Line data Source code
1 : pub mod chaos_injector;
2 : mod context_iterator;
3 :
4 : use hyper::Uri;
5 : use std::{
6 : borrow::Cow,
7 : cmp::Ordering,
8 : collections::{BTreeMap, HashMap, HashSet},
9 : error::Error,
10 : ops::Deref,
11 : path::PathBuf,
12 : str::FromStr,
13 : sync::Arc,
14 : time::{Duration, Instant},
15 : };
16 :
17 : use crate::{
18 : background_node_operations::{
19 : Drain, Fill, Operation, OperationError, OperationHandler, MAX_RECONCILES_PER_OPERATION,
20 : },
21 : compute_hook::{self, NotifyError},
22 : drain_utils::{self, TenantShardDrain, TenantShardIterator},
23 : id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, TracingExclusiveGuard},
24 : leadership::Leadership,
25 : metrics,
26 : peer_client::GlobalObservedState,
27 : persistence::{
28 : AbortShardSplitStatus, ControllerPersistence, DatabaseResult, MetadataHealthPersistence,
29 : ShardGenerationState, TenantFilter,
30 : },
31 : reconciler::{ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder},
32 : scheduler::{MaySchedule, ScheduleContext, ScheduleError, ScheduleMode},
33 : tenant_shard::{
34 : MigrateAttachment, ObservedStateDelta, ReconcileNeeded, ReconcilerStatus,
35 : ScheduleOptimization, ScheduleOptimizationAction,
36 : },
37 : };
38 : use anyhow::Context;
39 : use control_plane::storage_controller::{
40 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
41 : };
42 : use diesel::result::DatabaseErrorKind;
43 : use futures::{stream::FuturesUnordered, StreamExt};
44 : use itertools::Itertools;
45 : use pageserver_api::{
46 : controller_api::{
47 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
48 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
49 : SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest,
50 : ShardsPreferredAzsResponse, SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse,
51 : TenantCreateResponseShard, TenantDescribeResponse, TenantDescribeResponseShard,
52 : TenantLocateResponse, TenantPolicyRequest, TenantShardMigrateRequest,
53 : TenantShardMigrateResponse,
54 : },
55 : models::{
56 : SecondaryProgress, TenantConfigPatchRequest, TenantConfigRequest,
57 : TimelineArchivalConfigRequest, TopTenantShardsRequest,
58 : },
59 : };
60 : use reqwest::StatusCode;
61 : use tracing::{instrument, Instrument};
62 :
63 : use crate::pageserver_client::PageserverClient;
64 : use pageserver_api::{
65 : models::{
66 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
67 : PageserverUtilization, ShardParameters, TenantConfig, TenantLocationConfigRequest,
68 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
69 : TenantShardSplitResponse, TenantTimeTravelRequest, TimelineCreateRequest, TimelineInfo,
70 : },
71 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
72 : upcall_api::{
73 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
74 : ValidateResponse, ValidateResponseTenant,
75 : },
76 : };
77 : use pageserver_client::{mgmt_api, BlockUnblock};
78 : use tokio::sync::mpsc::error::TrySendError;
79 : use tokio_util::sync::CancellationToken;
80 : use utils::{
81 : completion::Barrier,
82 : failpoint_support,
83 : generation::Generation,
84 : http::error::ApiError,
85 : id::{NodeId, TenantId, TimelineId},
86 : pausable_failpoint,
87 : sync::gate::Gate,
88 : };
89 :
90 : use crate::{
91 : compute_hook::ComputeHook,
92 : heartbeater::{Heartbeater, PageserverState},
93 : node::{AvailabilityTransition, Node},
94 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
95 : reconciler::attached_location_conf,
96 : scheduler::Scheduler,
97 : tenant_shard::{
98 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
99 : ReconcilerWaiter, TenantShard,
100 : },
101 : };
102 :
103 : use context_iterator::TenantShardContextIterator;
104 :
105 : const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
106 :
107 : // For operations that should be quick, like attaching a new tenant
108 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
109 :
110 : // For operations that might be slow, like migrating a tenant with
111 : // some data in it.
112 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
113 :
114 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
115 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
116 : // input generation from future requests as authoritative.
117 : const INITIAL_GENERATION: Generation = Generation::new(0);
118 :
119 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
120 : /// up on unresponsive pageservers and proceed.
121 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
122 :
123 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
124 : /// This must be long enough to cover node restarts as well as normal operations: in future
125 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
126 :
127 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
128 : /// offline.
129 : ///
130 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
131 : /// handling of the re-attach response may take a long time and blocks heartbeats from
132 : /// being handled on the pageserver side.
133 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
134 :
135 : /// How often to send heartbeats to registered nodes?
136 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
137 :
138 : /// How long is too long for a reconciliation?
139 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
140 :
141 : #[derive(Clone, strum_macros::Display)]
142 : enum TenantOperations {
143 : Create,
144 : LocationConfig,
145 : ConfigSet,
146 : ConfigPatch,
147 : TimeTravelRemoteStorage,
148 : Delete,
149 : UpdatePolicy,
150 : ShardSplit,
151 : SecondaryDownload,
152 : TimelineCreate,
153 : TimelineDelete,
154 : AttachHook,
155 : TimelineArchivalConfig,
156 : TimelineDetachAncestor,
157 : TimelineGcBlockUnblock,
158 : DropDetached,
159 : }
160 :
161 : #[derive(Clone, strum_macros::Display)]
162 : enum NodeOperations {
163 : Register,
164 : Configure,
165 : Delete,
166 : }
167 :
168 : /// The leadership status for the storage controller process.
169 : /// Allowed transitions are:
170 : /// 1. Leader -> SteppedDown
171 : /// 2. Candidate -> Leader
172 : #[derive(
173 : Eq,
174 : PartialEq,
175 : Copy,
176 : Clone,
177 : strum_macros::Display,
178 0 : strum_macros::EnumIter,
179 : measured::FixedCardinalityLabel,
180 : )]
181 : #[strum(serialize_all = "snake_case")]
182 : pub(crate) enum LeadershipStatus {
183 : /// This is the steady state where the storage controller can produce
184 : /// side effects in the cluster.
185 : Leader,
186 : /// We've been notified to step down by another candidate. No reconciliations
187 : /// take place in this state.
188 : SteppedDown,
189 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
190 : #[allow(unused)]
191 : Candidate,
192 : }
193 :
194 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
195 :
196 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
197 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
198 : // than they're being pushed onto the queue.
199 : const MAX_DELAYED_RECONCILES: usize = 10000;
200 :
201 : // Top level state available to all HTTP handlers
202 : struct ServiceState {
203 : leadership_status: LeadershipStatus,
204 :
205 : tenants: BTreeMap<TenantShardId, TenantShard>,
206 :
207 : nodes: Arc<HashMap<NodeId, Node>>,
208 :
209 : scheduler: Scheduler,
210 :
211 : /// Ongoing background operation on the cluster if any is running.
212 : /// Note that only one such operation may run at any given time,
213 : /// hence the type choice.
214 : ongoing_operation: Option<OperationHandler>,
215 :
216 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
217 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
218 : }
219 :
220 : /// Transform an error from a pageserver into an error to return to callers of a storage
221 : /// controller API.
222 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
223 0 : match e {
224 0 : mgmt_api::Error::SendRequest(e) => {
225 0 : // Presume errors sending requests are connectivity/availability issues
226 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
227 : }
228 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
229 0 : // Presume errors receiving body are connectivity/availability issues
230 0 : ApiError::ResourceUnavailable(
231 0 : format!("{node} error receiving error body: {str}").into(),
232 0 : )
233 : }
234 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
235 0 : // Return 500 for decoding errors.
236 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
237 : }
238 0 : mgmt_api::Error::ReceiveBody(err) => {
239 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
240 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
241 0 : ApiError::ResourceUnavailable(
242 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
243 0 : )
244 : }
245 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
246 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
247 : }
248 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
249 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
250 : }
251 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
252 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
253 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
254 : // internal server errors, showing that something is wrong with the pageserver or
255 : // storage controller's auth configuration.
256 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
257 : }
258 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
259 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
260 0 : // turn ask our clients to wait + retry
261 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
262 : }
263 0 : mgmt_api::Error::ApiError(status, msg) => {
264 0 : // Presume general case of pageserver API errors is that we tried to do something
265 0 : // that can't be done right now.
266 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
267 : }
268 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
269 : }
270 0 : }
271 :
272 : impl ServiceState {
273 0 : fn new(
274 0 : nodes: HashMap<NodeId, Node>,
275 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
276 0 : scheduler: Scheduler,
277 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
278 0 : initial_leadership_status: LeadershipStatus,
279 0 : ) -> Self {
280 0 : metrics::update_leadership_status(initial_leadership_status);
281 0 :
282 0 : Self {
283 0 : leadership_status: initial_leadership_status,
284 0 : tenants,
285 0 : nodes: Arc::new(nodes),
286 0 : scheduler,
287 0 : ongoing_operation: None,
288 0 : delayed_reconcile_rx,
289 0 : }
290 0 : }
291 :
292 0 : fn parts_mut(
293 0 : &mut self,
294 0 : ) -> (
295 0 : &mut Arc<HashMap<NodeId, Node>>,
296 0 : &mut BTreeMap<TenantShardId, TenantShard>,
297 0 : &mut Scheduler,
298 0 : ) {
299 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
300 0 : }
301 :
302 0 : fn get_leadership_status(&self) -> LeadershipStatus {
303 0 : self.leadership_status
304 0 : }
305 :
306 0 : fn step_down(&mut self) {
307 0 : self.leadership_status = LeadershipStatus::SteppedDown;
308 0 : metrics::update_leadership_status(self.leadership_status);
309 0 : }
310 :
311 0 : fn become_leader(&mut self) {
312 0 : self.leadership_status = LeadershipStatus::Leader;
313 0 : metrics::update_leadership_status(self.leadership_status);
314 0 : }
315 : }
316 :
317 : #[derive(Clone)]
318 : pub struct Config {
319 : // All pageservers managed by one instance of this service must have
320 : // the same public key. This JWT token will be used to authenticate
321 : // this service to the pageservers it manages.
322 : pub jwt_token: Option<String>,
323 :
324 : // This JWT token will be used to authenticate this service to the control plane.
325 : pub control_plane_jwt_token: Option<String>,
326 :
327 : // This JWT token will be used to authenticate with other storage controller instances
328 : pub peer_jwt_token: Option<String>,
329 :
330 : /// Where the compute hook should send notifications of pageserver attachment locations
331 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
332 : /// assume it is running in a test environment and try to update neon_local.
333 : pub compute_hook_url: Option<String>,
334 :
335 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
336 : /// considered active. Once the grace period elapses, the next heartbeat failure will
337 : /// mark the pagseserver offline.
338 : pub max_offline_interval: Duration,
339 :
340 : /// Extended grace period within which pageserver may not respond to heartbeats.
341 : /// This extended grace period kicks in after the node has been drained for restart
342 : /// and/or upon handling the re-attach request from a node.
343 : pub max_warming_up_interval: Duration,
344 :
345 : /// How many Reconcilers may be spawned concurrently
346 : pub reconciler_concurrency: usize,
347 :
348 : /// How large must a shard grow in bytes before we split it?
349 : /// None disables auto-splitting.
350 : pub split_threshold: Option<u64>,
351 :
352 : // TODO: make this cfg(feature = "testing")
353 : pub neon_local_repo_dir: Option<PathBuf>,
354 :
355 : // Maximum acceptable download lag for the secondary location
356 : // while draining a node. If the secondary location is lagging
357 : // by more than the configured amount, then the secondary is not
358 : // upgraded to primary.
359 : pub max_secondary_lag_bytes: Option<u64>,
360 :
361 : pub heartbeat_interval: Duration,
362 :
363 : pub address_for_peers: Option<Uri>,
364 :
365 : pub start_as_candidate: bool,
366 :
367 : pub http_service_port: i32,
368 :
369 : pub long_reconcile_threshold: Duration,
370 : }
371 :
372 : impl From<DatabaseError> for ApiError {
373 0 : fn from(err: DatabaseError) -> ApiError {
374 0 : match err {
375 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
376 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
377 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
378 0 : ApiError::ShuttingDown
379 : }
380 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
381 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
382 : }
383 : }
384 0 : }
385 : }
386 :
387 : enum InitialShardScheduleOutcome {
388 : Scheduled(TenantCreateResponseShard),
389 : NotScheduled,
390 : ShardScheduleError(ScheduleError),
391 : }
392 :
393 : pub struct Service {
394 : inner: Arc<std::sync::RwLock<ServiceState>>,
395 : config: Config,
396 : persistence: Arc<Persistence>,
397 : compute_hook: Arc<ComputeHook>,
398 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
399 :
400 : heartbeater: Heartbeater,
401 :
402 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
403 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
404 :
405 : // Locking on a tenant granularity (covers all shards in the tenant):
406 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
407 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
408 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
409 :
410 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
411 : // that transition it to/from Active.
412 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
413 :
414 : // Limit how many Reconcilers we will spawn concurrently
415 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
416 :
417 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
418 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
419 : ///
420 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
421 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
422 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
423 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
424 :
425 : // Process shutdown will fire this token
426 : cancel: CancellationToken,
427 :
428 : // Child token of [`Service::cancel`] used by reconcilers
429 : reconcilers_cancel: CancellationToken,
430 :
431 : // Background tasks will hold this gate
432 : gate: Gate,
433 :
434 : // Reconcilers background tasks will hold this gate
435 : reconcilers_gate: Gate,
436 :
437 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
438 : /// passes, it isn't safe to do any actions that mutate tenants.
439 : pub(crate) startup_complete: Barrier,
440 : }
441 :
442 : impl From<ReconcileWaitError> for ApiError {
443 0 : fn from(value: ReconcileWaitError) -> Self {
444 0 : match value {
445 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
446 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
447 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
448 : }
449 0 : }
450 : }
451 :
452 : impl From<OperationError> for ApiError {
453 0 : fn from(value: OperationError) -> Self {
454 0 : match value {
455 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
456 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
457 : }
458 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
459 : }
460 0 : }
461 : }
462 :
463 : #[allow(clippy::large_enum_variant)]
464 : enum TenantCreateOrUpdate {
465 : Create(TenantCreateRequest),
466 : Update(Vec<ShardUpdate>),
467 : }
468 :
469 : struct ShardSplitParams {
470 : old_shard_count: ShardCount,
471 : new_shard_count: ShardCount,
472 : new_stripe_size: Option<ShardStripeSize>,
473 : targets: Vec<ShardSplitTarget>,
474 : policy: PlacementPolicy,
475 : config: TenantConfig,
476 : shard_ident: ShardIdentity,
477 : preferred_az_id: Option<AvailabilityZone>,
478 : }
479 :
480 : // When preparing for a shard split, we may either choose to proceed with the split,
481 : // or find that the work is already done and return NoOp.
482 : enum ShardSplitAction {
483 : Split(Box<ShardSplitParams>),
484 : NoOp(TenantShardSplitResponse),
485 : }
486 :
487 : // A parent shard which will be split
488 : struct ShardSplitTarget {
489 : parent_id: TenantShardId,
490 : node: Node,
491 : child_ids: Vec<TenantShardId>,
492 : }
493 :
494 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
495 : /// might not be available. We therefore use a queue of abort operations processed in the background.
496 : struct TenantShardSplitAbort {
497 : tenant_id: TenantId,
498 : /// The target values from the request that failed
499 : new_shard_count: ShardCount,
500 : new_stripe_size: Option<ShardStripeSize>,
501 : /// Until this abort op is complete, no other operations may be done on the tenant
502 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
503 : }
504 :
505 : #[derive(thiserror::Error, Debug)]
506 : enum TenantShardSplitAbortError {
507 : #[error(transparent)]
508 : Database(#[from] DatabaseError),
509 : #[error(transparent)]
510 : Remote(#[from] mgmt_api::Error),
511 : #[error("Unavailable")]
512 : Unavailable,
513 : }
514 :
515 : struct ShardUpdate {
516 : tenant_shard_id: TenantShardId,
517 : placement_policy: PlacementPolicy,
518 : tenant_config: TenantConfig,
519 :
520 : /// If this is None, generation is not updated.
521 : generation: Option<Generation>,
522 :
523 : /// If this is None, scheduling policy is not updated.
524 : scheduling_policy: Option<ShardSchedulingPolicy>,
525 : }
526 :
527 : enum StopReconciliationsReason {
528 : ShuttingDown,
529 : SteppingDown,
530 : }
531 :
532 : impl std::fmt::Display for StopReconciliationsReason {
533 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
534 0 : let s = match self {
535 0 : Self::ShuttingDown => "Shutting down",
536 0 : Self::SteppingDown => "Stepping down",
537 : };
538 0 : write!(writer, "{}", s)
539 0 : }
540 : }
541 :
542 : pub(crate) enum ReconcileResultRequest {
543 : ReconcileResult(ReconcileResult),
544 : Stop,
545 : }
546 :
547 : #[derive(Clone)]
548 : struct MutationLocation {
549 : node: Node,
550 : generation: Generation,
551 : }
552 :
553 : #[derive(Clone)]
554 : struct ShardMutationLocations {
555 : latest: MutationLocation,
556 : other: Vec<MutationLocation>,
557 : }
558 :
559 : #[derive(Default, Clone)]
560 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
561 :
562 : impl Service {
563 0 : pub fn get_config(&self) -> &Config {
564 0 : &self.config
565 0 : }
566 :
567 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
568 : /// view of the world, and determine which pageservers are responsive.
569 : #[instrument(skip_all)]
570 : async fn startup_reconcile(
571 : self: &Arc<Service>,
572 : current_leader: Option<ControllerPersistence>,
573 : leader_step_down_state: Option<GlobalObservedState>,
574 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
575 : Result<(), (TenantShardId, NotifyError)>,
576 : >,
577 : ) {
578 : // Startup reconciliation does I/O to other services: whether they
579 : // are responsive or not, we should aim to finish within our deadline, because:
580 : // - If we don't, a k8s readiness hook watching /ready will kill us.
581 : // - While we're waiting for startup reconciliation, we are not fully
582 : // available for end user operations like creating/deleting tenants and timelines.
583 : //
584 : // We set multiple deadlines to break up the time available between the phases of work: this is
585 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
586 : let start_at = Instant::now();
587 : let node_scan_deadline = start_at
588 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
589 : .expect("Reconcile timeout is a modest constant");
590 :
591 : let observed = if let Some(state) = leader_step_down_state {
592 : tracing::info!(
593 : "Using observed state received from leader at {}",
594 : current_leader.as_ref().unwrap().address
595 : );
596 :
597 : state
598 : } else {
599 : self.build_global_observed_state(node_scan_deadline).await
600 : };
601 :
602 : // Accumulate a list of any tenant locations that ought to be detached
603 : let mut cleanup = Vec::new();
604 :
605 : // Send initial heartbeat requests to all nodes loaded from the database
606 : let all_nodes = {
607 : let locked = self.inner.read().unwrap();
608 : locked.nodes.clone()
609 : };
610 : let mut nodes_online = self.initial_heartbeat_round(all_nodes.keys()).await;
611 :
612 : // List of tenants for which we will attempt to notify compute of their location at startup
613 : let mut compute_notifications = Vec::new();
614 :
615 : // Populate intent and observed states for all tenants, based on reported state on pageservers
616 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
617 : let shard_count = {
618 : let mut locked = self.inner.write().unwrap();
619 : let (nodes, tenants, scheduler) = locked.parts_mut();
620 :
621 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
622 : let mut new_nodes = (**nodes).clone();
623 : for (node_id, node) in new_nodes.iter_mut() {
624 : if let Some(utilization) = nodes_online.remove(node_id) {
625 : node.set_availability(NodeAvailability::Active(utilization));
626 : scheduler.node_upsert(node);
627 : }
628 : }
629 : *nodes = Arc::new(new_nodes);
630 :
631 : for (tenant_shard_id, observed_state) in observed.0 {
632 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
633 : for node_id in observed_state.locations.keys() {
634 : cleanup.push((tenant_shard_id, *node_id));
635 : }
636 :
637 : continue;
638 : };
639 :
640 : tenant_shard.observed = observed_state;
641 : }
642 :
643 : // Populate each tenant's intent state
644 : let mut schedule_context = ScheduleContext::default();
645 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
646 : if tenant_shard_id.shard_number == ShardNumber(0) {
647 : // Reset scheduling context each time we advance to the next Tenant
648 : schedule_context = ScheduleContext::default();
649 : }
650 :
651 : tenant_shard.intent_from_observed(scheduler);
652 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
653 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
654 : // not enough pageservers are available. The tenant may well still be available
655 : // to clients.
656 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
657 : } else {
658 : // If we're both intending and observed to be attached at a particular node, we will
659 : // emit a compute notification for this. In the case where our observed state does not
660 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
661 : if let Some(attached_at) = tenant_shard.stably_attached() {
662 : compute_notifications.push(compute_hook::ShardUpdate {
663 : tenant_shard_id: *tenant_shard_id,
664 : node_id: attached_at,
665 : stripe_size: tenant_shard.shard.stripe_size,
666 : preferred_az: tenant_shard
667 : .preferred_az()
668 0 : .map(|az| Cow::Owned(az.clone())),
669 : });
670 : }
671 : }
672 : }
673 :
674 : tenants.len()
675 : };
676 :
677 : // Before making any obeservable changes to the cluster, persist self
678 : // as leader in database and memory.
679 : let leadership = Leadership::new(
680 : self.persistence.clone(),
681 : self.config.clone(),
682 : self.cancel.child_token(),
683 : );
684 :
685 : if let Err(e) = leadership.become_leader(current_leader).await {
686 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
687 : std::process::exit(1);
688 : }
689 :
690 : self.inner.write().unwrap().become_leader();
691 :
692 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
693 : // generation_pageserver in the database.
694 :
695 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
696 : // will emit compute hook notifications when they reconcile.
697 : //
698 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
699 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
700 : // calls will be correctly ordered wrt these.
701 : //
702 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
703 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
704 : // unit and start doing I/O.
705 : tracing::info!(
706 : "Sending {} compute notifications",
707 : compute_notifications.len()
708 : );
709 : self.compute_hook.notify_background(
710 : compute_notifications,
711 : bg_compute_notify_result_tx.clone(),
712 : &self.cancel,
713 : );
714 :
715 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
716 : // which require it: under normal circumstances this should only include tenants that were in some
717 : // transient state before we restarted, or any tenants whose compute hooks failed above.
718 : tracing::info!("Checking for shards in need of reconciliation...");
719 : let reconcile_tasks = self.reconcile_all();
720 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
721 : // normal operations may proceed.
722 :
723 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
724 : // background because it does not need to complete in order to proceed with other work.
725 : if !cleanup.is_empty() {
726 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
727 : tokio::task::spawn({
728 : let cleanup_self = self.clone();
729 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
730 : });
731 : }
732 :
733 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
734 : }
735 :
736 0 : async fn initial_heartbeat_round<'a>(
737 0 : &self,
738 0 : node_ids: impl Iterator<Item = &'a NodeId>,
739 0 : ) -> HashMap<NodeId, PageserverUtilization> {
740 0 : assert!(!self.startup_complete.is_ready());
741 :
742 0 : let all_nodes = {
743 0 : let locked = self.inner.read().unwrap();
744 0 : locked.nodes.clone()
745 0 : };
746 0 :
747 0 : let mut nodes_to_heartbeat = HashMap::new();
748 0 : for node_id in node_ids {
749 0 : match all_nodes.get(node_id) {
750 0 : Some(node) => {
751 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
752 0 : }
753 : None => {
754 0 : tracing::warn!("Node {node_id} was removed during start-up");
755 : }
756 : }
757 : }
758 :
759 0 : tracing::info!("Sending initial heartbeats...");
760 0 : let res = self
761 0 : .heartbeater
762 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
763 0 : .await;
764 :
765 0 : let mut online_nodes = HashMap::new();
766 0 : if let Ok(deltas) = res {
767 0 : for (node_id, status) in deltas.0 {
768 0 : match status {
769 0 : PageserverState::Available { utilization, .. } => {
770 0 : online_nodes.insert(node_id, utilization);
771 0 : }
772 0 : PageserverState::Offline => {}
773 : PageserverState::WarmingUp { .. } => {
774 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
775 : }
776 : }
777 : }
778 0 : }
779 :
780 0 : online_nodes
781 0 : }
782 :
783 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
784 : ///
785 : /// The result includes only nodes which responded within the deadline
786 0 : async fn scan_node_locations(
787 0 : &self,
788 0 : deadline: Instant,
789 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
790 0 : let nodes = {
791 0 : let locked = self.inner.read().unwrap();
792 0 : locked.nodes.clone()
793 0 : };
794 0 :
795 0 : let mut node_results = HashMap::new();
796 0 :
797 0 : let mut node_list_futs = FuturesUnordered::new();
798 0 :
799 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
800 0 : for node in nodes.values() {
801 0 : node_list_futs.push({
802 0 : async move {
803 0 : tracing::info!("Scanning shards on node {node}...");
804 0 : let timeout = Duration::from_secs(5);
805 0 : let response = node
806 0 : .with_client_retries(
807 0 : |client| async move { client.list_location_config().await },
808 0 : &self.config.jwt_token,
809 0 : 1,
810 0 : 5,
811 0 : timeout,
812 0 : &self.cancel,
813 0 : )
814 0 : .await;
815 0 : (node.get_id(), response)
816 0 : }
817 0 : });
818 0 : }
819 :
820 : loop {
821 0 : let (node_id, result) = tokio::select! {
822 0 : next = node_list_futs.next() => {
823 0 : match next {
824 0 : Some(result) => result,
825 : None =>{
826 : // We got results for all our nodes
827 0 : break;
828 : }
829 :
830 : }
831 : },
832 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
833 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
834 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
835 0 : break;
836 : }
837 : };
838 :
839 0 : let Some(list_response) = result else {
840 0 : tracing::info!("Shutdown during startup_reconcile");
841 0 : break;
842 : };
843 :
844 0 : match list_response {
845 0 : Err(e) => {
846 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
847 : }
848 0 : Ok(listing) => {
849 0 : node_results.insert(node_id, listing);
850 0 : }
851 : }
852 : }
853 :
854 0 : node_results
855 0 : }
856 :
857 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
858 0 : let node_listings = self.scan_node_locations(deadline).await;
859 0 : let mut observed = GlobalObservedState::default();
860 :
861 0 : for (node_id, location_confs) in node_listings {
862 0 : tracing::info!(
863 0 : "Received {} shard statuses from pageserver {}",
864 0 : location_confs.tenant_shards.len(),
865 : node_id
866 : );
867 :
868 0 : for (tid, location_conf) in location_confs.tenant_shards {
869 0 : let entry = observed.0.entry(tid).or_default();
870 0 : entry.locations.insert(
871 0 : node_id,
872 0 : ObservedStateLocation {
873 0 : conf: location_conf,
874 0 : },
875 0 : );
876 0 : }
877 : }
878 :
879 0 : observed
880 0 : }
881 :
882 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
883 : ///
884 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
885 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
886 : /// other task trying to attach it.
887 : #[instrument(skip_all)]
888 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
889 : let nodes = self.inner.read().unwrap().nodes.clone();
890 :
891 : for (tenant_shard_id, node_id) in cleanup {
892 : // A node reported a tenant_shard_id which is unknown to us: detach it.
893 : let Some(node) = nodes.get(&node_id) else {
894 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
895 : // a location to clean up on a node that has since been removed.
896 : tracing::info!(
897 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
898 : );
899 : continue;
900 : };
901 :
902 : if self.cancel.is_cancelled() {
903 : break;
904 : }
905 :
906 : let client = PageserverClient::new(
907 : node.get_id(),
908 : node.base_url(),
909 : self.config.jwt_token.as_deref(),
910 : );
911 : match client
912 : .location_config(
913 : tenant_shard_id,
914 : LocationConfig {
915 : mode: LocationConfigMode::Detached,
916 : generation: None,
917 : secondary_conf: None,
918 : shard_number: tenant_shard_id.shard_number.0,
919 : shard_count: tenant_shard_id.shard_count.literal(),
920 : shard_stripe_size: 0,
921 : tenant_conf: models::TenantConfig::default(),
922 : },
923 : None,
924 : false,
925 : )
926 : .await
927 : {
928 : Ok(()) => {
929 : tracing::info!(
930 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
931 : );
932 : }
933 : Err(e) => {
934 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
935 : // break anything.
936 : tracing::error!(
937 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
938 : );
939 : }
940 : }
941 : }
942 : }
943 :
944 : /// Long running background task that periodically wakes up and looks for shards that need
945 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
946 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
947 : /// for those retries.
948 : #[instrument(skip_all)]
949 : async fn background_reconcile(self: &Arc<Self>) {
950 : self.startup_complete.clone().wait().await;
951 :
952 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
953 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
954 : while !self.reconcilers_cancel.is_cancelled() {
955 : tokio::select! {
956 : _ = interval.tick() => {
957 : let reconciles_spawned = self.reconcile_all();
958 : if reconciles_spawned == 0 {
959 : // Run optimizer only when we didn't find any other work to do
960 : let optimizations = self.optimize_all().await;
961 : if optimizations == 0 {
962 : // Run new splits only when no optimizations are pending
963 : self.autosplit_tenants().await;
964 : }
965 : }
966 : }
967 : _ = self.reconcilers_cancel.cancelled() => return
968 : }
969 : }
970 : }
971 : #[instrument(skip_all)]
972 : async fn spawn_heartbeat_driver(&self) {
973 : self.startup_complete.clone().wait().await;
974 :
975 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
976 : while !self.cancel.is_cancelled() {
977 : tokio::select! {
978 : _ = interval.tick() => { }
979 : _ = self.cancel.cancelled() => return
980 : };
981 :
982 : let nodes = {
983 : let locked = self.inner.read().unwrap();
984 : locked.nodes.clone()
985 : };
986 :
987 : let res = self.heartbeater.heartbeat(nodes).await;
988 : if let Ok(deltas) = res {
989 : let mut to_handle = Vec::default();
990 :
991 : for (node_id, state) in deltas.0 {
992 : let new_availability = match state {
993 : PageserverState::Available { utilization, .. } => {
994 : NodeAvailability::Active(utilization)
995 : }
996 : PageserverState::WarmingUp { started_at } => {
997 : NodeAvailability::WarmingUp(started_at)
998 : }
999 : PageserverState::Offline => {
1000 : // The node might have been placed in the WarmingUp state
1001 : // while the heartbeat round was on-going. Hence, filter out
1002 : // offline transitions for WarmingUp nodes that are still within
1003 : // their grace period.
