Line data Source code
1 : pub mod chaos_injector;
2 : mod context_iterator;
3 :
4 : use std::borrow::Cow;
5 : use std::cmp::Ordering;
6 : use std::collections::{BTreeMap, HashMap, HashSet};
7 : use std::error::Error;
8 : use std::ops::Deref;
9 : use std::path::PathBuf;
10 : use std::str::FromStr;
11 : use std::sync::Arc;
12 : use std::time::{Duration, Instant};
13 :
14 : use anyhow::Context;
15 : use context_iterator::TenantShardContextIterator;
16 : use control_plane::storage_controller::{
17 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
18 : };
19 : use diesel::result::DatabaseErrorKind;
20 : use futures::StreamExt;
21 : use futures::stream::FuturesUnordered;
22 : use http_utils::error::ApiError;
23 : use hyper::Uri;
24 : use itertools::Itertools;
25 : use pageserver_api::controller_api::{
26 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
27 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
28 : SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest,
29 : ShardsPreferredAzsResponse, SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse,
30 : TenantCreateResponseShard, TenantDescribeResponse, TenantDescribeResponseShard,
31 : TenantLocateResponse, TenantPolicyRequest, TenantShardMigrateRequest,
32 : TenantShardMigrateResponse,
33 : };
34 : use pageserver_api::models::{
35 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode, PageserverUtilization,
36 : SecondaryProgress, ShardParameters, TenantConfig, TenantConfigPatchRequest,
37 : TenantConfigRequest, TenantLocationConfigRequest, TenantLocationConfigResponse,
38 : TenantShardLocation, TenantShardSplitRequest, TenantShardSplitResponse,
39 : TenantTimeTravelRequest, TimelineArchivalConfigRequest, TimelineCreateRequest, TimelineInfo,
40 : TopTenantShardsRequest,
41 : };
42 : use pageserver_api::shard::{
43 : ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
44 : };
45 : use pageserver_api::upcall_api::{
46 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest, ValidateResponse,
47 : ValidateResponseTenant,
48 : };
49 : use pageserver_client::{BlockUnblock, mgmt_api};
50 : use reqwest::StatusCode;
51 : use safekeeper_api::models::SafekeeperUtilization;
52 : use tokio::sync::TryAcquireError;
53 : use tokio::sync::mpsc::error::TrySendError;
54 : use tokio_util::sync::CancellationToken;
55 : use tracing::{Instrument, instrument};
56 : use utils::completion::Barrier;
57 : use utils::generation::Generation;
58 : use utils::id::{NodeId, TenantId, TimelineId};
59 : use utils::sync::gate::Gate;
60 : use utils::{failpoint_support, pausable_failpoint};
61 :
62 : use crate::background_node_operations::{
63 : Drain, Fill, MAX_RECONCILES_PER_OPERATION, Operation, OperationError, OperationHandler,
64 : };
65 : use crate::compute_hook::{self, ComputeHook, NotifyError};
66 : use crate::drain_utils::{self, TenantShardDrain, TenantShardIterator};
67 : use crate::heartbeater::{Heartbeater, PageserverState, SafekeeperState};
68 : use crate::id_lock_map::{
69 : IdLockMap, TracingExclusiveGuard, trace_exclusive_lock, trace_shared_lock,
70 : };
71 : use crate::leadership::Leadership;
72 : use crate::metrics;
73 : use crate::node::{AvailabilityTransition, Node};
74 : use crate::pageserver_client::PageserverClient;
75 : use crate::peer_client::GlobalObservedState;
76 : use crate::persistence::split_state::SplitState;
77 : use crate::persistence::{
78 : AbortShardSplitStatus, ControllerPersistence, DatabaseError, DatabaseResult,
79 : MetadataHealthPersistence, Persistence, ShardGenerationState, TenantFilter,
80 : TenantShardPersistence,
81 : };
82 : use crate::reconciler::{
83 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder, ReconcilerPriority,
84 : attached_location_conf,
85 : };
86 : use crate::safekeeper::Safekeeper;
87 : use crate::scheduler::{MaySchedule, ScheduleContext, ScheduleError, ScheduleMode, Scheduler};
88 : use crate::tenant_shard::{
89 : IntentState, MigrateAttachment, ObservedState, ObservedStateDelta, ObservedStateLocation,
90 : ReconcileNeeded, ReconcileResult, ReconcileWaitError, ReconcilerStatus, ReconcilerWaiter,
91 : ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
92 : };
93 :
94 : const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
95 :
96 : // For operations that should be quick, like attaching a new tenant
97 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
98 :
99 : // For operations that might be slow, like migrating a tenant with
100 : // some data in it.
101 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
102 :
103 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
104 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
105 : // input generation from future requests as authoritative.
106 : const INITIAL_GENERATION: Generation = Generation::new(0);
107 :
108 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
109 : /// up on unresponsive pageservers and proceed.
110 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
111 :
112 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
113 : /// This must be long enough to cover node restarts as well as normal operations: in future
114 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
115 :
116 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
117 : /// offline.
118 : ///
119 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
120 : /// handling of the re-attach response may take a long time and blocks heartbeats from
121 : /// being handled on the pageserver side.
122 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
123 :
124 : /// How often to send heartbeats to registered nodes?
125 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
126 :
127 : /// How long is too long for a reconciliation?
128 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
129 :
130 : #[derive(Clone, strum_macros::Display)]
131 : enum TenantOperations {
132 : Create,
133 : LocationConfig,
134 : ConfigSet,
135 : ConfigPatch,
136 : TimeTravelRemoteStorage,
137 : Delete,
138 : UpdatePolicy,
139 : ShardSplit,
140 : SecondaryDownload,
141 : TimelineCreate,
142 : TimelineDelete,
143 : AttachHook,
144 : TimelineArchivalConfig,
145 : TimelineDetachAncestor,
146 : TimelineGcBlockUnblock,
147 : DropDetached,
148 : DownloadHeatmapLayers,
149 : }
150 :
151 : #[derive(Clone, strum_macros::Display)]
152 : enum NodeOperations {
153 : Register,
154 : Configure,
155 : Delete,
156 : }
157 :
158 : /// The leadership status for the storage controller process.
159 : /// Allowed transitions are:
160 : /// 1. Leader -> SteppedDown
161 : /// 2. Candidate -> Leader
162 : #[derive(
163 : Eq,
164 : PartialEq,
165 : Copy,
166 : Clone,
167 : strum_macros::Display,
168 0 : strum_macros::EnumIter,
169 : measured::FixedCardinalityLabel,
170 : )]
171 : #[strum(serialize_all = "snake_case")]
172 : pub(crate) enum LeadershipStatus {
173 : /// This is the steady state where the storage controller can produce
174 : /// side effects in the cluster.
175 : Leader,
176 : /// We've been notified to step down by another candidate. No reconciliations
177 : /// take place in this state.
178 : SteppedDown,
179 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
180 : #[allow(unused)]
181 : Candidate,
182 : }
183 :
184 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
185 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
186 :
187 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
188 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
189 : // than they're being pushed onto the queue.
190 : const MAX_DELAYED_RECONCILES: usize = 10000;
191 :
192 : // Top level state available to all HTTP handlers
193 : struct ServiceState {
194 : leadership_status: LeadershipStatus,
195 :
196 : tenants: BTreeMap<TenantShardId, TenantShard>,
197 :
198 : nodes: Arc<HashMap<NodeId, Node>>,
199 :
200 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
201 :
202 : scheduler: Scheduler,
203 :
204 : /// Ongoing background operation on the cluster if any is running.
205 : /// Note that only one such operation may run at any given time,
206 : /// hence the type choice.
207 : ongoing_operation: Option<OperationHandler>,
208 :
209 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
210 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
211 : }
212 :
213 : /// Transform an error from a pageserver into an error to return to callers of a storage
214 : /// controller API.
215 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
216 0 : match e {
217 0 : mgmt_api::Error::SendRequest(e) => {
218 0 : // Presume errors sending requests are connectivity/availability issues
219 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
220 : }
221 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
222 0 : // Presume errors receiving body are connectivity/availability issues
223 0 : ApiError::ResourceUnavailable(
224 0 : format!("{node} error receiving error body: {str}").into(),
225 0 : )
226 : }
227 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
228 0 : // Return 500 for decoding errors.
229 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
230 : }
231 0 : mgmt_api::Error::ReceiveBody(err) => {
232 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
233 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
234 0 : ApiError::ResourceUnavailable(
235 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
236 0 : )
237 : }
238 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
239 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
240 : }
241 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
242 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
243 : }
244 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
245 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
246 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
247 : // internal server errors, showing that something is wrong with the pageserver or
248 : // storage controller's auth configuration.
249 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
250 : }
251 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
252 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
253 0 : // turn ask our clients to wait + retry
254 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
255 : }
256 0 : mgmt_api::Error::ApiError(status, msg) => {
257 0 : // Presume general case of pageserver API errors is that we tried to do something
258 0 : // that can't be done right now.
259 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
260 : }
261 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
262 : }
263 0 : }
264 :
265 : impl ServiceState {
266 0 : fn new(
267 0 : nodes: HashMap<NodeId, Node>,
268 0 : safekeepers: HashMap<NodeId, Safekeeper>,
269 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
270 0 : scheduler: Scheduler,
271 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
272 0 : initial_leadership_status: LeadershipStatus,
273 0 : ) -> Self {
274 0 : metrics::update_leadership_status(initial_leadership_status);
275 0 :
276 0 : Self {
277 0 : leadership_status: initial_leadership_status,
278 0 : tenants,
279 0 : nodes: Arc::new(nodes),
280 0 : safekeepers: Arc::new(safekeepers),
281 0 : scheduler,
282 0 : ongoing_operation: None,
283 0 : delayed_reconcile_rx,
284 0 : }
285 0 : }
286 :
287 0 : fn parts_mut(
288 0 : &mut self,
289 0 : ) -> (
290 0 : &mut Arc<HashMap<NodeId, Node>>,
291 0 : &mut BTreeMap<TenantShardId, TenantShard>,
292 0 : &mut Scheduler,
293 0 : ) {
294 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
295 0 : }
296 :
297 : #[allow(clippy::type_complexity)]
298 0 : fn parts_mut_sk(
299 0 : &mut self,
300 0 : ) -> (
301 0 : &mut Arc<HashMap<NodeId, Node>>,
302 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
303 0 : &mut BTreeMap<TenantShardId, TenantShard>,
304 0 : &mut Scheduler,
305 0 : ) {
306 0 : (
307 0 : &mut self.nodes,
308 0 : &mut self.safekeepers,
309 0 : &mut self.tenants,
310 0 : &mut self.scheduler,
311 0 : )
312 0 : }
313 :
314 0 : fn get_leadership_status(&self) -> LeadershipStatus {
315 0 : self.leadership_status
316 0 : }
317 :
318 0 : fn step_down(&mut self) {
319 0 : self.leadership_status = LeadershipStatus::SteppedDown;
320 0 : metrics::update_leadership_status(self.leadership_status);
321 0 : }
322 :
323 0 : fn become_leader(&mut self) {
324 0 : self.leadership_status = LeadershipStatus::Leader;
325 0 : metrics::update_leadership_status(self.leadership_status);
326 0 : }
327 : }
328 :
329 : #[derive(Clone)]
330 : pub struct Config {
331 : // All pageservers managed by one instance of this service must have
332 : // the same public key. This JWT token will be used to authenticate
333 : // this service to the pageservers it manages.
334 : pub pageserver_jwt_token: Option<String>,
335 :
336 : // All safekeepers managed by one instance of this service must have
337 : // the same public key. This JWT token will be used to authenticate
338 : // this service to the safekeepers it manages.
339 : pub safekeeper_jwt_token: Option<String>,
340 :
341 : // This JWT token will be used to authenticate this service to the control plane.
342 : pub control_plane_jwt_token: Option<String>,
343 :
344 : // This JWT token will be used to authenticate with other storage controller instances
345 : pub peer_jwt_token: Option<String>,
346 :
347 : /// Where the compute hook should send notifications of pageserver attachment locations
348 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
349 : /// assume it is running in a test environment and try to update neon_local.
350 : pub compute_hook_url: Option<String>,
351 :
352 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
353 : /// considered active. Once the grace period elapses, the next heartbeat failure will
354 : /// mark the pagseserver offline.
355 : pub max_offline_interval: Duration,
356 :
357 : /// Extended grace period within which pageserver may not respond to heartbeats.
358 : /// This extended grace period kicks in after the node has been drained for restart
359 : /// and/or upon handling the re-attach request from a node.
360 : pub max_warming_up_interval: Duration,
361 :
362 : /// How many normal-priority Reconcilers may be spawned concurrently
363 : pub reconciler_concurrency: usize,
364 :
365 : /// How many high-priority Reconcilers may be spawned concurrently
366 : pub priority_reconciler_concurrency: usize,
367 :
368 : /// How large must a shard grow in bytes before we split it?
369 : /// None disables auto-splitting.
370 : pub split_threshold: Option<u64>,
371 :
372 : // TODO: make this cfg(feature = "testing")
373 : pub neon_local_repo_dir: Option<PathBuf>,
374 :
375 : // Maximum acceptable download lag for the secondary location
376 : // while draining a node. If the secondary location is lagging
377 : // by more than the configured amount, then the secondary is not
378 : // upgraded to primary.
379 : pub max_secondary_lag_bytes: Option<u64>,
380 :
381 : pub heartbeat_interval: Duration,
382 :
383 : pub address_for_peers: Option<Uri>,
384 :
385 : pub start_as_candidate: bool,
386 :
387 : pub http_service_port: i32,
388 :
389 : pub long_reconcile_threshold: Duration,
390 :
391 : pub use_https_pageserver_api: bool,
392 : }
393 :
394 : impl From<DatabaseError> for ApiError {
395 0 : fn from(err: DatabaseError) -> ApiError {
396 0 : match err {
397 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
398 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
399 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
400 0 : ApiError::ShuttingDown
401 : }
402 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
403 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
404 : }
405 : }
406 0 : }
407 : }
408 :
409 : enum InitialShardScheduleOutcome {
410 : Scheduled(TenantCreateResponseShard),
411 : NotScheduled,
412 : ShardScheduleError(ScheduleError),
413 : }
414 :
415 : pub struct Service {
416 : inner: Arc<std::sync::RwLock<ServiceState>>,
417 : config: Config,
418 : persistence: Arc<Persistence>,
419 : compute_hook: Arc<ComputeHook>,
420 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
421 :
422 : heartbeater_ps: Heartbeater<Node, PageserverState>,
423 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
424 :
425 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
426 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
427 :
428 : // Locking on a tenant granularity (covers all shards in the tenant):
429 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
430 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
431 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
432 :
433 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
434 : // that transition it to/from Active.
435 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
436 :
437 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
438 : // and reconciliation on startup.
439 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
440 :
441 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
442 : // a human user might be waiting for.
443 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
444 :
445 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
446 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
447 : ///
448 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
449 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
450 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
451 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
452 :
453 : // Process shutdown will fire this token
454 : cancel: CancellationToken,
455 :
456 : // Child token of [`Service::cancel`] used by reconcilers
457 : reconcilers_cancel: CancellationToken,
458 :
459 : // Background tasks will hold this gate
460 : gate: Gate,
461 :
462 : // Reconcilers background tasks will hold this gate
463 : reconcilers_gate: Gate,
464 :
465 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
466 : /// passes, it isn't safe to do any actions that mutate tenants.
467 : pub(crate) startup_complete: Barrier,
468 : }
469 :
470 : impl From<ReconcileWaitError> for ApiError {
471 0 : fn from(value: ReconcileWaitError) -> Self {
472 0 : match value {
473 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
474 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
475 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
476 : }
477 0 : }
478 : }
479 :
480 : impl From<OperationError> for ApiError {
481 0 : fn from(value: OperationError) -> Self {
482 0 : match value {
483 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
484 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
485 : }
486 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
487 : }
488 0 : }
489 : }
490 :
491 : #[allow(clippy::large_enum_variant)]
492 : enum TenantCreateOrUpdate {
493 : Create(TenantCreateRequest),
494 : Update(Vec<ShardUpdate>),
495 : }
496 :
497 : struct ShardSplitParams {
498 : old_shard_count: ShardCount,
499 : new_shard_count: ShardCount,
500 : new_stripe_size: Option<ShardStripeSize>,
501 : targets: Vec<ShardSplitTarget>,
502 : policy: PlacementPolicy,
503 : config: TenantConfig,
504 : shard_ident: ShardIdentity,
505 : preferred_az_id: Option<AvailabilityZone>,
506 : }
507 :
508 : // When preparing for a shard split, we may either choose to proceed with the split,
509 : // or find that the work is already done and return NoOp.
510 : enum ShardSplitAction {
511 : Split(Box<ShardSplitParams>),
512 : NoOp(TenantShardSplitResponse),
513 : }
514 :
515 : // A parent shard which will be split
516 : struct ShardSplitTarget {
517 : parent_id: TenantShardId,
518 : node: Node,
519 : child_ids: Vec<TenantShardId>,
520 : }
521 :
522 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
523 : /// might not be available. We therefore use a queue of abort operations processed in the background.
524 : struct TenantShardSplitAbort {
525 : tenant_id: TenantId,
526 : /// The target values from the request that failed
527 : new_shard_count: ShardCount,
528 : new_stripe_size: Option<ShardStripeSize>,
529 : /// Until this abort op is complete, no other operations may be done on the tenant
530 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
531 : }
532 :
533 : #[derive(thiserror::Error, Debug)]
534 : enum TenantShardSplitAbortError {
535 : #[error(transparent)]
536 : Database(#[from] DatabaseError),
537 : #[error(transparent)]
538 : Remote(#[from] mgmt_api::Error),
539 : #[error("Unavailable")]
540 : Unavailable,
541 : }
542 :
543 : struct ShardUpdate {
544 : tenant_shard_id: TenantShardId,
545 : placement_policy: PlacementPolicy,
546 : tenant_config: TenantConfig,
547 :
548 : /// If this is None, generation is not updated.
549 : generation: Option<Generation>,
550 :
551 : /// If this is None, scheduling policy is not updated.
552 : scheduling_policy: Option<ShardSchedulingPolicy>,
553 : }
554 :
555 : enum StopReconciliationsReason {
556 : ShuttingDown,
557 : SteppingDown,
558 : }
559 :
560 : impl std::fmt::Display for StopReconciliationsReason {
561 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
562 0 : let s = match self {
563 0 : Self::ShuttingDown => "Shutting down",
564 0 : Self::SteppingDown => "Stepping down",
565 : };
566 0 : write!(writer, "{}", s)
567 0 : }
568 : }
569 :
570 : pub(crate) enum ReconcileResultRequest {
571 : ReconcileResult(ReconcileResult),
572 : Stop,
573 : }
574 :
575 : #[derive(Clone)]
576 : struct MutationLocation {
577 : node: Node,
578 : generation: Generation,
579 : }
580 :
581 : #[derive(Clone)]
582 : struct ShardMutationLocations {
583 : latest: MutationLocation,
584 : other: Vec<MutationLocation>,
585 : }
586 :
587 : #[derive(Default, Clone)]
588 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
589 :
590 : impl Service {
591 0 : pub fn get_config(&self) -> &Config {
592 0 : &self.config
593 0 : }
594 :
595 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
596 : /// view of the world, and determine which pageservers are responsive.
597 : #[instrument(skip_all)]
598 : async fn startup_reconcile(
599 : self: &Arc<Service>,
600 : current_leader: Option<ControllerPersistence>,
601 : leader_step_down_state: Option<GlobalObservedState>,
602 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
603 : Result<(), (TenantShardId, NotifyError)>,
604 : >,
605 : ) {
606 : // Startup reconciliation does I/O to other services: whether they
607 : // are responsive or not, we should aim to finish within our deadline, because:
608 : // - If we don't, a k8s readiness hook watching /ready will kill us.
609 : // - While we're waiting for startup reconciliation, we are not fully
610 : // available for end user operations like creating/deleting tenants and timelines.
611 : //
612 : // We set multiple deadlines to break up the time available between the phases of work: this is
613 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
614 : let start_at = Instant::now();
615 : let node_scan_deadline = start_at
616 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
617 : .expect("Reconcile timeout is a modest constant");
618 :
619 : let observed = if let Some(state) = leader_step_down_state {
620 : tracing::info!(
621 : "Using observed state received from leader at {}",
622 : current_leader.as_ref().unwrap().address
623 : );
624 :
625 : state
626 : } else {
627 : self.build_global_observed_state(node_scan_deadline).await
628 : };
629 :
630 : // Accumulate a list of any tenant locations that ought to be detached
631 : let mut cleanup = Vec::new();
632 :
633 : // Send initial heartbeat requests to all nodes loaded from the database
634 : let all_nodes = {
635 : let locked = self.inner.read().unwrap();
636 : locked.nodes.clone()
637 : };
638 : let (mut nodes_online, mut sks_online) =
639 : self.initial_heartbeat_round(all_nodes.keys()).await;
640 :
641 : // List of tenants for which we will attempt to notify compute of their location at startup
642 : let mut compute_notifications = Vec::new();
643 :
644 : // Populate intent and observed states for all tenants, based on reported state on pageservers
645 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
646 : let shard_count = {
647 : let mut locked = self.inner.write().unwrap();
648 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
649 :
650 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
651 : let mut new_nodes = (**nodes).clone();
652 : for (node_id, node) in new_nodes.iter_mut() {
653 : if let Some(utilization) = nodes_online.remove(node_id) {
654 : node.set_availability(NodeAvailability::Active(utilization));
655 : scheduler.node_upsert(node);
656 : }
657 : }
658 : *nodes = Arc::new(new_nodes);
659 :
660 : let mut new_sks = (**safekeepers).clone();
661 : for (node_id, node) in new_sks.iter_mut() {
662 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
663 : node.set_availability(SafekeeperState::Available {
664 : utilization,
665 : last_seen_at,
666 : });
667 : }
668 : }
669 : *safekeepers = Arc::new(new_sks);
670 :
671 : for (tenant_shard_id, observed_state) in observed.0 {
672 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
673 : for node_id in observed_state.locations.keys() {
674 : cleanup.push((tenant_shard_id, *node_id));
675 : }
676 :
677 : continue;
678 : };
679 :
680 : tenant_shard.observed = observed_state;
681 : }
682 :
683 : // Populate each tenant's intent state
684 : let mut schedule_context = ScheduleContext::default();
685 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
686 : if tenant_shard_id.shard_number == ShardNumber(0) {
687 : // Reset scheduling context each time we advance to the next Tenant
688 : schedule_context = ScheduleContext::default();
689 : }
690 :
691 : tenant_shard.intent_from_observed(scheduler);
692 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
693 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
694 : // not enough pageservers are available. The tenant may well still be available
695 : // to clients.
696 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
697 : } else {
698 : // If we're both intending and observed to be attached at a particular node, we will
699 : // emit a compute notification for this. In the case where our observed state does not
700 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
701 : if let Some(attached_at) = tenant_shard.stably_attached() {
702 : compute_notifications.push(compute_hook::ShardUpdate {
703 : tenant_shard_id: *tenant_shard_id,
704 : node_id: attached_at,
705 : stripe_size: tenant_shard.shard.stripe_size,
706 : preferred_az: tenant_shard
707 : .preferred_az()
708 0 : .map(|az| Cow::Owned(az.clone())),
709 : });
710 : }
711 : }
712 : }
713 :
714 : tenants.len()
715 : };
716 :
717 : // Before making any obeservable changes to the cluster, persist self
718 : // as leader in database and memory.
719 : let leadership = Leadership::new(
720 : self.persistence.clone(),
721 : self.config.clone(),
722 : self.cancel.child_token(),
723 : );
724 :
725 : if let Err(e) = leadership.become_leader(current_leader).await {
726 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
727 : std::process::exit(1);
728 : }
729 :
730 : self.inner.write().unwrap().become_leader();
731 :
732 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
733 : // generation_pageserver in the database.
734 :
735 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
736 : // will emit compute hook notifications when they reconcile.
737 : //
738 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
739 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
740 : // calls will be correctly ordered wrt these.
741 : //
742 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
743 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
744 : // unit and start doing I/O.
745 : tracing::info!(
746 : "Sending {} compute notifications",
747 : compute_notifications.len()
748 : );
749 : self.compute_hook.notify_background(
750 : compute_notifications,
751 : bg_compute_notify_result_tx.clone(),
752 : &self.cancel,
753 : );
754 :
755 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
756 : // which require it: under normal circumstances this should only include tenants that were in some
757 : // transient state before we restarted, or any tenants whose compute hooks failed above.
758 : tracing::info!("Checking for shards in need of reconciliation...");
759 : let reconcile_tasks = self.reconcile_all();
760 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
761 : // normal operations may proceed.
762 :
763 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
764 : // background because it does not need to complete in order to proceed with other work.
765 : if !cleanup.is_empty() {
766 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
767 : tokio::task::spawn({
768 : let cleanup_self = self.clone();
769 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
770 : });
771 : }
772 :
773 : tracing::info!(
774 : "Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)"
775 : );
776 : }
777 :
778 0 : async fn initial_heartbeat_round<'a>(
779 0 : &self,
780 0 : node_ids: impl Iterator<Item = &'a NodeId>,
781 0 : ) -> (
782 0 : HashMap<NodeId, PageserverUtilization>,
783 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
784 0 : ) {
785 0 : assert!(!self.startup_complete.is_ready());
786 :
787 0 : let all_nodes = {
788 0 : let locked = self.inner.read().unwrap();
789 0 : locked.nodes.clone()
790 0 : };
791 0 :
792 0 : let mut nodes_to_heartbeat = HashMap::new();
793 0 : for node_id in node_ids {
794 0 : match all_nodes.get(node_id) {
795 0 : Some(node) => {
796 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
797 0 : }
798 : None => {
799 0 : tracing::warn!("Node {node_id} was removed during start-up");
800 : }
801 : }
802 : }
803 :
804 0 : let all_sks = {
805 0 : let locked = self.inner.read().unwrap();
806 0 : locked.safekeepers.clone()
807 0 : };
808 0 :
809 0 : tracing::info!("Sending initial heartbeats...");
810 : // Put a small, but reasonable timeout to get the initial heartbeats of the safekeepers to avoid a storage controller downtime
811 : const SK_TIMEOUT: Duration = Duration::from_secs(5);
812 0 : let (res_ps, res_sk) = tokio::join!(
813 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
814 0 : tokio::time::timeout(SK_TIMEOUT, self.heartbeater_sk.heartbeat(all_sks))
815 0 : );
816 :
817 0 : let mut online_nodes = HashMap::new();
818 0 : if let Ok(deltas) = res_ps {
819 0 : for (node_id, status) in deltas.0 {
820 0 : match status {
821 0 : PageserverState::Available { utilization, .. } => {
822 0 : online_nodes.insert(node_id, utilization);
823 0 : }
824 0 : PageserverState::Offline => {}
825 : PageserverState::WarmingUp { .. } => {
826 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
827 : }
828 : }
829 : }
830 0 : }
831 :
832 0 : let mut online_sks = HashMap::new();
833 0 : if let Ok(Ok(deltas)) = res_sk {
834 0 : for (node_id, status) in deltas.0 {
835 0 : match status {
836 : SafekeeperState::Available {
837 0 : utilization,
838 0 : last_seen_at,
839 0 : } => {
840 0 : online_sks.insert(node_id, (utilization, last_seen_at));
841 0 : }
842 0 : SafekeeperState::Offline => {}
843 : }
844 : }
845 0 : }
846 :
847 0 : (online_nodes, online_sks)
848 0 : }
849 :
850 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
851 : ///
852 : /// The result includes only nodes which responded within the deadline
853 0 : async fn scan_node_locations(
854 0 : &self,
855 0 : deadline: Instant,
856 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
857 0 : let nodes = {
858 0 : let locked = self.inner.read().unwrap();
859 0 : locked.nodes.clone()
860 0 : };
861 0 :
862 0 : let mut node_results = HashMap::new();
863 0 :
864 0 : let mut node_list_futs = FuturesUnordered::new();
865 0 :
866 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
867 0 : for node in nodes.values() {
868 0 : node_list_futs.push({
869 0 : async move {
870 0 : tracing::info!("Scanning shards on node {node}...");
871 0 : let timeout = Duration::from_secs(5);
872 0 : let response = node
873 0 : .with_client_retries(
874 0 : |client| async move { client.list_location_config().await },
875 0 : &self.config.pageserver_jwt_token,
876 0 : 1,
877 0 : 5,
878 0 : timeout,
879 0 : &self.cancel,
880 0 : )
881 0 : .await;
882 0 : (node.get_id(), response)
883 0 : }
884 0 : });
885 0 : }
886 :
887 : loop {
888 0 : let (node_id, result) = tokio::select! {
889 0 : next = node_list_futs.next() => {
890 0 : match next {
891 0 : Some(result) => result,
892 : None =>{
893 : // We got results for all our nodes
894 0 : break;
895 : }
896 :
897 : }
898 : },
899 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
900 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
901 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
902 0 : break;
903 : }
904 : };
905 :
906 0 : let Some(list_response) = result else {
907 0 : tracing::info!("Shutdown during startup_reconcile");
908 0 : break;
909 : };
910 :
911 0 : match list_response {
912 0 : Err(e) => {
913 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
914 : }
915 0 : Ok(listing) => {
916 0 : node_results.insert(node_id, listing);
917 0 : }
918 : }
919 : }
920 :
921 0 : node_results
922 0 : }
923 :
924 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
925 0 : let node_listings = self.scan_node_locations(deadline).await;
926 0 : let mut observed = GlobalObservedState::default();
927 :
928 0 : for (node_id, location_confs) in node_listings {
929 0 : tracing::info!(
930 0 : "Received {} shard statuses from pageserver {}",
931 0 : location_confs.tenant_shards.len(),
932 : node_id
933 : );
934 :
935 0 : for (tid, location_conf) in location_confs.tenant_shards {
936 0 : let entry = observed.0.entry(tid).or_default();
937 0 : entry.locations.insert(
938 0 : node_id,
939 0 : ObservedStateLocation {
940 0 : conf: location_conf,
941 0 : },
942 0 : );
943 0 : }
944 : }
945 :
946 0 : observed
947 0 : }
948 :
949 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
950 : ///
951 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
952 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
953 : /// other task trying to attach it.
954 : #[instrument(skip_all)]
955 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
956 : let nodes = self.inner.read().unwrap().nodes.clone();
957 :
958 : for (tenant_shard_id, node_id) in cleanup {
959 : // A node reported a tenant_shard_id which is unknown to us: detach it.
960 : let Some(node) = nodes.get(&node_id) else {
961 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
962 : // a location to clean up on a node that has since been removed.
963 : tracing::info!(
964 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
965 : );
966 : continue;
967 : };
968 :
969 : if self.cancel.is_cancelled() {
970 : break;
971 : }
972 :
973 : let client = PageserverClient::new(
974 : node.get_id(),
975 : node.base_url(),
976 : self.config.pageserver_jwt_token.as_deref(),
977 : );
978 : match client
979 : .location_config(
980 : tenant_shard_id,
981 : LocationConfig {
982 : mode: LocationConfigMode::Detached,
983 : generation: None,
984 : secondary_conf: None,
985 : shard_number: tenant_shard_id.shard_number.0,
986 : shard_count: tenant_shard_id.shard_count.literal(),
987 : shard_stripe_size: 0,
988 : tenant_conf: models::TenantConfig::default(),
989 : },
990 : None,
991 : false,
992 : )
993 : .await
994 : {
995 : Ok(()) => {
996 : tracing::info!(
997 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
998 : );
999 : }
1000 : Err(e) => {
1001 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1002 : // break anything.
1003 : tracing::error!(
1004 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1005 : );
1006 : }
1007 : }
1008 : }
1009 : }
1010 :
1011 : /// Long running background task that periodically wakes up and looks for shards that need
1012 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1013 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1014 : /// for those retries.
1015 : #[instrument(skip_all)]
1016 : async fn background_reconcile(self: &Arc<Self>) {
1017 : self.startup_complete.clone().wait().await;
1018 :
1019 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1020 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1021 : while !self.reconcilers_cancel.is_cancelled() {
1022 : tokio::select! {
1023 : _ = interval.tick() => {
1024 : let reconciles_spawned = self.reconcile_all();
1025 : if reconciles_spawned == 0 {
1026 : // Run optimizer only when we didn't find any other work to do
1027 : self.optimize_all().await;
1028 : }
1029 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1030 : // must be responsive when new projects begin ingesting and reach the threshold.
