Line data Source code
1 : pub mod chaos_injector;
2 : mod context_iterator;
3 : pub(crate) mod safekeeper_reconciler;
4 : mod safekeeper_service;
5 :
6 : use std::borrow::Cow;
7 : use std::cmp::Ordering;
8 : use std::collections::{BTreeMap, HashMap, HashSet};
9 : use std::error::Error;
10 : use std::num::NonZeroU32;
11 : use std::ops::{Deref, DerefMut};
12 : use std::path::PathBuf;
13 : use std::str::FromStr;
14 : use std::sync::Arc;
15 : use std::time::{Duration, Instant, SystemTime};
16 :
17 : use anyhow::Context;
18 : use context_iterator::TenantShardContextIterator;
19 : use control_plane::storage_controller::{
20 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
21 : };
22 : use diesel::result::DatabaseErrorKind;
23 : use futures::StreamExt;
24 : use futures::stream::FuturesUnordered;
25 : use http_utils::error::ApiError;
26 : use hyper::Uri;
27 : use itertools::Itertools;
28 : use pageserver_api::controller_api::{
29 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
30 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
31 : ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
32 : TenantCreateRequest, TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse,
33 : TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
34 : TenantShardMigrateRequest, TenantShardMigrateResponse,
35 : };
36 : use pageserver_api::models::{
37 : self, DetachBehavior, LocationConfig, LocationConfigListResponse, LocationConfigMode, LsnLease,
38 : PageserverUtilization, SecondaryProgress, ShardParameters, TenantConfig,
39 : TenantConfigPatchRequest, TenantConfigRequest, TenantLocationConfigRequest,
40 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
41 : TenantShardSplitResponse, TenantSorting, TenantTimeTravelRequest,
42 : TimelineArchivalConfigRequest, TimelineCreateRequest, TimelineCreateResponseStorcon,
43 : TimelineInfo, TopTenantShardItem, TopTenantShardsRequest,
44 : };
45 : use pageserver_api::shard::{
46 : DEFAULT_STRIPE_SIZE, ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
47 : };
48 : use pageserver_api::upcall_api::{
49 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest, ValidateResponse,
50 : ValidateResponseTenant,
51 : };
52 : use pageserver_client::{BlockUnblock, mgmt_api};
53 : use reqwest::{Certificate, StatusCode};
54 : use safekeeper_api::models::SafekeeperUtilization;
55 : use safekeeper_reconciler::SafekeeperReconcilers;
56 : use tokio::sync::TryAcquireError;
57 : use tokio::sync::mpsc::error::TrySendError;
58 : use tokio_util::sync::CancellationToken;
59 : use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
60 : use utils::completion::Barrier;
61 : use utils::generation::Generation;
62 : use utils::id::{NodeId, TenantId, TimelineId};
63 : use utils::lsn::Lsn;
64 : use utils::sync::gate::{Gate, GateGuard};
65 : use utils::{failpoint_support, pausable_failpoint};
66 :
67 : use crate::background_node_operations::{
68 : Drain, Fill, MAX_RECONCILES_PER_OPERATION, Operation, OperationError, OperationHandler,
69 : };
70 : use crate::compute_hook::{self, ComputeHook, NotifyError};
71 : use crate::drain_utils::{self, TenantShardDrain, TenantShardIterator};
72 : use crate::heartbeater::{Heartbeater, PageserverState, SafekeeperState};
73 : use crate::id_lock_map::{
74 : IdLockMap, TracingExclusiveGuard, trace_exclusive_lock, trace_shared_lock,
75 : };
76 : use crate::leadership::Leadership;
77 : use crate::metrics;
78 : use crate::node::{AvailabilityTransition, Node};
79 : use crate::pageserver_client::PageserverClient;
80 : use crate::peer_client::GlobalObservedState;
81 : use crate::persistence::split_state::SplitState;
82 : use crate::persistence::{
83 : AbortShardSplitStatus, ControllerPersistence, DatabaseError, DatabaseResult,
84 : MetadataHealthPersistence, Persistence, ShardGenerationState, TenantFilter,
85 : TenantShardPersistence,
86 : };
87 : use crate::reconciler::{
88 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder, ReconcilerPriority,
89 : attached_location_conf,
90 : };
91 : use crate::safekeeper::Safekeeper;
92 : use crate::scheduler::{
93 : AttachedShardTag, MaySchedule, ScheduleContext, ScheduleError, ScheduleMode, Scheduler,
94 : };
95 : use crate::tenant_shard::{
96 : IntentState, MigrateAttachment, ObservedState, ObservedStateDelta, ObservedStateLocation,
97 : ReconcileNeeded, ReconcileResult, ReconcileWaitError, ReconcilerStatus, ReconcilerWaiter,
98 : ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
99 : };
100 :
101 : const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
102 :
103 : // For operations that should be quick, like attaching a new tenant
104 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
105 :
106 : // For operations that might be slow, like migrating a tenant with
107 : // some data in it.
108 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
109 :
110 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
111 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
112 : // input generation from future requests as authoritative.
113 : const INITIAL_GENERATION: Generation = Generation::new(0);
114 :
115 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
116 : /// up on unresponsive pageservers and proceed.
117 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
118 :
119 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
120 : /// This must be long enough to cover node restarts as well as normal operations: in future
121 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
122 :
123 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
124 : /// offline.
125 : ///
126 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
127 : /// handling of the re-attach response may take a long time and blocks heartbeats from
128 : /// being handled on the pageserver side.
129 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
130 :
131 : /// How often to send heartbeats to registered nodes?
132 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
133 :
134 : /// How long is too long for a reconciliation?
135 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
136 :
137 : #[derive(Clone, strum_macros::Display)]
138 : enum TenantOperations {
139 : Create,
140 : LocationConfig,
141 : ConfigSet,
142 : ConfigPatch,
143 : TimeTravelRemoteStorage,
144 : Delete,
145 : UpdatePolicy,
146 : ShardSplit,
147 : SecondaryDownload,
148 : TimelineCreate,
149 : TimelineDelete,
150 : AttachHook,
151 : TimelineArchivalConfig,
152 : TimelineDetachAncestor,
153 : TimelineGcBlockUnblock,
154 : DropDetached,
155 : DownloadHeatmapLayers,
156 : TimelineLsnLease,
157 : }
158 :
159 : #[derive(Clone, strum_macros::Display)]
160 : enum NodeOperations {
161 : Register,
162 : Configure,
163 : Delete,
164 : }
165 :
166 : /// The leadership status for the storage controller process.
167 : /// Allowed transitions are:
168 : /// 1. Leader -> SteppedDown
169 : /// 2. Candidate -> Leader
170 : #[derive(
171 : Eq,
172 : PartialEq,
173 : Copy,
174 : Clone,
175 : strum_macros::Display,
176 0 : strum_macros::EnumIter,
177 : measured::FixedCardinalityLabel,
178 : )]
179 : #[strum(serialize_all = "snake_case")]
180 : pub(crate) enum LeadershipStatus {
181 : /// This is the steady state where the storage controller can produce
182 : /// side effects in the cluster.
183 : Leader,
184 : /// We've been notified to step down by another candidate. No reconciliations
185 : /// take place in this state.
186 : SteppedDown,
187 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
188 : #[allow(unused)]
189 : Candidate,
190 : }
191 :
192 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
193 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
194 :
195 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
196 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
197 : // than they're being pushed onto the queue.
198 : const MAX_DELAYED_RECONCILES: usize = 10000;
199 :
200 : // Top level state available to all HTTP handlers
201 : struct ServiceState {
202 : leadership_status: LeadershipStatus,
203 :
204 : tenants: BTreeMap<TenantShardId, TenantShard>,
205 :
206 : nodes: Arc<HashMap<NodeId, Node>>,
207 :
208 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
209 :
210 : safekeeper_reconcilers: SafekeeperReconcilers,
211 :
212 : scheduler: Scheduler,
213 :
214 : /// Ongoing background operation on the cluster if any is running.
215 : /// Note that only one such operation may run at any given time,
216 : /// hence the type choice.
217 : ongoing_operation: Option<OperationHandler>,
218 :
219 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
220 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
221 : }
222 :
223 : /// Transform an error from a pageserver into an error to return to callers of a storage
224 : /// controller API.
225 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
226 0 : match e {
227 0 : mgmt_api::Error::SendRequest(e) => {
228 0 : // Presume errors sending requests are connectivity/availability issues
229 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
230 : }
231 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
232 0 : // Presume errors receiving body are connectivity/availability issues
233 0 : ApiError::ResourceUnavailable(
234 0 : format!("{node} error receiving error body: {str}").into(),
235 0 : )
236 : }
237 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
238 0 : // Return 500 for decoding errors.
239 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
240 : }
241 0 : mgmt_api::Error::ReceiveBody(err) => {
242 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
243 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
244 0 : ApiError::ResourceUnavailable(
245 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
246 0 : )
247 : }
248 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
249 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
250 : }
251 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
252 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
253 : }
254 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
255 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
256 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
257 : // internal server errors, showing that something is wrong with the pageserver or
258 : // storage controller's auth configuration.
259 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
260 : }
261 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
262 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
263 0 : // turn ask our clients to wait + retry
264 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
265 : }
266 0 : mgmt_api::Error::ApiError(status, msg) => {
267 0 : // Presume general case of pageserver API errors is that we tried to do something
268 0 : // that can't be done right now.
269 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
270 : }
271 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
272 0 : mgmt_api::Error::Timeout(e) => ApiError::Timeout(e.into()),
273 : }
274 0 : }
275 :
276 : impl ServiceState {
277 0 : fn new(
278 0 : nodes: HashMap<NodeId, Node>,
279 0 : safekeepers: HashMap<NodeId, Safekeeper>,
280 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
281 0 : scheduler: Scheduler,
282 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
283 0 : initial_leadership_status: LeadershipStatus,
284 0 : reconcilers_cancel: CancellationToken,
285 0 : ) -> Self {
286 0 : metrics::update_leadership_status(initial_leadership_status);
287 0 :
288 0 : Self {
289 0 : leadership_status: initial_leadership_status,
290 0 : tenants,
291 0 : nodes: Arc::new(nodes),
292 0 : safekeepers: Arc::new(safekeepers),
293 0 : safekeeper_reconcilers: SafekeeperReconcilers::new(reconcilers_cancel),
294 0 : scheduler,
295 0 : ongoing_operation: None,
296 0 : delayed_reconcile_rx,
297 0 : }
298 0 : }
299 :
300 0 : fn parts_mut(
301 0 : &mut self,
302 0 : ) -> (
303 0 : &mut Arc<HashMap<NodeId, Node>>,
304 0 : &mut BTreeMap<TenantShardId, TenantShard>,
305 0 : &mut Scheduler,
306 0 : ) {
307 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
308 0 : }
309 :
310 : #[allow(clippy::type_complexity)]
311 0 : fn parts_mut_sk(
312 0 : &mut self,
313 0 : ) -> (
314 0 : &mut Arc<HashMap<NodeId, Node>>,
315 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
316 0 : &mut BTreeMap<TenantShardId, TenantShard>,
317 0 : &mut Scheduler,
318 0 : ) {
319 0 : (
320 0 : &mut self.nodes,
321 0 : &mut self.safekeepers,
322 0 : &mut self.tenants,
323 0 : &mut self.scheduler,
324 0 : )
325 0 : }
326 :
327 0 : fn get_leadership_status(&self) -> LeadershipStatus {
328 0 : self.leadership_status
329 0 : }
330 :
331 0 : fn step_down(&mut self) {
332 0 : self.leadership_status = LeadershipStatus::SteppedDown;
333 0 : metrics::update_leadership_status(self.leadership_status);
334 0 : }
335 :
336 0 : fn become_leader(&mut self) {
337 0 : self.leadership_status = LeadershipStatus::Leader;
338 0 : metrics::update_leadership_status(self.leadership_status);
339 0 : }
340 : }
341 :
342 : #[derive(Clone)]
343 : pub struct Config {
344 : // All pageservers managed by one instance of this service must have
345 : // the same public key. This JWT token will be used to authenticate
346 : // this service to the pageservers it manages.
347 : pub pageserver_jwt_token: Option<String>,
348 :
349 : // All safekeepers managed by one instance of this service must have
350 : // the same public key. This JWT token will be used to authenticate
351 : // this service to the safekeepers it manages.
352 : pub safekeeper_jwt_token: Option<String>,
353 :
354 : // This JWT token will be used to authenticate this service to the control plane.
355 : pub control_plane_jwt_token: Option<String>,
356 :
357 : // This JWT token will be used to authenticate with other storage controller instances
358 : pub peer_jwt_token: Option<String>,
359 :
360 : /// Where the compute hook should send notifications of pageserver attachment locations
361 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
362 : /// assume it is running in a test environment and try to update neon_local.
363 : pub compute_hook_url: Option<String>,
364 :
365 : /// Prefix for storage API endpoints of the control plane. We use this prefix to compute
366 : /// URLs that we use to send pageserver and safekeeper attachment locations.
367 : /// If this is None, the compute hook will assume it is running in a test environment
368 : /// and try to invoke neon_local instead.
