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
4205 : // No secondary location. Nothing for us to do.
4206 0 : continue;
4207 : };
4208 :
4209 0 : let attached_node = locked
4210 0 : .nodes
4211 0 : .get(node_id)
4212 0 : .expect("Pageservers may not be deleted while referenced");
4213 0 :
4214 0 : let secondary_node = locked
4215 0 : .nodes
4216 0 : .get(secondary_node_id)
4217 0 : .expect("Pageservers may not be deleted while referenced");
4218 0 :
4219 0 : attached.push((*tenant_shard_id, attached_node.clone()));
4220 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
4221 : }
4222 0 : (attached, secondary)
4223 0 : };
4224 0 :
4225 0 : if secondary.is_empty() {
4226 : // No secondary locations; nothing for us to do
4227 0 : return;
4228 0 : }
4229 :
4230 0 : for result in self
4231 0 : .tenant_for_shards_api(
4232 0 : attached,
4233 0 : |tenant_shard_id, client| async move {
4234 0 : client.tenant_heatmap_upload(tenant_shard_id).await
4235 0 : },
4236 0 : 1,
4237 0 : 1,
4238 0 : SHORT_RECONCILE_TIMEOUT,
4239 0 : &self.cancel,
4240 0 : )
4241 0 : .await
4242 : {
4243 0 : if let Err(e) = result {
4244 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
4245 0 : return;
4246 0 : }
4247 : }
4248 :
4249 0 : for result in self
4250 0 : .tenant_for_shards_api(
4251 0 : secondary,
4252 0 : |tenant_shard_id, client| async move {
4253 0 : client
4254 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
4255 0 : .await
4256 0 : },
4257 0 : 1,
4258 0 : 1,
4259 0 : SHORT_RECONCILE_TIMEOUT,
4260 0 : &self.cancel,
4261 0 : )
4262 0 : .await
4263 : {
4264 0 : if let Err(e) = result {
4265 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
4266 0 : return;
4267 0 : }
4268 : }
4269 0 : }
4270 :
4271 0 : pub(crate) async fn tenant_shard_split(
4272 0 : &self,
4273 0 : tenant_id: TenantId,
4274 0 : split_req: TenantShardSplitRequest,
4275 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
4276 : // TODO: return 503 if we get stuck waiting for this lock
4277 : // (issue https://github.com/neondatabase/neon/issues/7108)
4278 0 : let _tenant_lock = trace_exclusive_lock(
4279 0 : &self.tenant_op_locks,
4280 0 : tenant_id,
4281 0 : TenantOperations::ShardSplit,
4282 0 : )
4283 0 : .await;
4284 :
4285 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
4286 0 : let new_stripe_size = split_req.new_stripe_size;
4287 :
4288 : // Validate the request and construct parameters. This phase is fallible, but does not require
4289 : // rollback on errors, as it does no I/O and mutates no state.
4290 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
4291 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
4292 0 : ShardSplitAction::Split(params) => params,
4293 : };
4294 :
4295 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
4296 : // we must roll back.
4297 0 : let r = self
4298 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
4299 0 : .await;
4300 :
4301 0 : let (response, waiters) = match r {
4302 0 : Ok(r) => r,
4303 0 : Err(e) => {
4304 0 : // Split might be part-done, we must do work to abort it.
4305 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
4306 0 : self.abort_tx
4307 0 : .send(TenantShardSplitAbort {
4308 0 : tenant_id,
4309 0 : new_shard_count,
4310 0 : new_stripe_size,
4311 0 : _tenant_lock,
4312 0 : })
4313 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
4314 0 : .ok();
4315 0 : return Err(e);
4316 : }
4317 : };
4318 :
4319 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
4320 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
4321 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
4322 : // in [`Self::optimize_all`]
4323 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
4324 0 : .await;
4325 0 : Ok(response)
4326 0 : }
4327 :
4328 0 : fn prepare_tenant_shard_split(
4329 0 : &self,
4330 0 : tenant_id: TenantId,
4331 0 : split_req: TenantShardSplitRequest,
4332 0 : ) -> Result<ShardSplitAction, ApiError> {
4333 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
4334 0 : anyhow::anyhow!("failpoint")
4335 0 : )));
4336 :
4337 0 : let mut policy = None;
4338 0 : let mut config = None;
4339 0 : let mut shard_ident = None;
4340 : // Validate input, and calculate which shards we will create
4341 0 : let (old_shard_count, targets) =
4342 : {
4343 0 : let locked = self.inner.read().unwrap();
4344 0 :
4345 0 : let pageservers = locked.nodes.clone();
4346 0 :
4347 0 : let mut targets = Vec::new();
4348 0 :
4349 0 : // In case this is a retry, count how many already-split shards we found
4350 0 : let mut children_found = Vec::new();
4351 0 : let mut old_shard_count = None;
4352 :
4353 0 : for (tenant_shard_id, shard) in
4354 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4355 : {
4356 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
4357 : Ordering::Equal => {
4358 : // Already split this
4359 0 : children_found.push(*tenant_shard_id);
4360 0 : continue;
4361 : }
4362 : Ordering::Greater => {
4363 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4364 0 : "Requested count {} but already have shards at count {}",
4365 0 : split_req.new_shard_count,
4366 0 : shard.shard.count.count()
4367 0 : )));
4368 : }
4369 0 : Ordering::Less => {
4370 0 : // Fall through: this shard has lower count than requested,
4371 0 : // is a candidate for splitting.
4372 0 : }
4373 0 : }
4374 0 :
4375 0 : match old_shard_count {
4376 0 : None => old_shard_count = Some(shard.shard.count),
4377 0 : Some(old_shard_count) => {
4378 0 : if old_shard_count != shard.shard.count {
4379 : // We may hit this case if a caller asked for two splits to
4380 : // different sizes, before the first one is complete.
4381 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
4382 : // of shard_count=1 and shard_count=2 shards in the map.
4383 0 : return Err(ApiError::Conflict(
4384 0 : "Cannot split, currently mid-split".to_string(),
4385 0 : ));
4386 0 : }
4387 : }
4388 : }
4389 0 : if policy.is_none() {
4390 0 : policy = Some(shard.policy.clone());
4391 0 : }
4392 0 : if shard_ident.is_none() {
4393 0 : shard_ident = Some(shard.shard);
4394 0 : }
4395 0 : if config.is_none() {
4396 0 : config = Some(shard.config.clone());
4397 0 : }
4398 :
4399 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4400 0 : tracing::info!(
4401 0 : "Tenant shard {} already has shard count {}",
4402 : tenant_shard_id,
4403 : split_req.new_shard_count
4404 : );
4405 0 : continue;
4406 0 : }
4407 :
4408 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4409 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4410 0 : ))?;
4411 :
4412 0 : let node = pageservers
4413 0 : .get(&node_id)
4414 0 : .expect("Pageservers may not be deleted while referenced");
4415 0 :
4416 0 : targets.push(ShardSplitTarget {
4417 0 : parent_id: *tenant_shard_id,
4418 0 : node: node.clone(),
4419 0 : child_ids: tenant_shard_id
4420 0 : .split(ShardCount::new(split_req.new_shard_count)),
4421 0 : });
4422 : }
4423 :
4424 0 : if targets.is_empty() {
4425 0 : if children_found.len() == split_req.new_shard_count as usize {
4426 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4427 0 : new_shards: children_found,
4428 0 : }));
4429 : } else {
4430 : // No shards found to split, and no existing children found: the
4431 : // tenant doesn't exist at all.
4432 0 : return Err(ApiError::NotFound(
4433 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4434 0 : ));
4435 : }
4436 0 : }
4437 0 :
4438 0 : (old_shard_count, targets)
4439 0 : };
4440 0 :
4441 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
4442 0 : let old_shard_count = old_shard_count.unwrap();
4443 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
4444 : // This ShardIdentity will be used as the template for all children, so this implicitly
4445 : // applies the new stripe size to the children.
4446 0 : let mut shard_ident = shard_ident.unwrap();
4447 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
4448 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)));
4449 0 : }
4450 0 :
4451 0 : shard_ident.stripe_size = new_stripe_size;
4452 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
4453 0 : shard_ident
4454 : } else {
4455 0 : shard_ident.unwrap()
4456 : };
4457 0 : let policy = policy.unwrap();
4458 0 : let config = config.unwrap();
4459 0 :
4460 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
4461 0 : old_shard_count,
4462 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
4463 0 : new_stripe_size: split_req.new_stripe_size,
4464 0 : targets,
4465 0 : policy,
4466 0 : config,
4467 0 : shard_ident,
4468 0 : })))
4469 0 : }
4470 :
4471 0 : async fn do_tenant_shard_split(
4472 0 : &self,
4473 0 : tenant_id: TenantId,
4474 0 : params: Box<ShardSplitParams>,
4475 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
4476 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
4477 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
4478 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
4479 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
4480 0 : // (https://github.com/neondatabase/neon/issues/6676)
4481 0 :
4482 0 : let ShardSplitParams {
4483 0 : old_shard_count,
4484 0 : new_shard_count,
4485 0 : new_stripe_size,
4486 0 : mut targets,
4487 0 : policy,
4488 0 : config,
4489 0 : shard_ident,
4490 0 : } = *params;
4491 :
4492 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
4493 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
4494 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
4495 : // at the time of split.
4496 0 : let waiters = {
4497 0 : let mut locked = self.inner.write().unwrap();
4498 0 : let mut waiters = Vec::new();
4499 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4500 0 : for target in &mut targets {
4501 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
4502 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4503 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4504 0 : "Shard {} not found",
4505 0 : target.parent_id
4506 0 : )));
4507 : };
4508 :
4509 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
4510 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4511 0 : return Err(ApiError::Conflict(format!(
4512 0 : "Shard {} unexpectedly rescheduled during split",
4513 0 : target.parent_id
4514 0 : )));
4515 0 : }
4516 0 :
4517 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
4518 0 : shard.intent.clear_secondary(scheduler);
4519 :
4520 : // Run Reconciler to execute detach fo secondary locations.
4521 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
4522 0 : waiters.push(waiter);
4523 0 : }
4524 : }
4525 0 : waiters
4526 0 : };
4527 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
4528 :
4529 : // Before creating any new child shards in memory or on the pageservers, persist them: this
4530 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
4531 : // acts as the protection against two concurrent attempts to split: one of them will get a database
4532 : // error trying to insert the child shards.
4533 0 : let mut child_tsps = Vec::new();
4534 0 : for target in &targets {
4535 0 : let mut this_child_tsps = Vec::new();
4536 0 : for child in &target.child_ids {
4537 0 : let mut child_shard = shard_ident;
4538 0 : child_shard.number = child.shard_number;
4539 0 : child_shard.count = child.shard_count;
4540 0 :
4541 0 : tracing::info!(
4542 0 : "Create child shard persistence with stripe size {}",
4543 : shard_ident.stripe_size.0
4544 : );
4545 :
4546 0 : this_child_tsps.push(TenantShardPersistence {
4547 0 : tenant_id: child.tenant_id.to_string(),
4548 0 : shard_number: child.shard_number.0 as i32,
4549 0 : shard_count: child.shard_count.literal() as i32,
4550 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
4551 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
4552 0 : // populate the correct generation as part of its transaction, to protect us
4553 0 : // against racing with changes in the state of the parent.