1004 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1005 : .get_node(node_id)
1006 : .await
1007 : .as_ref()
1008 0 : .map(|n| n.get_availability())
1009 : {
1010 : let now = Instant::now();
1011 : if now - *started_at >= self.config.max_warming_up_interval {
1012 : NodeAvailability::Offline
1013 : } else {
1014 : NodeAvailability::WarmingUp(*started_at)
1015 : }
1016 : } else {
1017 : NodeAvailability::Offline
1018 : }
1019 : }
1020 : };
1021 :
1022 : let node_lock = trace_exclusive_lock(
1023 : &self.node_op_locks,
1024 : node_id,
1025 : NodeOperations::Configure,
1026 : )
1027 : .await;
1028 :
1029 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1030 :
1031 : // This is the code path for geniune availability transitions (i.e node
1032 : // goes unavailable and/or comes back online).
1033 : let res = self
1034 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1035 : .await;
1036 :
1037 : match res {
1038 : Ok(transition) => {
1039 : // Keep hold of the lock until the availability transitions
1040 : // have been handled in
1041 : // [`Service::handle_node_availability_transitions`] in order avoid
1042 : // racing with [`Service::external_node_configure`].
1043 : to_handle.push((node_id, node_lock, transition));
1044 : }
1045 : Err(ApiError::NotFound(_)) => {
1046 : // This should be rare, but legitimate since the heartbeats are done
1047 : // on a snapshot of the nodes.
1048 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1049 : }
1050 : Err(ApiError::ShuttingDown) => {
1051 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1052 : }
1053 : Err(err) => {
1054 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1055 : // becomes unavailable again, we may get an error here.
1056 : tracing::error!(
1057 : "Failed to update node state {} after heartbeat round: {}",
1058 : node_id,
1059 : err
1060 : );
1061 : }
1062 : }
1063 : }
1064 :
1065 : // We collected all the transitions above and now we handle them.
1066 : let res = self.handle_node_availability_transitions(to_handle).await;
1067 : if let Err(errs) = res {
1068 : for (node_id, err) in errs {
1069 : match err {
1070 : ApiError::NotFound(_) => {
1071 : // This should be rare, but legitimate since the heartbeats are done
1072 : // on a snapshot of the nodes.
1073 : tracing::info!(
1074 : "Node {} was not found after heartbeat round",
1075 : node_id
1076 : );
1077 : }
1078 : err => {
1079 : tracing::error!(
1080 : "Failed to handle availability transition for {} after heartbeat round: {}",
1081 : node_id,
1082 : err
1083 : );
1084 : }
1085 : }
1086 : }
1087 : }
1088 : }
1089 : }
1090 : }
1091 :
1092 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1093 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1094 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1095 : /// will indicate that reconciliation is not needed.
1096 : #[instrument(skip_all, fields(
1097 : seq=%result.sequence,
1098 : tenant_id=%result.tenant_shard_id.tenant_id,
1099 : shard_id=%result.tenant_shard_id.shard_slug(),
1100 : ))]
1101 : fn process_result(&self, result: ReconcileResult) {
1102 : let mut locked = self.inner.write().unwrap();
1103 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1104 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1105 : // A reconciliation result might race with removing a tenant: drop results for
1106 : // tenants that aren't in our map.
1107 : return;
1108 : };
1109 :
1110 : // Usually generation should only be updated via this path, so the max() isn't
1111 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1112 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1113 :
1114 : // If the reconciler signals that it failed to notify compute, set this state on
1115 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1116 : tenant.pending_compute_notification = result.pending_compute_notification;
1117 :
1118 : // Let the TenantShard know it is idle.
1119 : tenant.reconcile_complete(result.sequence);
1120 :
1121 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1122 : // make to the tenant
1123 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1124 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1125 : // make to the tenant
1126 0 : let node = nodes.get(delta.node_id())?;
1127 :
1128 0 : if node.is_available() {
1129 0 : return Some(delta);
1130 0 : }
1131 0 :
1132 0 : // In case a node became unavailable concurrently with the reconcile, observed
1133 0 : // locations on it are now uncertain. By convention, set them to None in order
1134 0 : // for them to get refreshed when the node comes back online.
1135 0 : Some(ObservedStateDelta::Upsert(Box::new((
1136 0 : node.get_id(),
1137 0 : ObservedStateLocation { conf: None },
1138 0 : ))))
1139 0 : });
1140 :
1141 : match result.result {
1142 : Ok(()) => {
1143 : tenant.apply_observed_deltas(deltas);
1144 : tenant.waiter.advance(result.sequence);
1145 : }
1146 : Err(e) => {
1147 : match e {
1148 : ReconcileError::Cancel => {
1149 : tracing::info!("Reconciler was cancelled");
1150 : }
1151 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1152 : // This might be due to the reconciler getting cancelled, or it might
1153 : // be due to the `Node` being marked offline.
1154 : tracing::info!("Reconciler cancelled during pageserver API call");
1155 : }
1156 : _ => {
1157 : tracing::warn!("Reconcile error: {}", e);
1158 : }
1159 : }
1160 :
1161 : // Ordering: populate last_error before advancing error_seq,
1162 : // so that waiters will see the correct error after waiting.
1163 : tenant.set_last_error(result.sequence, e);
1164 :
1165 : // Skip deletions on reconcile failures
1166 : let upsert_deltas =
1167 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1168 : tenant.apply_observed_deltas(upsert_deltas);
1169 : }
1170 : }
1171 :
1172 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1173 : if tenant.policy == PlacementPolicy::Detached {
1174 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1175 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1176 : // duration of the request.
1177 : let guard = self.tenant_op_locks.try_exclusive(
1178 : tenant.tenant_shard_id.tenant_id,
1179 : TenantOperations::DropDetached,
1180 : );
1181 : if let Some(guard) = guard {
1182 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1183 : }
1184 : }
1185 :
1186 : // Maybe some other work can proceed now that this job finished.
1187 : if self.reconciler_concurrency.available_permits() > 0 {
1188 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1189 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1190 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1191 : shard.delayed_reconcile = false;
1192 : self.maybe_reconcile_shard(shard, nodes);
1193 : }
1194 :
1195 : if self.reconciler_concurrency.available_permits() == 0 {
1196 : break;
1197 : }
1198 : }
1199 : }
1200 : }
1201 :
1202 0 : async fn process_results(
1203 0 : &self,
1204 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1205 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1206 0 : Result<(), (TenantShardId, NotifyError)>,
1207 0 : >,
1208 0 : ) {
1209 : loop {
1210 : // Wait for the next result, or for cancellation
1211 0 : tokio::select! {
1212 0 : r = result_rx.recv() => {
1213 0 : match r {
1214 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1215 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1216 : }
1217 : }
1218 0 : _ = async{
1219 0 : match bg_compute_hook_result_rx.recv().await {
1220 0 : Some(result) => {
1221 0 : if let Err((tenant_shard_id, notify_error)) = result {
1222 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1223 0 : let mut locked = self.inner.write().unwrap();
1224 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1225 0 : shard.pending_compute_notification = true;
1226 0 : }
1227 :
1228 0 : }
1229 : },
1230 : None => {
1231 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1232 0 : self.cancel.cancelled().await;
1233 : }
1234 : }
1235 0 : } => {},
1236 0 : _ = self.cancel.cancelled() => {
1237 0 : break;
1238 : }
1239 : };
1240 : }
1241 0 : }
1242 :
1243 0 : async fn process_aborts(
1244 0 : &self,
1245 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1246 0 : ) {
1247 : loop {
1248 : // Wait for the next result, or for cancellation
1249 0 : let op = tokio::select! {
1250 0 : r = abort_rx.recv() => {
1251 0 : match r {
1252 0 : Some(op) => {op},
1253 0 : None => {break;}
1254 : }
1255 : }
1256 0 : _ = self.cancel.cancelled() => {
1257 0 : break;
1258 : }
1259 : };
1260 :
1261 : // Retry until shutdown: we must keep this request object alive until it is properly
1262 : // processed, as it holds a lock guard that prevents other operations trying to do things
1263 : // to the tenant while it is in a weird part-split state.
1264 0 : while !self.cancel.is_cancelled() {
1265 0 : match self.abort_tenant_shard_split(&op).await {
1266 0 : Ok(_) => break,
1267 0 : Err(e) => {
1268 0 : tracing::warn!(
1269 0 : "Failed to abort shard split on {}, will retry: {e}",
1270 : op.tenant_id
1271 : );
1272 :
1273 : // If a node is unavailable, we hope that it has been properly marked Offline
1274 : // when we retry, so that the abort op will succeed. If the abort op is failing
1275 : // for some other reason, we will keep retrying forever, or until a human notices
1276 : // and does something about it (either fixing a pageserver or restarting the controller).
1277 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1278 0 : .await
1279 0 : .ok();
1280 : }
1281 : }
1282 : }
1283 : }
1284 0 : }
1285 :
1286 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1287 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1288 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1289 0 :
1290 0 : let leadership_cancel = CancellationToken::new();
1291 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1292 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1293 :
1294 : // Apply the migrations **after** the current leader has stepped down
1295 : // (or we've given up waiting for it), but **before** reading from the
1296 : // database. The only exception is reading the current leader before
1297 : // migrating.
1298 0 : persistence.migration_run().await?;
1299 :
1300 0 : tracing::info!("Loading nodes from database...");
1301 0 : let nodes = persistence
1302 0 : .list_nodes()
1303 0 : .await?
1304 0 : .into_iter()
1305 0 : .map(Node::from_persistent)
1306 0 : .collect::<Vec<_>>();
1307 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1308 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1309 0 : metrics::METRICS_REGISTRY
1310 0 : .metrics_group
1311 0 : .storage_controller_pageserver_nodes
1312 0 : .set(nodes.len() as i64);
1313 0 :
1314 0 : tracing::info!("Loading shards from database...");
1315 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1316 0 : tracing::info!(
1317 0 : "Loaded {} shards from database.",
1318 0 : tenant_shard_persistence.len()
1319 : );
1320 :
1321 : // If any shard splits were in progress, reset the database state to abort them
1322 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1323 0 : HashMap::new();
1324 0 : for tsp in &mut tenant_shard_persistence {
1325 0 : let shard = tsp.get_shard_identity()?;
1326 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1327 0 : let entry = tenant_shard_count_min_max
1328 0 : .entry(tenant_shard_id.tenant_id)
1329 0 : .or_insert_with(|| (shard.count, shard.count));
1330 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1331 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1332 0 : }
1333 :
1334 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1335 0 : if count_min != count_max {
1336 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1337 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1338 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1339 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1340 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1341 :
1342 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1343 : // identified this tenant has having mismatching min/max counts.
1344 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1345 :
1346 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1347 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1348 0 : // Set idle split state on those shards that we will retain.
1349 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1350 0 : if tsp_tenant_id == tenant_id
1351 0 : && tsp.get_shard_identity().unwrap().count == count_min
1352 0 : {
1353 0 : tsp.splitting = SplitState::Idle;
1354 0 : } else if tsp_tenant_id == tenant_id {
1355 : // Leave the splitting state on the child shards: this will be used next to
1356 : // drop them.
1357 0 : tracing::info!(
1358 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1359 : );
1360 0 : }
1361 0 : });
1362 0 :
1363 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1364 0 : tenant_shard_persistence.retain(|tsp| {
1365 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1366 0 : || tsp.splitting == SplitState::Idle
1367 0 : });
1368 0 : }
1369 : }
1370 :
1371 0 : let mut tenants = BTreeMap::new();
1372 0 :
1373 0 : let mut scheduler = Scheduler::new(nodes.values());
1374 :
1375 : #[cfg(feature = "testing")]
1376 : {
1377 : use pageserver_api::controller_api::AvailabilityZone;
1378 :
1379 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1380 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1381 : // after when pageservers start up and register.
1382 0 : let mut node_ids = HashSet::new();
1383 0 : for tsp in &tenant_shard_persistence {
1384 0 : if let Some(node_id) = tsp.generation_pageserver {
1385 0 : node_ids.insert(node_id);
1386 0 : }
1387 : }
1388 0 : for node_id in node_ids {
1389 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1390 0 : let node = Node::new(
1391 0 : NodeId(node_id as u64),
1392 0 : "".to_string(),
1393 0 : 123,
1394 0 : "".to_string(),
1395 0 : 123,
1396 0 : AvailabilityZone("test_az".to_string()),
1397 0 : );
1398 0 :
1399 0 : scheduler.node_upsert(&node);
1400 : }
1401 : }
1402 0 : for tsp in tenant_shard_persistence {
1403 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1404 :
1405 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1406 : // it with what we can infer: the node for which a generation was most recently issued.
1407 0 : let mut intent = IntentState::new(
1408 0 : tsp.preferred_az_id
1409 0 : .as_ref()
1410 0 : .map(|az| AvailabilityZone(az.clone())),
1411 0 : );
1412 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1413 : {
1414 0 : if nodes.contains_key(&generation_pageserver) {
1415 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1416 0 : } else {
1417 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1418 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1419 : // on different pageservers.
1420 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1421 : }
1422 0 : }
1423 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1424 :
1425 0 : tenants.insert(tenant_shard_id, new_tenant);
1426 : }
1427 :
1428 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1429 0 :
1430 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1431 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1432 0 : tokio::sync::mpsc::channel(512);
1433 0 :
1434 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1435 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1436 0 :
1437 0 : let cancel = CancellationToken::new();
1438 0 : let reconcilers_cancel = cancel.child_token();
1439 0 :
1440 0 : let heartbeater = Heartbeater::new(
1441 0 : config.jwt_token.clone(),
1442 0 : config.max_offline_interval,
1443 0 : config.max_warming_up_interval,
1444 0 : cancel.clone(),
1445 0 : );
1446 :
1447 0 : let initial_leadership_status = if config.start_as_candidate {
1448 0 : LeadershipStatus::Candidate
1449 : } else {
1450 0 : LeadershipStatus::Leader
1451 : };
1452 :
1453 0 : let this = Arc::new(Self {
1454 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1455 0 : nodes,
1456 0 : tenants,
1457 0 : scheduler,
1458 0 : delayed_reconcile_rx,
1459 0 : initial_leadership_status,
1460 0 : ))),
1461 0 : config: config.clone(),
1462 0 : persistence,
1463 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1464 0 : result_tx,
1465 0 : heartbeater,
1466 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1467 0 : config.reconciler_concurrency,
1468 0 : )),
1469 0 : delayed_reconcile_tx,
1470 0 : abort_tx,
1471 0 : startup_complete: startup_complete.clone(),
1472 0 : cancel,
1473 0 : reconcilers_cancel,
1474 0 : gate: Gate::default(),
1475 0 : reconcilers_gate: Gate::default(),
1476 0 : tenant_op_locks: Default::default(),
1477 0 : node_op_locks: Default::default(),
1478 0 : });
1479 0 :
1480 0 : let result_task_this = this.clone();
1481 0 : tokio::task::spawn(async move {
1482 : // Block shutdown until we're done (we must respect self.cancel)
1483 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1484 0 : result_task_this
1485 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1486 0 : .await
1487 0 : }
1488 0 : });
1489 0 :
1490 0 : tokio::task::spawn({
1491 0 : let this = this.clone();
1492 0 : async move {
1493 : // Block shutdown until we're done (we must respect self.cancel)
1494 0 : if let Ok(_gate) = this.gate.enter() {
1495 0 : this.process_aborts(abort_rx).await
1496 0 : }
1497 0 : }
1498 0 : });
1499 0 :
1500 0 : tokio::task::spawn({
1501 0 : let this = this.clone();
1502 0 : async move {
1503 0 : if let Ok(_gate) = this.gate.enter() {
1504 : loop {
1505 0 : tokio::select! {
1506 0 : _ = this.cancel.cancelled() => {
1507 0 : break;
1508 : },
1509 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1510 0 : };
1511 0 : this.tenant_op_locks.housekeeping();
1512 : }
1513 0 : }
1514 0 : }
1515 0 : });
1516 0 :
1517 0 : tokio::task::spawn({
1518 0 : let this = this.clone();
1519 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1520 0 : // is done.
1521 0 : let startup_completion = startup_completion.clone();
1522 0 : async move {
1523 : // Block shutdown until we're done (we must respect self.cancel)
1524 0 : let Ok(_gate) = this.gate.enter() else {
1525 0 : return;
1526 : };
1527 :
1528 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1529 0 : .await;
1530 :
1531 0 : drop(startup_completion);
1532 0 : }
1533 0 : });
1534 0 :
1535 0 : tokio::task::spawn({
1536 0 : let this = this.clone();
1537 0 : let startup_complete = startup_complete.clone();
1538 0 : async move {
1539 0 : startup_complete.wait().await;
1540 0 : this.background_reconcile().await;
1541 0 : }
1542 0 : });
1543 0 :
1544 0 : tokio::task::spawn({
1545 0 : let this = this.clone();
1546 0 : let startup_complete = startup_complete.clone();
1547 0 : async move {
1548 0 : startup_complete.wait().await;
1549 0 : this.spawn_heartbeat_driver().await;
1550 0 : }
1551 0 : });
1552 0 :
1553 0 : Ok(this)
1554 0 : }
1555 :
1556 0 : pub(crate) async fn attach_hook(
1557 0 : &self,
1558 0 : attach_req: AttachHookRequest,
1559 0 : ) -> anyhow::Result<AttachHookResponse> {
1560 0 : let _tenant_lock = trace_exclusive_lock(
1561 0 : &self.tenant_op_locks,
1562 0 : attach_req.tenant_shard_id.tenant_id,
1563 0 : TenantOperations::AttachHook,
1564 0 : )
1565 0 : .await;
1566 :
1567 : // This is a test hook. To enable using it on tenants that were created directly with
1568 : // the pageserver API (not via this service), we will auto-create any missing tenant
1569 : // shards with default state.
1570 0 : let insert = {
1571 0 : match self
1572 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1573 0 : .await
1574 : {
1575 0 : Ok(_) => false,
1576 0 : Err(ApiError::NotFound(_)) => true,
1577 0 : Err(e) => return Err(e.into()),
1578 : }
1579 : };
1580 :
1581 0 : if insert {
1582 0 : let tsp = TenantShardPersistence {
1583 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1584 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1585 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1586 0 : shard_stripe_size: 0,
1587 0 : generation: attach_req.generation_override.or(Some(0)),
1588 0 : generation_pageserver: None,
1589 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1590 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1591 0 : splitting: SplitState::default(),
1592 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1593 0 : .unwrap(),
1594 0 : preferred_az_id: None,
1595 0 : };
1596 0 :
1597 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1598 0 : Err(e) => match e {
1599 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1600 : DatabaseErrorKind::UniqueViolation,
1601 : _,
1602 : )) => {
1603 0 : tracing::info!(
1604 0 : "Raced with another request to insert tenant {}",
1605 : attach_req.tenant_shard_id
1606 : )
1607 : }
1608 0 : _ => return Err(e.into()),
1609 : },
1610 : Ok(()) => {
1611 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1612 :
1613 0 : let mut locked = self.inner.write().unwrap();
1614 0 : locked.tenants.insert(
1615 0 : attach_req.tenant_shard_id,
1616 0 : TenantShard::new(
1617 0 : attach_req.tenant_shard_id,
1618 0 : ShardIdentity::unsharded(),
1619 0 : PlacementPolicy::Attached(0),
1620 0 : None,
1621 0 : ),
1622 0 : );
1623 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1624 : }
1625 : }
1626 0 : }
1627 :
1628 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1629 0 : let maybe_tenant_conf = {
1630 0 : let locked = self.inner.write().unwrap();
1631 0 : locked
1632 0 : .tenants
1633 0 : .get(&attach_req.tenant_shard_id)
1634 0 : .map(|t| t.config.clone())
1635 0 : };
1636 0 :
1637 0 : match maybe_tenant_conf {
1638 0 : Some(conf) => {
1639 0 : let new_generation = self
1640 0 : .persistence
1641 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1642 0 : .await?;
1643 :
1644 : // Persist the placement policy update. This is required
1645 : // when we reattaching a detached tenant.
1646 0 : self.persistence
1647 0 : .update_tenant_shard(
1648 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1649 0 : Some(PlacementPolicy::Attached(0)),
1650 0 : Some(conf),
1651 0 : None,
1652 0 : None,
1653 0 : )
1654 0 : .await?;
1655 0 : Some(new_generation)
1656 : }
1657 : None => {
1658 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1659 : }
1660 : }
1661 : } else {
1662 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1663 0 : None
1664 : };
1665 :
1666 0 : let mut locked = self.inner.write().unwrap();
1667 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1668 0 :
1669 0 : let tenant_shard = tenants
1670 0 : .get_mut(&attach_req.tenant_shard_id)
1671 0 : .expect("Checked for existence above");
1672 :
1673 0 : if let Some(new_generation) = new_generation {
1674 0 : tenant_shard.generation = Some(new_generation);
1675 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1676 0 : } else {
1677 : // This is a detach notification. We must update placement policy to avoid re-attaching
1678 : // during background scheduling/reconciliation, or during storage controller restart.
1679 0 : assert!(attach_req.node_id.is_none());
1680 0 : tenant_shard.policy = PlacementPolicy::Detached;
1681 : }
1682 :
1683 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1684 0 : tracing::info!(
1685 : tenant_id = %attach_req.tenant_shard_id,
1686 : ps_id = %attaching_pageserver,
1687 : generation = ?tenant_shard.generation,
1688 0 : "issuing",
1689 : );
1690 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1691 0 : tracing::info!(
1692 : tenant_id = %attach_req.tenant_shard_id,
1693 : %ps_id,
1694 : generation = ?tenant_shard.generation,
1695 0 : "dropping",
1696 : );
1697 : } else {
1698 0 : tracing::info!(
1699 : tenant_id = %attach_req.tenant_shard_id,
1700 0 : "no-op: tenant already has no pageserver");
1701 : }
1702 0 : tenant_shard
1703 0 : .intent
1704 0 : .set_attached(scheduler, attach_req.node_id);
1705 0 :
1706 0 : tracing::info!(
1707 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1708 0 : attach_req.tenant_shard_id,
1709 0 : tenant_shard.generation,
1710 0 : // TODO: this is an odd number of 0xf's
1711 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1712 : );
1713 :
1714 : // Trick the reconciler into not doing anything for this tenant: this helps
1715 : // tests that manually configure a tenant on the pagesrever, and then call this
1716 : // attach hook: they don't want background reconciliation to modify what they
1717 : // did to the pageserver.
1718 : #[cfg(feature = "testing")]
1719 : {
1720 0 : if let Some(node_id) = attach_req.node_id {
1721 0 : tenant_shard.observed.locations = HashMap::from([(
1722 0 : node_id,
1723 0 : ObservedStateLocation {
1724 0 : conf: Some(attached_location_conf(
1725 0 : tenant_shard.generation.unwrap(),
1726 0 : &tenant_shard.shard,
1727 0 : &tenant_shard.config,
1728 0 : &PlacementPolicy::Attached(0),
1729 0 : )),
1730 0 : },
1731 0 : )]);
1732 0 : } else {
1733 0 : tenant_shard.observed.locations.clear();
1734 0 : }
1735 : }
1736 :
1737 0 : Ok(AttachHookResponse {
1738 0 : gen: attach_req
1739 0 : .node_id
1740 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1741 0 : })
1742 0 : }
1743 :
1744 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1745 0 : let locked = self.inner.read().unwrap();
1746 0 :
1747 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1748 0 :
1749 0 : InspectResponse {
1750 0 : attachment: tenant_shard.and_then(|s| {
1751 0 : s.intent
1752 0 : .get_attached()
1753 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1754 0 : }),
1755 0 : }
1756 0 : }
1757 :
1758 : // When the availability state of a node transitions to active, we must do a full reconciliation
1759 : // of LocationConfigs on that node. This is because while a node was offline:
1760 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1761 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1762 : //
1763 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1764 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1765 : // this function.
1766 : //
1767 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1768 : // for written for a single node rather than as a batch job for all nodes.
1769 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1770 : async fn node_activate_reconcile(
1771 : &self,
1772 : mut node: Node,
1773 : _lock: &TracingExclusiveGuard<NodeOperations>,
1774 : ) -> Result<(), ApiError> {
1775 : // This Node is a mutable local copy: we will set it active so that we can use its
1776 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1777 : // later.
1778 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1779 :
1780 : let configs = match node
1781 : .with_client_retries(
1782 0 : |client| async move { client.list_location_config().await },
1783 : &self.config.jwt_token,
1784 : 1,
1785 : 5,
1786 : SHORT_RECONCILE_TIMEOUT,
1787 : &self.cancel,
1788 : )
1789 : .await
1790 : {
1791 : None => {
1792 : // We're shutting down (the Node's cancellation token can't have fired, because
1793 : // we're the only scope that has a reference to it, and we didn't fire it).
1794 : return Err(ApiError::ShuttingDown);
1795 : }
1796 : Some(Err(e)) => {
1797 : // This node didn't succeed listing its locations: it may not proceed to active state
1798 : // as it is apparently unavailable.
1799 : return Err(ApiError::PreconditionFailed(
1800 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1801 : ));
1802 : }
1803 : Some(Ok(configs)) => configs,
1804 : };
1805 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1806 :
1807 : let mut cleanup = Vec::new();
1808 : {
1809 : let mut locked = self.inner.write().unwrap();
1810 :
1811 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1812 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1813 : cleanup.push(tenant_shard_id);
1814 : continue;
1815 : };
1816 : tenant_shard
1817 : .observed
1818 : .locations
1819 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1820 : }
1821 : }
1822 :
1823 : for tenant_shard_id in cleanup {
1824 : tracing::info!("Detaching {tenant_shard_id}");
1825 : match node
1826 : .with_client_retries(
1827 0 : |client| async move {
1828 0 : let config = LocationConfig {
1829 0 : mode: LocationConfigMode::Detached,
1830 0 : generation: None,
1831 0 : secondary_conf: None,
1832 0 : shard_number: tenant_shard_id.shard_number.0,
1833 0 : shard_count: tenant_shard_id.shard_count.literal(),
1834 0 : shard_stripe_size: 0,
1835 0 : tenant_conf: models::TenantConfig::default(),
1836 0 : };
1837 0 : client
1838 0 : .location_config(tenant_shard_id, config, None, false)
1839 0 : .await
1840 0 : },
1841 : &self.config.jwt_token,
1842 : 1,
1843 : 5,
1844 : SHORT_RECONCILE_TIMEOUT,
1845 : &self.cancel,
1846 : )
1847 : .await
1848 : {
1849 : None => {
1850 : // We're shutting down (the Node's cancellation token can't have fired, because
1851 : // we're the only scope that has a reference to it, and we didn't fire it).
1852 : return Err(ApiError::ShuttingDown);
1853 : }
1854 : Some(Err(e)) => {
1855 : // Do not let the node proceed to Active state if it is not responsive to requests
1856 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1857 : // detach completing: we should not let this node back into the set of nodes considered
1858 : // okay for scheduling.
1859 : return Err(ApiError::Conflict(format!(
1860 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1861 : )));
1862 : }
1863 : Some(Ok(_)) => {}
1864 : };
1865 : }
1866 :
1867 : Ok(())
1868 : }
1869 :
1870 0 : pub(crate) async fn re_attach(
1871 0 : &self,
1872 0 : reattach_req: ReAttachRequest,
1873 0 : ) -> Result<ReAttachResponse, ApiError> {
1874 0 : if let Some(register_req) = reattach_req.register {
1875 0 : self.node_register(register_req).await?;
1876 0 : }
1877 :
1878 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1879 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1880 :
1881 0 : tracing::info!(
1882 : node_id=%reattach_req.node_id,
1883 0 : "Incremented {} tenant shards' generations",
1884 0 : incremented_generations.len()
1885 : );
1886 :
1887 : // Apply the updated generation to our in-memory state, and
1888 : // gather discover secondary locations.
1889 0 : let mut locked = self.inner.write().unwrap();
1890 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1891 0 :
1892 0 : let mut response = ReAttachResponse {
1893 0 : tenants: Vec::new(),
1894 0 : };
1895 :
1896 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
1897 : // to call location_conf API with an old generation. Wait for cancellation to complete
1898 : // before responding to this request. Requires well implemented CancellationToken logic
1899 : // all the way to where we call location_conf. Even then, there can still be a location_conf
1900 : // request in flight over the network: TODO handle that by making location_conf API refuse
1901 : // to go backward in generations.
1902 :
1903 : // Scan through all shards, applying updates for ones where we updated generation
1904 : // and identifying shards that intend to have a secondary location on this node.
1905 0 : for (tenant_shard_id, shard) in tenants {
1906 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
1907 0 : let new_gen = *new_gen;
1908 0 : response.tenants.push(ReAttachResponseTenant {
1909 0 : id: *tenant_shard_id,
1910 0 : gen: Some(new_gen.into().unwrap()),
1911 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
1912 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
1913 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
1914 0 : // the stale/multi states at this point.
1915 0 : mode: LocationConfigMode::AttachedSingle,
1916 0 : });
1917 0 :
1918 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
1919 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
1920 : // Why can we update `observed` even though we're not sure our response will be received
1921 : // by the pageserver? Because the pageserver will not proceed with startup until
1922 : // it has processed response: if it loses it, we'll see another request and increment
1923 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
1924 0 : if let Some(conf) = observed.conf.as_mut() {
1925 0 : conf.generation = new_gen.into();
1926 0 : }
1927 0 : } else {
1928 0 : // This node has no observed state for the shard: perhaps it was offline
1929 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
1930 0 : // will be prompted to learn the location's state before it makes changes.
1931 0 : shard
1932 0 : .observed
1933 0 : .locations
1934 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
1935 0 : }
1936 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
1937 0 : // Ordering: pageserver will not accept /location_config requests until it has
1938 0 : // finished processing the response from re-attach. So we can update our in-memory state
1939 0 : // now, and be confident that we are not stamping on the result of some later location config.
1940 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
1941 0 : // so we might update observed state here, and then get over-written by some racing
1942 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
1943 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
1944 0 :
1945 0 : response.tenants.push(ReAttachResponseTenant {
1946 0 : id: *tenant_shard_id,
1947 0 : gen: None,
1948 0 : mode: LocationConfigMode::Secondary,
1949 0 : });
1950 0 :
1951 0 : // We must not update observed, because we have no guarantee that our
1952 0 : // response will be received by the pageserver. This could leave it
1953 0 : // falsely dirty, but the resulting reconcile should be idempotent.
1954 0 : }
1955 : }
1956 :
1957 : // We consider a node Active once we have composed a re-attach response, but we
1958 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
1959 : // implicitly synchronizes the LocationConfigs on the node.
1960 : //
1961 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
1962 : // but those requests will not be accepted by the node until it has finished processing
1963 : // the re-attach response.
1964 : //
1965 : // Additionally, reset the nodes scheduling policy to match the conditional update done
1966 : // in [`Persistence::re_attach`].