1031 : self.autosplit_tenants().await;
1032 : }
1033 : _ = self.reconcilers_cancel.cancelled() => return
1034 : }
1035 : }
1036 : }
1037 : #[instrument(skip_all)]
1038 : async fn spawn_heartbeat_driver(&self) {
1039 : self.startup_complete.clone().wait().await;
1040 :
1041 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1042 : while !self.cancel.is_cancelled() {
1043 : tokio::select! {
1044 : _ = interval.tick() => { }
1045 : _ = self.cancel.cancelled() => return
1046 : };
1047 :
1048 : let nodes = {
1049 : let locked = self.inner.read().unwrap();
1050 : locked.nodes.clone()
1051 : };
1052 :
1053 : let safekeepers = {
1054 : let locked = self.inner.read().unwrap();
1055 : locked.safekeepers.clone()
1056 : };
1057 :
1058 : const SK_TIMEOUT: Duration = Duration::from_secs(3);
1059 : let (res_ps, res_sk) = tokio::join!(
1060 : self.heartbeater_ps.heartbeat(nodes),
1061 : tokio::time::timeout(SK_TIMEOUT, self.heartbeater_sk.heartbeat(safekeepers))
1062 : );
1063 :
1064 : if let Ok(deltas) = res_ps {
1065 : let mut to_handle = Vec::default();
1066 :
1067 : for (node_id, state) in deltas.0 {
1068 : let new_availability = match state {
1069 : PageserverState::Available { utilization, .. } => {
1070 : NodeAvailability::Active(utilization)
1071 : }
1072 : PageserverState::WarmingUp { started_at } => {
1073 : NodeAvailability::WarmingUp(started_at)
1074 : }
1075 : PageserverState::Offline => {
1076 : // The node might have been placed in the WarmingUp state
1077 : // while the heartbeat round was on-going. Hence, filter out
1078 : // offline transitions for WarmingUp nodes that are still within
1079 : // their grace period.
1080 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1081 : .get_node(node_id)
1082 : .await
1083 : .as_ref()
1084 0 : .map(|n| n.get_availability())
1085 : {
1086 : let now = Instant::now();
1087 : if now - *started_at >= self.config.max_warming_up_interval {
1088 : NodeAvailability::Offline
1089 : } else {
1090 : NodeAvailability::WarmingUp(*started_at)
1091 : }
1092 : } else {
1093 : NodeAvailability::Offline
1094 : }
1095 : }
1096 : };
1097 :
1098 : let node_lock = trace_exclusive_lock(
1099 : &self.node_op_locks,
1100 : node_id,
1101 : NodeOperations::Configure,
1102 : )
1103 : .await;
1104 :
1105 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1106 :
1107 : // This is the code path for geniune availability transitions (i.e node
1108 : // goes unavailable and/or comes back online).
1109 : let res = self
1110 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1111 : .await;
1112 :
1113 : match res {
1114 : Ok(transition) => {
1115 : // Keep hold of the lock until the availability transitions
1116 : // have been handled in
1117 : // [`Service::handle_node_availability_transitions`] in order avoid
1118 : // racing with [`Service::external_node_configure`].
1119 : to_handle.push((node_id, node_lock, transition));
1120 : }
1121 : Err(ApiError::NotFound(_)) => {
1122 : // This should be rare, but legitimate since the heartbeats are done
1123 : // on a snapshot of the nodes.
1124 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1125 : }
1126 : Err(ApiError::ShuttingDown) => {
1127 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1128 : }
1129 : Err(err) => {
1130 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1131 : // becomes unavailable again, we may get an error here.
1132 : tracing::error!(
1133 : "Failed to update node state {} after heartbeat round: {}",
1134 : node_id,
1135 : err
1136 : );
1137 : }
1138 : }
1139 : }
1140 :
1141 : // We collected all the transitions above and now we handle them.
1142 : let res = self.handle_node_availability_transitions(to_handle).await;
1143 : if let Err(errs) = res {
1144 : for (node_id, err) in errs {
1145 : match err {
1146 : ApiError::NotFound(_) => {
1147 : // This should be rare, but legitimate since the heartbeats are done
1148 : // on a snapshot of the nodes.
1149 : tracing::info!(
1150 : "Node {} was not found after heartbeat round",
1151 : node_id
1152 : );
1153 : }
1154 : err => {
1155 : tracing::error!(
1156 : "Failed to handle availability transition for {} after heartbeat round: {}",
1157 : node_id,
1158 : err
1159 : );
1160 : }
1161 : }
1162 : }
1163 : }
1164 : }
1165 : if let Ok(Ok(deltas)) = res_sk {
1166 : let mut locked = self.inner.write().unwrap();
1167 : let mut safekeepers = (*locked.safekeepers).clone();
1168 : for (id, state) in deltas.0 {
1169 : let Some(sk) = safekeepers.get_mut(&id) else {
1170 : tracing::info!(
1171 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1172 : );
1173 : continue;
1174 : };
1175 : sk.set_availability(state);
1176 : }
1177 : locked.safekeepers = Arc::new(safekeepers);
1178 : }
1179 : }
1180 : }
1181 :
1182 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1183 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1184 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1185 : /// will indicate that reconciliation is not needed.
1186 : #[instrument(skip_all, fields(
1187 : seq=%result.sequence,
1188 : tenant_id=%result.tenant_shard_id.tenant_id,
1189 : shard_id=%result.tenant_shard_id.shard_slug(),
1190 : ))]
1191 : fn process_result(&self, result: ReconcileResult) {
1192 : let mut locked = self.inner.write().unwrap();
1193 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1194 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1195 : // A reconciliation result might race with removing a tenant: drop results for
1196 : // tenants that aren't in our map.
1197 : return;
1198 : };
1199 :
1200 : // Usually generation should only be updated via this path, so the max() isn't
1201 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1202 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1203 :
1204 : // If the reconciler signals that it failed to notify compute, set this state on
1205 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1206 : tenant.pending_compute_notification = result.pending_compute_notification;
1207 :
1208 : // Let the TenantShard know it is idle.
1209 : tenant.reconcile_complete(result.sequence);
1210 :
1211 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1212 : // make to the tenant
1213 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1214 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1215 : // make to the tenant
1216 0 : let node = nodes.get(delta.node_id())?;
1217 :
1218 0 : if node.is_available() {
1219 0 : return Some(delta);
1220 0 : }
1221 0 :
1222 0 : // In case a node became unavailable concurrently with the reconcile, observed
1223 0 : // locations on it are now uncertain. By convention, set them to None in order
1224 0 : // for them to get refreshed when the node comes back online.
1225 0 : Some(ObservedStateDelta::Upsert(Box::new((
1226 0 : node.get_id(),
1227 0 : ObservedStateLocation { conf: None },
1228 0 : ))))
1229 0 : });
1230 :
1231 : match result.result {
1232 : Ok(()) => {
1233 : tenant.apply_observed_deltas(deltas);
1234 : tenant.waiter.advance(result.sequence);
1235 : }
1236 : Err(e) => {
1237 : match e {
1238 : ReconcileError::Cancel => {
1239 : tracing::info!("Reconciler was cancelled");
1240 : }
1241 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1242 : // This might be due to the reconciler getting cancelled, or it might
1243 : // be due to the `Node` being marked offline.
1244 : tracing::info!("Reconciler cancelled during pageserver API call");
1245 : }
1246 : _ => {
1247 : tracing::warn!("Reconcile error: {}", e);
1248 : }
1249 : }
1250 :
1251 : // Ordering: populate last_error before advancing error_seq,
1252 : // so that waiters will see the correct error after waiting.
1253 : tenant.set_last_error(result.sequence, e);
1254 :
1255 : // Skip deletions on reconcile failures
1256 : let upsert_deltas =
1257 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1258 : tenant.apply_observed_deltas(upsert_deltas);
1259 : }
1260 : }
1261 :
1262 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1263 : if tenant.policy == PlacementPolicy::Detached {
1264 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1265 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1266 : // duration of the request.
1267 : let guard = self.tenant_op_locks.try_exclusive(
1268 : tenant.tenant_shard_id.tenant_id,
1269 : TenantOperations::DropDetached,
1270 : );
1271 : if let Some(guard) = guard {
1272 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1273 : }
1274 : }
1275 :
1276 : // Maybe some other work can proceed now that this job finished.
1277 : //
1278 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1279 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1280 : if self.reconciler_concurrency.available_permits() > 0 {
1281 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1282 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1283 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1284 : shard.delayed_reconcile = false;
1285 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1286 : }
1287 :
1288 : if self.reconciler_concurrency.available_permits() == 0 {
1289 : break;
1290 : }
1291 : }
1292 : }
1293 : }
1294 :
1295 0 : async fn process_results(
1296 0 : &self,
1297 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1298 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1299 0 : Result<(), (TenantShardId, NotifyError)>,
1300 0 : >,
1301 0 : ) {
1302 : loop {
1303 : // Wait for the next result, or for cancellation
1304 0 : tokio::select! {
1305 0 : r = result_rx.recv() => {
1306 0 : match r {
1307 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1308 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1309 : }
1310 : }
1311 0 : _ = async{
1312 0 : match bg_compute_hook_result_rx.recv().await {
1313 0 : Some(result) => {
1314 0 : if let Err((tenant_shard_id, notify_error)) = result {
1315 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1316 0 : let mut locked = self.inner.write().unwrap();
1317 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1318 0 : shard.pending_compute_notification = true;
1319 0 : }
1320 :
1321 0 : }
1322 : },
1323 : None => {
1324 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1325 0 : self.cancel.cancelled().await;
1326 : }
1327 : }
1328 0 : } => {},
1329 0 : _ = self.cancel.cancelled() => {
1330 0 : break;
1331 : }
1332 : };
1333 : }
1334 0 : }
1335 :
1336 0 : async fn process_aborts(
1337 0 : &self,
1338 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1339 0 : ) {
1340 : loop {
1341 : // Wait for the next result, or for cancellation
1342 0 : let op = tokio::select! {
1343 0 : r = abort_rx.recv() => {
1344 0 : match r {
1345 0 : Some(op) => {op},
1346 0 : None => {break;}
1347 : }
1348 : }
1349 0 : _ = self.cancel.cancelled() => {
1350 0 : break;
1351 : }
1352 : };
1353 :
1354 : // Retry until shutdown: we must keep this request object alive until it is properly
1355 : // processed, as it holds a lock guard that prevents other operations trying to do things
1356 : // to the tenant while it is in a weird part-split state.
1357 0 : while !self.cancel.is_cancelled() {
1358 0 : match self.abort_tenant_shard_split(&op).await {
1359 0 : Ok(_) => break,
1360 0 : Err(e) => {
1361 0 : tracing::warn!(
1362 0 : "Failed to abort shard split on {}, will retry: {e}",
1363 : op.tenant_id
1364 : );
1365 :
1366 : // If a node is unavailable, we hope that it has been properly marked Offline
1367 : // when we retry, so that the abort op will succeed. If the abort op is failing
1368 : // for some other reason, we will keep retrying forever, or until a human notices
1369 : // and does something about it (either fixing a pageserver or restarting the controller).
1370 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1371 0 : .await
1372 0 : .ok();
1373 : }
1374 : }
1375 : }
1376 : }
1377 0 : }
1378 :
1379 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1380 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1381 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1382 0 :
1383 0 : let leadership_cancel = CancellationToken::new();
1384 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1385 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1386 :
1387 : // Apply the migrations **after** the current leader has stepped down
1388 : // (or we've given up waiting for it), but **before** reading from the
1389 : // database. The only exception is reading the current leader before
1390 : // migrating.
1391 0 : persistence.migration_run().await?;
1392 :
1393 0 : tracing::info!("Loading nodes from database...");
1394 0 : let nodes = persistence
1395 0 : .list_nodes()
1396 0 : .await?
1397 0 : .into_iter()
1398 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1399 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1400 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1401 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1402 0 : metrics::METRICS_REGISTRY
1403 0 : .metrics_group
1404 0 : .storage_controller_pageserver_nodes
1405 0 : .set(nodes.len() as i64);
1406 0 :
1407 0 : tracing::info!("Loading safekeepers from database...");
1408 0 : let safekeepers = persistence
1409 0 : .list_safekeepers()
1410 0 : .await?
1411 0 : .into_iter()
1412 0 : .map(|skp| Safekeeper::from_persistence(skp, CancellationToken::new()))
1413 0 : .collect::<Vec<_>>();
1414 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1415 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1416 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1417 :
1418 0 : tracing::info!("Loading shards from database...");
1419 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1420 0 : tracing::info!(
1421 0 : "Loaded {} shards from database.",
1422 0 : tenant_shard_persistence.len()
1423 : );
1424 :
1425 : // If any shard splits were in progress, reset the database state to abort them
1426 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1427 0 : HashMap::new();
1428 0 : for tsp in &mut tenant_shard_persistence {
1429 0 : let shard = tsp.get_shard_identity()?;
1430 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1431 0 : let entry = tenant_shard_count_min_max
1432 0 : .entry(tenant_shard_id.tenant_id)
1433 0 : .or_insert_with(|| (shard.count, shard.count));
1434 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1435 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1436 0 : }
1437 :
1438 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1439 0 : if count_min != count_max {
1440 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1441 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1442 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1443 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1444 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1445 :
1446 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1447 : // identified this tenant has having mismatching min/max counts.
1448 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1449 :
1450 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1451 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1452 0 : // Set idle split state on those shards that we will retain.
1453 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1454 0 : if tsp_tenant_id == tenant_id
1455 0 : && tsp.get_shard_identity().unwrap().count == count_min
1456 0 : {
1457 0 : tsp.splitting = SplitState::Idle;
1458 0 : } else if tsp_tenant_id == tenant_id {
1459 : // Leave the splitting state on the child shards: this will be used next to
1460 : // drop them.
1461 0 : tracing::info!(
1462 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1463 : );
1464 0 : }
1465 0 : });
1466 0 :
1467 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1468 0 : tenant_shard_persistence.retain(|tsp| {
1469 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1470 0 : || tsp.splitting == SplitState::Idle
1471 0 : });
1472 0 : }
1473 : }
1474 :
1475 0 : let mut tenants = BTreeMap::new();
1476 0 :
1477 0 : let mut scheduler = Scheduler::new(nodes.values());
1478 :
1479 : #[cfg(feature = "testing")]
1480 : {
1481 : use pageserver_api::controller_api::AvailabilityZone;
1482 :
1483 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1484 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1485 : // after when pageservers start up and register.
1486 0 : let mut node_ids = HashSet::new();
1487 0 : for tsp in &tenant_shard_persistence {
1488 0 : if let Some(node_id) = tsp.generation_pageserver {
1489 0 : node_ids.insert(node_id);
1490 0 : }
1491 : }
1492 0 : for node_id in node_ids {
1493 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1494 0 : let node = Node::new(
1495 0 : NodeId(node_id as u64),
1496 0 : "".to_string(),
1497 0 : 123,
1498 0 : None,
1499 0 : "".to_string(),
1500 0 : 123,
1501 0 : AvailabilityZone("test_az".to_string()),
1502 0 : false,
1503 0 : )
1504 0 : .unwrap();
1505 0 :
1506 0 : scheduler.node_upsert(&node);
1507 : }
1508 : }
1509 0 : for tsp in tenant_shard_persistence {
1510 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1511 :
1512 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1513 : // it with what we can infer: the node for which a generation was most recently issued.
1514 0 : let mut intent = IntentState::new(
1515 0 : tsp.preferred_az_id
1516 0 : .as_ref()
1517 0 : .map(|az| AvailabilityZone(az.clone())),
1518 0 : );
1519 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1520 : {
1521 0 : if nodes.contains_key(&generation_pageserver) {
1522 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1523 0 : } else {
1524 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1525 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1526 : // on different pageservers.
1527 0 : tracing::warn!(
1528 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1529 : );
1530 : }
1531 0 : }
1532 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1533 :
1534 0 : tenants.insert(tenant_shard_id, new_tenant);
1535 : }
1536 :
1537 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1538 0 :
1539 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1540 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1541 0 : tokio::sync::mpsc::channel(512);
1542 0 :
1543 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1544 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1545 0 :
1546 0 : let cancel = CancellationToken::new();
1547 0 : let reconcilers_cancel = cancel.child_token();
1548 0 :
1549 0 : let heartbeater_ps = Heartbeater::new(
1550 0 : config.pageserver_jwt_token.clone(),
1551 0 : config.max_offline_interval,
1552 0 : config.max_warming_up_interval,
1553 0 : cancel.clone(),
1554 0 : );
1555 0 :
1556 0 : let heartbeater_sk = Heartbeater::new(
1557 0 : config.safekeeper_jwt_token.clone(),
1558 0 : config.max_offline_interval,
1559 0 : config.max_warming_up_interval,
1560 0 : cancel.clone(),
1561 0 : );
1562 :
1563 0 : let initial_leadership_status = if config.start_as_candidate {
1564 0 : LeadershipStatus::Candidate
1565 : } else {
1566 0 : LeadershipStatus::Leader
1567 : };
1568 :
1569 0 : let this = Arc::new(Self {
1570 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1571 0 : nodes,
1572 0 : safekeepers,
1573 0 : tenants,
1574 0 : scheduler,
1575 0 : delayed_reconcile_rx,
1576 0 : initial_leadership_status,
1577 0 : ))),
1578 0 : config: config.clone(),
1579 0 : persistence,
1580 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1581 0 : result_tx,
1582 0 : heartbeater_ps,
1583 0 : heartbeater_sk,
1584 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1585 0 : config.reconciler_concurrency,
1586 0 : )),
1587 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1588 0 : config.priority_reconciler_concurrency,
1589 0 : )),
1590 0 : delayed_reconcile_tx,
1591 0 : abort_tx,
1592 0 : startup_complete: startup_complete.clone(),
1593 0 : cancel,
1594 0 : reconcilers_cancel,
1595 0 : gate: Gate::default(),
1596 0 : reconcilers_gate: Gate::default(),
1597 0 : tenant_op_locks: Default::default(),
1598 0 : node_op_locks: Default::default(),
1599 0 : });
1600 0 :
1601 0 : let result_task_this = this.clone();
1602 0 : tokio::task::spawn(async move {
1603 : // Block shutdown until we're done (we must respect self.cancel)
1604 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1605 0 : result_task_this
1606 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1607 0 : .await
1608 0 : }
1609 0 : });
1610 0 :
1611 0 : tokio::task::spawn({
1612 0 : let this = this.clone();
1613 0 : async move {
1614 : // Block shutdown until we're done (we must respect self.cancel)
1615 0 : if let Ok(_gate) = this.gate.enter() {
1616 0 : this.process_aborts(abort_rx).await
1617 0 : }
1618 0 : }
1619 0 : });
1620 0 :
1621 0 : tokio::task::spawn({
1622 0 : let this = this.clone();
1623 0 : async move {
1624 0 : if let Ok(_gate) = this.gate.enter() {
1625 : loop {
1626 0 : tokio::select! {
1627 0 : _ = this.cancel.cancelled() => {
1628 0 : break;
1629 : },
1630 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1631 0 : };
1632 0 : this.tenant_op_locks.housekeeping();
1633 : }
1634 0 : }
1635 0 : }
1636 0 : });
1637 0 :
1638 0 : tokio::task::spawn({
1639 0 : let this = this.clone();
1640 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1641 0 : // is done.
1642 0 : let startup_completion = startup_completion.clone();
1643 0 : async move {
1644 : // Block shutdown until we're done (we must respect self.cancel)
1645 0 : let Ok(_gate) = this.gate.enter() else {
1646 0 : return;
1647 : };
1648 :
1649 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1650 0 : .await;
1651 :
1652 0 : drop(startup_completion);
1653 0 : }
1654 0 : });
1655 0 :
1656 0 : tokio::task::spawn({
1657 0 : let this = this.clone();
1658 0 : let startup_complete = startup_complete.clone();
1659 0 : async move {
1660 0 : startup_complete.wait().await;
1661 0 : this.background_reconcile().await;
1662 0 : }
1663 0 : });
1664 0 :
1665 0 : tokio::task::spawn({
1666 0 : let this = this.clone();
1667 0 : let startup_complete = startup_complete.clone();
1668 0 : async move {
1669 0 : startup_complete.wait().await;
1670 0 : this.spawn_heartbeat_driver().await;
1671 0 : }
1672 0 : });
1673 0 :
1674 0 : Ok(this)
1675 0 : }
1676 :
1677 0 : pub(crate) async fn attach_hook(
1678 0 : &self,
1679 0 : attach_req: AttachHookRequest,
1680 0 : ) -> anyhow::Result<AttachHookResponse> {
1681 0 : let _tenant_lock = trace_exclusive_lock(
1682 0 : &self.tenant_op_locks,
1683 0 : attach_req.tenant_shard_id.tenant_id,
1684 0 : TenantOperations::AttachHook,
1685 0 : )
1686 0 : .await;
1687 :
1688 : // This is a test hook. To enable using it on tenants that were created directly with
1689 : // the pageserver API (not via this service), we will auto-create any missing tenant
1690 : // shards with default state.
1691 0 : let insert = {
1692 0 : match self
1693 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1694 0 : .await
1695 : {
1696 0 : Ok(_) => false,
1697 0 : Err(ApiError::NotFound(_)) => true,
1698 0 : Err(e) => return Err(e.into()),
1699 : }
1700 : };
1701 :
1702 0 : if insert {
1703 0 : let tsp = TenantShardPersistence {
1704 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1705 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1706 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1707 0 : shard_stripe_size: 0,
1708 0 : generation: attach_req.generation_override.or(Some(0)),
1709 0 : generation_pageserver: None,
1710 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1711 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1712 0 : splitting: SplitState::default(),
1713 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1714 0 : .unwrap(),
1715 0 : preferred_az_id: None,
1716 0 : };
1717 0 :
1718 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1719 0 : Err(e) => match e {
1720 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1721 : DatabaseErrorKind::UniqueViolation,
1722 : _,
1723 : )) => {
1724 0 : tracing::info!(
1725 0 : "Raced with another request to insert tenant {}",
1726 : attach_req.tenant_shard_id
1727 : )
1728 : }
1729 0 : _ => return Err(e.into()),
1730 : },
1731 : Ok(()) => {
1732 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1733 :
1734 0 : let mut locked = self.inner.write().unwrap();
1735 0 : locked.tenants.insert(
1736 0 : attach_req.tenant_shard_id,
1737 0 : TenantShard::new(
1738 0 : attach_req.tenant_shard_id,
1739 0 : ShardIdentity::unsharded(),
1740 0 : PlacementPolicy::Attached(0),
1741 0 : None,
1742 0 : ),
1743 0 : );
1744 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1745 : }
1746 : }
1747 0 : }
1748 :
1749 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1750 0 : let maybe_tenant_conf = {
1751 0 : let locked = self.inner.write().unwrap();
1752 0 : locked
1753 0 : .tenants
1754 0 : .get(&attach_req.tenant_shard_id)
1755 0 : .map(|t| t.config.clone())
1756 0 : };
1757 0 :
1758 0 : match maybe_tenant_conf {
1759 0 : Some(conf) => {
1760 0 : let new_generation = self
1761 0 : .persistence
1762 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1763 0 : .await?;
1764 :
1765 : // Persist the placement policy update. This is required
1766 : // when we reattaching a detached tenant.
1767 0 : self.persistence
1768 0 : .update_tenant_shard(
1769 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1770 0 : Some(PlacementPolicy::Attached(0)),
1771 0 : Some(conf),
1772 0 : None,
1773 0 : None,
1774 0 : )
1775 0 : .await?;
1776 0 : Some(new_generation)
1777 : }
1778 : None => {
1779 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1780 : }
1781 : }
1782 : } else {
1783 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1784 0 : None
1785 : };
1786 :
1787 0 : let mut locked = self.inner.write().unwrap();
1788 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1789 0 :
1790 0 : let tenant_shard = tenants
1791 0 : .get_mut(&attach_req.tenant_shard_id)
1792 0 : .expect("Checked for existence above");
1793 :
1794 0 : if let Some(new_generation) = new_generation {
1795 0 : tenant_shard.generation = Some(new_generation);
1796 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1797 0 : } else {
1798 : // This is a detach notification. We must update placement policy to avoid re-attaching
1799 : // during background scheduling/reconciliation, or during storage controller restart.
1800 0 : assert!(attach_req.node_id.is_none());
1801 0 : tenant_shard.policy = PlacementPolicy::Detached;
1802 : }
1803 :
1804 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1805 0 : tracing::info!(
1806 : tenant_id = %attach_req.tenant_shard_id,
1807 : ps_id = %attaching_pageserver,
1808 : generation = ?tenant_shard.generation,
1809 0 : "issuing",
1810 : );
1811 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1812 0 : tracing::info!(
1813 : tenant_id = %attach_req.tenant_shard_id,
1814 : %ps_id,
1815 : generation = ?tenant_shard.generation,
1816 0 : "dropping",
1817 : );
1818 : } else {
1819 0 : tracing::info!(
1820 : tenant_id = %attach_req.tenant_shard_id,
1821 0 : "no-op: tenant already has no pageserver");
1822 : }
1823 0 : tenant_shard
1824 0 : .intent
1825 0 : .set_attached(scheduler, attach_req.node_id);
1826 0 :
1827 0 : tracing::info!(
1828 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1829 0 : attach_req.tenant_shard_id,
1830 0 : tenant_shard.generation,
1831 0 : // TODO: this is an odd number of 0xf's
1832 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1833 : );
1834 :
1835 : // Trick the reconciler into not doing anything for this tenant: this helps
1836 : // tests that manually configure a tenant on the pagesrever, and then call this
1837 : // attach hook: they don't want background reconciliation to modify what they
1838 : // did to the pageserver.
1839 : #[cfg(feature = "testing")]
1840 : {
1841 0 : if let Some(node_id) = attach_req.node_id {
1842 0 : tenant_shard.observed.locations = HashMap::from([(
1843 0 : node_id,
1844 0 : ObservedStateLocation {
1845 0 : conf: Some(attached_location_conf(
1846 0 : tenant_shard.generation.unwrap(),
1847 0 : &tenant_shard.shard,
1848 0 : &tenant_shard.config,
1849 0 : &PlacementPolicy::Attached(0),
1850 0 : )),
1851 0 : },
1852 0 : )]);
1853 0 : } else {
1854 0 : tenant_shard.observed.locations.clear();
1855 0 : }
1856 : }
1857 :
1858 0 : Ok(AttachHookResponse {
1859 0 : generation: attach_req
1860 0 : .node_id
1861 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1862 0 : })
1863 0 : }
1864 :
1865 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1866 0 : let locked = self.inner.read().unwrap();
1867 0 :
1868 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1869 0 :
1870 0 : InspectResponse {
1871 0 : attachment: tenant_shard.and_then(|s| {
1872 0 : s.intent
1873 0 : .get_attached()
1874 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1875 0 : }),
1876 0 : }
1877 0 : }
1878 :
1879 : // When the availability state of a node transitions to active, we must do a full reconciliation
1880 : // of LocationConfigs on that node. This is because while a node was offline:
1881 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1882 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1883 : //
1884 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1885 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1886 : // this function.
1887 : //
1888 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1889 : // for written for a single node rather than as a batch job for all nodes.
1890 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1891 : async fn node_activate_reconcile(
1892 : &self,
1893 : mut node: Node,
1894 : _lock: &TracingExclusiveGuard<NodeOperations>,
1895 : ) -> Result<(), ApiError> {
1896 : // This Node is a mutable local copy: we will set it active so that we can use its
1897 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1898 : // later.
1899 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1900 :
1901 : let configs = match node
1902 : .with_client_retries(
1903 0 : |client| async move { client.list_location_config().await },
1904 : &self.config.pageserver_jwt_token,
1905 : 1,
1906 : 5,
1907 : SHORT_RECONCILE_TIMEOUT,
1908 : &self.cancel,
1909 : )
1910 : .await
1911 : {
1912 : None => {
1913 : // We're shutting down (the Node's cancellation token can't have fired, because
1914 : // we're the only scope that has a reference to it, and we didn't fire it).
1915 : return Err(ApiError::ShuttingDown);
1916 : }
1917 : Some(Err(e)) => {
1918 : // This node didn't succeed listing its locations: it may not proceed to active state
1919 : // as it is apparently unavailable.
1920 : return Err(ApiError::PreconditionFailed(
1921 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1922 : ));
1923 : }
1924 : Some(Ok(configs)) => configs,
1925 : };
1926 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1927 :
1928 : let mut cleanup = Vec::new();
1929 : {
1930 : let mut locked = self.inner.write().unwrap();
1931 :
1932 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1933 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1934 : cleanup.push(tenant_shard_id);
1935 : continue;
1936 : };
1937 : tenant_shard
1938 : .observed
1939 : .locations
1940 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1941 : }
1942 : }
1943 :
1944 : for tenant_shard_id in cleanup {
1945 : tracing::info!("Detaching {tenant_shard_id}");
1946 : match node
1947 : .with_client_retries(
1948 0 : |client| async move {
1949 0 : let config = LocationConfig {
1950 0 : mode: LocationConfigMode::Detached,
1951 0 : generation: None,
1952 0 : secondary_conf: None,
1953 0 : shard_number: tenant_shard_id.shard_number.0,
1954 0 : shard_count: tenant_shard_id.shard_count.literal(),
1955 0 : shard_stripe_size: 0,
1956 0 : tenant_conf: models::TenantConfig::default(),
1957 0 : };
1958 0 : client
1959 0 : .location_config(tenant_shard_id, config, None, false)
1960 0 : .await
1961 0 : },
1962 : &self.config.pageserver_jwt_token,
1963 : 1,
1964 : 5,
1965 : SHORT_RECONCILE_TIMEOUT,
1966 : &self.cancel,
1967 : )
1968 : .await
1969 : {
1970 : None => {
1971 : // We're shutting down (the Node's cancellation token can't have fired, because
1972 : // we're the only scope that has a reference to it, and we didn't fire it).
1973 : return Err(ApiError::ShuttingDown);
1974 : }
1975 : Some(Err(e)) => {
1976 : // Do not let the node proceed to Active state if it is not responsive to requests
1977 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1978 : // detach completing: we should not let this node back into the set of nodes considered
1979 : // okay for scheduling.
1980 : return Err(ApiError::Conflict(format!(
1981 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1982 : )));
1983 : }
1984 : Some(Ok(_)) => {}
1985 : };
1986 : }
1987 :
1988 : Ok(())
1989 : }
1990 :
1991 0 : pub(crate) async fn re_attach(
1992 0 : &self,
1993 0 : reattach_req: ReAttachRequest,
1994 0 : ) -> Result<ReAttachResponse, ApiError> {
1995 0 : if let Some(register_req) = reattach_req.register {
1996 0 : self.node_register(register_req).await?;
1997 0 : }
1998 :
1999 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2000 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2001 :
2002 0 : tracing::info!(
2003 : node_id=%reattach_req.node_id,
2004 0 : "Incremented {} tenant shards' generations",
2005 0 : incremented_generations.len()
2006 : );
2007 :
2008 : // Apply the updated generation to our in-memory state, and
2009 : // gather discover secondary locations.
2010 0 : let mut locked = self.inner.write().unwrap();
2011 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2012 0 :
2013 0 : let mut response = ReAttachResponse {
2014 0 : tenants: Vec::new(),
2015 0 : };
2016 :
2017 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2018 : // to call location_conf API with an old generation. Wait for cancellation to complete
2019 : // before responding to this request. Requires well implemented CancellationToken logic
2020 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2021 : // request in flight over the network: TODO handle that by making location_conf API refuse
2022 : // to go backward in generations.
2023 :
2024 : // Scan through all shards, applying updates for ones where we updated generation
2025 : // and identifying shards that intend to have a secondary location on this node.
2026 0 : for (tenant_shard_id, shard) in tenants {
2027 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2028 0 : let new_gen = *new_gen;
2029 0 : response.tenants.push(ReAttachResponseTenant {
2030 0 : id: *tenant_shard_id,
2031 0 : r#gen: Some(new_gen.into().unwrap()),
2032 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2033 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2034 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2035 0 : // the stale/multi states at this point.
2036 0 : mode: LocationConfigMode::AttachedSingle,
2037 0 : });
2038 0 :
2039 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2040 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2041 : // Why can we update `observed` even though we're not sure our response will be received
2042 : // by the pageserver? Because the pageserver will not proceed with startup until
2043 : // it has processed response: if it loses it, we'll see another request and increment
2044 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2045 0 : if let Some(conf) = observed.conf.as_mut() {
2046 0 : conf.generation = new_gen.into();
2047 0 : }
2048 0 : } else {
2049 0 : // This node has no observed state for the shard: perhaps it was offline
2050 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2051 0 : // will be prompted to learn the location's state before it makes changes.