4554 0 : generation: None,
4555 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
4556 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
4557 0 : config: serde_json::to_string(&config).unwrap(),
4558 0 : splitting: SplitState::Splitting,
4559 0 :
4560 0 : // Scheduling policies and preferred AZ do not carry through to children
4561 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
4562 0 : .unwrap(),
4563 0 : preferred_az_id: None,
4564 0 : });
4565 : }
4566 :
4567 0 : child_tsps.push((target.parent_id, this_child_tsps));
4568 : }
4569 :
4570 0 : if let Err(e) = self
4571 0 : .persistence
4572 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
4573 0 : .await
4574 : {
4575 0 : match e {
4576 : DatabaseError::Query(diesel::result::Error::DatabaseError(
4577 : DatabaseErrorKind::UniqueViolation,
4578 : _,
4579 : )) => {
4580 : // Inserting a child shard violated a unique constraint: we raced with another call to
4581 : // this function
4582 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
4583 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
4584 : }
4585 0 : _ => return Err(ApiError::InternalServerError(e.into())),
4586 : }
4587 0 : }
4588 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
4589 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4590 0 : ));
4591 :
4592 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
4593 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
4594 : // is not set in memory, then it was not persisted.
4595 : {
4596 0 : let mut locked = self.inner.write().unwrap();
4597 0 : for target in &targets {
4598 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
4599 0 : parent_shard.splitting = SplitState::Splitting;
4600 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
4601 0 : // split operation.
4602 0 : parent_shard
4603 0 : .observed
4604 0 : .locations
4605 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
4606 0 : }
4607 : }
4608 : }
4609 :
4610 : // TODO: issue split calls concurrently (this only matters once we're splitting
4611 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
4612 :
4613 0 : for target in &targets {
4614 : let ShardSplitTarget {
4615 0 : parent_id,
4616 0 : node,
4617 0 : child_ids,
4618 0 : } = target;
4619 0 : let client = PageserverClient::new(
4620 0 : node.get_id(),
4621 0 : node.base_url(),
4622 0 : self.config.jwt_token.as_deref(),
4623 0 : );
4624 0 : let response = client
4625 0 : .tenant_shard_split(
4626 0 : *parent_id,
4627 0 : TenantShardSplitRequest {
4628 0 : new_shard_count: new_shard_count.literal(),
4629 0 : new_stripe_size,
4630 0 : },
4631 0 : )
4632 0 : .await
4633 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
4634 :
4635 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
4636 0 : "failpoint".to_string()
4637 0 : )));
4638 :
4639 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
4640 :
4641 0 : tracing::info!(
4642 0 : "Split {} into {}",
4643 0 : parent_id,
4644 0 : response
4645 0 : .new_shards
4646 0 : .iter()
4647 0 : .map(|s| format!("{:?}", s))
4648 0 : .collect::<Vec<_>>()
4649 0 : .join(",")
4650 : );
4651 :
4652 0 : if &response.new_shards != child_ids {
4653 : // This should never happen: the pageserver should agree with us on how shard splits work.
4654 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4655 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
4656 0 : parent_id,
4657 0 : response.new_shards,
4658 0 : child_ids
4659 0 : )));
4660 0 : }
4661 : }
4662 :
4663 : // TODO: if the pageserver restarted concurrently with our split API call,
4664 : // the actual generation of the child shard might differ from the generation
4665 : // we expect it to have. In order for our in-database generation to end up
4666 : // correct, we should carry the child generation back in the response and apply it here
4667 : // in complete_shard_split (and apply the correct generation in memory)
4668 : // (or, we can carry generation in the request and reject the request if
4669 : // it doesn't match, but that requires more retry logic on this side)
4670 :
4671 0 : self.persistence
4672 0 : .complete_shard_split(tenant_id, old_shard_count)
4673 0 : .await?;
4674 :
4675 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
4676 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4677 0 : ));
4678 :
4679 : // Replace all the shards we just split with their children: this phase is infallible.
4680 0 : let (response, child_locations, waiters) =
4681 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
4682 0 :
4683 0 : // Now that we have scheduled the child shards, attempt to set their preferred AZ
4684 0 : // to that of the pageserver they've been attached on.
4685 0 : let preferred_azs = {
4686 0 : let locked = self.inner.read().unwrap();
4687 0 : child_locations
4688 0 : .iter()
4689 0 : .filter_map(|(tid, node_id, _stripe_size)| {
4690 0 : let az_id = locked
4691 0 : .nodes
4692 0 : .get(node_id)
4693 0 : .map(|n| n.get_availability_zone_id().clone())?;
4694 :
4695 0 : Some((*tid, az_id))
4696 0 : })
4697 0 : .collect::<Vec<_>>()
4698 : };
4699 :
4700 0 : let updated = self
4701 0 : .persistence
4702 0 : .set_tenant_shard_preferred_azs(preferred_azs)
4703 0 : .await
4704 0 : .map_err(|err| {
4705 0 : ApiError::InternalServerError(anyhow::anyhow!(
4706 0 : "Failed to persist preferred az ids: {err}"
4707 0 : ))
4708 0 : });
4709 0 :
4710 0 : match updated {
4711 0 : Ok(updated) => {
4712 0 : let mut locked = self.inner.write().unwrap();
4713 0 : for (tid, az_id) in updated {
4714 0 : if let Some(shard) = locked.tenants.get_mut(&tid) {
4715 0 : shard.set_preferred_az(az_id);
4716 0 : }
4717 : }
4718 : }
4719 0 : Err(err) => {
4720 0 : tracing::warn!("Failed to persist preferred AZs after split: {err}");
4721 : }
4722 : }
4723 :
4724 : // Send compute notifications for all the new shards
4725 0 : let mut failed_notifications = Vec::new();
4726 0 : for (child_id, child_ps, stripe_size) in child_locations {
4727 0 : if let Err(e) = self
4728 0 : .compute_hook
4729 0 : .notify(child_id, child_ps, stripe_size, &self.cancel)
4730 0 : .await
4731 : {
4732 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
4733 : child_id, child_ps);
4734 0 : failed_notifications.push(child_id);
4735 0 : }
4736 : }
4737 :
4738 : // If we failed any compute notifications, make a note to retry later.
4739 0 : if !failed_notifications.is_empty() {
4740 0 : let mut locked = self.inner.write().unwrap();
4741 0 : for failed in failed_notifications {
4742 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
4743 0 : shard.pending_compute_notification = true;
4744 0 : }
4745 : }
4746 0 : }
4747 :
4748 0 : Ok((response, waiters))
4749 0 : }
4750 :
4751 0 : pub(crate) async fn tenant_shard_migrate(
4752 0 : &self,
4753 0 : tenant_shard_id: TenantShardId,
4754 0 : migrate_req: TenantShardMigrateRequest,
4755 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
4756 0 : let waiter = {
4757 0 : let mut locked = self.inner.write().unwrap();
4758 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4759 :
4760 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
4761 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4762 0 : "Node {} not found",
4763 0 : migrate_req.node_id
4764 0 : )));
4765 : };
4766 :
4767 0 : if !node.is_available() {
4768 : // Warn but proceed: the caller may intend to manually adjust the placement of
4769 : // a shard even if the node is down, e.g. if intervening during an incident.
4770 0 : tracing::warn!("Migrating to unavailable node {node}");
4771 0 : }
4772 :
4773 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
4774 0 : return Err(ApiError::NotFound(
4775 0 : anyhow::anyhow!("Tenant shard not found").into(),
4776 0 : ));
4777 : };
4778 :
4779 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
4780 : // No-op case: we will still proceed to wait for reconciliation in case it is
4781 : // incomplete from an earlier update to the intent.
4782 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
4783 : } else {
4784 0 : let old_attached = *shard.intent.get_attached();
4785 0 :
4786 0 : match shard.policy {
4787 0 : PlacementPolicy::Attached(n) => {
4788 0 : // If our new attached node was a secondary, it no longer should be.
4789 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
4790 :
4791 : // If we were already attached to something, demote that to a secondary
4792 0 : if let Some(old_attached) = old_attached {
4793 0 : if n > 0 {
4794 : // Remove other secondaries to make room for the location we'll demote
4795 0 : while shard.intent.get_secondary().len() >= n {
4796 0 : shard.intent.pop_secondary(scheduler);
4797 0 : }
4798 :
4799 0 : shard.intent.push_secondary(scheduler, old_attached);
4800 0 : }
4801 0 : }
4802 :
4803 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
4804 : }
4805 0 : PlacementPolicy::Secondary => {
4806 0 : shard.intent.clear(scheduler);
4807 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
4808 0 : }
4809 : PlacementPolicy::Detached => {
4810 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4811 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
4812 0 : )))
4813 : }
4814 : }
4815 :
4816 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
4817 0 : shard.sequence = shard.sequence.next();
4818 : }
4819 :
4820 0 : self.maybe_reconcile_shard(shard, nodes)
4821 : };
4822 :
4823 0 : if let Some(waiter) = waiter {
4824 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
4825 : } else {
4826 0 : tracing::info!("Migration is a no-op");
4827 : }
4828 :
4829 0 : Ok(TenantShardMigrateResponse {})
4830 0 : }
4831 :
4832 : /// This is for debug/support only: we simply drop all state for a tenant, without
4833 : /// detaching or deleting it on pageservers.
4834 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
4835 0 : self.persistence.delete_tenant(tenant_id).await?;
4836 :
4837 0 : let mut locked = self.inner.write().unwrap();
4838 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
4839 0 : let mut shards = Vec::new();
4840 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
4841 0 : shards.push(*tenant_shard_id);
4842 0 : }
4843 :
4844 0 : for shard_id in shards {
4845 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
4846 0 : shard.intent.clear(scheduler);
4847 0 : }
4848 : }
4849 :
4850 0 : Ok(())
4851 0 : }
4852 :
4853 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
4854 : /// tenant with a very high generation number so that it will see the existing data.
4855 0 : pub(crate) async fn tenant_import(
4856 0 : &self,
4857 0 : tenant_id: TenantId,
4858 0 : ) -> Result<TenantCreateResponse, ApiError> {
4859 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
4860 0 : let maybe_node = {
4861 0 : self.inner
4862 0 : .read()
4863 0 : .unwrap()
4864 0 : .nodes
4865 0 : .values()
4866 0 : .find(|n| n.is_available())
4867 0 : .cloned()
4868 : };
4869 0 : let Some(node) = maybe_node else {
4870 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
4871 : };
4872 :
4873 0 : let client = PageserverClient::new(
4874 0 : node.get_id(),
4875 0 : node.base_url(),
4876 0 : self.config.jwt_token.as_deref(),
4877 0 : );
4878 :
4879 0 : let scan_result = client
4880 0 : .tenant_scan_remote_storage(tenant_id)
4881 0 : .await
4882 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4883 :
4884 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
4885 0 : let Some(shard_count) = scan_result
4886 0 : .shards
4887 0 : .iter()
4888 0 : .map(|s| s.tenant_shard_id.shard_count)
4889 0 : .max()
4890 : else {
4891 0 : return Err(ApiError::NotFound(
4892 0 : anyhow::anyhow!("No shards found").into(),
4893 0 : ));
4894 : };
4895 :
4896 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
4897 : // to
4898 0 : let generation = scan_result
4899 0 : .shards
4900 0 : .iter()
4901 0 : .map(|s| s.generation)
4902 0 : .max()
4903 0 : .expect("We already validated >0 shards");
4904 0 :
4905 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
4906 0 : // only work if they were using the default stripe size.