369 : ///
370 : /// For now, there is also `compute_hook_url` which allows configuration of the pageserver
371 : /// specific endpoint, but it is in the process of being phased out.
372 : pub control_plane_url: Option<String>,
373 :
374 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
375 : /// considered active. Once the grace period elapses, the next heartbeat failure will
376 : /// mark the pagseserver offline.
377 : pub max_offline_interval: Duration,
378 :
379 : /// Extended grace period within which pageserver may not respond to heartbeats.
380 : /// This extended grace period kicks in after the node has been drained for restart
381 : /// and/or upon handling the re-attach request from a node.
382 : pub max_warming_up_interval: Duration,
383 :
384 : /// How many normal-priority Reconcilers may be spawned concurrently
385 : pub reconciler_concurrency: usize,
386 :
387 : /// How many high-priority Reconcilers may be spawned concurrently
388 : pub priority_reconciler_concurrency: usize,
389 :
390 : /// How many API requests per second to allow per tenant, across all
391 : /// tenant-scoped API endpoints. Further API requests queue until ready.
392 : pub tenant_rate_limit: NonZeroU32,
393 :
394 : /// If a tenant shard's largest timeline (max_logical_size) exceeds this value, all tenant
395 : /// shards will be split in 2 until they fall below split_threshold (up to max_split_shards).
396 : ///
397 : /// This will greedily split into as many shards as necessary to fall below split_threshold, as
398 : /// powers of 2: if a tenant shard is 7 times larger than split_threshold, it will split into 8
399 : /// immediately, rather than first 2 then 4 then 8.
400 : ///
401 : /// None or 0 disables auto-splitting.
402 : ///
403 : /// TODO: consider using total logical size of all timelines instead.
404 : pub split_threshold: Option<u64>,
405 :
406 : /// The maximum number of shards a tenant can be split into during autosplits. Does not affect
407 : /// manual split requests. 0 or 1 disables autosplits, as we already have 1 shard.
408 : pub max_split_shards: u8,
409 :
410 : /// The size at which an unsharded tenant should initially split. Ingestion is significantly
411 : /// faster with multiple shards, so eagerly splitting below split_threshold will typically speed
412 : /// up initial ingestion of large tenants.
413 : ///
414 : /// This should be below split_threshold, but it is not required. If both split_threshold and
415 : /// initial_split_threshold qualify, the largest number of target shards will be used.
416 : ///
417 : /// Does not apply to already sharded tenants: changing initial_split_threshold or
418 : /// initial_split_shards is not retroactive for already-sharded tenants.
419 : ///
420 : /// None or 0 disables initial splits.
421 : pub initial_split_threshold: Option<u64>,
422 :
423 : /// The number of shards to split into when reaching initial_split_threshold. Will
424 : /// be clamped to max_split_shards.
425 : ///
426 : /// 0 or 1 disables initial splits. Has no effect if initial_split_threshold is disabled.
427 : pub initial_split_shards: u8,
428 :
429 : // TODO: make this cfg(feature = "testing")
430 : pub neon_local_repo_dir: Option<PathBuf>,
431 :
432 : // Maximum acceptable download lag for the secondary location
433 : // while draining a node. If the secondary location is lagging
434 : // by more than the configured amount, then the secondary is not
435 : // upgraded to primary.
436 : pub max_secondary_lag_bytes: Option<u64>,
437 :
438 : pub heartbeat_interval: Duration,
439 :
440 : pub address_for_peers: Option<Uri>,
441 :
442 : pub start_as_candidate: bool,
443 :
444 : pub long_reconcile_threshold: Duration,
445 :
446 : pub use_https_pageserver_api: bool,
447 :
448 : pub use_https_safekeeper_api: bool,
449 :
450 : pub ssl_ca_certs: Vec<Certificate>,
451 :
452 : pub timelines_onto_safekeepers: bool,
453 :
454 : pub use_local_compute_notifications: bool,
455 : }
456 :
457 : impl From<DatabaseError> for ApiError {
458 0 : fn from(err: DatabaseError) -> ApiError {
459 0 : match err {
460 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
461 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
462 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
463 0 : ApiError::ShuttingDown
464 : }
465 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
466 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
467 : }
468 : }
469 0 : }
470 : }
471 :
472 : enum InitialShardScheduleOutcome {
473 : Scheduled(TenantCreateResponseShard),
474 : NotScheduled,
475 : ShardScheduleError(ScheduleError),
476 : }
477 :
478 : pub struct Service {
479 : inner: Arc<std::sync::RwLock<ServiceState>>,
480 : config: Config,
481 : persistence: Arc<Persistence>,
482 : compute_hook: Arc<ComputeHook>,
483 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
484 :
485 : heartbeater_ps: Heartbeater<Node, PageserverState>,
486 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
487 :
488 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
489 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
490 :
491 : // Locking on a tenant granularity (covers all shards in the tenant):
492 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
493 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
494 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
495 :
496 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
497 : // that transition it to/from Active.
498 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
499 :
500 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
501 : // and reconciliation on startup.
502 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
503 :
504 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
505 : // a human user might be waiting for.
506 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
507 :
508 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
509 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
510 : ///
511 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
512 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
513 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
514 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
515 :
516 : // Process shutdown will fire this token
517 : cancel: CancellationToken,
518 :
519 : // Child token of [`Service::cancel`] used by reconcilers
520 : reconcilers_cancel: CancellationToken,
521 :
522 : // Background tasks will hold this gate
523 : gate: Gate,
524 :
525 : // Reconcilers background tasks will hold this gate
526 : reconcilers_gate: Gate,
527 :
528 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
529 : /// passes, it isn't safe to do any actions that mutate tenants.
530 : pub(crate) startup_complete: Barrier,
531 :
532 : /// HTTP client with proper CA certs.
533 : http_client: reqwest::Client,
534 : }
535 :
536 : impl From<ReconcileWaitError> for ApiError {
537 0 : fn from(value: ReconcileWaitError) -> Self {
538 0 : match value {
539 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
540 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
541 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
542 : }
543 0 : }
544 : }
545 :
546 : impl From<OperationError> for ApiError {
547 0 : fn from(value: OperationError) -> Self {
548 0 : match value {
549 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
550 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
551 : }
552 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
553 : }
554 0 : }
555 : }
556 :
557 : #[allow(clippy::large_enum_variant)]
558 : enum TenantCreateOrUpdate {
559 : Create(TenantCreateRequest),
560 : Update(Vec<ShardUpdate>),
561 : }
562 :
563 : struct ShardSplitParams {
564 : old_shard_count: ShardCount,
565 : new_shard_count: ShardCount,
566 : new_stripe_size: Option<ShardStripeSize>,
567 : targets: Vec<ShardSplitTarget>,
568 : policy: PlacementPolicy,
569 : config: TenantConfig,
570 : shard_ident: ShardIdentity,
571 : preferred_az_id: Option<AvailabilityZone>,
572 : }
573 :
574 : // When preparing for a shard split, we may either choose to proceed with the split,
575 : // or find that the work is already done and return NoOp.
576 : enum ShardSplitAction {
577 : Split(Box<ShardSplitParams>),
578 : NoOp(TenantShardSplitResponse),
579 : }
580 :
581 : // A parent shard which will be split
582 : struct ShardSplitTarget {
583 : parent_id: TenantShardId,
584 : node: Node,
585 : child_ids: Vec<TenantShardId>,
586 : }
587 :
588 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
589 : /// might not be available. We therefore use a queue of abort operations processed in the background.
590 : struct TenantShardSplitAbort {
591 : tenant_id: TenantId,
592 : /// The target values from the request that failed
593 : new_shard_count: ShardCount,
594 : new_stripe_size: Option<ShardStripeSize>,
595 : /// Until this abort op is complete, no other operations may be done on the tenant
596 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
597 : /// The reconciler gate for the duration of the split operation, and any included abort.
598 : _gate: GateGuard,
599 : }
600 :
601 : #[derive(thiserror::Error, Debug)]
602 : enum TenantShardSplitAbortError {
603 : #[error(transparent)]
604 : Database(#[from] DatabaseError),
605 : #[error(transparent)]
606 : Remote(#[from] mgmt_api::Error),
607 : #[error("Unavailable")]
608 : Unavailable,
609 : }
610 :
611 : /// Inputs for computing a target shard count for a tenant.
612 : struct ShardSplitInputs {
613 : /// Current shard count.
614 : shard_count: ShardCount,
615 : /// Total size of largest timeline summed across all shards.
616 : max_logical_size: u64,
617 : /// Size-based split threshold. Zero if size-based splits are disabled.
618 : split_threshold: u64,
619 : /// Upper bound on target shards. 0 or 1 disables splits.
620 : max_split_shards: u8,
621 : /// Initial split threshold. Zero if initial splits are disabled.
622 : initial_split_threshold: u64,
623 : /// Number of shards for initial splits. 0 or 1 disables initial splits.
624 : initial_split_shards: u8,
625 : }
626 :
627 : struct ShardUpdate {
628 : tenant_shard_id: TenantShardId,
629 : placement_policy: PlacementPolicy,
630 : tenant_config: TenantConfig,
631 :
632 : /// If this is None, generation is not updated.
633 : generation: Option<Generation>,
634 :
635 : /// If this is None, scheduling policy is not updated.
636 : scheduling_policy: Option<ShardSchedulingPolicy>,
637 : }
638 :
639 : enum StopReconciliationsReason {
640 : ShuttingDown,
641 : SteppingDown,
642 : }
643 :
644 : impl std::fmt::Display for StopReconciliationsReason {
645 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
646 0 : let s = match self {
647 0 : Self::ShuttingDown => "Shutting down",
648 0 : Self::SteppingDown => "Stepping down",
649 : };
650 0 : write!(writer, "{}", s)
651 0 : }
652 : }
653 :
654 : pub(crate) enum ReconcileResultRequest {
655 : ReconcileResult(ReconcileResult),
656 : Stop,
657 : }
658 :
659 : #[derive(Clone)]
660 : struct MutationLocation {
661 : node: Node,
662 : generation: Generation,
663 : }
664 :
665 : #[derive(Clone)]
666 : struct ShardMutationLocations {
667 : latest: MutationLocation,
668 : other: Vec<MutationLocation>,
669 : }
670 :
671 : #[derive(Default, Clone)]
672 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
673 :
674 : impl Service {
675 0 : pub fn get_config(&self) -> &Config {
676 0 : &self.config
677 0 : }
678 :
679 0 : pub fn get_http_client(&self) -> &reqwest::Client {
680 0 : &self.http_client
681 0 : }
682 :
683 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
684 : /// view of the world, and determine which pageservers are responsive.
685 : #[instrument(skip_all)]
686 : async fn startup_reconcile(
687 : self: &Arc<Service>,
688 : current_leader: Option<ControllerPersistence>,
689 : leader_step_down_state: Option<GlobalObservedState>,
690 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
691 : Result<(), (TenantShardId, NotifyError)>,
692 : >,
693 : ) {
694 : // Startup reconciliation does I/O to other services: whether they
695 : // are responsive or not, we should aim to finish within our deadline, because:
696 : // - If we don't, a k8s readiness hook watching /ready will kill us.
697 : // - While we're waiting for startup reconciliation, we are not fully
698 : // available for end user operations like creating/deleting tenants and timelines.
699 : //
700 : // We set multiple deadlines to break up the time available between the phases of work: this is
701 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
702 : let start_at = Instant::now();
703 : let node_scan_deadline = start_at
704 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
705 : .expect("Reconcile timeout is a modest constant");
706 :
707 : let observed = if let Some(state) = leader_step_down_state {
708 : tracing::info!(
709 : "Using observed state received from leader at {}",
710 : current_leader.as_ref().unwrap().address
711 : );
712 :
713 : state
714 : } else {
715 : self.build_global_observed_state(node_scan_deadline).await
716 : };
717 :
718 : // Accumulate a list of any tenant locations that ought to be detached
719 : let mut cleanup = Vec::new();
720 :
721 : // Send initial heartbeat requests to all nodes loaded from the database
722 : let all_nodes = {
723 : let locked = self.inner.read().unwrap();
724 : locked.nodes.clone()
725 : };
726 : let (mut nodes_online, mut sks_online) =
727 : self.initial_heartbeat_round(all_nodes.keys()).await;
728 :
729 : // List of tenants for which we will attempt to notify compute of their location at startup
730 : let mut compute_notifications = Vec::new();
731 :
732 : // Populate intent and observed states for all tenants, based on reported state on pageservers
733 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
734 : let shard_count = {
735 : let mut locked = self.inner.write().unwrap();
736 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
737 :
738 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
739 : let mut new_nodes = (**nodes).clone();
740 : for (node_id, node) in new_nodes.iter_mut() {
741 : if let Some(utilization) = nodes_online.remove(node_id) {
742 : node.set_availability(NodeAvailability::Active(utilization));
743 : scheduler.node_upsert(node);
744 : }
745 : }
746 : *nodes = Arc::new(new_nodes);
747 :
748 : let mut new_sks = (**safekeepers).clone();
749 : for (node_id, node) in new_sks.iter_mut() {
750 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
751 : node.set_availability(SafekeeperState::Available {
752 : utilization,
753 : last_seen_at,
754 : });
755 : }
756 : }
757 : *safekeepers = Arc::new(new_sks);
758 :
759 : for (tenant_shard_id, observed_state) in observed.0 {
760 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
761 : for node_id in observed_state.locations.keys() {
762 : cleanup.push((tenant_shard_id, *node_id));
763 : }
764 :
765 : continue;
766 : };
767 :
768 : tenant_shard.observed = observed_state;
769 : }
770 :
771 : // Populate each tenant's intent state
772 : let mut schedule_context = ScheduleContext::default();
773 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
774 : if tenant_shard_id.shard_number == ShardNumber(0) {
775 : // Reset scheduling context each time we advance to the next Tenant
776 : schedule_context = ScheduleContext::default();
777 : }
778 :
779 : tenant_shard.intent_from_observed(scheduler);
780 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
781 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
782 : // not enough pageservers are available. The tenant may well still be available
783 : // to clients.