1967 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
1968 0 : let reset_scheduling = matches!(
1969 0 : node.get_scheduling(),
1970 : NodeSchedulingPolicy::PauseForRestart
1971 : | NodeSchedulingPolicy::Draining
1972 : | NodeSchedulingPolicy::Filling
1973 : );
1974 :
1975 0 : let mut new_nodes = (**nodes).clone();
1976 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
1977 0 : if reset_scheduling {
1978 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
1979 0 : }
1980 :
1981 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
1982 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
1983 0 :
1984 0 : scheduler.node_upsert(node);
1985 0 : let new_nodes = Arc::new(new_nodes);
1986 0 : *nodes = new_nodes;
1987 : } else {
1988 0 : tracing::error!(
1989 0 : "Reattaching node {} was removed while processing the request",
1990 : reattach_req.node_id
1991 : );
1992 : }
1993 0 : }
1994 :
1995 0 : Ok(response)
1996 0 : }
1997 :
1998 0 : pub(crate) async fn validate(
1999 0 : &self,
2000 0 : validate_req: ValidateRequest,
2001 0 : ) -> Result<ValidateResponse, DatabaseError> {
2002 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2003 : // in-memory generation for a shard
2004 0 : let in_memory_result = {
2005 0 : let mut in_memory_result = Vec::new();
2006 0 : let locked = self.inner.read().unwrap();
2007 0 : for req_tenant in validate_req.tenants {
2008 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2009 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
2010 0 : tracing::info!(
2011 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2012 : req_tenant.id,
2013 : req_tenant.gen,
2014 : tenant_shard.generation
2015 : );
2016 :
2017 0 : in_memory_result.push((req_tenant.id, Generation::new(req_tenant.gen), valid));
2018 : } else {
2019 : // This is legal: for example during a shard split the pageserver may still
2020 : // have deletions in its queue from the old pre-split shard, or after deletion
2021 : // of a tenant that was busy with compaction/gc while being deleted.
2022 0 : tracing::info!(
2023 0 : "Refusing deletion validation for missing shard {}",
2024 : req_tenant.id
2025 : );
2026 : }
2027 : }
2028 :
2029 0 : in_memory_result
2030 : };
2031 :
2032 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2033 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2034 0 : let db_generations = self
2035 0 : .persistence
2036 0 : .shard_generations(in_memory_result.iter().filter_map(|i| {
2037 0 : if i.2 {
2038 0 : Some(&i.0)
2039 : } else {
2040 0 : None
2041 : }
2042 0 : }))
2043 0 : .await?;
2044 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2045 0 :
2046 0 : let mut response = ValidateResponse {
2047 0 : tenants: Vec::new(),
2048 0 : };
2049 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2050 0 : let valid = if valid {
2051 0 : let db_generation = db_generations.get(&tenant_shard_id);
2052 0 : db_generation == Some(&Some(validate_generation))
2053 : } else {
2054 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2055 : // this prevents a pageserver from cleaning up an object in S3.
2056 0 : false
2057 : };
2058 :
2059 0 : response.tenants.push(ValidateResponseTenant {
2060 0 : id: tenant_shard_id,
2061 0 : valid,
2062 0 : })
2063 : }
2064 :
2065 0 : Ok(response)
2066 0 : }
2067 :
2068 0 : pub(crate) async fn tenant_create(
2069 0 : &self,
2070 0 : create_req: TenantCreateRequest,
2071 0 : ) -> Result<TenantCreateResponse, ApiError> {
2072 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2073 :
2074 : // Exclude any concurrent attempts to create/access the same tenant ID
2075 0 : let _tenant_lock = trace_exclusive_lock(
2076 0 : &self.tenant_op_locks,
2077 0 : create_req.new_tenant_id.tenant_id,
2078 0 : TenantOperations::Create,
2079 0 : )
2080 0 : .await;
2081 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2082 :
2083 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2084 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2085 : // accept compute notifications while it is in the process of creating. Reconciliation will
2086 : // be retried in the background.
2087 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2088 0 : }
2089 0 : Ok(response)
2090 0 : }
2091 :
2092 0 : pub(crate) async fn do_tenant_create(
2093 0 : &self,
2094 0 : create_req: TenantCreateRequest,
2095 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2096 0 : let placement_policy = create_req
2097 0 : .placement_policy
2098 0 : .clone()
2099 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2100 0 : .unwrap_or(PlacementPolicy::Attached(0));
2101 :
2102 : // This service expects to handle sharding itself: it is an error to try and directly create
2103 : // a particular shard here.
2104 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2105 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2106 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2107 0 : )));
2108 : } else {
2109 0 : create_req.new_tenant_id.tenant_id
2110 0 : };
2111 0 :
2112 0 : tracing::info!(
2113 0 : "Creating tenant {}, shard_count={:?}",
2114 : create_req.new_tenant_id,
2115 : create_req.shard_parameters.count,
2116 : );
2117 :
2118 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2119 0 : .map(|i| TenantShardId {
2120 0 : tenant_id,
2121 0 : shard_number: ShardNumber(i),
2122 0 : shard_count: create_req.shard_parameters.count,
2123 0 : })
2124 0 : .collect::<Vec<_>>();
2125 :
2126 : // If the caller specifies a None generation, it means "start from default". This is different
2127 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2128 : // an incompletely-onboarded tenant.
2129 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2130 0 : tracing::info!(
2131 0 : "tenant_create: secondary mode, generation is_some={}",
2132 0 : create_req.generation.is_some()
2133 : );
2134 0 : create_req.generation.map(Generation::new)
2135 : } else {
2136 0 : tracing::info!(
2137 0 : "tenant_create: not secondary mode, generation is_some={}",
2138 0 : create_req.generation.is_some()
2139 : );
2140 0 : Some(
2141 0 : create_req
2142 0 : .generation
2143 0 : .map(Generation::new)
2144 0 : .unwrap_or(INITIAL_GENERATION),
2145 0 : )
2146 : };
2147 :
2148 0 : let preferred_az_id = {
2149 0 : let locked = self.inner.read().unwrap();
2150 : // Idempotency: take the existing value if the tenant already exists
2151 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2152 0 : shard.preferred_az().cloned()
2153 : } else {
2154 0 : locked.scheduler.get_az_for_new_tenant()
2155 : }
2156 : };
2157 :
2158 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2159 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2160 : // during the creation, rather than risking leaving orphan objects in S3.
2161 0 : let persist_tenant_shards = create_ids
2162 0 : .iter()
2163 0 : .map(|tenant_shard_id| TenantShardPersistence {
2164 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2165 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2166 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2167 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2168 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2169 0 : // The pageserver is not known until scheduling happens: we will set this column when
2170 0 : // incrementing the generation the first time we attach to a pageserver.
2171 0 : generation_pageserver: None,
2172 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2173 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2174 0 : splitting: SplitState::default(),
2175 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2176 0 : .unwrap(),
2177 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2178 0 : })
2179 0 : .collect();
2180 0 :
2181 0 : match self
2182 0 : .persistence
2183 0 : .insert_tenant_shards(persist_tenant_shards)
2184 0 : .await
2185 : {
2186 0 : Ok(_) => {}
2187 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2188 : DatabaseErrorKind::UniqueViolation,
2189 : _,
2190 : ))) => {
2191 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2192 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2193 : // creation's shard count.
2194 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
2195 : }
2196 : // Any other database error is unexpected and a bug.
2197 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2198 : };
2199 :
2200 0 : let mut schedule_context = ScheduleContext::default();
2201 0 : let mut schedule_error = None;
2202 0 : let mut response_shards = Vec::new();
2203 0 : for tenant_shard_id in create_ids {
2204 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2205 :
2206 0 : let outcome = self
2207 0 : .do_initial_shard_scheduling(
2208 0 : tenant_shard_id,
2209 0 : initial_generation,
2210 0 : &create_req.shard_parameters,
2211 0 : create_req.config.clone(),
2212 0 : placement_policy.clone(),
2213 0 : preferred_az_id.as_ref(),
2214 0 : &mut schedule_context,
2215 0 : )
2216 0 : .await;
2217 :
2218 0 : match outcome {
2219 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2220 0 : InitialShardScheduleOutcome::NotScheduled => {}
2221 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2222 0 : schedule_error = Some(err);
2223 0 : }
2224 : }
2225 : }
2226 :
2227 : // If we failed to schedule shards, then they are still created in the controller,
2228 : // but we return an error to the requester to avoid a silent failure when someone
2229 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2230 : // are present in the system. We do this here rather than in the above loop, to
2231 : // avoid situations where we only create a subset of shards in the tenant.
2232 0 : if let Some(e) = schedule_error {
2233 0 : return Err(ApiError::Conflict(format!(
2234 0 : "Failed to schedule shard(s): {e}"
2235 0 : )));
2236 0 : }
2237 0 :
2238 0 : let waiters = {
2239 0 : let mut locked = self.inner.write().unwrap();
2240 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2241 0 : tenants
2242 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2243 0 : .filter_map(|(_shard_id, shard)| self.maybe_reconcile_shard(shard, nodes))
2244 0 : .collect::<Vec<_>>()
2245 0 : };
2246 0 :
2247 0 : Ok((
2248 0 : TenantCreateResponse {
2249 0 : shards: response_shards,
2250 0 : },
2251 0 : waiters,
2252 0 : ))
2253 0 : }
2254 :
2255 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2256 : /// case of a new tenant and a pre-existing one.
2257 : #[allow(clippy::too_many_arguments)]
2258 0 : async fn do_initial_shard_scheduling(
2259 0 : &self,
2260 0 : tenant_shard_id: TenantShardId,
2261 0 : initial_generation: Option<Generation>,
2262 0 : shard_params: &ShardParameters,
2263 0 : config: TenantConfig,
2264 0 : placement_policy: PlacementPolicy,
2265 0 : preferred_az_id: Option<&AvailabilityZone>,
2266 0 : schedule_context: &mut ScheduleContext,
2267 0 : ) -> InitialShardScheduleOutcome {
2268 0 : let mut locked = self.inner.write().unwrap();
2269 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2270 :
2271 : use std::collections::btree_map::Entry;
2272 0 : match tenants.entry(tenant_shard_id) {
2273 0 : Entry::Occupied(mut entry) => {
2274 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2275 :
2276 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2277 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2278 0 : }
2279 :
2280 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2281 0 : let generation = entry
2282 0 : .get()
2283 0 : .generation
2284 0 : .expect("Generation is set when in attached mode");
2285 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2286 0 : shard_id: tenant_shard_id,
2287 0 : node_id: *node_id,
2288 0 : generation: generation.into().unwrap(),
2289 0 : })
2290 : } else {
2291 0 : InitialShardScheduleOutcome::NotScheduled
2292 : }
2293 : }
2294 0 : Entry::Vacant(entry) => {
2295 0 : let state = entry.insert(TenantShard::new(
2296 0 : tenant_shard_id,
2297 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2298 0 : placement_policy,
2299 0 : preferred_az_id.cloned(),
2300 0 : ));
2301 0 :
2302 0 : state.generation = initial_generation;
2303 0 : state.config = config;
2304 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2305 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2306 0 : }
2307 :
2308 : // Only include shards in result if we are attaching: the purpose
2309 : // of the response is to tell the caller where the shards are attached.
2310 0 : if let Some(node_id) = state.intent.get_attached() {
2311 0 : let generation = state
2312 0 : .generation
2313 0 : .expect("Generation is set when in attached mode");
2314 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2315 0 : shard_id: tenant_shard_id,
2316 0 : node_id: *node_id,
2317 0 : generation: generation.into().unwrap(),
2318 0 : })
2319 : } else {
2320 0 : InitialShardScheduleOutcome::NotScheduled
2321 : }
2322 : }
2323 : }
2324 0 : }
2325 :
2326 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2327 : /// wait for reconciliation to complete before responding.
2328 0 : async fn await_waiters(
2329 0 : &self,
2330 0 : waiters: Vec<ReconcilerWaiter>,
2331 0 : timeout: Duration,
2332 0 : ) -> Result<(), ReconcileWaitError> {
2333 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2334 0 : for waiter in waiters {
2335 0 : let timeout = deadline.duration_since(Instant::now());
2336 0 : waiter.wait_timeout(timeout).await?;
2337 : }
2338 :
2339 0 : Ok(())
2340 0 : }
2341 :
2342 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2343 : /// in progress
2344 0 : async fn await_waiters_remainder(
2345 0 : &self,
2346 0 : waiters: Vec<ReconcilerWaiter>,
2347 0 : timeout: Duration,
2348 0 : ) -> Vec<ReconcilerWaiter> {
2349 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2350 0 : for waiter in waiters.iter() {
2351 0 : let timeout = deadline.duration_since(Instant::now());
2352 0 : let _ = waiter.wait_timeout(timeout).await;
2353 : }
2354 :
2355 0 : waiters
2356 0 : .into_iter()
2357 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2358 0 : .collect::<Vec<_>>()
2359 0 : }
2360 :
2361 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2362 : /// and transform it into either a tenant creation of a series of shard updates.
2363 : ///
2364 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2365 : /// still be returned.
2366 0 : fn tenant_location_config_prepare(
2367 0 : &self,
2368 0 : tenant_id: TenantId,
2369 0 : req: TenantLocationConfigRequest,
2370 0 : ) -> TenantCreateOrUpdate {
2371 0 : let mut updates = Vec::new();
2372 0 : let mut locked = self.inner.write().unwrap();
2373 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2374 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2375 :
2376 : // Use location config mode as an indicator of policy.
2377 0 : let placement_policy = match req.config.mode {
2378 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2379 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2380 : LocationConfigMode::AttachedMulti
2381 : | LocationConfigMode::AttachedSingle
2382 : | LocationConfigMode::AttachedStale => {
2383 0 : if nodes.len() > 1 {
2384 0 : PlacementPolicy::Attached(1)
2385 : } else {
2386 : // Convenience for dev/test: if we just have one pageserver, import
2387 : // tenants into non-HA mode so that scheduling will succeed.
2388 0 : PlacementPolicy::Attached(0)
2389 : }
2390 : }
2391 : };
2392 :
2393 : // Ordinarily we do not update scheduling policy, but when making major changes
2394 : // like detaching or demoting to secondary-only, we need to force the scheduling
2395 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2396 0 : let scheduling_policy = match req.config.mode {
2397 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2398 : // Special case: when making major changes like detaching or demoting to secondary-only,
2399 : // we need to force the scheduling mode to Active, or nothing will happen.
2400 0 : Some(ShardSchedulingPolicy::Active)
2401 : }
2402 : LocationConfigMode::AttachedMulti
2403 : | LocationConfigMode::AttachedSingle
2404 : | LocationConfigMode::AttachedStale => {
2405 : // While attached, continue to respect whatever the existing scheduling mode is.
2406 0 : None
2407 : }
2408 : };
2409 :
2410 0 : let mut create = true;
2411 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2412 : // Saw an existing shard: this is not a creation
2413 0 : create = false;
2414 :
2415 : // Shards may have initially been created by a Secondary request, where we
2416 : // would have left generation as None.
2417 : //
2418 : // We only update generation the first time we see an attached-mode request,
2419 : // and if there is no existing generation set. The caller is responsible for
2420 : // ensuring that no non-storage-controller pageserver ever uses a higher
2421 : // generation than they passed in here.
2422 : use LocationConfigMode::*;
2423 0 : let set_generation = match req.config.mode {
2424 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2425 0 : req.config.generation.map(Generation::new)
2426 : }
2427 0 : _ => None,
2428 : };
2429 :
2430 0 : updates.push(ShardUpdate {
2431 0 : tenant_shard_id: *shard_id,
2432 0 : placement_policy: placement_policy.clone(),
2433 0 : tenant_config: req.config.tenant_conf.clone(),
2434 0 : generation: set_generation,
2435 0 : scheduling_policy,
2436 0 : });
2437 : }
2438 :
2439 0 : if create {
2440 : use LocationConfigMode::*;
2441 0 : let generation = match req.config.mode {
2442 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2443 : // If a caller provided a generation in a non-attached request, ignore it
2444 : // and leave our generation as None: this enables a subsequent update to set
2445 : // the generation when setting an attached mode for the first time.
2446 0 : _ => None,
2447 : };
2448 :
2449 0 : TenantCreateOrUpdate::Create(
2450 0 : // Synthesize a creation request
2451 0 : TenantCreateRequest {
2452 0 : new_tenant_id: tenant_shard_id,
2453 0 : generation,
2454 0 : shard_parameters: ShardParameters {
2455 0 : count: tenant_shard_id.shard_count,
2456 0 : // We only import un-sharded or single-sharded tenants, so stripe
2457 0 : // size can be made up arbitrarily here.
2458 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2459 0 : },
2460 0 : placement_policy: Some(placement_policy),
2461 0 : config: req.config.tenant_conf,
2462 0 : },
2463 0 : )
2464 : } else {
2465 0 : assert!(!updates.is_empty());
2466 0 : TenantCreateOrUpdate::Update(updates)
2467 : }
2468 0 : }
2469 :
2470 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2471 : /// the tenant is present in memory. If not, load it from the database. If it is found
2472 : /// in neither location, return a NotFound error.
2473 : ///
2474 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2475 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2476 0 : async fn maybe_load_tenant(
2477 0 : &self,
2478 0 : tenant_id: TenantId,
2479 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2480 0 : ) -> Result<(), ApiError> {
2481 : // Check if the tenant is present in memory, and select an AZ to use when loading
2482 : // if we will load it.
2483 0 : let load_in_az = {
2484 0 : let locked = self.inner.read().unwrap();
2485 0 : let existing = locked
2486 0 : .tenants
2487 0 : .range(TenantShardId::tenant_range(tenant_id))
2488 0 : .next();
2489 0 :
2490 0 : // If the tenant is not present in memory, we expect to load it from database,
2491 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2492 0 : if existing.is_none() {
2493 0 : locked
2494 0 : .scheduler
2495 0 : .get_az_for_new_tenant()
2496 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2497 0 : "No AZ with nodes found to load tenant"
2498 0 : )))?
2499 : } else {
2500 : // We already have this tenant in memory
2501 0 : return Ok(());
2502 : }
2503 : };
2504 :
2505 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2506 0 : if tenant_shards.is_empty() {
2507 0 : return Err(ApiError::NotFound(
2508 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2509 0 : ));
2510 0 : }
2511 0 :
2512 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2513 0 : self.persistence
2514 0 : .set_tenant_shard_preferred_azs(
2515 0 : tenant_shards
2516 0 : .iter()
2517 0 : .map(|t| {
2518 0 : (
2519 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2520 0 : Some(load_in_az.clone()),
2521 0 : )
2522 0 : })
2523 0 : .collect(),
2524 0 : )
2525 0 : .await?;
2526 :
2527 0 : let mut locked = self.inner.write().unwrap();
2528 0 : tracing::info!(
2529 0 : "Loaded {} shards for tenant {}",
2530 0 : tenant_shards.len(),
2531 : tenant_id
2532 : );
2533 :
2534 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2535 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2536 0 : let shard =
2537 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2538 0 :
2539 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2540 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2541 0 : if shard.policy != PlacementPolicy::Detached {
2542 0 : tracing::error!(
2543 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2544 : shard.tenant_shard_id,
2545 : shard.policy
2546 : );
2547 0 : }
2548 :
2549 0 : (shard.tenant_shard_id, shard)
2550 0 : }));
2551 0 :
2552 0 : Ok(())
2553 0 : }
2554 :
2555 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2556 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2557 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2558 : ///
2559 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2560 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2561 0 : fn maybe_drop_tenant(
2562 0 : &self,
2563 0 : tenant_id: TenantId,
2564 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2565 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2566 0 : ) {
2567 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2568 0 : if tenant_shards.all(|(_id, shard)| {
2569 0 : shard.policy == PlacementPolicy::Detached
2570 0 : && shard.reconciler.is_none()
2571 0 : && shard.observed.is_empty()
2572 0 : }) {
2573 0 : let keys = locked
2574 0 : .tenants
2575 0 : .range(TenantShardId::tenant_range(tenant_id))
2576 0 : .map(|(id, _)| id)
2577 0 : .copied()
2578 0 : .collect::<Vec<_>>();
2579 0 : for key in keys {
2580 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2581 0 : locked.tenants.remove(&key);
2582 : }
2583 0 : }
2584 0 : }
2585 :
2586 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2587 : /// directly with pageservers into this service.
2588 : ///
2589 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2590 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2591 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2592 : /// tenant's source of generation numbers.
2593 : ///
2594 : /// The mode in this request coarse-grained control of tenants:
2595 : /// - Call with mode Attached* to upsert the tenant.
2596 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2597 : /// to set an existing tenant to PolicyMode::Secondary
2598 : /// - Call with mode Detached to switch to PolicyMode::Detached
2599 0 : pub(crate) async fn tenant_location_config(
2600 0 : &self,
2601 0 : tenant_shard_id: TenantShardId,
2602 0 : req: TenantLocationConfigRequest,
2603 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2604 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2605 0 : let _tenant_lock = trace_exclusive_lock(
2606 0 : &self.tenant_op_locks,
2607 0 : tenant_shard_id.tenant_id,
2608 0 : TenantOperations::LocationConfig,
2609 0 : )
2610 0 : .await;
2611 :
2612 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2613 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2614 0 : "This API is for importing single-sharded or unsharded tenants"
2615 0 : )));
2616 : } else {
2617 0 : tenant_shard_id.tenant_id
2618 0 : };
2619 0 :
2620 0 : // In case we are waking up a Detached tenant
2621 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2622 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2623 0 : // This is a creation or an update
2624 0 : }
2625 0 : Err(e) => {
2626 0 : return Err(e);
2627 : }
2628 : };
2629 :
2630 : // First check if this is a creation or an update
2631 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2632 0 :
2633 0 : let mut result = TenantLocationConfigResponse {
2634 0 : shards: Vec::new(),
2635 0 : stripe_size: None,
2636 0 : };
2637 0 : let waiters = match create_or_update {
2638 0 : TenantCreateOrUpdate::Create(create_req) => {
2639 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2640 0 : result.shards = create_resp
2641 0 : .shards
2642 0 : .into_iter()
2643 0 : .map(|s| TenantShardLocation {
2644 0 : node_id: s.node_id,
2645 0 : shard_id: s.shard_id,
2646 0 : })
2647 0 : .collect();
2648 0 : waiters
2649 : }
2650 0 : TenantCreateOrUpdate::Update(updates) => {
2651 0 : // Persist updates
2652 0 : // Ordering: write to the database before applying changes in-memory, so that
2653 0 : // we will not appear time-travel backwards on a restart.
2654 0 :
2655 0 : let mut schedule_context = ScheduleContext::default();
2656 : for ShardUpdate {
2657 0 : tenant_shard_id,
2658 0 : placement_policy,
2659 0 : tenant_config,
2660 0 : generation,
2661 0 : scheduling_policy,
2662 0 : } in &updates
2663 : {
2664 0 : self.persistence
2665 0 : .update_tenant_shard(
2666 0 : TenantFilter::Shard(*tenant_shard_id),
2667 0 : Some(placement_policy.clone()),
2668 0 : Some(tenant_config.clone()),
2669 0 : *generation,
2670 0 : *scheduling_policy,
2671 0 : )
2672 0 : .await?;
2673 : }
2674 :
2675 : // Apply updates in-memory
2676 0 : let mut waiters = Vec::new();
2677 0 : {
2678 0 : let mut locked = self.inner.write().unwrap();
2679 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2680 :
2681 : for ShardUpdate {
2682 0 : tenant_shard_id,
2683 0 : placement_policy,
2684 0 : tenant_config,
2685 0 : generation: update_generation,
2686 0 : scheduling_policy,
2687 0 : } in updates
2688 : {
2689 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2690 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2691 0 : continue;
2692 : };
2693 :
2694 : // Update stripe size
2695 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2696 0 : result.stripe_size = Some(shard.shard.stripe_size);
2697 0 : }
2698 :
2699 0 : shard.policy = placement_policy;
2700 0 : shard.config = tenant_config;
2701 0 : if let Some(generation) = update_generation {
2702 0 : shard.generation = Some(generation);
2703 0 : }
2704 :
2705 0 : if let Some(scheduling_policy) = scheduling_policy {
2706 0 : shard.set_scheduling_policy(scheduling_policy);
2707 0 : }
2708 :
2709 0 : shard.schedule(scheduler, &mut schedule_context)?;
2710 :
2711 0 : let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
2712 0 : if let Some(waiter) = maybe_waiter {
2713 0 : waiters.push(waiter);
2714 0 : }
2715 :
2716 0 : if let Some(node_id) = shard.intent.get_attached() {
2717 0 : result.shards.push(TenantShardLocation {
2718 0 : shard_id: tenant_shard_id,
2719 0 : node_id: *node_id,
2720 0 : })
2721 0 : }
2722 : }
2723 : }
2724 0 : waiters
2725 : }
2726 : };
2727 :
2728 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2729 : // Do not treat a reconcile error as fatal: we have already applied any requested
2730 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2731 : // compute notification API. In these cases, it is important that we do not
2732 : // cause the cloud control plane to retry forever on this API.
2733 0 : tracing::warn!(
2734 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2735 : );
2736 0 : }
2737 :
2738 : // Logging the full result is useful because it lets us cross-check what the cloud control
2739 : // plane's tenant_shards table should contain.
2740 0 : tracing::info!("Complete, returning {result:?}");
2741 :
2742 0 : Ok(result)
2743 0 : }
2744 :
2745 0 : pub(crate) async fn tenant_config_patch(
2746 0 : &self,
2747 0 : req: TenantConfigPatchRequest,
2748 0 : ) -> Result<(), ApiError> {
2749 0 : let _tenant_lock = trace_exclusive_lock(
2750 0 : &self.tenant_op_locks,
2751 0 : req.tenant_id,
2752 0 : TenantOperations::ConfigPatch,
2753 0 : )
2754 0 : .await;
2755 :
2756 0 : let tenant_id = req.tenant_id;
2757 0 : let patch = req.config;
2758 0 :
2759 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
2760 :
2761 0 : let base = {
2762 0 : let locked = self.inner.read().unwrap();
2763 0 : let shards = locked
2764 0 : .tenants
2765 0 : .range(TenantShardId::tenant_range(req.tenant_id));
2766 0 :
2767 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
2768 :
2769 0 : let first = match configs.peek() {
2770 0 : Some(first) => (*first).clone(),
2771 : None => {
2772 0 : return Err(ApiError::NotFound(
2773 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
2774 0 : ));
2775 : }
2776 : };
2777 :
2778 0 : if !configs.all_equal() {
2779 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
2780 : // This can't happen because we atomically update the database records
2781 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
2782 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2783 0 : "Tenant configs for {} are mismatched",
2784 0 : req.tenant_id
2785 0 : )));
2786 0 : }
2787 0 :
2788 0 : first
2789 0 : };
2790 0 :
2791 0 : let updated_config = base.apply_patch(patch);
2792 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
2793 0 : .await
2794 0 : }
2795 :
2796 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2797 : // We require an exclusive lock, because we are updating persistent and in-memory state
2798 0 : let _tenant_lock = trace_exclusive_lock(
2799 0 : &self.tenant_op_locks,
2800 0 : req.tenant_id,
2801 0 : TenantOperations::ConfigSet,
2802 0 : )
2803 0 : .await;
2804 :
2805 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
2806 :
2807 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
2808 0 : .await
2809 0 : }
2810 :
2811 0 : async fn set_tenant_config_and_reconcile(
2812 0 : &self,
2813 0 : tenant_id: TenantId,
2814 0 : config: TenantConfig,
2815 0 : ) -> Result<(), ApiError> {
2816 0 : self.persistence
2817 0 : .update_tenant_shard(
2818 0 : TenantFilter::Tenant(tenant_id),
2819 0 : None,
2820 0 : Some(config.clone()),
2821 0 : None,
2822 0 : None,
2823 0 : )
2824 0 : .await?;
2825 :
2826 0 : let waiters = {
2827 0 : let mut waiters = Vec::new();
2828 0 : let mut locked = self.inner.write().unwrap();
2829 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2830 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2831 0 : shard.config = config.clone();
2832 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2833 0 : waiters.push(waiter);
2834 0 : }
2835 : }
2836 0 : waiters
2837 : };
2838 :
2839 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2840 : // Treat this as success because we have stored the configuration. If e.g.
2841 : // a node was unavailable at this time, it should not stop us accepting a
2842 : // configuration change.
2843 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2844 0 : }
2845 :
2846 0 : Ok(())
2847 0 : }
2848 :
2849 0 : pub(crate) fn tenant_config_get(
2850 0 : &self,
2851 0 : tenant_id: TenantId,
2852 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2853 0 : let config = {
2854 0 : let locked = self.inner.read().unwrap();
2855 0 :
2856 0 : match locked
2857 0 : .tenants
2858 0 : .range(TenantShardId::tenant_range(tenant_id))
2859 0 : .next()
2860 : {
2861 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2862 : None => {
2863 0 : return Err(ApiError::NotFound(
2864 0 : anyhow::anyhow!("Tenant not found").into(),
2865 0 : ))
2866 : }
2867 : }
2868 : };
2869 :
2870 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2871 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2872 : // and `effective_config` in the response is meaningless, but we retain that syntax
2873 : // in order to remain compatible with the pageserver API.
2874 :
2875 0 : let response = HashMap::from([
2876 : (
2877 : "tenant_specific_overrides",
2878 0 : serde_json::to_value(&config)
2879 0 : .context("serializing tenant specific overrides")
2880 0 : .map_err(ApiError::InternalServerError)?,
2881 : ),
2882 : (
2883 0 : "effective_config",
2884 0 : serde_json::to_value(&config)
2885 0 : .context("serializing effective config")
2886 0 : .map_err(ApiError::InternalServerError)?,
2887 : ),
2888 : ]);
2889 :
2890 0 : Ok(response)
2891 0 : }
2892 :
2893 0 : pub(crate) async fn tenant_time_travel_remote_storage(
2894 0 : &self,
2895 0 : time_travel_req: &TenantTimeTravelRequest,
2896 0 : tenant_id: TenantId,
2897 0 : timestamp: Cow<'_, str>,
2898 0 : done_if_after: Cow<'_, str>,
2899 0 : ) -> Result<(), ApiError> {
2900 0 : let _tenant_lock = trace_exclusive_lock(
2901 0 : &self.tenant_op_locks,
2902 0 : tenant_id,
2903 0 : TenantOperations::TimeTravelRemoteStorage,
2904 0 : )
2905 0 : .await;
2906 :
2907 0 : let node = {
2908 0 : let mut locked = self.inner.write().unwrap();
2909 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
2910 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
2911 : // but only at the start of the process, so it's really just to prevent operator
2912 : // mistakes.