4907 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
4908 :
4909 0 : let (response, waiters) = self
4910 0 : .do_tenant_create(TenantCreateRequest {
4911 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
4912 0 : generation,
4913 0 :
4914 0 : shard_parameters: ShardParameters {
4915 0 : count: shard_count,
4916 0 : stripe_size,
4917 0 : },
4918 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
4919 0 :
4920 0 : // There is no way to know what the tenant's config was: revert to defaults
4921 0 : //
4922 0 : // TODO: remove `switch_aux_file_policy` once we finish auxv2 migration
4923 0 : //
4924 0 : // we write to both v1+v2 storage, so that the test case can use either storage format for testing
4925 0 : config: TenantConfig {
4926 0 : switch_aux_file_policy: Some(models::AuxFilePolicy::CrossValidation),
4927 0 : ..TenantConfig::default()
4928 0 : },
4929 0 : })
4930 0 : .await?;
4931 :
4932 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
4933 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
4934 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
4935 : // reconcile, as reconciliation includes notifying compute.
4936 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
4937 0 : }
4938 :
4939 0 : Ok(response)
4940 0 : }
4941 :
4942 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
4943 : /// we don't have to make TenantShard clonable in the return path.
4944 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4945 0 : let serialized = {
4946 0 : let locked = self.inner.read().unwrap();
4947 0 : let result = locked.tenants.values().collect::<Vec<_>>();
4948 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
4949 : };
4950 :
4951 0 : hyper::Response::builder()
4952 0 : .status(hyper::StatusCode::OK)
4953 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4954 0 : .body(hyper::Body::from(serialized))
4955 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4956 0 : }
4957 :
4958 : /// Check the consistency of in-memory state vs. persistent state, and check that the
4959 : /// scheduler's statistics are up to date.
4960 : ///
4961 : /// These consistency checks expect an **idle** system. If changes are going on while
4962 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
4963 : /// checks, but not suitable for running continuously in the background in the field.
4964 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
4965 0 : let (mut expect_nodes, mut expect_shards) = {
4966 0 : let locked = self.inner.read().unwrap();
4967 0 :
4968 0 : locked
4969 0 : .scheduler
4970 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
4971 0 : .context("Scheduler checks")
4972 0 : .map_err(ApiError::InternalServerError)?;
4973 :
4974 0 : let expect_nodes = locked
4975 0 : .nodes
4976 0 : .values()
4977 0 : .map(|n| n.to_persistent())
4978 0 : .collect::<Vec<_>>();
4979 0 :
4980 0 : let expect_shards = locked
4981 0 : .tenants
4982 0 : .values()
4983 0 : .map(|t| t.to_persistent())
4984 0 : .collect::<Vec<_>>();
4985 :
4986 : // This method can only validate the state of an idle system: if a reconcile is in
4987 : // progress, fail out early to avoid giving false errors on state that won't match
4988 : // between database and memory under a ReconcileResult is processed.
4989 0 : for t in locked.tenants.values() {
4990 0 : if t.reconciler.is_some() {
4991 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4992 0 : "Shard {} reconciliation in progress",
4993 0 : t.tenant_shard_id
4994 0 : )));
4995 0 : }
4996 : }
4997 :
4998 0 : (expect_nodes, expect_shards)
4999 : };
5000 :
5001 0 : let mut nodes = self.persistence.list_nodes().await?;
5002 0 : expect_nodes.sort_by_key(|n| n.node_id);
5003 0 : nodes.sort_by_key(|n| n.node_id);
5004 0 :
5005 0 : if nodes != expect_nodes {
5006 0 : tracing::error!("Consistency check failed on nodes.");
5007 0 : tracing::error!(
5008 0 : "Nodes in memory: {}",
5009 0 : serde_json::to_string(&expect_nodes)
5010 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5011 : );
5012 0 : tracing::error!(
5013 0 : "Nodes in database: {}",
5014 0 : serde_json::to_string(&nodes)
5015 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5016 : );
5017 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5018 0 : "Node consistency failure"
5019 0 : )));
5020 0 : }
5021 :
5022 0 : let mut shards = self.persistence.list_tenant_shards().await?;
5023 0 : shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5024 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5025 0 :
5026 0 : if shards != expect_shards {
5027 0 : tracing::error!("Consistency check failed on shards.");
5028 0 : tracing::error!(
5029 0 : "Shards in memory: {}",
5030 0 : serde_json::to_string(&expect_shards)
5031 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5032 : );
5033 0 : tracing::error!(
5034 0 : "Shards in database: {}",
5035 0 : serde_json::to_string(&shards)
5036 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5037 : );
5038 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5039 0 : "Shard consistency failure"
5040 0 : )));
5041 0 : }
5042 0 :
5043 0 : Ok(())
5044 0 : }
5045 :
5046 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
5047 : /// we don't have to make TenantShard clonable in the return path.
5048 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5049 0 : let serialized = {
5050 0 : let locked = self.inner.read().unwrap();
5051 0 : serde_json::to_string(&locked.scheduler)
5052 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5053 : };
5054 :
5055 0 : hyper::Response::builder()
5056 0 : .status(hyper::StatusCode::OK)
5057 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5058 0 : .body(hyper::Body::from(serialized))
5059 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5060 0 : }
5061 :
5062 : /// This is for debug/support only: we simply drop all state for a tenant, without
5063 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
5064 : /// tenants that were on this node.
5065 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
5066 0 : self.persistence.delete_node(node_id).await?;
5067 :
5068 0 : let mut locked = self.inner.write().unwrap();
5069 :
5070 0 : for shard in locked.tenants.values_mut() {
5071 0 : shard.deref_node(node_id);
5072 0 : shard.observed.locations.remove(&node_id);
5073 0 : }
5074 :
5075 0 : let mut nodes = (*locked.nodes).clone();
5076 0 : nodes.remove(&node_id);
5077 0 : locked.nodes = Arc::new(nodes);
5078 0 :
5079 0 : locked.scheduler.node_remove(node_id);
5080 0 :
5081 0 : Ok(())
5082 0 : }
5083 :
5084 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
5085 : /// that we don't leave any bad state behind in the storage controller, but unclean
5086 : /// in the sense that we are not carefully draining the node.
5087 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
5088 0 : let _node_lock =
5089 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
5090 :
5091 : // 1. Atomically update in-memory state:
5092 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
5093 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
5094 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
5095 : // re-insert references to this node into the ObservedState of shards
5096 : // - drop the node from the scheduler
5097 : {
5098 0 : let mut locked = self.inner.write().unwrap();
5099 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5100 0 :
5101 0 : {
5102 0 : let mut nodes_mut = (*nodes).deref().clone();
5103 0 : match nodes_mut.get_mut(&node_id) {
5104 0 : Some(node) => {
5105 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
5106 0 : // if we crash it would not be desirable to leave the node paused after a restart.
5107 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
5108 0 : }
5109 : None => {
5110 0 : tracing::info!(
5111 0 : "Node not found: presuming this is a retry and returning success"
5112 : );
5113 0 : return Ok(());
5114 : }
5115 : }
5116 :
5117 0 : *nodes = Arc::new(nodes_mut);
5118 : }
5119 :
5120 0 : for (tenant_shard_id, shard) in tenants {
5121 0 : if shard.deref_node(node_id) {
5122 : // FIXME: we need to build a ScheduleContext that reflects this shard's peers, otherwise
5123 : // it won't properly do anti-affinity.
5124 0 : let mut schedule_context = ScheduleContext::default();
5125 :
5126 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
5127 : // TODO: implement force flag to remove a node even if we can't reschedule
5128 : // a tenant
5129 0 : tracing::error!("Refusing to delete node, shard {tenant_shard_id} can't be rescheduled: {e}");
5130 0 : return Err(e.into());
5131 : } else {
5132 0 : tracing::info!(
5133 0 : "Rescheduled shard {tenant_shard_id} away from node during deletion"
5134 : )
5135 : }
5136 :
5137 0 : self.maybe_reconcile_shard(shard, nodes);
5138 0 : }
5139 :
5140 : // Here we remove an existing observed location for the node we're removing, and it will
5141 : // not be re-added by a reconciler's completion because we filter out removed nodes in
5142 : // process_result.
5143 : //
5144 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
5145 : // means any reconciles we spawned will know about the node we're deleting, enabling them
5146 : // to do live migrations if it's still online.
5147 0 : shard.observed.locations.remove(&node_id);
5148 : }
5149 :
5150 0 : scheduler.node_remove(node_id);
5151 0 :
5152 0 : {
5153 0 : let mut nodes_mut = (**nodes).clone();
5154 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
5155 0 : // Ensure that any reconciler holding an Arc<> to this node will
5156 0 : // drop out when trying to RPC to it (setting Offline state sets the
5157 0 : // cancellation token on the Node object).
5158 0 : removed_node.set_availability(NodeAvailability::Offline);
5159 0 : }
5160 0 : *nodes = Arc::new(nodes_mut);
5161 0 : }
5162 0 : }
5163 0 :
5164 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
5165 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
5166 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
5167 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
5168 0 : // that exists.
5169 0 :
5170 0 : // 2. Actually delete the node from the database and from in-memory state
5171 0 : tracing::info!("Deleting node from database");
5172 0 : self.persistence.delete_node(node_id).await?;
5173 :
5174 0 : Ok(())
5175 0 : }
5176 :
5177 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
5178 0 : let nodes = {
5179 0 : self.inner
5180 0 : .read()
5181 0 : .unwrap()
5182 0 : .nodes
5183 0 : .values()
5184 0 : .cloned()
5185 0 : .collect::<Vec<_>>()
5186 0 : };
5187 0 :
5188 0 : Ok(nodes)
5189 0 : }
5190 :
5191 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
5192 0 : self.inner
5193 0 : .read()
5194 0 : .unwrap()
5195 0 : .nodes
5196 0 : .get(&node_id)
5197 0 : .cloned()
5198 0 : .ok_or(ApiError::NotFound(
5199 0 : format!("Node {node_id} not registered").into(),
5200 0 : ))
5201 0 : }
5202 :
5203 0 : pub(crate) async fn get_node_shards(
5204 0 : &self,
5205 0 : node_id: NodeId,
5206 0 : ) -> Result<NodeShardResponse, ApiError> {
5207 0 : let locked = self.inner.read().unwrap();
5208 0 : let mut shards = Vec::new();
5209 0 : for (tid, tenant) in locked.tenants.iter() {
5210 0 : let is_intended_secondary = match (
5211 0 : tenant.intent.get_attached() == &Some(node_id),
5212 0 : tenant.intent.get_secondary().contains(&node_id),
5213 0 : ) {
5214 : (true, true) => {
5215 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5216 0 : "{} attached as primary+secondary on the same node",
5217 0 : tid
5218 0 : )))
5219 : }
5220 0 : (true, false) => Some(false),
5221 0 : (false, true) => Some(true),
5222 0 : (false, false) => None,
5223 : };
5224 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
5225 0 : tenant.observed.locations.get(&node_id)
5226 : {
5227 0 : Some(conf.secondary_conf.is_some())
5228 : } else {
5229 0 : None
5230 : };
5231 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
5232 0 : shards.push(NodeShard {
5233 0 : tenant_shard_id: *tid,
5234 0 : is_intended_secondary,
5235 0 : is_observed_secondary,
5236 0 : });
5237 0 : }
5238 : }
5239 0 : Ok(NodeShardResponse { node_id, shards })
5240 0 : }
5241 :
5242 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
5243 0 : self.persistence.get_leader().await
5244 0 : }
5245 :
5246 0 : pub(crate) async fn node_register(
5247 0 : &self,
5248 0 : register_req: NodeRegisterRequest,
5249 0 : ) -> Result<(), ApiError> {
5250 0 : let _node_lock = trace_exclusive_lock(
5251 0 : &self.node_op_locks,
5252 0 : register_req.node_id,
5253 0 : NodeOperations::Register,
5254 0 : )
5255 0 : .await;
5256 :
5257 : enum RegistrationStatus {
5258 : Matched,
5259 : Mismatched,
5260 : New,
5261 : }
5262 :
5263 0 : let registration_status = {
5264 0 : let locked = self.inner.read().unwrap();
5265 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
5266 0 : if node.registration_match(®ister_req) {
5267 0 : RegistrationStatus::Matched
5268 : } else {
5269 0 : RegistrationStatus::Mismatched
5270 : }
5271 : } else {
5272 0 : RegistrationStatus::New
5273 : }
5274 : };
5275 :
5276 0 : match registration_status {
5277 : RegistrationStatus::Matched => {
5278 0 : tracing::info!(
5279 0 : "Node {} re-registered with matching address",
5280 : register_req.node_id
5281 : );
5282 :
5283 0 : return Ok(());
5284 : }
5285 : RegistrationStatus::Mismatched => {
5286 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
5287 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
5288 : // a fixed address through the lifetime of a node.