784 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
785 : } else {
786 : // If we're both intending and observed to be attached at a particular node, we will
787 : // emit a compute notification for this. In the case where our observed state does not
788 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
789 : if let Some(attached_at) = tenant_shard.stably_attached() {
790 : compute_notifications.push(compute_hook::ShardUpdate {
791 : tenant_shard_id: *tenant_shard_id,
792 : node_id: attached_at,
793 : stripe_size: tenant_shard.shard.stripe_size,
794 : preferred_az: tenant_shard
795 : .preferred_az()
796 0 : .map(|az| Cow::Owned(az.clone())),
797 : });
798 : }
799 : }
800 : }
801 :
802 : tenants.len()
803 : };
804 :
805 : // Before making any obeservable changes to the cluster, persist self
806 : // as leader in database and memory.
807 : let leadership = Leadership::new(
808 : self.persistence.clone(),
809 : self.config.clone(),
810 : self.cancel.child_token(),
811 : );
812 :
813 : if let Err(e) = leadership.become_leader(current_leader).await {
814 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
815 : std::process::exit(1);
816 : }
817 :
818 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
819 : let sk_schedule_requests =
820 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
821 : Ok(v) => v,
822 : Err(e) => {
823 : tracing::warn!(
824 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
825 : );
826 : // std::process::exit(1);
827 : Vec::new()
828 : }
829 : };
830 :
831 : {
832 : let mut locked = self.inner.write().unwrap();
833 : locked.become_leader();
834 :
835 : locked
836 : .safekeeper_reconcilers
837 : .schedule_request_vec(self, sk_schedule_requests);
838 : }
839 :
840 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
841 : // generation_pageserver in the database.
842 :
843 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
844 : // will emit compute hook notifications when they reconcile.
845 : //
846 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
847 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
848 : // calls will be correctly ordered wrt these.
849 : //
850 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
851 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
852 : // unit and start doing I/O.
853 : tracing::info!(
854 : "Sending {} compute notifications",
855 : compute_notifications.len()
856 : );
857 : self.compute_hook.notify_background(
858 : compute_notifications,
859 : bg_compute_notify_result_tx.clone(),
860 : &self.cancel,
861 : );
862 :
863 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
864 : // which require it: under normal circumstances this should only include tenants that were in some
865 : // transient state before we restarted, or any tenants whose compute hooks failed above.
866 : tracing::info!("Checking for shards in need of reconciliation...");
867 : let reconcile_tasks = self.reconcile_all();
868 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
869 : // normal operations may proceed.
870 :
871 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
872 : // background because it does not need to complete in order to proceed with other work.
873 : if !cleanup.is_empty() {
874 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
875 : tokio::task::spawn({
876 : let cleanup_self = self.clone();
877 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
878 : });
879 : }
880 :
881 : tracing::info!(
882 : "Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)"
883 : );
884 : }
885 :
886 0 : async fn initial_heartbeat_round<'a>(
887 0 : &self,
888 0 : node_ids: impl Iterator<Item = &'a NodeId>,
889 0 : ) -> (
890 0 : HashMap<NodeId, PageserverUtilization>,
891 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
892 0 : ) {
893 0 : assert!(!self.startup_complete.is_ready());
894 :
895 0 : let all_nodes = {
896 0 : let locked = self.inner.read().unwrap();
897 0 : locked.nodes.clone()
898 0 : };
899 0 :
900 0 : let mut nodes_to_heartbeat = HashMap::new();
901 0 : for node_id in node_ids {
902 0 : match all_nodes.get(node_id) {
903 0 : Some(node) => {
904 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
905 0 : }
906 : None => {
907 0 : tracing::warn!("Node {node_id} was removed during start-up");
908 : }
909 : }
910 : }
911 :
912 0 : let all_sks = {
913 0 : let locked = self.inner.read().unwrap();
914 0 : locked.safekeepers.clone()
915 0 : };
916 0 :
917 0 : tracing::info!("Sending initial heartbeats...");
918 0 : let (res_ps, res_sk) = tokio::join!(
919 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
920 0 : self.heartbeater_sk.heartbeat(all_sks)
921 0 : );
922 :
923 0 : let mut online_nodes = HashMap::new();
924 0 : if let Ok(deltas) = res_ps {
925 0 : for (node_id, status) in deltas.0 {
926 0 : match status {
927 0 : PageserverState::Available { utilization, .. } => {
928 0 : online_nodes.insert(node_id, utilization);
929 0 : }
930 0 : PageserverState::Offline => {}
931 : PageserverState::WarmingUp { .. } => {
932 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
933 : }
934 : }
935 : }
936 0 : }
937 :
938 0 : let mut online_sks = HashMap::new();
939 0 : if let Ok(deltas) = res_sk {
940 0 : for (node_id, status) in deltas.0 {
941 0 : match status {
942 : SafekeeperState::Available {
943 0 : utilization,
944 0 : last_seen_at,
945 0 : } => {
946 0 : online_sks.insert(node_id, (utilization, last_seen_at));
947 0 : }
948 0 : SafekeeperState::Offline => {}
949 : }
950 : }
951 0 : }
952 :
953 0 : (online_nodes, online_sks)
954 0 : }
955 :
956 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
957 : ///
958 : /// The result includes only nodes which responded within the deadline
959 0 : async fn scan_node_locations(
960 0 : &self,
961 0 : deadline: Instant,
962 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
963 0 : let nodes = {
964 0 : let locked = self.inner.read().unwrap();
965 0 : locked.nodes.clone()
966 0 : };
967 0 :
968 0 : let mut node_results = HashMap::new();
969 0 :
970 0 : let mut node_list_futs = FuturesUnordered::new();
971 0 :
972 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
973 0 : for node in nodes.values() {
974 0 : node_list_futs.push({
975 0 : async move {
976 0 : tracing::info!("Scanning shards on node {node}...");
977 0 : let timeout = Duration::from_secs(5);
978 0 : let response = node
979 0 : .with_client_retries(
980 0 : |client| async move { client.list_location_config().await },
981 0 : &self.http_client,
982 0 : &self.config.pageserver_jwt_token,
983 0 : 1,
984 0 : 5,
985 0 : timeout,
986 0 : &self.cancel,
987 0 : )
988 0 : .await;
989 0 : (node.get_id(), response)
990 0 : }
991 0 : });
992 0 : }
993 :
994 : loop {
995 0 : let (node_id, result) = tokio::select! {
996 0 : next = node_list_futs.next() => {
997 0 : match next {
998 0 : Some(result) => result,
999 : None =>{
1000 : // We got results for all our nodes
1001 0 : break;
1002 : }
1003 :
1004 : }
1005 : },
1006 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
1007 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
1008 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
1009 0 : break;
1010 : }
1011 : };
1012 :
1013 0 : let Some(list_response) = result else {
1014 0 : tracing::info!("Shutdown during startup_reconcile");
1015 0 : break;
1016 : };
1017 :
1018 0 : match list_response {
1019 0 : Err(e) => {
1020 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
1021 : }
1022 0 : Ok(listing) => {
1023 0 : node_results.insert(node_id, listing);
1024 0 : }
1025 : }
1026 : }
1027 :
1028 0 : node_results
1029 0 : }
1030 :
1031 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
1032 0 : let node_listings = self.scan_node_locations(deadline).await;
1033 0 : let mut observed = GlobalObservedState::default();
1034 :
1035 0 : for (node_id, location_confs) in node_listings {
1036 0 : tracing::info!(
1037 0 : "Received {} shard statuses from pageserver {}",
1038 0 : location_confs.tenant_shards.len(),
1039 : node_id
1040 : );
1041 :
1042 0 : for (tid, location_conf) in location_confs.tenant_shards {
1043 0 : let entry = observed.0.entry(tid).or_default();
1044 0 : entry.locations.insert(
1045 0 : node_id,
1046 0 : ObservedStateLocation {
1047 0 : conf: location_conf,
1048 0 : },
1049 0 : );
1050 0 : }
1051 : }
1052 :
1053 0 : observed
1054 0 : }
1055 :
1056 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
1057 : ///
1058 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1059 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1060 : /// other task trying to attach it.
1061 : #[instrument(skip_all)]
1062 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1063 : let nodes = self.inner.read().unwrap().nodes.clone();
1064 :
1065 : for (tenant_shard_id, node_id) in cleanup {
1066 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1067 : let Some(node) = nodes.get(&node_id) else {
1068 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1069 : // a location to clean up on a node that has since been removed.
1070 : tracing::info!(
1071 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1072 : );
1073 : continue;
1074 : };
1075 :
1076 : if self.cancel.is_cancelled() {
1077 : break;
1078 : }
1079 :
1080 : let client = PageserverClient::new(
1081 : node.get_id(),
1082 : self.http_client.clone(),
1083 : node.base_url(),
1084 : self.config.pageserver_jwt_token.as_deref(),
1085 : );
1086 : match client
1087 : .location_config(
1088 : tenant_shard_id,
1089 : LocationConfig {
1090 : mode: LocationConfigMode::Detached,
1091 : generation: None,
1092 : secondary_conf: None,
1093 : shard_number: tenant_shard_id.shard_number.0,
1094 : shard_count: tenant_shard_id.shard_count.literal(),
1095 : shard_stripe_size: 0,
1096 : tenant_conf: models::TenantConfig::default(),
1097 : },
1098 : None,
1099 : false,
1100 : )
1101 : .await
1102 : {
1103 : Ok(()) => {
1104 : tracing::info!(
1105 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1106 : );
1107 : }
1108 : Err(e) => {
1109 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1110 : // break anything.
1111 : tracing::error!(
1112 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1113 : );
1114 : }
1115 : }
1116 : }
1117 : }
1118 :
1119 : /// Long running background task that periodically wakes up and looks for shards that need
1120 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1121 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1122 : /// for those retries.
1123 : #[instrument(skip_all)]
1124 : async fn background_reconcile(self: &Arc<Self>) {
1125 : self.startup_complete.clone().wait().await;
1126 :
1127 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1128 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1129 : while !self.reconcilers_cancel.is_cancelled() {
1130 : tokio::select! {
1131 : _ = interval.tick() => {
1132 : let reconciles_spawned = self.reconcile_all();
1133 : if reconciles_spawned == 0 {
1134 : // Run optimizer only when we didn't find any other work to do
1135 : self.optimize_all().await;
1136 : }
1137 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1138 : // must be responsive when new projects begin ingesting and reach the threshold.
1139 : self.autosplit_tenants().await;
1140 : }
1141 : _ = self.reconcilers_cancel.cancelled() => return
1142 : }
1143 : }
1144 : }
1145 : /// Heartbeat all storage nodes once in a while.
1146 : #[instrument(skip_all)]
1147 : async fn spawn_heartbeat_driver(&self) {
1148 : self.startup_complete.clone().wait().await;
1149 :
1150 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1151 : while !self.cancel.is_cancelled() {
1152 : tokio::select! {
1153 : _ = interval.tick() => { }
1154 : _ = self.cancel.cancelled() => return
1155 : };
1156 :
1157 : let nodes = {
1158 : let locked = self.inner.read().unwrap();
1159 : locked.nodes.clone()
1160 : };
1161 :
1162 : let safekeepers = {
1163 : let locked = self.inner.read().unwrap();
1164 : locked.safekeepers.clone()
1165 : };
1166 :
1167 : let (res_ps, res_sk) = tokio::join!(
1168 : self.heartbeater_ps.heartbeat(nodes),
1169 : self.heartbeater_sk.heartbeat(safekeepers)
1170 : );
1171 :
1172 : if let Ok(deltas) = res_ps {
1173 : let mut to_handle = Vec::default();
1174 :
1175 : for (node_id, state) in deltas.0 {
1176 : let new_availability = match state {
1177 : PageserverState::Available { utilization, .. } => {
1178 : NodeAvailability::Active(utilization)
1179 : }
1180 : PageserverState::WarmingUp { started_at } => {
1181 : NodeAvailability::WarmingUp(started_at)
1182 : }
1183 : PageserverState::Offline => {
1184 : // The node might have been placed in the WarmingUp state
1185 : // while the heartbeat round was on-going. Hence, filter out
1186 : // offline transitions for WarmingUp nodes that are still within
1187 : // their grace period.