2052 0 : shard
2053 0 : .observed
2054 0 : .locations
2055 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2056 0 : }
2057 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2058 0 : // Ordering: pageserver will not accept /location_config requests until it has
2059 0 : // finished processing the response from re-attach. So we can update our in-memory state
2060 0 : // now, and be confident that we are not stamping on the result of some later location config.
2061 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2062 0 : // so we might update observed state here, and then get over-written by some racing
2063 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2064 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2065 0 :
2066 0 : response.tenants.push(ReAttachResponseTenant {
2067 0 : id: *tenant_shard_id,
2068 0 : r#gen: None,
2069 0 : mode: LocationConfigMode::Secondary,
2070 0 : });
2071 0 :
2072 0 : // We must not update observed, because we have no guarantee that our
2073 0 : // response will be received by the pageserver. This could leave it
2074 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2075 0 : }
2076 : }
2077 :
2078 : // We consider a node Active once we have composed a re-attach response, but we
2079 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2080 : // implicitly synchronizes the LocationConfigs on the node.
2081 : //
2082 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2083 : // but those requests will not be accepted by the node until it has finished processing
2084 : // the re-attach response.
2085 : //
2086 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2087 : // in [`Persistence::re_attach`].
2088 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2089 0 : let reset_scheduling = matches!(
2090 0 : node.get_scheduling(),
2091 : NodeSchedulingPolicy::PauseForRestart
2092 : | NodeSchedulingPolicy::Draining
2093 : | NodeSchedulingPolicy::Filling
2094 : );
2095 :
2096 0 : let mut new_nodes = (**nodes).clone();
2097 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2098 0 : if reset_scheduling {
2099 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2100 0 : }
2101 :
2102 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2103 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2104 0 :
2105 0 : scheduler.node_upsert(node);
2106 0 : let new_nodes = Arc::new(new_nodes);
2107 0 : *nodes = new_nodes;
2108 : } else {
2109 0 : tracing::error!(
2110 0 : "Reattaching node {} was removed while processing the request",
2111 : reattach_req.node_id
2112 : );
2113 : }
2114 0 : }
2115 :
2116 0 : Ok(response)
2117 0 : }
2118 :
2119 0 : pub(crate) async fn validate(
2120 0 : &self,
2121 0 : validate_req: ValidateRequest,
2122 0 : ) -> Result<ValidateResponse, DatabaseError> {
2123 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2124 : // in-memory generation for a shard
2125 0 : let in_memory_result = {
2126 0 : let mut in_memory_result = Vec::new();
2127 0 : let locked = self.inner.read().unwrap();
2128 0 : for req_tenant in validate_req.tenants {
2129 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2130 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2131 0 : tracing::info!(
2132 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2133 : req_tenant.id,
2134 : req_tenant.r#gen,
2135 : tenant_shard.generation
2136 : );
2137 :
2138 0 : in_memory_result.push((
2139 0 : req_tenant.id,
2140 0 : Generation::new(req_tenant.r#gen),
2141 0 : valid,
2142 0 : ));
2143 : } else {
2144 : // This is legal: for example during a shard split the pageserver may still
2145 : // have deletions in its queue from the old pre-split shard, or after deletion
2146 : // of a tenant that was busy with compaction/gc while being deleted.
2147 0 : tracing::info!(
2148 0 : "Refusing deletion validation for missing shard {}",
2149 : req_tenant.id
2150 : );
2151 : }
2152 : }
2153 :
2154 0 : in_memory_result
2155 : };
2156 :
2157 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2158 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2159 0 : let db_generations = self
2160 0 : .persistence
2161 0 : .shard_generations(
2162 0 : in_memory_result
2163 0 : .iter()
2164 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2165 0 : )
2166 0 : .await?;
2167 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2168 0 :
2169 0 : let mut response = ValidateResponse {
2170 0 : tenants: Vec::new(),
2171 0 : };
2172 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2173 0 : let valid = if valid {
2174 0 : let db_generation = db_generations.get(&tenant_shard_id);
2175 0 : db_generation == Some(&Some(validate_generation))
2176 : } else {
2177 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2178 : // this prevents a pageserver from cleaning up an object in S3.
2179 0 : false
2180 : };
2181 :
2182 0 : response.tenants.push(ValidateResponseTenant {
2183 0 : id: tenant_shard_id,
2184 0 : valid,
2185 0 : })
2186 : }
2187 :
2188 0 : Ok(response)
2189 0 : }
2190 :
2191 0 : pub(crate) async fn tenant_create(
2192 0 : &self,
2193 0 : create_req: TenantCreateRequest,
2194 0 : ) -> Result<TenantCreateResponse, ApiError> {
2195 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2196 :
2197 : // Exclude any concurrent attempts to create/access the same tenant ID
2198 0 : let _tenant_lock = trace_exclusive_lock(
2199 0 : &self.tenant_op_locks,
2200 0 : create_req.new_tenant_id.tenant_id,
2201 0 : TenantOperations::Create,
2202 0 : )
2203 0 : .await;
2204 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2205 :
2206 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2207 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2208 : // accept compute notifications while it is in the process of creating. Reconciliation will
2209 : // be retried in the background.
2210 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2211 0 : }
2212 0 : Ok(response)
2213 0 : }
2214 :
2215 0 : pub(crate) async fn do_tenant_create(
2216 0 : &self,
2217 0 : create_req: TenantCreateRequest,
2218 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2219 0 : let placement_policy = create_req
2220 0 : .placement_policy
2221 0 : .clone()
2222 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2223 0 : .unwrap_or(PlacementPolicy::Attached(0));
2224 :
2225 : // This service expects to handle sharding itself: it is an error to try and directly create
2226 : // a particular shard here.
2227 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2228 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2229 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2230 0 : )));
2231 : } else {
2232 0 : create_req.new_tenant_id.tenant_id
2233 0 : };
2234 0 :
2235 0 : tracing::info!(
2236 0 : "Creating tenant {}, shard_count={:?}",
2237 : create_req.new_tenant_id,
2238 : create_req.shard_parameters.count,
2239 : );
2240 :
2241 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2242 0 : .map(|i| TenantShardId {
2243 0 : tenant_id,
2244 0 : shard_number: ShardNumber(i),
2245 0 : shard_count: create_req.shard_parameters.count,
2246 0 : })
2247 0 : .collect::<Vec<_>>();
2248 :
2249 : // If the caller specifies a None generation, it means "start from default". This is different
2250 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2251 : // an incompletely-onboarded tenant.
2252 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2253 0 : tracing::info!(
2254 0 : "tenant_create: secondary mode, generation is_some={}",
2255 0 : create_req.generation.is_some()
2256 : );
2257 0 : create_req.generation.map(Generation::new)
2258 : } else {
2259 0 : tracing::info!(
2260 0 : "tenant_create: not secondary mode, generation is_some={}",
2261 0 : create_req.generation.is_some()
2262 : );
2263 0 : Some(
2264 0 : create_req
2265 0 : .generation
2266 0 : .map(Generation::new)
2267 0 : .unwrap_or(INITIAL_GENERATION),
2268 0 : )
2269 : };
2270 :
2271 0 : let preferred_az_id = {
2272 0 : let locked = self.inner.read().unwrap();
2273 : // Idempotency: take the existing value if the tenant already exists
2274 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2275 0 : shard.preferred_az().cloned()
2276 : } else {
2277 0 : locked.scheduler.get_az_for_new_tenant()
2278 : }
2279 : };
2280 :
2281 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2282 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2283 : // during the creation, rather than risking leaving orphan objects in S3.
2284 0 : let persist_tenant_shards = create_ids
2285 0 : .iter()
2286 0 : .map(|tenant_shard_id| TenantShardPersistence {
2287 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2288 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2289 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2290 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2291 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2292 0 : // The pageserver is not known until scheduling happens: we will set this column when
2293 0 : // incrementing the generation the first time we attach to a pageserver.
2294 0 : generation_pageserver: None,
2295 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2296 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2297 0 : splitting: SplitState::default(),
2298 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2299 0 : .unwrap(),
2300 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2301 0 : })
2302 0 : .collect();
2303 0 :
2304 0 : match self
2305 0 : .persistence
2306 0 : .insert_tenant_shards(persist_tenant_shards)
2307 0 : .await
2308 : {
2309 0 : Ok(_) => {}
2310 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2311 : DatabaseErrorKind::UniqueViolation,
2312 : _,
2313 : ))) => {
2314 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2315 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2316 : // creation's shard count.
2317 0 : tracing::info!(
2318 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2319 : );
2320 : }
2321 : // Any other database error is unexpected and a bug.
2322 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2323 : };
2324 :
2325 0 : let mut schedule_context = ScheduleContext::default();
2326 0 : let mut schedule_error = None;
2327 0 : let mut response_shards = Vec::new();
2328 0 : for tenant_shard_id in create_ids {
2329 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2330 :
2331 0 : let outcome = self
2332 0 : .do_initial_shard_scheduling(
2333 0 : tenant_shard_id,
2334 0 : initial_generation,
2335 0 : &create_req.shard_parameters,
2336 0 : create_req.config.clone(),
2337 0 : placement_policy.clone(),
2338 0 : preferred_az_id.as_ref(),
2339 0 : &mut schedule_context,
2340 0 : )
2341 0 : .await;
2342 :
2343 0 : match outcome {
2344 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2345 0 : InitialShardScheduleOutcome::NotScheduled => {}
2346 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2347 0 : schedule_error = Some(err);
2348 0 : }
2349 : }
2350 : }
2351 :
2352 : // If we failed to schedule shards, then they are still created in the controller,
2353 : // but we return an error to the requester to avoid a silent failure when someone
2354 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2355 : // are present in the system. We do this here rather than in the above loop, to
2356 : // avoid situations where we only create a subset of shards in the tenant.
2357 0 : if let Some(e) = schedule_error {
2358 0 : return Err(ApiError::Conflict(format!(
2359 0 : "Failed to schedule shard(s): {e}"
2360 0 : )));
2361 0 : }
2362 0 :
2363 0 : let waiters = {
2364 0 : let mut locked = self.inner.write().unwrap();
2365 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2366 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2367 0 : .tenant_creation_hint(true)
2368 0 : .build();
2369 0 : tenants
2370 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2371 0 : .filter_map(|(_shard_id, shard)| {
2372 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2373 0 : })
2374 0 : .collect::<Vec<_>>()
2375 0 : };
2376 0 :
2377 0 : Ok((
2378 0 : TenantCreateResponse {
2379 0 : shards: response_shards,
2380 0 : },
2381 0 : waiters,
2382 0 : ))
2383 0 : }
2384 :
2385 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2386 : /// case of a new tenant and a pre-existing one.
2387 : #[allow(clippy::too_many_arguments)]
2388 0 : async fn do_initial_shard_scheduling(
2389 0 : &self,
2390 0 : tenant_shard_id: TenantShardId,
2391 0 : initial_generation: Option<Generation>,
2392 0 : shard_params: &ShardParameters,
2393 0 : config: TenantConfig,
2394 0 : placement_policy: PlacementPolicy,
2395 0 : preferred_az_id: Option<&AvailabilityZone>,
2396 0 : schedule_context: &mut ScheduleContext,
2397 0 : ) -> InitialShardScheduleOutcome {
2398 0 : let mut locked = self.inner.write().unwrap();
2399 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2400 :
2401 : use std::collections::btree_map::Entry;
2402 0 : match tenants.entry(tenant_shard_id) {
2403 0 : Entry::Occupied(mut entry) => {
2404 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2405 :
2406 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2407 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2408 0 : }
2409 :
2410 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2411 0 : let generation = entry
2412 0 : .get()
2413 0 : .generation
2414 0 : .expect("Generation is set when in attached mode");
2415 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2416 0 : shard_id: tenant_shard_id,
2417 0 : node_id: *node_id,
2418 0 : generation: generation.into().unwrap(),
2419 0 : })
2420 : } else {
2421 0 : InitialShardScheduleOutcome::NotScheduled
2422 : }
2423 : }
2424 0 : Entry::Vacant(entry) => {
2425 0 : let state = entry.insert(TenantShard::new(
2426 0 : tenant_shard_id,
2427 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2428 0 : placement_policy,
2429 0 : preferred_az_id.cloned(),
2430 0 : ));
2431 0 :
2432 0 : state.generation = initial_generation;
2433 0 : state.config = config;
2434 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2435 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2436 0 : }
2437 :
2438 : // Only include shards in result if we are attaching: the purpose
2439 : // of the response is to tell the caller where the shards are attached.
2440 0 : if let Some(node_id) = state.intent.get_attached() {
2441 0 : let generation = state
2442 0 : .generation
2443 0 : .expect("Generation is set when in attached mode");
2444 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2445 0 : shard_id: tenant_shard_id,
2446 0 : node_id: *node_id,
2447 0 : generation: generation.into().unwrap(),
2448 0 : })
2449 : } else {
2450 0 : InitialShardScheduleOutcome::NotScheduled
2451 : }
2452 : }
2453 : }
2454 0 : }
2455 :
2456 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2457 : /// wait for reconciliation to complete before responding.
2458 0 : async fn await_waiters(
2459 0 : &self,
2460 0 : waiters: Vec<ReconcilerWaiter>,
2461 0 : timeout: Duration,
2462 0 : ) -> Result<(), ReconcileWaitError> {
2463 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2464 0 : for waiter in waiters {
2465 0 : let timeout = deadline.duration_since(Instant::now());
2466 0 : waiter.wait_timeout(timeout).await?;
2467 : }
2468 :
2469 0 : Ok(())
2470 0 : }
2471 :
2472 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2473 : /// in progress
2474 0 : async fn await_waiters_remainder(
2475 0 : &self,
2476 0 : waiters: Vec<ReconcilerWaiter>,
2477 0 : timeout: Duration,
2478 0 : ) -> Vec<ReconcilerWaiter> {
2479 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2480 0 : for waiter in waiters.iter() {
2481 0 : let timeout = deadline.duration_since(Instant::now());
2482 0 : let _ = waiter.wait_timeout(timeout).await;
2483 : }
2484 :
2485 0 : waiters
2486 0 : .into_iter()
2487 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2488 0 : .collect::<Vec<_>>()
2489 0 : }
2490 :
2491 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2492 : /// and transform it into either a tenant creation of a series of shard updates.
2493 : ///
2494 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2495 : /// still be returned.
2496 0 : fn tenant_location_config_prepare(
2497 0 : &self,
2498 0 : tenant_id: TenantId,
2499 0 : req: TenantLocationConfigRequest,
2500 0 : ) -> TenantCreateOrUpdate {
2501 0 : let mut updates = Vec::new();
2502 0 : let mut locked = self.inner.write().unwrap();
2503 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2504 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2505 :
2506 : // Use location config mode as an indicator of policy.
2507 0 : let placement_policy = match req.config.mode {
2508 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2509 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2510 : LocationConfigMode::AttachedMulti
2511 : | LocationConfigMode::AttachedSingle
2512 : | LocationConfigMode::AttachedStale => {
2513 0 : if nodes.len() > 1 {
2514 0 : PlacementPolicy::Attached(1)
2515 : } else {
2516 : // Convenience for dev/test: if we just have one pageserver, import
2517 : // tenants into non-HA mode so that scheduling will succeed.
2518 0 : PlacementPolicy::Attached(0)
2519 : }
2520 : }
2521 : };
2522 :
2523 : // Ordinarily we do not update scheduling policy, but when making major changes
2524 : // like detaching or demoting to secondary-only, we need to force the scheduling
2525 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2526 0 : let scheduling_policy = match req.config.mode {
2527 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2528 : // Special case: when making major changes like detaching or demoting to secondary-only,
2529 : // we need to force the scheduling mode to Active, or nothing will happen.
2530 0 : Some(ShardSchedulingPolicy::Active)
2531 : }
2532 : LocationConfigMode::AttachedMulti
2533 : | LocationConfigMode::AttachedSingle
2534 : | LocationConfigMode::AttachedStale => {
2535 : // While attached, continue to respect whatever the existing scheduling mode is.
2536 0 : None
2537 : }
2538 : };
2539 :
2540 0 : let mut create = true;
2541 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2542 : // Saw an existing shard: this is not a creation
2543 0 : create = false;
2544 :
2545 : // Shards may have initially been created by a Secondary request, where we
2546 : // would have left generation as None.
2547 : //
2548 : // We only update generation the first time we see an attached-mode request,
2549 : // and if there is no existing generation set. The caller is responsible for
2550 : // ensuring that no non-storage-controller pageserver ever uses a higher
2551 : // generation than they passed in here.
2552 : use LocationConfigMode::*;
2553 0 : let set_generation = match req.config.mode {
2554 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2555 0 : req.config.generation.map(Generation::new)
2556 : }
2557 0 : _ => None,
2558 : };
2559 :
2560 0 : updates.push(ShardUpdate {
2561 0 : tenant_shard_id: *shard_id,
2562 0 : placement_policy: placement_policy.clone(),
2563 0 : tenant_config: req.config.tenant_conf.clone(),
2564 0 : generation: set_generation,
2565 0 : scheduling_policy,
2566 0 : });
2567 : }
2568 :
2569 0 : if create {
2570 : use LocationConfigMode::*;
2571 0 : let generation = match req.config.mode {
2572 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2573 : // If a caller provided a generation in a non-attached request, ignore it
2574 : // and leave our generation as None: this enables a subsequent update to set
2575 : // the generation when setting an attached mode for the first time.
2576 0 : _ => None,
2577 : };
2578 :
2579 0 : TenantCreateOrUpdate::Create(
2580 0 : // Synthesize a creation request
2581 0 : TenantCreateRequest {
2582 0 : new_tenant_id: tenant_shard_id,
2583 0 : generation,
2584 0 : shard_parameters: ShardParameters {
2585 0 : count: tenant_shard_id.shard_count,
2586 0 : // We only import un-sharded or single-sharded tenants, so stripe
2587 0 : // size can be made up arbitrarily here.
2588 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2589 0 : },
2590 0 : placement_policy: Some(placement_policy),
2591 0 : config: req.config.tenant_conf,
2592 0 : },
2593 0 : )
2594 : } else {
2595 0 : assert!(!updates.is_empty());
2596 0 : TenantCreateOrUpdate::Update(updates)
2597 : }
2598 0 : }
2599 :
2600 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2601 : /// the tenant is present in memory. If not, load it from the database. If it is found
2602 : /// in neither location, return a NotFound error.
2603 : ///
2604 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2605 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2606 0 : async fn maybe_load_tenant(
2607 0 : &self,
2608 0 : tenant_id: TenantId,
2609 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2610 0 : ) -> Result<(), ApiError> {
2611 : // Check if the tenant is present in memory, and select an AZ to use when loading
2612 : // if we will load it.
2613 0 : let load_in_az = {
2614 0 : let locked = self.inner.read().unwrap();
2615 0 : let existing = locked
2616 0 : .tenants
2617 0 : .range(TenantShardId::tenant_range(tenant_id))
2618 0 : .next();
2619 0 :
2620 0 : // If the tenant is not present in memory, we expect to load it from database,
2621 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2622 0 : if existing.is_none() {
2623 0 : locked
2624 0 : .scheduler
2625 0 : .get_az_for_new_tenant()
2626 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2627 0 : "No AZ with nodes found to load tenant"
2628 0 : )))?
2629 : } else {
2630 : // We already have this tenant in memory
2631 0 : return Ok(());
2632 : }
2633 : };
2634 :
2635 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2636 0 : if tenant_shards.is_empty() {
2637 0 : return Err(ApiError::NotFound(
2638 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2639 0 : ));
2640 0 : }
2641 0 :
2642 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2643 0 : self.persistence
2644 0 : .set_tenant_shard_preferred_azs(
2645 0 : tenant_shards
2646 0 : .iter()
2647 0 : .map(|t| {
2648 0 : (
2649 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2650 0 : Some(load_in_az.clone()),
2651 0 : )
2652 0 : })
2653 0 : .collect(),
2654 0 : )
2655 0 : .await?;
2656 :
2657 0 : let mut locked = self.inner.write().unwrap();
2658 0 : tracing::info!(
2659 0 : "Loaded {} shards for tenant {}",
2660 0 : tenant_shards.len(),
2661 : tenant_id
2662 : );
2663 :
2664 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2665 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2666 0 : let shard =
2667 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2668 0 :
2669 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2670 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2671 0 : if shard.policy != PlacementPolicy::Detached {
2672 0 : tracing::error!(
2673 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2674 : shard.tenant_shard_id,
2675 : shard.policy
2676 : );
2677 0 : }
2678 :
2679 0 : (shard.tenant_shard_id, shard)
2680 0 : }));
2681 0 :
2682 0 : Ok(())
2683 0 : }
2684 :
2685 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2686 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2687 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2688 : ///
2689 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2690 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2691 0 : fn maybe_drop_tenant(
2692 0 : &self,
2693 0 : tenant_id: TenantId,
2694 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2695 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2696 0 : ) {
2697 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2698 0 : if tenant_shards.all(|(_id, shard)| {
2699 0 : shard.policy == PlacementPolicy::Detached
2700 0 : && shard.reconciler.is_none()
2701 0 : && shard.observed.is_empty()
2702 0 : }) {
2703 0 : let keys = locked
2704 0 : .tenants
2705 0 : .range(TenantShardId::tenant_range(tenant_id))
2706 0 : .map(|(id, _)| id)
2707 0 : .copied()
2708 0 : .collect::<Vec<_>>();
2709 0 : for key in keys {
2710 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2711 0 : locked.tenants.remove(&key);
2712 : }
2713 0 : }
2714 0 : }
2715 :
2716 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2717 : /// directly with pageservers into this service.
2718 : ///
2719 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2720 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2721 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2722 : /// tenant's source of generation numbers.
2723 : ///
2724 : /// The mode in this request coarse-grained control of tenants:
2725 : /// - Call with mode Attached* to upsert the tenant.
2726 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2727 : /// to set an existing tenant to PolicyMode::Secondary
2728 : /// - Call with mode Detached to switch to PolicyMode::Detached
2729 0 : pub(crate) async fn tenant_location_config(
2730 0 : &self,
2731 0 : tenant_shard_id: TenantShardId,
2732 0 : req: TenantLocationConfigRequest,
2733 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2734 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2735 0 : let _tenant_lock = trace_exclusive_lock(
2736 0 : &self.tenant_op_locks,
2737 0 : tenant_shard_id.tenant_id,
2738 0 : TenantOperations::LocationConfig,
2739 0 : )
2740 0 : .await;
2741 :
2742 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2743 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2744 0 : "This API is for importing single-sharded or unsharded tenants"
2745 0 : )));
2746 : } else {
2747 0 : tenant_shard_id.tenant_id
2748 0 : };
2749 0 :
2750 0 : // In case we are waking up a Detached tenant
2751 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2752 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2753 0 : // This is a creation or an update
2754 0 : }
2755 0 : Err(e) => {
2756 0 : return Err(e);
2757 : }
2758 : };
2759 :
2760 : // First check if this is a creation or an update
2761 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2762 0 :
2763 0 : let mut result = TenantLocationConfigResponse {
2764 0 : shards: Vec::new(),
2765 0 : stripe_size: None,
2766 0 : };
2767 0 : let waiters = match create_or_update {
2768 0 : TenantCreateOrUpdate::Create(create_req) => {
2769 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2770 0 : result.shards = create_resp
2771 0 : .shards
2772 0 : .into_iter()
2773 0 : .map(|s| TenantShardLocation {
2774 0 : node_id: s.node_id,
2775 0 : shard_id: s.shard_id,
2776 0 : })
2777 0 : .collect();
2778 0 : waiters
2779 : }
2780 0 : TenantCreateOrUpdate::Update(updates) => {
2781 0 : // Persist updates
2782 0 : // Ordering: write to the database before applying changes in-memory, so that
2783 0 : // we will not appear time-travel backwards on a restart.
2784 0 :
2785 0 : let mut schedule_context = ScheduleContext::default();
2786 : for ShardUpdate {
2787 0 : tenant_shard_id,
2788 0 : placement_policy,
2789 0 : tenant_config,
2790 0 : generation,
2791 0 : scheduling_policy,
2792 0 : } in &updates
2793 : {
2794 0 : self.persistence
2795 0 : .update_tenant_shard(
2796 0 : TenantFilter::Shard(*tenant_shard_id),
2797 0 : Some(placement_policy.clone()),
2798 0 : Some(tenant_config.clone()),
2799 0 : *generation,
2800 0 : *scheduling_policy,
2801 0 : )
2802 0 : .await?;
2803 : }
2804 :
2805 : // Apply updates in-memory
2806 0 : let mut waiters = Vec::new();
2807 0 : {
2808 0 : let mut locked = self.inner.write().unwrap();
2809 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2810 :
2811 : for ShardUpdate {
2812 0 : tenant_shard_id,
2813 0 : placement_policy,
2814 0 : tenant_config,
2815 0 : generation: update_generation,
2816 0 : scheduling_policy,
2817 0 : } in updates
2818 : {
2819 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2820 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2821 0 : continue;
2822 : };
2823 :
2824 : // Update stripe size
2825 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2826 0 : result.stripe_size = Some(shard.shard.stripe_size);
2827 0 : }
2828 :
2829 0 : shard.policy = placement_policy;
2830 0 : shard.config = tenant_config;
2831 0 : if let Some(generation) = update_generation {
2832 0 : shard.generation = Some(generation);
2833 0 : }
2834 :
2835 0 : if let Some(scheduling_policy) = scheduling_policy {
2836 0 : shard.set_scheduling_policy(scheduling_policy);
2837 0 : }
2838 :
2839 0 : shard.schedule(scheduler, &mut schedule_context)?;
2840 :
2841 0 : let maybe_waiter =
2842 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
2843 0 : if let Some(waiter) = maybe_waiter {
2844 0 : waiters.push(waiter);
2845 0 : }
2846 :
2847 0 : if let Some(node_id) = shard.intent.get_attached() {
2848 0 : result.shards.push(TenantShardLocation {
2849 0 : shard_id: tenant_shard_id,
2850 0 : node_id: *node_id,
2851 0 : })
2852 0 : }
2853 : }
2854 : }
2855 0 : waiters
2856 : }
2857 : };
2858 :
2859 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2860 : // Do not treat a reconcile error as fatal: we have already applied any requested
2861 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2862 : // compute notification API. In these cases, it is important that we do not
2863 : // cause the cloud control plane to retry forever on this API.
2864 0 : tracing::warn!(
2865 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2866 : );
2867 0 : }
2868 :
2869 : // Logging the full result is useful because it lets us cross-check what the cloud control
2870 : // plane's tenant_shards table should contain.
2871 0 : tracing::info!("Complete, returning {result:?}");
2872 :
2873 0 : Ok(result)
2874 0 : }
2875 :
2876 0 : pub(crate) async fn tenant_config_patch(
2877 0 : &self,
2878 0 : req: TenantConfigPatchRequest,
2879 0 : ) -> Result<(), ApiError> {
2880 0 : let _tenant_lock = trace_exclusive_lock(
2881 0 : &self.tenant_op_locks,
2882 0 : req.tenant_id,
2883 0 : TenantOperations::ConfigPatch,
2884 0 : )
2885 0 : .await;
2886 :
2887 0 : let tenant_id = req.tenant_id;
2888 0 : let patch = req.config;
2889 0 :
2890 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
2891 :
2892 0 : let base = {
2893 0 : let locked = self.inner.read().unwrap();
2894 0 : let shards = locked
2895 0 : .tenants
2896 0 : .range(TenantShardId::tenant_range(req.tenant_id));
2897 0 :
2898 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
2899 :
2900 0 : let first = match configs.peek() {
2901 0 : Some(first) => (*first).clone(),
2902 : None => {
2903 0 : return Err(ApiError::NotFound(
2904 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
2905 0 : ));
2906 : }
2907 : };
2908 :
2909 0 : if !configs.all_equal() {
2910 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
2911 : // This can't happen because we atomically update the database records
2912 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
2913 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2914 0 : "Tenant configs for {} are mismatched",
2915 0 : req.tenant_id
2916 0 : )));
2917 0 : }
2918 0 :
2919 0 : first
2920 : };
2921 :
2922 0 : let updated_config = base
2923 0 : .apply_patch(patch)
2924 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
2925 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
2926 0 : .await
2927 0 : }
2928 :
2929 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2930 : // We require an exclusive lock, because we are updating persistent and in-memory state
2931 0 : let _tenant_lock = trace_exclusive_lock(
2932 0 : &self.tenant_op_locks,
2933 0 : req.tenant_id,
2934 0 : TenantOperations::ConfigSet,
2935 0 : )
2936 0 : .await;
2937 :
2938 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
2939 :
2940 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
2941 0 : .await
2942 0 : }
2943 :
2944 0 : async fn set_tenant_config_and_reconcile(
2945 0 : &self,
2946 0 : tenant_id: TenantId,
2947 0 : config: TenantConfig,
2948 0 : ) -> Result<(), ApiError> {
2949 0 : self.persistence
2950 0 : .update_tenant_shard(
2951 0 : TenantFilter::Tenant(tenant_id),
2952 0 : None,
2953 0 : Some(config.clone()),
2954 0 : None,
2955 0 : None,
2956 0 : )
2957 0 : .await?;
2958 :
2959 0 : let waiters = {
2960 0 : let mut waiters = Vec::new();
2961 0 : let mut locked = self.inner.write().unwrap();
2962 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2963 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2964 0 : shard.config = config.clone();
2965 0 : if let Some(waiter) =
2966 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
2967 0 : {
2968 0 : waiters.push(waiter);
2969 0 : }
2970 : }
2971 0 : waiters
2972 : };
2973 :
2974 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2975 : // Treat this as success because we have stored the configuration. If e.g.
2976 : // a node was unavailable at this time, it should not stop us accepting a
2977 : // configuration change.
2978 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2979 0 : }
2980 :
2981 0 : Ok(())
2982 0 : }
2983 :
2984 0 : pub(crate) fn tenant_config_get(
2985 0 : &self,
2986 0 : tenant_id: TenantId,
2987 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2988 0 : let config = {
2989 0 : let locked = self.inner.read().unwrap();
2990 0 :
2991 0 : match locked
2992 0 : .tenants
2993 0 : .range(TenantShardId::tenant_range(tenant_id))
2994 0 : .next()
2995 : {
2996 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2997 : None => {
2998 0 : return Err(ApiError::NotFound(
2999 0 : anyhow::anyhow!("Tenant not found").into(),
3000 0 : ));
3001 : }
3002 : }
3003 : };
3004 :
3005 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3006 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3007 : // and `effective_config` in the response is meaningless, but we retain that syntax
3008 : // in order to remain compatible with the pageserver API.
3009 :
3010 0 : let response = HashMap::from([
3011 : (
3012 : "tenant_specific_overrides",
3013 0 : serde_json::to_value(&config)
3014 0 : .context("serializing tenant specific overrides")
3015 0 : .map_err(ApiError::InternalServerError)?,
3016 : ),
3017 : (
3018 0 : "effective_config",
3019 0 : serde_json::to_value(&config)
3020 0 : .context("serializing effective config")
3021 0 : .map_err(ApiError::InternalServerError)?,
3022 : ),
3023 : ]);
3024 :
3025 0 : Ok(response)
3026 0 : }
3027 :
3028 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3029 0 : &self,
3030 0 : time_travel_req: &TenantTimeTravelRequest,
3031 0 : tenant_id: TenantId,
3032 0 : timestamp: Cow<'_, str>,
3033 0 : done_if_after: Cow<'_, str>,
3034 0 : ) -> Result<(), ApiError> {
3035 0 : let _tenant_lock = trace_exclusive_lock(
3036 0 : &self.tenant_op_locks,
3037 0 : tenant_id,
3038 0 : TenantOperations::TimeTravelRemoteStorage,
3039 0 : )
3040 0 : .await;
3041 :
3042 0 : let node = {
3043 0 : let mut locked = self.inner.write().unwrap();
3044 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3045 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3046 : // but only at the start of the process, so it's really just to prevent operator
3047 : // mistakes.