5289 0 : tracing::warn!(
5290 0 : "Node {} tried to register with different address",
5291 : register_req.node_id
5292 : );
5293 0 : return Err(ApiError::Conflict(
5294 0 : "Node is already registered with different address".to_string(),
5295 0 : ));
5296 : }
5297 0 : RegistrationStatus::New => {
5298 0 : // fallthrough
5299 0 : }
5300 0 : }
5301 0 :
5302 0 : // We do not require that a node is actually online when registered (it will start life
5303 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
5304 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
5305 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
5306 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
5307 0 : if tokio::net::lookup_host(format!(
5308 0 : "{}:{}",
5309 0 : register_req.listen_http_addr, register_req.listen_http_port
5310 0 : ))
5311 0 : .await
5312 0 : .is_err()
5313 : {
5314 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
5315 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
5316 : // we return a soft 503 error, to encourage callers to retry past transient issues.
5317 0 : return Err(ApiError::ResourceUnavailable(
5318 0 : format!(
5319 0 : "Node {} tried to register with unknown DNS name '{}'",
5320 0 : register_req.node_id, register_req.listen_http_addr
5321 0 : )
5322 0 : .into(),
5323 0 : ));
5324 0 : }
5325 0 :
5326 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
5327 0 : // This ensures that before we use it for anything or expose it via any external
5328 0 : // API, it is guaranteed to be available after a restart.
5329 0 : let new_node = Node::new(
5330 0 : register_req.node_id,
5331 0 : register_req.listen_http_addr,
5332 0 : register_req.listen_http_port,
5333 0 : register_req.listen_pg_addr,
5334 0 : register_req.listen_pg_port,
5335 0 : register_req.availability_zone_id,
5336 0 : );
5337 0 :
5338 0 : // TODO: idempotency if the node already exists in the database
5339 0 : self.persistence.insert_node(&new_node).await?;
5340 :
5341 0 : let mut locked = self.inner.write().unwrap();
5342 0 : let mut new_nodes = (*locked.nodes).clone();
5343 0 :
5344 0 : locked.scheduler.node_upsert(&new_node);
5345 0 : new_nodes.insert(register_req.node_id, new_node);
5346 0 :
5347 0 : locked.nodes = Arc::new(new_nodes);
5348 0 :
5349 0 : tracing::info!(
5350 0 : "Registered pageserver {}, now have {} pageservers",
5351 0 : register_req.node_id,
5352 0 : locked.nodes.len()
5353 : );
5354 0 : Ok(())
5355 0 : }
5356 :
5357 : /// Configure in-memory and persistent state of a node as requested
5358 : ///
5359 : /// Note that this function does not trigger any immediate side effects in response
5360 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
5361 0 : async fn node_state_configure(
5362 0 : &self,
5363 0 : node_id: NodeId,
5364 0 : availability: Option<NodeAvailability>,
5365 0 : scheduling: Option<NodeSchedulingPolicy>,
5366 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
5367 0 : ) -> Result<AvailabilityTransition, ApiError> {
5368 0 : if let Some(scheduling) = scheduling {
5369 : // Scheduling is a persistent part of Node: we must write updates to the database before
5370 : // applying them in memory
5371 0 : self.persistence.update_node(node_id, scheduling).await?;
5372 0 : }
5373 :
5374 : // If we're activating a node, then before setting it active we must reconcile any shard locations
5375 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
5376 : // by calling [`Self::node_activate_reconcile`]
5377 : //
5378 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
5379 : // nothing else can mutate its availability while we run.
5380 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
5381 0 : let (activate_node, availability_transition) = {
5382 0 : let locked = self.inner.read().unwrap();
5383 0 : let Some(node) = locked.nodes.get(&node_id) else {
5384 0 : return Err(ApiError::NotFound(
5385 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5386 0 : ));
5387 : };
5388 :
5389 0 : (
5390 0 : node.clone(),
5391 0 : node.get_availability_transition(input_availability),
5392 0 : )
5393 : };
5394 :
5395 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
5396 0 : self.node_activate_reconcile(activate_node, node_lock)
5397 0 : .await?;
5398 0 : }
5399 0 : availability_transition
5400 : } else {
5401 0 : AvailabilityTransition::Unchanged
5402 : };
5403 :
5404 : // Apply changes from the request to our in-memory state for the Node
5405 0 : let mut locked = self.inner.write().unwrap();
5406 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
5407 0 :
5408 0 : let mut new_nodes = (**nodes).clone();
5409 :
5410 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
5411 0 : return Err(ApiError::NotFound(
5412 0 : anyhow::anyhow!("Node not registered").into(),
5413 0 : ));
5414 : };
5415 :
5416 0 : if let Some(availability) = availability {
5417 0 : node.set_availability(availability);
5418 0 : }
5419 :
5420 0 : if let Some(scheduling) = scheduling {
5421 0 : node.set_scheduling(scheduling);
5422 0 : }
5423 :
5424 : // Update the scheduler, in case the elegibility of the node for new shards has changed
5425 0 : scheduler.node_upsert(node);
5426 0 :
5427 0 : let new_nodes = Arc::new(new_nodes);
5428 0 : locked.nodes = new_nodes;
5429 0 :
5430 0 : Ok(availability_transition)
5431 0 : }
5432 :
5433 : /// Handle availability transition of one node
5434 : ///
5435 : /// Note that you should first call [`Self::node_state_configure`] to update
5436 : /// the in-memory state referencing that node. If you need to handle more than one transition
5437 : /// consider using [`Self::handle_node_availability_transitions`].
5438 0 : async fn handle_node_availability_transition(
5439 0 : &self,
5440 0 : node_id: NodeId,
5441 0 : transition: AvailabilityTransition,
5442 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
5443 0 : ) -> Result<(), ApiError> {
5444 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
5445 0 : match transition {
5446 : AvailabilityTransition::ToOffline => {
5447 0 : tracing::info!("Node {} transition to offline", node_id);
5448 :
5449 0 : let mut locked = self.inner.write().unwrap();
5450 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5451 0 :
5452 0 : let mut tenants_affected: usize = 0;
5453 :
5454 0 : for (tenant_shard_id, tenant_shard) in tenants {
5455 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
5456 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
5457 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
5458 0 : observed_loc.conf = None;
5459 0 : }
5460 :
5461 0 : if nodes.len() == 1 {
5462 : // Special case for single-node cluster: there is no point trying to reschedule
5463 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
5464 : // failures to schedule them.
5465 0 : continue;
5466 0 : }
5467 0 :
5468 0 : if !nodes
5469 0 : .values()
5470 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
5471 : {
5472 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
5473 : // trying to reschedule since there's nowhere else to go. Without this
5474 : // branch we incorrectly detach tenants in response to node unavailability.
5475 0 : continue;
5476 0 : }
5477 0 :
5478 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
5479 0 : tenant_shard.sequence = tenant_shard.sequence.next();
5480 0 :
5481 0 : // TODO: populate a ScheduleContext including all shards in the same tenant_id (only matters
5482 0 : // for tenants without secondary locations: if they have a secondary location, then this
5483 0 : // schedule() call is just promoting an existing secondary)
5484 0 : let mut schedule_context = ScheduleContext::default();
5485 0 :
5486 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
5487 0 : Err(e) => {
5488 0 : // It is possible that some tenants will become unschedulable when too many pageservers
5489 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
5490 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
5491 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
5492 : }
5493 : Ok(()) => {
5494 0 : if self.maybe_reconcile_shard(tenant_shard, nodes).is_some() {
5495 0 : tenants_affected += 1;
5496 0 : };
5497 : }
5498 : }
5499 0 : }
5500 : }
5501 0 : tracing::info!(
5502 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
5503 : tenants_affected,
5504 : node_id
5505 : )
5506 : }
5507 : AvailabilityTransition::ToActive => {
5508 0 : tracing::info!("Node {} transition to active", node_id);
5509 :
5510 0 : let mut locked = self.inner.write().unwrap();
5511 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
5512 :
5513 : // When a node comes back online, we must reconcile any tenant that has a None observed
5514 : // location on the node.
5515 0 : for tenant_shard in tenants.values_mut() {
5516 : // If a reconciliation is already in progress, rely on the previous scheduling
5517 : // decision and skip triggering a new reconciliation.
5518 0 : if tenant_shard.reconciler.is_some() {
5519 0 : continue;
5520 0 : }
5521 :
5522 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
5523 0 : if observed_loc.conf.is_none() {
5524 0 : self.maybe_reconcile_shard(tenant_shard, nodes);
5525 0 : }
5526 0 : }
5527 : }
5528 :
5529 : // TODO: in the background, we should balance work back onto this pageserver
5530 : }
5531 : // No action required for the intermediate unavailable state.
5532 : // When we transition into active or offline from the unavailable state,
5533 : // the correct handling above will kick in.
5534 : AvailabilityTransition::ToWarmingUpFromActive => {
5535 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
5536 : }
5537 : AvailabilityTransition::ToWarmingUpFromOffline => {
5538 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
5539 : }
5540 : AvailabilityTransition::Unchanged => {
5541 0 : tracing::debug!("Node {} no availability change during config", node_id);
5542 : }
5543 : }
5544 :
5545 0 : Ok(())
5546 0 : }
5547 :
5548 : /// Handle availability transition for multiple nodes
5549 : ///
5550 : /// Note that you should first call [`Self::node_state_configure`] for
5551 : /// all nodes being handled here for the handling to use fresh in-memory state.
5552 0 : async fn handle_node_availability_transitions(
5553 0 : &self,
5554 0 : transitions: Vec<(
5555 0 : NodeId,
5556 0 : TracingExclusiveGuard<NodeOperations>,
5557 0 : AvailabilityTransition,
5558 0 : )>,
5559 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
5560 0 : let mut errors = Vec::default();
5561 0 : for (node_id, node_lock, transition) in transitions {
5562 0 : let res = self
5563 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
5564 0 : .await;
5565 0 : if let Err(err) = res {
5566 0 : errors.push((node_id, err));
5567 0 : }
5568 : }
5569 :
5570 0 : if errors.is_empty() {
5571 0 : Ok(())
5572 : } else {
5573 0 : Err(errors)
5574 : }
5575 0 : }
5576 :
5577 0 : pub(crate) async fn node_configure(
5578 0 : &self,
5579 0 : node_id: NodeId,
5580 0 : availability: Option<NodeAvailability>,
5581 0 : scheduling: Option<NodeSchedulingPolicy>,
5582 0 : ) -> Result<(), ApiError> {
5583 0 : let node_lock =
5584 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
5585 :
5586 0 : let transition = self
5587 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
5588 0 : .await?;
5589 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
5590 0 : .await
5591 0 : }
5592 :
5593 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
5594 : /// operation for HTTP api.