2913 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
2914 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
2915 : {
2916 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2917 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
2918 0 : )));
2919 0 : }
2920 0 : let maybe_attached = shard
2921 0 : .observed
2922 0 : .locations
2923 0 : .iter()
2924 0 : .filter_map(|(node_id, observed_location)| {
2925 0 : observed_location
2926 0 : .conf
2927 0 : .as_ref()
2928 0 : .map(|loc| (node_id, observed_location, loc.mode))
2929 0 : })
2930 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
2931 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
2932 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
2933 0 : }
2934 : }
2935 0 : let scheduler = &mut locked.scheduler;
2936 : // Right now we only perform the operation on a single node without parallelization
2937 : // TODO fan out the operation to multiple nodes for better performance
2938 0 : let node_id = scheduler.any_available_node()?;
2939 0 : let node = locked
2940 0 : .nodes
2941 0 : .get(&node_id)
2942 0 : .expect("Pageservers may not be deleted while lock is active");
2943 0 : node.clone()
2944 0 : };
2945 0 :
2946 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
2947 0 : let mut counts = time_travel_req
2948 0 : .shard_counts
2949 0 : .iter()
2950 0 : .copied()
2951 0 : .collect::<HashSet<_>>()
2952 0 : .into_iter()
2953 0 : .collect::<Vec<_>>();
2954 0 : counts.sort_unstable();
2955 :
2956 0 : for count in counts {
2957 0 : let shard_ids = (0..count.count())
2958 0 : .map(|i| TenantShardId {
2959 0 : tenant_id,
2960 0 : shard_number: ShardNumber(i),
2961 0 : shard_count: count,
2962 0 : })
2963 0 : .collect::<Vec<_>>();
2964 0 : for tenant_shard_id in shard_ids {
2965 0 : let client = PageserverClient::new(
2966 0 : node.get_id(),
2967 0 : node.base_url(),
2968 0 : self.config.jwt_token.as_deref(),
2969 0 : );
2970 0 :
2971 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
2972 :
2973 0 : client
2974 0 : .tenant_time_travel_remote_storage(
2975 0 : tenant_shard_id,
2976 0 : ×tamp,
2977 0 : &done_if_after,
2978 0 : )
2979 0 : .await
2980 0 : .map_err(|e| {
2981 0 : ApiError::InternalServerError(anyhow::anyhow!(
2982 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
2983 0 : node
2984 0 : ))
2985 0 : })?;
2986 : }
2987 : }
2988 0 : Ok(())
2989 0 : }
2990 :
2991 0 : pub(crate) async fn tenant_secondary_download(
2992 0 : &self,
2993 0 : tenant_id: TenantId,
2994 0 : wait: Option<Duration>,
2995 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
2996 0 : let _tenant_lock = trace_shared_lock(
2997 0 : &self.tenant_op_locks,
2998 0 : tenant_id,
2999 0 : TenantOperations::SecondaryDownload,
3000 0 : )
3001 0 : .await;
3002 :
3003 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3004 0 : let targets = {
3005 0 : let locked = self.inner.read().unwrap();
3006 0 : let mut targets = Vec::new();
3007 :
3008 0 : for (tenant_shard_id, shard) in
3009 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3010 : {
3011 0 : for node_id in shard.intent.get_secondary() {
3012 0 : let node = locked
3013 0 : .nodes
3014 0 : .get(node_id)
3015 0 : .expect("Pageservers may not be deleted while referenced");
3016 0 :
3017 0 : targets.push((*tenant_shard_id, node.clone()));
3018 0 : }
3019 : }
3020 0 : targets
3021 0 : };
3022 0 :
3023 0 : // Issue concurrent requests to all shards' locations
3024 0 : let mut futs = FuturesUnordered::new();
3025 0 : for (tenant_shard_id, node) in targets {
3026 0 : let client = PageserverClient::new(
3027 0 : node.get_id(),
3028 0 : node.base_url(),
3029 0 : self.config.jwt_token.as_deref(),
3030 0 : );
3031 0 : futs.push(async move {
3032 0 : let result = client
3033 0 : .tenant_secondary_download(tenant_shard_id, wait)
3034 0 : .await;
3035 0 : (result, node, tenant_shard_id)
3036 0 : })
3037 : }
3038 :
3039 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3040 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3041 : // well as more general cases like 503s, 500s, or timeouts.
3042 0 : let mut aggregate_progress = SecondaryProgress::default();
3043 0 : let mut aggregate_status: Option<StatusCode> = None;
3044 0 : let mut error: Option<mgmt_api::Error> = None;
3045 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3046 0 : match result {
3047 0 : Err(e) => {
3048 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3049 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3050 0 : // than they had hoped for.
3051 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3052 0 : error = Some(e)
3053 : }
3054 0 : Ok((status_code, progress)) => {
3055 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3056 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3057 0 : aggregate_progress.layers_total += progress.layers_total;
3058 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3059 0 : aggregate_progress.bytes_total += progress.bytes_total;
3060 0 : aggregate_progress.heatmap_mtime =
3061 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3062 0 : aggregate_status = match aggregate_status {
3063 0 : None => Some(status_code),
3064 0 : Some(StatusCode::OK) => Some(status_code),
3065 0 : Some(cur) => {
3066 0 : // Other status codes (e.g. 202) -- do not overwrite.
3067 0 : Some(cur)
3068 : }
3069 : };
3070 : }
3071 : }
3072 : }
3073 :
3074 : // If any of the shards return 202, indicate our result as 202.
3075 0 : match aggregate_status {
3076 : None => {
3077 0 : match error {
3078 0 : Some(e) => {
3079 0 : // No successes, and an error: surface it
3080 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3081 : }
3082 : None => {
3083 : // No shards found
3084 0 : Err(ApiError::NotFound(
3085 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3086 0 : ))
3087 : }
3088 : }
3089 : }
3090 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3091 : }
3092 0 : }
3093 :
3094 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
3095 0 : let _tenant_lock =
3096 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3097 :
3098 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3099 :
3100 : // Detach all shards. This also deletes local pageserver shard data.
3101 0 : let (detach_waiters, node) = {
3102 0 : let mut detach_waiters = Vec::new();
3103 0 : let mut locked = self.inner.write().unwrap();
3104 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3105 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3106 : // Update the tenant's intent to remove all attachments
3107 0 : shard.policy = PlacementPolicy::Detached;
3108 0 : shard
3109 0 : .schedule(scheduler, &mut ScheduleContext::default())
3110 0 : .expect("De-scheduling is infallible");
3111 0 : debug_assert!(shard.intent.get_attached().is_none());
3112 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3113 :
3114 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
3115 0 : detach_waiters.push(waiter);
3116 0 : }
3117 : }
3118 :
3119 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3120 : // was attached, just has to be able to see the S3 content)
3121 0 : let node_id = scheduler.any_available_node()?;
3122 0 : let node = nodes
3123 0 : .get(&node_id)
3124 0 : .expect("Pageservers may not be deleted while lock is active");
3125 0 : (detach_waiters, node.clone())
3126 0 : };
3127 0 :
3128 0 : // This reconcile wait can fail in a few ways:
3129 0 : // A there is a very long queue for the reconciler semaphore
3130 0 : // B some pageserver is failing to handle a detach promptly
3131 0 : // C some pageserver goes offline right at the moment we send it a request.
3132 0 : //
3133 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3134 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3135 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3136 0 : // deleting the underlying data).
3137 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3138 0 : .await?;
3139 :
3140 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3141 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3142 : // the tenant ID prefix, including all shards (even possibly stale ones).
3143 0 : match node
3144 0 : .with_client_retries(
3145 0 : |client| async move {
3146 0 : client
3147 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3148 0 : .await
3149 0 : },
3150 0 : &self.config.jwt_token,
3151 0 : 1,
3152 0 : 3,
3153 0 : RECONCILE_TIMEOUT,
3154 0 : &self.cancel,
3155 0 : )
3156 0 : .await
3157 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3158 : {
3159 0 : Ok(_) => {}
3160 : Err(mgmt_api::Error::Cancelled) => {
3161 0 : return Err(ApiError::ShuttingDown);
3162 : }
3163 0 : Err(e) => {
3164 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3165 0 : // at large is unavailable.
3166 0 : tracing::error!("Error deleting via node {node}: {e}");
3167 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3168 : }
3169 : }
3170 :
3171 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3172 : // our in-memory state and database state.
3173 :
3174 : // Ordering: we delete persistent state first: if we then
3175 : // crash, we will drop the in-memory state.
3176 :
3177 : // Drop persistent state.
3178 0 : self.persistence.delete_tenant(tenant_id).await?;
3179 :
3180 : // Drop in-memory state
3181 : {
3182 0 : let mut locked = self.inner.write().unwrap();
3183 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3184 :
3185 : // Dereference Scheduler from shards before dropping them
3186 0 : for (_tenant_shard_id, shard) in
3187 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3188 0 : {
3189 0 : shard.intent.clear(scheduler);
3190 0 : }
3191 :
3192 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3193 0 : tracing::info!(
3194 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3195 0 : locked.tenants.len()
3196 : );
3197 : };
3198 :
3199 : // Success is represented as 404, to imitate the existing pageserver deletion API
3200 0 : Ok(StatusCode::NOT_FOUND)
3201 0 : }
3202 :
3203 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3204 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3205 : /// the tenant's policies (configuration) within the storage controller
3206 0 : pub(crate) async fn tenant_update_policy(
3207 0 : &self,
3208 0 : tenant_id: TenantId,
3209 0 : req: TenantPolicyRequest,
3210 0 : ) -> Result<(), ApiError> {
3211 : // We require an exclusive lock, because we are updating persistent and in-memory state
3212 0 : let _tenant_lock = trace_exclusive_lock(
3213 0 : &self.tenant_op_locks,
3214 0 : tenant_id,
3215 0 : TenantOperations::UpdatePolicy,
3216 0 : )
3217 0 : .await;
3218 :
3219 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3220 :
3221 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3222 :
3223 : let TenantPolicyRequest {
3224 0 : placement,
3225 0 : mut scheduling,
3226 0 : } = req;
3227 :
3228 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3229 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3230 : // this to take effect.
3231 0 : if scheduling.is_none() {
3232 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3233 0 : }
3234 0 : }
3235 :
3236 0 : self.persistence
3237 0 : .update_tenant_shard(
3238 0 : TenantFilter::Tenant(tenant_id),
3239 0 : placement.clone(),
3240 0 : None,
3241 0 : None,
3242 0 : scheduling,
3243 0 : )
3244 0 : .await?;
3245 :
3246 0 : let mut schedule_context = ScheduleContext::default();
3247 0 : let mut locked = self.inner.write().unwrap();
3248 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3249 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3250 0 : if let Some(placement) = &placement {
3251 0 : shard.policy = placement.clone();
3252 0 :
3253 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3254 0 : "Updated placement policy to {placement:?}");
3255 0 : }
3256 :
3257 0 : if let Some(scheduling) = &scheduling {
3258 0 : shard.set_scheduling_policy(*scheduling);
3259 0 :
3260 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3261 0 : "Updated scheduling policy to {scheduling:?}");
3262 0 : }
3263 :
3264 : // In case scheduling is being switched back on, try it now.
3265 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3266 0 : self.maybe_reconcile_shard(shard, nodes);
3267 : }
3268 :
3269 0 : Ok(())
3270 0 : }
3271 :
3272 0 : pub(crate) async fn tenant_timeline_create(
3273 0 : &self,
3274 0 : tenant_id: TenantId,
3275 0 : mut create_req: TimelineCreateRequest,
3276 0 : ) -> Result<TimelineInfo, ApiError> {
3277 0 : tracing::info!(
3278 0 : "Creating timeline {}/{}",
3279 : tenant_id,
3280 : create_req.new_timeline_id,
3281 : );
3282 :
3283 0 : let _tenant_lock = trace_shared_lock(
3284 0 : &self.tenant_op_locks,
3285 0 : tenant_id,
3286 0 : TenantOperations::TimelineCreate,
3287 0 : )
3288 0 : .await;
3289 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3290 :
3291 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3292 0 : if targets.0.is_empty() {
3293 0 : return Err(ApiError::NotFound(
3294 0 : anyhow::anyhow!("Tenant not found").into(),
3295 0 : ));
3296 0 : };
3297 0 :
3298 0 : let (shard_zero_tid, shard_zero_locations) =
3299 0 : targets.0.pop_first().expect("Must have at least one shard");
3300 0 : assert!(shard_zero_tid.is_shard_zero());
3301 :
3302 0 : async fn create_one(
3303 0 : tenant_shard_id: TenantShardId,
3304 0 : locations: ShardMutationLocations,
3305 0 : jwt: Option<String>,
3306 0 : create_req: TimelineCreateRequest,
3307 0 : ) -> Result<TimelineInfo, ApiError> {
3308 0 : let latest = locations.latest.node;
3309 0 :
3310 0 : tracing::info!(
3311 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3312 : tenant_shard_id,
3313 : create_req.new_timeline_id,
3314 : locations.latest.generation
3315 : );
3316 :
3317 0 : let client =
3318 0 : PageserverClient::new(latest.get_id(), latest.base_url(), jwt.as_deref());
3319 :
3320 0 : let timeline_info = client
3321 0 : .timeline_create(tenant_shard_id, &create_req)
3322 0 : .await
3323 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3324 :
3325 : // We propagate timeline creations to all attached locations such that a compute
3326 : // for the new timeline is able to start regardless of the current state of the
3327 : // tenant shard reconciliation.
3328 0 : for location in locations.other {
3329 0 : tracing::info!(
3330 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3331 : tenant_shard_id,
3332 : create_req.new_timeline_id,
3333 : location.node,
3334 : location.generation
3335 : );
3336 :
3337 0 : let client = PageserverClient::new(
3338 0 : location.node.get_id(),
3339 0 : location.node.base_url(),
3340 0 : jwt.as_deref(),
3341 0 : );
3342 :
3343 0 : let res = client
3344 0 : .timeline_create(tenant_shard_id, &create_req)
3345 0 : .await;
3346 :
3347 0 : if let Err(e) = res {
3348 0 : match e {
3349 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3350 0 : // Tenant might have been detached from the stale location,
3351 0 : // so ignore 404s.
3352 0 : },
3353 : _ => {
3354 0 : return Err(passthrough_api_error(&location.node, e));
3355 : }
3356 : }
3357 0 : }
3358 : }
3359 :
3360 0 : Ok(timeline_info)
3361 0 : }
3362 :
3363 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3364 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3365 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3366 0 : let timeline_info = create_one(
3367 0 : shard_zero_tid,
3368 0 : shard_zero_locations,
3369 0 : self.config.jwt_token.clone(),
3370 0 : create_req.clone(),
3371 0 : )
3372 0 : .await?;
3373 :
3374 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3375 0 : match &mut create_req.mode {
3376 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3377 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3378 0 : },
3379 0 : _ => {}
3380 : }
3381 :
3382 : // Create timeline on remaining shards with number >0
3383 0 : if !targets.0.is_empty() {
3384 : // If we had multiple shards, issue requests for the remainder now.
3385 0 : let jwt = &self.config.jwt_token;
3386 0 : self.tenant_for_shards(
3387 0 : targets
3388 0 : .0
3389 0 : .iter()
3390 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3391 0 : .collect(),
3392 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3393 0 : let create_req = create_req.clone();
3394 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3395 0 : Box::pin(create_one(
3396 0 : tenant_shard_id,
3397 0 : mutation_locations,
3398 0 : jwt.clone(),
3399 0 : create_req,
3400 0 : ))
3401 0 : },
3402 0 : )
3403 0 : .await?;
3404 0 : }
3405 :
3406 0 : Ok(timeline_info)
3407 0 : })
3408 0 : .await?
3409 0 : }
3410 :
3411 0 : pub(crate) async fn tenant_timeline_archival_config(
3412 0 : &self,
3413 0 : tenant_id: TenantId,
3414 0 : timeline_id: TimelineId,
3415 0 : req: TimelineArchivalConfigRequest,
3416 0 : ) -> Result<(), ApiError> {
3417 0 : tracing::info!(
3418 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3419 : req.state
3420 : );
3421 :
3422 0 : let _tenant_lock = trace_shared_lock(
3423 0 : &self.tenant_op_locks,
3424 0 : tenant_id,
3425 0 : TenantOperations::TimelineArchivalConfig,
3426 0 : )
3427 0 : .await;
3428 :
3429 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3430 0 : if targets.0.is_empty() {
3431 0 : return Err(ApiError::NotFound(
3432 0 : anyhow::anyhow!("Tenant not found").into(),
3433 0 : ));
3434 0 : }
3435 0 : async fn config_one(
3436 0 : tenant_shard_id: TenantShardId,
3437 0 : timeline_id: TimelineId,
3438 0 : node: Node,
3439 0 : jwt: Option<String>,
3440 0 : req: TimelineArchivalConfigRequest,
3441 0 : ) -> Result<(), ApiError> {
3442 0 : tracing::info!(
3443 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3444 : );
3445 :
3446 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3447 0 :
3448 0 : client
3449 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3450 0 : .await
3451 0 : .map_err(|e| match e {
3452 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3453 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3454 : }
3455 0 : _ => passthrough_api_error(&node, e),
3456 0 : })
3457 0 : }
3458 :
3459 : // no shard needs to go first/last; the operation should be idempotent
3460 : // TODO: it would be great to ensure that all shards return the same error
3461 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3462 0 : let results = self
3463 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3464 0 : futures::FutureExt::boxed(config_one(
3465 0 : tenant_shard_id,
3466 0 : timeline_id,
3467 0 : node,
3468 0 : self.config.jwt_token.clone(),
3469 0 : req.clone(),
3470 0 : ))
3471 0 : })
3472 0 : .await?;
3473 0 : assert!(!results.is_empty(), "must have at least one result");
3474 :
3475 0 : Ok(())
3476 0 : }).await?
3477 0 : }
3478 :
3479 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3480 0 : &self,
3481 0 : tenant_id: TenantId,
3482 0 : timeline_id: TimelineId,
3483 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3484 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3485 :
3486 0 : let _tenant_lock = trace_shared_lock(
3487 0 : &self.tenant_op_locks,
3488 0 : tenant_id,
3489 0 : TenantOperations::TimelineDetachAncestor,
3490 0 : )
3491 0 : .await;
3492 :
3493 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3494 0 : if targets.0.is_empty() {
3495 0 : return Err(ApiError::NotFound(
3496 0 : anyhow::anyhow!("Tenant not found").into(),
3497 0 : ));
3498 0 : }
3499 :
3500 0 : async fn detach_one(
3501 0 : tenant_shard_id: TenantShardId,
3502 0 : timeline_id: TimelineId,
3503 0 : node: Node,
3504 0 : jwt: Option<String>,
3505 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3506 0 : tracing::info!(
3507 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3508 : );
3509 :
3510 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3511 0 :
3512 0 : client
3513 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3514 0 : .await
3515 0 : .map_err(|e| {
3516 : use mgmt_api::Error;
3517 :
3518 0 : match e {
3519 : // no ancestor (ever)
3520 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3521 0 : "{node}: {}",
3522 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3523 0 : )),
3524 : // too many ancestors
3525 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3526 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3527 : }
3528 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3529 0 : // avoid turning these into conflicts to remain compatible with
3530 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3531 0 : // detach
3532 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3533 : }
3534 : // rest can be mapped as usual
3535 0 : other => passthrough_api_error(&node, other),
3536 : }
3537 0 : })
3538 0 : .map(|res| (tenant_shard_id.shard_number, res))
3539 0 : }
3540 :
3541 : // no shard needs to go first/last; the operation should be idempotent
3542 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3543 0 : let mut results = self
3544 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3545 0 : futures::FutureExt::boxed(detach_one(
3546 0 : tenant_shard_id,
3547 0 : timeline_id,
3548 0 : node,
3549 0 : self.config.jwt_token.clone(),
3550 0 : ))
3551 0 : })
3552 0 : .await?;
3553 :
3554 0 : let any = results.pop().expect("we must have at least one response");
3555 0 :
3556 0 : let mismatching = results
3557 0 : .iter()
3558 0 : .filter(|(_, res)| res != &any.1)
3559 0 : .collect::<Vec<_>>();
3560 0 : if !mismatching.is_empty() {
3561 : // this can be hit by races which should not happen because operation lock on cplane
3562 0 : let matching = results.len() - mismatching.len();
3563 0 : tracing::error!(
3564 : matching,
3565 : compared_against=?any,
3566 : ?mismatching,
3567 0 : "shards returned different results"
3568 : );
3569 :
3570 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3571 0 : }
3572 0 :
3573 0 : Ok(any.1)
3574 0 : }).await?
3575 0 : }
3576 :
3577 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3578 0 : &self,
3579 0 : tenant_id: TenantId,
3580 0 : timeline_id: TimelineId,
3581 0 : dir: BlockUnblock,
3582 0 : ) -> Result<(), ApiError> {
3583 0 : let _tenant_lock = trace_shared_lock(
3584 0 : &self.tenant_op_locks,
3585 0 : tenant_id,
3586 0 : TenantOperations::TimelineGcBlockUnblock,
3587 0 : )
3588 0 : .await;
3589 :
3590 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3591 0 : if targets.0.is_empty() {
3592 0 : return Err(ApiError::NotFound(
3593 0 : anyhow::anyhow!("Tenant not found").into(),
3594 0 : ));
3595 0 : }
3596 :
3597 0 : async fn do_one(
3598 0 : tenant_shard_id: TenantShardId,
3599 0 : timeline_id: TimelineId,
3600 0 : node: Node,
3601 0 : jwt: Option<String>,
3602 0 : dir: BlockUnblock,
3603 0 : ) -> Result<(), ApiError> {
3604 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3605 0 :
3606 0 : client
3607 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3608 0 : .await
3609 0 : .map_err(|e| passthrough_api_error(&node, e))
3610 0 : }
3611 :
3612 : // no shard needs to go first/last; the operation should be idempotent
3613 0 : let locations = targets
3614 0 : .0
3615 0 : .iter()
3616 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3617 0 : .collect();
3618 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
3619 0 : futures::FutureExt::boxed(do_one(
3620 0 : tenant_shard_id,
3621 0 : timeline_id,
3622 0 : node,
3623 0 : self.config.jwt_token.clone(),
3624 0 : dir,
3625 0 : ))
3626 0 : })
3627 0 : .await
3628 0 : })
3629 0 : .await??;
3630 0 : Ok(())
3631 0 : }
3632 :
3633 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3634 : ///
3635 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3636 0 : async fn tenant_for_shards<F, R>(
3637 0 : &self,
3638 0 : locations: Vec<(TenantShardId, Node)>,
3639 0 : mut req_fn: F,
3640 0 : ) -> Result<Vec<R>, ApiError>
3641 0 : where
3642 0 : F: FnMut(
3643 0 : TenantShardId,
3644 0 : Node,
3645 0 : )
3646 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3647 0 : {
3648 0 : let mut futs = FuturesUnordered::new();
3649 0 : let mut results = Vec::with_capacity(locations.len());
3650 :
3651 0 : for (tenant_shard_id, node) in locations {
3652 0 : futs.push(req_fn(tenant_shard_id, node));
3653 0 : }
3654 :
3655 0 : while let Some(r) = futs.next().await {
3656 0 : results.push(r?);
3657 : }
3658 :
3659 0 : Ok(results)
3660 0 : }
3661 :
3662 : /// Concurrently invoke a pageserver API call on many shards at once
3663 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3664 0 : &self,
3665 0 : locations: Vec<(TenantShardId, Node)>,
3666 0 : op: O,
3667 0 : warn_threshold: u32,
3668 0 : max_retries: u32,
3669 0 : timeout: Duration,
3670 0 : cancel: &CancellationToken,
3671 0 : ) -> Vec<mgmt_api::Result<T>>
3672 0 : where
3673 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3674 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3675 0 : {
3676 0 : let mut futs = FuturesUnordered::new();
3677 0 : let mut results = Vec::with_capacity(locations.len());
3678 :
3679 0 : for (tenant_shard_id, node) in locations {
3680 0 : futs.push(async move {
3681 0 : node.with_client_retries(
3682 0 : |client| op(tenant_shard_id, client),
3683 0 : &self.config.jwt_token,
3684 0 : warn_threshold,
3685 0 : max_retries,
3686 0 : timeout,
3687 0 : cancel,
3688 0 : )
3689 0 : .await
3690 0 : });
3691 0 : }
3692 :
3693 0 : while let Some(r) = futs.next().await {
3694 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3695 0 : results.push(r);
3696 0 : }
3697 :
3698 0 : results
3699 0 : }
3700 :
3701 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3702 : /// when creating and deleting timelines:
3703 : /// - Makes sure shards are attached somewhere if they weren't already
3704 : /// - Looks up the shards and the nodes where they were most recently attached
3705 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3706 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3707 0 : async fn tenant_remote_mutation<R, O, F>(
3708 0 : &self,
3709 0 : tenant_id: TenantId,
3710 0 : op: O,
3711 0 : ) -> Result<R, ApiError>
3712 0 : where
3713 0 : O: FnOnce(TenantMutationLocations) -> F,
3714 0 : F: std::future::Future<Output = R>,
3715 0 : {
3716 0 : let mutation_locations = {
3717 0 : let mut locations = TenantMutationLocations::default();
3718 :
3719 : // Load the currently attached pageservers for the latest generation of each shard. This can
3720 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3721 : // will still be the latest when we're done: we will check generations again at the end of
3722 : // this function to handle that.
3723 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
3724 :
3725 0 : if generations
3726 0 : .iter()
3727 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3728 : {
3729 0 : let shard_generations = generations
3730 0 : .into_iter()
3731 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
3732 0 : .collect::<HashMap<_, _>>();
3733 0 :
3734 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3735 0 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
3736 0 : let locked = self.inner.read().unwrap();
3737 0 : for (shard_id, shard) in
3738 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3739 : {
3740 0 : match shard.policy {
3741 : PlacementPolicy::Attached(_) => {
3742 : // This shard is meant to be attached: the caller is not wrong to try and
3743 : // use this function, but we can't service the request right now.
3744 0 : let Some(generation) = shard_generations.get(shard_id) else {
3745 : // This can only happen if there is a split brain controller modifying the database. This should
3746 : // never happen when testing, and if it happens in production we can only log the issue.
3747 0 : debug_assert!(false);
3748 0 : tracing::error!("Shard {shard_id} not found in generation state! Is another rogue controller running?");
3749 0 : continue;
3750 : };
3751 0 : let (generation, generation_pageserver) = generation;
3752 0 : if let Some(generation) = generation {
3753 0 : if generation_pageserver.is_none() {
3754 : // This is legitimate only in a very narrow window where the shard was only just configured into
3755 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
3756 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
3757 0 : tracing::warn!("Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?");
3758 0 : }
3759 : } else {
3760 : // This should never happen: a shard with no generation is only permitted when it was created in some state
3761 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
3762 0 : debug_assert!(false);
3763 0 : tracing::error!("Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!");
3764 0 : continue;
3765 : }
3766 : }
3767 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3768 0 : return Err(ApiError::Conflict(format!(
3769 0 : "Shard {shard_id} tenant has policy {:?}",
3770 0 : shard.policy
3771 0 : )));
3772 : }
3773 : }
3774 : }
3775 :
3776 0 : return Err(ApiError::ResourceUnavailable(
3777 0 : "One or more shards in tenant is not yet attached".into(),
3778 0 : ));
3779 0 : }
3780 0 :
3781 0 : let locked = self.inner.read().unwrap();
3782 : for ShardGenerationState {
3783 0 : tenant_shard_id,
3784 0 : generation,
3785 0 : generation_pageserver,
3786 0 : } in generations
3787 : {
3788 0 : let node_id = generation_pageserver.expect("We checked for None above");
3789 0 : let node = locked
3790 0 : .nodes
3791 0 : .get(&node_id)
3792 0 : .ok_or(ApiError::Conflict(format!(
3793 0 : "Raced with removal of node {node_id}"
3794 0 : )))?;
3795 0 : let generation = generation.expect("Checked above");
3796 0 :
3797 0 : let tenant = locked.tenants.get(&tenant_shard_id);
3798 :
3799 : // TODO(vlad): Abstract the logic that finds stale attached locations
3800 : // from observed state into a [`Service`] method.
3801 0 : let other_locations = match tenant {
3802 0 : Some(tenant) => {
3803 0 : let mut other = tenant.attached_locations();
3804 0 : let latest_location_index =
3805 0 : other.iter().position(|&l| l == (node.get_id(), generation));
3806 0 : if let Some(idx) = latest_location_index {
3807 0 : other.remove(idx);
3808 0 : }
3809 :
3810 0 : other
3811 : }
3812 0 : None => Vec::default(),
3813 : };
3814 :
3815 0 : let location = ShardMutationLocations {
3816 0 : latest: MutationLocation {
3817 0 : node: node.clone(),
3818 0 : generation,
3819 0 : },
3820 0 : other: other_locations
3821 0 : .into_iter()
3822 0 : .filter_map(|(node_id, generation)| {
3823 0 : let node = locked.nodes.get(&node_id)?;
3824 :
3825 0 : Some(MutationLocation {
3826 0 : node: node.clone(),
3827 0 : generation,
3828 0 : })
3829 0 : })
3830 0 : .collect(),
3831 0 : };
3832 0 : locations.0.insert(tenant_shard_id, location);
3833 0 : }
3834 :
3835 0 : locations
3836 : };
3837 :
3838 0 : let result = op(mutation_locations.clone()).await;
3839 :
3840 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
3841 : // our remote operation executed on the latest generation and is therefore persistent.
3842 : {
3843 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
3844 0 : if latest_generations
3845 0 : .into_iter()
3846 0 : .map(
3847 0 : |ShardGenerationState {
3848 : tenant_shard_id,
3849 : generation,
3850 : generation_pageserver: _,
3851 0 : }| (tenant_shard_id, generation),
3852 0 : )
3853 0 : .collect::<Vec<_>>()
3854 0 : != mutation_locations
3855 0 : .0
3856 0 : .into_iter()
3857 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
3858 0 : .collect::<Vec<_>>()
3859 : {
3860 : // We raced with something that incremented the generation, and therefore cannot be
3861 : // confident that our actions are persistent (they might have hit an old generation).