1188 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1189 : .get_node(node_id)
1190 : .await
1191 : .as_ref()
1192 0 : .map(|n| n.get_availability())
1193 : {
1194 : let now = Instant::now();
1195 : if now - *started_at >= self.config.max_warming_up_interval {
1196 : NodeAvailability::Offline
1197 : } else {
1198 : NodeAvailability::WarmingUp(*started_at)
1199 : }
1200 : } else {
1201 : NodeAvailability::Offline
1202 : }
1203 : }
1204 : };
1205 :
1206 : let node_lock = trace_exclusive_lock(
1207 : &self.node_op_locks,
1208 : node_id,
1209 : NodeOperations::Configure,
1210 : )
1211 : .await;
1212 :
1213 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1214 :
1215 : // This is the code path for geniune availability transitions (i.e node
1216 : // goes unavailable and/or comes back online).
1217 : let res = self
1218 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1219 : .await;
1220 :
1221 : match res {
1222 : Ok(transition) => {
1223 : // Keep hold of the lock until the availability transitions
1224 : // have been handled in
1225 : // [`Service::handle_node_availability_transitions`] in order avoid
1226 : // racing with [`Service::external_node_configure`].
1227 : to_handle.push((node_id, node_lock, transition));
1228 : }
1229 : Err(ApiError::NotFound(_)) => {
1230 : // This should be rare, but legitimate since the heartbeats are done
1231 : // on a snapshot of the nodes.
1232 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1233 : }
1234 : Err(ApiError::ShuttingDown) => {
1235 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1236 : }
1237 : Err(err) => {
1238 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1239 : // becomes unavailable again, we may get an error here.
1240 : tracing::error!(
1241 : "Failed to update node state {} after heartbeat round: {}",
1242 : node_id,
1243 : err
1244 : );
1245 : }
1246 : }
1247 : }
1248 :
1249 : // We collected all the transitions above and now we handle them.
1250 : let res = self.handle_node_availability_transitions(to_handle).await;
1251 : if let Err(errs) = res {
1252 : for (node_id, err) in errs {
1253 : match err {
1254 : ApiError::NotFound(_) => {
1255 : // This should be rare, but legitimate since the heartbeats are done
1256 : // on a snapshot of the nodes.
1257 : tracing::info!(
1258 : "Node {} was not found after heartbeat round",
1259 : node_id
1260 : );
1261 : }
1262 : err => {
1263 : tracing::error!(
1264 : "Failed to handle availability transition for {} after heartbeat round: {}",
1265 : node_id,
1266 : err
1267 : );
1268 : }
1269 : }
1270 : }
1271 : }
1272 : }
1273 : if let Ok(deltas) = res_sk {
1274 : let mut locked = self.inner.write().unwrap();
1275 : let mut safekeepers = (*locked.safekeepers).clone();
1276 : for (id, state) in deltas.0 {
1277 : let Some(sk) = safekeepers.get_mut(&id) else {
1278 : tracing::info!(
1279 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1280 : );
1281 : continue;
1282 : };
1283 : sk.set_availability(state);
1284 : }
1285 : locked.safekeepers = Arc::new(safekeepers);
1286 : }
1287 : }
1288 : }
1289 :
1290 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1291 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1292 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1293 : /// will indicate that reconciliation is not needed.
1294 : #[instrument(skip_all, fields(
1295 : seq=%result.sequence,
1296 : tenant_id=%result.tenant_shard_id.tenant_id,
1297 : shard_id=%result.tenant_shard_id.shard_slug(),
1298 : ))]
1299 : fn process_result(&self, result: ReconcileResult) {
1300 : let mut locked = self.inner.write().unwrap();
1301 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1302 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1303 : // A reconciliation result might race with removing a tenant: drop results for
1304 : // tenants that aren't in our map.
1305 : return;
1306 : };
1307 :
1308 : // Usually generation should only be updated via this path, so the max() isn't
1309 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1310 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1311 :
1312 : // If the reconciler signals that it failed to notify compute, set this state on
1313 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1314 : tenant.pending_compute_notification = result.pending_compute_notification;
1315 :
1316 : // Let the TenantShard know it is idle.
1317 : tenant.reconcile_complete(result.sequence);
1318 :
1319 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1320 : // make to the tenant
1321 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1322 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1323 : // make to the tenant
1324 0 : let node = nodes.get(delta.node_id())?;
1325 :
1326 0 : if node.is_available() {
1327 0 : return Some(delta);
1328 0 : }
1329 0 :
1330 0 : // In case a node became unavailable concurrently with the reconcile, observed
1331 0 : // locations on it are now uncertain. By convention, set them to None in order
1332 0 : // for them to get refreshed when the node comes back online.
1333 0 : Some(ObservedStateDelta::Upsert(Box::new((
1334 0 : node.get_id(),
1335 0 : ObservedStateLocation { conf: None },
1336 0 : ))))
1337 0 : });
1338 :
1339 : match result.result {
1340 : Ok(()) => {
1341 : tenant.apply_observed_deltas(deltas);
1342 : tenant.waiter.advance(result.sequence);
1343 : }
1344 : Err(e) => {
1345 : match e {
1346 : ReconcileError::Cancel => {
1347 : tracing::info!("Reconciler was cancelled");
1348 : }
1349 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1350 : // This might be due to the reconciler getting cancelled, or it might
1351 : // be due to the `Node` being marked offline.
1352 : tracing::info!("Reconciler cancelled during pageserver API call");
1353 : }
1354 : _ => {
1355 : tracing::warn!("Reconcile error: {}", e);
1356 : }
1357 : }
1358 :
1359 : // Ordering: populate last_error before advancing error_seq,
1360 : // so that waiters will see the correct error after waiting.
1361 : tenant.set_last_error(result.sequence, e);
1362 :
1363 : // Skip deletions on reconcile failures
1364 : let upsert_deltas =
1365 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1366 : tenant.apply_observed_deltas(upsert_deltas);
1367 : }
1368 : }
1369 :
1370 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1371 : if tenant.policy == PlacementPolicy::Detached {
1372 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1373 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1374 : // duration of the request.
1375 : let guard = self.tenant_op_locks.try_exclusive(
1376 : tenant.tenant_shard_id.tenant_id,
1377 : TenantOperations::DropDetached,
1378 : );
1379 : if let Some(guard) = guard {
1380 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1381 : }
1382 : }
1383 :
1384 : // Maybe some other work can proceed now that this job finished.
1385 : //
1386 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1387 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1388 : if self.reconciler_concurrency.available_permits() > 0 {
1389 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1390 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1391 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1392 : shard.delayed_reconcile = false;
1393 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1394 : }
1395 :
1396 : if self.reconciler_concurrency.available_permits() == 0 {
1397 : break;
1398 : }
1399 : }
1400 : }
1401 : }
1402 :
1403 0 : async fn process_results(
1404 0 : &self,
1405 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1406 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1407 0 : Result<(), (TenantShardId, NotifyError)>,
1408 0 : >,
1409 0 : ) {
1410 : loop {
1411 : // Wait for the next result, or for cancellation
1412 0 : tokio::select! {
1413 0 : r = result_rx.recv() => {
1414 0 : match r {
1415 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1416 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1417 : }
1418 : }
1419 0 : _ = async{
1420 0 : match bg_compute_hook_result_rx.recv().await {
1421 0 : Some(result) => {
1422 0 : if let Err((tenant_shard_id, notify_error)) = result {
1423 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1424 0 : let mut locked = self.inner.write().unwrap();
1425 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1426 0 : shard.pending_compute_notification = true;
1427 0 : }
1428 :
1429 0 : }
1430 : },
1431 : None => {
1432 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1433 0 : self.cancel.cancelled().await;
1434 : }
1435 : }
1436 0 : } => {},
1437 0 : _ = self.cancel.cancelled() => {
1438 0 : break;
1439 : }
1440 : };
1441 : }
1442 0 : }
1443 :
1444 0 : async fn process_aborts(
1445 0 : &self,
1446 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1447 0 : ) {
1448 : loop {
1449 : // Wait for the next result, or for cancellation
1450 0 : let op = tokio::select! {
1451 0 : r = abort_rx.recv() => {
1452 0 : match r {
1453 0 : Some(op) => {op},
1454 0 : None => {break;}
1455 : }
1456 : }
1457 0 : _ = self.cancel.cancelled() => {
1458 0 : break;
1459 : }
1460 : };
1461 :
1462 : // Retry until shutdown: we must keep this request object alive until it is properly
1463 : // processed, as it holds a lock guard that prevents other operations trying to do things
1464 : // to the tenant while it is in a weird part-split state.
1465 0 : while !self.reconcilers_cancel.is_cancelled() {
1466 0 : match self.abort_tenant_shard_split(&op).await {
1467 0 : Ok(_) => break,
1468 0 : Err(e) => {
1469 0 : tracing::warn!(
1470 0 : "Failed to abort shard split on {}, will retry: {e}",
1471 : op.tenant_id
1472 : );
1473 :
1474 : // If a node is unavailable, we hope that it has been properly marked Offline
1475 : // when we retry, so that the abort op will succeed. If the abort op is failing
1476 : // for some other reason, we will keep retrying forever, or until a human notices
1477 : // and does something about it (either fixing a pageserver or restarting the controller).
1478 0 : tokio::time::timeout(
1479 0 : Duration::from_secs(5),
1480 0 : self.reconcilers_cancel.cancelled(),
1481 0 : )
1482 0 : .await
1483 0 : .ok();
1484 : }
1485 : }
1486 : }
1487 : }
1488 0 : }
1489 :
1490 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1491 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1492 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1493 0 :
1494 0 : let leadership_cancel = CancellationToken::new();
1495 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1496 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1497 :
1498 : // Apply the migrations **after** the current leader has stepped down
1499 : // (or we've given up waiting for it), but **before** reading from the
1500 : // database. The only exception is reading the current leader before
1501 : // migrating.
1502 0 : persistence.migration_run().await?;
1503 :
1504 0 : tracing::info!("Loading nodes from database...");
1505 0 : let nodes = persistence
1506 0 : .list_nodes()
1507 0 : .await?
1508 0 : .into_iter()
1509 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1510 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1511 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1512 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1513 0 : metrics::METRICS_REGISTRY
1514 0 : .metrics_group
1515 0 : .storage_controller_pageserver_nodes
1516 0 : .set(nodes.len() as i64);
1517 0 : metrics::METRICS_REGISTRY
1518 0 : .metrics_group
1519 0 : .storage_controller_https_pageserver_nodes
1520 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
1521 0 :
1522 0 : tracing::info!("Loading safekeepers from database...");
1523 0 : let safekeepers = persistence
1524 0 : .list_safekeepers()
1525 0 : .await?
1526 0 : .into_iter()
1527 0 : .map(|skp| {
1528 0 : Safekeeper::from_persistence(
1529 0 : skp,
1530 0 : CancellationToken::new(),
1531 0 : config.use_https_safekeeper_api,
1532 0 : )
1533 0 : })
1534 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1535 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1536 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1537 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1538 0 : metrics::METRICS_REGISTRY
1539 0 : .metrics_group
1540 0 : .storage_controller_safekeeper_nodes
1541 0 : .set(safekeepers.len() as i64);
1542 0 : metrics::METRICS_REGISTRY
1543 0 : .metrics_group
1544 0 : .storage_controller_https_safekeeper_nodes
1545 0 : .set(safekeepers.values().filter(|s| s.has_https_port()).count() as i64);
1546 0 :
1547 0 : tracing::info!("Loading shards from database...");
1548 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1549 0 : tracing::info!(
1550 0 : "Loaded {} shards from database.",
1551 0 : tenant_shard_persistence.len()
1552 : );
1553 :
1554 : // If any shard splits were in progress, reset the database state to abort them
1555 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1556 0 : HashMap::new();
1557 0 : for tsp in &mut tenant_shard_persistence {
1558 0 : let shard = tsp.get_shard_identity()?;
1559 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1560 0 : let entry = tenant_shard_count_min_max
1561 0 : .entry(tenant_shard_id.tenant_id)
1562 0 : .or_insert_with(|| (shard.count, shard.count));
1563 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1564 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1565 0 : }
1566 :
1567 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1568 0 : if count_min != count_max {
1569 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1570 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1571 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1572 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1573 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1574 :
1575 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1576 : // identified this tenant has having mismatching min/max counts.
1577 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1578 :
1579 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1580 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1581 0 : // Set idle split state on those shards that we will retain.
1582 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1583 0 : if tsp_tenant_id == tenant_id
1584 0 : && tsp.get_shard_identity().unwrap().count == count_min
1585 0 : {
1586 0 : tsp.splitting = SplitState::Idle;
1587 0 : } else if tsp_tenant_id == tenant_id {
1588 : // Leave the splitting state on the child shards: this will be used next to
1589 : // drop them.
1590 0 : tracing::info!(
1591 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1592 : );
1593 0 : }
1594 0 : });
1595 0 :
1596 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1597 0 : tenant_shard_persistence.retain(|tsp| {
1598 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1599 0 : || tsp.splitting == SplitState::Idle
1600 0 : });
1601 0 : }
1602 : }
1603 :
1604 0 : let mut tenants = BTreeMap::new();
1605 0 :
1606 0 : let mut scheduler = Scheduler::new(nodes.values());
1607 :
1608 : #[cfg(feature = "testing")]
1609 : {
1610 : use pageserver_api::controller_api::AvailabilityZone;
1611 :
1612 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1613 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1614 : // after when pageservers start up and register.