5595 0 : pub(crate) async fn external_node_configure(
5596 0 : &self,
5597 0 : node_id: NodeId,
5598 0 : availability: Option<NodeAvailability>,
5599 0 : scheduling: Option<NodeSchedulingPolicy>,
5600 0 : ) -> Result<(), ApiError> {
5601 0 : {
5602 0 : let locked = self.inner.read().unwrap();
5603 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
5604 0 : return Err(ApiError::PreconditionFailed(
5605 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
5606 0 : ));
5607 0 : }
5608 0 : }
5609 0 :
5610 0 : self.node_configure(node_id, availability, scheduling).await
5611 0 : }
5612 :
5613 0 : pub(crate) async fn start_node_drain(
5614 0 : self: &Arc<Self>,
5615 0 : node_id: NodeId,
5616 0 : ) -> Result<(), ApiError> {
5617 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
5618 0 : let locked = self.inner.read().unwrap();
5619 0 : let nodes = &locked.nodes;
5620 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5621 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5622 0 : ))?;
5623 0 : let schedulable_nodes_count = nodes
5624 0 : .iter()
5625 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
5626 0 : .count();
5627 0 :
5628 0 : (
5629 0 : locked
5630 0 : .ongoing_operation
5631 0 : .as_ref()
5632 0 : .map(|ongoing| ongoing.operation),
5633 0 : node.is_available(),
5634 0 : node.get_scheduling(),
5635 0 : schedulable_nodes_count,
5636 0 : )
5637 0 : };
5638 :
5639 0 : if let Some(ongoing) = ongoing_op {
5640 0 : return Err(ApiError::PreconditionFailed(
5641 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
5642 0 : ));
5643 0 : }
5644 0 :
5645 0 : if !node_available {
5646 0 : return Err(ApiError::ResourceUnavailable(
5647 0 : format!("Node {node_id} is currently unavailable").into(),
5648 0 : ));
5649 0 : }
5650 0 :
5651 0 : if schedulable_nodes_count == 0 {
5652 0 : return Err(ApiError::PreconditionFailed(
5653 0 : "No other schedulable nodes to drain to".into(),
5654 0 : ));
5655 0 : }
5656 0 :
5657 0 : match node_policy {
5658 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
5659 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
5660 0 : .await?;
5661 :
5662 0 : let cancel = self.cancel.child_token();
5663 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
5664 :
5665 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
5666 0 : operation: Operation::Drain(Drain { node_id }),
5667 0 : cancel: cancel.clone(),
5668 0 : });
5669 :
5670 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
5671 :
5672 0 : tokio::task::spawn({
5673 0 : let service = self.clone();
5674 0 : let cancel = cancel.clone();
5675 0 : async move {
5676 0 : let _gate_guard = gate_guard;
5677 0 :
5678 0 : scopeguard::defer! {
5679 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
5680 0 :
5681 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
5682 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
5683 0 : } else {
5684 0 : panic!("We always remove the same operation")
5685 0 : }
5686 0 : }
5687 0 :
5688 0 : tracing::info!("Drain background operation starting");
5689 0 : let res = service.drain_node(node_id, cancel).await;
5690 0 : match res {
5691 : Ok(()) => {
5692 0 : tracing::info!("Drain background operation completed successfully");
5693 : }
5694 : Err(OperationError::Cancelled) => {
5695 0 : tracing::info!("Drain background operation was cancelled");
5696 : }
5697 0 : Err(err) => {
5698 0 : tracing::error!("Drain background operation encountered: {err}")
5699 : }
5700 : }
5701 0 : }
5702 0 : }.instrument(span));
5703 0 : }
5704 : NodeSchedulingPolicy::Draining => {
5705 0 : return Err(ApiError::Conflict(format!(
5706 0 : "Node {node_id} has drain in progress"
5707 0 : )));
5708 : }
5709 0 : policy => {
5710 0 : return Err(ApiError::PreconditionFailed(
5711 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
5712 0 : ));
5713 : }
5714 : }
5715 :
5716 0 : Ok(())
5717 0 : }
5718 :
5719 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
5720 0 : let node_available = {
5721 0 : let locked = self.inner.read().unwrap();
5722 0 : let nodes = &locked.nodes;
5723 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5724 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5725 0 : ))?;
5726 :
5727 0 : node.is_available()
5728 0 : };
5729 0 :
5730 0 : if !node_available {
5731 0 : return Err(ApiError::ResourceUnavailable(
5732 0 : format!("Node {node_id} is currently unavailable").into(),
5733 0 : ));
5734 0 : }
5735 :
5736 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
5737 0 : if let Operation::Drain(drain) = op_handler.operation {
5738 0 : if drain.node_id == node_id {
5739 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
5740 0 : op_handler.cancel.cancel();
5741 0 : return Ok(());
5742 0 : }
5743 0 : }
5744 0 : }
5745 :
5746 0 : Err(ApiError::PreconditionFailed(
5747 0 : format!("Node {node_id} has no drain in progress").into(),
5748 0 : ))
5749 0 : }
5750 :
5751 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
5752 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
5753 0 : let locked = self.inner.read().unwrap();
5754 0 : let nodes = &locked.nodes;
5755 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5756 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5757 0 : ))?;
5758 :
5759 0 : (
5760 0 : locked
5761 0 : .ongoing_operation
5762 0 : .as_ref()
5763 0 : .map(|ongoing| ongoing.operation),
5764 0 : node.is_available(),
5765 0 : node.get_scheduling(),
5766 0 : nodes.len(),
5767 0 : )
5768 0 : };
5769 :
5770 0 : if let Some(ongoing) = ongoing_op {
5771 0 : return Err(ApiError::PreconditionFailed(
5772 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
5773 0 : ));
5774 0 : }
5775 0 :
5776 0 : if !node_available {
5777 0 : return Err(ApiError::ResourceUnavailable(
5778 0 : format!("Node {node_id} is currently unavailable").into(),
5779 0 : ));
5780 0 : }
5781 0 :
5782 0 : if total_nodes_count <= 1 {
5783 0 : return Err(ApiError::PreconditionFailed(
5784 0 : "No other nodes to fill from".into(),
5785 0 : ));
5786 0 : }
5787 0 :
5788 0 : match node_policy {
5789 : NodeSchedulingPolicy::Active => {
5790 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
5791 0 : .await?;
5792 :
5793 0 : let cancel = self.cancel.child_token();
5794 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
5795 :
5796 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
5797 0 : operation: Operation::Fill(Fill { node_id }),
5798 0 : cancel: cancel.clone(),
5799 0 : });
5800 :
5801 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
5802 :
5803 0 : tokio::task::spawn({
5804 0 : let service = self.clone();
5805 0 : let cancel = cancel.clone();
5806 0 : async move {
5807 0 : let _gate_guard = gate_guard;
5808 0 :
5809 0 : scopeguard::defer! {
5810 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
5811 0 :
5812 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
5813 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
5814 0 : } else {
5815 0 : panic!("We always remove the same operation")
5816 0 : }
5817 0 : }
5818 0 :
5819 0 : tracing::info!("Fill background operation starting");
5820 0 : let res = service.fill_node(node_id, cancel).await;
5821 0 : match res {
5822 : Ok(()) => {
5823 0 : tracing::info!("Fill background operation completed successfully");
5824 : }
5825 : Err(OperationError::Cancelled) => {
5826 0 : tracing::info!("Fill background operation was cancelled");
5827 : }
5828 0 : Err(err) => {
5829 0 : tracing::error!("Fill background operation encountered: {err}")
5830 : }
5831 : }
5832 0 : }
5833 0 : }.instrument(span));
5834 0 : }
5835 : NodeSchedulingPolicy::Filling => {
5836 0 : return Err(ApiError::Conflict(format!(
5837 0 : "Node {node_id} has fill in progress"
5838 0 : )));
5839 : }
5840 0 : policy => {
5841 0 : return Err(ApiError::PreconditionFailed(
5842 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
5843 0 : ));
5844 : }
5845 : }
5846 :
5847 0 : Ok(())
5848 0 : }
5849 :
5850 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
5851 0 : let node_available = {
5852 0 : let locked = self.inner.read().unwrap();
5853 0 : let nodes = &locked.nodes;
5854 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5855 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5856 0 : ))?;
5857 :
5858 0 : node.is_available()
5859 0 : };
5860 0 :
5861 0 : if !node_available {
5862 0 : return Err(ApiError::ResourceUnavailable(
5863 0 : format!("Node {node_id} is currently unavailable").into(),
5864 0 : ));
5865 0 : }
5866 :
5867 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
5868 0 : if let Operation::Fill(fill) = op_handler.operation {
5869 0 : if fill.node_id == node_id {
5870 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
5871 0 : op_handler.cancel.cancel();
5872 0 : return Ok(());
5873 0 : }
5874 0 : }
5875 0 : }
5876 :
5877 0 : Err(ApiError::PreconditionFailed(
5878 0 : format!("Node {node_id} has no fill in progress").into(),
5879 0 : ))
5880 0 : }
5881 :
5882 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
5883 : /// configuration
5884 0 : fn maybe_reconcile_shard(
5885 0 : &self,
5886 0 : shard: &mut TenantShard,
5887 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5888 0 : ) -> Option<ReconcilerWaiter> {
5889 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::default())
5890 0 : }
5891 :
5892 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
5893 0 : fn maybe_configured_reconcile_shard(
5894 0 : &self,
5895 0 : shard: &mut TenantShard,
5896 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5897 0 : reconciler_config: ReconcilerConfig,
5898 0 : ) -> Option<ReconcilerWaiter> {
5899 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
5900 0 :
5901 0 : match reconcile_needed {
5902 0 : ReconcileNeeded::No => return None,
5903 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
5904 0 : ReconcileNeeded::Yes => {
5905 0 : // Fall through to try and acquire units for spawning reconciler
5906 0 : }
5907 : };
5908 :
5909 0 : let units = match self.reconciler_concurrency.clone().try_acquire_owned() {
5910 0 : Ok(u) => ReconcileUnits::new(u),
5911 : Err(_) => {
5912 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
5913 0 : "Concurrency limited: enqueued for reconcile later");
5914 0 : if !shard.delayed_reconcile {
5915 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
5916 0 : Err(TrySendError::Closed(_)) => {
5917 0 : // Weird mid-shutdown case?
5918 0 : }
5919 : Err(TrySendError::Full(_)) => {
5920 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
5921 0 : tracing::warn!(
5922 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
5923 : );
5924 : }
5925 0 : Ok(()) => {
5926 0 : shard.delayed_reconcile = true;
5927 0 : }
5928 : }
5929 0 : }
5930 :
5931 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
5932 : // number to advance. When this function is eventually called again and succeeds in getting units,
5933 : // it will spawn a reconciler that makes this waiter complete.
5934 0 : return Some(shard.future_reconcile_waiter());
5935 : }
5936 : };
5937 :
5938 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
5939 : // Gate closed: we're shutting down, drop out.
5940 0 : return None;
5941 : };
5942 :
5943 0 : shard.spawn_reconciler(
5944 0 : &self.result_tx,
5945 0 : nodes,
5946 0 : &self.compute_hook,
5947 0 : reconciler_config,
5948 0 : &self.config,
5949 0 : &self.persistence,
5950 0 : units,
5951 0 : gate_guard,
5952 0 : &self.reconcilers_cancel,
5953 0 : )
5954 0 : }
5955 :
5956 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
5957 : /// Additionally, reschedule tenants that require it.
5958 : ///
5959 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
5960 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
5961 : /// available. A return value of 0 indicates that everything is fully reconciled already.