3862 : //
3863 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
3864 0 : return Err(ApiError::ResourceUnavailable(
3865 0 : "Tenant attachment changed, please retry".into(),
3866 0 : ));
3867 0 : }
3868 0 : }
3869 0 :
3870 0 : Ok(result)
3871 0 : }
3872 :
3873 0 : pub(crate) async fn tenant_timeline_delete(
3874 0 : &self,
3875 0 : tenant_id: TenantId,
3876 0 : timeline_id: TimelineId,
3877 0 : ) -> Result<StatusCode, ApiError> {
3878 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
3879 0 : let _tenant_lock = trace_shared_lock(
3880 0 : &self.tenant_op_locks,
3881 0 : tenant_id,
3882 0 : TenantOperations::TimelineDelete,
3883 0 : )
3884 0 : .await;
3885 :
3886 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3887 0 : if targets.0.is_empty() {
3888 0 : return Err(ApiError::NotFound(
3889 0 : anyhow::anyhow!("Tenant not found").into(),
3890 0 : ));
3891 0 : }
3892 0 :
3893 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
3894 0 : assert!(shard_zero_tid.is_shard_zero());
3895 :
3896 0 : async fn delete_one(
3897 0 : tenant_shard_id: TenantShardId,
3898 0 : timeline_id: TimelineId,
3899 0 : node: Node,
3900 0 : jwt: Option<String>,
3901 0 : ) -> Result<StatusCode, ApiError> {
3902 0 : tracing::info!(
3903 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3904 : );
3905 :
3906 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3907 0 : let res = client
3908 0 : .timeline_delete(tenant_shard_id, timeline_id)
3909 0 : .await;
3910 :
3911 0 : match res {
3912 0 : Ok(ok) => Ok(ok),
3913 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
3914 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
3915 0 : Err(e) => {
3916 0 : Err(
3917 0 : ApiError::InternalServerError(anyhow::anyhow!(
3918 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
3919 0 : ))
3920 0 : )
3921 : }
3922 : }
3923 0 : }
3924 :
3925 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3926 0 : let statuses = self
3927 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
3928 0 : Box::pin(delete_one(
3929 0 : tenant_shard_id,
3930 0 : timeline_id,
3931 0 : node,
3932 0 : self.config.jwt_token.clone(),
3933 0 : ))
3934 0 : })
3935 0 : .await?;
3936 :
3937 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
3938 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
3939 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
3940 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
3941 0 : return Ok(StatusCode::CONFLICT);
3942 0 : }
3943 0 :
3944 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
3945 0 : return Ok(StatusCode::ACCEPTED);
3946 0 : }
3947 :
3948 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
3949 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
3950 0 : let shard_zero_status = delete_one(
3951 0 : shard_zero_tid,
3952 0 : timeline_id,
3953 0 : shard_zero_locations.latest.node,
3954 0 : self.config.jwt_token.clone(),
3955 0 : )
3956 0 : .await?;
3957 0 : Ok(shard_zero_status)
3958 0 : }).await?
3959 0 : }
3960 :
3961 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
3962 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
3963 0 : pub(crate) async fn tenant_shard0_node(
3964 0 : &self,
3965 0 : tenant_id: TenantId,
3966 0 : ) -> Result<(Node, TenantShardId), ApiError> {
3967 0 : // Look up in-memory state and maybe use the node from there.
3968 0 : {
3969 0 : let locked = self.inner.read().unwrap();
3970 0 : let Some((tenant_shard_id, shard)) = locked
3971 0 : .tenants
3972 0 : .range(TenantShardId::tenant_range(tenant_id))
3973 0 : .next()
3974 : else {
3975 0 : return Err(ApiError::NotFound(
3976 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
3977 0 : ));
3978 : };
3979 :
3980 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
3981 0 : tracing::warn!(
3982 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
3983 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
3984 : shard.policy
3985 : );
3986 0 : return Err(ApiError::Conflict(
3987 0 : "Cannot call timeline API on non-attached tenant".to_string(),
3988 0 : ));
3989 : };
3990 :
3991 0 : if shard.reconciler.is_none() {
3992 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
3993 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
3994 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
3995 : // This should never happen
3996 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3997 0 : "Shard refers to nonexistent node"
3998 0 : )));
3999 : };
4000 0 : return Ok((node.clone(), *tenant_shard_id));
4001 0 : }
4002 : };
4003 :
4004 : // Look up the latest attached pageserver location from the database
4005 : // generation state: this will reflect the progress of any ongoing migration.
4006 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4007 : // the case where they call through to the pageserver and get a 404.
4008 0 : let db_result = self.persistence.tenant_generations(tenant_id).await?;
4009 : let Some(ShardGenerationState {
4010 0 : tenant_shard_id,
4011 0 : generation: _,
4012 0 : generation_pageserver: Some(node_id),
4013 0 : }) = db_result.first()
4014 : else {
4015 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4016 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4017 : // or a conflict response (case where tenant was detached in background)
4018 0 : return Err(ApiError::ResourceUnavailable(
4019 0 : "Shard {} not found in database, or is not attached".into(),
4020 0 : ));
4021 : };
4022 0 : let locked = self.inner.read().unwrap();
4023 0 : let Some(node) = locked.nodes.get(node_id) else {
4024 : // This should never happen
4025 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4026 0 : "Shard refers to nonexistent node"
4027 0 : )));
4028 : };
4029 :
4030 0 : Ok((node.clone(), *tenant_shard_id))
4031 0 : }
4032 :
4033 0 : pub(crate) fn tenant_locate(
4034 0 : &self,
4035 0 : tenant_id: TenantId,
4036 0 : ) -> Result<TenantLocateResponse, ApiError> {
4037 0 : let locked = self.inner.read().unwrap();
4038 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4039 :
4040 0 : let mut result = Vec::new();
4041 0 : let mut shard_params: Option<ShardParameters> = None;
4042 :
4043 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4044 : {
4045 0 : let node_id =
4046 0 : shard
4047 0 : .intent
4048 0 : .get_attached()
4049 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4050 0 : "Cannot locate a tenant that is not attached"
4051 0 : )))?;
4052 :
4053 0 : let node = locked
4054 0 : .nodes
4055 0 : .get(&node_id)
4056 0 : .expect("Pageservers may not be deleted while referenced");
4057 0 :
4058 0 : result.push(node.shard_location(*tenant_shard_id));
4059 0 :
4060 0 : match &shard_params {
4061 0 : None => {
4062 0 : shard_params = Some(ShardParameters {
4063 0 : stripe_size: shard.shard.stripe_size,
4064 0 : count: shard.shard.count,
4065 0 : });
4066 0 : }
4067 0 : Some(params) => {
4068 0 : if params.stripe_size != shard.shard.stripe_size {
4069 : // This should never happen. We enforce at runtime because it's simpler than
4070 : // adding an extra per-tenant data structure to store the things that should be the same
4071 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4072 0 : "Inconsistent shard stripe size parameters!"
4073 0 : )));
4074 0 : }
4075 : }
4076 : }
4077 : }
4078 :
4079 0 : if result.is_empty() {
4080 0 : return Err(ApiError::NotFound(
4081 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4082 0 : ));
4083 0 : }
4084 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4085 0 : tracing::info!(
4086 0 : "Located tenant {} with params {:?} on shards {}",
4087 0 : tenant_id,
4088 0 : shard_params,
4089 0 : result
4090 0 : .iter()
4091 0 : .map(|s| format!("{:?}", s))
4092 0 : .collect::<Vec<_>>()
4093 0 : .join(",")
4094 : );
4095 :
4096 0 : Ok(TenantLocateResponse {
4097 0 : shards: result,
4098 0 : shard_params,
4099 0 : })
4100 0 : }
4101 :
4102 : /// Returns None if the input iterator of shards does not include a shard with number=0
4103 0 : fn tenant_describe_impl<'a>(
4104 0 : &self,
4105 0 : shards: impl Iterator<Item = &'a TenantShard>,
4106 0 : ) -> Option<TenantDescribeResponse> {
4107 0 : let mut shard_zero = None;
4108 0 : let mut describe_shards = Vec::new();
4109 :
4110 0 : for shard in shards {
4111 0 : if shard.tenant_shard_id.is_shard_zero() {
4112 0 : shard_zero = Some(shard);
4113 0 : }
4114 :
4115 0 : describe_shards.push(TenantDescribeResponseShard {
4116 0 : tenant_shard_id: shard.tenant_shard_id,
4117 0 : node_attached: *shard.intent.get_attached(),
4118 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4119 0 : last_error: shard
4120 0 : .last_error
4121 0 : .lock()
4122 0 : .unwrap()
4123 0 : .as_ref()
4124 0 : .map(|e| format!("{e}"))
4125 0 : .unwrap_or("".to_string())
4126 0 : .clone(),
4127 0 : is_reconciling: shard.reconciler.is_some(),
4128 0 : is_pending_compute_notification: shard.pending_compute_notification,
4129 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4130 0 : scheduling_policy: *shard.get_scheduling_policy(),
4131 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4132 : })
4133 : }
4134 :
4135 0 : let shard_zero = shard_zero?;
4136 :
4137 0 : Some(TenantDescribeResponse {
4138 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
4139 0 : shards: describe_shards,
4140 0 : stripe_size: shard_zero.shard.stripe_size,
4141 0 : policy: shard_zero.policy.clone(),
4142 0 : config: shard_zero.config.clone(),
4143 0 : })
4144 0 : }
4145 :
4146 0 : pub(crate) fn tenant_describe(
4147 0 : &self,
4148 0 : tenant_id: TenantId,
4149 0 : ) -> Result<TenantDescribeResponse, ApiError> {
4150 0 : let locked = self.inner.read().unwrap();
4151 0 :
4152 0 : self.tenant_describe_impl(
4153 0 : locked
4154 0 : .tenants
4155 0 : .range(TenantShardId::tenant_range(tenant_id))
4156 0 : .map(|(_k, v)| v),
4157 0 : )
4158 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
4159 0 : }
4160 :
4161 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
4162 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
4163 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
4164 : /// in our external API.
4165 0 : pub(crate) fn tenant_list(
4166 0 : &self,
4167 0 : limit: Option<usize>,
4168 0 : start_after: Option<TenantId>,
4169 0 : ) -> Vec<TenantDescribeResponse> {
4170 0 : let locked = self.inner.read().unwrap();
4171 :
4172 : // Apply start_from parameter
4173 0 : let shard_range = match start_after {
4174 0 : None => locked.tenants.range(..),
4175 0 : Some(tenant_id) => locked.tenants.range(
4176 0 : TenantShardId {
4177 0 : tenant_id,
4178 0 : shard_number: ShardNumber(u8::MAX),
4179 0 : shard_count: ShardCount(u8::MAX),
4180 0 : }..,
4181 0 : ),
4182 : };
4183 :
4184 0 : let mut result = Vec::new();
4185 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
4186 0 : result.push(
4187 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
4188 0 : .expect("Groups are always non-empty"),
4189 0 : );
4190 :
4191 : // Enforce `limit` parameter
4192 0 : if let Some(limit) = limit {
4193 0 : if result.len() >= limit {
4194 0 : break;
4195 0 : }
4196 0 : }
4197 : }
4198 :
4199 0 : result
4200 0 : }
4201 :
4202 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
4203 : async fn abort_tenant_shard_split(
4204 : &self,
4205 : op: &TenantShardSplitAbort,
4206 : ) -> Result<(), TenantShardSplitAbortError> {
4207 : // Cleaning up a split:
4208 : // - Parent shards are not destroyed during a split, just detached.
4209 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
4210 : // just the children attached, or both.
4211 : //
4212 : // Therefore our work to do is to:
4213 : // 1. Clean up storage controller's internal state to just refer to parents, no children
4214 : // 2. Call out to pageservers to ensure that children are detached
4215 : // 3. Call out to pageservers to ensure that parents are attached.
4216 : //
4217 : // Crash safety:
4218 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
4219 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
4220 : // and detach them.
4221 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
4222 : // from our database, then we will re-enter this cleanup routine on startup.
4223 :
4224 : let TenantShardSplitAbort {
4225 : tenant_id,
4226 : new_shard_count,
4227 : new_stripe_size,
4228 : ..
4229 : } = op;
4230 :
4231 : // First abort persistent state, if any exists.
4232 : match self
4233 : .persistence
4234 : .abort_shard_split(*tenant_id, *new_shard_count)
4235 : .await?
4236 : {
4237 : AbortShardSplitStatus::Aborted => {
4238 : // Proceed to roll back any child shards created on pageservers
4239 : }
4240 : AbortShardSplitStatus::Complete => {
4241 : // The split completed (we might hit that path if e.g. our database transaction
4242 : // to write the completion landed in the database, but we dropped connection
4243 : // before seeing the result).
4244 : //
4245 : // We must update in-memory state to reflect the successful split.
4246 : self.tenant_shard_split_commit_inmem(
4247 : *tenant_id,
4248 : *new_shard_count,
4249 : *new_stripe_size,
4250 : );
4251 : return Ok(());
4252 : }
4253 : }
4254 :
4255 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
4256 : let detach_locations: Vec<(Node, TenantShardId)> = {
4257 : let mut detach_locations = Vec::new();
4258 : let mut locked = self.inner.write().unwrap();
4259 : let (nodes, tenants, scheduler) = locked.parts_mut();
4260 :
4261 : for (tenant_shard_id, shard) in
4262 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
4263 : {
4264 : if shard.shard.count == op.new_shard_count {
4265 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
4266 : // is infallible, so if we got an error we shouldn't have got that far.
4267 : tracing::warn!(
4268 : "During split abort, child shard {tenant_shard_id} found in-memory"
4269 : );
4270 : continue;
4271 : }
4272 :
4273 : // Add the children of this shard to this list of things to detach
4274 : if let Some(node_id) = shard.intent.get_attached() {
4275 : for child_id in tenant_shard_id.split(*new_shard_count) {
4276 : detach_locations.push((
4277 : nodes
4278 : .get(node_id)
4279 : .expect("Intent references nonexistent node")
4280 : .clone(),
4281 : child_id,
4282 : ));
4283 : }
4284 : } else {
4285 : tracing::warn!(
4286 : "During split abort, shard {tenant_shard_id} has no attached location"
4287 : );
4288 : }
4289 :
4290 : tracing::info!("Restoring parent shard {tenant_shard_id}");
4291 :
4292 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
4293 : // if the original attachment location is offline.
4294 : if let Some(node_id) = shard.intent.get_attached() {
4295 : if !nodes.get(node_id).unwrap().is_available() {
4296 : tracing::info!("Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}");
4297 : shard.intent.demote_attached(scheduler, *node_id);
4298 : }
4299 : }
4300 : for node_id in shard.intent.get_secondary().clone() {
4301 : if !nodes.get(&node_id).unwrap().is_available() {
4302 : tracing::info!("Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}");
4303 : shard.intent.remove_secondary(scheduler, node_id);
4304 : }
4305 : }
4306 :
4307 : shard.splitting = SplitState::Idle;
4308 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
4309 : // If this shard can't be scheduled now (perhaps due to offline nodes or
4310 : // capacity issues), that must not prevent us rolling back a split. In this
4311 : // case it should be eventually scheduled in the background.
4312 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
4313 : }
4314 :
4315 : self.maybe_reconcile_shard(shard, nodes);
4316 : }
4317 :
4318 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
4319 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
4320 :
4321 : detach_locations
4322 : };
4323 :
4324 : for (node, child_id) in detach_locations {
4325 : if !node.is_available() {
4326 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
4327 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
4328 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
4329 : // them from the node.
4330 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
4331 : continue;
4332 : }
4333 :
4334 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
4335 : // a 503 and retry, up to our limit.
4336 : tracing::info!("Detaching {child_id} on {node}...");
4337 : match node
4338 : .with_client_retries(
4339 0 : |client| async move {
4340 0 : let config = LocationConfig {
4341 0 : mode: LocationConfigMode::Detached,
4342 0 : generation: None,
4343 0 : secondary_conf: None,
4344 0 : shard_number: child_id.shard_number.0,
4345 0 : shard_count: child_id.shard_count.literal(),
4346 0 : // Stripe size and tenant config don't matter when detaching
4347 0 : shard_stripe_size: 0,
4348 0 : tenant_conf: TenantConfig::default(),
4349 0 : };
4350 0 :
4351 0 : client.location_config(child_id, config, None, false).await
4352 0 : },
4353 : &self.config.jwt_token,
4354 : 1,
4355 : 10,
4356 : Duration::from_secs(5),
4357 : &self.cancel,
4358 : )
4359 : .await
4360 : {
4361 : Some(Ok(_)) => {}
4362 : Some(Err(e)) => {
4363 : // We failed to communicate with the remote node. This is problematic: we may be
4364 : // leaving it with a rogue child shard.
4365 : tracing::warn!(
4366 : "Failed to detach child {child_id} from node {node} during abort"
4367 : );
4368 : return Err(e.into());
4369 : }
4370 : None => {
4371 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
4372 : // clean up on restart. The node going offline requires a retry.
4373 : return Err(TenantShardSplitAbortError::Unavailable);
4374 : }
4375 : };
4376 : }
4377 :
4378 : tracing::info!("Successfully aborted split");
4379 : Ok(())
4380 : }
4381 :
4382 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
4383 : /// of the tenant map to reflect the child shards that exist after the split.
4384 0 : fn tenant_shard_split_commit_inmem(
4385 0 : &self,
4386 0 : tenant_id: TenantId,
4387 0 : new_shard_count: ShardCount,
4388 0 : new_stripe_size: Option<ShardStripeSize>,
4389 0 : ) -> (
4390 0 : TenantShardSplitResponse,
4391 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
4392 0 : Vec<ReconcilerWaiter>,
4393 0 : ) {
4394 0 : let mut response = TenantShardSplitResponse {
4395 0 : new_shards: Vec::new(),
4396 0 : };
4397 0 : let mut child_locations = Vec::new();
4398 0 : let mut waiters = Vec::new();
4399 0 :
4400 0 : {
4401 0 : let mut locked = self.inner.write().unwrap();
4402 0 :
4403 0 : let parent_ids = locked
4404 0 : .tenants
4405 0 : .range(TenantShardId::tenant_range(tenant_id))
4406 0 : .map(|(shard_id, _)| *shard_id)
4407 0 : .collect::<Vec<_>>();
4408 0 :
4409 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4410 0 : for parent_id in parent_ids {
4411 0 : let child_ids = parent_id.split(new_shard_count);
4412 :
4413 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
4414 0 : let mut old_state = tenants
4415 0 : .remove(&parent_id)
4416 0 : .expect("It was present, we just split it");
4417 0 :
4418 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
4419 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
4420 0 : // nothing else can clear this.
4421 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
4422 :
4423 0 : let old_attached = old_state.intent.get_attached().unwrap();
4424 0 : old_state.intent.clear(scheduler);
4425 0 : let generation = old_state.generation.expect("Shard must have been attached");
4426 0 : (
4427 0 : old_attached,
4428 0 : generation,
4429 0 : old_state.policy.clone(),
4430 0 : old_state.shard,
4431 0 : old_state.config.clone(),
4432 0 : old_state.preferred_az().cloned(),
4433 0 : )
4434 0 : };
4435 0 :
4436 0 : let mut schedule_context = ScheduleContext::default();
4437 0 : for child in child_ids {
4438 0 : let mut child_shard = parent_ident;
4439 0 : child_shard.number = child.shard_number;
4440 0 : child_shard.count = child.shard_count;
4441 0 : if let Some(stripe_size) = new_stripe_size {
4442 0 : child_shard.stripe_size = stripe_size;
4443 0 : }
4444 :
4445 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
4446 0 : child_observed.insert(
4447 0 : pageserver,
4448 0 : ObservedStateLocation {
4449 0 : conf: Some(attached_location_conf(
4450 0 : generation,
4451 0 : &child_shard,
4452 0 : &config,
4453 0 : &policy,
4454 0 : )),
4455 0 : },
4456 0 : );
4457 0 :
4458 0 : let mut child_state =
4459 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
4460 0 : child_state.intent =
4461 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
4462 0 : child_state.observed = ObservedState {
4463 0 : locations: child_observed,
4464 0 : };
4465 0 : child_state.generation = Some(generation);
4466 0 : child_state.config = config.clone();
4467 0 :
4468 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
4469 0 : // as at this point in the split process we have succeeded and this part is infallible:
4470 0 : // we will never need to do any special recovery from this state.
4471 0 :
4472 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
4473 :
4474 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
4475 : // This is not fatal, because we've implicitly already got an attached
4476 : // location for the child shard. Failure here just means we couldn't
4477 : // find a secondary (e.g. because cluster is overloaded).
4478 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
4479 0 : }
4480 : // In the background, attach secondary locations for the new shards
4481 0 : if let Some(waiter) = self.maybe_reconcile_shard(&mut child_state, nodes) {
4482 0 : waiters.push(waiter);
4483 0 : }
4484 :
4485 0 : tenants.insert(child, child_state);
4486 0 : response.new_shards.push(child);
4487 : }
4488 : }
4489 0 : (response, child_locations, waiters)
4490 0 : }
4491 0 : }
4492 :
4493 0 : async fn tenant_shard_split_start_secondaries(
4494 0 : &self,
4495 0 : tenant_id: TenantId,
4496 0 : waiters: Vec<ReconcilerWaiter>,
4497 0 : ) {
4498 : // Wait for initial reconcile of child shards, this creates the secondary locations
4499 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
4500 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
4501 : // their secondaries couldn't be attached.
4502 0 : tracing::warn!("Failed to reconcile after split: {e}");
4503 0 : return;
4504 0 : }
4505 :
4506 : // Take the state lock to discover the attached & secondary intents for all shards
4507 0 : let (attached, secondary) = {
4508 0 : let locked = self.inner.read().unwrap();
4509 0 : let mut attached = Vec::new();
4510 0 : let mut secondary = Vec::new();
4511 :
4512 0 : for (tenant_shard_id, shard) in
4513 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4514 : {
4515 0 : let Some(node_id) = shard.intent.get_attached() else {
4516 : // Unexpected. Race with a PlacementPolicy change?
4517 0 : tracing::warn!(
4518 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
4519 : );
4520 0 : continue;
4521 : };
4522 :
4523 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
4524 : // No secondary location. Nothing for us to do.
4525 0 : continue;
4526 : };
4527 :
4528 0 : let attached_node = locked
4529 0 : .nodes
4530 0 : .get(node_id)
4531 0 : .expect("Pageservers may not be deleted while referenced");
4532 0 :
4533 0 : let secondary_node = locked
4534 0 : .nodes
4535 0 : .get(secondary_node_id)
4536 0 : .expect("Pageservers may not be deleted while referenced");
4537 0 :
4538 0 : attached.push((*tenant_shard_id, attached_node.clone()));
4539 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
4540 : }
4541 0 : (attached, secondary)
4542 0 : };
4543 0 :
4544 0 : if secondary.is_empty() {
4545 : // No secondary locations; nothing for us to do
4546 0 : return;
4547 0 : }
4548 :
4549 0 : for result in self
4550 0 : .tenant_for_shards_api(
4551 0 : attached,
4552 0 : |tenant_shard_id, client| async move {
4553 0 : client.tenant_heatmap_upload(tenant_shard_id).await
4554 0 : },
4555 0 : 1,
4556 0 : 1,
4557 0 : SHORT_RECONCILE_TIMEOUT,
4558 0 : &self.cancel,
4559 0 : )
4560 0 : .await
4561 : {
4562 0 : if let Err(e) = result {
4563 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
4564 0 : return;
4565 0 : }
4566 : }
4567 :
4568 0 : for result in self
4569 0 : .tenant_for_shards_api(
4570 0 : secondary,
4571 0 : |tenant_shard_id, client| async move {
4572 0 : client
4573 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
4574 0 : .await
4575 0 : },
4576 0 : 1,
4577 0 : 1,
4578 0 : SHORT_RECONCILE_TIMEOUT,
4579 0 : &self.cancel,
4580 0 : )
4581 0 : .await
4582 : {
4583 0 : if let Err(e) = result {
4584 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
4585 0 : return;
4586 0 : }
4587 : }
4588 0 : }
4589 :
4590 0 : pub(crate) async fn tenant_shard_split(
4591 0 : &self,
4592 0 : tenant_id: TenantId,
4593 0 : split_req: TenantShardSplitRequest,
4594 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
4595 : // TODO: return 503 if we get stuck waiting for this lock
4596 : // (issue https://github.com/neondatabase/neon/issues/7108)
4597 0 : let _tenant_lock = trace_exclusive_lock(
4598 0 : &self.tenant_op_locks,
4599 0 : tenant_id,
4600 0 : TenantOperations::ShardSplit,
4601 0 : )
4602 0 : .await;
4603 :
4604 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
4605 0 : let new_stripe_size = split_req.new_stripe_size;
4606 :
4607 : // Validate the request and construct parameters. This phase is fallible, but does not require
4608 : // rollback on errors, as it does no I/O and mutates no state.
4609 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
4610 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
4611 0 : ShardSplitAction::Split(params) => params,
4612 : };
4613 :
4614 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
4615 : // we must roll back.
4616 0 : let r = self
4617 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
4618 0 : .await;
4619 :
4620 0 : let (response, waiters) = match r {
4621 0 : Ok(r) => r,
4622 0 : Err(e) => {
4623 0 : // Split might be part-done, we must do work to abort it.
4624 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
4625 0 : self.abort_tx
4626 0 : .send(TenantShardSplitAbort {
4627 0 : tenant_id,
4628 0 : new_shard_count,
4629 0 : new_stripe_size,
4630 0 : _tenant_lock,
4631 0 : })
4632 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
4633 0 : .ok();
4634 0 : return Err(e);
4635 : }
4636 : };
4637 :
4638 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
4639 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
4640 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
4641 : // in [`Self::optimize_all`]
4642 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
4643 0 : .await;
4644 0 : Ok(response)
4645 0 : }
4646 :
4647 0 : fn prepare_tenant_shard_split(
4648 0 : &self,
4649 0 : tenant_id: TenantId,
4650 0 : split_req: TenantShardSplitRequest,
4651 0 : ) -> Result<ShardSplitAction, ApiError> {
4652 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
4653 0 : anyhow::anyhow!("failpoint")
4654 0 : )));
4655 :
4656 0 : let mut policy = None;
4657 0 : let mut config = None;
4658 0 : let mut shard_ident = None;
4659 0 : let mut preferred_az_id = None;
4660 : // Validate input, and calculate which shards we will create
4661 0 : let (old_shard_count, targets) =
4662 : {
4663 0 : let locked = self.inner.read().unwrap();
4664 0 :
4665 0 : let pageservers = locked.nodes.clone();
4666 0 :
4667 0 : let mut targets = Vec::new();
4668 0 :
4669 0 : // In case this is a retry, count how many already-split shards we found
4670 0 : let mut children_found = Vec::new();
4671 0 : let mut old_shard_count = None;
4672 :
4673 0 : for (tenant_shard_id, shard) in
4674 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4675 : {
4676 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
4677 : Ordering::Equal => {
4678 : // Already split this
4679 0 : children_found.push(*tenant_shard_id);
4680 0 : continue;
4681 : }
4682 : Ordering::Greater => {
4683 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4684 0 : "Requested count {} but already have shards at count {}",
4685 0 : split_req.new_shard_count,
4686 0 : shard.shard.count.count()
4687 0 : )));
4688 : }
4689 0 : Ordering::Less => {
4690 0 : // Fall through: this shard has lower count than requested,
4691 0 : // is a candidate for splitting.
4692 0 : }
4693 0 : }
4694 0 :
4695 0 : match old_shard_count {
4696 0 : None => old_shard_count = Some(shard.shard.count),
4697 0 : Some(old_shard_count) => {
4698 0 : if old_shard_count != shard.shard.count {
4699 : // We may hit this case if a caller asked for two splits to
4700 : // different sizes, before the first one is complete.
4701 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
4702 : // of shard_count=1 and shard_count=2 shards in the map.
4703 0 : return Err(ApiError::Conflict(
4704 0 : "Cannot split, currently mid-split".to_string(),
4705 0 : ));
4706 0 : }
4707 : }
4708 : }
4709 0 : if policy.is_none() {
4710 0 : policy = Some(shard.policy.clone());
4711 0 : }
4712 0 : if shard_ident.is_none() {
4713 0 : shard_ident = Some(shard.shard);
4714 0 : }
4715 0 : if config.is_none() {
4716 0 : config = Some(shard.config.clone());
4717 0 : }
4718 0 : if preferred_az_id.is_none() {
4719 0 : preferred_az_id = shard.preferred_az().cloned();
4720 0 : }
4721 :
4722 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4723 0 : tracing::info!(
4724 0 : "Tenant shard {} already has shard count {}",
4725 : tenant_shard_id,
4726 : split_req.new_shard_count
4727 : );
4728 0 : continue;
4729 0 : }
4730 :
4731 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4732 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4733 0 : ))?;
4734 :
4735 0 : let node = pageservers
4736 0 : .get(&node_id)
4737 0 : .expect("Pageservers may not be deleted while referenced");
4738 0 :
4739 0 : targets.push(ShardSplitTarget {
4740 0 : parent_id: *tenant_shard_id,
4741 0 : node: node.clone(),
4742 0 : child_ids: tenant_shard_id
4743 0 : .split(ShardCount::new(split_req.new_shard_count)),
4744 0 : });
4745 : }
4746 :
4747 0 : if targets.is_empty() {
4748 0 : if children_found.len() == split_req.new_shard_count as usize {
4749 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4750 0 : new_shards: children_found,
4751 0 : }));
4752 : } else {
4753 : // No shards found to split, and no existing children found: the
4754 : // tenant doesn't exist at all.
4755 0 : return Err(ApiError::NotFound(
4756 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4757 0 : ));
4758 : }
4759 0 : }
4760 0 :
4761 0 : (old_shard_count, targets)
4762 0 : };
4763 0 :
4764 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
4765 0 : let old_shard_count = old_shard_count.unwrap();
4766 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
4767 : // This ShardIdentity will be used as the template for all children, so this implicitly
4768 : // applies the new stripe size to the children.
4769 0 : let mut shard_ident = shard_ident.unwrap();
4770 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
4771 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)));
4772 0 : }
4773 0 :
4774 0 : shard_ident.stripe_size = new_stripe_size;
4775 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
4776 0 : shard_ident
4777 : } else {
4778 0 : shard_ident.unwrap()
4779 : };
4780 0 : let policy = policy.unwrap();
4781 0 : let config = config.unwrap();
4782 0 :
4783 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
4784 0 : old_shard_count,
4785 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
4786 0 : new_stripe_size: split_req.new_stripe_size,
4787 0 : targets,
4788 0 : policy,
4789 0 : config,
4790 0 : shard_ident,
4791 0 : preferred_az_id,
4792 0 : })))
4793 0 : }
4794 :
4795 0 : async fn do_tenant_shard_split(
4796 0 : &self,
4797 0 : tenant_id: TenantId,
4798 0 : params: Box<ShardSplitParams>,
4799 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
4800 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
4801 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
4802 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
4803 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
4804 0 : // (https://github.com/neondatabase/neon/issues/6676)
4805 0 :
4806 0 : let ShardSplitParams {
4807 0 : old_shard_count,
4808 0 : new_shard_count,
4809 0 : new_stripe_size,
4810 0 : mut targets,
4811 0 : policy,
4812 0 : config,
4813 0 : shard_ident,
4814 0 : preferred_az_id,
4815 0 : } = *params;
4816 :
4817 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
4818 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
4819 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
4820 : // at the time of split.