3048 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3049 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3050 : {
3051 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3052 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3053 0 : )));
3054 0 : }
3055 0 : let maybe_attached = shard
3056 0 : .observed
3057 0 : .locations
3058 0 : .iter()
3059 0 : .filter_map(|(node_id, observed_location)| {
3060 0 : observed_location
3061 0 : .conf
3062 0 : .as_ref()
3063 0 : .map(|loc| (node_id, observed_location, loc.mode))
3064 0 : })
3065 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3066 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3067 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3068 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3069 0 : )));
3070 0 : }
3071 : }
3072 0 : let scheduler = &mut locked.scheduler;
3073 : // Right now we only perform the operation on a single node without parallelization
3074 : // TODO fan out the operation to multiple nodes for better performance
3075 0 : let node_id = scheduler.any_available_node()?;
3076 0 : let node = locked
3077 0 : .nodes
3078 0 : .get(&node_id)
3079 0 : .expect("Pageservers may not be deleted while lock is active");
3080 0 : node.clone()
3081 0 : };
3082 0 :
3083 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3084 0 : let mut counts = time_travel_req
3085 0 : .shard_counts
3086 0 : .iter()
3087 0 : .copied()
3088 0 : .collect::<HashSet<_>>()
3089 0 : .into_iter()
3090 0 : .collect::<Vec<_>>();
3091 0 : counts.sort_unstable();
3092 :
3093 0 : for count in counts {
3094 0 : let shard_ids = (0..count.count())
3095 0 : .map(|i| TenantShardId {
3096 0 : tenant_id,
3097 0 : shard_number: ShardNumber(i),
3098 0 : shard_count: count,
3099 0 : })
3100 0 : .collect::<Vec<_>>();
3101 0 : for tenant_shard_id in shard_ids {
3102 0 : let client = PageserverClient::new(
3103 0 : node.get_id(),
3104 0 : node.base_url(),
3105 0 : self.config.pageserver_jwt_token.as_deref(),
3106 0 : );
3107 0 :
3108 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3109 :
3110 0 : client
3111 0 : .tenant_time_travel_remote_storage(
3112 0 : tenant_shard_id,
3113 0 : ×tamp,
3114 0 : &done_if_after,
3115 0 : )
3116 0 : .await
3117 0 : .map_err(|e| {
3118 0 : ApiError::InternalServerError(anyhow::anyhow!(
3119 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3120 0 : node
3121 0 : ))
3122 0 : })?;
3123 : }
3124 : }
3125 0 : Ok(())
3126 0 : }
3127 :
3128 0 : pub(crate) async fn tenant_secondary_download(
3129 0 : &self,
3130 0 : tenant_id: TenantId,
3131 0 : wait: Option<Duration>,
3132 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3133 0 : let _tenant_lock = trace_shared_lock(
3134 0 : &self.tenant_op_locks,
3135 0 : tenant_id,
3136 0 : TenantOperations::SecondaryDownload,
3137 0 : )
3138 0 : .await;
3139 :
3140 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3141 0 : let targets = {
3142 0 : let locked = self.inner.read().unwrap();
3143 0 : let mut targets = Vec::new();
3144 :
3145 0 : for (tenant_shard_id, shard) in
3146 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3147 : {
3148 0 : for node_id in shard.intent.get_secondary() {
3149 0 : let node = locked
3150 0 : .nodes
3151 0 : .get(node_id)
3152 0 : .expect("Pageservers may not be deleted while referenced");
3153 0 :
3154 0 : targets.push((*tenant_shard_id, node.clone()));
3155 0 : }
3156 : }
3157 0 : targets
3158 0 : };
3159 0 :
3160 0 : // Issue concurrent requests to all shards' locations
3161 0 : let mut futs = FuturesUnordered::new();
3162 0 : for (tenant_shard_id, node) in targets {
3163 0 : let client = PageserverClient::new(
3164 0 : node.get_id(),
3165 0 : node.base_url(),
3166 0 : self.config.pageserver_jwt_token.as_deref(),
3167 0 : );
3168 0 : futs.push(async move {
3169 0 : let result = client
3170 0 : .tenant_secondary_download(tenant_shard_id, wait)
3171 0 : .await;
3172 0 : (result, node, tenant_shard_id)
3173 0 : })
3174 : }
3175 :
3176 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3177 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3178 : // well as more general cases like 503s, 500s, or timeouts.
3179 0 : let mut aggregate_progress = SecondaryProgress::default();
3180 0 : let mut aggregate_status: Option<StatusCode> = None;
3181 0 : let mut error: Option<mgmt_api::Error> = None;
3182 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3183 0 : match result {
3184 0 : Err(e) => {
3185 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3186 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3187 0 : // than they had hoped for.
3188 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3189 0 : error = Some(e)
3190 : }
3191 0 : Ok((status_code, progress)) => {
3192 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3193 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3194 0 : aggregate_progress.layers_total += progress.layers_total;
3195 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3196 0 : aggregate_progress.bytes_total += progress.bytes_total;
3197 0 : aggregate_progress.heatmap_mtime =
3198 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3199 0 : aggregate_status = match aggregate_status {
3200 0 : None => Some(status_code),
3201 0 : Some(StatusCode::OK) => Some(status_code),
3202 0 : Some(cur) => {
3203 0 : // Other status codes (e.g. 202) -- do not overwrite.
3204 0 : Some(cur)
3205 : }
3206 : };
3207 : }
3208 : }
3209 : }
3210 :
3211 : // If any of the shards return 202, indicate our result as 202.
3212 0 : match aggregate_status {
3213 : None => {
3214 0 : match error {
3215 0 : Some(e) => {
3216 0 : // No successes, and an error: surface it
3217 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3218 : }
3219 : None => {
3220 : // No shards found
3221 0 : Err(ApiError::NotFound(
3222 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3223 0 : ))
3224 : }
3225 : }
3226 : }
3227 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3228 : }
3229 0 : }
3230 :
3231 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
3232 0 : let _tenant_lock =
3233 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3234 :
3235 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3236 :
3237 : // Detach all shards. This also deletes local pageserver shard data.
3238 0 : let (detach_waiters, node) = {
3239 0 : let mut detach_waiters = Vec::new();
3240 0 : let mut locked = self.inner.write().unwrap();
3241 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3242 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3243 : // Update the tenant's intent to remove all attachments
3244 0 : shard.policy = PlacementPolicy::Detached;
3245 0 : shard
3246 0 : .schedule(scheduler, &mut ScheduleContext::default())
3247 0 : .expect("De-scheduling is infallible");
3248 0 : debug_assert!(shard.intent.get_attached().is_none());
3249 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3250 :
3251 0 : if let Some(waiter) =
3252 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3253 0 : {
3254 0 : detach_waiters.push(waiter);
3255 0 : }
3256 : }
3257 :
3258 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3259 : // was attached, just has to be able to see the S3 content)
3260 0 : let node_id = scheduler.any_available_node()?;
3261 0 : let node = nodes
3262 0 : .get(&node_id)
3263 0 : .expect("Pageservers may not be deleted while lock is active");
3264 0 : (detach_waiters, node.clone())
3265 0 : };
3266 0 :
3267 0 : // This reconcile wait can fail in a few ways:
3268 0 : // A there is a very long queue for the reconciler semaphore
3269 0 : // B some pageserver is failing to handle a detach promptly
3270 0 : // C some pageserver goes offline right at the moment we send it a request.
3271 0 : //
3272 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3273 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3274 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3275 0 : // deleting the underlying data).
3276 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3277 0 : .await?;
3278 :
3279 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3280 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3281 : // the tenant ID prefix, including all shards (even possibly stale ones).
3282 0 : match node
3283 0 : .with_client_retries(
3284 0 : |client| async move {
3285 0 : client
3286 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3287 0 : .await
3288 0 : },
3289 0 : &self.config.pageserver_jwt_token,
3290 0 : 1,
3291 0 : 3,
3292 0 : RECONCILE_TIMEOUT,
3293 0 : &self.cancel,
3294 0 : )
3295 0 : .await
3296 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3297 : {
3298 0 : Ok(_) => {}
3299 : Err(mgmt_api::Error::Cancelled) => {
3300 0 : return Err(ApiError::ShuttingDown);
3301 : }
3302 0 : Err(e) => {
3303 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3304 0 : // at large is unavailable.
3305 0 : tracing::error!("Error deleting via node {node}: {e}");
3306 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3307 : }
3308 : }
3309 :
3310 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3311 : // our in-memory state and database state.
3312 :
3313 : // Ordering: we delete persistent state first: if we then
3314 : // crash, we will drop the in-memory state.
3315 :
3316 : // Drop persistent state.
3317 0 : self.persistence.delete_tenant(tenant_id).await?;
3318 :
3319 : // Drop in-memory state
3320 : {
3321 0 : let mut locked = self.inner.write().unwrap();
3322 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3323 :
3324 : // Dereference Scheduler from shards before dropping them
3325 0 : for (_tenant_shard_id, shard) in
3326 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3327 0 : {
3328 0 : shard.intent.clear(scheduler);
3329 0 : }
3330 :
3331 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3332 0 : tracing::info!(
3333 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3334 0 : locked.tenants.len()
3335 : );
3336 : };
3337 :
3338 : // Success is represented as 404, to imitate the existing pageserver deletion API
3339 0 : Ok(StatusCode::NOT_FOUND)
3340 0 : }
3341 :
3342 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3343 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3344 : /// the tenant's policies (configuration) within the storage controller
3345 0 : pub(crate) async fn tenant_update_policy(
3346 0 : &self,
3347 0 : tenant_id: TenantId,
3348 0 : req: TenantPolicyRequest,
3349 0 : ) -> Result<(), ApiError> {
3350 : // We require an exclusive lock, because we are updating persistent and in-memory state
3351 0 : let _tenant_lock = trace_exclusive_lock(
3352 0 : &self.tenant_op_locks,
3353 0 : tenant_id,
3354 0 : TenantOperations::UpdatePolicy,
3355 0 : )
3356 0 : .await;
3357 :
3358 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3359 :
3360 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3361 :
3362 : let TenantPolicyRequest {
3363 0 : placement,
3364 0 : mut scheduling,
3365 0 : } = req;
3366 :
3367 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3368 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3369 : // this to take effect.
3370 0 : if scheduling.is_none() {
3371 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3372 0 : }
3373 0 : }
3374 :
3375 0 : self.persistence
3376 0 : .update_tenant_shard(
3377 0 : TenantFilter::Tenant(tenant_id),
3378 0 : placement.clone(),
3379 0 : None,
3380 0 : None,
3381 0 : scheduling,
3382 0 : )
3383 0 : .await?;
3384 :
3385 0 : let mut schedule_context = ScheduleContext::default();
3386 0 : let mut locked = self.inner.write().unwrap();
3387 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3388 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3389 0 : if let Some(placement) = &placement {
3390 0 : shard.policy = placement.clone();
3391 0 :
3392 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3393 0 : "Updated placement policy to {placement:?}");
3394 0 : }
3395 :
3396 0 : if let Some(scheduling) = &scheduling {
3397 0 : shard.set_scheduling_policy(*scheduling);
3398 0 :
3399 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3400 0 : "Updated scheduling policy to {scheduling:?}");
3401 0 : }
3402 :
3403 : // In case scheduling is being switched back on, try it now.
3404 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3405 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3406 : }
3407 :
3408 0 : Ok(())
3409 0 : }
3410 :
3411 0 : pub(crate) async fn tenant_timeline_create(
3412 0 : &self,
3413 0 : tenant_id: TenantId,
3414 0 : mut create_req: TimelineCreateRequest,
3415 0 : ) -> Result<TimelineInfo, ApiError> {
3416 0 : tracing::info!(
3417 0 : "Creating timeline {}/{}",
3418 : tenant_id,
3419 : create_req.new_timeline_id,
3420 : );
3421 :
3422 0 : let _tenant_lock = trace_shared_lock(
3423 0 : &self.tenant_op_locks,
3424 0 : tenant_id,
3425 0 : TenantOperations::TimelineCreate,
3426 0 : )
3427 0 : .await;
3428 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3429 :
3430 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3431 0 : if targets.0.is_empty() {
3432 0 : return Err(ApiError::NotFound(
3433 0 : anyhow::anyhow!("Tenant not found").into(),
3434 0 : ));
3435 0 : };
3436 0 :
3437 0 : let (shard_zero_tid, shard_zero_locations) =
3438 0 : targets.0.pop_first().expect("Must have at least one shard");
3439 0 : assert!(shard_zero_tid.is_shard_zero());
3440 :
3441 0 : async fn create_one(
3442 0 : tenant_shard_id: TenantShardId,
3443 0 : locations: ShardMutationLocations,
3444 0 : jwt: Option<String>,
3445 0 : create_req: TimelineCreateRequest,
3446 0 : ) -> Result<TimelineInfo, ApiError> {
3447 0 : let latest = locations.latest.node;
3448 0 :
3449 0 : tracing::info!(
3450 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3451 : tenant_shard_id,
3452 : create_req.new_timeline_id,
3453 : locations.latest.generation
3454 : );
3455 :
3456 0 : let client =
3457 0 : PageserverClient::new(latest.get_id(), latest.base_url(), jwt.as_deref());
3458 :
3459 0 : let timeline_info = client
3460 0 : .timeline_create(tenant_shard_id, &create_req)
3461 0 : .await
3462 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3463 :
3464 : // We propagate timeline creations to all attached locations such that a compute
3465 : // for the new timeline is able to start regardless of the current state of the
3466 : // tenant shard reconciliation.
3467 0 : for location in locations.other {
3468 0 : tracing::info!(
3469 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3470 : tenant_shard_id,
3471 : create_req.new_timeline_id,
3472 : location.node,
3473 : location.generation
3474 : );
3475 :
3476 0 : let client = PageserverClient::new(
3477 0 : location.node.get_id(),
3478 0 : location.node.base_url(),
3479 0 : jwt.as_deref(),
3480 0 : );
3481 :
3482 0 : let res = client
3483 0 : .timeline_create(tenant_shard_id, &create_req)
3484 0 : .await;
3485 :
3486 0 : if let Err(e) = res {
3487 0 : match e {
3488 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3489 0 : // Tenant might have been detached from the stale location,
3490 0 : // so ignore 404s.
3491 0 : },
3492 : _ => {
3493 0 : return Err(passthrough_api_error(&location.node, e));
3494 : }
3495 : }
3496 0 : }
3497 : }
3498 :
3499 0 : Ok(timeline_info)
3500 0 : }
3501 :
3502 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3503 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3504 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3505 0 : let timeline_info = create_one(
3506 0 : shard_zero_tid,
3507 0 : shard_zero_locations,
3508 0 : self.config.pageserver_jwt_token.clone(),
3509 0 : create_req.clone(),
3510 0 : )
3511 0 : .await?;
3512 :
3513 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3514 0 : match &mut create_req.mode {
3515 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3516 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3517 0 : },
3518 0 : _ => {}
3519 : }
3520 :
3521 : // Create timeline on remaining shards with number >0
3522 0 : if !targets.0.is_empty() {
3523 : // If we had multiple shards, issue requests for the remainder now.
3524 0 : let jwt = &self.config.pageserver_jwt_token;
3525 0 : self.tenant_for_shards(
3526 0 : targets
3527 0 : .0
3528 0 : .iter()
3529 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3530 0 : .collect(),
3531 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3532 0 : let create_req = create_req.clone();
3533 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3534 0 : Box::pin(create_one(
3535 0 : tenant_shard_id,
3536 0 : mutation_locations,
3537 0 : jwt.clone(),
3538 0 : create_req,
3539 0 : ))
3540 0 : },
3541 0 : )
3542 0 : .await?;
3543 0 : }
3544 :
3545 0 : Ok(timeline_info)
3546 0 : })
3547 0 : .await?
3548 0 : }
3549 :
3550 0 : pub(crate) async fn tenant_timeline_archival_config(
3551 0 : &self,
3552 0 : tenant_id: TenantId,
3553 0 : timeline_id: TimelineId,
3554 0 : req: TimelineArchivalConfigRequest,
3555 0 : ) -> Result<(), ApiError> {
3556 0 : tracing::info!(
3557 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3558 : req.state
3559 : );
3560 :
3561 0 : let _tenant_lock = trace_shared_lock(
3562 0 : &self.tenant_op_locks,
3563 0 : tenant_id,
3564 0 : TenantOperations::TimelineArchivalConfig,
3565 0 : )
3566 0 : .await;
3567 :
3568 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3569 0 : if targets.0.is_empty() {
3570 0 : return Err(ApiError::NotFound(
3571 0 : anyhow::anyhow!("Tenant not found").into(),
3572 0 : ));
3573 0 : }
3574 0 : async fn config_one(
3575 0 : tenant_shard_id: TenantShardId,
3576 0 : timeline_id: TimelineId,
3577 0 : node: Node,
3578 0 : jwt: Option<String>,
3579 0 : req: TimelineArchivalConfigRequest,
3580 0 : ) -> Result<(), ApiError> {
3581 0 : tracing::info!(
3582 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3583 : );
3584 :
3585 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3586 0 :
3587 0 : client
3588 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3589 0 : .await
3590 0 : .map_err(|e| match e {
3591 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3592 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3593 : }
3594 0 : _ => passthrough_api_error(&node, e),
3595 0 : })
3596 0 : }
3597 :
3598 : // no shard needs to go first/last; the operation should be idempotent
3599 : // TODO: it would be great to ensure that all shards return the same error
3600 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3601 0 : let results = self
3602 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3603 0 : futures::FutureExt::boxed(config_one(
3604 0 : tenant_shard_id,
3605 0 : timeline_id,
3606 0 : node,
3607 0 : self.config.pageserver_jwt_token.clone(),
3608 0 : req.clone(),
3609 0 : ))
3610 0 : })
3611 0 : .await?;
3612 0 : assert!(!results.is_empty(), "must have at least one result");
3613 :
3614 0 : Ok(())
3615 0 : }).await?
3616 0 : }
3617 :
3618 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3619 0 : &self,
3620 0 : tenant_id: TenantId,
3621 0 : timeline_id: TimelineId,
3622 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3623 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3624 :
3625 0 : let _tenant_lock = trace_shared_lock(
3626 0 : &self.tenant_op_locks,
3627 0 : tenant_id,
3628 0 : TenantOperations::TimelineDetachAncestor,
3629 0 : )
3630 0 : .await;
3631 :
3632 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3633 0 : if targets.0.is_empty() {
3634 0 : return Err(ApiError::NotFound(
3635 0 : anyhow::anyhow!("Tenant not found").into(),
3636 0 : ));
3637 0 : }
3638 :
3639 0 : async fn detach_one(
3640 0 : tenant_shard_id: TenantShardId,
3641 0 : timeline_id: TimelineId,
3642 0 : node: Node,
3643 0 : jwt: Option<String>,
3644 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3645 0 : tracing::info!(
3646 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3647 : );
3648 :
3649 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3650 0 :
3651 0 : client
3652 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3653 0 : .await
3654 0 : .map_err(|e| {
3655 : use mgmt_api::Error;
3656 :
3657 0 : match e {
3658 : // no ancestor (ever)
3659 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3660 0 : "{node}: {}",
3661 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3662 0 : )),
3663 : // too many ancestors
3664 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3665 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3666 : }
3667 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3668 0 : // avoid turning these into conflicts to remain compatible with
3669 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3670 0 : // detach
3671 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3672 : }
3673 : // rest can be mapped as usual
3674 0 : other => passthrough_api_error(&node, other),
3675 : }
3676 0 : })
3677 0 : .map(|res| (tenant_shard_id.shard_number, res))
3678 0 : }
3679 :
3680 : // no shard needs to go first/last; the operation should be idempotent
3681 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3682 0 : let mut results = self
3683 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3684 0 : futures::FutureExt::boxed(detach_one(
3685 0 : tenant_shard_id,
3686 0 : timeline_id,
3687 0 : node,
3688 0 : self.config.pageserver_jwt_token.clone(),
3689 0 : ))
3690 0 : })
3691 0 : .await?;
3692 :
3693 0 : let any = results.pop().expect("we must have at least one response");
3694 0 :
3695 0 : let mismatching = results
3696 0 : .iter()
3697 0 : .filter(|(_, res)| res != &any.1)
3698 0 : .collect::<Vec<_>>();
3699 0 : if !mismatching.is_empty() {
3700 : // this can be hit by races which should not happen because operation lock on cplane
3701 0 : let matching = results.len() - mismatching.len();
3702 0 : tracing::error!(
3703 : matching,
3704 : compared_against=?any,
3705 : ?mismatching,
3706 0 : "shards returned different results"
3707 : );
3708 :
3709 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3710 0 : }
3711 0 :
3712 0 : Ok(any.1)
3713 0 : }).await?
3714 0 : }
3715 :
3716 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3717 0 : &self,
3718 0 : tenant_id: TenantId,
3719 0 : timeline_id: TimelineId,
3720 0 : dir: BlockUnblock,
3721 0 : ) -> Result<(), ApiError> {
3722 0 : let _tenant_lock = trace_shared_lock(
3723 0 : &self.tenant_op_locks,
3724 0 : tenant_id,
3725 0 : TenantOperations::TimelineGcBlockUnblock,
3726 0 : )
3727 0 : .await;
3728 :
3729 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3730 0 : if targets.0.is_empty() {
3731 0 : return Err(ApiError::NotFound(
3732 0 : anyhow::anyhow!("Tenant not found").into(),
3733 0 : ));
3734 0 : }
3735 :
3736 0 : async fn do_one(
3737 0 : tenant_shard_id: TenantShardId,
3738 0 : timeline_id: TimelineId,
3739 0 : node: Node,
3740 0 : jwt: Option<String>,
3741 0 : dir: BlockUnblock,
3742 0 : ) -> Result<(), ApiError> {
3743 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3744 0 :
3745 0 : client
3746 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3747 0 : .await
3748 0 : .map_err(|e| passthrough_api_error(&node, e))
3749 0 : }
3750 :
3751 : // no shard needs to go first/last; the operation should be idempotent
3752 0 : let locations = targets
3753 0 : .0
3754 0 : .iter()
3755 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3756 0 : .collect();
3757 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
3758 0 : futures::FutureExt::boxed(do_one(
3759 0 : tenant_shard_id,
3760 0 : timeline_id,
3761 0 : node,
3762 0 : self.config.pageserver_jwt_token.clone(),
3763 0 : dir,
3764 0 : ))
3765 0 : })
3766 0 : .await
3767 0 : })
3768 0 : .await??;
3769 0 : Ok(())
3770 0 : }
3771 :
3772 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
3773 0 : &self,
3774 0 : tenant_shard_id: TenantShardId,
3775 0 : timeline_id: TimelineId,
3776 0 : concurrency: Option<usize>,
3777 0 : recurse: bool,
3778 0 : ) -> Result<(), ApiError> {
3779 0 : let _tenant_lock = trace_shared_lock(
3780 0 : &self.tenant_op_locks,
3781 0 : tenant_shard_id.tenant_id,
3782 0 : TenantOperations::DownloadHeatmapLayers,
3783 0 : )
3784 0 : .await;
3785 :
3786 0 : let targets = {
3787 0 : let locked = self.inner.read().unwrap();
3788 0 : let mut targets = Vec::new();
3789 :
3790 : // If the request got an unsharded tenant id, then apply
3791 : // the operation to all shards. Otherwise, apply it to a specific shard.
3792 0 : let shards_range = if tenant_shard_id.is_unsharded() {
3793 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
3794 : } else {
3795 0 : tenant_shard_id.range()
3796 : };
3797 :
3798 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
3799 0 : if let Some(node_id) = shard.intent.get_attached() {
3800 0 : let node = locked
3801 0 : .nodes
3802 0 : .get(node_id)
3803 0 : .expect("Pageservers may not be deleted while referenced");
3804 0 :
3805 0 : targets.push((*tenant_shard_id, node.clone()));
3806 0 : }
3807 : }
3808 0 : targets
3809 0 : };
3810 0 :
3811 0 : self.tenant_for_shards_api(
3812 0 : targets,
3813 0 : |tenant_shard_id, client| async move {
3814 0 : client
3815 0 : .timeline_download_heatmap_layers(
3816 0 : tenant_shard_id,
3817 0 : timeline_id,
3818 0 : concurrency,
3819 0 : recurse,
3820 0 : )
3821 0 : .await
3822 0 : },
3823 0 : 1,
3824 0 : 1,
3825 0 : SHORT_RECONCILE_TIMEOUT,
3826 0 : &self.cancel,
3827 0 : )
3828 0 : .await;
3829 :
3830 0 : Ok(())
3831 0 : }
3832 :
3833 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3834 : ///
3835 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3836 0 : async fn tenant_for_shards<F, R>(
3837 0 : &self,
3838 0 : locations: Vec<(TenantShardId, Node)>,
3839 0 : mut req_fn: F,
3840 0 : ) -> Result<Vec<R>, ApiError>
3841 0 : where
3842 0 : F: FnMut(
3843 0 : TenantShardId,
3844 0 : Node,
3845 0 : )
3846 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3847 0 : {
3848 0 : let mut futs = FuturesUnordered::new();
3849 0 : let mut results = Vec::with_capacity(locations.len());
3850 :
3851 0 : for (tenant_shard_id, node) in locations {
3852 0 : futs.push(req_fn(tenant_shard_id, node));
3853 0 : }
3854 :
3855 0 : while let Some(r) = futs.next().await {
3856 0 : results.push(r?);
3857 : }
3858 :
3859 0 : Ok(results)
3860 0 : }
3861 :
3862 : /// Concurrently invoke a pageserver API call on many shards at once
3863 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3864 0 : &self,
3865 0 : locations: Vec<(TenantShardId, Node)>,
3866 0 : op: O,
3867 0 : warn_threshold: u32,
3868 0 : max_retries: u32,
3869 0 : timeout: Duration,
3870 0 : cancel: &CancellationToken,
3871 0 : ) -> Vec<mgmt_api::Result<T>>
3872 0 : where
3873 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3874 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3875 0 : {
3876 0 : let mut futs = FuturesUnordered::new();
3877 0 : let mut results = Vec::with_capacity(locations.len());
3878 :
3879 0 : for (tenant_shard_id, node) in locations {
3880 0 : futs.push(async move {
3881 0 : node.with_client_retries(
3882 0 : |client| op(tenant_shard_id, client),
3883 0 : &self.config.pageserver_jwt_token,
3884 0 : warn_threshold,
3885 0 : max_retries,
3886 0 : timeout,
3887 0 : cancel,
3888 0 : )
3889 0 : .await
3890 0 : });
3891 0 : }
3892 :
3893 0 : while let Some(r) = futs.next().await {
3894 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3895 0 : results.push(r);
3896 0 : }
3897 :
3898 0 : results
3899 0 : }
3900 :
3901 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3902 : /// when creating and deleting timelines:
3903 : /// - Makes sure shards are attached somewhere if they weren't already
3904 : /// - Looks up the shards and the nodes where they were most recently attached
3905 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3906 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3907 0 : async fn tenant_remote_mutation<R, O, F>(
3908 0 : &self,
3909 0 : tenant_id: TenantId,
3910 0 : op: O,
3911 0 : ) -> Result<R, ApiError>
3912 0 : where
3913 0 : O: FnOnce(TenantMutationLocations) -> F,
3914 0 : F: std::future::Future<Output = R>,
3915 0 : {
3916 0 : let mutation_locations = {
3917 0 : let mut locations = TenantMutationLocations::default();
3918 :
3919 : // Load the currently attached pageservers for the latest generation of each shard. This can
3920 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3921 : // will still be the latest when we're done: we will check generations again at the end of
3922 : // this function to handle that.
3923 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
3924 :
3925 0 : if generations
3926 0 : .iter()
3927 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3928 : {
3929 0 : let shard_generations = generations
3930 0 : .into_iter()
3931 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
3932 0 : .collect::<HashMap<_, _>>();
3933 0 :
3934 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3935 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)
3936 0 : let locked = self.inner.read().unwrap();
3937 0 : for (shard_id, shard) in
3938 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3939 : {
3940 0 : match shard.policy {
3941 : PlacementPolicy::Attached(_) => {
3942 : // This shard is meant to be attached: the caller is not wrong to try and
3943 : // use this function, but we can't service the request right now.
3944 0 : let Some(generation) = shard_generations.get(shard_id) else {
3945 : // This can only happen if there is a split brain controller modifying the database. This should
3946 : // never happen when testing, and if it happens in production we can only log the issue.
3947 0 : debug_assert!(false);
3948 0 : tracing::error!(
3949 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
3950 : );
3951 0 : continue;
3952 : };
3953 0 : let (generation, generation_pageserver) = generation;
3954 0 : if let Some(generation) = generation {
3955 0 : if generation_pageserver.is_none() {
3956 : // This is legitimate only in a very narrow window where the shard was only just configured into
3957 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
3958 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
3959 0 : tracing::warn!(
3960 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
3961 : );
3962 0 : }
3963 : } else {
3964 : // This should never happen: a shard with no generation is only permitted when it was created in some state
3965 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
3966 0 : debug_assert!(false);
3967 0 : tracing::error!(
3968 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
3969 : );
3970 0 : continue;
3971 : }
3972 : }
3973 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3974 0 : return Err(ApiError::Conflict(format!(
3975 0 : "Shard {shard_id} tenant has policy {:?}",
3976 0 : shard.policy
3977 0 : )));
3978 : }
3979 : }
3980 : }
3981 :
3982 0 : return Err(ApiError::ResourceUnavailable(
3983 0 : "One or more shards in tenant is not yet attached".into(),
3984 0 : ));
3985 0 : }
3986 0 :
3987 0 : let locked = self.inner.read().unwrap();
3988 : for ShardGenerationState {
3989 0 : tenant_shard_id,
3990 0 : generation,
3991 0 : generation_pageserver,
3992 0 : } in generations
3993 : {
3994 0 : let node_id = generation_pageserver.expect("We checked for None above");
3995 0 : let node = locked
3996 0 : .nodes
3997 0 : .get(&node_id)
3998 0 : .ok_or(ApiError::Conflict(format!(
3999 0 : "Raced with removal of node {node_id}"
4000 0 : )))?;
4001 0 : let generation = generation.expect("Checked above");
4002 0 :
4003 0 : let tenant = locked.tenants.get(&tenant_shard_id);
4004 :
4005 : // TODO(vlad): Abstract the logic that finds stale attached locations
4006 : // from observed state into a [`Service`] method.
4007 0 : let other_locations = match tenant {
4008 0 : Some(tenant) => {
4009 0 : let mut other = tenant.attached_locations();
4010 0 : let latest_location_index =
4011 0 : other.iter().position(|&l| l == (node.get_id(), generation));
4012 0 : if let Some(idx) = latest_location_index {
4013 0 : other.remove(idx);
4014 0 : }
4015 :
4016 0 : other
4017 : }
4018 0 : None => Vec::default(),
4019 : };
4020 :
4021 0 : let location = ShardMutationLocations {
4022 0 : latest: MutationLocation {
4023 0 : node: node.clone(),
4024 0 : generation,
4025 0 : },
4026 0 : other: other_locations
4027 0 : .into_iter()
4028 0 : .filter_map(|(node_id, generation)| {
4029 0 : let node = locked.nodes.get(&node_id)?;
4030 :
4031 0 : Some(MutationLocation {
4032 0 : node: node.clone(),
4033 0 : generation,
4034 0 : })
4035 0 : })
4036 0 : .collect(),
4037 0 : };
4038 0 : locations.0.insert(tenant_shard_id, location);
4039 0 : }
4040 :
4041 0 : locations
4042 : };
4043 :
4044 0 : let result = op(mutation_locations.clone()).await;
4045 :
4046 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
4047 : // our remote operation executed on the latest generation and is therefore persistent.
4048 : {
4049 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
4050 0 : if latest_generations
4051 0 : .into_iter()
4052 0 : .map(
4053 0 : |ShardGenerationState {
4054 : tenant_shard_id,
4055 : generation,
4056 : generation_pageserver: _,
4057 0 : }| (tenant_shard_id, generation),
4058 0 : )
4059 0 : .collect::<Vec<_>>()
4060 0 : != mutation_locations
4061 0 : .0
4062 0 : .into_iter()
4063 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
4064 0 : .collect::<Vec<_>>()
4065 : {
4066 : // We raced with something that incremented the generation, and therefore cannot be
4067 : // confident that our actions are persistent (they might have hit an old generation).