1615 0 : let mut node_ids = HashSet::new();
1616 0 : for tsp in &tenant_shard_persistence {
1617 0 : if let Some(node_id) = tsp.generation_pageserver {
1618 0 : node_ids.insert(node_id);
1619 0 : }
1620 : }
1621 0 : for node_id in node_ids {
1622 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1623 0 : let node = Node::new(
1624 0 : NodeId(node_id as u64),
1625 0 : "".to_string(),
1626 0 : 123,
1627 0 : None,
1628 0 : "".to_string(),
1629 0 : 123,
1630 0 : AvailabilityZone("test_az".to_string()),
1631 0 : false,
1632 0 : )
1633 0 : .unwrap();
1634 0 :
1635 0 : scheduler.node_upsert(&node);
1636 : }
1637 : }
1638 0 : for tsp in tenant_shard_persistence {
1639 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1640 :
1641 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1642 : // it with what we can infer: the node for which a generation was most recently issued.
1643 0 : let mut intent = IntentState::new(
1644 0 : tsp.preferred_az_id
1645 0 : .as_ref()
1646 0 : .map(|az| AvailabilityZone(az.clone())),
1647 0 : );
1648 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1649 : {
1650 0 : if nodes.contains_key(&generation_pageserver) {
1651 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1652 0 : } else {
1653 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1654 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1655 : // on different pageservers.
1656 0 : tracing::warn!(
1657 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1658 : );
1659 : }
1660 0 : }
1661 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1662 :
1663 0 : tenants.insert(tenant_shard_id, new_tenant);
1664 : }
1665 :
1666 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1667 0 :
1668 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1669 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1670 0 : tokio::sync::mpsc::channel(512);
1671 0 :
1672 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1673 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1674 0 :
1675 0 : let cancel = CancellationToken::new();
1676 0 : let reconcilers_cancel = cancel.child_token();
1677 0 :
1678 0 : let mut http_client = reqwest::Client::builder();
1679 0 : // We intentionally disable the connection pool, so every request will create its own TCP connection.
1680 0 : // It's especially important for heartbeaters to notice more network problems.
1681 0 : //
1682 0 : // TODO: It makes sense to use this client only in heartbeaters and create a second one with
1683 0 : // connection pooling for everything else. But reqwest::Client may create a connection without
1684 0 : // ever using it (it uses hyper's Client under the hood):
1685 0 : // https://github.com/hyperium/hyper-util/blob/d51318df3461d40e5f5e5ca163cb3905ac960209/src/client/legacy/client.rs#L415
1686 0 : //
1687 0 : // Because of a bug in hyper0::Connection::graceful_shutdown such connections hang during
1688 0 : // graceful server shutdown: https://github.com/hyperium/hyper/issues/2730
1689 0 : //
1690 0 : // The bug has been fixed in hyper v1, so keep alive may be enabled only after we migrate to hyper1.
1691 0 : http_client = http_client.pool_max_idle_per_host(0);
1692 0 : for ssl_ca_cert in &config.ssl_ca_certs {
1693 0 : http_client = http_client.add_root_certificate(ssl_ca_cert.clone());
1694 0 : }
1695 0 : let http_client = http_client.build()?;
1696 :
1697 0 : let heartbeater_ps = Heartbeater::new(
1698 0 : http_client.clone(),
1699 0 : config.pageserver_jwt_token.clone(),
1700 0 : config.max_offline_interval,
1701 0 : config.max_warming_up_interval,
1702 0 : cancel.clone(),
1703 0 : );
1704 0 :
1705 0 : let heartbeater_sk = Heartbeater::new(
1706 0 : http_client.clone(),
1707 0 : config.safekeeper_jwt_token.clone(),
1708 0 : config.max_offline_interval,
1709 0 : config.max_warming_up_interval,
1710 0 : cancel.clone(),
1711 0 : );
1712 :
1713 0 : let initial_leadership_status = if config.start_as_candidate {
1714 0 : LeadershipStatus::Candidate
1715 : } else {
1716 0 : LeadershipStatus::Leader
1717 : };
1718 :
1719 0 : let this = Arc::new(Self {
1720 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1721 0 : nodes,
1722 0 : safekeepers,
1723 0 : tenants,
1724 0 : scheduler,
1725 0 : delayed_reconcile_rx,
1726 0 : initial_leadership_status,
1727 0 : reconcilers_cancel.clone(),
1728 0 : ))),
1729 0 : config: config.clone(),
1730 0 : persistence,
1731 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())?),
1732 0 : result_tx,
1733 0 : heartbeater_ps,
1734 0 : heartbeater_sk,
1735 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1736 0 : config.reconciler_concurrency,
1737 0 : )),
1738 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1739 0 : config.priority_reconciler_concurrency,
1740 0 : )),
1741 0 : delayed_reconcile_tx,
1742 0 : abort_tx,
1743 0 : startup_complete: startup_complete.clone(),
1744 0 : cancel,
1745 0 : reconcilers_cancel,
1746 0 : gate: Gate::default(),
1747 0 : reconcilers_gate: Gate::default(),
1748 0 : tenant_op_locks: Default::default(),
1749 0 : node_op_locks: Default::default(),
1750 0 : http_client,
1751 0 : });
1752 0 :
1753 0 : let result_task_this = this.clone();
1754 0 : tokio::task::spawn(async move {
1755 : // Block shutdown until we're done (we must respect self.cancel)
1756 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1757 0 : result_task_this
1758 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1759 0 : .await
1760 0 : }
1761 0 : });
1762 0 :
1763 0 : tokio::task::spawn({
1764 0 : let this = this.clone();
1765 0 : async move {
1766 : // Block shutdown until we're done (we must respect self.cancel)
1767 0 : if let Ok(_gate) = this.gate.enter() {
1768 0 : this.process_aborts(abort_rx).await
1769 0 : }
1770 0 : }
1771 0 : });
1772 0 :
1773 0 : tokio::task::spawn({
1774 0 : let this = this.clone();
1775 0 : async move {
1776 0 : if let Ok(_gate) = this.gate.enter() {
1777 : loop {
1778 0 : tokio::select! {
1779 0 : _ = this.cancel.cancelled() => {
1780 0 : break;
1781 : },
1782 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1783 0 : };
1784 0 : this.tenant_op_locks.housekeeping();
1785 : }
1786 0 : }
1787 0 : }
1788 0 : });
1789 0 :
1790 0 : tokio::task::spawn({
1791 0 : let this = this.clone();
1792 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1793 0 : // is done.
1794 0 : let startup_completion = startup_completion.clone();
1795 0 : async move {
1796 : // Block shutdown until we're done (we must respect self.cancel)
1797 0 : let Ok(_gate) = this.gate.enter() else {
1798 0 : return;
1799 : };
1800 :
1801 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1802 0 : .await;
1803 :
1804 0 : drop(startup_completion);
1805 0 : }
1806 0 : });
1807 0 :
1808 0 : tokio::task::spawn({
1809 0 : let this = this.clone();
1810 0 : let startup_complete = startup_complete.clone();
1811 0 : async move {
1812 0 : startup_complete.wait().await;
1813 0 : this.background_reconcile().await;
1814 0 : }
1815 0 : });
1816 0 :
1817 0 : tokio::task::spawn({
1818 0 : let this = this.clone();
1819 0 : let startup_complete = startup_complete.clone();
1820 0 : async move {
1821 0 : startup_complete.wait().await;
1822 0 : this.spawn_heartbeat_driver().await;
1823 0 : }
1824 0 : });
1825 0 :
1826 0 : Ok(this)
1827 0 : }
1828 :
1829 0 : pub(crate) async fn attach_hook(
1830 0 : &self,
1831 0 : attach_req: AttachHookRequest,
1832 0 : ) -> anyhow::Result<AttachHookResponse> {
1833 0 : let _tenant_lock = trace_exclusive_lock(
1834 0 : &self.tenant_op_locks,
1835 0 : attach_req.tenant_shard_id.tenant_id,
1836 0 : TenantOperations::AttachHook,
1837 0 : )
1838 0 : .await;
1839 :
1840 : // This is a test hook. To enable using it on tenants that were created directly with
1841 : // the pageserver API (not via this service), we will auto-create any missing tenant
1842 : // shards with default state.
1843 0 : let insert = {
1844 0 : match self
1845 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1846 0 : .await
1847 : {
1848 0 : Ok(_) => false,
1849 0 : Err(ApiError::NotFound(_)) => true,
1850 0 : Err(e) => return Err(e.into()),
1851 : }
1852 : };
1853 :
1854 0 : if insert {
1855 0 : let config = attach_req.config.clone().unwrap_or_default();
1856 0 : let tsp = TenantShardPersistence {
1857 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1858 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1859 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1860 0 : shard_stripe_size: 0,
1861 0 : generation: attach_req.generation_override.or(Some(0)),
1862 0 : generation_pageserver: None,
1863 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1864 0 : config: serde_json::to_string(&config).unwrap(),
1865 0 : splitting: SplitState::default(),
1866 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1867 0 : .unwrap(),
1868 0 : preferred_az_id: None,
1869 0 : };
1870 0 :
1871 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1872 0 : Err(e) => match e {
1873 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1874 : DatabaseErrorKind::UniqueViolation,
1875 : _,
1876 : )) => {
1877 0 : tracing::info!(
1878 0 : "Raced with another request to insert tenant {}",
1879 : attach_req.tenant_shard_id
1880 : )
1881 : }
1882 0 : _ => return Err(e.into()),
1883 : },
1884 : Ok(()) => {
1885 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1886 :
1887 0 : let mut shard = TenantShard::new(
1888 0 : attach_req.tenant_shard_id,
1889 0 : ShardIdentity::unsharded(),
1890 0 : PlacementPolicy::Attached(0),
1891 0 : None,
1892 0 : );
1893 0 : shard.config = config;
1894 0 :
1895 0 : let mut locked = self.inner.write().unwrap();
1896 0 : locked.tenants.insert(attach_req.tenant_shard_id, shard);
1897 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1898 : }
1899 : }
1900 0 : }
1901 :
1902 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1903 0 : let maybe_tenant_conf = {
1904 0 : let locked = self.inner.write().unwrap();
1905 0 : locked
1906 0 : .tenants
1907 0 : .get(&attach_req.tenant_shard_id)
1908 0 : .map(|t| t.config.clone())
1909 0 : };
1910 0 :
1911 0 : match maybe_tenant_conf {
1912 0 : Some(conf) => {
1913 0 : let new_generation = self
1914 0 : .persistence
1915 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1916 0 : .await?;
1917 :
1918 : // Persist the placement policy update. This is required
1919 : // when we reattaching a detached tenant.
1920 0 : self.persistence
1921 0 : .update_tenant_shard(
1922 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1923 0 : Some(PlacementPolicy::Attached(0)),
1924 0 : Some(conf),
1925 0 : None,
1926 0 : None,
1927 0 : )
1928 0 : .await?;
1929 0 : Some(new_generation)
1930 : }
1931 : None => {
1932 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1933 : }
1934 : }
1935 : } else {
1936 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1937 0 : None
1938 : };
1939 :
1940 0 : let mut locked = self.inner.write().unwrap();
1941 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1942 0 :
1943 0 : let tenant_shard = tenants
1944 0 : .get_mut(&attach_req.tenant_shard_id)
1945 0 : .expect("Checked for existence above");
1946 :
1947 0 : if let Some(new_generation) = new_generation {
1948 0 : tenant_shard.generation = Some(new_generation);
1949 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1950 0 : } else {
1951 : // This is a detach notification. We must update placement policy to avoid re-attaching
1952 : // during background scheduling/reconciliation, or during storage controller restart.
1953 0 : assert!(attach_req.node_id.is_none());
1954 0 : tenant_shard.policy = PlacementPolicy::Detached;
1955 : }
1956 :
1957 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1958 0 : tracing::info!(
1959 : tenant_id = %attach_req.tenant_shard_id,
1960 : ps_id = %attaching_pageserver,
1961 : generation = ?tenant_shard.generation,
1962 0 : "issuing",
1963 : );
1964 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1965 0 : tracing::info!(
1966 : tenant_id = %attach_req.tenant_shard_id,
1967 : %ps_id,
1968 : generation = ?tenant_shard.generation,
1969 0 : "dropping",
1970 : );
1971 : } else {
1972 0 : tracing::info!(
1973 : tenant_id = %attach_req.tenant_shard_id,
1974 0 : "no-op: tenant already has no pageserver");
1975 : }
1976 0 : tenant_shard
1977 0 : .intent
1978 0 : .set_attached(scheduler, attach_req.node_id);
1979 0 :
1980 0 : tracing::info!(
1981 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}, config {:?}",
1982 0 : attach_req.tenant_shard_id,
1983 0 : tenant_shard.generation,
1984 0 : // TODO: this is an odd number of 0xf's
1985 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff)),
1986 : attach_req.config,
1987 : );
1988 :
1989 : // Trick the reconciler into not doing anything for this tenant: this helps
1990 : // tests that manually configure a tenant on the pagesrever, and then call this
1991 : // attach hook: they don't want background reconciliation to modify what they
1992 : // did to the pageserver.