4821 0 : let waiters = {
4822 0 : let mut locked = self.inner.write().unwrap();
4823 0 : let mut waiters = Vec::new();
4824 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4825 0 : for target in &mut targets {
4826 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
4827 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4828 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4829 0 : "Shard {} not found",
4830 0 : target.parent_id
4831 0 : )));
4832 : };
4833 :
4834 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
4835 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4836 0 : return Err(ApiError::Conflict(format!(
4837 0 : "Shard {} unexpectedly rescheduled during split",
4838 0 : target.parent_id
4839 0 : )));
4840 0 : }
4841 0 :
4842 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
4843 0 : shard.intent.clear_secondary(scheduler);
4844 :
4845 : // Run Reconciler to execute detach fo secondary locations.
4846 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
4847 0 : waiters.push(waiter);
4848 0 : }
4849 : }
4850 0 : waiters
4851 0 : };
4852 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
4853 :
4854 : // Before creating any new child shards in memory or on the pageservers, persist them: this
4855 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
4856 : // acts as the protection against two concurrent attempts to split: one of them will get a database
4857 : // error trying to insert the child shards.
4858 0 : let mut child_tsps = Vec::new();
4859 0 : for target in &targets {
4860 0 : let mut this_child_tsps = Vec::new();
4861 0 : for child in &target.child_ids {
4862 0 : let mut child_shard = shard_ident;
4863 0 : child_shard.number = child.shard_number;
4864 0 : child_shard.count = child.shard_count;
4865 0 :
4866 0 : tracing::info!(
4867 0 : "Create child shard persistence with stripe size {}",
4868 : shard_ident.stripe_size.0
4869 : );
4870 :
4871 0 : this_child_tsps.push(TenantShardPersistence {
4872 0 : tenant_id: child.tenant_id.to_string(),
4873 0 : shard_number: child.shard_number.0 as i32,
4874 0 : shard_count: child.shard_count.literal() as i32,
4875 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
4876 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
4877 0 : // populate the correct generation as part of its transaction, to protect us
4878 0 : // against racing with changes in the state of the parent.
4879 0 : generation: None,
4880 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
4881 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
4882 0 : config: serde_json::to_string(&config).unwrap(),
4883 0 : splitting: SplitState::Splitting,
4884 0 :
4885 0 : // Scheduling policies and preferred AZ do not carry through to children
4886 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
4887 0 : .unwrap(),
4888 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
4889 0 : });
4890 0 : }
4891 :
4892 0 : child_tsps.push((target.parent_id, this_child_tsps));
4893 : }
4894 :
4895 0 : if let Err(e) = self
4896 0 : .persistence
4897 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
4898 0 : .await
4899 : {
4900 0 : match e {
4901 : DatabaseError::Query(diesel::result::Error::DatabaseError(
4902 : DatabaseErrorKind::UniqueViolation,
4903 : _,
4904 : )) => {
4905 : // Inserting a child shard violated a unique constraint: we raced with another call to
4906 : // this function
4907 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
4908 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
4909 : }
4910 0 : _ => return Err(ApiError::InternalServerError(e.into())),
4911 : }
4912 0 : }
4913 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
4914 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4915 0 : ));
4916 :
4917 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
4918 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
4919 : // is not set in memory, then it was not persisted.
4920 : {
4921 0 : let mut locked = self.inner.write().unwrap();
4922 0 : for target in &targets {
4923 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
4924 0 : parent_shard.splitting = SplitState::Splitting;
4925 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
4926 0 : // split operation.
4927 0 : parent_shard
4928 0 : .observed
4929 0 : .locations
4930 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
4931 0 : }
4932 : }
4933 : }
4934 :
4935 : // TODO: issue split calls concurrently (this only matters once we're splitting
4936 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
4937 :
4938 0 : for target in &targets {
4939 : let ShardSplitTarget {
4940 0 : parent_id,
4941 0 : node,
4942 0 : child_ids,
4943 0 : } = target;
4944 0 : let client = PageserverClient::new(
4945 0 : node.get_id(),
4946 0 : node.base_url(),
4947 0 : self.config.jwt_token.as_deref(),
4948 0 : );
4949 0 : let response = client
4950 0 : .tenant_shard_split(
4951 0 : *parent_id,
4952 0 : TenantShardSplitRequest {
4953 0 : new_shard_count: new_shard_count.literal(),
4954 0 : new_stripe_size,
4955 0 : },
4956 0 : )
4957 0 : .await
4958 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
4959 :
4960 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
4961 0 : "failpoint".to_string()
4962 0 : )));
4963 :
4964 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
4965 :
4966 0 : tracing::info!(
4967 0 : "Split {} into {}",
4968 0 : parent_id,
4969 0 : response
4970 0 : .new_shards
4971 0 : .iter()
4972 0 : .map(|s| format!("{:?}", s))
4973 0 : .collect::<Vec<_>>()
4974 0 : .join(",")
4975 : );
4976 :
4977 0 : if &response.new_shards != child_ids {
4978 : // This should never happen: the pageserver should agree with us on how shard splits work.
4979 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4980 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
4981 0 : parent_id,
4982 0 : response.new_shards,
4983 0 : child_ids
4984 0 : )));
4985 0 : }
4986 : }
4987 :
4988 : // TODO: if the pageserver restarted concurrently with our split API call,
4989 : // the actual generation of the child shard might differ from the generation
4990 : // we expect it to have. In order for our in-database generation to end up
4991 : // correct, we should carry the child generation back in the response and apply it here
4992 : // in complete_shard_split (and apply the correct generation in memory)
4993 : // (or, we can carry generation in the request and reject the request if
4994 : // it doesn't match, but that requires more retry logic on this side)
4995 :
4996 0 : self.persistence
4997 0 : .complete_shard_split(tenant_id, old_shard_count)
4998 0 : .await?;
4999 :
5000 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5001 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5002 0 : ));
5003 :
5004 : // Replace all the shards we just split with their children: this phase is infallible.
5005 0 : let (response, child_locations, waiters) =
5006 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5007 0 :
5008 0 : // Send compute notifications for all the new shards
5009 0 : let mut failed_notifications = Vec::new();
5010 0 : for (child_id, child_ps, stripe_size) in child_locations {
5011 0 : if let Err(e) = self
5012 0 : .compute_hook
5013 0 : .notify(
5014 0 : compute_hook::ShardUpdate {
5015 0 : tenant_shard_id: child_id,
5016 0 : node_id: child_ps,
5017 0 : stripe_size,
5018 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5019 0 : },
5020 0 : &self.cancel,
5021 0 : )
5022 0 : .await
5023 : {
5024 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5025 : child_id, child_ps);
5026 0 : failed_notifications.push(child_id);
5027 0 : }
5028 : }
5029 :
5030 : // If we failed any compute notifications, make a note to retry later.
5031 0 : if !failed_notifications.is_empty() {
5032 0 : let mut locked = self.inner.write().unwrap();
5033 0 : for failed in failed_notifications {
5034 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5035 0 : shard.pending_compute_notification = true;
5036 0 : }
5037 : }
5038 0 : }
5039 :
5040 0 : Ok((response, waiters))
5041 0 : }
5042 :
5043 0 : pub(crate) async fn tenant_shard_migrate(
5044 0 : &self,
5045 0 : tenant_shard_id: TenantShardId,
5046 0 : migrate_req: TenantShardMigrateRequest,
5047 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5048 0 : let waiter = {
5049 0 : let mut locked = self.inner.write().unwrap();
5050 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5051 :
5052 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5053 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5054 0 : "Node {} not found",
5055 0 : migrate_req.node_id
5056 0 : )));
5057 : };
5058 :
5059 0 : if !node.is_available() {
5060 : // Warn but proceed: the caller may intend to manually adjust the placement of
5061 : // a shard even if the node is down, e.g. if intervening during an incident.
5062 0 : tracing::warn!("Migrating to unavailable node {node}");
5063 0 : }
5064 :
5065 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5066 0 : return Err(ApiError::NotFound(
5067 0 : anyhow::anyhow!("Tenant shard not found").into(),
5068 0 : ));
5069 : };
5070 :
5071 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5072 : // No-op case: we will still proceed to wait for reconciliation in case it is
5073 : // incomplete from an earlier update to the intent.
5074 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
5075 : } else {
5076 0 : let old_attached = *shard.intent.get_attached();
5077 0 :
5078 0 : match shard.policy {
5079 0 : PlacementPolicy::Attached(n) => {
5080 0 : // If our new attached node was a secondary, it no longer should be.
5081 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
5082 0 :
5083 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
5084 :
5085 : // If we were already attached to something, demote that to a secondary
5086 0 : if let Some(old_attached) = old_attached {
5087 0 : if n > 0 {
5088 : // Remove other secondaries to make room for the location we'll demote
5089 0 : while shard.intent.get_secondary().len() >= n {
5090 0 : shard.intent.pop_secondary(scheduler);
5091 0 : }
5092 :
5093 0 : shard.intent.push_secondary(scheduler, old_attached);
5094 0 : }
5095 0 : }
5096 : }
5097 0 : PlacementPolicy::Secondary => {
5098 0 : shard.intent.clear(scheduler);
5099 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5100 0 : }
5101 : PlacementPolicy::Detached => {
5102 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5103 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
5104 0 : )))
5105 : }
5106 : }
5107 :
5108 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
5109 0 : shard.sequence = shard.sequence.next();
5110 : }
5111 :
5112 0 : self.maybe_reconcile_shard(shard, nodes)
5113 : };
5114 :
5115 0 : if let Some(waiter) = waiter {
5116 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5117 : } else {
5118 0 : tracing::info!("Migration is a no-op");
5119 : }
5120 :
5121 0 : Ok(TenantShardMigrateResponse {})
5122 0 : }
5123 :
5124 0 : pub(crate) async fn tenant_shard_migrate_secondary(
5125 0 : &self,
5126 0 : tenant_shard_id: TenantShardId,
5127 0 : migrate_req: TenantShardMigrateRequest,
5128 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5129 0 : let waiter = {
5130 0 : let mut locked = self.inner.write().unwrap();
5131 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5132 :
5133 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5134 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5135 0 : "Node {} not found",
5136 0 : migrate_req.node_id
5137 0 : )));
5138 : };
5139 :
5140 0 : if !node.is_available() {
5141 : // Warn but proceed: the caller may intend to manually adjust the placement of
5142 : // a shard even if the node is down, e.g. if intervening during an incident.
5143 0 : tracing::warn!("Migrating to unavailable node {node}");
5144 0 : }
5145 :
5146 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5147 0 : return Err(ApiError::NotFound(
5148 0 : anyhow::anyhow!("Tenant shard not found").into(),
5149 0 : ));
5150 : };
5151 :
5152 0 : if shard.intent.get_secondary().len() == 1
5153 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
5154 : {
5155 0 : tracing::info!(
5156 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
5157 : shard.intent
5158 : );
5159 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5160 0 : tracing::info!("Migrating secondary to {node}: already attached where we were asked to create a secondary");
5161 : } else {
5162 0 : let old_secondaries = shard.intent.get_secondary().clone();
5163 0 : for secondary in old_secondaries {
5164 0 : shard.intent.remove_secondary(scheduler, secondary);
5165 0 : }
5166 :
5167 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5168 0 : shard.sequence = shard.sequence.next();
5169 0 : tracing::info!(
5170 0 : "Migrating secondary to {node}: new intent {:?}",
5171 : shard.intent
5172 : );
5173 : }
5174 :
5175 0 : self.maybe_reconcile_shard(shard, nodes)
5176 : };
5177 :
5178 0 : if let Some(waiter) = waiter {
5179 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5180 : } else {
5181 0 : tracing::info!("Migration is a no-op");
5182 : }
5183 :
5184 0 : Ok(TenantShardMigrateResponse {})
5185 0 : }
5186 :
5187 : /// 'cancel' in this context means cancel any ongoing reconcile
5188 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
5189 0 : &self,
5190 0 : tenant_shard_id: TenantShardId,
5191 0 : ) -> Result<(), ApiError> {
5192 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
5193 0 : let waiter = {
5194 0 : let locked = self.inner.write().unwrap();
5195 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5196 0 : return Err(ApiError::NotFound(
5197 0 : anyhow::anyhow!("Tenant shard not found").into(),
5198 0 : ));
5199 : };
5200 :
5201 0 : let waiter = shard.get_waiter();
5202 0 : match waiter {
5203 : None => {
5204 0 : tracing::info!("Shard does not have an ongoing Reconciler");
5205 0 : return Ok(());
5206 : }
5207 0 : Some(waiter) => {
5208 0 : tracing::info!("Cancelling Reconciler");
5209 0 : shard.cancel_reconciler();
5210 0 : waiter
5211 0 : }
5212 0 : }
5213 0 : };
5214 0 :
5215 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
5216 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
5217 0 : // the Reconciler is misbehaving and not respecting the cancellation token
5218 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
5219 0 : .await?;
5220 :
5221 0 : Ok(())
5222 0 : }
5223 :
5224 : /// This is for debug/support only: we simply drop all state for a tenant, without
5225 : /// detaching or deleting it on pageservers.
5226 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
5227 0 : self.persistence.delete_tenant(tenant_id).await?;
5228 :
5229 0 : let mut locked = self.inner.write().unwrap();
5230 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
5231 0 : let mut shards = Vec::new();
5232 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
5233 0 : shards.push(*tenant_shard_id);
5234 0 : }
5235 :
5236 0 : for shard_id in shards {
5237 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
5238 0 : shard.intent.clear(scheduler);
5239 0 : }
5240 : }
5241 :
5242 0 : Ok(())
5243 0 : }
5244 :
5245 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
5246 : /// tenant with a very high generation number so that it will see the existing data.
5247 0 : pub(crate) async fn tenant_import(
5248 0 : &self,
5249 0 : tenant_id: TenantId,
5250 0 : ) -> Result<TenantCreateResponse, ApiError> {
5251 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
5252 0 : let maybe_node = {
5253 0 : self.inner
5254 0 : .read()
5255 0 : .unwrap()
5256 0 : .nodes
5257 0 : .values()
5258 0 : .find(|n| n.is_available())
5259 0 : .cloned()
5260 : };
5261 0 : let Some(node) = maybe_node else {
5262 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
5263 : };
5264 :
5265 0 : let client = PageserverClient::new(
5266 0 : node.get_id(),
5267 0 : node.base_url(),
5268 0 : self.config.jwt_token.as_deref(),
5269 0 : );
5270 :
5271 0 : let scan_result = client
5272 0 : .tenant_scan_remote_storage(tenant_id)
5273 0 : .await
5274 0 : .map_err(|e| passthrough_api_error(&node, e))?;
5275 :
5276 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
5277 0 : let Some(shard_count) = scan_result
5278 0 : .shards
5279 0 : .iter()
5280 0 : .map(|s| s.tenant_shard_id.shard_count)
5281 0 : .max()
5282 : else {
5283 0 : return Err(ApiError::NotFound(
5284 0 : anyhow::anyhow!("No shards found").into(),
5285 0 : ));
5286 : };
5287 :
5288 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
5289 : // to
5290 0 : let generation = scan_result
5291 0 : .shards
5292 0 : .iter()
5293 0 : .map(|s| s.generation)
5294 0 : .max()
5295 0 : .expect("We already validated >0 shards");
5296 0 :
5297 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
5298 0 : // only work if they were using the default stripe size.
5299 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
5300 :
5301 0 : let (response, waiters) = self
5302 0 : .do_tenant_create(TenantCreateRequest {
5303 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
5304 0 : generation,
5305 0 :
5306 0 : shard_parameters: ShardParameters {
5307 0 : count: shard_count,
5308 0 : stripe_size,
5309 0 : },
5310 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
5311 0 : config: TenantConfig::default(),
5312 0 : })
5313 0 : .await?;
5314 :
5315 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
5316 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
5317 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
5318 : // reconcile, as reconciliation includes notifying compute.
5319 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
5320 0 : }
5321 :
5322 0 : Ok(response)
5323 0 : }
5324 :
5325 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
5326 : /// we don't have to make TenantShard clonable in the return path.
5327 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5328 0 : let serialized = {
5329 0 : let locked = self.inner.read().unwrap();
5330 0 : let result = locked.tenants.values().collect::<Vec<_>>();
5331 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
5332 : };
5333 :
5334 0 : hyper::Response::builder()
5335 0 : .status(hyper::StatusCode::OK)
5336 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5337 0 : .body(hyper::Body::from(serialized))
5338 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5339 0 : }
5340 :
5341 : /// Check the consistency of in-memory state vs. persistent state, and check that the
5342 : /// scheduler's statistics are up to date.
5343 : ///
5344 : /// These consistency checks expect an **idle** system. If changes are going on while
5345 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
5346 : /// checks, but not suitable for running continuously in the background in the field.
5347 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
5348 0 : let (mut expect_nodes, mut expect_shards) = {
5349 0 : let locked = self.inner.read().unwrap();
5350 0 :
5351 0 : locked
5352 0 : .scheduler
5353 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
5354 0 : .context("Scheduler checks")
5355 0 : .map_err(ApiError::InternalServerError)?;
5356 :
5357 0 : let expect_nodes = locked
5358 0 : .nodes
5359 0 : .values()
5360 0 : .map(|n| n.to_persistent())
5361 0 : .collect::<Vec<_>>();
5362 0 :
5363 0 : let expect_shards = locked
5364 0 : .tenants
5365 0 : .values()
5366 0 : .map(|t| t.to_persistent())
5367 0 : .collect::<Vec<_>>();
5368 :
5369 : // This method can only validate the state of an idle system: if a reconcile is in
5370 : // progress, fail out early to avoid giving false errors on state that won't match
5371 : // between database and memory under a ReconcileResult is processed.
5372 0 : for t in locked.tenants.values() {
5373 0 : if t.reconciler.is_some() {
5374 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5375 0 : "Shard {} reconciliation in progress",
5376 0 : t.tenant_shard_id
5377 0 : )));
5378 0 : }
5379 : }
5380 :
5381 0 : (expect_nodes, expect_shards)
5382 : };
5383 :
5384 0 : let mut nodes = self.persistence.list_nodes().await?;
5385 0 : expect_nodes.sort_by_key(|n| n.node_id);
5386 0 : nodes.sort_by_key(|n| n.node_id);
5387 :
5388 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
5389 0 : let node_result = if nodes != expect_nodes {
5390 0 : tracing::error!("Consistency check failed on nodes.");
5391 0 : tracing::error!(
5392 0 : "Nodes in memory: {}",
5393 0 : serde_json::to_string(&expect_nodes)
5394 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5395 : );
5396 0 : tracing::error!(
5397 0 : "Nodes in database: {}",
5398 0 : serde_json::to_string(&nodes)
5399 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5400 : );
5401 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
5402 0 : "Node consistency failure"
5403 0 : )))
5404 : } else {
5405 0 : Ok(())
5406 : };
5407 :
5408 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
5409 0 : persistent_shards
5410 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5411 0 :
5412 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5413 :
5414 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
5415 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
5416 : // fields are added, before doing a comparison.
5417 0 : for tsp in &mut persistent_shards {
5418 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
5419 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
5420 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
5421 : }
5422 :
5423 0 : if persistent_shards != expect_shards {
5424 0 : tracing::error!("Consistency check failed on shards.");
5425 :
5426 0 : tracing::error!(
5427 0 : "Shards in memory: {}",
5428 0 : serde_json::to_string(&expect_shards)
5429 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5430 : );
5431 0 : tracing::error!(
5432 0 : "Shards in database: {}",
5433 0 : serde_json::to_string(&persistent_shards)
5434 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5435 : );
5436 :
5437 : // The total dump log lines above are useful in testing but in the field grafana will
5438 : // usually just drop them because they're so large. So we also do some explicit logging
5439 : // of just the diffs.
5440 0 : let persistent_shards = persistent_shards
5441 0 : .into_iter()
5442 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5443 0 : .collect::<HashMap<_, _>>();
5444 0 : let expect_shards = expect_shards
5445 0 : .into_iter()
5446 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5447 0 : .collect::<HashMap<_, _>>();
5448 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
5449 0 : match expect_shards.get(tenant_shard_id) {
5450 : None => {
5451 0 : tracing::error!(
5452 0 : "Shard {} found in database but not in memory",
5453 : tenant_shard_id
5454 : );
5455 : }
5456 0 : Some(expect_tsp) => {
5457 0 : if expect_tsp != persistent_tsp {
5458 0 : tracing::error!(
5459 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
5460 0 : tenant_shard_id,
5461 0 : serde_json::to_string(expect_tsp).unwrap(),
5462 0 : serde_json::to_string(&persistent_tsp).unwrap()
5463 : );
5464 0 : }
5465 : }
5466 : }
5467 : }
5468 :
5469 : // Having already logged any differences, log any shards that simply aren't present in the database
5470 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
5471 0 : if !persistent_shards.contains_key(tenant_shard_id) {
5472 0 : tracing::error!(
5473 0 : "Shard {} found in memory but not in database: {}",
5474 0 : tenant_shard_id,
5475 0 : serde_json::to_string(memory_tsp)
5476 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5477 : );
5478 0 : }
5479 : }
5480 :
5481 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5482 0 : "Shard consistency failure"
5483 0 : )));
5484 0 : }
5485 0 :
5486 0 : node_result
5487 0 : }
5488 :
5489 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
5490 : /// we don't have to make TenantShard clonable in the return path.
5491 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5492 0 : let serialized = {
5493 0 : let locked = self.inner.read().unwrap();
5494 0 : serde_json::to_string(&locked.scheduler)
5495 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5496 : };
5497 :
5498 0 : hyper::Response::builder()
5499 0 : .status(hyper::StatusCode::OK)
5500 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5501 0 : .body(hyper::Body::from(serialized))
5502 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5503 0 : }
5504 :
5505 : /// This is for debug/support only: we simply drop all state for a tenant, without
5506 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
5507 : /// tenants that were on this node.
5508 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
5509 0 : self.persistence.delete_node(node_id).await?;
5510 :
5511 0 : let mut locked = self.inner.write().unwrap();
5512 :
5513 0 : for shard in locked.tenants.values_mut() {
5514 0 : shard.deref_node(node_id);
5515 0 : shard.observed.locations.remove(&node_id);
5516 0 : }
5517 :
5518 0 : let mut nodes = (*locked.nodes).clone();
5519 0 : nodes.remove(&node_id);
5520 0 : locked.nodes = Arc::new(nodes);
5521 0 : metrics::METRICS_REGISTRY
5522 0 : .metrics_group
5523 0 : .storage_controller_pageserver_nodes
5524 0 : .set(locked.nodes.len() as i64);
5525 0 :
5526 0 : locked.scheduler.node_remove(node_id);
5527 0 :
5528 0 : Ok(())
5529 0 : }
5530 :
5531 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
5532 : /// that we don't leave any bad state behind in the storage controller, but unclean
5533 : /// in the sense that we are not carefully draining the node.
5534 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
5535 0 : let _node_lock =
5536 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
5537 :
5538 : // 1. Atomically update in-memory state:
5539 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
5540 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
5541 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
5542 : // re-insert references to this node into the ObservedState of shards
5543 : // - drop the node from the scheduler
5544 : {
5545 0 : let mut locked = self.inner.write().unwrap();
5546 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5547 0 :
5548 0 : {
5549 0 : let mut nodes_mut = (*nodes).deref().clone();
5550 0 : match nodes_mut.get_mut(&node_id) {
5551 0 : Some(node) => {
5552 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
5553 0 : // if we crash it would not be desirable to leave the node paused after a restart.
5554 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
5555 0 : }
5556 : None => {
5557 0 : tracing::info!(
5558 0 : "Node not found: presuming this is a retry and returning success"
5559 : );
5560 0 : return Ok(());
5561 : }
5562 : }
5563 :
5564 0 : *nodes = Arc::new(nodes_mut);
5565 : }
5566 :
5567 0 : for (_tenant_id, mut schedule_context, shards) in
5568 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
5569 : {
5570 0 : for shard in shards {
5571 0 : if shard.deref_node(node_id) {
5572 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
5573 : // TODO: implement force flag to remove a node even if we can't reschedule
5574 : // a tenant
5575 0 : tracing::error!(
5576 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
5577 : shard.tenant_shard_id
5578 : );
5579 0 : return Err(e.into());
5580 : } else {
5581 0 : tracing::info!(
5582 0 : "Rescheduled shard {} away from node during deletion",
5583 : shard.tenant_shard_id
5584 : )
5585 : }
5586 :
5587 0 : self.maybe_reconcile_shard(shard, nodes);
5588 0 : }
5589 :
5590 : // Here we remove an existing observed location for the node we're removing, and it will
5591 : // not be re-added by a reconciler's completion because we filter out removed nodes in
5592 : // process_result.
5593 : //
5594 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
5595 : // means any reconciles we spawned will know about the node we're deleting, enabling them
5596 : // to do live migrations if it's still online.
5597 0 : shard.observed.locations.remove(&node_id);
5598 : }
5599 : }
5600 :
5601 0 : scheduler.node_remove(node_id);
5602 0 :
5603 0 : {
5604 0 : let mut nodes_mut = (**nodes).clone();
5605 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
5606 0 : // Ensure that any reconciler holding an Arc<> to this node will
5607 0 : // drop out when trying to RPC to it (setting Offline state sets the
5608 0 : // cancellation token on the Node object).
5609 0 : removed_node.set_availability(NodeAvailability::Offline);
5610 0 : }
5611 0 : *nodes = Arc::new(nodes_mut);
5612 0 : metrics::METRICS_REGISTRY
5613 0 : .metrics_group
5614 0 : .storage_controller_pageserver_nodes
5615 0 : .set(nodes.len() as i64);
5616 0 : }
5617 0 : }
5618 0 :
5619 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
5620 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
5621 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
5622 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
5623 0 : // that exists.
5624 0 :
5625 0 : // 2. Actually delete the node from the database and from in-memory state
5626 0 : tracing::info!("Deleting node from database");
5627 0 : self.persistence.delete_node(node_id).await?;
5628 :
5629 0 : Ok(())
5630 0 : }
5631 :
5632 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
5633 0 : let nodes = {
5634 0 : self.inner
5635 0 : .read()
5636 0 : .unwrap()
5637 0 : .nodes
5638 0 : .values()
5639 0 : .cloned()
5640 0 : .collect::<Vec<_>>()
5641 0 : };
5642 0 :
5643 0 : Ok(nodes)
5644 0 : }
5645 :
5646 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
5647 0 : self.inner
5648 0 : .read()
5649 0 : .unwrap()
5650 0 : .nodes
5651 0 : .get(&node_id)
5652 0 : .cloned()
5653 0 : .ok_or(ApiError::NotFound(
5654 0 : format!("Node {node_id} not registered").into(),
5655 0 : ))
5656 0 : }
5657 :
5658 0 : pub(crate) async fn get_node_shards(
5659 0 : &self,
5660 0 : node_id: NodeId,
5661 0 : ) -> Result<NodeShardResponse, ApiError> {
5662 0 : let locked = self.inner.read().unwrap();
5663 0 : let mut shards = Vec::new();
5664 0 : for (tid, tenant) in locked.tenants.iter() {
5665 0 : let is_intended_secondary = match (
5666 0 : tenant.intent.get_attached() == &Some(node_id),
5667 0 : tenant.intent.get_secondary().contains(&node_id),
5668 0 : ) {
5669 : (true, true) => {
5670 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5671 0 : "{} attached as primary+secondary on the same node",
5672 0 : tid
5673 0 : )))
5674 : }
5675 0 : (true, false) => Some(false),
5676 0 : (false, true) => Some(true),
5677 0 : (false, false) => None,
5678 : };
5679 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
5680 0 : tenant.observed.locations.get(&node_id)
5681 : {
5682 0 : Some(conf.secondary_conf.is_some())
5683 : } else {
5684 0 : None
5685 : };
5686 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
5687 0 : shards.push(NodeShard {
5688 0 : tenant_shard_id: *tid,
5689 0 : is_intended_secondary,
5690 0 : is_observed_secondary,
5691 0 : });
5692 0 : }
5693 : }
5694 0 : Ok(NodeShardResponse { node_id, shards })
5695 0 : }
5696 :
5697 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
5698 0 : self.persistence.get_leader().await
5699 0 : }
5700 :
5701 0 : pub(crate) async fn node_register(
5702 0 : &self,
5703 0 : register_req: NodeRegisterRequest,
5704 0 : ) -> Result<(), ApiError> {
5705 0 : let _node_lock = trace_exclusive_lock(
5706 0 : &self.node_op_locks,
5707 0 : register_req.node_id,
5708 0 : NodeOperations::Register,
5709 0 : )
5710 0 : .await;
5711 :
5712 : enum RegistrationStatus {
5713 : Matched,
5714 : Mismatched,
5715 : New,
5716 : }
5717 :
5718 0 : let registration_status = {
5719 0 : let locked = self.inner.read().unwrap();
5720 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
5721 0 : if node.registration_match(®ister_req) {
5722 0 : RegistrationStatus::Matched
5723 : } else {
5724 0 : RegistrationStatus::Mismatched
5725 : }
5726 : } else {
5727 0 : RegistrationStatus::New
5728 : }
5729 : };
5730 :
5731 0 : match registration_status {
5732 : RegistrationStatus::Matched => {
5733 0 : tracing::info!(
5734 0 : "Node {} re-registered with matching address",
5735 : register_req.node_id
5736 : );
5737 :
5738 0 : return Ok(());
5739 : }
5740 : RegistrationStatus::Mismatched => {
5741 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
5742 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
5743 : // a fixed address through the lifetime of a node.
5744 0 : tracing::warn!(
5745 0 : "Node {} tried to register with different address",
5746 : register_req.node_id
5747 : );
5748 0 : return Err(ApiError::Conflict(
5749 0 : "Node is already registered with different address".to_string(),
5750 0 : ));
5751 : }
5752 0 : RegistrationStatus::New => {
5753 0 : // fallthrough
5754 0 : }
5755 0 : }
5756 0 :
5757 0 : // We do not require that a node is actually online when registered (it will start life
5758 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
5759 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
5760 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
5761 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
5762 0 : if tokio::net::lookup_host(format!(
5763 0 : "{}:{}",
5764 0 : register_req.listen_http_addr, register_req.listen_http_port
5765 0 : ))
5766 0 : .await
5767 0 : .is_err()
5768 : {
5769 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
5770 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
5771 : // we return a soft 503 error, to encourage callers to retry past transient issues.
5772 0 : return Err(ApiError::ResourceUnavailable(
5773 0 : format!(
5774 0 : "Node {} tried to register with unknown DNS name '{}'",
5775 0 : register_req.node_id, register_req.listen_http_addr
5776 0 : )
5777 0 : .into(),
5778 0 : ));
5779 0 : }
5780 0 :
5781 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
5782 0 : // This ensures that before we use it for anything or expose it via any external
5783 0 : // API, it is guaranteed to be available after a restart.