4068 : //
4069 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
4070 0 : return Err(ApiError::ResourceUnavailable(
4071 0 : "Tenant attachment changed, please retry".into(),
4072 0 : ));
4073 0 : }
4074 0 : }
4075 0 :
4076 0 : Ok(result)
4077 0 : }
4078 :
4079 0 : pub(crate) async fn tenant_timeline_delete(
4080 0 : &self,
4081 0 : tenant_id: TenantId,
4082 0 : timeline_id: TimelineId,
4083 0 : ) -> Result<StatusCode, ApiError> {
4084 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
4085 0 : let _tenant_lock = trace_shared_lock(
4086 0 : &self.tenant_op_locks,
4087 0 : tenant_id,
4088 0 : TenantOperations::TimelineDelete,
4089 0 : )
4090 0 : .await;
4091 :
4092 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
4093 0 : if targets.0.is_empty() {
4094 0 : return Err(ApiError::NotFound(
4095 0 : anyhow::anyhow!("Tenant not found").into(),
4096 0 : ));
4097 0 : }
4098 0 :
4099 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
4100 0 : assert!(shard_zero_tid.is_shard_zero());
4101 :
4102 0 : async fn delete_one(
4103 0 : tenant_shard_id: TenantShardId,
4104 0 : timeline_id: TimelineId,
4105 0 : node: Node,
4106 0 : jwt: Option<String>,
4107 0 : ) -> Result<StatusCode, ApiError> {
4108 0 : tracing::info!(
4109 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4110 : );
4111 :
4112 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
4113 0 : let res = client
4114 0 : .timeline_delete(tenant_shard_id, timeline_id)
4115 0 : .await;
4116 :
4117 0 : match res {
4118 0 : Ok(ok) => Ok(ok),
4119 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
4120 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
4121 0 : Err(e) => {
4122 0 : Err(
4123 0 : ApiError::InternalServerError(anyhow::anyhow!(
4124 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
4125 0 : ))
4126 0 : )
4127 : }
4128 : }
4129 0 : }
4130 :
4131 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4132 0 : let statuses = self
4133 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
4134 0 : Box::pin(delete_one(
4135 0 : tenant_shard_id,
4136 0 : timeline_id,
4137 0 : node,
4138 0 : self.config.pageserver_jwt_token.clone(),
4139 0 : ))
4140 0 : })
4141 0 : .await?;
4142 :
4143 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
4144 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
4145 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
4146 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
4147 0 : return Ok(StatusCode::CONFLICT);
4148 0 : }
4149 0 :
4150 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
4151 0 : return Ok(StatusCode::ACCEPTED);
4152 0 : }
4153 :
4154 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
4155 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
4156 0 : let shard_zero_status = delete_one(
4157 0 : shard_zero_tid,
4158 0 : timeline_id,
4159 0 : shard_zero_locations.latest.node,
4160 0 : self.config.pageserver_jwt_token.clone(),
4161 0 : )
4162 0 : .await?;
4163 0 : Ok(shard_zero_status)
4164 0 : }).await?
4165 0 : }
4166 :
4167 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
4168 0 : pub(crate) async fn tenant_shard0_node(
4169 0 : &self,
4170 0 : tenant_id: TenantId,
4171 0 : ) -> Result<(Node, TenantShardId), ApiError> {
4172 0 : let tenant_shard_id = {
4173 0 : let locked = self.inner.read().unwrap();
4174 0 : let Some((tenant_shard_id, _shard)) = locked
4175 0 : .tenants
4176 0 : .range(TenantShardId::tenant_range(tenant_id))
4177 0 : .next()
4178 : else {
4179 0 : return Err(ApiError::NotFound(
4180 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
4181 0 : ));
4182 : };
4183 :
4184 0 : *tenant_shard_id
4185 0 : };
4186 0 :
4187 0 : self.tenant_shard_node(tenant_shard_id)
4188 0 : .await
4189 0 : .map(|node| (node, tenant_shard_id))
4190 0 : }
4191 :
4192 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
4193 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
4194 0 : pub(crate) async fn tenant_shard_node(
4195 0 : &self,
4196 0 : tenant_shard_id: TenantShardId,
4197 0 : ) -> Result<Node, ApiError> {
4198 0 : // Look up in-memory state and maybe use the node from there.
4199 0 : {
4200 0 : let locked = self.inner.read().unwrap();
4201 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
4202 0 : return Err(ApiError::NotFound(
4203 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
4204 0 : ));
4205 : };
4206 :
4207 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
4208 0 : tracing::warn!(
4209 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
4210 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
4211 : shard.policy
4212 : );
4213 0 : return Err(ApiError::Conflict(
4214 0 : "Cannot call timeline API on non-attached tenant".to_string(),
4215 0 : ));
4216 : };
4217 :
4218 0 : if shard.reconciler.is_none() {
4219 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
4220 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
4221 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
4222 : // This should never happen
4223 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4224 0 : "Shard refers to nonexistent node"
4225 0 : )));
4226 : };
4227 0 : return Ok(node.clone());
4228 0 : }
4229 : };
4230 :
4231 : // Look up the latest attached pageserver location from the database
4232 : // generation state: this will reflect the progress of any ongoing migration.
4233 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4234 : // the case where they call through to the pageserver and get a 404.
4235 0 : let db_result = self
4236 0 : .persistence
4237 0 : .tenant_generations(tenant_shard_id.tenant_id)
4238 0 : .await?;
4239 : let Some(ShardGenerationState {
4240 : tenant_shard_id: _,
4241 : generation: _,
4242 0 : generation_pageserver: Some(node_id),
4243 0 : }) = db_result
4244 0 : .into_iter()
4245 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
4246 : else {
4247 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4248 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4249 : // or a conflict response (case where tenant was detached in background)
4250 0 : return Err(ApiError::ResourceUnavailable(
4251 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
4252 0 : ));
4253 : };
4254 0 : let locked = self.inner.read().unwrap();
4255 0 : let Some(node) = locked.nodes.get(&node_id) else {
4256 : // This should never happen
4257 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4258 0 : "Shard refers to nonexistent node"
4259 0 : )));
4260 : };
4261 :
4262 0 : Ok(node.clone())
4263 0 : }
4264 :
4265 0 : pub(crate) fn tenant_locate(
4266 0 : &self,
4267 0 : tenant_id: TenantId,
4268 0 : ) -> Result<TenantLocateResponse, ApiError> {
4269 0 : let locked = self.inner.read().unwrap();
4270 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4271 :
4272 0 : let mut result = Vec::new();
4273 0 : let mut shard_params: Option<ShardParameters> = None;
4274 :
4275 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4276 : {
4277 0 : let node_id =
4278 0 : shard
4279 0 : .intent
4280 0 : .get_attached()
4281 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4282 0 : "Cannot locate a tenant that is not attached"
4283 0 : )))?;
4284 :
4285 0 : let node = locked
4286 0 : .nodes
4287 0 : .get(&node_id)
4288 0 : .expect("Pageservers may not be deleted while referenced");
4289 0 :
4290 0 : result.push(node.shard_location(*tenant_shard_id));
4291 0 :
4292 0 : match &shard_params {
4293 0 : None => {
4294 0 : shard_params = Some(ShardParameters {
4295 0 : stripe_size: shard.shard.stripe_size,
4296 0 : count: shard.shard.count,
4297 0 : });
4298 0 : }
4299 0 : Some(params) => {
4300 0 : if params.stripe_size != shard.shard.stripe_size {
4301 : // This should never happen. We enforce at runtime because it's simpler than
4302 : // adding an extra per-tenant data structure to store the things that should be the same
4303 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4304 0 : "Inconsistent shard stripe size parameters!"
4305 0 : )));
4306 0 : }
4307 : }
4308 : }
4309 : }
4310 :
4311 0 : if result.is_empty() {
4312 0 : return Err(ApiError::NotFound(
4313 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4314 0 : ));
4315 0 : }
4316 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4317 0 : tracing::info!(
4318 0 : "Located tenant {} with params {:?} on shards {}",
4319 0 : tenant_id,
4320 0 : shard_params,
4321 0 : result
4322 0 : .iter()
4323 0 : .map(|s| format!("{:?}", s))
4324 0 : .collect::<Vec<_>>()
4325 0 : .join(",")
4326 : );
4327 :
4328 0 : Ok(TenantLocateResponse {
4329 0 : shards: result,
4330 0 : shard_params,
4331 0 : })
4332 0 : }
4333 :
4334 : /// Returns None if the input iterator of shards does not include a shard with number=0
4335 0 : fn tenant_describe_impl<'a>(
4336 0 : &self,
4337 0 : shards: impl Iterator<Item = &'a TenantShard>,
4338 0 : ) -> Option<TenantDescribeResponse> {
4339 0 : let mut shard_zero = None;
4340 0 : let mut describe_shards = Vec::new();
4341 :
4342 0 : for shard in shards {
4343 0 : if shard.tenant_shard_id.is_shard_zero() {
4344 0 : shard_zero = Some(shard);
4345 0 : }
4346 :
4347 0 : describe_shards.push(TenantDescribeResponseShard {
4348 0 : tenant_shard_id: shard.tenant_shard_id,
4349 0 : node_attached: *shard.intent.get_attached(),
4350 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4351 0 : last_error: shard
4352 0 : .last_error
4353 0 : .lock()
4354 0 : .unwrap()
4355 0 : .as_ref()
4356 0 : .map(|e| format!("{e}"))
4357 0 : .unwrap_or("".to_string())
4358 0 : .clone(),
4359 0 : is_reconciling: shard.reconciler.is_some(),
4360 0 : is_pending_compute_notification: shard.pending_compute_notification,
4361 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4362 0 : scheduling_policy: *shard.get_scheduling_policy(),
4363 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4364 : })
4365 : }
4366 :
4367 0 : let shard_zero = shard_zero?;
4368 :
4369 0 : Some(TenantDescribeResponse {
4370 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
4371 0 : shards: describe_shards,
4372 0 : stripe_size: shard_zero.shard.stripe_size,
4373 0 : policy: shard_zero.policy.clone(),
4374 0 : config: shard_zero.config.clone(),
4375 0 : })
4376 0 : }
4377 :
4378 0 : pub(crate) fn tenant_describe(
4379 0 : &self,
4380 0 : tenant_id: TenantId,
4381 0 : ) -> Result<TenantDescribeResponse, ApiError> {
4382 0 : let locked = self.inner.read().unwrap();
4383 0 :
4384 0 : self.tenant_describe_impl(
4385 0 : locked
4386 0 : .tenants
4387 0 : .range(TenantShardId::tenant_range(tenant_id))
4388 0 : .map(|(_k, v)| v),
4389 0 : )
4390 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
4391 0 : }
4392 :
4393 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
4394 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
4395 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
4396 : /// in our external API.
4397 0 : pub(crate) fn tenant_list(
4398 0 : &self,
4399 0 : limit: Option<usize>,
4400 0 : start_after: Option<TenantId>,
4401 0 : ) -> Vec<TenantDescribeResponse> {
4402 0 : let locked = self.inner.read().unwrap();
4403 :
4404 : // Apply start_from parameter
4405 0 : let shard_range = match start_after {
4406 0 : None => locked.tenants.range(..),
4407 0 : Some(tenant_id) => locked.tenants.range(
4408 0 : TenantShardId {
4409 0 : tenant_id,
4410 0 : shard_number: ShardNumber(u8::MAX),
4411 0 : shard_count: ShardCount(u8::MAX),
4412 0 : }..,
4413 0 : ),
4414 : };
4415 :
4416 0 : let mut result = Vec::new();
4417 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
4418 0 : result.push(
4419 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
4420 0 : .expect("Groups are always non-empty"),
4421 0 : );
4422 :
4423 : // Enforce `limit` parameter
4424 0 : if let Some(limit) = limit {
4425 0 : if result.len() >= limit {
4426 0 : break;
4427 0 : }
4428 0 : }
4429 : }
4430 :
4431 0 : result
4432 0 : }
4433 :
4434 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
4435 : async fn abort_tenant_shard_split(
4436 : &self,
4437 : op: &TenantShardSplitAbort,
4438 : ) -> Result<(), TenantShardSplitAbortError> {
4439 : // Cleaning up a split:
4440 : // - Parent shards are not destroyed during a split, just detached.
4441 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
4442 : // just the children attached, or both.
4443 : //
4444 : // Therefore our work to do is to:
4445 : // 1. Clean up storage controller's internal state to just refer to parents, no children
4446 : // 2. Call out to pageservers to ensure that children are detached
4447 : // 3. Call out to pageservers to ensure that parents are attached.
4448 : //
4449 : // Crash safety:
4450 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
4451 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
4452 : // and detach them.
4453 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
4454 : // from our database, then we will re-enter this cleanup routine on startup.
4455 :
4456 : let TenantShardSplitAbort {
4457 : tenant_id,
4458 : new_shard_count,
4459 : new_stripe_size,
4460 : ..
4461 : } = op;
4462 :
4463 : // First abort persistent state, if any exists.
4464 : match self
4465 : .persistence
4466 : .abort_shard_split(*tenant_id, *new_shard_count)
4467 : .await?
4468 : {
4469 : AbortShardSplitStatus::Aborted => {
4470 : // Proceed to roll back any child shards created on pageservers
4471 : }
4472 : AbortShardSplitStatus::Complete => {
4473 : // The split completed (we might hit that path if e.g. our database transaction
4474 : // to write the completion landed in the database, but we dropped connection
4475 : // before seeing the result).
4476 : //
4477 : // We must update in-memory state to reflect the successful split.
4478 : self.tenant_shard_split_commit_inmem(
4479 : *tenant_id,
4480 : *new_shard_count,
4481 : *new_stripe_size,
4482 : );
4483 : return Ok(());
4484 : }
4485 : }
4486 :
4487 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
4488 : let detach_locations: Vec<(Node, TenantShardId)> = {
4489 : let mut detach_locations = Vec::new();
4490 : let mut locked = self.inner.write().unwrap();
4491 : let (nodes, tenants, scheduler) = locked.parts_mut();
4492 :
4493 : for (tenant_shard_id, shard) in
4494 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
4495 : {
4496 : if shard.shard.count == op.new_shard_count {
4497 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
4498 : // is infallible, so if we got an error we shouldn't have got that far.
4499 : tracing::warn!(
4500 : "During split abort, child shard {tenant_shard_id} found in-memory"
4501 : );
4502 : continue;
4503 : }
4504 :
4505 : // Add the children of this shard to this list of things to detach
4506 : if let Some(node_id) = shard.intent.get_attached() {
4507 : for child_id in tenant_shard_id.split(*new_shard_count) {
4508 : detach_locations.push((
4509 : nodes
4510 : .get(node_id)
4511 : .expect("Intent references nonexistent node")
4512 : .clone(),
4513 : child_id,
4514 : ));
4515 : }
4516 : } else {
4517 : tracing::warn!(
4518 : "During split abort, shard {tenant_shard_id} has no attached location"
4519 : );
4520 : }
4521 :
4522 : tracing::info!("Restoring parent shard {tenant_shard_id}");
4523 :
4524 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
4525 : // if the original attachment location is offline.
4526 : if let Some(node_id) = shard.intent.get_attached() {
4527 : if !nodes.get(node_id).unwrap().is_available() {
4528 : tracing::info!(
4529 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
4530 : );
4531 : shard.intent.demote_attached(scheduler, *node_id);
4532 : }
4533 : }
4534 : for node_id in shard.intent.get_secondary().clone() {
4535 : if !nodes.get(&node_id).unwrap().is_available() {
4536 : tracing::info!(
4537 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
4538 : );
4539 : shard.intent.remove_secondary(scheduler, node_id);
4540 : }
4541 : }
4542 :
4543 : shard.splitting = SplitState::Idle;
4544 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
4545 : // If this shard can't be scheduled now (perhaps due to offline nodes or
4546 : // capacity issues), that must not prevent us rolling back a split. In this
4547 : // case it should be eventually scheduled in the background.
4548 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
4549 : }
4550 :
4551 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
4552 : }
4553 :
4554 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
4555 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
4556 :
4557 : detach_locations
4558 : };
4559 :
4560 : for (node, child_id) in detach_locations {
4561 : if !node.is_available() {
4562 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
4563 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
4564 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
4565 : // them from the node.
4566 : tracing::warn!(
4567 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
4568 : );
4569 : continue;
4570 : }
4571 :
4572 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
4573 : // a 503 and retry, up to our limit.
4574 : tracing::info!("Detaching {child_id} on {node}...");
4575 : match node
4576 : .with_client_retries(
4577 0 : |client| async move {
4578 0 : let config = LocationConfig {
4579 0 : mode: LocationConfigMode::Detached,
4580 0 : generation: None,
4581 0 : secondary_conf: None,
4582 0 : shard_number: child_id.shard_number.0,
4583 0 : shard_count: child_id.shard_count.literal(),
4584 0 : // Stripe size and tenant config don't matter when detaching
4585 0 : shard_stripe_size: 0,
4586 0 : tenant_conf: TenantConfig::default(),
4587 0 : };
4588 0 :
4589 0 : client.location_config(child_id, config, None, false).await
4590 0 : },
4591 : &self.config.pageserver_jwt_token,
4592 : 1,
4593 : 10,
4594 : Duration::from_secs(5),
4595 : &self.cancel,
4596 : )
4597 : .await
4598 : {
4599 : Some(Ok(_)) => {}
4600 : Some(Err(e)) => {
4601 : // We failed to communicate with the remote node. This is problematic: we may be
4602 : // leaving it with a rogue child shard.
4603 : tracing::warn!(
4604 : "Failed to detach child {child_id} from node {node} during abort"
4605 : );
4606 : return Err(e.into());
4607 : }
4608 : None => {
4609 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
4610 : // clean up on restart. The node going offline requires a retry.
4611 : return Err(TenantShardSplitAbortError::Unavailable);
4612 : }
4613 : };
4614 : }
4615 :
4616 : tracing::info!("Successfully aborted split");
4617 : Ok(())
4618 : }
4619 :
4620 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
4621 : /// of the tenant map to reflect the child shards that exist after the split.
4622 0 : fn tenant_shard_split_commit_inmem(
4623 0 : &self,
4624 0 : tenant_id: TenantId,
4625 0 : new_shard_count: ShardCount,
4626 0 : new_stripe_size: Option<ShardStripeSize>,
4627 0 : ) -> (
4628 0 : TenantShardSplitResponse,
4629 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
4630 0 : Vec<ReconcilerWaiter>,
4631 0 : ) {
4632 0 : let mut response = TenantShardSplitResponse {
4633 0 : new_shards: Vec::new(),
4634 0 : };
4635 0 : let mut child_locations = Vec::new();
4636 0 : let mut waiters = Vec::new();
4637 0 :
4638 0 : {
4639 0 : let mut locked = self.inner.write().unwrap();
4640 0 :
4641 0 : let parent_ids = locked
4642 0 : .tenants
4643 0 : .range(TenantShardId::tenant_range(tenant_id))
4644 0 : .map(|(shard_id, _)| *shard_id)
4645 0 : .collect::<Vec<_>>();
4646 0 :
4647 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4648 0 : for parent_id in parent_ids {
4649 0 : let child_ids = parent_id.split(new_shard_count);
4650 :
4651 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
4652 0 : let mut old_state = tenants
4653 0 : .remove(&parent_id)
4654 0 : .expect("It was present, we just split it");
4655 0 :
4656 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
4657 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
4658 0 : // nothing else can clear this.
4659 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
4660 :
4661 0 : let old_attached = old_state.intent.get_attached().unwrap();
4662 0 : old_state.intent.clear(scheduler);
4663 0 : let generation = old_state.generation.expect("Shard must have been attached");
4664 0 : (
4665 0 : old_attached,
4666 0 : generation,
4667 0 : old_state.policy.clone(),
4668 0 : old_state.shard,
4669 0 : old_state.config.clone(),
4670 0 : old_state.preferred_az().cloned(),
4671 0 : )
4672 0 : };
4673 0 :
4674 0 : let mut schedule_context = ScheduleContext::default();
4675 0 : for child in child_ids {
4676 0 : let mut child_shard = parent_ident;
4677 0 : child_shard.number = child.shard_number;
4678 0 : child_shard.count = child.shard_count;
4679 0 : if let Some(stripe_size) = new_stripe_size {
4680 0 : child_shard.stripe_size = stripe_size;
4681 0 : }
4682 :
4683 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
4684 0 : child_observed.insert(
4685 0 : pageserver,
4686 0 : ObservedStateLocation {
4687 0 : conf: Some(attached_location_conf(
4688 0 : generation,
4689 0 : &child_shard,
4690 0 : &config,
4691 0 : &policy,
4692 0 : )),
4693 0 : },
4694 0 : );
4695 0 :
4696 0 : let mut child_state =
4697 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
4698 0 : child_state.intent =
4699 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
4700 0 : child_state.observed = ObservedState {
4701 0 : locations: child_observed,
4702 0 : };
4703 0 : child_state.generation = Some(generation);
4704 0 : child_state.config = config.clone();
4705 0 :
4706 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
4707 0 : // as at this point in the split process we have succeeded and this part is infallible:
4708 0 : // we will never need to do any special recovery from this state.
4709 0 :
4710 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
4711 :
4712 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
4713 : // This is not fatal, because we've implicitly already got an attached
4714 : // location for the child shard. Failure here just means we couldn't
4715 : // find a secondary (e.g. because cluster is overloaded).
4716 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
4717 0 : }
4718 : // In the background, attach secondary locations for the new shards
4719 0 : if let Some(waiter) = self.maybe_reconcile_shard(
4720 0 : &mut child_state,
4721 0 : nodes,
4722 0 : ReconcilerPriority::High,
4723 0 : ) {
4724 0 : waiters.push(waiter);
4725 0 : }
4726 :
4727 0 : tenants.insert(child, child_state);
4728 0 : response.new_shards.push(child);
4729 : }
4730 : }
4731 0 : (response, child_locations, waiters)
4732 0 : }
4733 0 : }
4734 :
4735 0 : async fn tenant_shard_split_start_secondaries(
4736 0 : &self,
4737 0 : tenant_id: TenantId,
4738 0 : waiters: Vec<ReconcilerWaiter>,
4739 0 : ) {
4740 : // Wait for initial reconcile of child shards, this creates the secondary locations
4741 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
4742 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
4743 : // their secondaries couldn't be attached.
4744 0 : tracing::warn!("Failed to reconcile after split: {e}");
4745 0 : return;
4746 0 : }
4747 :
4748 : // Take the state lock to discover the attached & secondary intents for all shards
4749 0 : let (attached, secondary) = {
4750 0 : let locked = self.inner.read().unwrap();
4751 0 : let mut attached = Vec::new();
4752 0 : let mut secondary = Vec::new();
4753 :
4754 0 : for (tenant_shard_id, shard) in
4755 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4756 : {
4757 0 : let Some(node_id) = shard.intent.get_attached() else {
4758 : // Unexpected. Race with a PlacementPolicy change?
4759 0 : tracing::warn!(
4760 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
4761 : );
4762 0 : continue;
4763 : };
4764 :
4765 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
4766 : // No secondary location. Nothing for us to do.
4767 0 : continue;
4768 : };
4769 :
4770 0 : let attached_node = locked
4771 0 : .nodes
4772 0 : .get(node_id)
4773 0 : .expect("Pageservers may not be deleted while referenced");
4774 0 :
4775 0 : let secondary_node = locked
4776 0 : .nodes
4777 0 : .get(secondary_node_id)
4778 0 : .expect("Pageservers may not be deleted while referenced");
4779 0 :
4780 0 : attached.push((*tenant_shard_id, attached_node.clone()));
4781 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
4782 : }
4783 0 : (attached, secondary)
4784 0 : };
4785 0 :
4786 0 : if secondary.is_empty() {
4787 : // No secondary locations; nothing for us to do
4788 0 : return;
4789 0 : }
4790 :
4791 0 : for result in self
4792 0 : .tenant_for_shards_api(
4793 0 : attached,
4794 0 : |tenant_shard_id, client| async move {
4795 0 : client.tenant_heatmap_upload(tenant_shard_id).await
4796 0 : },
4797 0 : 1,
4798 0 : 1,
4799 0 : SHORT_RECONCILE_TIMEOUT,
4800 0 : &self.cancel,
4801 0 : )
4802 0 : .await
4803 : {
4804 0 : if let Err(e) = result {
4805 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
4806 0 : return;
4807 0 : }
4808 : }
4809 :
4810 0 : for result in self
4811 0 : .tenant_for_shards_api(
4812 0 : secondary,
4813 0 : |tenant_shard_id, client| async move {
4814 0 : client
4815 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
4816 0 : .await
4817 0 : },
4818 0 : 1,
4819 0 : 1,
4820 0 : SHORT_RECONCILE_TIMEOUT,
4821 0 : &self.cancel,
4822 0 : )
4823 0 : .await
4824 : {
4825 0 : if let Err(e) = result {
4826 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
4827 0 : return;
4828 0 : }
4829 : }
4830 0 : }
4831 :
4832 0 : pub(crate) async fn tenant_shard_split(
4833 0 : &self,
4834 0 : tenant_id: TenantId,
4835 0 : split_req: TenantShardSplitRequest,
4836 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
4837 : // TODO: return 503 if we get stuck waiting for this lock
4838 : // (issue https://github.com/neondatabase/neon/issues/7108)
4839 0 : let _tenant_lock = trace_exclusive_lock(
4840 0 : &self.tenant_op_locks,
4841 0 : tenant_id,
4842 0 : TenantOperations::ShardSplit,
4843 0 : )
4844 0 : .await;
4845 :
4846 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
4847 0 : let new_stripe_size = split_req.new_stripe_size;
4848 :
4849 : // Validate the request and construct parameters. This phase is fallible, but does not require
4850 : // rollback on errors, as it does no I/O and mutates no state.
4851 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
4852 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
4853 0 : ShardSplitAction::Split(params) => params,
4854 : };
4855 :
4856 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
4857 : // we must roll back.
4858 0 : let r = self
4859 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
4860 0 : .await;
4861 :
4862 0 : let (response, waiters) = match r {
4863 0 : Ok(r) => r,
4864 0 : Err(e) => {
4865 0 : // Split might be part-done, we must do work to abort it.
4866 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
4867 0 : self.abort_tx
4868 0 : .send(TenantShardSplitAbort {
4869 0 : tenant_id,
4870 0 : new_shard_count,
4871 0 : new_stripe_size,
4872 0 : _tenant_lock,
4873 0 : })
4874 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
4875 0 : .ok();
4876 0 : return Err(e);
4877 : }
4878 : };
4879 :
4880 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
4881 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
4882 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
4883 : // in [`Self::optimize_all`]
4884 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
4885 0 : .await;
4886 0 : Ok(response)
4887 0 : }
4888 :
4889 0 : fn prepare_tenant_shard_split(
4890 0 : &self,
4891 0 : tenant_id: TenantId,
4892 0 : split_req: TenantShardSplitRequest,
4893 0 : ) -> Result<ShardSplitAction, ApiError> {
4894 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
4895 0 : anyhow::anyhow!("failpoint")
4896 0 : )));
4897 :
4898 0 : let mut policy = None;
4899 0 : let mut config = None;
4900 0 : let mut shard_ident = None;
4901 0 : let mut preferred_az_id = None;
4902 : // Validate input, and calculate which shards we will create
4903 0 : let (old_shard_count, targets) =
4904 : {
4905 0 : let locked = self.inner.read().unwrap();
4906 0 :
4907 0 : let pageservers = locked.nodes.clone();
4908 0 :
4909 0 : let mut targets = Vec::new();
4910 0 :
4911 0 : // In case this is a retry, count how many already-split shards we found
4912 0 : let mut children_found = Vec::new();
4913 0 : let mut old_shard_count = None;
4914 :
4915 0 : for (tenant_shard_id, shard) in
4916 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4917 : {
4918 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
4919 : Ordering::Equal => {
4920 : // Already split this
4921 0 : children_found.push(*tenant_shard_id);
4922 0 : continue;
4923 : }
4924 : Ordering::Greater => {
4925 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4926 0 : "Requested count {} but already have shards at count {}",
4927 0 : split_req.new_shard_count,
4928 0 : shard.shard.count.count()
4929 0 : )));
4930 : }
4931 0 : Ordering::Less => {
4932 0 : // Fall through: this shard has lower count than requested,
4933 0 : // is a candidate for splitting.
4934 0 : }
4935 0 : }
4936 0 :
4937 0 : match old_shard_count {
4938 0 : None => old_shard_count = Some(shard.shard.count),
4939 0 : Some(old_shard_count) => {
4940 0 : if old_shard_count != shard.shard.count {
4941 : // We may hit this case if a caller asked for two splits to
4942 : // different sizes, before the first one is complete.
4943 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
4944 : // of shard_count=1 and shard_count=2 shards in the map.
4945 0 : return Err(ApiError::Conflict(
4946 0 : "Cannot split, currently mid-split".to_string(),
4947 0 : ));
4948 0 : }
4949 : }
4950 : }
4951 0 : if policy.is_none() {
4952 0 : policy = Some(shard.policy.clone());
4953 0 : }
4954 0 : if shard_ident.is_none() {
4955 0 : shard_ident = Some(shard.shard);
4956 0 : }
4957 0 : if config.is_none() {
4958 0 : config = Some(shard.config.clone());
4959 0 : }
4960 0 : if preferred_az_id.is_none() {
4961 0 : preferred_az_id = shard.preferred_az().cloned();
4962 0 : }
4963 :
4964 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4965 0 : tracing::info!(
4966 0 : "Tenant shard {} already has shard count {}",
4967 : tenant_shard_id,
4968 : split_req.new_shard_count
4969 : );
4970 0 : continue;
4971 0 : }
4972 :
4973 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4974 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4975 0 : ))?;
4976 :
4977 0 : let node = pageservers
4978 0 : .get(&node_id)
4979 0 : .expect("Pageservers may not be deleted while referenced");
4980 0 :
4981 0 : targets.push(ShardSplitTarget {
4982 0 : parent_id: *tenant_shard_id,
4983 0 : node: node.clone(),
4984 0 : child_ids: tenant_shard_id
4985 0 : .split(ShardCount::new(split_req.new_shard_count)),
4986 0 : });
4987 : }
4988 :
4989 0 : if targets.is_empty() {
4990 0 : if children_found.len() == split_req.new_shard_count as usize {
4991 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4992 0 : new_shards: children_found,
4993 0 : }));
4994 : } else {
4995 : // No shards found to split, and no existing children found: the
4996 : // tenant doesn't exist at all.
4997 0 : return Err(ApiError::NotFound(
4998 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4999 0 : ));
5000 : }
5001 0 : }
5002 0 :
5003 0 : (old_shard_count, targets)
5004 0 : };
5005 0 :
5006 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
5007 0 : let old_shard_count = old_shard_count.unwrap();
5008 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
5009 : // This ShardIdentity will be used as the template for all children, so this implicitly
5010 : // applies the new stripe size to the children.
5011 0 : let mut shard_ident = shard_ident.unwrap();
5012 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
5013 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5014 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
5015 0 : shard_ident.stripe_size
5016 0 : )));
5017 0 : }
5018 0 :
5019 0 : shard_ident.stripe_size = new_stripe_size;
5020 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
5021 0 : shard_ident
5022 : } else {
5023 0 : shard_ident.unwrap()
5024 : };
5025 0 : let policy = policy.unwrap();
5026 0 : let config = config.unwrap();
5027 0 :
5028 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
5029 0 : old_shard_count,
5030 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
5031 0 : new_stripe_size: split_req.new_stripe_size,
5032 0 : targets,
5033 0 : policy,
5034 0 : config,
5035 0 : shard_ident,
5036 0 : preferred_az_id,
5037 0 : })))
5038 0 : }
5039 :
5040 0 : async fn do_tenant_shard_split(
5041 0 : &self,
5042 0 : tenant_id: TenantId,
5043 0 : params: Box<ShardSplitParams>,
5044 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
5045 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
5046 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
5047 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
5048 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
5049 0 : // (https://github.com/neondatabase/neon/issues/6676)
5050 0 :
5051 0 : let ShardSplitParams {
5052 0 : old_shard_count,
5053 0 : new_shard_count,
5054 0 : new_stripe_size,
5055 0 : mut targets,
5056 0 : policy,
5057 0 : config,
5058 0 : shard_ident,
5059 0 : preferred_az_id,
5060 0 : } = *params;
5061 :
5062 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
5063 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
5064 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
5065 : // at the time of split.