5962 0 : fn reconcile_all(&self) -> usize {
5963 0 : let mut locked = self.inner.write().unwrap();
5964 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
5965 0 : let pageservers = nodes.clone();
5966 0 :
5967 0 : let mut schedule_context = ScheduleContext::default();
5968 0 :
5969 0 : let mut reconciles_spawned = 0;
5970 0 : for (tenant_shard_id, shard) in tenants.iter_mut() {
5971 0 : if tenant_shard_id.is_shard_zero() {
5972 0 : schedule_context = ScheduleContext::default();
5973 0 : }
5974 :
5975 : // Skip checking if this shard is already enqueued for reconciliation
5976 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
5977 : // If there is something delayed, then return a nonzero count so that
5978 : // callers like reconcile_all_now do not incorrectly get the impression
5979 : // that the system is in a quiescent state.
5980 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
5981 0 : continue;
5982 0 : }
5983 0 :
5984 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
5985 0 : // dirty, spawn another rone
5986 0 : if self.maybe_reconcile_shard(shard, &pageservers).is_some() {
5987 0 : reconciles_spawned += 1;
5988 0 : }
5989 :
5990 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5991 : }
5992 :
5993 0 : reconciles_spawned
5994 0 : }
5995 :
5996 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
5997 : /// could be scheduled somewhere better:
5998 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
5999 : /// * e.g. after a node fails then recovers, to move some work back to it
6000 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
6001 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
6002 : /// we did the split, but are probably better placed elsewhere.
6003 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
6004 : /// * e.g. after a shard split, some locations will be on the same node (where the split
6005 : /// happened), and will probably be better placed elsewhere.
6006 : ///
6007 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
6008 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
6009 : /// according to those same soft constraints.
6010 0 : async fn optimize_all(&self) -> usize {
6011 : // Limit on how many shards' optmizations each call to this function will execute. Combined
6012 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
6013 : // trickle of optimizations in the background, rather than executing a large number in parallel
6014 : // when a change occurs.
6015 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
6016 :
6017 : // Synchronous prepare: scan shards for possible scheduling optimizations
6018 0 : let candidate_work = self.optimize_all_plan();
6019 0 : let candidate_work_len = candidate_work.len();
6020 :
6021 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
6022 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
6023 :
6024 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
6025 0 :
6026 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
6027 0 : let mut reconciles_spawned = 0;
6028 0 : let mut optimizations_applied = 0;
6029 0 : let mut locked = self.inner.write().unwrap();
6030 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6031 0 : for (tenant_shard_id, optimization) in validated_work {
6032 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6033 : // Shard was dropped between planning and execution;
6034 0 : continue;
6035 : };
6036 0 : if shard.apply_optimization(scheduler, optimization) {
6037 0 : optimizations_applied += 1;
6038 0 : if self.maybe_reconcile_shard(shard, nodes).is_some() {
6039 0 : reconciles_spawned += 1;
6040 0 : }
6041 0 : }
6042 :
6043 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
6044 0 : break;
6045 0 : }
6046 : }
6047 :
6048 0 : if was_work_filtered {
6049 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
6050 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
6051 0 : // as these validations start passing.
6052 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
6053 0 : }
6054 :
6055 0 : reconciles_spawned
6056 0 : }
6057 :
6058 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
6059 0 : let mut schedule_context = ScheduleContext::default();
6060 0 :
6061 0 : let mut tenant_shards: Vec<&TenantShard> = Vec::new();
6062 :
6063 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
6064 : // this higher than the execution limit gives us a chance to execute some work even if the first
6065 : // few optimizations we find are not ready.
6066 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
6067 :
6068 0 : let mut work = Vec::new();
6069 0 :
6070 0 : let mut locked = self.inner.write().unwrap();
6071 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6072 0 : for (tenant_shard_id, shard) in tenants.iter() {
6073 0 : if tenant_shard_id.is_shard_zero() {
6074 0 : // Reset accumulators on the first shard in a tenant
6075 0 : schedule_context = ScheduleContext::default();
6076 0 : schedule_context.mode = ScheduleMode::Speculative;
6077 0 : tenant_shards.clear();
6078 0 : }
6079 :
6080 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
6081 0 : break;
6082 0 : }
6083 0 :
6084 0 : match shard.get_scheduling_policy() {
6085 0 : ShardSchedulingPolicy::Active => {
6086 0 : // Ok to do optimization
6087 0 : }
6088 : ShardSchedulingPolicy::Essential
6089 : | ShardSchedulingPolicy::Pause
6090 : | ShardSchedulingPolicy::Stop => {
6091 : // Policy prevents optimizing this shard.
6092 0 : continue;
6093 : }
6094 : }
6095 :
6096 : // Accumulate the schedule context for all the shards in a tenant: we must have
6097 : // the total view of all shards before we can try to optimize any of them.
6098 0 : schedule_context.avoid(&shard.intent.all_pageservers());
6099 0 : if let Some(attached) = shard.intent.get_attached() {
6100 0 : schedule_context.push_attached(*attached);
6101 0 : }
6102 0 : tenant_shards.push(shard);
6103 0 :
6104 0 : // Once we have seen the last shard in the tenant, proceed to search across all shards
6105 0 : // in the tenant for optimizations
6106 0 : if shard.shard.number.0 == shard.shard.count.count() - 1 {
6107 0 : if tenant_shards.iter().any(|s| s.reconciler.is_some()) {
6108 : // Do not start any optimizations while another change to the tenant is ongoing: this
6109 : // is not necessary for correctness, but simplifies operations and implicitly throttles
6110 : // optimization changes to happen in a "trickle" over time.
6111 0 : continue;
6112 0 : }
6113 0 :
6114 0 : if tenant_shards.iter().any(|s| {
6115 0 : !matches!(s.splitting, SplitState::Idle)
6116 0 : || matches!(s.policy, PlacementPolicy::Detached)
6117 0 : }) {
6118 : // Never attempt to optimize a tenant that is currently being split, or
6119 : // a tenant that is meant to be detached
6120 0 : continue;
6121 0 : }
6122 :
6123 : // TODO: optimization calculations are relatively expensive: create some fast-path for
6124 : // the common idle case (avoiding the search on tenants that we have recently checked)
6125 :
6126 0 : for shard in &tenant_shards {
6127 0 : if let Some(optimization) =
6128 : // If idle, maybe ptimize attachments: if a shard has a secondary location that is preferable to
6129 : // its primary location based on soft constraints, cut it over.
6130 0 : shard.optimize_attachment(nodes, &schedule_context)
6131 : {
6132 0 : work.push((shard.tenant_shard_id, optimization));
6133 0 : break;
6134 0 : } else if let Some(optimization) =
6135 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
6136 : // better placed on another node, based on ScheduleContext, then adjust it. This
6137 : // covers cases like after a shard split, where we might have too many shards
6138 : // in the same tenant with secondary locations on the node where they originally split.
6139 0 : shard.optimize_secondary(scheduler, &schedule_context)
6140 : {
6141 0 : work.push((shard.tenant_shard_id, optimization));
6142 0 : break;
6143 0 : }
6144 :
6145 : // TODO: extend this mechanism to prefer attaching on nodes with fewer attached
6146 : // tenants (i.e. extend schedule state to distinguish attached from secondary counts),
6147 : // for the total number of attachments on a node (not just within a tenant.)
6148 : }
6149 0 : }
6150 : }
6151 :
6152 0 : work
6153 0 : }
6154 :
6155 0 : async fn optimize_all_validate(
6156 0 : &self,
6157 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
6158 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
6159 0 : // Take a clone of the node map to use outside the lock in async validation phase
6160 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
6161 0 :
6162 0 : let mut want_secondary_status = Vec::new();
6163 0 :
6164 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
6165 0 : // check that the state of locations is acceptable to run the optimization, such as
6166 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
6167 0 : // in a live migration.
6168 0 : let mut validated_work = Vec::new();
6169 0 : for (tenant_shard_id, optimization) in candidate_work {
6170 0 : match optimization.action {
6171 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
6172 : old_attached_node_id: _,
6173 0 : new_attached_node_id,
6174 0 : }) => {
6175 0 : match validation_nodes.get(&new_attached_node_id) {
6176 0 : None => {
6177 0 : // Node was dropped between planning and validation
6178 0 : }
6179 0 : Some(node) => {
6180 0 : if !node.is_available() {
6181 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
6182 0 : } else {
6183 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
6184 0 : // remote API requests concurrently.
6185 0 : want_secondary_status.push((
6186 0 : tenant_shard_id,
6187 0 : node.clone(),
6188 0 : optimization,
6189 0 : ));
6190 0 : }
6191 : }
6192 : }
6193 : }
6194 : ScheduleOptimizationAction::ReplaceSecondary(_) => {
6195 : // No extra checks needed to replace a secondary: this does not interrupt client access
6196 0 : validated_work.push((tenant_shard_id, optimization))
6197 : }
6198 : };
6199 : }
6200 :
6201 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
6202 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
6203 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
6204 0 : let results = self
6205 0 : .tenant_for_shards_api(
6206 0 : want_secondary_status
6207 0 : .iter()
6208 0 : .map(|i| (i.0, i.1.clone()))
6209 0 : .collect(),
6210 0 : |tenant_shard_id, client| async move {
6211 0 : client.tenant_secondary_status(tenant_shard_id).await
6212 0 : },
6213 0 : 1,
6214 0 : 1,
6215 0 : SHORT_RECONCILE_TIMEOUT,
6216 0 : &self.cancel,
6217 0 : )
6218 0 : .await;
6219 :
6220 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
6221 0 : want_secondary_status.into_iter().zip(results.into_iter())
6222 : {
6223 0 : match secondary_status {
6224 0 : Err(e) => {
6225 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
6226 : }
6227 0 : Ok(progress) => {
6228 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
6229 : // them in an optimization
6230 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
6231 :
6232 0 : if progress.heatmap_mtime.is_none()
6233 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
6234 0 : && progress.bytes_downloaded != progress.bytes_total
6235 0 : || progress.bytes_total - progress.bytes_downloaded
6236 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
6237 : {
6238 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
6239 : } else {
6240 : // Location looks ready: proceed
6241 0 : tracing::info!(
6242 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
6243 : );
6244 0 : validated_work.push((tenant_shard_id, optimization))
6245 : }
6246 : }
6247 : }
6248 : }
6249 :
6250 0 : validated_work
6251 0 : }
6252 :
6253 : /// Look for shards which are oversized and in need of splitting
6254 0 : async fn autosplit_tenants(self: &Arc<Self>) {
6255 0 : let Some(split_threshold) = self.config.split_threshold else {
6256 : // Auto-splitting is disabled
6257 0 : return;
6258 : };
6259 :
6260 0 : let nodes = self.inner.read().unwrap().nodes.clone();
6261 :
6262 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
6263 :
6264 0 : let mut top_n = Vec::new();
6265 0 :
6266 0 : // Call into each node to look for big tenants
6267 0 : let top_n_request = TopTenantShardsRequest {
6268 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
6269 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
6270 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
6271 0 : // disk space impact of one shard.