1993 : #[cfg(feature = "testing")]
1994 : {
1995 0 : if let Some(node_id) = attach_req.node_id {
1996 0 : tenant_shard.observed.locations = HashMap::from([(
1997 0 : node_id,
1998 0 : ObservedStateLocation {
1999 0 : conf: Some(attached_location_conf(
2000 0 : tenant_shard.generation.unwrap(),
2001 0 : &tenant_shard.shard,
2002 0 : &tenant_shard.config,
2003 0 : &PlacementPolicy::Attached(0),
2004 0 : )),
2005 0 : },
2006 0 : )]);
2007 0 : } else {
2008 0 : tenant_shard.observed.locations.clear();
2009 0 : }
2010 : }
2011 :
2012 0 : Ok(AttachHookResponse {
2013 0 : generation: attach_req
2014 0 : .node_id
2015 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
2016 0 : })
2017 0 : }
2018 :
2019 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
2020 0 : let locked = self.inner.read().unwrap();
2021 0 :
2022 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
2023 0 :
2024 0 : InspectResponse {
2025 0 : attachment: tenant_shard.and_then(|s| {
2026 0 : s.intent
2027 0 : .get_attached()
2028 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
2029 0 : }),
2030 0 : }
2031 0 : }
2032 :
2033 : // When the availability state of a node transitions to active, we must do a full reconciliation
2034 : // of LocationConfigs on that node. This is because while a node was offline:
2035 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
2036 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
2037 : //
2038 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
2039 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
2040 : // this function.
2041 : //
2042 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
2043 : // for written for a single node rather than as a batch job for all nodes.
2044 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
2045 : async fn node_activate_reconcile(
2046 : &self,
2047 : mut node: Node,
2048 : _lock: &TracingExclusiveGuard<NodeOperations>,
2049 : ) -> Result<(), ApiError> {
2050 : // This Node is a mutable local copy: we will set it active so that we can use its
2051 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
2052 : // later.
2053 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
2054 :
2055 : let configs = match node
2056 : .with_client_retries(
2057 0 : |client| async move { client.list_location_config().await },
2058 : &self.http_client,
2059 : &self.config.pageserver_jwt_token,
2060 : 1,
2061 : 5,
2062 : SHORT_RECONCILE_TIMEOUT,
2063 : &self.cancel,
2064 : )
2065 : .await
2066 : {
2067 : None => {
2068 : // We're shutting down (the Node's cancellation token can't have fired, because
2069 : // we're the only scope that has a reference to it, and we didn't fire it).
2070 : return Err(ApiError::ShuttingDown);
2071 : }
2072 : Some(Err(e)) => {
2073 : // This node didn't succeed listing its locations: it may not proceed to active state
2074 : // as it is apparently unavailable.
2075 : return Err(ApiError::PreconditionFailed(
2076 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2077 : ));
2078 : }
2079 : Some(Ok(configs)) => configs,
2080 : };
2081 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2082 :
2083 : let mut cleanup = Vec::new();
2084 : let mut mismatched_locations = 0;
2085 : {
2086 : let mut locked = self.inner.write().unwrap();
2087 :
2088 : for (tenant_shard_id, reported) in configs.tenant_shards {
2089 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2090 : cleanup.push(tenant_shard_id);
2091 : continue;
2092 : };
2093 :
2094 : let on_record = &mut tenant_shard
2095 : .observed
2096 : .locations
2097 : .entry(node.get_id())
2098 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2099 : .conf;
2100 :
2101 : // If the location reported by the node does not match our observed state,
2102 : // then we mark it as uncertain and let the background reconciliation loop
2103 : // deal with it.
2104 : //
2105 : // Note that this also covers net new locations reported by the node.
2106 : if *on_record != reported {
2107 : mismatched_locations += 1;
2108 : *on_record = None;
2109 : }
2110 : }
2111 : }
2112 :
2113 : if mismatched_locations > 0 {
2114 : tracing::info!(
2115 : "Set observed state to None for {mismatched_locations} mismatched locations"
2116 : );
2117 : }
2118 :
2119 : for tenant_shard_id in cleanup {
2120 : tracing::info!("Detaching {tenant_shard_id}");
2121 : match node
2122 : .with_client_retries(
2123 0 : |client| async move {
2124 0 : let config = LocationConfig {
2125 0 : mode: LocationConfigMode::Detached,
2126 0 : generation: None,
2127 0 : secondary_conf: None,
2128 0 : shard_number: tenant_shard_id.shard_number.0,
2129 0 : shard_count: tenant_shard_id.shard_count.literal(),
2130 0 : shard_stripe_size: 0,
2131 0 : tenant_conf: models::TenantConfig::default(),
2132 0 : };
2133 0 : client
2134 0 : .location_config(tenant_shard_id, config, None, false)
2135 0 : .await
2136 0 : },
2137 : &self.http_client,
2138 : &self.config.pageserver_jwt_token,
2139 : 1,
2140 : 5,
2141 : SHORT_RECONCILE_TIMEOUT,
2142 : &self.cancel,
2143 : )
2144 : .await
2145 : {
2146 : None => {
2147 : // We're shutting down (the Node's cancellation token can't have fired, because
2148 : // we're the only scope that has a reference to it, and we didn't fire it).
2149 : return Err(ApiError::ShuttingDown);
2150 : }
2151 : Some(Err(e)) => {
2152 : // Do not let the node proceed to Active state if it is not responsive to requests
2153 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2154 : // detach completing: we should not let this node back into the set of nodes considered
2155 : // okay for scheduling.
2156 : return Err(ApiError::Conflict(format!(
2157 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2158 : )));
2159 : }
2160 : Some(Ok(_)) => {}
2161 : };
2162 : }
2163 :
2164 : Ok(())
2165 : }
2166 :
2167 0 : pub(crate) async fn re_attach(
2168 0 : &self,
2169 0 : reattach_req: ReAttachRequest,
2170 0 : ) -> Result<ReAttachResponse, ApiError> {
2171 0 : if let Some(register_req) = reattach_req.register {
2172 0 : self.node_register(register_req).await?;
2173 0 : }
2174 :
2175 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2176 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2177 :
2178 0 : tracing::info!(
2179 : node_id=%reattach_req.node_id,
2180 0 : "Incremented {} tenant shards' generations",
2181 0 : incremented_generations.len()
2182 : );
2183 :
2184 : // Apply the updated generation to our in-memory state, and
2185 : // gather discover secondary locations.
2186 0 : let mut locked = self.inner.write().unwrap();
2187 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2188 0 :
2189 0 : let mut response = ReAttachResponse {
2190 0 : tenants: Vec::new(),
2191 0 : };
2192 :
2193 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2194 : // to call location_conf API with an old generation. Wait for cancellation to complete
2195 : // before responding to this request. Requires well implemented CancellationToken logic
2196 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2197 : // request in flight over the network: TODO handle that by making location_conf API refuse
2198 : // to go backward in generations.
2199 :
2200 : // Scan through all shards, applying updates for ones where we updated generation
2201 : // and identifying shards that intend to have a secondary location on this node.
2202 0 : for (tenant_shard_id, shard) in tenants {
2203 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2204 0 : let new_gen = *new_gen;
2205 0 : response.tenants.push(ReAttachResponseTenant {
2206 0 : id: *tenant_shard_id,
2207 0 : r#gen: Some(new_gen.into().unwrap()),
2208 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2209 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2210 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2211 0 : // the stale/multi states at this point.
2212 0 : mode: LocationConfigMode::AttachedSingle,
2213 0 : });
2214 0 :
2215 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2216 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2217 : // Why can we update `observed` even though we're not sure our response will be received
2218 : // by the pageserver? Because the pageserver will not proceed with startup until
2219 : // it has processed response: if it loses it, we'll see another request and increment
2220 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2221 0 : if let Some(conf) = observed.conf.as_mut() {
2222 0 : conf.generation = new_gen.into();
2223 0 : }
2224 0 : } else {
2225 0 : // This node has no observed state for the shard: perhaps it was offline
2226 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2227 0 : // will be prompted to learn the location's state before it makes changes.
2228 0 : shard
2229 0 : .observed
2230 0 : .locations
2231 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2232 0 : }
2233 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2234 0 : // Ordering: pageserver will not accept /location_config requests until it has
2235 0 : // finished processing the response from re-attach. So we can update our in-memory state
2236 0 : // now, and be confident that we are not stamping on the result of some later location config.
2237 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2238 0 : // so we might update observed state here, and then get over-written by some racing
2239 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2240 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2241 0 :
2242 0 : response.tenants.push(ReAttachResponseTenant {
2243 0 : id: *tenant_shard_id,
2244 0 : r#gen: None,
2245 0 : mode: LocationConfigMode::Secondary,
2246 0 : });
2247 0 :
2248 0 : // We must not update observed, because we have no guarantee that our
2249 0 : // response will be received by the pageserver. This could leave it
2250 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2251 0 : }
2252 : }
2253 :
2254 : // We consider a node Active once we have composed a re-attach response, but we
2255 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2256 : // implicitly synchronizes the LocationConfigs on the node.
2257 : //
2258 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2259 : // but those requests will not be accepted by the node until it has finished processing
2260 : // the re-attach response.
2261 : //
2262 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2263 : // in [`Persistence::re_attach`].
2264 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2265 0 : let reset_scheduling = matches!(
2266 0 : node.get_scheduling(),
2267 : NodeSchedulingPolicy::PauseForRestart
2268 : | NodeSchedulingPolicy::Draining
2269 : | NodeSchedulingPolicy::Filling
2270 : );
2271 :
2272 0 : let mut new_nodes = (**nodes).clone();
2273 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2274 0 : if reset_scheduling {
2275 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2276 0 : }
2277 :
2278 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2279 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2280 0 :
2281 0 : scheduler.node_upsert(node);
2282 0 : let new_nodes = Arc::new(new_nodes);
2283 0 : *nodes = new_nodes;
2284 : } else {
2285 0 : tracing::error!(
2286 0 : "Reattaching node {} was removed while processing the request",
2287 : reattach_req.node_id
2288 : );
2289 : }
2290 0 : }
2291 :
2292 0 : Ok(response)
2293 0 : }
2294 :
2295 0 : pub(crate) async fn validate(
2296 0 : &self,
2297 0 : validate_req: ValidateRequest,
2298 0 : ) -> Result<ValidateResponse, DatabaseError> {
2299 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2300 : // in-memory generation for a shard
2301 0 : let in_memory_result = {
2302 0 : let mut in_memory_result = Vec::new();
2303 0 : let locked = self.inner.read().unwrap();
2304 0 : for req_tenant in validate_req.tenants {
2305 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2306 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2307 0 : tracing::info!(
2308 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2309 : req_tenant.id,
2310 : req_tenant.r#gen,
2311 : tenant_shard.generation
2312 : );
2313 :
2314 0 : in_memory_result.push((
2315 0 : req_tenant.id,
2316 0 : Generation::new(req_tenant.r#gen),
2317 0 : valid,
2318 0 : ));
2319 : } else {
2320 : // This is legal: for example during a shard split the pageserver may still
2321 : // have deletions in its queue from the old pre-split shard, or after deletion
2322 : // of a tenant that was busy with compaction/gc while being deleted.
2323 0 : tracing::info!(
2324 0 : "Refusing deletion validation for missing shard {}",
2325 : req_tenant.id
2326 : );
2327 : }
2328 : }
2329 :
2330 0 : in_memory_result
2331 : };
2332 :
2333 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2334 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2335 0 : let db_generations = self
2336 0 : .persistence
2337 0 : .shard_generations(
2338 0 : in_memory_result
2339 0 : .iter()
2340 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2341 0 : )
2342 0 : .await?;
2343 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2344 0 :
2345 0 : let mut response = ValidateResponse {
2346 0 : tenants: Vec::new(),
2347 0 : };
2348 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2349 0 : let valid = if valid {
2350 0 : let db_generation = db_generations.get(&tenant_shard_id);
2351 0 : db_generation == Some(&Some(validate_generation))
2352 : } else {
2353 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2354 : // this prevents a pageserver from cleaning up an object in S3.
2355 0 : false
2356 : };
2357 :
2358 0 : response.tenants.push(ValidateResponseTenant {
2359 0 : id: tenant_shard_id,
2360 0 : valid,
2361 0 : })
2362 : }
2363 :
2364 0 : Ok(response)
2365 0 : }
2366 :
2367 0 : pub(crate) async fn tenant_create(
2368 0 : &self,
2369 0 : create_req: TenantCreateRequest,
2370 0 : ) -> Result<TenantCreateResponse, ApiError> {
2371 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2372 :
2373 : // Exclude any concurrent attempts to create/access the same tenant ID
2374 0 : let _tenant_lock = trace_exclusive_lock(
2375 0 : &self.tenant_op_locks,
2376 0 : create_req.new_tenant_id.tenant_id,
2377 0 : TenantOperations::Create,
2378 0 : )
2379 0 : .await;
2380 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2381 :
2382 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2383 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2384 : // accept compute notifications while it is in the process of creating. Reconciliation will
2385 : // be retried in the background.