5066 0 : let waiters = {
5067 0 : let mut locked = self.inner.write().unwrap();
5068 0 : let mut waiters = Vec::new();
5069 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5070 0 : for target in &mut targets {
5071 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
5072 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5073 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5074 0 : "Shard {} not found",
5075 0 : target.parent_id
5076 0 : )));
5077 : };
5078 :
5079 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
5080 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5081 0 : return Err(ApiError::Conflict(format!(
5082 0 : "Shard {} unexpectedly rescheduled during split",
5083 0 : target.parent_id
5084 0 : )));
5085 0 : }
5086 0 :
5087 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
5088 0 : shard.intent.clear_secondary(scheduler);
5089 :
5090 : // Run Reconciler to execute detach fo secondary locations.
5091 0 : if let Some(waiter) =
5092 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5093 0 : {
5094 0 : waiters.push(waiter);
5095 0 : }
5096 : }
5097 0 : waiters
5098 0 : };
5099 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
5100 :
5101 : // Before creating any new child shards in memory or on the pageservers, persist them: this
5102 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
5103 : // acts as the protection against two concurrent attempts to split: one of them will get a database
5104 : // error trying to insert the child shards.
5105 0 : let mut child_tsps = Vec::new();
5106 0 : for target in &targets {
5107 0 : let mut this_child_tsps = Vec::new();
5108 0 : for child in &target.child_ids {
5109 0 : let mut child_shard = shard_ident;
5110 0 : child_shard.number = child.shard_number;
5111 0 : child_shard.count = child.shard_count;
5112 0 :
5113 0 : tracing::info!(
5114 0 : "Create child shard persistence with stripe size {}",
5115 : shard_ident.stripe_size.0
5116 : );
5117 :
5118 0 : this_child_tsps.push(TenantShardPersistence {
5119 0 : tenant_id: child.tenant_id.to_string(),
5120 0 : shard_number: child.shard_number.0 as i32,
5121 0 : shard_count: child.shard_count.literal() as i32,
5122 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
5123 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
5124 0 : // populate the correct generation as part of its transaction, to protect us
5125 0 : // against racing with changes in the state of the parent.
5126 0 : generation: None,
5127 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
5128 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
5129 0 : config: serde_json::to_string(&config).unwrap(),
5130 0 : splitting: SplitState::Splitting,
5131 0 :
5132 0 : // Scheduling policies and preferred AZ do not carry through to children
5133 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
5134 0 : .unwrap(),
5135 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
5136 0 : });
5137 0 : }
5138 :
5139 0 : child_tsps.push((target.parent_id, this_child_tsps));
5140 : }
5141 :
5142 0 : if let Err(e) = self
5143 0 : .persistence
5144 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
5145 0 : .await
5146 : {
5147 0 : match e {
5148 : DatabaseError::Query(diesel::result::Error::DatabaseError(
5149 : DatabaseErrorKind::UniqueViolation,
5150 : _,
5151 : )) => {
5152 : // Inserting a child shard violated a unique constraint: we raced with another call to
5153 : // this function
5154 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
5155 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
5156 : }
5157 0 : _ => return Err(ApiError::InternalServerError(e.into())),
5158 : }
5159 0 : }
5160 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
5161 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5162 0 : ));
5163 :
5164 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
5165 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
5166 : // is not set in memory, then it was not persisted.
5167 : {
5168 0 : let mut locked = self.inner.write().unwrap();
5169 0 : for target in &targets {
5170 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
5171 0 : parent_shard.splitting = SplitState::Splitting;
5172 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
5173 0 : // split operation.
5174 0 : parent_shard
5175 0 : .observed
5176 0 : .locations
5177 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
5178 0 : }
5179 : }
5180 : }
5181 :
5182 : // TODO: issue split calls concurrently (this only matters once we're splitting
5183 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
5184 :
5185 0 : for target in &targets {
5186 : let ShardSplitTarget {
5187 0 : parent_id,
5188 0 : node,
5189 0 : child_ids,
5190 0 : } = target;
5191 0 : let client = PageserverClient::new(
5192 0 : node.get_id(),
5193 0 : node.base_url(),
5194 0 : self.config.pageserver_jwt_token.as_deref(),
5195 0 : );
5196 0 : let response = client
5197 0 : .tenant_shard_split(
5198 0 : *parent_id,
5199 0 : TenantShardSplitRequest {
5200 0 : new_shard_count: new_shard_count.literal(),
5201 0 : new_stripe_size,
5202 0 : },
5203 0 : )
5204 0 : .await
5205 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
5206 :
5207 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
5208 0 : "failpoint".to_string()
5209 0 : )));
5210 :
5211 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
5212 :
5213 0 : tracing::info!(
5214 0 : "Split {} into {}",
5215 0 : parent_id,
5216 0 : response
5217 0 : .new_shards
5218 0 : .iter()
5219 0 : .map(|s| format!("{:?}", s))
5220 0 : .collect::<Vec<_>>()
5221 0 : .join(",")
5222 : );
5223 :
5224 0 : if &response.new_shards != child_ids {
5225 : // This should never happen: the pageserver should agree with us on how shard splits work.
5226 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5227 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
5228 0 : parent_id,
5229 0 : response.new_shards,
5230 0 : child_ids
5231 0 : )));
5232 0 : }
5233 : }
5234 :
5235 : // TODO: if the pageserver restarted concurrently with our split API call,
5236 : // the actual generation of the child shard might differ from the generation
5237 : // we expect it to have. In order for our in-database generation to end up
5238 : // correct, we should carry the child generation back in the response and apply it here
5239 : // in complete_shard_split (and apply the correct generation in memory)
5240 : // (or, we can carry generation in the request and reject the request if
5241 : // it doesn't match, but that requires more retry logic on this side)
5242 :
5243 0 : self.persistence
5244 0 : .complete_shard_split(tenant_id, old_shard_count)
5245 0 : .await?;
5246 :
5247 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5248 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5249 0 : ));
5250 :
5251 : // Replace all the shards we just split with their children: this phase is infallible.
5252 0 : let (response, child_locations, waiters) =
5253 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5254 0 :
5255 0 : // Send compute notifications for all the new shards
5256 0 : let mut failed_notifications = Vec::new();
5257 0 : for (child_id, child_ps, stripe_size) in child_locations {
5258 0 : if let Err(e) = self
5259 0 : .compute_hook
5260 0 : .notify(
5261 0 : compute_hook::ShardUpdate {
5262 0 : tenant_shard_id: child_id,
5263 0 : node_id: child_ps,
5264 0 : stripe_size,
5265 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5266 0 : },
5267 0 : &self.cancel,
5268 0 : )
5269 0 : .await
5270 : {
5271 0 : tracing::warn!(
5272 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5273 : child_id,
5274 : child_ps
5275 : );
5276 0 : failed_notifications.push(child_id);
5277 0 : }
5278 : }
5279 :
5280 : // If we failed any compute notifications, make a note to retry later.
5281 0 : if !failed_notifications.is_empty() {
5282 0 : let mut locked = self.inner.write().unwrap();
5283 0 : for failed in failed_notifications {
5284 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5285 0 : shard.pending_compute_notification = true;
5286 0 : }
5287 : }
5288 0 : }
5289 :
5290 0 : Ok((response, waiters))
5291 0 : }
5292 :
5293 0 : pub(crate) async fn tenant_shard_migrate(
5294 0 : &self,
5295 0 : tenant_shard_id: TenantShardId,
5296 0 : migrate_req: TenantShardMigrateRequest,
5297 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5298 0 : let waiter = {
5299 0 : let mut locked = self.inner.write().unwrap();
5300 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5301 :
5302 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5303 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5304 0 : "Node {} not found",
5305 0 : migrate_req.node_id
5306 0 : )));
5307 : };
5308 :
5309 0 : if !node.is_available() {
5310 : // Warn but proceed: the caller may intend to manually adjust the placement of
5311 : // a shard even if the node is down, e.g. if intervening during an incident.
5312 0 : tracing::warn!("Migrating to unavailable node {node}");
5313 0 : }
5314 :
5315 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5316 0 : return Err(ApiError::NotFound(
5317 0 : anyhow::anyhow!("Tenant shard not found").into(),
5318 0 : ));
5319 : };
5320 :
5321 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5322 : // No-op case: we will still proceed to wait for reconciliation in case it is
5323 : // incomplete from an earlier update to the intent.
5324 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
5325 : } else {
5326 0 : let old_attached = *shard.intent.get_attached();
5327 0 :
5328 0 : match shard.policy {
5329 0 : PlacementPolicy::Attached(n) => {
5330 0 : // If our new attached node was a secondary, it no longer should be.
5331 0 : shard
5332 0 : .intent
5333 0 : .remove_secondary(scheduler, migrate_req.node_id);
5334 0 :
5335 0 : shard
5336 0 : .intent
5337 0 : .set_attached(scheduler, Some(migrate_req.node_id));
5338 :
5339 : // If we were already attached to something, demote that to a secondary
5340 0 : if let Some(old_attached) = old_attached {
5341 0 : if n > 0 {
5342 : // Remove other secondaries to make room for the location we'll demote
5343 0 : while shard.intent.get_secondary().len() >= n {
5344 0 : shard.intent.pop_secondary(scheduler);
5345 0 : }
5346 :
5347 0 : shard.intent.push_secondary(scheduler, old_attached);
5348 0 : }
5349 0 : }
5350 : }
5351 0 : PlacementPolicy::Secondary => {
5352 0 : shard.intent.clear(scheduler);
5353 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5354 0 : }
5355 : PlacementPolicy::Detached => {
5356 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5357 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
5358 0 : )));
5359 : }
5360 : }
5361 :
5362 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
5363 0 : shard.sequence = shard.sequence.next();
5364 : }
5365 :
5366 0 : let reconciler_config = match migrate_req.migration_config {
5367 0 : Some(cfg) => (&cfg).into(),
5368 0 : None => ReconcilerConfig::new(ReconcilerPriority::High),
5369 : };
5370 :
5371 0 : self.maybe_configured_reconcile_shard(shard, nodes, reconciler_config)
5372 : };
5373 :
5374 0 : if let Some(waiter) = waiter {
5375 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5376 : } else {
5377 0 : tracing::info!("Migration is a no-op");
5378 : }
5379 :
5380 0 : Ok(TenantShardMigrateResponse {})
5381 0 : }
5382 :
5383 0 : pub(crate) async fn tenant_shard_migrate_secondary(
5384 0 : &self,
5385 0 : tenant_shard_id: TenantShardId,
5386 0 : migrate_req: TenantShardMigrateRequest,
5387 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5388 0 : let waiter = {
5389 0 : let mut locked = self.inner.write().unwrap();
5390 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5391 :
5392 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5393 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5394 0 : "Node {} not found",
5395 0 : migrate_req.node_id
5396 0 : )));
5397 : };
5398 :
5399 0 : if !node.is_available() {
5400 : // Warn but proceed: the caller may intend to manually adjust the placement of
5401 : // a shard even if the node is down, e.g. if intervening during an incident.
5402 0 : tracing::warn!("Migrating to unavailable node {node}");
5403 0 : }
5404 :
5405 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5406 0 : return Err(ApiError::NotFound(
5407 0 : anyhow::anyhow!("Tenant shard not found").into(),
5408 0 : ));
5409 : };
5410 :
5411 0 : if shard.intent.get_secondary().len() == 1
5412 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
5413 : {
5414 0 : tracing::info!(
5415 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
5416 : shard.intent
5417 : );
5418 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5419 0 : tracing::info!(
5420 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
5421 : );
5422 : } else {
5423 0 : let old_secondaries = shard.intent.get_secondary().clone();
5424 0 : for secondary in old_secondaries {
5425 0 : shard.intent.remove_secondary(scheduler, secondary);
5426 0 : }
5427 :
5428 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5429 0 : shard.sequence = shard.sequence.next();
5430 0 : tracing::info!(
5431 0 : "Migrating secondary to {node}: new intent {:?}",
5432 : shard.intent
5433 : );
5434 : }
5435 :
5436 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5437 : };
5438 :
5439 0 : if let Some(waiter) = waiter {
5440 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5441 : } else {
5442 0 : tracing::info!("Migration is a no-op");
5443 : }
5444 :
5445 0 : Ok(TenantShardMigrateResponse {})
5446 0 : }
5447 :
5448 : /// 'cancel' in this context means cancel any ongoing reconcile
5449 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
5450 0 : &self,
5451 0 : tenant_shard_id: TenantShardId,
5452 0 : ) -> Result<(), ApiError> {
5453 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
5454 0 : let waiter = {
5455 0 : let locked = self.inner.write().unwrap();
5456 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5457 0 : return Err(ApiError::NotFound(
5458 0 : anyhow::anyhow!("Tenant shard not found").into(),
5459 0 : ));
5460 : };
5461 :
5462 0 : let waiter = shard.get_waiter();
5463 0 : match waiter {
5464 : None => {
5465 0 : tracing::info!("Shard does not have an ongoing Reconciler");
5466 0 : return Ok(());
5467 : }
5468 0 : Some(waiter) => {
5469 0 : tracing::info!("Cancelling Reconciler");
5470 0 : shard.cancel_reconciler();
5471 0 : waiter
5472 0 : }
5473 0 : }
5474 0 : };
5475 0 :
5476 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
5477 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
5478 0 : // the Reconciler is misbehaving and not respecting the cancellation token
5479 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
5480 0 : .await?;
5481 :
5482 0 : Ok(())
5483 0 : }
5484 :
5485 : /// This is for debug/support only: we simply drop all state for a tenant, without
5486 : /// detaching or deleting it on pageservers.
5487 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
5488 0 : self.persistence.delete_tenant(tenant_id).await?;
5489 :
5490 0 : let mut locked = self.inner.write().unwrap();
5491 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
5492 0 : let mut shards = Vec::new();
5493 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
5494 0 : shards.push(*tenant_shard_id);
5495 0 : }
5496 :
5497 0 : for shard_id in shards {
5498 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
5499 0 : shard.intent.clear(scheduler);
5500 0 : }
5501 : }
5502 :
5503 0 : Ok(())
5504 0 : }
5505 :
5506 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
5507 : /// tenant with a very high generation number so that it will see the existing data.
5508 0 : pub(crate) async fn tenant_import(
5509 0 : &self,
5510 0 : tenant_id: TenantId,
5511 0 : ) -> Result<TenantCreateResponse, ApiError> {
5512 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
5513 0 : let maybe_node = {
5514 0 : self.inner
5515 0 : .read()
5516 0 : .unwrap()
5517 0 : .nodes
5518 0 : .values()
5519 0 : .find(|n| n.is_available())
5520 0 : .cloned()
5521 : };
5522 0 : let Some(node) = maybe_node else {
5523 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
5524 : };
5525 :
5526 0 : let client = PageserverClient::new(
5527 0 : node.get_id(),
5528 0 : node.base_url(),
5529 0 : self.config.pageserver_jwt_token.as_deref(),
5530 0 : );
5531 :
5532 0 : let scan_result = client
5533 0 : .tenant_scan_remote_storage(tenant_id)
5534 0 : .await
5535 0 : .map_err(|e| passthrough_api_error(&node, e))?;
5536 :
5537 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
5538 0 : let Some(shard_count) = scan_result
5539 0 : .shards
5540 0 : .iter()
5541 0 : .map(|s| s.tenant_shard_id.shard_count)
5542 0 : .max()
5543 : else {
5544 0 : return Err(ApiError::NotFound(
5545 0 : anyhow::anyhow!("No shards found").into(),
5546 0 : ));
5547 : };
5548 :
5549 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
5550 : // to
5551 0 : let generation = scan_result
5552 0 : .shards
5553 0 : .iter()
5554 0 : .map(|s| s.generation)
5555 0 : .max()
5556 0 : .expect("We already validated >0 shards");
5557 0 :
5558 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
5559 0 : // only work if they were using the default stripe size.
5560 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
5561 :
5562 0 : let (response, waiters) = self
5563 0 : .do_tenant_create(TenantCreateRequest {
5564 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
5565 0 : generation,
5566 0 :
5567 0 : shard_parameters: ShardParameters {
5568 0 : count: shard_count,
5569 0 : stripe_size,
5570 0 : },
5571 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
5572 0 : config: TenantConfig::default(),
5573 0 : })
5574 0 : .await?;
5575 :
5576 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
5577 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
5578 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
5579 : // reconcile, as reconciliation includes notifying compute.
5580 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
5581 0 : }
5582 :
5583 0 : Ok(response)
5584 0 : }
5585 :
5586 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
5587 : /// we don't have to make TenantShard clonable in the return path.
5588 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5589 0 : let serialized = {
5590 0 : let locked = self.inner.read().unwrap();
5591 0 : let result = locked.tenants.values().collect::<Vec<_>>();
5592 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
5593 : };
5594 :
5595 0 : hyper::Response::builder()
5596 0 : .status(hyper::StatusCode::OK)
5597 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5598 0 : .body(hyper::Body::from(serialized))
5599 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5600 0 : }
5601 :
5602 : /// Check the consistency of in-memory state vs. persistent state, and check that the
5603 : /// scheduler's statistics are up to date.
5604 : ///
5605 : /// These consistency checks expect an **idle** system. If changes are going on while
5606 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
5607 : /// checks, but not suitable for running continuously in the background in the field.
5608 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
5609 0 : let (mut expect_nodes, mut expect_shards) = {
5610 0 : let locked = self.inner.read().unwrap();
5611 0 :
5612 0 : locked
5613 0 : .scheduler
5614 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
5615 0 : .context("Scheduler checks")
5616 0 : .map_err(ApiError::InternalServerError)?;
5617 :
5618 0 : let expect_nodes = locked
5619 0 : .nodes
5620 0 : .values()
5621 0 : .map(|n| n.to_persistent())
5622 0 : .collect::<Vec<_>>();
5623 0 :
5624 0 : let expect_shards = locked
5625 0 : .tenants
5626 0 : .values()
5627 0 : .map(|t| t.to_persistent())
5628 0 : .collect::<Vec<_>>();
5629 :
5630 : // This method can only validate the state of an idle system: if a reconcile is in
5631 : // progress, fail out early to avoid giving false errors on state that won't match
5632 : // between database and memory under a ReconcileResult is processed.
5633 0 : for t in locked.tenants.values() {
5634 0 : if t.reconciler.is_some() {
5635 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5636 0 : "Shard {} reconciliation in progress",
5637 0 : t.tenant_shard_id
5638 0 : )));
5639 0 : }
5640 : }
5641 :
5642 0 : (expect_nodes, expect_shards)
5643 : };
5644 :
5645 0 : let mut nodes = self.persistence.list_nodes().await?;
5646 0 : expect_nodes.sort_by_key(|n| n.node_id);
5647 0 : nodes.sort_by_key(|n| n.node_id);
5648 :
5649 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
5650 0 : let node_result = if nodes != expect_nodes {
5651 0 : tracing::error!("Consistency check failed on nodes.");
5652 0 : tracing::error!(
5653 0 : "Nodes in memory: {}",
5654 0 : serde_json::to_string(&expect_nodes)
5655 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5656 : );
5657 0 : tracing::error!(
5658 0 : "Nodes in database: {}",
5659 0 : serde_json::to_string(&nodes)
5660 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5661 : );
5662 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
5663 0 : "Node consistency failure"
5664 0 : )))
5665 : } else {
5666 0 : Ok(())
5667 : };
5668 :
5669 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
5670 0 : persistent_shards
5671 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5672 0 :
5673 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
5674 :
5675 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
5676 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
5677 : // fields are added, before doing a comparison.
5678 0 : for tsp in &mut persistent_shards {
5679 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
5680 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
5681 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
5682 : }
5683 :
5684 0 : if persistent_shards != expect_shards {
5685 0 : tracing::error!("Consistency check failed on shards.");
5686 :
5687 0 : tracing::error!(
5688 0 : "Shards in memory: {}",
5689 0 : serde_json::to_string(&expect_shards)
5690 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5691 : );
5692 0 : tracing::error!(
5693 0 : "Shards in database: {}",
5694 0 : serde_json::to_string(&persistent_shards)
5695 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5696 : );
5697 :
5698 : // The total dump log lines above are useful in testing but in the field grafana will
5699 : // usually just drop them because they're so large. So we also do some explicit logging
5700 : // of just the diffs.
5701 0 : let persistent_shards = persistent_shards
5702 0 : .into_iter()
5703 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5704 0 : .collect::<HashMap<_, _>>();
5705 0 : let expect_shards = expect_shards
5706 0 : .into_iter()
5707 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
5708 0 : .collect::<HashMap<_, _>>();
5709 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
5710 0 : match expect_shards.get(tenant_shard_id) {
5711 : None => {
5712 0 : tracing::error!(
5713 0 : "Shard {} found in database but not in memory",
5714 : tenant_shard_id
5715 : );
5716 : }
5717 0 : Some(expect_tsp) => {
5718 0 : if expect_tsp != persistent_tsp {
5719 0 : tracing::error!(
5720 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
5721 0 : tenant_shard_id,
5722 0 : serde_json::to_string(expect_tsp).unwrap(),
5723 0 : serde_json::to_string(&persistent_tsp).unwrap()
5724 : );
5725 0 : }
5726 : }
5727 : }
5728 : }
5729 :
5730 : // Having already logged any differences, log any shards that simply aren't present in the database
5731 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
5732 0 : if !persistent_shards.contains_key(tenant_shard_id) {
5733 0 : tracing::error!(
5734 0 : "Shard {} found in memory but not in database: {}",
5735 0 : tenant_shard_id,
5736 0 : serde_json::to_string(memory_tsp)
5737 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5738 : );
5739 0 : }
5740 : }
5741 :
5742 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5743 0 : "Shard consistency failure"
5744 0 : )));
5745 0 : }
5746 0 :
5747 0 : node_result
5748 0 : }
5749 :
5750 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
5751 : /// we don't have to make TenantShard clonable in the return path.
5752 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
5753 0 : let serialized = {
5754 0 : let locked = self.inner.read().unwrap();
5755 0 : serde_json::to_string(&locked.scheduler)
5756 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
5757 : };
5758 :
5759 0 : hyper::Response::builder()
5760 0 : .status(hyper::StatusCode::OK)
5761 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
5762 0 : .body(hyper::Body::from(serialized))
5763 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
5764 0 : }
5765 :
5766 : /// This is for debug/support only: we simply drop all state for a tenant, without
5767 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
5768 : /// tenants that were on this node.
5769 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
5770 0 : self.persistence.delete_node(node_id).await?;
5771 :
5772 0 : let mut locked = self.inner.write().unwrap();
5773 :
5774 0 : for shard in locked.tenants.values_mut() {
5775 0 : shard.deref_node(node_id);
5776 0 : shard.observed.locations.remove(&node_id);
5777 0 : }
5778 :
5779 0 : let mut nodes = (*locked.nodes).clone();
5780 0 : nodes.remove(&node_id);
5781 0 : locked.nodes = Arc::new(nodes);
5782 0 : metrics::METRICS_REGISTRY
5783 0 : .metrics_group
5784 0 : .storage_controller_pageserver_nodes
5785 0 : .set(locked.nodes.len() as i64);
5786 0 :
5787 0 : locked.scheduler.node_remove(node_id);
5788 0 :
5789 0 : Ok(())
5790 0 : }
5791 :
5792 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
5793 : /// that we don't leave any bad state behind in the storage controller, but unclean
5794 : /// in the sense that we are not carefully draining the node.
5795 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
5796 0 : let _node_lock =
5797 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
5798 :
5799 : // 1. Atomically update in-memory state:
5800 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
5801 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
5802 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
5803 : // re-insert references to this node into the ObservedState of shards
5804 : // - drop the node from the scheduler
5805 : {
5806 0 : let mut locked = self.inner.write().unwrap();
5807 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5808 0 :
5809 0 : {
5810 0 : let mut nodes_mut = (*nodes).deref().clone();
5811 0 : match nodes_mut.get_mut(&node_id) {
5812 0 : Some(node) => {
5813 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
5814 0 : // if we crash it would not be desirable to leave the node paused after a restart.
5815 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
5816 0 : }
5817 : None => {
5818 0 : tracing::info!(
5819 0 : "Node not found: presuming this is a retry and returning success"
5820 : );
5821 0 : return Ok(());
5822 : }
5823 : }
5824 :
5825 0 : *nodes = Arc::new(nodes_mut);
5826 : }
5827 :
5828 0 : for (_tenant_id, mut schedule_context, shards) in
5829 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
5830 : {
5831 0 : for shard in shards {
5832 0 : if shard.deref_node(node_id) {
5833 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
5834 : // TODO: implement force flag to remove a node even if we can't reschedule
5835 : // a tenant
5836 0 : tracing::error!(
5837 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
5838 : shard.tenant_shard_id
5839 : );
5840 0 : return Err(e.into());
5841 : } else {
5842 0 : tracing::info!(
5843 0 : "Rescheduled shard {} away from node during deletion",
5844 : shard.tenant_shard_id
5845 : )
5846 : }
5847 :
5848 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
5849 0 : }
5850 :
5851 : // Here we remove an existing observed location for the node we're removing, and it will
5852 : // not be re-added by a reconciler's completion because we filter out removed nodes in
5853 : // process_result.
5854 : //
5855 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
5856 : // means any reconciles we spawned will know about the node we're deleting, enabling them
5857 : // to do live migrations if it's still online.
5858 0 : shard.observed.locations.remove(&node_id);
5859 : }
5860 : }
5861 :
5862 0 : scheduler.node_remove(node_id);
5863 0 :
5864 0 : {
5865 0 : let mut nodes_mut = (**nodes).clone();
5866 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
5867 0 : // Ensure that any reconciler holding an Arc<> to this node will
5868 0 : // drop out when trying to RPC to it (setting Offline state sets the
5869 0 : // cancellation token on the Node object).
5870 0 : removed_node.set_availability(NodeAvailability::Offline);
5871 0 : }
5872 0 : *nodes = Arc::new(nodes_mut);
5873 0 : metrics::METRICS_REGISTRY
5874 0 : .metrics_group
5875 0 : .storage_controller_pageserver_nodes
5876 0 : .set(nodes.len() as i64);
5877 0 : }
5878 0 : }
5879 0 :
5880 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
5881 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
5882 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
5883 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
5884 0 : // that exists.
5885 0 :
5886 0 : // 2. Actually delete the node from the database and from in-memory state
5887 0 : tracing::info!("Deleting node from database");
5888 0 : self.persistence.delete_node(node_id).await?;
5889 :
5890 0 : Ok(())
5891 0 : }
5892 :
5893 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
5894 0 : let nodes = {
5895 0 : self.inner
5896 0 : .read()
5897 0 : .unwrap()
5898 0 : .nodes
5899 0 : .values()
5900 0 : .cloned()
5901 0 : .collect::<Vec<_>>()
5902 0 : };
5903 0 :
5904 0 : Ok(nodes)
5905 0 : }
5906 :
5907 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
5908 0 : self.inner
5909 0 : .read()
5910 0 : .unwrap()
5911 0 : .nodes
5912 0 : .get(&node_id)
5913 0 : .cloned()
5914 0 : .ok_or(ApiError::NotFound(
5915 0 : format!("Node {node_id} not registered").into(),
5916 0 : ))
5917 0 : }
5918 :
5919 0 : pub(crate) async fn get_node_shards(
5920 0 : &self,
5921 0 : node_id: NodeId,
5922 0 : ) -> Result<NodeShardResponse, ApiError> {
5923 0 : let locked = self.inner.read().unwrap();
5924 0 : let mut shards = Vec::new();
5925 0 : for (tid, tenant) in locked.tenants.iter() {
5926 0 : let is_intended_secondary = match (
5927 0 : tenant.intent.get_attached() == &Some(node_id),
5928 0 : tenant.intent.get_secondary().contains(&node_id),
5929 0 : ) {
5930 : (true, true) => {
5931 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5932 0 : "{} attached as primary+secondary on the same node",
5933 0 : tid
5934 0 : )));
5935 : }
5936 0 : (true, false) => Some(false),
5937 0 : (false, true) => Some(true),
5938 0 : (false, false) => None,
5939 : };
5940 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
5941 0 : tenant.observed.locations.get(&node_id)
5942 : {
5943 0 : Some(conf.secondary_conf.is_some())
5944 : } else {
5945 0 : None
5946 : };
5947 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
5948 0 : shards.push(NodeShard {
5949 0 : tenant_shard_id: *tid,
5950 0 : is_intended_secondary,
5951 0 : is_observed_secondary,
5952 0 : });
5953 0 : }
5954 : }
5955 0 : Ok(NodeShardResponse { node_id, shards })
5956 0 : }
5957 :
5958 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
5959 0 : self.persistence.get_leader().await
5960 0 : }
5961 :
5962 0 : pub(crate) async fn node_register(
5963 0 : &self,
5964 0 : register_req: NodeRegisterRequest,
5965 0 : ) -> Result<(), ApiError> {
5966 0 : let _node_lock = trace_exclusive_lock(
5967 0 : &self.node_op_locks,
5968 0 : register_req.node_id,
5969 0 : NodeOperations::Register,
5970 0 : )
5971 0 : .await;
5972 :
5973 : #[derive(PartialEq)]
5974 : enum RegistrationStatus {
5975 : UpToDate,
5976 : NeedUpdate,
5977 : Mismatched,
5978 : New,
5979 : }
5980 :
5981 0 : let registration_status = {
5982 0 : let locked = self.inner.read().unwrap();
5983 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
5984 0 : if node.registration_match(®ister_req) {
5985 0 : if node.need_update(®ister_req) {
5986 0 : RegistrationStatus::NeedUpdate
5987 : } else {
5988 0 : RegistrationStatus::UpToDate
5989 : }
5990 : } else {
5991 0 : RegistrationStatus::Mismatched
5992 : }
5993 : } else {
5994 0 : RegistrationStatus::New
5995 : }
5996 : };
5997 :
5998 0 : match registration_status {
5999 : RegistrationStatus::UpToDate => {
6000 0 : tracing::info!(
6001 0 : "Node {} re-registered with matching address and is up to date",
6002 : register_req.node_id
6003 : );
6004 :
6005 0 : return Ok(());
6006 : }
6007 : RegistrationStatus::Mismatched => {
6008 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
6009 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
6010 : // a fixed address through the lifetime of a node.
6011 0 : tracing::warn!(
6012 0 : "Node {} tried to register with different address",
6013 : register_req.node_id
6014 : );
6015 0 : return Err(ApiError::Conflict(
6016 0 : "Node is already registered with different address".to_string(),
6017 0 : ));
6018 : }
6019 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
6020 0 : // fallthrough
6021 0 : }
6022 0 : }
6023 0 :
6024 0 : // We do not require that a node is actually online when registered (it will start life
6025 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
6026 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
6027 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
6028 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
6029 0 : if tokio::net::lookup_host(format!(
6030 0 : "{}:{}",
6031 0 : register_req.listen_http_addr, register_req.listen_http_port
6032 0 : ))
6033 0 : .await
6034 0 : .is_err()
6035 : {
6036 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
6037 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
6038 : // we return a soft 503 error, to encourage callers to retry past transient issues.
6039 0 : return Err(ApiError::ResourceUnavailable(
6040 0 : format!(
6041 0 : "Node {} tried to register with unknown DNS name '{}'",
6042 0 : register_req.node_id, register_req.listen_http_addr
6043 0 : )
6044 0 : .into(),
6045 0 : ));
6046 0 : }
6047 0 :
6048 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
6049 0 : return Err(ApiError::PreconditionFailed(
6050 0 : format!(
6051 0 : "Node {} has no https port, but use_https is enabled",
6052 0 : register_req.node_id
6053 0 : )
6054 0 : .into(),
6055 0 : ));
6056 0 : }
6057 0 :
6058 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
6059 0 : // This ensures that before we use it for anything or expose it via any external
6060 0 : // API, it is guaranteed to be available after a restart.
6061 0 : let new_node = Node::new(
6062 0 : register_req.node_id,
6063 0 : register_req.listen_http_addr,
6064 0 : register_req.listen_http_port,
6065 0 : register_req.listen_https_port,
6066 0 : register_req.listen_pg_addr,
6067 0 : register_req.listen_pg_port,
6068 0 : register_req.availability_zone_id.clone(),
6069 0 : self.config.use_https_pageserver_api,
6070 0 : );
6071 0 : let new_node = match new_node {
6072 0 : Ok(new_node) => new_node,
6073 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
6074 : };
6075 :
6076 0 : match registration_status {
6077 0 : RegistrationStatus::New => self.persistence.insert_node(&new_node).await?,
6078 : RegistrationStatus::NeedUpdate => {
6079 0 : self.persistence
6080 0 : .update_node_on_registration(
6081 0 : register_req.node_id,
6082 0 : register_req.listen_https_port,
6083 0 : )
6084 0 : .await?