6272 0 : order_by: models::TenantSorting::MaxLogicalSize,
6273 0 : limit: 10,
6274 0 : where_shards_lt: Some(SPLIT_TO_MAX),
6275 0 : where_gt: Some(split_threshold),
6276 0 : };
6277 0 : for node in nodes.values() {
6278 0 : let request_ref = &top_n_request;
6279 0 : match node
6280 0 : .with_client_retries(
6281 0 : |client| async move {
6282 0 : let request = request_ref.clone();
6283 0 : client.top_tenant_shards(request.clone()).await
6284 0 : },
6285 0 : &self.config.jwt_token,
6286 0 : 3,
6287 0 : 3,
6288 0 : Duration::from_secs(5),
6289 0 : &self.cancel,
6290 0 : )
6291 0 : .await
6292 : {
6293 0 : Some(Ok(node_top_n)) => {
6294 0 : top_n.extend(node_top_n.shards.into_iter());
6295 0 : }
6296 : Some(Err(mgmt_api::Error::Cancelled)) => {
6297 0 : continue;
6298 : }
6299 0 : Some(Err(e)) => {
6300 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
6301 0 : continue;
6302 : }
6303 : None => {
6304 : // Node is shutting down
6305 0 : continue;
6306 : }
6307 : };
6308 : }
6309 :
6310 : // Pick the biggest tenant to split first
6311 0 : top_n.sort_by_key(|i| i.resident_size);
6312 0 : let Some(split_candidate) = top_n.into_iter().next() else {
6313 0 : tracing::debug!("No split-elegible shards found");
6314 0 : return;
6315 : };
6316 :
6317 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
6318 : // want to block the background reconcile loop on this.
6319 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
6320 :
6321 0 : let this = self.clone();
6322 0 : tokio::spawn(
6323 0 : async move {
6324 0 : match this
6325 0 : .tenant_shard_split(
6326 0 : split_candidate.id.tenant_id,
6327 0 : TenantShardSplitRequest {
6328 0 : // Always split to the max number of shards: this avoids stepping through
6329 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
6330 0 : // each time as a tenant grows, and is not too expensive because our max shard
6331 0 : // count is relatively low anyway.
6332 0 : // This policy will be adjusted in future once we support higher shard count.
6333 0 : new_shard_count: SPLIT_TO_MAX.literal(),
6334 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
6335 0 : },
6336 0 : )
6337 0 : .await
6338 : {
6339 : Ok(_) => {
6340 0 : tracing::info!("Successful auto-split");
6341 : }
6342 0 : Err(e) => {
6343 0 : tracing::error!("Auto-split failed: {e}");
6344 : }
6345 : }
6346 0 : }
6347 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
6348 : );
6349 0 : }
6350 :
6351 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
6352 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
6353 : /// put the system into a quiescent state where future background reconciliations won't do anything.
6354 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
6355 0 : let reconciles_spawned = self.reconcile_all();
6356 0 : let reconciles_spawned = if reconciles_spawned == 0 {
6357 : // Only optimize when we are otherwise idle
6358 0 : self.optimize_all().await
6359 : } else {
6360 0 : reconciles_spawned
6361 : };
6362 :
6363 0 : let waiters = {
6364 0 : let mut waiters = Vec::new();
6365 0 : let locked = self.inner.read().unwrap();
6366 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
6367 0 : if let Some(waiter) = shard.get_waiter() {
6368 0 : waiters.push(waiter);
6369 0 : }
6370 : }
6371 0 : waiters
6372 0 : };
6373 0 :
6374 0 : let waiter_count = waiters.len();
6375 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
6376 0 : Ok(()) => {}
6377 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
6378 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
6379 0 : {
6380 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
6381 0 : // because some other change superceded it. We will return a nonzero number,
6382 0 : // so the caller knows they might have to call again to quiesce the system.
6383 0 : }
6384 0 : Err(e) => {
6385 0 : return Err(e);
6386 : }
6387 : };
6388 :
6389 0 : tracing::info!(
6390 0 : "{} reconciles in reconcile_all, {} waiters",
6391 : reconciles_spawned,
6392 : waiter_count
6393 : );
6394 :
6395 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
6396 0 : }
6397 :
6398 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
6399 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
6400 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
6401 0 : self.reconcilers_cancel.cancel();
6402 0 : self.reconcilers_gate.close().await;
6403 :
6404 : // Signal the background loop in [`Service::process_results`] to exit once
6405 : // it has proccessed the results from all the reconciles we cancelled earlier.
6406 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
6407 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
6408 0 : self.result_tx.closed().await;
6409 0 : }
6410 :
6411 0 : pub async fn shutdown(&self) {
6412 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
6413 0 : .await;
6414 :
6415 : // Background tasks hold gate guards: this notifies them of the cancellation and
6416 : // waits for them all to complete.
6417 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
6418 0 : self.cancel.cancel();
6419 0 : self.gate.close().await;
6420 0 : }
6421 :
6422 : /// Spot check the download lag for a secondary location of a shard.
6423 : /// Should be used as a heuristic, since it's not always precise: the
6424 : /// secondary might have not downloaded the new heat map yet and, hence,
6425 : /// is not aware of the lag.
6426 : ///
6427 : /// Returns:
6428 : /// * Ok(None) if the lag could not be determined from the status,
6429 : /// * Ok(Some(_)) if the lag could be determind
6430 : /// * Err on failures to query the pageserver.
6431 0 : async fn secondary_lag(
6432 0 : &self,
6433 0 : secondary: &NodeId,
6434 0 : tenant_shard_id: TenantShardId,
6435 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
6436 0 : let nodes = self.inner.read().unwrap().nodes.clone();
6437 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
6438 0 : StatusCode::NOT_FOUND,
6439 0 : format!("Node with id {} not found", secondary),
6440 0 : ))?;
6441 :
6442 0 : match node
6443 0 : .with_client_retries(
6444 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
6445 0 : &self.config.jwt_token,
6446 0 : 1,
6447 0 : 3,
6448 0 : Duration::from_millis(250),
6449 0 : &self.cancel,
6450 0 : )
6451 0 : .await
6452 : {
6453 0 : Some(Ok(status)) => match status.heatmap_mtime {
6454 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
6455 0 : None => Ok(None),
6456 : },
6457 0 : Some(Err(e)) => Err(e),
6458 0 : None => Err(mgmt_api::Error::Cancelled),
6459 : }
6460 0 : }
6461 :
6462 : /// Drain a node by moving the shards attached to it as primaries.
6463 : /// This is a long running operation and it should run as a separate Tokio task.
6464 0 : pub(crate) async fn drain_node(
6465 0 : self: &Arc<Self>,
6466 0 : node_id: NodeId,
6467 0 : cancel: CancellationToken,
6468 0 : ) -> Result<(), OperationError> {
6469 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
6470 0 : let max_secondary_lag_bytes = self
6471 0 : .config
6472 0 : .max_secondary_lag_bytes
6473 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
6474 :
6475 : // By default, live migrations are generous about the wait time for getting
6476 : // the secondary location up to speed. When draining, give up earlier in order
6477 : // to not stall the operation when a cold secondary is encountered.
6478 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
6479 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
6480 0 : let reconciler_config = ReconcilerConfigBuilder::new()
6481 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
6482 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
6483 0 : .build();
6484 0 :
6485 0 : let mut waiters = Vec::new();
6486 0 :
6487 0 : let mut tid_iter = TenantShardIterator::new({
6488 0 : let service = self.clone();
6489 0 : move |last_inspected_shard: Option<TenantShardId>| {
6490 0 : let locked = &service.inner.read().unwrap();
6491 0 : let tenants = &locked.tenants;
6492 0 : let entry = match last_inspected_shard {
6493 0 : Some(skip_past) => {
6494 0 : // Skip to the last seen tenant shard id
6495 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
6496 0 :
6497 0 : // Skip past the last seen
6498 0 : cursor.nth(1)
6499 : }
6500 0 : None => tenants.first_key_value(),
6501 : };
6502 :
6503 0 : entry.map(|(tid, _)| tid).copied()
6504 0 : }
6505 0 : });
6506 :
6507 0 : while !tid_iter.finished() {
6508 0 : if cancel.is_cancelled() {
6509 0 : match self
6510 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6511 0 : .await
6512 : {
6513 0 : Ok(()) => return Err(OperationError::Cancelled),
6514 0 : Err(err) => {
6515 0 : return Err(OperationError::FinalizeError(
6516 0 : format!(
6517 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6518 0 : node_id, err
6519 0 : )
6520 0 : .into(),
6521 0 : ));
6522 : }
6523 : }
6524 0 : }
6525 0 :
6526 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
6527 :
6528 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
6529 0 : let tid = match tid_iter.next() {
6530 0 : Some(tid) => tid,
6531 : None => {
6532 0 : break;
6533 : }
6534 : };
6535 :
6536 0 : let tid_drain = TenantShardDrain {
6537 0 : drained_node: node_id,
6538 0 : tenant_shard_id: tid,
6539 0 : };
6540 :
6541 0 : let dest_node_id = {
6542 0 : let locked = self.inner.read().unwrap();
6543 0 :
6544 0 : match tid_drain
6545 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
6546 : {
6547 0 : Some(node_id) => node_id,
6548 : None => {
6549 0 : continue;
6550 : }
6551 : }
6552 : };
6553 :
6554 0 : match self.secondary_lag(&dest_node_id, tid).await {
6555 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
6556 0 : // The secondary is reasonably up to date.
6557 0 : // Migrate to it
6558 0 : }
6559 0 : Ok(Some(lag)) => {
6560 0 : tracing::info!(
6561 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6562 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
6563 : );
6564 0 : continue;
6565 : }
6566 : Ok(None) => {
6567 0 : tracing::info!(
6568 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6569 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
6570 : );
6571 0 : continue;
6572 : }
6573 0 : Err(err) => {
6574 0 : tracing::warn!(
6575 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6576 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
6577 : );
6578 0 : continue;
6579 : }
6580 : }
6581 :
6582 : {
6583 0 : let mut locked = self.inner.write().unwrap();
6584 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6585 0 : let rescheduled = tid_drain.reschedule_to_secondary(
6586 0 : dest_node_id,
6587 0 : tenants,
6588 0 : scheduler,
6589 0 : nodes,
6590 0 : )?;
6591 :
6592 0 : if let Some(tenant_shard) = rescheduled {
6593 0 : let waiter = self.maybe_configured_reconcile_shard(
6594 0 : tenant_shard,
6595 0 : nodes,
6596 0 : reconciler_config,
6597 0 : );
6598 0 : if let Some(some) = waiter {
6599 0 : waiters.push(some);
6600 0 : }
6601 0 : }
6602 : }
6603 : }
6604 :
6605 0 : waiters = self
6606 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6607 0 : .await;
6608 :
6609 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
6610 : }
6611 :
6612 0 : while !waiters.is_empty() {
6613 0 : if cancel.is_cancelled() {
6614 0 : match self
6615 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6616 0 : .await
6617 : {
6618 0 : Ok(()) => return Err(OperationError::Cancelled),
6619 0 : Err(err) => {
6620 0 : return Err(OperationError::FinalizeError(
6621 0 : format!(
6622 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6623 0 : node_id, err
6624 0 : )
6625 0 : .into(),
6626 0 : ));
6627 : }
6628 : }
6629 0 : }
6630 0 :
6631 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
6632 :
6633 0 : waiters = self
6634 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6635 0 : .await;
6636 : }
6637 :
6638 : // At this point we have done the best we could to drain shards from this node.
6639 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
6640 : // to complete the drain.
6641 0 : if let Err(err) = self
6642 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
6643 0 : .await
6644 : {
6645 : // This is not fatal. Anything that is polling the node scheduling policy to detect
6646 : // the end of the drain operations will hang, but all such places should enforce an
6647 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
6648 : // by the counterpart fill operation.
6649 0 : return Err(OperationError::FinalizeError(
6650 0 : format!(
6651 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
6652 0 : )
6653 0 : .into(),
6654 0 : ));
6655 0 : }
6656 0 :
6657 0 : Ok(())
6658 0 : }
6659 :
6660 : /// Create a node fill plan (pick secondaries to promote) that meets the following requirements:
6661 : /// 1. The node should be filled until it reaches the expected cluster average of
6662 : /// attached shards. If there are not enough secondaries on the node, the plan stops early.