2386 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2387 0 : }
2388 0 : Ok(response)
2389 0 : }
2390 :
2391 0 : pub(crate) async fn do_tenant_create(
2392 0 : &self,
2393 0 : create_req: TenantCreateRequest,
2394 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2395 0 : let placement_policy = create_req
2396 0 : .placement_policy
2397 0 : .clone()
2398 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2399 0 : .unwrap_or(PlacementPolicy::Attached(0));
2400 :
2401 : // This service expects to handle sharding itself: it is an error to try and directly create
2402 : // a particular shard here.
2403 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2404 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2405 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2406 0 : )));
2407 : } else {
2408 0 : create_req.new_tenant_id.tenant_id
2409 0 : };
2410 0 :
2411 0 : tracing::info!(
2412 0 : "Creating tenant {}, shard_count={:?}",
2413 : create_req.new_tenant_id,
2414 : create_req.shard_parameters.count,
2415 : );
2416 :
2417 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2418 0 : .map(|i| TenantShardId {
2419 0 : tenant_id,
2420 0 : shard_number: ShardNumber(i),
2421 0 : shard_count: create_req.shard_parameters.count,
2422 0 : })
2423 0 : .collect::<Vec<_>>();
2424 :
2425 : // If the caller specifies a None generation, it means "start from default". This is different
2426 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2427 : // an incompletely-onboarded tenant.
2428 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2429 0 : tracing::info!(
2430 0 : "tenant_create: secondary mode, generation is_some={}",
2431 0 : create_req.generation.is_some()
2432 : );
2433 0 : create_req.generation.map(Generation::new)
2434 : } else {
2435 0 : tracing::info!(
2436 0 : "tenant_create: not secondary mode, generation is_some={}",
2437 0 : create_req.generation.is_some()
2438 : );
2439 0 : Some(
2440 0 : create_req
2441 0 : .generation
2442 0 : .map(Generation::new)
2443 0 : .unwrap_or(INITIAL_GENERATION),
2444 0 : )
2445 : };
2446 :
2447 0 : let preferred_az_id = {
2448 0 : let locked = self.inner.read().unwrap();
2449 : // Idempotency: take the existing value if the tenant already exists
2450 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2451 0 : shard.preferred_az().cloned()
2452 : } else {
2453 0 : locked.scheduler.get_az_for_new_tenant()
2454 : }
2455 : };
2456 :
2457 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2458 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2459 : // during the creation, rather than risking leaving orphan objects in S3.
2460 0 : let persist_tenant_shards = create_ids
2461 0 : .iter()
2462 0 : .map(|tenant_shard_id| TenantShardPersistence {
2463 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2464 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2465 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2466 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2467 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2468 0 : // The pageserver is not known until scheduling happens: we will set this column when
2469 0 : // incrementing the generation the first time we attach to a pageserver.
2470 0 : generation_pageserver: None,
2471 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2472 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2473 0 : splitting: SplitState::default(),
2474 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2475 0 : .unwrap(),
2476 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2477 0 : })
2478 0 : .collect();
2479 0 :
2480 0 : match self
2481 0 : .persistence
2482 0 : .insert_tenant_shards(persist_tenant_shards)
2483 0 : .await
2484 : {
2485 0 : Ok(_) => {}
2486 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2487 : DatabaseErrorKind::UniqueViolation,
2488 : _,
2489 : ))) => {
2490 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2491 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2492 : // creation's shard count.
2493 0 : tracing::info!(
2494 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2495 : );
2496 : }
2497 : // Any other database error is unexpected and a bug.
2498 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2499 : };
2500 :
2501 0 : let mut schedule_context = ScheduleContext::default();
2502 0 : let mut schedule_error = None;
2503 0 : let mut response_shards = Vec::new();
2504 0 : for tenant_shard_id in create_ids {
2505 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2506 :
2507 0 : let outcome = self
2508 0 : .do_initial_shard_scheduling(
2509 0 : tenant_shard_id,
2510 0 : initial_generation,
2511 0 : &create_req.shard_parameters,
2512 0 : create_req.config.clone(),
2513 0 : placement_policy.clone(),
2514 0 : preferred_az_id.as_ref(),
2515 0 : &mut schedule_context,
2516 0 : )
2517 0 : .await;
2518 :
2519 0 : match outcome {
2520 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2521 0 : InitialShardScheduleOutcome::NotScheduled => {}
2522 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2523 0 : schedule_error = Some(err);
2524 0 : }
2525 : }
2526 : }
2527 :
2528 : // If we failed to schedule shards, then they are still created in the controller,
2529 : // but we return an error to the requester to avoid a silent failure when someone
2530 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2531 : // are present in the system. We do this here rather than in the above loop, to
2532 : // avoid situations where we only create a subset of shards in the tenant.
2533 0 : if let Some(e) = schedule_error {
2534 0 : return Err(ApiError::Conflict(format!(
2535 0 : "Failed to schedule shard(s): {e}"
2536 0 : )));
2537 0 : }
2538 0 :
2539 0 : let waiters = {
2540 0 : let mut locked = self.inner.write().unwrap();
2541 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2542 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2543 0 : .tenant_creation_hint(true)
2544 0 : .build();
2545 0 : tenants
2546 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2547 0 : .filter_map(|(_shard_id, shard)| {
2548 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2549 0 : })
2550 0 : .collect::<Vec<_>>()
2551 0 : };
2552 0 :
2553 0 : Ok((
2554 0 : TenantCreateResponse {
2555 0 : shards: response_shards,
2556 0 : },
2557 0 : waiters,
2558 0 : ))
2559 0 : }
2560 :
2561 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2562 : /// case of a new tenant and a pre-existing one.
2563 : #[allow(clippy::too_many_arguments)]
2564 0 : async fn do_initial_shard_scheduling(
2565 0 : &self,
2566 0 : tenant_shard_id: TenantShardId,
2567 0 : initial_generation: Option<Generation>,
2568 0 : shard_params: &ShardParameters,
2569 0 : config: TenantConfig,
2570 0 : placement_policy: PlacementPolicy,
2571 0 : preferred_az_id: Option<&AvailabilityZone>,
2572 0 : schedule_context: &mut ScheduleContext,
2573 0 : ) -> InitialShardScheduleOutcome {
2574 0 : let mut locked = self.inner.write().unwrap();
2575 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2576 :
2577 : use std::collections::btree_map::Entry;
2578 0 : match tenants.entry(tenant_shard_id) {
2579 0 : Entry::Occupied(mut entry) => {
2580 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2581 :
2582 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2583 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2584 0 : }
2585 :
2586 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2587 0 : let generation = entry
2588 0 : .get()
2589 0 : .generation
2590 0 : .expect("Generation is set when in attached mode");
2591 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2592 0 : shard_id: tenant_shard_id,
2593 0 : node_id: *node_id,
2594 0 : generation: generation.into().unwrap(),
2595 0 : })
2596 : } else {
2597 0 : InitialShardScheduleOutcome::NotScheduled
2598 : }
2599 : }
2600 0 : Entry::Vacant(entry) => {
2601 0 : let state = entry.insert(TenantShard::new(
2602 0 : tenant_shard_id,
2603 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2604 0 : placement_policy,
2605 0 : preferred_az_id.cloned(),
2606 0 : ));
2607 0 :
2608 0 : state.generation = initial_generation;
2609 0 : state.config = config;
2610 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2611 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2612 0 : }
2613 :
2614 : // Only include shards in result if we are attaching: the purpose
2615 : // of the response is to tell the caller where the shards are attached.
2616 0 : if let Some(node_id) = state.intent.get_attached() {
2617 0 : let generation = state
2618 0 : .generation
2619 0 : .expect("Generation is set when in attached mode");
2620 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2621 0 : shard_id: tenant_shard_id,
2622 0 : node_id: *node_id,
2623 0 : generation: generation.into().unwrap(),
2624 0 : })
2625 : } else {
2626 0 : InitialShardScheduleOutcome::NotScheduled
2627 : }
2628 : }
2629 : }
2630 0 : }
2631 :
2632 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2633 : /// wait for reconciliation to complete before responding.
2634 0 : async fn await_waiters(
2635 0 : &self,
2636 0 : waiters: Vec<ReconcilerWaiter>,
2637 0 : timeout: Duration,
2638 0 : ) -> Result<(), ReconcileWaitError> {
2639 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2640 0 : for waiter in waiters {
2641 0 : let timeout = deadline.duration_since(Instant::now());
2642 0 : waiter.wait_timeout(timeout).await?;
2643 : }
2644 :
2645 0 : Ok(())
2646 0 : }
2647 :
2648 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2649 : /// in progress
2650 0 : async fn await_waiters_remainder(
2651 0 : &self,
2652 0 : waiters: Vec<ReconcilerWaiter>,
2653 0 : timeout: Duration,
2654 0 : ) -> Vec<ReconcilerWaiter> {
2655 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2656 0 : for waiter in waiters.iter() {
2657 0 : let timeout = deadline.duration_since(Instant::now());
2658 0 : let _ = waiter.wait_timeout(timeout).await;
2659 : }
2660 :
2661 0 : waiters
2662 0 : .into_iter()
2663 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2664 0 : .collect::<Vec<_>>()
2665 0 : }
2666 :
2667 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2668 : /// and transform it into either a tenant creation of a series of shard updates.
2669 : ///
2670 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2671 : /// still be returned.
2672 0 : fn tenant_location_config_prepare(
2673 0 : &self,
2674 0 : tenant_id: TenantId,
2675 0 : req: TenantLocationConfigRequest,
2676 0 : ) -> TenantCreateOrUpdate {
2677 0 : let mut updates = Vec::new();
2678 0 : let mut locked = self.inner.write().unwrap();
2679 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2680 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2681 :
2682 : // Use location config mode as an indicator of policy.
2683 0 : let placement_policy = match req.config.mode {
2684 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2685 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2686 : LocationConfigMode::AttachedMulti
2687 : | LocationConfigMode::AttachedSingle
2688 : | LocationConfigMode::AttachedStale => {
2689 0 : if nodes.len() > 1 {
2690 0 : PlacementPolicy::Attached(1)
2691 : } else {
2692 : // Convenience for dev/test: if we just have one pageserver, import
2693 : // tenants into non-HA mode so that scheduling will succeed.
2694 0 : PlacementPolicy::Attached(0)
2695 : }
2696 : }
2697 : };
2698 :
2699 : // Ordinarily we do not update scheduling policy, but when making major changes
2700 : // like detaching or demoting to secondary-only, we need to force the scheduling
2701 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2702 0 : let scheduling_policy = match req.config.mode {
2703 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2704 : // Special case: when making major changes like detaching or demoting to secondary-only,
2705 : // we need to force the scheduling mode to Active, or nothing will happen.
2706 0 : Some(ShardSchedulingPolicy::Active)
2707 : }
2708 : LocationConfigMode::AttachedMulti
2709 : | LocationConfigMode::AttachedSingle
2710 : | LocationConfigMode::AttachedStale => {
2711 : // While attached, continue to respect whatever the existing scheduling mode is.
2712 0 : None
2713 : }
2714 : };
2715 :
2716 0 : let mut create = true;
2717 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2718 : // Saw an existing shard: this is not a creation
2719 0 : create = false;
2720 :
2721 : // Shards may have initially been created by a Secondary request, where we
2722 : // would have left generation as None.
2723 : //
2724 : // We only update generation the first time we see an attached-mode request,
2725 : // and if there is no existing generation set. The caller is responsible for
2726 : // ensuring that no non-storage-controller pageserver ever uses a higher
2727 : // generation than they passed in here.
2728 : use LocationConfigMode::*;
2729 0 : let set_generation = match req.config.mode {
2730 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2731 0 : req.config.generation.map(Generation::new)
2732 : }
2733 0 : _ => None,
2734 : };
2735 :
2736 0 : updates.push(ShardUpdate {
2737 0 : tenant_shard_id: *shard_id,
2738 0 : placement_policy: placement_policy.clone(),
2739 0 : tenant_config: req.config.tenant_conf.clone(),
2740 0 : generation: set_generation,
2741 0 : scheduling_policy,
2742 0 : });
2743 : }
2744 :
2745 0 : if create {
2746 : use LocationConfigMode::*;
2747 0 : let generation = match req.config.mode {
2748 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2749 : // If a caller provided a generation in a non-attached request, ignore it
2750 : // and leave our generation as None: this enables a subsequent update to set
2751 : // the generation when setting an attached mode for the first time.
2752 0 : _ => None,
2753 : };
2754 :
2755 0 : TenantCreateOrUpdate::Create(
2756 0 : // Synthesize a creation request
2757 0 : TenantCreateRequest {
2758 0 : new_tenant_id: tenant_shard_id,
2759 0 : generation,
2760 0 : shard_parameters: ShardParameters {
2761 0 : count: tenant_shard_id.shard_count,
2762 0 : // We only import un-sharded or single-sharded tenants, so stripe
2763 0 : // size can be made up arbitrarily here.