6085 : }
6086 0 : _ => unreachable!("Other statuses have been processed earlier"),
6087 : }
6088 :
6089 0 : let mut locked = self.inner.write().unwrap();
6090 0 : let mut new_nodes = (*locked.nodes).clone();
6091 0 :
6092 0 : locked.scheduler.node_upsert(&new_node);
6093 0 : new_nodes.insert(register_req.node_id, new_node);
6094 0 :
6095 0 : locked.nodes = Arc::new(new_nodes);
6096 0 :
6097 0 : metrics::METRICS_REGISTRY
6098 0 : .metrics_group
6099 0 : .storage_controller_pageserver_nodes
6100 0 : .set(locked.nodes.len() as i64);
6101 0 :
6102 0 : match registration_status {
6103 : RegistrationStatus::New => {
6104 0 : tracing::info!(
6105 0 : "Registered pageserver {} ({}), now have {} pageservers",
6106 0 : register_req.node_id,
6107 0 : register_req.availability_zone_id,
6108 0 : locked.nodes.len()
6109 : );
6110 : }
6111 : RegistrationStatus::NeedUpdate => {
6112 0 : tracing::info!(
6113 0 : "Re-registered and updated node {} ({})",
6114 : register_req.node_id,
6115 : register_req.availability_zone_id,
6116 : );
6117 : }
6118 0 : _ => unreachable!("Other statuses have been processed earlier"),
6119 : }
6120 0 : Ok(())
6121 0 : }
6122 :
6123 : /// Configure in-memory and persistent state of a node as requested
6124 : ///
6125 : /// Note that this function does not trigger any immediate side effects in response
6126 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
6127 0 : async fn node_state_configure(
6128 0 : &self,
6129 0 : node_id: NodeId,
6130 0 : availability: Option<NodeAvailability>,
6131 0 : scheduling: Option<NodeSchedulingPolicy>,
6132 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
6133 0 : ) -> Result<AvailabilityTransition, ApiError> {
6134 0 : if let Some(scheduling) = scheduling {
6135 : // Scheduling is a persistent part of Node: we must write updates to the database before
6136 : // applying them in memory
6137 0 : self.persistence
6138 0 : .update_node_scheduling_policy(node_id, scheduling)
6139 0 : .await?;
6140 0 : }
6141 :
6142 : // If we're activating a node, then before setting it active we must reconcile any shard locations
6143 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
6144 : // by calling [`Self::node_activate_reconcile`]
6145 : //
6146 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
6147 : // nothing else can mutate its availability while we run.
6148 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
6149 0 : let (activate_node, availability_transition) = {
6150 0 : let locked = self.inner.read().unwrap();
6151 0 : let Some(node) = locked.nodes.get(&node_id) else {
6152 0 : return Err(ApiError::NotFound(
6153 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6154 0 : ));
6155 : };
6156 :
6157 0 : (
6158 0 : node.clone(),
6159 0 : node.get_availability_transition(input_availability),
6160 0 : )
6161 : };
6162 :
6163 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
6164 0 : self.node_activate_reconcile(activate_node, node_lock)
6165 0 : .await?;
6166 0 : }
6167 0 : availability_transition
6168 : } else {
6169 0 : AvailabilityTransition::Unchanged
6170 : };
6171 :
6172 : // Apply changes from the request to our in-memory state for the Node
6173 0 : let mut locked = self.inner.write().unwrap();
6174 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
6175 0 :
6176 0 : let mut new_nodes = (**nodes).clone();
6177 :
6178 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
6179 0 : return Err(ApiError::NotFound(
6180 0 : anyhow::anyhow!("Node not registered").into(),
6181 0 : ));
6182 : };
6183 :
6184 0 : if let Some(availability) = availability {
6185 0 : node.set_availability(availability);
6186 0 : }
6187 :
6188 0 : if let Some(scheduling) = scheduling {
6189 0 : node.set_scheduling(scheduling);
6190 0 : }
6191 :
6192 : // Update the scheduler, in case the elegibility of the node for new shards has changed
6193 0 : scheduler.node_upsert(node);
6194 0 :
6195 0 : let new_nodes = Arc::new(new_nodes);
6196 0 : locked.nodes = new_nodes;
6197 0 :
6198 0 : Ok(availability_transition)
6199 0 : }
6200 :
6201 : /// Handle availability transition of one node
6202 : ///
6203 : /// Note that you should first call [`Self::node_state_configure`] to update
6204 : /// the in-memory state referencing that node. If you need to handle more than one transition
6205 : /// consider using [`Self::handle_node_availability_transitions`].
6206 0 : async fn handle_node_availability_transition(
6207 0 : &self,
6208 0 : node_id: NodeId,
6209 0 : transition: AvailabilityTransition,
6210 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
6211 0 : ) -> Result<(), ApiError> {
6212 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
6213 0 : match transition {
6214 : AvailabilityTransition::ToOffline => {
6215 0 : tracing::info!("Node {} transition to offline", node_id);
6216 :
6217 0 : let mut locked = self.inner.write().unwrap();
6218 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6219 0 :
6220 0 : let mut tenants_affected: usize = 0;
6221 :
6222 0 : for (_tenant_id, mut schedule_context, shards) in
6223 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6224 : {
6225 0 : for tenant_shard in shards {
6226 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
6227 0 : if let Some(observed_loc) =
6228 0 : tenant_shard.observed.locations.get_mut(&node_id)
6229 0 : {
6230 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
6231 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
6232 0 : observed_loc.conf = None;
6233 0 : }
6234 :
6235 0 : if nodes.len() == 1 {
6236 : // Special case for single-node cluster: there is no point trying to reschedule
6237 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
6238 : // failures to schedule them.
6239 0 : continue;
6240 0 : }
6241 0 :
6242 0 : if !nodes
6243 0 : .values()
6244 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6245 : {
6246 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
6247 : // trying to reschedule since there's nowhere else to go. Without this
6248 : // branch we incorrectly detach tenants in response to node unavailability.
6249 0 : continue;
6250 0 : }
6251 0 :
6252 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
6253 0 : tenant_shard.sequence = tenant_shard.sequence.next();
6254 0 :
6255 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
6256 0 : Err(e) => {
6257 0 : // It is possible that some tenants will become unschedulable when too many pageservers
6258 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
6259 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
6260 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
6261 : }
6262 : Ok(()) => {
6263 0 : if self
6264 0 : .maybe_reconcile_shard(
6265 0 : tenant_shard,
6266 0 : nodes,
6267 0 : ReconcilerPriority::Normal,
6268 0 : )
6269 0 : .is_some()
6270 0 : {
6271 0 : tenants_affected += 1;
6272 0 : };
6273 : }
6274 : }
6275 0 : }
6276 : }
6277 : }
6278 0 : tracing::info!(
6279 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
6280 : tenants_affected,
6281 : node_id
6282 : )
6283 : }
6284 : AvailabilityTransition::ToActive => {
6285 0 : tracing::info!("Node {} transition to active", node_id);
6286 :
6287 0 : let mut locked = self.inner.write().unwrap();
6288 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
6289 :
6290 : // When a node comes back online, we must reconcile any tenant that has a None observed
6291 : // location on the node.
6292 0 : for tenant_shard in tenants.values_mut() {
6293 : // If a reconciliation is already in progress, rely on the previous scheduling
6294 : // decision and skip triggering a new reconciliation.
6295 0 : if tenant_shard.reconciler.is_some() {
6296 0 : continue;
6297 0 : }
6298 :
6299 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
6300 0 : if observed_loc.conf.is_none() {
6301 0 : self.maybe_reconcile_shard(
6302 0 : tenant_shard,
6303 0 : nodes,
6304 0 : ReconcilerPriority::Normal,
6305 0 : );
6306 0 : }
6307 0 : }
6308 : }
6309 :
6310 : // TODO: in the background, we should balance work back onto this pageserver
6311 : }
6312 : // No action required for the intermediate unavailable state.
6313 : // When we transition into active or offline from the unavailable state,
6314 : // the correct handling above will kick in.
6315 : AvailabilityTransition::ToWarmingUpFromActive => {
6316 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
6317 : }
6318 : AvailabilityTransition::ToWarmingUpFromOffline => {
6319 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
6320 : }
6321 : AvailabilityTransition::Unchanged => {
6322 0 : tracing::debug!("Node {} no availability change during config", node_id);
6323 : }
6324 : }
6325 :
6326 0 : Ok(())
6327 0 : }
6328 :
6329 : /// Handle availability transition for multiple nodes
6330 : ///
6331 : /// Note that you should first call [`Self::node_state_configure`] for
6332 : /// all nodes being handled here for the handling to use fresh in-memory state.
6333 0 : async fn handle_node_availability_transitions(
6334 0 : &self,
6335 0 : transitions: Vec<(
6336 0 : NodeId,
6337 0 : TracingExclusiveGuard<NodeOperations>,
6338 0 : AvailabilityTransition,
6339 0 : )>,
6340 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
6341 0 : let mut errors = Vec::default();
6342 0 : for (node_id, node_lock, transition) in transitions {
6343 0 : let res = self
6344 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
6345 0 : .await;
6346 0 : if let Err(err) = res {
6347 0 : errors.push((node_id, err));
6348 0 : }
6349 : }
6350 :
6351 0 : if errors.is_empty() {
6352 0 : Ok(())
6353 : } else {
6354 0 : Err(errors)
6355 : }
6356 0 : }
6357 :
6358 0 : pub(crate) async fn node_configure(
6359 0 : &self,
6360 0 : node_id: NodeId,
6361 0 : availability: Option<NodeAvailability>,
6362 0 : scheduling: Option<NodeSchedulingPolicy>,
6363 0 : ) -> Result<(), ApiError> {
6364 0 : let node_lock =
6365 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
6366 :
6367 0 : let transition = self
6368 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
6369 0 : .await?;
6370 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
6371 0 : .await
6372 0 : }
6373 :
6374 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
6375 : /// operation for HTTP api.
6376 0 : pub(crate) async fn external_node_configure(
6377 0 : &self,
6378 0 : node_id: NodeId,
6379 0 : availability: Option<NodeAvailability>,
6380 0 : scheduling: Option<NodeSchedulingPolicy>,
6381 0 : ) -> Result<(), ApiError> {
6382 0 : {
6383 0 : let locked = self.inner.read().unwrap();
6384 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
6385 0 : return Err(ApiError::PreconditionFailed(
6386 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
6387 0 : ));
6388 0 : }
6389 0 : }
6390 0 :
6391 0 : self.node_configure(node_id, availability, scheduling).await
6392 0 : }
6393 :
6394 0 : pub(crate) async fn start_node_drain(
6395 0 : self: &Arc<Self>,
6396 0 : node_id: NodeId,
6397 0 : ) -> Result<(), ApiError> {
6398 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
6399 0 : let locked = self.inner.read().unwrap();
6400 0 : let nodes = &locked.nodes;
6401 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6402 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6403 0 : ))?;
6404 0 : let schedulable_nodes_count = nodes
6405 0 : .iter()
6406 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6407 0 : .count();
6408 0 :
6409 0 : (
6410 0 : locked
6411 0 : .ongoing_operation
6412 0 : .as_ref()
6413 0 : .map(|ongoing| ongoing.operation),
6414 0 : node.is_available(),
6415 0 : node.get_scheduling(),
6416 0 : schedulable_nodes_count,
6417 0 : )
6418 0 : };
6419 :
6420 0 : if let Some(ongoing) = ongoing_op {
6421 0 : return Err(ApiError::PreconditionFailed(
6422 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6423 0 : ));
6424 0 : }
6425 0 :
6426 0 : if !node_available {
6427 0 : return Err(ApiError::ResourceUnavailable(
6428 0 : format!("Node {node_id} is currently unavailable").into(),
6429 0 : ));
6430 0 : }
6431 0 :
6432 0 : if schedulable_nodes_count == 0 {
6433 0 : return Err(ApiError::PreconditionFailed(
6434 0 : "No other schedulable nodes to drain to".into(),
6435 0 : ));
6436 0 : }
6437 0 :
6438 0 : match node_policy {
6439 : NodeSchedulingPolicy::Active => {
6440 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
6441 0 : .await?;
6442 :
6443 0 : let cancel = self.cancel.child_token();
6444 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6445 :
6446 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6447 0 : operation: Operation::Drain(Drain { node_id }),
6448 0 : cancel: cancel.clone(),
6449 0 : });
6450 :
6451 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
6452 :
6453 0 : tokio::task::spawn({
6454 0 : let service = self.clone();
6455 0 : let cancel = cancel.clone();
6456 0 : async move {
6457 0 : let _gate_guard = gate_guard;
6458 0 :
6459 0 : scopeguard::defer! {
6460 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6461 0 :
6462 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
6463 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
6464 0 : } else {
6465 0 : panic!("We always remove the same operation")
6466 0 : }
6467 0 : }
6468 0 :
6469 0 : tracing::info!("Drain background operation starting");
6470 0 : let res = service.drain_node(node_id, cancel).await;
6471 0 : match res {
6472 : Ok(()) => {
6473 0 : tracing::info!("Drain background operation completed successfully");
6474 : }
6475 : Err(OperationError::Cancelled) => {
6476 0 : tracing::info!("Drain background operation was cancelled");
6477 : }
6478 0 : Err(err) => {
6479 0 : tracing::error!("Drain background operation encountered: {err}")
6480 : }
6481 : }
6482 0 : }
6483 0 : }.instrument(span));
6484 0 : }
6485 : NodeSchedulingPolicy::Draining => {
6486 0 : return Err(ApiError::Conflict(format!(
6487 0 : "Node {node_id} has drain in progress"
6488 0 : )));
6489 : }
6490 0 : policy => {
6491 0 : return Err(ApiError::PreconditionFailed(
6492 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
6493 0 : ));
6494 : }
6495 : }
6496 :
6497 0 : Ok(())
6498 0 : }
6499 :
6500 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
6501 0 : let node_available = {
6502 0 : let locked = self.inner.read().unwrap();
6503 0 : let nodes = &locked.nodes;
6504 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6505 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6506 0 : ))?;
6507 :
6508 0 : node.is_available()
6509 0 : };
6510 0 :
6511 0 : if !node_available {
6512 0 : return Err(ApiError::ResourceUnavailable(
6513 0 : format!("Node {node_id} is currently unavailable").into(),
6514 0 : ));
6515 0 : }
6516 :
6517 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6518 0 : if let Operation::Drain(drain) = op_handler.operation {
6519 0 : if drain.node_id == node_id {
6520 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6521 0 : op_handler.cancel.cancel();
6522 0 : return Ok(());
6523 0 : }
6524 0 : }
6525 0 : }
6526 :
6527 0 : Err(ApiError::PreconditionFailed(
6528 0 : format!("Node {node_id} has no drain in progress").into(),
6529 0 : ))
6530 0 : }
6531 :
6532 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
6533 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
6534 0 : let locked = self.inner.read().unwrap();
6535 0 : let nodes = &locked.nodes;
6536 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6537 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6538 0 : ))?;
6539 :
6540 0 : (
6541 0 : locked
6542 0 : .ongoing_operation
6543 0 : .as_ref()
6544 0 : .map(|ongoing| ongoing.operation),
6545 0 : node.is_available(),
6546 0 : node.get_scheduling(),
6547 0 : nodes.len(),
6548 0 : )
6549 0 : };
6550 :
6551 0 : if let Some(ongoing) = ongoing_op {
6552 0 : return Err(ApiError::PreconditionFailed(
6553 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6554 0 : ));
6555 0 : }
6556 0 :
6557 0 : if !node_available {
6558 0 : return Err(ApiError::ResourceUnavailable(
6559 0 : format!("Node {node_id} is currently unavailable").into(),
6560 0 : ));
6561 0 : }
6562 0 :
6563 0 : if total_nodes_count <= 1 {
6564 0 : return Err(ApiError::PreconditionFailed(
6565 0 : "No other nodes to fill from".into(),
6566 0 : ));
6567 0 : }
6568 0 :
6569 0 : match node_policy {
6570 : NodeSchedulingPolicy::Active => {
6571 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
6572 0 : .await?;
6573 :
6574 0 : let cancel = self.cancel.child_token();
6575 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6576 :
6577 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6578 0 : operation: Operation::Fill(Fill { node_id }),
6579 0 : cancel: cancel.clone(),
6580 0 : });
6581 :
6582 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
6583 :
6584 0 : tokio::task::spawn({
6585 0 : let service = self.clone();
6586 0 : let cancel = cancel.clone();
6587 0 : async move {
6588 0 : let _gate_guard = gate_guard;
6589 0 :
6590 0 : scopeguard::defer! {
6591 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6592 0 :
6593 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
6594 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
6595 0 : } else {
6596 0 : panic!("We always remove the same operation")
6597 0 : }
6598 0 : }
6599 0 :
6600 0 : tracing::info!("Fill background operation starting");
6601 0 : let res = service.fill_node(node_id, cancel).await;
6602 0 : match res {
6603 : Ok(()) => {
6604 0 : tracing::info!("Fill background operation completed successfully");
6605 : }
6606 : Err(OperationError::Cancelled) => {
6607 0 : tracing::info!("Fill background operation was cancelled");
6608 : }
6609 0 : Err(err) => {
6610 0 : tracing::error!("Fill background operation encountered: {err}")
6611 : }
6612 : }
6613 0 : }
6614 0 : }.instrument(span));
6615 0 : }
6616 : NodeSchedulingPolicy::Filling => {
6617 0 : return Err(ApiError::Conflict(format!(
6618 0 : "Node {node_id} has fill in progress"
6619 0 : )));
6620 : }
6621 0 : policy => {
6622 0 : return Err(ApiError::PreconditionFailed(
6623 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
6624 0 : ));
6625 : }
6626 : }
6627 :
6628 0 : Ok(())
6629 0 : }
6630 :
6631 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
6632 0 : let node_available = {
6633 0 : let locked = self.inner.read().unwrap();
6634 0 : let nodes = &locked.nodes;
6635 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6636 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6637 0 : ))?;
6638 :
6639 0 : node.is_available()
6640 0 : };
6641 0 :
6642 0 : if !node_available {
6643 0 : return Err(ApiError::ResourceUnavailable(
6644 0 : format!("Node {node_id} is currently unavailable").into(),
6645 0 : ));
6646 0 : }
6647 :
6648 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
6649 0 : if let Operation::Fill(fill) = op_handler.operation {
6650 0 : if fill.node_id == node_id {
6651 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
6652 0 : op_handler.cancel.cancel();
6653 0 : return Ok(());
6654 0 : }
6655 0 : }
6656 0 : }
6657 :
6658 0 : Err(ApiError::PreconditionFailed(
6659 0 : format!("Node {node_id} has no fill in progress").into(),
6660 0 : ))
6661 0 : }
6662 :
6663 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
6664 : /// configuration
6665 0 : fn maybe_reconcile_shard(
6666 0 : &self,
6667 0 : shard: &mut TenantShard,
6668 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6669 0 : priority: ReconcilerPriority,
6670 0 : ) -> Option<ReconcilerWaiter> {
6671 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
6672 0 : }
6673 :
6674 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
6675 0 : fn get_reconciler_units(
6676 0 : &self,
6677 0 : priority: ReconcilerPriority,
6678 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
6679 0 : let units = match priority {
6680 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
6681 : ReconcilerPriority::High => {
6682 0 : match self
6683 0 : .priority_reconciler_concurrency
6684 0 : .clone()
6685 0 : .try_acquire_owned()
6686 : {
6687 0 : Ok(u) => Ok(u),
6688 : Err(TryAcquireError::NoPermits) => {
6689 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
6690 : // the normal priority semaphore.
6691 0 : self.reconciler_concurrency.clone().try_acquire_owned()
6692 : }
6693 0 : Err(e) => Err(e),
6694 : }
6695 : }
6696 : };
6697 :
6698 0 : units.map(ReconcileUnits::new)
6699 0 : }
6700 :
6701 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
6702 0 : fn maybe_configured_reconcile_shard(
6703 0 : &self,
6704 0 : shard: &mut TenantShard,
6705 0 : nodes: &Arc<HashMap<NodeId, Node>>,
6706 0 : reconciler_config: ReconcilerConfig,
6707 0 : ) -> Option<ReconcilerWaiter> {
6708 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
6709 :
6710 0 : let reconcile_reason = match reconcile_needed {
6711 0 : ReconcileNeeded::No => return None,
6712 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
6713 0 : ReconcileNeeded::Yes(reason) => {
6714 0 : // Fall through to try and acquire units for spawning reconciler
6715 0 : reason
6716 : }
6717 : };
6718 :
6719 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
6720 0 : Ok(u) => u,
6721 : Err(_) => {
6722 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
6723 0 : "Concurrency limited: enqueued for reconcile later");
6724 0 : if !shard.delayed_reconcile {
6725 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
6726 0 : Err(TrySendError::Closed(_)) => {
6727 0 : // Weird mid-shutdown case?
6728 0 : }
6729 : Err(TrySendError::Full(_)) => {
6730 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
6731 0 : tracing::warn!(
6732 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
6733 : );
6734 : }
6735 0 : Ok(()) => {
6736 0 : shard.delayed_reconcile = true;
6737 0 : }
6738 : }
6739 0 : }
6740 :
6741 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
6742 : // number to advance. When this function is eventually called again and succeeds in getting units,
6743 : // it will spawn a reconciler that makes this waiter complete.
6744 0 : return Some(shard.future_reconcile_waiter());
6745 : }
6746 : };
6747 :
6748 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
6749 : // Gate closed: we're shutting down, drop out.
6750 0 : return None;
6751 : };
6752 :
6753 0 : shard.spawn_reconciler(
6754 0 : reconcile_reason,
6755 0 : &self.result_tx,
6756 0 : nodes,
6757 0 : &self.compute_hook,
6758 0 : reconciler_config,
6759 0 : &self.config,
6760 0 : &self.persistence,
6761 0 : units,
6762 0 : gate_guard,
6763 0 : &self.reconcilers_cancel,
6764 0 : )
6765 0 : }
6766 :
6767 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
6768 : /// Additionally, reschedule tenants that require it.
6769 : ///
6770 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
6771 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
6772 : /// available. A return value of 0 indicates that everything is fully reconciled already.
6773 0 : fn reconcile_all(&self) -> usize {
6774 0 : let mut locked = self.inner.write().unwrap();
6775 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6776 0 : let pageservers = nodes.clone();
6777 0 :
6778 0 : // This function is an efficient place to update lazy statistics, since we are walking
6779 0 : // all tenants.
6780 0 : let mut pending_reconciles = 0;
6781 0 : let mut az_violations = 0;
6782 0 :
6783 0 : // If we find any tenants to drop from memory, stash them to offload after
6784 0 : // we're done traversing the map of tenants.
6785 0 : let mut drop_detached_tenants = Vec::new();
6786 0 :
6787 0 : let mut reconciles_spawned = 0;
6788 0 : for shard in tenants.values_mut() {
6789 : // Accumulate scheduling statistics
6790 0 : if let (Some(attached), Some(preferred)) =
6791 0 : (shard.intent.get_attached(), shard.preferred_az())
6792 : {
6793 0 : let node_az = nodes
6794 0 : .get(attached)
6795 0 : .expect("Nodes exist if referenced")
6796 0 : .get_availability_zone_id();
6797 0 : if node_az != preferred {
6798 0 : az_violations += 1;
6799 0 : }
6800 0 : }
6801 :
6802 : // Skip checking if this shard is already enqueued for reconciliation
6803 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
6804 : // If there is something delayed, then return a nonzero count so that
6805 : // callers like reconcile_all_now do not incorrectly get the impression
6806 : // that the system is in a quiescent state.
6807 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
6808 0 : pending_reconciles += 1;
6809 0 : continue;
6810 0 : }
6811 0 :
6812 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
6813 0 : // dirty, spawn another rone
6814 0 : if self
6815 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
6816 0 : .is_some()
6817 0 : {
6818 0 : reconciles_spawned += 1;
6819 0 : } else if shard.delayed_reconcile {
6820 0 : // Shard wanted to reconcile but for some reason couldn't.
6821 0 : pending_reconciles += 1;
6822 0 : }
6823 :
6824 : // If this tenant is detached, try dropping it from memory. This is usually done
6825 : // proactively in [`Self::process_results`], but we do it here to handle the edge
6826 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
6827 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
6828 0 : && shard.policy == PlacementPolicy::Detached
6829 : {
6830 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
6831 0 : shard.tenant_shard_id.tenant_id,
6832 0 : TenantOperations::DropDetached,
6833 0 : ) {
6834 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
6835 0 : }
6836 0 : }
6837 : }
6838 :
6839 : // Some metrics are calculated from SchedulerNode state, update these periodically
6840 0 : scheduler.update_metrics();
6841 :
6842 : // Process any deferred tenant drops
6843 0 : for (tenant_id, guard) in drop_detached_tenants {
6844 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
6845 0 : }
6846 :
6847 0 : metrics::METRICS_REGISTRY
6848 0 : .metrics_group
6849 0 : .storage_controller_schedule_az_violation
6850 0 : .set(az_violations as i64);
6851 0 :
6852 0 : metrics::METRICS_REGISTRY
6853 0 : .metrics_group
6854 0 : .storage_controller_pending_reconciles
6855 0 : .set(pending_reconciles as i64);
6856 0 :
6857 0 : reconciles_spawned
6858 0 : }
6859 :
6860 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
6861 : /// could be scheduled somewhere better:
6862 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
6863 : /// * e.g. after a node fails then recovers, to move some work back to it
6864 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
6865 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
6866 : /// we did the split, but are probably better placed elsewhere.
6867 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
6868 : /// * e.g. after a shard split, some locations will be on the same node (where the split
6869 : /// happened), and will probably be better placed elsewhere.
6870 : ///
6871 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
6872 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
6873 : /// according to those same soft constraints.
6874 0 : async fn optimize_all(&self) -> usize {
6875 : // Limit on how many shards' optmizations each call to this function will execute. Combined
6876 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
6877 : // trickle of optimizations in the background, rather than executing a large number in parallel
6878 : // when a change occurs.
6879 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
6880 :
6881 : // Synchronous prepare: scan shards for possible scheduling optimizations
6882 0 : let candidate_work = self.optimize_all_plan();
6883 0 : let candidate_work_len = candidate_work.len();
6884 :
6885 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
6886 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
6887 :
6888 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
6889 0 :
6890 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
6891 0 : let mut reconciles_spawned = 0;
6892 0 : let mut optimizations_applied = 0;
6893 0 : let mut locked = self.inner.write().unwrap();
6894 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6895 0 : for (tenant_shard_id, optimization) in validated_work {
6896 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6897 : // Shard was dropped between planning and execution;
6898 0 : continue;
6899 : };
6900 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
6901 0 : if shard.apply_optimization(scheduler, optimization) {
6902 0 : optimizations_applied += 1;
6903 0 : if self
6904 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
6905 0 : .is_some()
6906 0 : {
6907 0 : reconciles_spawned += 1;
6908 0 : }
6909 0 : }
6910 :
6911 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
6912 0 : break;
6913 0 : }
6914 : }
6915 :
6916 0 : if was_work_filtered {
6917 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
6918 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
6919 0 : // as these validations start passing.
6920 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
6921 0 : }
6922 :
6923 0 : reconciles_spawned
6924 0 : }
6925 :
6926 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
6927 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
6928 : // this higher than the execution limit gives us a chance to execute some work even if the first
6929 : // few optimizations we find are not ready.
6930 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
6931 :
6932 0 : let mut work = Vec::new();
6933 0 : let mut locked = self.inner.write().unwrap();
6934 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6935 :
6936 : // We are going to plan a bunch of optimisations before applying any of them, so the
6937 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
6938 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
6939 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
6940 : // to ignore the utilisation component of the score.
6941 :
6942 0 : for (_tenant_id, schedule_context, shards) in
6943 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
6944 : {
6945 0 : for shard in shards {
6946 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
6947 0 : break;
6948 0 : }
6949 0 : match shard.get_scheduling_policy() {
6950 0 : ShardSchedulingPolicy::Active => {
6951 0 : // Ok to do optimization
6952 0 : }
6953 : ShardSchedulingPolicy::Essential
6954 : | ShardSchedulingPolicy::Pause
6955 : | ShardSchedulingPolicy::Stop => {
6956 : // Policy prevents optimizing this shard.
6957 0 : continue;
6958 : }
6959 : }
6960 :
6961 0 : if !matches!(shard.splitting, SplitState::Idle)
6962 0 : || matches!(shard.policy, PlacementPolicy::Detached)
6963 0 : || shard.reconciler.is_some()
6964 : {
6965 : // Do not start any optimizations while another change to the tenant is ongoing: this
6966 : // is not necessary for correctness, but simplifies operations and implicitly throttles
6967 : // optimization changes to happen in a "trickle" over time.
6968 0 : continue;
6969 0 : }
6970 0 :
6971 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
6972 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
6973 0 : if cfg!(feature = "testing") {
6974 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
6975 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
6976 : // panic in prod if we hit this, or spend cycles on it in prod.
6977 0 : assert!(
6978 0 : shard
6979 0 : .optimize_attachment(scheduler, &schedule_context)
6980 0 : .is_none()
6981 0 : );
6982 0 : assert!(
6983 0 : shard
6984 0 : .optimize_secondary(scheduler, &schedule_context)
6985 0 : .is_none()
6986 0 : );
6987 0 : }
6988 0 : continue;
6989 0 : }
6990 :
6991 0 : if let Some(optimization) =
6992 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
6993 : // its primary location based on soft constraints, cut it over.
6994 0 : shard.optimize_attachment(scheduler, &schedule_context)
6995 : {
6996 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
6997 0 : work.push((shard.tenant_shard_id, optimization));
6998 0 : break;
6999 0 : } else if let Some(optimization) =
7000 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
7001 : // better placed on another node, based on ScheduleContext, then adjust it. This
7002 : // covers cases like after a shard split, where we might have too many shards
7003 : // in the same tenant with secondary locations on the node where they originally split.
7004 0 : shard.optimize_secondary(scheduler, &schedule_context)
7005 : {
7006 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
7007 0 : work.push((shard.tenant_shard_id, optimization));
7008 0 : break;
7009 0 : }
7010 : }
7011 : }
7012 :
7013 0 : work
7014 0 : }
7015 :
7016 0 : async fn optimize_all_validate(
7017 0 : &self,
7018 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
7019 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
7020 0 : // Take a clone of the node map to use outside the lock in async validation phase
7021 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
7022 0 :
7023 0 : let mut want_secondary_status = Vec::new();
7024 0 :
7025 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
7026 0 : // check that the state of locations is acceptable to run the optimization, such as
7027 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
7028 0 : // in a live migration.
7029 0 : let mut validated_work = Vec::new();
7030 0 : for (tenant_shard_id, optimization) in candidate_work {
7031 0 : match optimization.action {
7032 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
7033 : old_attached_node_id: _,
7034 0 : new_attached_node_id,
7035 0 : }) => {
7036 0 : match validation_nodes.get(&new_attached_node_id) {
7037 0 : None => {
7038 0 : // Node was dropped between planning and validation
7039 0 : }
7040 0 : Some(node) => {
7041 0 : if !node.is_available() {
7042 0 : tracing::info!(
7043 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
7044 : );
7045 0 : } else {
7046 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
7047 0 : // remote API requests concurrently.
7048 0 : want_secondary_status.push((
7049 0 : tenant_shard_id,
7050 0 : node.clone(),
7051 0 : optimization,
7052 0 : ));
7053 0 : }
7054 : }
7055 : }
7056 : }
7057 : ScheduleOptimizationAction::ReplaceSecondary(_)
7058 : | ScheduleOptimizationAction::CreateSecondary(_)
7059 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
7060 : // No extra checks needed to manage secondaries: this does not interrupt client access
7061 0 : validated_work.push((tenant_shard_id, optimization))
7062 : }
7063 : };
7064 : }
7065 :
7066 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
7067 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
7068 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
7069 0 : let results = self
7070 0 : .tenant_for_shards_api(
7071 0 : want_secondary_status
7072 0 : .iter()
7073 0 : .map(|i| (i.0, i.1.clone()))
7074 0 : .collect(),
7075 0 : |tenant_shard_id, client| async move {
7076 0 : client.tenant_secondary_status(tenant_shard_id).await
7077 0 : },
7078 0 : 1,
7079 0 : 1,
7080 0 : SHORT_RECONCILE_TIMEOUT,
7081 0 : &self.cancel,
7082 0 : )
7083 0 : .await;
7084 :
7085 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
7086 0 : want_secondary_status.into_iter().zip(results.into_iter())
7087 : {
7088 0 : match secondary_status {
7089 0 : Err(e) => {
7090 0 : tracing::info!(
7091 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
7092 : );
7093 : }
7094 0 : Ok(progress) => {
7095 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
7096 : // them in an optimization
7097 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
7098 :
7099 0 : if progress.heatmap_mtime.is_none()
7100 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
7101 0 : && progress.bytes_downloaded != progress.bytes_total
7102 0 : || progress.bytes_total - progress.bytes_downloaded
7103 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
7104 : {
7105 0 : tracing::info!(
7106 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
7107 : );
7108 :
7109 : #[cfg(feature = "testing")]
7110 0 : if progress.heatmap_mtime.is_none() {
7111 : // No heatmap might mean the attached location has never uploaded one, or that
7112 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
7113 : // but fairly common in tests.