6663 : /// 2. Select tenant shards to promote such that the number of attached shards is balanced
6664 : /// throughout the cluster. We achieve this by picking tenant shards from each node,
6665 : /// starting from the ones with the largest number of attached shards, until the node
6666 : /// reaches the expected cluster average.
6667 : /// 3. Avoid promoting more shards of the same tenant than required. The upper bound
6668 : /// for the number of tenants from the same shard promoted to the node being filled is:
6669 : /// shard count for the tenant divided by the number of nodes in the cluster.
6670 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
6671 0 : let mut locked = self.inner.write().unwrap();
6672 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
6673 0 :
6674 0 : let mut tids_by_node = locked
6675 0 : .tenants
6676 0 : .iter_mut()
6677 0 : .filter_map(|(tid, tenant_shard)| {
6678 0 : if tenant_shard.intent.get_secondary().contains(&node_id) {
6679 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
6680 0 : return Some((*primary, *tid));
6681 0 : }
6682 0 : }
6683 :
6684 0 : None
6685 0 : })
6686 0 : .into_group_map();
6687 0 :
6688 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
6689 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
6690 0 :
6691 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
6692 0 : let mut plan = Vec::new();
6693 :
6694 0 : for (node_id, attached) in nodes_by_load {
6695 0 : let available = locked
6696 0 : .nodes
6697 0 : .get(&node_id)
6698 0 : .map_or(false, |n| n.is_available());
6699 0 : if !available {
6700 0 : continue;
6701 0 : }
6702 0 :
6703 0 : if plan.len() >= fill_requirement
6704 0 : || tids_by_node.is_empty()
6705 0 : || attached <= expected_attached
6706 : {
6707 0 : break;
6708 0 : }
6709 0 :
6710 0 : let can_take = attached - expected_attached;
6711 0 : let needed = fill_requirement - plan.len();
6712 0 : let mut take = std::cmp::min(can_take, needed);
6713 0 :
6714 0 : let mut remove_node = false;
6715 0 : while take > 0 {
6716 0 : match tids_by_node.get_mut(&node_id) {
6717 0 : Some(tids) => match tids.pop() {
6718 0 : Some(tid) => {
6719 0 : let max_promote_for_tenant = std::cmp::max(
6720 0 : tid.shard_count.count() as usize / locked.nodes.len(),
6721 0 : 1,
6722 0 : );
6723 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
6724 0 : if *promoted < max_promote_for_tenant {
6725 0 : plan.push(tid);
6726 0 : *promoted += 1;
6727 0 : take -= 1;
6728 0 : }
6729 : }
6730 : None => {
6731 0 : remove_node = true;
6732 0 : break;
6733 : }
6734 : },
6735 : None => {
6736 0 : break;
6737 : }
6738 : }
6739 : }
6740 :
6741 0 : if remove_node {
6742 0 : tids_by_node.remove(&node_id);
6743 0 : }
6744 : }
6745 :
6746 0 : plan
6747 0 : }
6748 :
6749 : /// Fill a node by promoting its secondaries until the cluster is balanced
6750 : /// with regards to attached shard counts. Note that this operation only
6751 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
6752 : /// This is a long running operation and it should run as a separate Tokio task.
6753 0 : pub(crate) async fn fill_node(
6754 0 : &self,
6755 0 : node_id: NodeId,
6756 0 : cancel: CancellationToken,
6757 0 : ) -> Result<(), OperationError> {
6758 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
6759 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
6760 0 : let reconciler_config = ReconcilerConfigBuilder::new()
6761 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
6762 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
6763 0 : .build();
6764 0 :
6765 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
6766 0 : let mut waiters = Vec::new();
6767 :
6768 : // Execute the plan we've composed above. Before aplying each move from the plan,
6769 : // we validate to ensure that it has not gone stale in the meantime.
6770 0 : while !tids_to_promote.is_empty() {
6771 0 : if cancel.is_cancelled() {
6772 0 : match self
6773 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6774 0 : .await
6775 : {
6776 0 : Ok(()) => return Err(OperationError::Cancelled),
6777 0 : Err(err) => {
6778 0 : return Err(OperationError::FinalizeError(
6779 0 : format!(
6780 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6781 0 : node_id, err
6782 0 : )
6783 0 : .into(),
6784 0 : ));
6785 : }
6786 : }
6787 0 : }
6788 0 :
6789 0 : {
6790 0 : let mut locked = self.inner.write().unwrap();
6791 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6792 :
6793 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
6794 0 : format!("node {node_id} was removed").into(),
6795 0 : ))?;
6796 :
6797 0 : let current_policy = node.get_scheduling();
6798 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
6799 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
6800 : // about it
6801 0 : return Err(OperationError::NodeStateChanged(
6802 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
6803 0 : ));
6804 0 : }
6805 :
6806 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
6807 0 : if let Some(tid) = tids_to_promote.pop() {
6808 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
6809 : // If the node being filled is not a secondary anymore,
6810 : // skip the promotion.
6811 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
6812 0 : continue;
6813 0 : }
6814 0 :
6815 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
6816 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
6817 0 : Err(e) => {
6818 0 : tracing::warn!(
6819 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6820 0 : "Scheduling error when filling pageserver {} : {e}", node_id
6821 : );
6822 : }
6823 : Ok(()) => {
6824 0 : tracing::info!(
6825 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6826 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
6827 : node_id,
6828 : previously_attached_to,
6829 : node_id
6830 : );
6831 :
6832 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
6833 0 : tenant_shard,
6834 0 : nodes,
6835 0 : reconciler_config,
6836 0 : ) {
6837 0 : waiters.push(waiter);
6838 0 : }
6839 : }
6840 : }
6841 0 : }
6842 : } else {
6843 0 : break;
6844 : }
6845 : }
6846 : }
6847 :
6848 0 : waiters = self
6849 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6850 0 : .await;
6851 : }
6852 :
6853 0 : while !waiters.is_empty() {
6854 0 : if cancel.is_cancelled() {
6855 0 : match self
6856 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6857 0 : .await
6858 : {
6859 0 : Ok(()) => return Err(OperationError::Cancelled),
6860 0 : Err(err) => {
6861 0 : return Err(OperationError::FinalizeError(
6862 0 : format!(
6863 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6864 0 : node_id, err
6865 0 : )
6866 0 : .into(),
6867 0 : ));
6868 : }
6869 : }
6870 0 : }
6871 0 :
6872 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
6873 :
6874 0 : waiters = self
6875 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6876 0 : .await;
6877 : }
6878 :
6879 0 : if let Err(err) = self
6880 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6881 0 : .await
6882 : {
6883 : // This isn't a huge issue since the filling process starts upon request. However, it
6884 : // will prevent the next drain from starting. The only case in which this can fail
6885 : // is database unavailability. Such a case will require manual intervention.
6886 0 : return Err(OperationError::FinalizeError(
6887 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
6888 0 : .into(),
6889 0 : ));
6890 0 : }
6891 0 :
6892 0 : Ok(())
6893 0 : }
6894 :
6895 : /// Updates scrubber metadata health check results.
6896 0 : pub(crate) async fn metadata_health_update(
6897 0 : &self,
6898 0 : update_req: MetadataHealthUpdateRequest,
6899 0 : ) -> Result<(), ApiError> {
6900 0 : let now = chrono::offset::Utc::now();
6901 0 : let (healthy_records, unhealthy_records) = {
6902 0 : let locked = self.inner.read().unwrap();
6903 0 : let healthy_records = update_req
6904 0 : .healthy_tenant_shards
6905 0 : .into_iter()
6906 0 : // Retain only health records associated with tenant shards managed by storage controller.
6907 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6908 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
6909 0 : .collect();
6910 0 : let unhealthy_records = update_req
6911 0 : .unhealthy_tenant_shards
6912 0 : .into_iter()
6913 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6914 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
6915 0 : .collect();
6916 0 :
6917 0 : (healthy_records, unhealthy_records)
6918 0 : };
6919 0 :
6920 0 : self.persistence
6921 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
6922 0 : .await?;
6923 0 : Ok(())
6924 0 : }
6925 :
6926 : /// Lists the tenant shards that has unhealthy metadata status.
6927 0 : pub(crate) async fn metadata_health_list_unhealthy(
6928 0 : &self,
6929 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
6930 0 : let result = self
6931 0 : .persistence
6932 0 : .list_unhealthy_metadata_health_records()
6933 0 : .await?
6934 0 : .iter()
6935 0 : .map(|p| p.get_tenant_shard_id().unwrap())
6936 0 : .collect();
6937 0 :
6938 0 : Ok(result)
6939 0 : }
6940 :
6941 : /// Lists the tenant shards that have not been scrubbed for some duration.
6942 0 : pub(crate) async fn metadata_health_list_outdated(
6943 0 : &self,
6944 0 : not_scrubbed_for: Duration,
6945 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
6946 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
6947 0 : let result = self
6948 0 : .persistence
6949 0 : .list_outdated_metadata_health_records(earlier)
6950 0 : .await?
6951 0 : .into_iter()
6952 0 : .map(|record| record.into())
6953 0 : .collect();
6954 0 : Ok(result)
6955 0 : }
6956 :
6957 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
6958 0 : self.inner.read().unwrap().get_leadership_status()
6959 0 : }
6960 :
6961 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
6962 0 : tracing::info!("Received step down request from peer");
6963 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
6964 :
6965 0 : self.inner.write().unwrap().step_down();
6966 0 : // TODO: would it make sense to have a time-out for this?
6967 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
6968 0 : .await;
6969 :
6970 0 : let mut global_observed = GlobalObservedState::default();
6971 0 : let locked = self.inner.read().unwrap();
6972 0 : for (tid, tenant_shard) in locked.tenants.iter() {
6973 0 : global_observed
6974 0 : .0
6975 0 : .insert(*tid, tenant_shard.observed.clone());
6976 0 : }
6977 :
6978 0 : global_observed
6979 0 : }
6980 :
6981 0 : pub(crate) async fn get_safekeeper(
6982 0 : &self,
6983 0 : id: i64,
6984 0 : ) -> Result<crate::persistence::SafekeeperPersistence, DatabaseError> {
6985 0 : self.persistence.safekeeper_get(id).await
6986 0 : }
6987 :
6988 0 : pub(crate) async fn upsert_safekeeper(
6989 0 : &self,
6990 0 : record: crate::persistence::SafekeeperPersistence,
6991 0 : ) -> Result<(), DatabaseError> {
6992 0 : self.persistence.safekeeper_upsert(record).await
6993 0 : }
6994 :
6995 0 : pub(crate) async fn update_shards_preferred_azs(
6996 0 : &self,
6997 0 : req: ShardsPreferredAzsRequest,
6998 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
6999 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
7000 0 : let updated = self
7001 0 : .persistence
7002 0 : .set_tenant_shard_preferred_azs(preferred_azs)
7003 0 : .await
7004 0 : .map_err(|err| {
7005 0 : ApiError::InternalServerError(anyhow::anyhow!(
7006 0 : "Failed to persist preferred AZs: {err}"
7007 0 : ))
7008 0 : })?;
7009 :
7010 0 : let mut updated_in_mem_and_db = Vec::default();
7011 0 :
7012 0 : let mut locked = self.inner.write().unwrap();
7013 0 : for (tid, az_id) in updated {
7014 0 : let shard = locked.tenants.get_mut(&tid);
7015 0 : if let Some(shard) = shard {
7016 0 : shard.set_preferred_az(az_id);
7017 0 : updated_in_mem_and_db.push(tid);
7018 0 : }
7019 : }
7020 :
7021 0 : Ok(ShardsPreferredAzsResponse {
7022 0 : updated: updated_in_mem_and_db,
7023 0 : })
7024 0 : }
7025 : }
|
