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