7114 0 : self.kick_secondary_download(tenant_shard_id).await;
7115 0 : }
7116 : } else {
7117 : // Location looks ready: proceed
7118 0 : tracing::info!(
7119 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
7120 : );
7121 0 : validated_work.push((tenant_shard_id, optimization))
7122 : }
7123 : }
7124 : }
7125 : }
7126 :
7127 0 : validated_work
7128 0 : }
7129 :
7130 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
7131 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
7132 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
7133 : /// we have this helper to move things along faster.
7134 : #[cfg(feature = "testing")]
7135 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
7136 0 : let (attached_node, secondaries) = {
7137 0 : let locked = self.inner.read().unwrap();
7138 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
7139 0 : tracing::warn!(
7140 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
7141 : );
7142 0 : return;
7143 : };
7144 :
7145 0 : let Some(attached) = shard.intent.get_attached() else {
7146 0 : tracing::warn!(
7147 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
7148 : );
7149 0 : return;
7150 : };
7151 :
7152 0 : let secondaries = shard
7153 0 : .intent
7154 0 : .get_secondary()
7155 0 : .iter()
7156 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
7157 0 : .collect::<Vec<_>>();
7158 0 :
7159 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
7160 0 : };
7161 0 :
7162 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
7163 0 : // a 'kick' to let scheduling optimisation run more promptly.
7164 0 : match attached_node
7165 0 : .with_client_retries(
7166 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
7167 0 : &self.config.pageserver_jwt_token,
7168 0 : 3,
7169 0 : 10,
7170 0 : SHORT_RECONCILE_TIMEOUT,
7171 0 : &self.cancel,
7172 0 : )
7173 0 : .await
7174 : {
7175 0 : Some(Err(e)) => {
7176 0 : tracing::info!(
7177 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
7178 : );
7179 : }
7180 : None => {
7181 0 : tracing::info!(
7182 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
7183 : );
7184 : }
7185 : Some(Ok(_)) => {
7186 0 : tracing::info!(
7187 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
7188 : );
7189 : }
7190 : }
7191 :
7192 0 : for secondary_node in secondaries {
7193 0 : match secondary_node
7194 0 : .with_client_retries(
7195 0 : |client| async move {
7196 0 : client
7197 0 : .tenant_secondary_download(
7198 0 : tenant_shard_id,
7199 0 : Some(Duration::from_secs(1)),
7200 0 : )
7201 0 : .await
7202 0 : },
7203 0 : &self.config.pageserver_jwt_token,
7204 0 : 3,
7205 0 : 10,
7206 0 : SHORT_RECONCILE_TIMEOUT,
7207 0 : &self.cancel,
7208 0 : )
7209 0 : .await
7210 : {
7211 0 : Some(Err(e)) => {
7212 0 : tracing::info!(
7213 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
7214 : );
7215 : }
7216 : None => {
7217 0 : tracing::info!(
7218 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
7219 : );
7220 : }
7221 0 : Some(Ok(progress)) => {
7222 0 : tracing::info!(
7223 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
7224 : );
7225 : }
7226 : }
7227 : }
7228 0 : }
7229 :
7230 : /// Look for shards which are oversized and in need of splitting
7231 0 : async fn autosplit_tenants(self: &Arc<Self>) {
7232 0 : let Some(split_threshold) = self.config.split_threshold else {
7233 : // Auto-splitting is disabled
7234 0 : return;
7235 : };
7236 :
7237 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7238 :
7239 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
7240 :
7241 0 : let mut top_n = Vec::new();
7242 0 :
7243 0 : // Call into each node to look for big tenants
7244 0 : let top_n_request = TopTenantShardsRequest {
7245 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
7246 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
7247 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
7248 0 : // disk space impact of one shard.
7249 0 : order_by: models::TenantSorting::MaxLogicalSize,
7250 0 : limit: 10,
7251 0 : where_shards_lt: Some(SPLIT_TO_MAX),
7252 0 : where_gt: Some(split_threshold),
7253 0 : };
7254 0 : for node in nodes.values() {
7255 0 : let request_ref = &top_n_request;
7256 0 : match node
7257 0 : .with_client_retries(
7258 0 : |client| async move {
7259 0 : let request = request_ref.clone();
7260 0 : client.top_tenant_shards(request.clone()).await
7261 0 : },
7262 0 : &self.config.pageserver_jwt_token,
7263 0 : 3,
7264 0 : 3,
7265 0 : Duration::from_secs(5),
7266 0 : &self.cancel,
7267 0 : )
7268 0 : .await
7269 : {
7270 0 : Some(Ok(node_top_n)) => {
7271 0 : top_n.extend(node_top_n.shards.into_iter());
7272 0 : }
7273 : Some(Err(mgmt_api::Error::Cancelled)) => {
7274 0 : continue;
7275 : }
7276 0 : Some(Err(e)) => {
7277 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
7278 0 : continue;
7279 : }
7280 : None => {
7281 : // Node is shutting down
7282 0 : continue;
7283 : }
7284 : };
7285 : }
7286 :
7287 : // Pick the biggest tenant to split first
7288 0 : top_n.sort_by_key(|i| i.resident_size);
7289 0 :
7290 0 : // Filter out tenants in a prohibiting scheduling mode
7291 0 : {
7292 0 : let locked = self.inner.read().unwrap();
7293 0 : top_n.retain(|i| {
7294 0 : if let Some(shard) = locked.tenants.get(&i.id) {
7295 0 : matches!(shard.get_scheduling_policy(), ShardSchedulingPolicy::Active)
7296 : } else {
7297 0 : false
7298 : }
7299 0 : });
7300 0 : }
7301 :
7302 0 : let Some(split_candidate) = top_n.into_iter().next() else {
7303 0 : tracing::debug!("No split-elegible shards found");
7304 0 : return;
7305 : };
7306 :
7307 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
7308 : // want to block the background reconcile loop on this.
7309 0 : tracing::info!(
7310 0 : "Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}"
7311 : );
7312 :
7313 0 : let this = self.clone();
7314 0 : tokio::spawn(
7315 0 : async move {
7316 0 : match this
7317 0 : .tenant_shard_split(
7318 0 : split_candidate.id.tenant_id,
7319 0 : TenantShardSplitRequest {
7320 0 : // Always split to the max number of shards: this avoids stepping through
7321 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
7322 0 : // each time as a tenant grows, and is not too expensive because our max shard
7323 0 : // count is relatively low anyway.
7324 0 : // This policy will be adjusted in future once we support higher shard count.
7325 0 : new_shard_count: SPLIT_TO_MAX.literal(),
7326 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
7327 0 : },
7328 0 : )
7329 0 : .await
7330 : {
7331 : Ok(_) => {
7332 0 : tracing::info!("Successful auto-split");
7333 : }
7334 0 : Err(e) => {
7335 0 : tracing::error!("Auto-split failed: {e}");
7336 : }
7337 : }
7338 0 : }
7339 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
7340 : );
7341 0 : }
7342 :
7343 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
7344 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
7345 : /// put the system into a quiescent state where future background reconciliations won't do anything.
7346 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
7347 0 : let reconciles_spawned = self.reconcile_all();
7348 0 : let reconciles_spawned = if reconciles_spawned == 0 {
7349 : // Only optimize when we are otherwise idle
7350 0 : self.optimize_all().await
7351 : } else {
7352 0 : reconciles_spawned
7353 : };
7354 :
7355 0 : let waiters = {
7356 0 : let mut waiters = Vec::new();
7357 0 : let locked = self.inner.read().unwrap();
7358 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
7359 0 : if let Some(waiter) = shard.get_waiter() {
7360 0 : waiters.push(waiter);
7361 0 : }
7362 : }
7363 0 : waiters
7364 0 : };
7365 0 :
7366 0 : let waiter_count = waiters.len();
7367 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
7368 0 : Ok(()) => {}
7369 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
7370 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
7371 0 : {
7372 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
7373 0 : // because some other change superceded it. We will return a nonzero number,
7374 0 : // so the caller knows they might have to call again to quiesce the system.
7375 0 : }
7376 0 : Err(e) => {
7377 0 : return Err(e);
7378 : }
7379 : };
7380 :
7381 0 : tracing::info!(
7382 0 : "{} reconciles in reconcile_all, {} waiters",
7383 : reconciles_spawned,
7384 : waiter_count
7385 : );
7386 :
7387 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
7388 0 : }
7389 :
7390 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
7391 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
7392 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
7393 0 : self.reconcilers_cancel.cancel();
7394 0 : self.reconcilers_gate.close().await;
7395 :
7396 : // Signal the background loop in [`Service::process_results`] to exit once
7397 : // it has proccessed the results from all the reconciles we cancelled earlier.
7398 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
7399 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
7400 0 : self.result_tx.closed().await;
7401 0 : }
7402 :
7403 0 : pub async fn shutdown(&self) {
7404 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
7405 0 : .await;
7406 :
7407 : // Background tasks hold gate guards: this notifies them of the cancellation and
7408 : // waits for them all to complete.
7409 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
7410 0 : self.cancel.cancel();
7411 0 : self.gate.close().await;
7412 0 : }
7413 :
7414 : /// Spot check the download lag for a secondary location of a shard.
7415 : /// Should be used as a heuristic, since it's not always precise: the
7416 : /// secondary might have not downloaded the new heat map yet and, hence,
7417 : /// is not aware of the lag.
7418 : ///
7419 : /// Returns:
7420 : /// * Ok(None) if the lag could not be determined from the status,
7421 : /// * Ok(Some(_)) if the lag could be determind
7422 : /// * Err on failures to query the pageserver.
7423 0 : async fn secondary_lag(
7424 0 : &self,
7425 0 : secondary: &NodeId,
7426 0 : tenant_shard_id: TenantShardId,
7427 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
7428 0 : let nodes = self.inner.read().unwrap().nodes.clone();
7429 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
7430 0 : StatusCode::NOT_FOUND,
7431 0 : format!("Node with id {} not found", secondary),
7432 0 : ))?;
7433 :
7434 0 : match node
7435 0 : .with_client_retries(
7436 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
7437 0 : &self.config.pageserver_jwt_token,
7438 0 : 1,
7439 0 : 3,
7440 0 : Duration::from_millis(250),
7441 0 : &self.cancel,
7442 0 : )
7443 0 : .await
7444 : {
7445 0 : Some(Ok(status)) => match status.heatmap_mtime {
7446 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
7447 0 : None => Ok(None),
7448 : },
7449 0 : Some(Err(e)) => Err(e),
7450 0 : None => Err(mgmt_api::Error::Cancelled),
7451 : }
7452 0 : }
7453 :
7454 : /// Drain a node by moving the shards attached to it as primaries.
7455 : /// This is a long running operation and it should run as a separate Tokio task.
7456 0 : pub(crate) async fn drain_node(
7457 0 : self: &Arc<Self>,
7458 0 : node_id: NodeId,
7459 0 : cancel: CancellationToken,
7460 0 : ) -> Result<(), OperationError> {
7461 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
7462 0 : let max_secondary_lag_bytes = self
7463 0 : .config
7464 0 : .max_secondary_lag_bytes
7465 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
7466 :
7467 : // By default, live migrations are generous about the wait time for getting
7468 : // the secondary location up to speed. When draining, give up earlier in order
7469 : // to not stall the operation when a cold secondary is encountered.
7470 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7471 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7472 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
7473 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7474 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7475 0 : .build();
7476 0 :
7477 0 : let mut waiters = Vec::new();
7478 0 :
7479 0 : let mut tid_iter = TenantShardIterator::new({
7480 0 : let service = self.clone();
7481 0 : move |last_inspected_shard: Option<TenantShardId>| {
7482 0 : let locked = &service.inner.read().unwrap();
7483 0 : let tenants = &locked.tenants;
7484 0 : let entry = match last_inspected_shard {
7485 0 : Some(skip_past) => {
7486 0 : // Skip to the last seen tenant shard id
7487 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
7488 0 :
7489 0 : // Skip past the last seen
7490 0 : cursor.nth(1)
7491 : }
7492 0 : None => tenants.first_key_value(),
7493 : };
7494 :
7495 0 : entry.map(|(tid, _)| tid).copied()
7496 0 : }
7497 0 : });
7498 :
7499 0 : while !tid_iter.finished() {
7500 0 : if cancel.is_cancelled() {
7501 0 : match self
7502 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7503 0 : .await
7504 : {
7505 0 : Ok(()) => return Err(OperationError::Cancelled),
7506 0 : Err(err) => {
7507 0 : return Err(OperationError::FinalizeError(
7508 0 : format!(
7509 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7510 0 : node_id, err
7511 0 : )
7512 0 : .into(),
7513 0 : ));
7514 : }
7515 : }
7516 0 : }
7517 0 :
7518 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
7519 :
7520 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7521 0 : let tid = match tid_iter.next() {
7522 0 : Some(tid) => tid,
7523 : None => {
7524 0 : break;
7525 : }
7526 : };
7527 :
7528 0 : let tid_drain = TenantShardDrain {
7529 0 : drained_node: node_id,
7530 0 : tenant_shard_id: tid,
7531 0 : };
7532 :
7533 0 : let dest_node_id = {
7534 0 : let locked = self.inner.read().unwrap();
7535 0 :
7536 0 : match tid_drain
7537 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
7538 : {
7539 0 : Some(node_id) => node_id,
7540 : None => {
7541 0 : continue;
7542 : }
7543 : }
7544 : };
7545 :
7546 0 : match self.secondary_lag(&dest_node_id, tid).await {
7547 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
7548 0 : // The secondary is reasonably up to date.
7549 0 : // Migrate to it
7550 0 : }
7551 0 : Ok(Some(lag)) => {
7552 0 : tracing::info!(
7553 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7554 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
7555 : );
7556 0 : continue;
7557 : }
7558 : Ok(None) => {
7559 0 : tracing::info!(
7560 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7561 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
7562 : );
7563 0 : continue;
7564 : }
7565 0 : Err(err) => {
7566 0 : tracing::warn!(
7567 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7568 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
7569 : );
7570 0 : continue;
7571 : }
7572 : }
7573 :
7574 : {
7575 0 : let mut locked = self.inner.write().unwrap();
7576 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7577 0 : let rescheduled = tid_drain.reschedule_to_secondary(
7578 0 : dest_node_id,
7579 0 : tenants,
7580 0 : scheduler,
7581 0 : nodes,
7582 0 : )?;
7583 :
7584 0 : if let Some(tenant_shard) = rescheduled {
7585 0 : let waiter = self.maybe_configured_reconcile_shard(
7586 0 : tenant_shard,
7587 0 : nodes,
7588 0 : reconciler_config,
7589 0 : );
7590 0 : if let Some(some) = waiter {
7591 0 : waiters.push(some);
7592 0 : }
7593 0 : }
7594 : }
7595 : }
7596 :
7597 0 : waiters = self
7598 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7599 0 : .await;
7600 :
7601 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
7602 : }
7603 :
7604 0 : while !waiters.is_empty() {
7605 0 : if cancel.is_cancelled() {
7606 0 : match self
7607 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7608 0 : .await
7609 : {
7610 0 : Ok(()) => return Err(OperationError::Cancelled),
7611 0 : Err(err) => {
7612 0 : return Err(OperationError::FinalizeError(
7613 0 : format!(
7614 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7615 0 : node_id, err
7616 0 : )
7617 0 : .into(),
7618 0 : ));
7619 : }
7620 : }
7621 0 : }
7622 0 :
7623 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
7624 :
7625 0 : waiters = self
7626 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7627 0 : .await;
7628 : }
7629 :
7630 : // At this point we have done the best we could to drain shards from this node.
7631 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
7632 : // to complete the drain.
7633 0 : if let Err(err) = self
7634 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
7635 0 : .await
7636 : {
7637 : // This is not fatal. Anything that is polling the node scheduling policy to detect
7638 : // the end of the drain operations will hang, but all such places should enforce an
7639 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
7640 : // by the counterpart fill operation.
7641 0 : return Err(OperationError::FinalizeError(
7642 0 : format!(
7643 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
7644 0 : )
7645 0 : .into(),
7646 0 : ));
7647 0 : }
7648 0 :
7649 0 : Ok(())
7650 0 : }
7651 :
7652 : /// Create a node fill plan (pick secondaries to promote), based on:
7653 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
7654 : /// outside their home AZ, should be migrated back here.
7655 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
7656 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
7657 : /// those shards have a home AZ that doesn't match the node we're filling.
7658 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
7659 0 : let mut locked = self.inner.write().unwrap();
7660 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
7661 0 :
7662 0 : let node_az = nodes
7663 0 : .get(&node_id)
7664 0 : .expect("Node must exist")
7665 0 : .get_availability_zone_id()
7666 0 : .clone();
7667 0 :
7668 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
7669 0 : let mut plan = Vec::new();
7670 0 :
7671 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
7672 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
7673 0 :
7674 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
7675 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
7676 0 : let respect_azs = nodes
7677 0 : .values()
7678 0 : .map(|n| n.get_availability_zone_id())
7679 0 : .unique()
7680 0 : .count()
7681 0 : > 1;
7682 :
7683 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
7684 : // requires it.
7685 0 : for (tsid, tenant_shard) in tenants {
7686 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7687 : // Shard doesn't have a secondary on this node, ignore it.
7688 0 : continue;
7689 0 : }
7690 0 :
7691 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
7692 0 : // shards which belong on this node, not to promote shards whose scheduling preference
7693 0 : // would be on their currently attached node. So will avoid promoting shards whose
7694 0 : // home AZ doesn't match the AZ of the node we're filling.
7695 0 : match tenant_shard.preferred_az() {
7696 0 : _ if !respect_azs => {
7697 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7698 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7699 0 : }
7700 : }
7701 : None => {
7702 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
7703 : // will only do so if our target shard count requires it.
7704 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7705 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
7706 0 : }
7707 : }
7708 0 : Some(az) if az == &node_az => {
7709 : // This shard's home AZ is equal to the node we're filling: it should
7710 : // be moved back to this node as part of filling, unless its currently
7711 : // attached location is also in its home AZ.
7712 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
7713 0 : if nodes
7714 0 : .get(primary)
7715 0 : .expect("referenced node must exist")
7716 0 : .get_availability_zone_id()
7717 0 : != tenant_shard
7718 0 : .preferred_az()
7719 0 : .expect("tenant must have an AZ preference")
7720 : {
7721 0 : plan.push(*tsid)
7722 0 : }
7723 : } else {
7724 0 : plan.push(*tsid)
7725 : }
7726 : }
7727 0 : Some(_) => {
7728 0 : // This shard's home AZ is somewhere other than the node we're filling,
7729 0 : // it may not be moved back to this node as part of filling. Ignore it
7730 0 : }
7731 : }
7732 : }
7733 :
7734 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
7735 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
7736 0 :
7737 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
7738 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
7739 0 :
7740 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
7741 :
7742 0 : for (node_id, attached) in nodes_by_load {
7743 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
7744 0 : if !available {
7745 0 : continue;
7746 0 : }
7747 0 :
7748 0 : if plan.len() >= fill_requirement
7749 0 : || free_tids_by_node.is_empty()
7750 0 : || attached <= expected_attached
7751 : {
7752 0 : break;
7753 0 : }
7754 0 :
7755 0 : let can_take = attached - expected_attached;
7756 0 : let needed = fill_requirement - plan.len();
7757 0 : let mut take = std::cmp::min(can_take, needed);
7758 0 :
7759 0 : let mut remove_node = false;
7760 0 : while take > 0 {
7761 0 : match free_tids_by_node.get_mut(&node_id) {
7762 0 : Some(tids) => match tids.pop() {
7763 0 : Some(tid) => {
7764 0 : let max_promote_for_tenant = std::cmp::max(
7765 0 : tid.shard_count.count() as usize / locked.nodes.len(),
7766 0 : 1,
7767 0 : );
7768 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
7769 0 : if *promoted < max_promote_for_tenant {
7770 0 : plan.push(tid);
7771 0 : *promoted += 1;
7772 0 : take -= 1;
7773 0 : }
7774 : }
7775 : None => {
7776 0 : remove_node = true;
7777 0 : break;
7778 : }
7779 : },
7780 : None => {
7781 0 : break;
7782 : }
7783 : }
7784 : }
7785 :
7786 0 : if remove_node {
7787 0 : free_tids_by_node.remove(&node_id);
7788 0 : }
7789 : }
7790 :
7791 0 : plan
7792 0 : }
7793 :
7794 : /// Fill a node by promoting its secondaries until the cluster is balanced
7795 : /// with regards to attached shard counts. Note that this operation only
7796 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
7797 : /// This is a long running operation and it should run as a separate Tokio task.
7798 0 : pub(crate) async fn fill_node(
7799 0 : &self,
7800 0 : node_id: NodeId,
7801 0 : cancel: CancellationToken,
7802 0 : ) -> Result<(), OperationError> {
7803 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
7804 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
7805 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
7806 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
7807 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
7808 0 : .build();
7809 0 :
7810 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
7811 0 : let mut waiters = Vec::new();
7812 :
7813 : // Execute the plan we've composed above. Before aplying each move from the plan,
7814 : // we validate to ensure that it has not gone stale in the meantime.
7815 0 : while !tids_to_promote.is_empty() {
7816 0 : if cancel.is_cancelled() {
7817 0 : match self
7818 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7819 0 : .await
7820 : {
7821 0 : Ok(()) => return Err(OperationError::Cancelled),
7822 0 : Err(err) => {
7823 0 : return Err(OperationError::FinalizeError(
7824 0 : format!(
7825 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7826 0 : node_id, err
7827 0 : )
7828 0 : .into(),
7829 0 : ));
7830 : }
7831 : }
7832 0 : }
7833 0 :
7834 0 : {
7835 0 : let mut locked = self.inner.write().unwrap();
7836 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7837 :
7838 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
7839 0 : format!("node {node_id} was removed").into(),
7840 0 : ))?;
7841 :
7842 0 : let current_policy = node.get_scheduling();
7843 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
7844 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
7845 : // about it
7846 0 : return Err(OperationError::NodeStateChanged(
7847 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
7848 0 : ));
7849 0 : }
7850 :
7851 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
7852 0 : if let Some(tid) = tids_to_promote.pop() {
7853 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
7854 : // If the node being filled is not a secondary anymore,
7855 : // skip the promotion.
7856 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
7857 0 : continue;
7858 0 : }
7859 0 :
7860 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
7861 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
7862 0 : Err(e) => {
7863 0 : tracing::warn!(
7864 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7865 0 : "Scheduling error when filling pageserver {} : {e}", node_id
7866 : );
7867 : }
7868 : Ok(()) => {
7869 0 : tracing::info!(
7870 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
7871 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
7872 : node_id,
7873 : previously_attached_to,
7874 : node_id
7875 : );
7876 :
7877 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
7878 0 : tenant_shard,
7879 0 : nodes,
7880 0 : reconciler_config,
7881 0 : ) {
7882 0 : waiters.push(waiter);
7883 0 : }
7884 : }
7885 : }
7886 0 : }
7887 : } else {
7888 0 : break;
7889 : }
7890 : }
7891 : }
7892 :
7893 0 : waiters = self
7894 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
7895 0 : .await;
7896 : }
7897 :
7898 0 : while !waiters.is_empty() {
7899 0 : if cancel.is_cancelled() {
7900 0 : match self
7901 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7902 0 : .await
7903 : {
7904 0 : Ok(()) => return Err(OperationError::Cancelled),
7905 0 : Err(err) => {
7906 0 : return Err(OperationError::FinalizeError(
7907 0 : format!(
7908 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
7909 0 : node_id, err
7910 0 : )
7911 0 : .into(),
7912 0 : ));
7913 : }
7914 : }
7915 0 : }
7916 0 :
7917 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
7918 :
7919 0 : waiters = self
7920 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
7921 0 : .await;
7922 : }
7923 :
7924 0 : if let Err(err) = self
7925 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
7926 0 : .await
7927 : {
7928 : // This isn't a huge issue since the filling process starts upon request. However, it
7929 : // will prevent the next drain from starting. The only case in which this can fail
7930 : // is database unavailability. Such a case will require manual intervention.
7931 0 : return Err(OperationError::FinalizeError(
7932 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
7933 0 : .into(),
7934 0 : ));
7935 0 : }
7936 0 :
7937 0 : Ok(())
7938 0 : }
7939 :
7940 : /// Updates scrubber metadata health check results.
7941 0 : pub(crate) async fn metadata_health_update(
7942 0 : &self,
7943 0 : update_req: MetadataHealthUpdateRequest,
7944 0 : ) -> Result<(), ApiError> {
7945 0 : let now = chrono::offset::Utc::now();
7946 0 : let (healthy_records, unhealthy_records) = {
7947 0 : let locked = self.inner.read().unwrap();
7948 0 : let healthy_records = update_req
7949 0 : .healthy_tenant_shards
7950 0 : .into_iter()
7951 0 : // Retain only health records associated with tenant shards managed by storage controller.
7952 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7953 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
7954 0 : .collect();
7955 0 : let unhealthy_records = update_req
7956 0 : .unhealthy_tenant_shards
7957 0 : .into_iter()
7958 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
7959 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
7960 0 : .collect();
7961 0 :
7962 0 : (healthy_records, unhealthy_records)
7963 0 : };
7964 0 :
7965 0 : self.persistence
7966 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
7967 0 : .await?;
7968 0 : Ok(())
7969 0 : }
7970 :
7971 : /// Lists the tenant shards that has unhealthy metadata status.
7972 0 : pub(crate) async fn metadata_health_list_unhealthy(
7973 0 : &self,
7974 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
7975 0 : let result = self
7976 0 : .persistence
7977 0 : .list_unhealthy_metadata_health_records()
7978 0 : .await?
7979 0 : .iter()
7980 0 : .map(|p| p.get_tenant_shard_id().unwrap())
7981 0 : .collect();
7982 0 :
7983 0 : Ok(result)
7984 0 : }
7985 :
7986 : /// Lists the tenant shards that have not been scrubbed for some duration.
7987 0 : pub(crate) async fn metadata_health_list_outdated(
7988 0 : &self,
7989 0 : not_scrubbed_for: Duration,
7990 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
7991 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
7992 0 : let result = self
7993 0 : .persistence
7994 0 : .list_outdated_metadata_health_records(earlier)
7995 0 : .await?
7996 0 : .into_iter()
7997 0 : .map(|record| record.into())
7998 0 : .collect();
7999 0 : Ok(result)
8000 0 : }
8001 :
8002 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
8003 0 : self.inner.read().unwrap().get_leadership_status()
8004 0 : }
8005 :
8006 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
8007 0 : tracing::info!("Received step down request from peer");
8008 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
8009 :
8010 0 : self.inner.write().unwrap().step_down();
8011 0 : // TODO: would it make sense to have a time-out for this?
8012 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
8013 0 : .await;
8014 :
8015 0 : let mut global_observed = GlobalObservedState::default();
8016 0 : let locked = self.inner.read().unwrap();
8017 0 : for (tid, tenant_shard) in locked.tenants.iter() {
8018 0 : global_observed
8019 0 : .0
8020 0 : .insert(*tid, tenant_shard.observed.clone());
8021 0 : }
8022 :
8023 0 : global_observed
8024 0 : }
8025 :
8026 0 : pub(crate) async fn safekeepers_list(
8027 0 : &self,
8028 0 : ) -> Result<Vec<SafekeeperDescribeResponse>, DatabaseError> {
8029 0 : let locked = self.inner.read().unwrap();
8030 0 : let mut list = locked
8031 0 : .safekeepers
8032 0 : .iter()
8033 0 : .map(|sk| sk.1.describe_response())
8034 0 : .collect::<Result<Vec<_>, _>>()?;
8035 0 : list.sort_by_key(|v| v.id);
8036 0 : Ok(list)
8037 0 : }
8038 :
8039 0 : pub(crate) async fn get_safekeeper(
8040 0 : &self,
8041 0 : id: i64,
8042 0 : ) -> Result<SafekeeperDescribeResponse, DatabaseError> {
8043 0 : let locked = self.inner.read().unwrap();
8044 0 : let sk = locked
8045 0 : .safekeepers
8046 0 : .get(&NodeId(id as u64))
8047 0 : .ok_or(diesel::result::Error::NotFound)?;
8048 0 : sk.describe_response()
8049 0 : }
8050 :
8051 0 : pub(crate) async fn upsert_safekeeper(
8052 0 : &self,
8053 0 : record: crate::persistence::SafekeeperUpsert,
8054 0 : ) -> Result<(), DatabaseError> {
8055 0 : let node_id = NodeId(record.id as u64);
8056 0 : self.persistence.safekeeper_upsert(record.clone()).await?;
8057 : {
8058 0 : let mut locked = self.inner.write().unwrap();
8059 0 : let mut safekeepers = (*locked.safekeepers).clone();
8060 0 : match safekeepers.entry(node_id) {
8061 0 : std::collections::hash_map::Entry::Occupied(mut entry) => {
8062 0 : entry.get_mut().update_from_record(record);
8063 0 : }
8064 0 : std::collections::hash_map::Entry::Vacant(entry) => {
8065 0 : entry.insert(Safekeeper::from_persistence(
8066 0 : crate::persistence::SafekeeperPersistence::from_upsert(
8067 0 : record,
8068 0 : SkSchedulingPolicy::Pause,
8069 0 : ),
8070 0 : CancellationToken::new(),
8071 0 : ));
8072 0 : }
8073 : }
8074 0 : locked.safekeepers = Arc::new(safekeepers);
8075 0 : }
8076 0 : Ok(())
8077 0 : }
8078 :
8079 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
8080 0 : &self,
8081 0 : id: i64,
8082 0 : scheduling_policy: SkSchedulingPolicy,
8083 0 : ) -> Result<(), DatabaseError> {
8084 0 : self.persistence
8085 0 : .set_safekeeper_scheduling_policy(id, scheduling_policy)
8086 0 : .await?;
8087 0 : let node_id = NodeId(id as u64);
8088 0 : // After the change has been persisted successfully, update the in-memory state
8089 0 : {
8090 0 : let mut locked = self.inner.write().unwrap();
8091 0 : let mut safekeepers = (*locked.safekeepers).clone();
8092 0 : let sk = safekeepers
8093 0 : .get_mut(&node_id)
8094 0 : .ok_or(DatabaseError::Logical("Not found".to_string()))?;
8095 0 : sk.set_scheduling_policy(scheduling_policy);
8096 0 :
8097 0 : locked.safekeepers = Arc::new(safekeepers);
8098 0 : }
8099 0 : Ok(())
8100 0 : }
8101 :
8102 0 : pub(crate) async fn update_shards_preferred_azs(
8103 0 : &self,
8104 0 : req: ShardsPreferredAzsRequest,
8105 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
8106 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
8107 0 : let updated = self
8108 0 : .persistence
8109 0 : .set_tenant_shard_preferred_azs(preferred_azs)
8110 0 : .await
8111 0 : .map_err(|err| {
8112 0 : ApiError::InternalServerError(anyhow::anyhow!(
8113 0 : "Failed to persist preferred AZs: {err}"
8114 0 : ))
8115 0 : })?;
8116 :
8117 0 : let mut updated_in_mem_and_db = Vec::default();
8118 0 :
8119 0 : let mut locked = self.inner.write().unwrap();
8120 0 : for (tid, az_id) in updated {
8121 0 : let shard = locked.tenants.get_mut(&tid);
8122 0 : if let Some(shard) = shard {
8123 0 : shard.set_preferred_az(az_id);
8124 0 : updated_in_mem_and_db.push(tid);
8125 0 : }
8126 : }
8127 :
8128 0 : Ok(ShardsPreferredAzsResponse {
8129 0 : updated: updated_in_mem_and_db,
8130 0 : })
8131 0 : }
8132 : }
|