5784 0 : let new_node = Node::new(
5785 0 : register_req.node_id,
5786 0 : register_req.listen_http_addr,
5787 0 : register_req.listen_http_port,
5788 0 : register_req.listen_pg_addr,
5789 0 : register_req.listen_pg_port,
5790 0 : register_req.availability_zone_id.clone(),
5791 0 : );
5792 0 :
5793 0 : // TODO: idempotency if the node already exists in the database
5794 0 : self.persistence.insert_node(&new_node).await?;
5795 :
5796 0 : let mut locked = self.inner.write().unwrap();
5797 0 : let mut new_nodes = (*locked.nodes).clone();
5798 0 :
5799 0 : locked.scheduler.node_upsert(&new_node);
5800 0 : new_nodes.insert(register_req.node_id, new_node);
5801 0 :
5802 0 : locked.nodes = Arc::new(new_nodes);
5803 0 :
5804 0 : metrics::METRICS_REGISTRY
5805 0 : .metrics_group
5806 0 : .storage_controller_pageserver_nodes
5807 0 : .set(locked.nodes.len() as i64);
5808 0 :
5809 0 : tracing::info!(
5810 0 : "Registered pageserver {} ({}), now have {} pageservers",
5811 0 : register_req.node_id,
5812 0 : register_req.availability_zone_id,
5813 0 : locked.nodes.len()
5814 : );
5815 0 : Ok(())
5816 0 : }
5817 :
5818 : /// Configure in-memory and persistent state of a node as requested
5819 : ///
5820 : /// Note that this function does not trigger any immediate side effects in response
5821 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
5822 0 : async fn node_state_configure(
5823 0 : &self,
5824 0 : node_id: NodeId,
5825 0 : availability: Option<NodeAvailability>,
5826 0 : scheduling: Option<NodeSchedulingPolicy>,
5827 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
5828 0 : ) -> Result<AvailabilityTransition, ApiError> {
5829 0 : if let Some(scheduling) = scheduling {
5830 : // Scheduling is a persistent part of Node: we must write updates to the database before
5831 : // applying them in memory
5832 0 : self.persistence.update_node(node_id, scheduling).await?;
5833 0 : }
5834 :
5835 : // If we're activating a node, then before setting it active we must reconcile any shard locations
5836 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
5837 : // by calling [`Self::node_activate_reconcile`]
5838 : //
5839 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
5840 : // nothing else can mutate its availability while we run.
5841 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
5842 0 : let (activate_node, availability_transition) = {
5843 0 : let locked = self.inner.read().unwrap();
5844 0 : let Some(node) = locked.nodes.get(&node_id) else {
5845 0 : return Err(ApiError::NotFound(
5846 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5847 0 : ));
5848 : };
5849 :
5850 0 : (
5851 0 : node.clone(),
5852 0 : node.get_availability_transition(input_availability),
5853 0 : )
5854 : };
5855 :
5856 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
5857 0 : self.node_activate_reconcile(activate_node, node_lock)
5858 0 : .await?;
5859 0 : }
5860 0 : availability_transition
5861 : } else {
5862 0 : AvailabilityTransition::Unchanged
5863 : };
5864 :
5865 : // Apply changes from the request to our in-memory state for the Node
5866 0 : let mut locked = self.inner.write().unwrap();
5867 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
5868 0 :
5869 0 : let mut new_nodes = (**nodes).clone();
5870 :
5871 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
5872 0 : return Err(ApiError::NotFound(
5873 0 : anyhow::anyhow!("Node not registered").into(),
5874 0 : ));
5875 : };
5876 :
5877 0 : if let Some(availability) = availability {
5878 0 : node.set_availability(availability);
5879 0 : }
5880 :
5881 0 : if let Some(scheduling) = scheduling {
5882 0 : node.set_scheduling(scheduling);
5883 0 : }
5884 :
5885 : // Update the scheduler, in case the elegibility of the node for new shards has changed
5886 0 : scheduler.node_upsert(node);
5887 0 :
5888 0 : let new_nodes = Arc::new(new_nodes);
5889 0 : locked.nodes = new_nodes;
5890 0 :
5891 0 : Ok(availability_transition)
5892 0 : }
5893 :
5894 : /// Handle availability transition of one node
5895 : ///
5896 : /// Note that you should first call [`Self::node_state_configure`] to update
5897 : /// the in-memory state referencing that node. If you need to handle more than one transition
5898 : /// consider using [`Self::handle_node_availability_transitions`].
5899 0 : async fn handle_node_availability_transition(
5900 0 : &self,
5901 0 : node_id: NodeId,
5902 0 : transition: AvailabilityTransition,
5903 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
5904 0 : ) -> Result<(), ApiError> {
5905 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
5906 0 : match transition {
5907 : AvailabilityTransition::ToOffline => {
5908 0 : tracing::info!("Node {} transition to offline", node_id);
5909 :
5910 0 : let mut locked = self.inner.write().unwrap();
5911 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5912 0 :
5913 0 : let mut tenants_affected: usize = 0;
5914 :
5915 0 : for (_tenant_id, mut schedule_context, shards) in
5916 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
5917 : {
5918 0 : for tenant_shard in shards {
5919 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
5920 0 : if let Some(observed_loc) =
5921 0 : tenant_shard.observed.locations.get_mut(&node_id)
5922 0 : {
5923 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
5924 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
5925 0 : observed_loc.conf = None;
5926 0 : }
5927 :
5928 0 : if nodes.len() == 1 {
5929 : // Special case for single-node cluster: there is no point trying to reschedule
5930 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
5931 : // failures to schedule them.
5932 0 : continue;
5933 0 : }
5934 0 :
5935 0 : if !nodes
5936 0 : .values()
5937 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
5938 : {
5939 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
5940 : // trying to reschedule since there's nowhere else to go. Without this
5941 : // branch we incorrectly detach tenants in response to node unavailability.
5942 0 : continue;
5943 0 : }
5944 0 :
5945 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
5946 0 : tenant_shard.sequence = tenant_shard.sequence.next();
5947 0 :
5948 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
5949 0 : Err(e) => {
5950 0 : // It is possible that some tenants will become unschedulable when too many pageservers
5951 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
5952 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
5953 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
5954 : }
5955 : Ok(()) => {
5956 0 : if self.maybe_reconcile_shard(tenant_shard, nodes).is_some() {
5957 0 : tenants_affected += 1;
5958 0 : };
5959 : }
5960 : }
5961 0 : }
5962 : }
5963 : }
5964 0 : tracing::info!(
5965 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
5966 : tenants_affected,
5967 : node_id
5968 : )
5969 : }
5970 : AvailabilityTransition::ToActive => {
5971 0 : tracing::info!("Node {} transition to active", node_id);
5972 :
5973 0 : let mut locked = self.inner.write().unwrap();
5974 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
5975 :
5976 : // When a node comes back online, we must reconcile any tenant that has a None observed
5977 : // location on the node.
5978 0 : for tenant_shard in tenants.values_mut() {
5979 : // If a reconciliation is already in progress, rely on the previous scheduling
5980 : // decision and skip triggering a new reconciliation.
5981 0 : if tenant_shard.reconciler.is_some() {
5982 0 : continue;
5983 0 : }
5984 :
5985 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
5986 0 : if observed_loc.conf.is_none() {
5987 0 : self.maybe_reconcile_shard(tenant_shard, nodes);
5988 0 : }
5989 0 : }
5990 : }
5991 :
5992 : // TODO: in the background, we should balance work back onto this pageserver
5993 : }
5994 : // No action required for the intermediate unavailable state.
5995 : // When we transition into active or offline from the unavailable state,
5996 : // the correct handling above will kick in.
5997 : AvailabilityTransition::ToWarmingUpFromActive => {
5998 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
5999 : }
6000 : AvailabilityTransition::ToWarmingUpFromOffline => {
6001 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
6002 : }
6003 : AvailabilityTransition::Unchanged => {
6004 0 : tracing::debug!("Node {} no availability change during config", node_id);
6005 : }
6006 : }
6007 :
6008 0 : Ok(())
6009 0 : }
6010 :
6011 : /// Handle availability transition for multiple nodes
6012 : ///
6013 : /// Note that you should first call [`Self::node_state_configure`] for
6014 : /// all nodes being handled here for the handling to use fresh in-memory state.
6015 0 : async fn handle_node_availability_transitions(
6016 0 : &self,
6017 0 : transitions: Vec<(
6018 0 : NodeId,
6019 0 : TracingExclusiveGuard<NodeOperations>,
6020 0 : AvailabilityTransition,
6021 0 : )>,
6022 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
6023 0 : let mut errors = Vec::default();
6024 0 : for (node_id, node_lock, transition) in transitions {
6025 0 : let res = self
6026 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
6027 0 : .await;
6028 0 : if let Err(err) = res {
6029 0 : errors.push((node_id, err));
6030 0 : }
6031 : }
6032 :
6033 0 : if errors.is_empty() {
6034 0 : Ok(())
6035 : } else {
6036 0 : Err(errors)
6037 : }
6038 0 : }
6039 :
6040 0 : pub(crate) async fn node_configure(
6041 0 : &self,
6042 0 : node_id: NodeId,
6043 0 : availability: Option<NodeAvailability>,
6044 0 : scheduling: Option<NodeSchedulingPolicy>,
6045 0 : ) -> Result<(), ApiError> {
6046 0 : let node_lock =
6047 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
6048 :
6049 0 : let transition = self
6050 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
6051 0 : .await?;
6052 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
6053 0 : .await
6054 0 : }
6055 :
6056 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
6057 : /// operation for HTTP api.
6058 0 : pub(crate) async fn external_node_configure(
6059 0 : &self,
6060 0 : node_id: NodeId,
6061 0 : availability: Option<NodeAvailability>,
6062 0 : scheduling: Option<NodeSchedulingPolicy>,
6063 0 : ) -> Result<(), ApiError> {
6064 0 : {
6065 0 : let locked = self.inner.read().unwrap();
6066 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
6067 0 : return Err(ApiError::PreconditionFailed(
6068 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
6069 0 : ));
6070 0 : }
6071 0 : }
6072 0 :
6073 0 : self.node_configure(node_id, availability, scheduling).await
6074 0 : }
6075 :
6076 0 : pub(crate) async fn start_node_drain(
6077 0 : self: &Arc<Self>,
6078 0 : node_id: NodeId,
6079 0 : ) -> Result<(), ApiError> {
6080 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
6081 0 : let locked = self.inner.read().unwrap();
6082 0 : let nodes = &locked.nodes;
6083 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6084 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6085 0 : ))?;
6086 0 : let schedulable_nodes_count = nodes
6087 0 : .iter()
6088 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6089 0 : .count();
6090 0 :
6091 0 : (
6092 0 : locked
6093 0 : .ongoing_operation
6094 0 : .as_ref()
6095 0 : .map(|ongoing| ongoing.operation),
6096 0 : node.is_available(),
6097 0 : node.get_scheduling(),
6098 0 : schedulable_nodes_count,
6099 0 : )
6100 0 : };
6101 :
6102 0 : if let Some(ongoing) = ongoing_op {
6103 0 : return Err(ApiError::PreconditionFailed(
6104 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6105 0 : ));
6106 0 : }
6107 0 :
6108 0 : if !node_available {
6109 0 : return Err(ApiError::ResourceUnavailable(
6110 0 : format!("Node {node_id} is currently unavailable").into(),
6111 0 : ));
6112 0 : }
6113 0 :
6114 0 : if schedulable_nodes_count == 0 {
6115 0 : return Err(ApiError::PreconditionFailed(
6116 0 : "No other schedulable nodes to drain to".into(),
6117 0 : ));
6118 0 : }
6119 0 :
6120 0 : match node_policy {
6121 : NodeSchedulingPolicy::Active => {
6122 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
6123 0 : .await?;
6124 :
6125 0 : let cancel = self.cancel.child_token();
6126 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6127 :
6128 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6129 0 : operation: Operation::Drain(Drain { node_id }),
6130 0 : cancel: cancel.clone(),
6131 0 : });
6132 :
6133 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
6134 :
6135 0 : tokio::task::spawn({
6136 0 : let service = self.clone();
6137 0 : let cancel = cancel.clone();
6138 0 : async move {
6139 0 : let _gate_guard = gate_guard;
6140 0 :
6141 0 : scopeguard::defer! {
6142 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6143 0 :
6144 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
6145 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
6146 0 : } else {
6147 0 : panic!("We always remove the same operation")
6148 0 : }
6149 0 : }
6150 0 :
6151 0 : tracing::info!("Drain background operation starting");
6152 0 : let res = service.drain_node(node_id, cancel).await;
6153 0 : match res {
6154 : Ok(()) => {
6155 0 : tracing::info!("Drain background operation completed successfully");
6156 : }
6157 : Err(OperationError::Cancelled) => {
6158 0 : tracing::info!("Drain background operation was cancelled");
6159 : }
6160 0 : Err(err) => {
6161 0 : tracing::error!("Drain background operation encountered: {err}")
6162 : }
6163 : }
6164 0 : }
6165 0 : }.instrument(span));
6166 0 : }
6167 : NodeSchedulingPolicy::Draining => {
6168 0 : return Err(ApiError::Conflict(format!(
6169 0 : "Node {node_id} has drain in progress"
6170 0 : )));
6171 : }
6172 0 : policy => {
6173 0 : return Err(ApiError::PreconditionFailed(
6174 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
6175 0 : ));
6176 : }
6177 : }
6178 :
6179 0 : Ok(())
6180 0 : }
6181 :
6182 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
6183 0 : let node_available = {
6184 0 : let locked = self.inner.read().unwrap();
6185 0 : let nodes = &locked.nodes;
6186 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6187 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6188 0 : ))?;
6189 :
6190 0 : node.is_available()
6191 0 : };
6192 0 :
6193 0 : if !node_available {
6194 0 : return Err(ApiError::ResourceUnavailable(
6195 0 : format!("Node {node_id} is currently unavailable").into(),
6196 0 : ));
6197 0 : }
6198 :
6199 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6200 0 : if let Operation::Drain(drain) = op_handler.operation {
6201 0 : if drain.node_id == node_id {
6202 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6203 0 : op_handler.cancel.cancel();
6204 0 : return Ok(());
6205 0 : }
6206 0 : }
6207 0 : }
6208 :
6209 0 : Err(ApiError::PreconditionFailed(
6210 0 : format!("Node {node_id} has no drain in progress").into(),
6211 0 : ))
6212 0 : }
6213 :
6214 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
6215 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
6216 0 : let locked = self.inner.read().unwrap();
6217 0 : let nodes = &locked.nodes;
6218 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6219 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6220 0 : ))?;
6221 :
6222 0 : (
6223 0 : locked
6224 0 : .ongoing_operation
6225 0 : .as_ref()
6226 0 : .map(|ongoing| ongoing.operation),
6227 0 : node.is_available(),
6228 0 : node.get_scheduling(),
6229 0 : nodes.len(),
6230 0 : )
6231 0 : };
6232 :
6233 0 : if let Some(ongoing) = ongoing_op {
6234 0 : return Err(ApiError::PreconditionFailed(
6235 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6236 0 : ));
6237 0 : }
6238 0 :
6239 0 : if !node_available {
6240 0 : return Err(ApiError::ResourceUnavailable(
6241 0 : format!("Node {node_id} is currently unavailable").into(),
6242 0 : ));
6243 0 : }
6244 0 :
6245 0 : if total_nodes_count <= 1 {
6246 0 : return Err(ApiError::PreconditionFailed(
6247 0 : "No other nodes to fill from".into(),
6248 0 : ));
6249 0 : }
6250 0 :
6251 0 : match node_policy {
6252 : NodeSchedulingPolicy::Active => {
6253 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
6254 0 : .await?;
6255 :
6256 0 : let cancel = self.cancel.child_token();
6257 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6258 :
6259 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6260 0 : operation: Operation::Fill(Fill { node_id }),
6261 0 : cancel: cancel.clone(),
6262 0 : });
6263 :
6264 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
6265 :
6266 0 : tokio::task::spawn({
6267 0 : let service = self.clone();
6268 0 : let cancel = cancel.clone();
6269 0 : async move {
6270 0 : let _gate_guard = gate_guard;
6271 0 :
6272 0 : scopeguard::defer! {
6273 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6274 0 :
6275 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
6276 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
6277 0 : } else {
6278 0 : panic!("We always remove the same operation")
6279 0 : }
6280 0 : }
6281 0 :
6282 0 : tracing::info!("Fill background operation starting");
6283 0 : let res = service.fill_node(node_id, cancel).await;
6284 0 : match res {
6285 : Ok(()) => {
6286 0 : tracing::info!("Fill background operation completed successfully");
6287 : }
6288 : Err(OperationError::Cancelled) => {
6289 0 : tracing::info!("Fill background operation was cancelled");
6290 : }
6291 0 : Err(err) => {
6292 0 : tracing::error!("Fill background operation encountered: {err}")
6293 : }
6294 : }
6295 0 : }
6296 0 : }.instrument(span));
6297 0 : }
6298 : NodeSchedulingPolicy::Filling => {
6299 0 : return Err(ApiError::Conflict(format!(
6300 0 : "Node {node_id} has fill in progress"
6301 0 : )));
6302 : }
6303 0 : policy => {
6304 0 : return Err(ApiError::PreconditionFailed(
6305 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
6306 0 : ));
6307 : }
6308 : }
6309 :
6310 0 : Ok(())
6311 0 : }
6312 :
6313 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
6314 0 : let node_available = {
6315 0 : let locked = self.inner.read().unwrap();
6316 0 : let nodes = &locked.nodes;
6317 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6318 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6319 0 : ))?;
6320 :
6321 0 : node.is_available()
6322 0 : };
6323 0 :
6324 0 : if !node_available {
6325 0 : return Err(ApiError::ResourceUnavailable(
6326 0 : format!("Node {node_id} is currently unavailable").into(),
6327 0 : ));
6328 0 : }
6329 :
6330 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6331 0 : if let Operation::Fill(fill) = op_handler.operation {
6332 0 : if fill.node_id == node_id {
6333 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6334 0 : op_handler.cancel.cancel();
6335 0 : return Ok(());
6336 0 : }
6337 0 : }
6338 0 : }
6339 :
6340 0 : Err(ApiError::PreconditionFailed(
6341 0 : format!("Node {node_id} has no fill in progress").into(),
6342 0 : ))
6343 0 : }
6344 :
6345 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
6346 : /// configuration
6347 0 : fn maybe_reconcile_shard(
6348 0 : &self,
6349 0 : shard: &mut TenantShard,
6350 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6351 0 : ) -> Option<ReconcilerWaiter> {
6352 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::default())
6353 0 : }
6354 :
6355 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
6356 0 : fn maybe_configured_reconcile_shard(
6357 0 : &self,
6358 0 : shard: &mut TenantShard,
6359 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6360 0 : reconciler_config: ReconcilerConfig,
6361 0 : ) -> Option<ReconcilerWaiter> {
6362 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
6363 0 :
6364 0 : match reconcile_needed {
6365 0 : ReconcileNeeded::No => return None,
6366 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
6367 0 : ReconcileNeeded::Yes => {
6368 0 : // Fall through to try and acquire units for spawning reconciler
6369 0 : }
6370 : };
6371 :
6372 0 : let units = match self.reconciler_concurrency.clone().try_acquire_owned() {
6373 0 : Ok(u) => ReconcileUnits::new(u),
6374 : Err(_) => {
6375 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
6376 0 : "Concurrency limited: enqueued for reconcile later");
6377 0 : if !shard.delayed_reconcile {
6378 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
6379 0 : Err(TrySendError::Closed(_)) => {
6380 0 : // Weird mid-shutdown case?
6381 0 : }
6382 : Err(TrySendError::Full(_)) => {
6383 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
6384 0 : tracing::warn!(
6385 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
6386 : );
6387 : }
6388 0 : Ok(()) => {
6389 0 : shard.delayed_reconcile = true;
6390 0 : }
6391 : }
6392 0 : }
6393 :
6394 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
6395 : // number to advance. When this function is eventually called again and succeeds in getting units,
6396 : // it will spawn a reconciler that makes this waiter complete.
6397 0 : return Some(shard.future_reconcile_waiter());
6398 : }
6399 : };
6400 :
6401 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
6402 : // Gate closed: we're shutting down, drop out.
6403 0 : return None;
6404 : };
6405 :
6406 0 : shard.spawn_reconciler(
6407 0 : &self.result_tx,
6408 0 : nodes,
6409 0 : &self.compute_hook,
6410 0 : reconciler_config,
6411 0 : &self.config,
6412 0 : &self.persistence,
6413 0 : units,
6414 0 : gate_guard,
6415 0 : &self.reconcilers_cancel,
6416 0 : )
6417 0 : }
6418 :
6419 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
6420 : /// Additionally, reschedule tenants that require it.
6421 : ///
6422 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
6423 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
6424 : /// available. A return value of 0 indicates that everything is fully reconciled already.
6425 0 : fn reconcile_all(&self) -> usize {
6426 0 : let mut locked = self.inner.write().unwrap();
6427 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6428 0 : let pageservers = nodes.clone();
6429 0 :
6430 0 : // This function is an efficient place to update lazy statistics, since we are walking
6431 0 : // all tenants.
6432 0 : let mut pending_reconciles = 0;
6433 0 : let mut az_violations = 0;
6434 0 :
6435 0 : // If we find any tenants to drop from memory, stash them to offload after
6436 0 : // we're done traversing the map of tenants.
6437 0 : let mut drop_detached_tenants = Vec::new();
6438 0 :
6439 0 : let mut reconciles_spawned = 0;
6440 0 : for shard in tenants.values_mut() {
6441 : // Accumulate scheduling statistics
6442 0 : if let (Some(attached), Some(preferred)) =
6443 0 : (shard.intent.get_attached(), shard.preferred_az())
6444 : {
6445 0 : let node_az = nodes
6446 0 : .get(attached)
6447 0 : .expect("Nodes exist if referenced")
6448 0 : .get_availability_zone_id();
6449 0 : if node_az != preferred {
6450 0 : az_violations += 1;
6451 0 : }
6452 0 : }
6453 :
6454 : // Skip checking if this shard is already enqueued for reconciliation
6455 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
6456 : // If there is something delayed, then return a nonzero count so that
6457 : // callers like reconcile_all_now do not incorrectly get the impression
6458 : // that the system is in a quiescent state.
6459 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
6460 0 : pending_reconciles += 1;
6461 0 : continue;
6462 0 : }
6463 0 :
6464 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
6465 0 : // dirty, spawn another rone
6466 0 : if self.maybe_reconcile_shard(shard, &pageservers).is_some() {
6467 0 : reconciles_spawned += 1;
6468 0 : } else if shard.delayed_reconcile {
6469 0 : // Shard wanted to reconcile but for some reason couldn't.
6470 0 : pending_reconciles += 1;
6471 0 : }
6472 :
6473 : // If this tenant is detached, try dropping it from memory. This is usually done
6474 : // proactively in [`Self::process_results`], but we do it here to handle the edge
6475 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
6476 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
6477 0 : && shard.policy == PlacementPolicy::Detached
6478 : {
6479 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
6480 0 : shard.tenant_shard_id.tenant_id,
6481 0 : TenantOperations::DropDetached,
6482 0 : ) {
6483 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
6484 0 : }
6485 0 : }
6486 : }
6487 :
6488 : // Some metrics are calculated from SchedulerNode state, update these periodically
6489 0 : scheduler.update_metrics();
6490 :
6491 : // Process any deferred tenant drops
6492 0 : for (tenant_id, guard) in drop_detached_tenants {
6493 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
6494 0 : }
6495 :
6496 0 : metrics::METRICS_REGISTRY
6497 0 : .metrics_group
6498 0 : .storage_controller_schedule_az_violation
6499 0 : .set(az_violations as i64);
6500 0 :
6501 0 : metrics::METRICS_REGISTRY
6502 0 : .metrics_group
6503 0 : .storage_controller_pending_reconciles
6504 0 : .set(pending_reconciles as i64);
6505 0 :
6506 0 : reconciles_spawned
6507 0 : }
6508 :
6509 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
6510 : /// could be scheduled somewhere better:
6511 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
6512 : /// * e.g. after a node fails then recovers, to move some work back to it
6513 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
6514 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
6515 : /// we did the split, but are probably better placed elsewhere.
6516 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
6517 : /// * e.g. after a shard split, some locations will be on the same node (where the split
6518 : /// happened), and will probably be better placed elsewhere.
6519 : ///
6520 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
6521 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
6522 : /// according to those same soft constraints.
6523 0 : async fn optimize_all(&self) -> usize {
6524 : // Limit on how many shards' optmizations each call to this function will execute. Combined
6525 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
6526 : // trickle of optimizations in the background, rather than executing a large number in parallel
6527 : // when a change occurs.
6528 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
6529 :
6530 : // Synchronous prepare: scan shards for possible scheduling optimizations
6531 0 : let candidate_work = self.optimize_all_plan();
6532 0 : let candidate_work_len = candidate_work.len();
6533 :
6534 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
6535 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
6536 :
6537 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
6538 0 :
6539 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
6540 0 : let mut reconciles_spawned = 0;
6541 0 : let mut optimizations_applied = 0;
6542 0 : let mut locked = self.inner.write().unwrap();
6543 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6544 0 : for (tenant_shard_id, optimization) in validated_work {
6545 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6546 : // Shard was dropped between planning and execution;
6547 0 : continue;
6548 : };
6549 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
6550 0 : if shard.apply_optimization(scheduler, optimization) {
6551 0 : optimizations_applied += 1;
6552 0 : if self.maybe_reconcile_shard(shard, nodes).is_some() {
6553 0 : reconciles_spawned += 1;
6554 0 : }
6555 0 : }
6556 :
6557 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
6558 0 : break;
6559 0 : }
6560 : }
6561 :
6562 0 : if was_work_filtered {
6563 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
6564 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
6565 0 : // as these validations start passing.
6566 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
6567 0 : }
6568 :
6569 0 : reconciles_spawned
6570 0 : }
6571 :
6572 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
6573 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
6574 : // this higher than the execution limit gives us a chance to execute some work even if the first
6575 : // few optimizations we find are not ready.
6576 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
6577 :
6578 0 : let mut work = Vec::new();
6579 0 : let mut locked = self.inner.write().unwrap();
6580 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6581 :
6582 : // We are going to plan a bunch of optimisations before applying any of them, so the
6583 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
6584 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
6585 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
6586 : // to ignore the utilisation component of the score.
6587 :
6588 0 : for (_tenant_id, schedule_context, shards) in
6589 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
6590 : {
6591 0 : for shard in shards {
6592 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
6593 0 : break;
6594 0 : }
6595 0 : match shard.get_scheduling_policy() {
6596 0 : ShardSchedulingPolicy::Active => {
6597 0 : // Ok to do optimization
6598 0 : }
6599 : ShardSchedulingPolicy::Essential
6600 : | ShardSchedulingPolicy::Pause
6601 : | ShardSchedulingPolicy::Stop => {
6602 : // Policy prevents optimizing this shard.
6603 0 : continue;
6604 : }
6605 : }
6606 :
6607 0 : if !matches!(shard.splitting, SplitState::Idle)
6608 0 : || matches!(shard.policy, PlacementPolicy::Detached)
6609 0 : || shard.reconciler.is_some()
6610 : {
6611 : // Do not start any optimizations while another change to the tenant is ongoing: this
6612 : // is not necessary for correctness, but simplifies operations and implicitly throttles
6613 : // optimization changes to happen in a "trickle" over time.
6614 0 : continue;
6615 0 : }
6616 0 :
6617 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
6618 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
6619 0 : if cfg!(feature = "testing") {
6620 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
6621 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
6622 : // panic in prod if we hit this, or spend cycles on it in prod.
6623 0 : assert!(shard
6624 0 : .optimize_attachment(scheduler, &schedule_context)
6625 0 : .is_none());
6626 0 : assert!(shard
6627 0 : .optimize_secondary(scheduler, &schedule_context)
6628 0 : .is_none());
6629 0 : }
6630 0 : continue;
6631 0 : }
6632 :
6633 0 : if let Some(optimization) =
6634 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
6635 : // its primary location based on soft constraints, cut it over.
6636 0 : shard.optimize_attachment(scheduler, &schedule_context)
6637 : {
6638 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
6639 0 : work.push((shard.tenant_shard_id, optimization));
6640 0 : break;
6641 0 : } else if let Some(optimization) =
6642 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
6643 : // better placed on another node, based on ScheduleContext, then adjust it. This
6644 : // covers cases like after a shard split, where we might have too many shards
6645 : // in the same tenant with secondary locations on the node where they originally split.
6646 0 : shard.optimize_secondary(scheduler, &schedule_context)
6647 : {
6648 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
6649 0 : work.push((shard.tenant_shard_id, optimization));
6650 0 : break;
6651 0 : }
6652 : }
6653 : }
6654 :
6655 0 : work
6656 0 : }
6657 :
6658 0 : async fn optimize_all_validate(
6659 0 : &self,
6660 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
6661 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
6662 0 : // Take a clone of the node map to use outside the lock in async validation phase
6663 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
6664 0 :
6665 0 : let mut want_secondary_status = Vec::new();
6666 0 :
6667 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
6668 0 : // check that the state of locations is acceptable to run the optimization, such as
6669 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
6670 0 : // in a live migration.
6671 0 : let mut validated_work = Vec::new();
6672 0 : for (tenant_shard_id, optimization) in candidate_work {
6673 0 : match optimization.action {
6674 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
6675 : old_attached_node_id: _,
6676 0 : new_attached_node_id,
6677 0 : }) => {
6678 0 : match validation_nodes.get(&new_attached_node_id) {
6679 0 : None => {
6680 0 : // Node was dropped between planning and validation
6681 0 : }
6682 0 : Some(node) => {
6683 0 : if !node.is_available() {
6684 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
6685 0 : } else {
6686 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
6687 0 : // remote API requests concurrently.
6688 0 : want_secondary_status.push((
6689 0 : tenant_shard_id,
6690 0 : node.clone(),
6691 0 : optimization,
6692 0 : ));
6693 0 : }
6694 : }
6695 : }
6696 : }
6697 : ScheduleOptimizationAction::ReplaceSecondary(_)
6698 : | ScheduleOptimizationAction::CreateSecondary(_)
6699 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
6700 : // No extra checks needed to manage secondaries: this does not interrupt client access
6701 0 : validated_work.push((tenant_shard_id, optimization))
6702 : }
6703 : };
6704 : }
6705 :
6706 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
6707 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
6708 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
6709 0 : let results = self
6710 0 : .tenant_for_shards_api(
6711 0 : want_secondary_status
6712 0 : .iter()
6713 0 : .map(|i| (i.0, i.1.clone()))
6714 0 : .collect(),
6715 0 : |tenant_shard_id, client| async move {
6716 0 : client.tenant_secondary_status(tenant_shard_id).await
6717 0 : },
6718 0 : 1,
6719 0 : 1,
6720 0 : SHORT_RECONCILE_TIMEOUT,
6721 0 : &self.cancel,
6722 0 : )
6723 0 : .await;
6724 :
6725 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
6726 0 : want_secondary_status.into_iter().zip(results.into_iter())
6727 : {
6728 0 : match secondary_status {
6729 0 : Err(e) => {
6730 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
6731 : }
6732 0 : Ok(progress) => {
6733 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
6734 : // them in an optimization
6735 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
6736 :
6737 0 : if progress.heatmap_mtime.is_none()
6738 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
6739 0 : && progress.bytes_downloaded != progress.bytes_total
6740 0 : || progress.bytes_total - progress.bytes_downloaded
6741 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
6742 : {
6743 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
6744 :
6745 : #[cfg(feature = "testing")]
6746 0 : if progress.heatmap_mtime.is_none() {
6747 : // No heatmap might mean the attached location has never uploaded one, or that
6748 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
6749 : // but fairly common in tests.
6750 0 : self.kick_secondary_download(tenant_shard_id).await;
6751 0 : }
6752 : } else {
6753 : // Location looks ready: proceed
6754 0 : tracing::info!(
6755 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
6756 : );
6757 0 : validated_work.push((tenant_shard_id, optimization))
6758 : }
6759 : }
6760 : }
6761 : }
6762 :
6763 0 : validated_work
6764 0 : }
6765 :
6766 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
6767 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
6768 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
6769 : /// we have this helper to move things along faster.
6770 : #[cfg(feature = "testing")]
6771 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
6772 0 : let (attached_node, secondaries) = {
6773 0 : let locked = self.inner.read().unwrap();
6774 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6775 0 : tracing::warn!(
6776 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
6777 : );
6778 0 : return;
6779 : };
6780 :
6781 0 : let Some(attached) = shard.intent.get_attached() else {
6782 0 : tracing::warn!(
6783 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
6784 : );
6785 0 : return;
6786 : };
6787 :
6788 0 : let secondaries = shard
6789 0 : .intent
6790 0 : .get_secondary()
6791 0 : .iter()
6792 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
6793 0 : .collect::<Vec<_>>();
6794 0 :
6795 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
6796 0 : };
6797 0 :
6798 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
6799 0 : // a 'kick' to let scheduling optimisation run more promptly.
6800 0 : match attached_node
6801 0 : .with_client_retries(
6802 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
6803 0 : &self.config.jwt_token,
6804 0 : 3,
6805 0 : 10,
6806 0 : SHORT_RECONCILE_TIMEOUT,
6807 0 : &self.cancel,
6808 0 : )
6809 0 : .await
6810 : {
6811 0 : Some(Err(e)) => {
6812 0 : tracing::info!(
6813 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
6814 : );
6815 : }
6816 : None => {
6817 0 : tracing::info!(
6818 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
6819 : );
6820 : }
6821 : Some(Ok(_)) => {
6822 0 : tracing::info!(
6823 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
6824 : );
6825 : }
6826 : }
6827 :
6828 0 : for secondary_node in secondaries {
6829 0 : match secondary_node
6830 0 : .with_client_retries(
6831 0 : |client| async move {
6832 0 : client
6833 0 : .tenant_secondary_download(
6834 0 : tenant_shard_id,
6835 0 : Some(Duration::from_secs(1)),
6836 0 : )
6837 0 : .await
6838 0 : },
6839 0 : &self.config.jwt_token,
6840 0 : 3,
6841 0 : 10,
6842 0 : SHORT_RECONCILE_TIMEOUT,
6843 0 : &self.cancel,
6844 0 : )
6845 0 : .await
6846 : {
6847 0 : Some(Err(e)) => {
6848 0 : tracing::info!(
6849 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
6850 : );
6851 : }
6852 : None => {
6853 0 : tracing::info!("Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}");
6854 : }
6855 0 : Some(Ok(progress)) => {
6856 0 : tracing::info!("Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}");
6857 : }
6858 : }
6859 : }
6860 0 : }
6861 :
6862 : /// Look for shards which are oversized and in need of splitting
6863 0 : async fn autosplit_tenants(self: &Arc<Self>) {
6864 0 : let Some(split_threshold) = self.config.split_threshold else {
6865 : // Auto-splitting is disabled
6866 0 : return;
6867 : };
6868 :
6869 0 : let nodes = self.inner.read().unwrap().nodes.clone();
6870 :
6871 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
6872 :
6873 0 : let mut top_n = Vec::new();
6874 0 :
6875 0 : // Call into each node to look for big tenants
6876 0 : let top_n_request = TopTenantShardsRequest {
6877 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
6878 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
6879 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
6880 0 : // disk space impact of one shard.
6881 0 : order_by: models::TenantSorting::MaxLogicalSize,
6882 0 : limit: 10,
6883 0 : where_shards_lt: Some(SPLIT_TO_MAX),
6884 0 : where_gt: Some(split_threshold),
6885 0 : };
6886 0 : for node in nodes.values() {
6887 0 : let request_ref = &top_n_request;
6888 0 : match node
6889 0 : .with_client_retries(
6890 0 : |client| async move {
6891 0 : let request = request_ref.clone();
6892 0 : client.top_tenant_shards(request.clone()).await
6893 0 : },
6894 0 : &self.config.jwt_token,
6895 0 : 3,
6896 0 : 3,
6897 0 : Duration::from_secs(5),
6898 0 : &self.cancel,
6899 0 : )
6900 0 : .await
6901 : {
6902 0 : Some(Ok(node_top_n)) => {
6903 0 : top_n.extend(node_top_n.shards.into_iter());
6904 0 : }
6905 : Some(Err(mgmt_api::Error::Cancelled)) => {
6906 0 : continue;
6907 : }
6908 0 : Some(Err(e)) => {
6909 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
6910 0 : continue;
6911 : }
6912 : None => {
6913 : // Node is shutting down
6914 0 : continue;
6915 : }
6916 : };
6917 : }
6918 :
6919 : // Pick the biggest tenant to split first
6920 0 : top_n.sort_by_key(|i| i.resident_size);
6921 0 :
6922 0 : // Filter out tenants in a prohibiting scheduling mode
6923 0 : {
6924 0 : let locked = self.inner.read().unwrap();
6925 0 : top_n.retain(|i| {
6926 0 : if let Some(shard) = locked.tenants.get(&i.id) {
6927 0 : matches!(shard.get_scheduling_policy(), ShardSchedulingPolicy::Active)
6928 : } else {
6929 0 : false
6930 : }
6931 0 : });
6932 0 : }
6933 :
6934 0 : let Some(split_candidate) = top_n.into_iter().next() else {
6935 0 : tracing::debug!("No split-elegible shards found");
6936 0 : return;
6937 : };
6938 :
6939 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
6940 : // want to block the background reconcile loop on this.
6941 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
6942 :
6943 0 : let this = self.clone();
6944 0 : tokio::spawn(
6945 0 : async move {
6946 0 : match this
6947 0 : .tenant_shard_split(
6948 0 : split_candidate.id.tenant_id,
6949 0 : TenantShardSplitRequest {
6950 0 : // Always split to the max number of shards: this avoids stepping through
6951 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
6952 0 : // each time as a tenant grows, and is not too expensive because our max shard
6953 0 : // count is relatively low anyway.
6954 0 : // This policy will be adjusted in future once we support higher shard count.
6955 0 : new_shard_count: SPLIT_TO_MAX.literal(),
6956 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
6957 0 : },
6958 0 : )
6959 0 : .await
6960 : {
6961 : Ok(_) => {
6962 0 : tracing::info!("Successful auto-split");
6963 : }
6964 0 : Err(e) => {
6965 0 : tracing::error!("Auto-split failed: {e}");
6966 : }
6967 : }
6968 0 : }
6969 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
6970 : );
6971 0 : }
6972 :
6973 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
6974 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
6975 : /// put the system into a quiescent state where future background reconciliations won't do anything.
6976 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
6977 0 : let reconciles_spawned = self.reconcile_all();
6978 0 : let reconciles_spawned = if reconciles_spawned == 0 {
6979 : // Only optimize when we are otherwise idle
6980 0 : self.optimize_all().await
6981 : } else {
6982 0 : reconciles_spawned
6983 : };
6984 :
6985 0 : let waiters = {
6986 0 : let mut waiters = Vec::new();
6987 0 : let locked = self.inner.read().unwrap();
6988 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
6989 0 : if let Some(waiter) = shard.get_waiter() {
6990 0 : waiters.push(waiter);
6991 0 : }
6992 : }
6993 0 : waiters
6994 0 : };
6995 0 :
6996 0 : let waiter_count = waiters.len();
6997 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
6998 0 : Ok(()) => {}
6999 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
7000 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
7001 0 : {
7002 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
7003 0 : // because some other change superceded it. We will return a nonzero number,
7004 0 : // so the caller knows they might have to call again to quiesce the system.
7005 0 : }
7006 0 : Err(e) => {
7007 0 : return Err(e);
7008 : }
7009 : };
7010 :
7011 0 : tracing::info!(
7012 0 : "{} reconciles in reconcile_all, {} waiters",
7013 : reconciles_spawned,
7014 : waiter_count
7015 : );
7016 :
7017 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
7018 0 : }
7019 :
7020 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
7021 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
7022 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
7023 0 : self.reconcilers_cancel.cancel();
7024 0 : self.reconcilers_gate.close().await;
7025 :
7026 : // Signal the background loop in [`Service::process_results`] to exit once
7027 : // it has proccessed the results from all the reconciles we cancelled earlier.
7028 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
7029 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
7030 0 : self.result_tx.closed().await;
7031 0 : }
7032 :
7033 0 : pub async fn shutdown(&self) {
7034 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
7035 0 : .await;
7036 :
7037 : // Background tasks hold gate guards: this notifies them of the cancellation and
7038 : // waits for them all to complete.
7039 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
7040 0 : self.cancel.cancel();
7041 0 : self.gate.close().await;
7042 0 : }
7043 :
7044 : /// Spot check the download lag for a secondary location of a shard.
7045 : /// Should be used as a heuristic, since it's not always precise: the
7046 : /// secondary might have not downloaded the new heat map yet and, hence,
7047 : /// is not aware of the lag.
7048 : ///
7049 : /// Returns:
7050 : /// * Ok(None) if the lag could not be determined from the status,
7051 : /// * Ok(Some(_)) if the lag could be determind
7052 : /// * Err on failures to query the pageserver.
7053 0 : async fn secondary_lag(
7054 0 : &self,
7055 0 : secondary: &NodeId,
7056 0 : tenant_shard_id: TenantShardId,
7057 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
7058 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7059 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
7060 0 : StatusCode::NOT_FOUND,
7061 0 : format!("Node with id {} not found", secondary),
7062 0 : ))?;
7063 :
7064 0 : match node
7065 0 : .with_client_retries(
7066 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
7067 0 : &self.config.jwt_token,
7068 0 : 1,
7069 0 : 3,
7070 0 : Duration::from_millis(250),
7071 0 : &self.cancel,
7072 0 : )
7073 0 : .await
7074 : {
7075 0 : Some(Ok(status)) => match status.heatmap_mtime {
7076 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
7077 0 : None => Ok(None),
7078 : },
7079 0 : Some(Err(e)) => Err(e),
7080 0 : None => Err(mgmt_api::Error::Cancelled),
7081 : }
7082 0 : }
7083 :
7084 : /// Drain a node by moving the shards attached to it as primaries.
7085 : /// This is a long running operation and it should run as a separate Tokio task.
7086 0 : pub(crate) async fn drain_node(
7087 0 : self: &Arc<Self>,
7088 0 : node_id: NodeId,
7089 0 : cancel: CancellationToken,
7090 0 : ) -> Result<(), OperationError> {
7091 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
7092 0 : let max_secondary_lag_bytes = self
7093 0 : .config
7094 0 : .max_secondary_lag_bytes
7095 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
7096 :
7097 : // By default, live migrations are generous about the wait time for getting
7098 : // the secondary location up to speed. When draining, give up earlier in order
7099 : // to not stall the operation when a cold secondary is encountered.
7100 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7101 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7102 0 : let reconciler_config = ReconcilerConfigBuilder::new()
7103 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7104 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7105 0 : .build();
7106 0 :
7107 0 : let mut waiters = Vec::new();
7108 0 :
7109 0 : let mut tid_iter = TenantShardIterator::new({
7110 0 : let service = self.clone();
7111 0 : move |last_inspected_shard: Option<TenantShardId>| {
7112 0 : let locked = &service.inner.read().unwrap();
7113 0 : let tenants = &locked.tenants;
7114 0 : let entry = match last_inspected_shard {
7115 0 : Some(skip_past) => {
7116 0 : // Skip to the last seen tenant shard id
7117 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
7118 0 :
7119 0 : // Skip past the last seen
7120 0 : cursor.nth(1)
7121 : }
7122 0 : None => tenants.first_key_value(),
7123 : };
7124 :
7125 0 : entry.map(|(tid, _)| tid).copied()
7126 0 : }
7127 0 : });
7128 :
7129 0 : while !tid_iter.finished() {
7130 0 : if cancel.is_cancelled() {
7131 0 : match self
7132 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7133 0 : .await
7134 : {
7135 0 : Ok(()) => return Err(OperationError::Cancelled),
7136 0 : Err(err) => {
7137 0 : return Err(OperationError::FinalizeError(
7138 0 : format!(
7139 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7140 0 : node_id, err
7141 0 : )
7142 0 : .into(),
7143 0 : ));
7144 : }
7145 : }
7146 0 : }
7147 0 :
7148 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
7149 :
7150 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7151 0 : let tid = match tid_iter.next() {
7152 0 : Some(tid) => tid,
7153 : None => {
7154 0 : break;
7155 : }
7156 : };
7157 :
7158 0 : let tid_drain = TenantShardDrain {
7159 0 : drained_node: node_id,
7160 0 : tenant_shard_id: tid,
7161 0 : };
7162 :
7163 0 : let dest_node_id = {
7164 0 : let locked = self.inner.read().unwrap();
7165 0 :
7166 0 : match tid_drain
7167 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
7168 : {
7169 0 : Some(node_id) => node_id,
7170 : None => {
7171 0 : continue;
7172 : }
7173 : }
7174 : };
7175 :
7176 0 : match self.secondary_lag(&dest_node_id, tid).await {
7177 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
7178 0 : // The secondary is reasonably up to date.
7179 0 : // Migrate to it
7180 0 : }
7181 0 : Ok(Some(lag)) => {
7182 0 : tracing::info!(
7183 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7184 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
7185 : );
7186 0 : continue;
7187 : }
7188 : Ok(None) => {
7189 0 : tracing::info!(
7190 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7191 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
7192 : );
7193 0 : continue;
7194 : }
7195 0 : Err(err) => {
7196 0 : tracing::warn!(
7197 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7198 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
7199 : );
7200 0 : continue;
7201 : }
7202 : }
7203 :
7204 : {
7205 0 : let mut locked = self.inner.write().unwrap();
7206 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7207 0 : let rescheduled = tid_drain.reschedule_to_secondary(
7208 0 : dest_node_id,
7209 0 : tenants,
7210 0 : scheduler,
7211 0 : nodes,
7212 0 : )?;
7213 :
7214 0 : if let Some(tenant_shard) = rescheduled {
7215 0 : let waiter = self.maybe_configured_reconcile_shard(
7216 0 : tenant_shard,
7217 0 : nodes,
7218 0 : reconciler_config,
7219 0 : );
7220 0 : if let Some(some) = waiter {
7221 0 : waiters.push(some);
7222 0 : }
7223 0 : }
7224 : }
7225 : }
7226 :
7227 0 : waiters = self
7228 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7229 0 : .await;
7230 :
7231 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
7232 : }
7233 :
7234 0 : while !waiters.is_empty() {
7235 0 : if cancel.is_cancelled() {
7236 0 : match self
7237 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7238 0 : .await
7239 : {
7240 0 : Ok(()) => return Err(OperationError::Cancelled),
7241 0 : Err(err) => {
7242 0 : return Err(OperationError::FinalizeError(
7243 0 : format!(
7244 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7245 0 : node_id, err
7246 0 : )
7247 0 : .into(),
7248 0 : ));
7249 : }
7250 : }
7251 0 : }
7252 0 :
7253 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
7254 :
7255 0 : waiters = self
7256 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7257 0 : .await;
7258 : }
7259 :
7260 : // At this point we have done the best we could to drain shards from this node.
7261 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
7262 : // to complete the drain.
7263 0 : if let Err(err) = self
7264 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
7265 0 : .await
7266 : {
7267 : // This is not fatal. Anything that is polling the node scheduling policy to detect
7268 : // the end of the drain operations will hang, but all such places should enforce an
7269 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
7270 : // by the counterpart fill operation.
7271 0 : return Err(OperationError::FinalizeError(
7272 0 : format!(
7273 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
7274 0 : )
7275 0 : .into(),
7276 0 : ));
7277 0 : }
7278 0 :
7279 0 : Ok(())
7280 0 : }
7281 :
7282 : /// Create a node fill plan (pick secondaries to promote), based on:
7283 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
7284 : /// outside their home AZ, should be migrated back here.
7285 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
7286 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
7287 : /// those shards have a home AZ that doesn't match the node we're filling.
7288 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
7289 0 : let mut locked = self.inner.write().unwrap();
7290 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7291 0 :
7292 0 : let node_az = nodes
7293 0 : .get(&node_id)
7294 0 : .expect("Node must exist")
7295 0 : .get_availability_zone_id()
7296 0 : .clone();
7297 0 :
7298 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
7299 0 : let mut plan = Vec::new();
7300 0 :
7301 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
7302 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
7303 0 :
7304 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
7305 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
7306 0 : let respect_azs = nodes
7307 0 : .values()
7308 0 : .map(|n| n.get_availability_zone_id())
7309 0 : .unique()
7310 0 : .count()
7311 0 : > 1;
7312 :
7313 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
7314 : // requires it.
7315 0 : for (tsid, tenant_shard) in tenants {
7316 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7317 : // Shard doesn't have a secondary on this node, ignore it.
7318 0 : continue;
7319 0 : }
7320 0 :
7321 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
7322 0 : // shards which belong on this node, not to promote shards whose scheduling preference
7323 0 : // would be on their currently attached node. So will avoid promoting shards whose
7324 0 : // home AZ doesn't match the AZ of the node we're filling.
7325 0 : match tenant_shard.preferred_az() {
7326 0 : _ if !respect_azs => {
7327 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7328 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7329 0 : }
7330 : }
7331 : None => {
7332 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
7333 : // will only do so if our target shard count requires it.
7334 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7335 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7336 0 : }
7337 : }
7338 0 : Some(az) if az == &node_az => {
7339 : // This shard's home AZ is equal to the node we're filling: it should
7340 : // be moved back to this node as part of filling, unless its currently
7341 : // attached location is also in its home AZ.
7342 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7343 0 : if nodes
7344 0 : .get(primary)
7345 0 : .expect("referenced node must exist")
7346 0 : .get_availability_zone_id()
7347 0 : != tenant_shard
7348 0 : .preferred_az()
7349 0 : .expect("tenant must have an AZ preference")
7350 : {
7351 0 : plan.push(*tsid)
7352 0 : }
7353 : } else {
7354 0 : plan.push(*tsid)
7355 : }
7356 : }
7357 0 : Some(_) => {
7358 0 : // This shard's home AZ is somewhere other than the node we're filling,
7359 0 : // it may not be moved back to this node as part of filling. Ignore it
7360 0 : }
7361 : }
7362 : }
7363 :
7364 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
7365 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
7366 0 :
7367 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
7368 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
7369 0 :
7370 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
7371 :
7372 0 : for (node_id, attached) in nodes_by_load {
7373 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
7374 0 : if !available {
7375 0 : continue;
7376 0 : }
7377 0 :
7378 0 : if plan.len() >= fill_requirement
7379 0 : || free_tids_by_node.is_empty()
7380 0 : || attached <= expected_attached
7381 : {
7382 0 : break;
7383 0 : }
7384 0 :
7385 0 : let can_take = attached - expected_attached;
7386 0 : let needed = fill_requirement - plan.len();
7387 0 : let mut take = std::cmp::min(can_take, needed);
7388 0 :
7389 0 : let mut remove_node = false;
7390 0 : while take > 0 {
7391 0 : match free_tids_by_node.get_mut(&node_id) {
7392 0 : Some(tids) => match tids.pop() {
7393 0 : Some(tid) => {
7394 0 : let max_promote_for_tenant = std::cmp::max(
7395 0 : tid.shard_count.count() as usize / locked.nodes.len(),
7396 0 : 1,
7397 0 : );
7398 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
7399 0 : if *promoted < max_promote_for_tenant {
7400 0 : plan.push(tid);
7401 0 : *promoted += 1;
7402 0 : take -= 1;
7403 0 : }
7404 : }
7405 : None => {
7406 0 : remove_node = true;
7407 0 : break;
7408 : }
7409 : },
7410 : None => {
7411 0 : break;
7412 : }
7413 : }
7414 : }
7415 :
7416 0 : if remove_node {
7417 0 : free_tids_by_node.remove(&node_id);
7418 0 : }
7419 : }
7420 :
7421 0 : plan
7422 0 : }
7423 :
7424 : /// Fill a node by promoting its secondaries until the cluster is balanced
7425 : /// with regards to attached shard counts. Note that this operation only
7426 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
7427 : /// This is a long running operation and it should run as a separate Tokio task.
7428 0 : pub(crate) async fn fill_node(
7429 0 : &self,
7430 0 : node_id: NodeId,
7431 0 : cancel: CancellationToken,
7432 0 : ) -> Result<(), OperationError> {
7433 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7434 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7435 0 : let reconciler_config = ReconcilerConfigBuilder::new()
7436 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7437 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7438 0 : .build();
7439 0 :
7440 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
7441 0 : let mut waiters = Vec::new();
7442 :
7443 : // Execute the plan we've composed above. Before aplying each move from the plan,
7444 : // we validate to ensure that it has not gone stale in the meantime.
7445 0 : while !tids_to_promote.is_empty() {
7446 0 : if cancel.is_cancelled() {
7447 0 : match self
7448 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7449 0 : .await
7450 : {
7451 0 : Ok(()) => return Err(OperationError::Cancelled),
7452 0 : Err(err) => {
7453 0 : return Err(OperationError::FinalizeError(
7454 0 : format!(
7455 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7456 0 : node_id, err
7457 0 : )
7458 0 : .into(),
7459 0 : ));
7460 : }
7461 : }
7462 0 : }
7463 0 :
7464 0 : {
7465 0 : let mut locked = self.inner.write().unwrap();
7466 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7467 :
7468 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
7469 0 : format!("node {node_id} was removed").into(),
7470 0 : ))?;
7471 :
7472 0 : let current_policy = node.get_scheduling();
7473 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
7474 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
7475 : // about it
7476 0 : return Err(OperationError::NodeStateChanged(
7477 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
7478 0 : ));
7479 0 : }
7480 :
7481 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7482 0 : if let Some(tid) = tids_to_promote.pop() {
7483 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
7484 : // If the node being filled is not a secondary anymore,
7485 : // skip the promotion.
7486 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7487 0 : continue;
7488 0 : }
7489 0 :
7490 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
7491 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
7492 0 : Err(e) => {
7493 0 : tracing::warn!(
7494 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7495 0 : "Scheduling error when filling pageserver {} : {e}", node_id
7496 : );
7497 : }
7498 : Ok(()) => {
7499 0 : tracing::info!(
7500 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7501 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
7502 : node_id,
7503 : previously_attached_to,
7504 : node_id
7505 : );
7506 :
7507 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
7508 0 : tenant_shard,
7509 0 : nodes,
7510 0 : reconciler_config,
7511 0 : ) {
7512 0 : waiters.push(waiter);
7513 0 : }
7514 : }
7515 : }
7516 0 : }
7517 : } else {
7518 0 : break;
7519 : }
7520 : }
7521 : }
7522 :
7523 0 : waiters = self
7524 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7525 0 : .await;
7526 : }
7527 :
7528 0 : while !waiters.is_empty() {
7529 0 : if cancel.is_cancelled() {
7530 0 : match self
7531 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7532 0 : .await
7533 : {
7534 0 : Ok(()) => return Err(OperationError::Cancelled),
7535 0 : Err(err) => {
7536 0 : return Err(OperationError::FinalizeError(
7537 0 : format!(
7538 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7539 0 : node_id, err
7540 0 : )
7541 0 : .into(),
7542 0 : ));
7543 : }
7544 : }
7545 0 : }
7546 0 :
7547 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
7548 :
7549 0 : waiters = self
7550 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7551 0 : .await;
7552 : }
7553 :
7554 0 : if let Err(err) = self
7555 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7556 0 : .await
7557 : {
7558 : // This isn't a huge issue since the filling process starts upon request. However, it
7559 : // will prevent the next drain from starting. The only case in which this can fail
7560 : // is database unavailability. Such a case will require manual intervention.
7561 0 : return Err(OperationError::FinalizeError(
7562 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
7563 0 : .into(),
7564 0 : ));
7565 0 : }
7566 0 :
7567 0 : Ok(())
7568 0 : }
7569 :
7570 : /// Updates scrubber metadata health check results.
7571 0 : pub(crate) async fn metadata_health_update(
7572 0 : &self,
7573 0 : update_req: MetadataHealthUpdateRequest,
7574 0 : ) -> Result<(), ApiError> {
7575 0 : let now = chrono::offset::Utc::now();
7576 0 : let (healthy_records, unhealthy_records) = {
7577 0 : let locked = self.inner.read().unwrap();
7578 0 : let healthy_records = update_req
7579 0 : .healthy_tenant_shards
7580 0 : .into_iter()
7581 0 : // Retain only health records associated with tenant shards managed by storage controller.
7582 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7583 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
7584 0 : .collect();
7585 0 : let unhealthy_records = update_req
7586 0 : .unhealthy_tenant_shards
7587 0 : .into_iter()
7588 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7589 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
7590 0 : .collect();
7591 0 :
7592 0 : (healthy_records, unhealthy_records)
7593 0 : };
7594 0 :
7595 0 : self.persistence
7596 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
7597 0 : .await?;
7598 0 : Ok(())
7599 0 : }
7600 :
7601 : /// Lists the tenant shards that has unhealthy metadata status.
7602 0 : pub(crate) async fn metadata_health_list_unhealthy(
7603 0 : &self,
7604 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
7605 0 : let result = self
7606 0 : .persistence
7607 0 : .list_unhealthy_metadata_health_records()
7608 0 : .await?
7609 0 : .iter()
7610 0 : .map(|p| p.get_tenant_shard_id().unwrap())
7611 0 : .collect();
7612 0 :
7613 0 : Ok(result)
7614 0 : }
7615 :
7616 : /// Lists the tenant shards that have not been scrubbed for some duration.
7617 0 : pub(crate) async fn metadata_health_list_outdated(
7618 0 : &self,
7619 0 : not_scrubbed_for: Duration,
7620 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
7621 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
7622 0 : let result = self
7623 0 : .persistence
7624 0 : .list_outdated_metadata_health_records(earlier)
7625 0 : .await?
7626 0 : .into_iter()
7627 0 : .map(|record| record.into())
7628 0 : .collect();
7629 0 : Ok(result)
7630 0 : }
7631 :
7632 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
7633 0 : self.inner.read().unwrap().get_leadership_status()
7634 0 : }
7635 :
7636 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
7637 0 : tracing::info!("Received step down request from peer");
7638 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
7639 :
7640 0 : self.inner.write().unwrap().step_down();
7641 0 : // TODO: would it make sense to have a time-out for this?
7642 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
7643 0 : .await;
7644 :
7645 0 : let mut global_observed = GlobalObservedState::default();
7646 0 : let locked = self.inner.read().unwrap();
7647 0 : for (tid, tenant_shard) in locked.tenants.iter() {
7648 0 : global_observed
7649 0 : .0
7650 0 : .insert(*tid, tenant_shard.observed.clone());
7651 0 : }
7652 :
7653 0 : global_observed
7654 0 : }
7655 :
7656 0 : pub(crate) async fn safekeepers_list(
7657 0 : &self,
7658 0 : ) -> Result<Vec<SafekeeperDescribeResponse>, DatabaseError> {
7659 0 : self.persistence
7660 0 : .list_safekeepers()
7661 0 : .await?
7662 0 : .into_iter()
7663 0 : .map(|v| v.as_describe_response())
7664 0 : .collect::<Result<Vec<_>, _>>()
7665 0 : }
7666 :
7667 0 : pub(crate) async fn get_safekeeper(
7668 0 : &self,
7669 0 : id: i64,
7670 0 : ) -> Result<SafekeeperDescribeResponse, DatabaseError> {
7671 0 : self.persistence
7672 0 : .safekeeper_get(id)
7673 0 : .await
7674 0 : .and_then(|v| v.as_describe_response())
7675 0 : }
7676 :
7677 0 : pub(crate) async fn upsert_safekeeper(
7678 0 : &self,
7679 0 : record: crate::persistence::SafekeeperUpsert,
7680 0 : ) -> Result<(), DatabaseError> {
7681 0 : self.persistence.safekeeper_upsert(record).await
7682 0 : }
7683 :
7684 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
7685 0 : &self,
7686 0 : id: i64,
7687 0 : scheduling_policy: SkSchedulingPolicy,
7688 0 : ) -> Result<(), DatabaseError> {
7689 0 : self.persistence
7690 0 : .set_safekeeper_scheduling_policy(id, scheduling_policy)
7691 0 : .await
7692 0 : }
7693 :
7694 0 : pub(crate) async fn update_shards_preferred_azs(
7695 0 : &self,
7696 0 : req: ShardsPreferredAzsRequest,
7697 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
7698 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
7699 0 : let updated = self
7700 0 : .persistence
7701 0 : .set_tenant_shard_preferred_azs(preferred_azs)
7702 0 : .await
7703 0 : .map_err(|err| {
7704 0 : ApiError::InternalServerError(anyhow::anyhow!(
7705 0 : "Failed to persist preferred AZs: {err}"
7706 0 : ))
7707 0 : })?;
7708 :
7709 0 : let mut updated_in_mem_and_db = Vec::default();
7710 0 :
7711 0 : let mut locked = self.inner.write().unwrap();
7712 0 : for (tid, az_id) in updated {
7713 0 : let shard = locked.tenants.get_mut(&tid);
7714 0 : if let Some(shard) = shard {
7715 0 : shard.set_preferred_az(az_id);
7716 0 : updated_in_mem_and_db.push(tid);
7717 0 : }
7718 : }
7719 :
7720 0 : Ok(ShardsPreferredAzsResponse {
7721 0 : updated: updated_in_mem_and_db,
7722 0 : })
7723 0 : }
7724 : }
|