2764 0 : stripe_size: DEFAULT_STRIPE_SIZE,
2765 0 : },
2766 0 : placement_policy: Some(placement_policy),
2767 0 : config: req.config.tenant_conf,
2768 0 : },
2769 0 : )
2770 : } else {
2771 0 : assert!(!updates.is_empty());
2772 0 : TenantCreateOrUpdate::Update(updates)
2773 : }
2774 0 : }
2775 :
2776 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2777 : /// the tenant is present in memory. If not, load it from the database. If it is found
2778 : /// in neither location, return a NotFound error.
2779 : ///
2780 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2781 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2782 0 : async fn maybe_load_tenant(
2783 0 : &self,
2784 0 : tenant_id: TenantId,
2785 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2786 0 : ) -> Result<(), ApiError> {
2787 : // Check if the tenant is present in memory, and select an AZ to use when loading
2788 : // if we will load it.
2789 0 : let load_in_az = {
2790 0 : let locked = self.inner.read().unwrap();
2791 0 : let existing = locked
2792 0 : .tenants
2793 0 : .range(TenantShardId::tenant_range(tenant_id))
2794 0 : .next();
2795 0 :
2796 0 : // If the tenant is not present in memory, we expect to load it from database,
2797 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2798 0 : if existing.is_none() {
2799 0 : locked
2800 0 : .scheduler
2801 0 : .get_az_for_new_tenant()
2802 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2803 0 : "No AZ with nodes found to load tenant"
2804 0 : )))?
2805 : } else {
2806 : // We already have this tenant in memory
2807 0 : return Ok(());
2808 : }
2809 : };
2810 :
2811 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2812 0 : if tenant_shards.is_empty() {
2813 0 : return Err(ApiError::NotFound(
2814 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2815 0 : ));
2816 0 : }
2817 0 :
2818 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2819 0 : self.persistence
2820 0 : .set_tenant_shard_preferred_azs(
2821 0 : tenant_shards
2822 0 : .iter()
2823 0 : .map(|t| {
2824 0 : (
2825 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2826 0 : Some(load_in_az.clone()),
2827 0 : )
2828 0 : })
2829 0 : .collect(),
2830 0 : )
2831 0 : .await?;
2832 :
2833 0 : let mut locked = self.inner.write().unwrap();
2834 0 : tracing::info!(
2835 0 : "Loaded {} shards for tenant {}",
2836 0 : tenant_shards.len(),
2837 : tenant_id
2838 : );
2839 :
2840 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2841 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2842 0 : let shard =
2843 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2844 0 :
2845 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2846 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2847 0 : if shard.policy != PlacementPolicy::Detached {
2848 0 : tracing::error!(
2849 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2850 : shard.tenant_shard_id,
2851 : shard.policy
2852 : );
2853 0 : }
2854 :
2855 0 : (shard.tenant_shard_id, shard)
2856 0 : }));
2857 0 :
2858 0 : Ok(())
2859 0 : }
2860 :
2861 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2862 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2863 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2864 : ///
2865 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2866 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2867 0 : fn maybe_drop_tenant(
2868 0 : &self,
2869 0 : tenant_id: TenantId,
2870 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2871 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2872 0 : ) {
2873 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2874 0 : if tenant_shards.all(|(_id, shard)| {
2875 0 : shard.policy == PlacementPolicy::Detached
2876 0 : && shard.reconciler.is_none()
2877 0 : && shard.observed.is_empty()
2878 0 : }) {
2879 0 : let keys = locked
2880 0 : .tenants
2881 0 : .range(TenantShardId::tenant_range(tenant_id))
2882 0 : .map(|(id, _)| id)
2883 0 : .copied()
2884 0 : .collect::<Vec<_>>();
2885 0 : for key in keys {
2886 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2887 0 : locked.tenants.remove(&key);
2888 : }
2889 0 : }
2890 0 : }
2891 :
2892 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2893 : /// directly with pageservers into this service.
2894 : ///
2895 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2896 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2897 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2898 : /// tenant's source of generation numbers.
2899 : ///
2900 : /// The mode in this request coarse-grained control of tenants:
2901 : /// - Call with mode Attached* to upsert the tenant.
2902 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2903 : /// to set an existing tenant to PolicyMode::Secondary
2904 : /// - Call with mode Detached to switch to PolicyMode::Detached
2905 0 : pub(crate) async fn tenant_location_config(
2906 0 : &self,
2907 0 : tenant_shard_id: TenantShardId,
2908 0 : req: TenantLocationConfigRequest,
2909 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2910 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2911 0 : let _tenant_lock = trace_exclusive_lock(
2912 0 : &self.tenant_op_locks,
2913 0 : tenant_shard_id.tenant_id,
2914 0 : TenantOperations::LocationConfig,
2915 0 : )
2916 0 : .await;
2917 :
2918 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2919 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2920 0 : "This API is for importing single-sharded or unsharded tenants"
2921 0 : )));
2922 : } else {
2923 0 : tenant_shard_id.tenant_id
2924 0 : };
2925 0 :
2926 0 : // In case we are waking up a Detached tenant
2927 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2928 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2929 0 : // This is a creation or an update
2930 0 : }
2931 0 : Err(e) => {
2932 0 : return Err(e);
2933 : }
2934 : };
2935 :
2936 : // First check if this is a creation or an update
2937 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2938 0 :
2939 0 : let mut result = TenantLocationConfigResponse {
2940 0 : shards: Vec::new(),
2941 0 : stripe_size: None,
2942 0 : };
2943 0 : let waiters = match create_or_update {
2944 0 : TenantCreateOrUpdate::Create(create_req) => {
2945 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2946 0 : result.shards = create_resp
2947 0 : .shards
2948 0 : .into_iter()
2949 0 : .map(|s| TenantShardLocation {
2950 0 : node_id: s.node_id,
2951 0 : shard_id: s.shard_id,
2952 0 : })
2953 0 : .collect();
2954 0 : waiters
2955 : }
2956 0 : TenantCreateOrUpdate::Update(updates) => {
2957 0 : // Persist updates
2958 0 : // Ordering: write to the database before applying changes in-memory, so that
2959 0 : // we will not appear time-travel backwards on a restart.
2960 0 :
2961 0 : let mut schedule_context = ScheduleContext::default();
2962 : for ShardUpdate {
2963 0 : tenant_shard_id,
2964 0 : placement_policy,
2965 0 : tenant_config,
2966 0 : generation,
2967 0 : scheduling_policy,
2968 0 : } in &updates
2969 : {
2970 0 : self.persistence
2971 0 : .update_tenant_shard(
2972 0 : TenantFilter::Shard(*tenant_shard_id),
2973 0 : Some(placement_policy.clone()),
2974 0 : Some(tenant_config.clone()),
2975 0 : *generation,
2976 0 : *scheduling_policy,
2977 0 : )
2978 0 : .await?;
2979 : }
2980 :
2981 : // Apply updates in-memory
2982 0 : let mut waiters = Vec::new();
2983 0 : {
2984 0 : let mut locked = self.inner.write().unwrap();
2985 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2986 :
2987 : for ShardUpdate {
2988 0 : tenant_shard_id,
2989 0 : placement_policy,
2990 0 : tenant_config,
2991 0 : generation: update_generation,
2992 0 : scheduling_policy,
2993 0 : } in updates
2994 : {
2995 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2996 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2997 0 : continue;
2998 : };
2999 :
3000 : // Update stripe size
3001 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
3002 0 : result.stripe_size = Some(shard.shard.stripe_size);
3003 0 : }
3004 :
3005 0 : shard.policy = placement_policy;
3006 0 : shard.config = tenant_config;
3007 0 : if let Some(generation) = update_generation {
3008 0 : shard.generation = Some(generation);
3009 0 : }
3010 :
3011 0 : if let Some(scheduling_policy) = scheduling_policy {
3012 0 : shard.set_scheduling_policy(scheduling_policy);
3013 0 : }
3014 :
3015 0 : shard.schedule(scheduler, &mut schedule_context)?;
3016 :
3017 0 : let maybe_waiter =
3018 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3019 0 : if let Some(waiter) = maybe_waiter {
3020 0 : waiters.push(waiter);
3021 0 : }
3022 :
3023 0 : if let Some(node_id) = shard.intent.get_attached() {
3024 0 : result.shards.push(TenantShardLocation {
3025 0 : shard_id: tenant_shard_id,
3026 0 : node_id: *node_id,
3027 0 : })
3028 0 : }
3029 : }
3030 : }
3031 0 : waiters
3032 : }
3033 : };
3034 :
3035 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3036 : // Do not treat a reconcile error as fatal: we have already applied any requested
3037 : // Intent changes, and the reconcile can fail for external reasons like unavailable
3038 : // compute notification API. In these cases, it is important that we do not
3039 : // cause the cloud control plane to retry forever on this API.
3040 0 : tracing::warn!(
3041 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
3042 : );
3043 0 : }
3044 :
3045 : // Logging the full result is useful because it lets us cross-check what the cloud control
3046 : // plane's tenant_shards table should contain.
3047 0 : tracing::info!("Complete, returning {result:?}");
3048 :
3049 0 : Ok(result)
3050 0 : }
3051 :
3052 0 : pub(crate) async fn tenant_config_patch(
3053 0 : &self,
3054 0 : req: TenantConfigPatchRequest,
3055 0 : ) -> Result<(), ApiError> {
3056 0 : let _tenant_lock = trace_exclusive_lock(
3057 0 : &self.tenant_op_locks,
3058 0 : req.tenant_id,
3059 0 : TenantOperations::ConfigPatch,
3060 0 : )
3061 0 : .await;
3062 :
3063 0 : let tenant_id = req.tenant_id;
3064 0 : let patch = req.config;
3065 0 :
3066 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3067 :
3068 0 : let base = {
3069 0 : let locked = self.inner.read().unwrap();
3070 0 : let shards = locked
3071 0 : .tenants
3072 0 : .range(TenantShardId::tenant_range(req.tenant_id));
3073 0 :
3074 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
3075 :
3076 0 : let first = match configs.peek() {
3077 0 : Some(first) => (*first).clone(),
3078 : None => {
3079 0 : return Err(ApiError::NotFound(
3080 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3081 0 : ));
3082 : }
3083 : };
3084 :
3085 0 : if !configs.all_equal() {
3086 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3087 : // This can't happen because we atomically update the database records
3088 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3089 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3090 0 : "Tenant configs for {} are mismatched",
3091 0 : req.tenant_id
3092 0 : )));
3093 0 : }
3094 0 :
3095 0 : first
3096 : };
3097 :
3098 0 : let updated_config = base
3099 0 : .apply_patch(patch)
3100 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3101 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3102 0 : .await
3103 0 : }
3104 :
3105 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3106 : // We require an exclusive lock, because we are updating persistent and in-memory state
3107 0 : let _tenant_lock = trace_exclusive_lock(
3108 0 : &self.tenant_op_locks,
3109 0 : req.tenant_id,
3110 0 : TenantOperations::ConfigSet,
3111 0 : )
3112 0 : .await;
3113 :
3114 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3115 :
3116 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3117 0 : .await
3118 0 : }
3119 :
3120 0 : async fn set_tenant_config_and_reconcile(
3121 0 : &self,
3122 0 : tenant_id: TenantId,
3123 0 : config: TenantConfig,
3124 0 : ) -> Result<(), ApiError> {
3125 0 : self.persistence
3126 0 : .update_tenant_shard(
3127 0 : TenantFilter::Tenant(tenant_id),
3128 0 : None,
3129 0 : Some(config.clone()),
3130 0 : None,
3131 0 : None,
3132 0 : )
3133 0 : .await?;
3134 :
3135 0 : let waiters = {
3136 0 : let mut waiters = Vec::new();
3137 0 : let mut locked = self.inner.write().unwrap();
3138 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3139 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3140 0 : shard.config = config.clone();
3141 0 : if let Some(waiter) =
3142 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3143 0 : {
3144 0 : waiters.push(waiter);
3145 0 : }
3146 : }
3147 0 : waiters
3148 : };
3149 :
3150 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3151 : // Treat this as success because we have stored the configuration. If e.g.
3152 : // a node was unavailable at this time, it should not stop us accepting a
3153 : // configuration change.
3154 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3155 0 : }
3156 :
3157 0 : Ok(())
3158 0 : }
3159 :
3160 0 : pub(crate) fn tenant_config_get(
3161 0 : &self,
3162 0 : tenant_id: TenantId,
3163 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3164 0 : let config = {
3165 0 : let locked = self.inner.read().unwrap();
3166 0 :
3167 0 : match locked
3168 0 : .tenants
3169 0 : .range(TenantShardId::tenant_range(tenant_id))
3170 0 : .next()
3171 : {
3172 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3173 : None => {
3174 0 : return Err(ApiError::NotFound(
3175 0 : anyhow::anyhow!("Tenant not found").into(),
3176 0 : ));
3177 : }
3178 : }
3179 : };
3180 :
3181 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3182 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3183 : // and `effective_config` in the response is meaningless, but we retain that syntax
3184 : // in order to remain compatible with the pageserver API.
3185 :
3186 0 : let response = HashMap::from([
3187 : (
3188 : "tenant_specific_overrides",
3189 0 : serde_json::to_value(&config)
3190 0 : .context("serializing tenant specific overrides")
3191 0 : .map_err(ApiError::InternalServerError)?,
3192 : ),
3193 : (
3194 0 : "effective_config",
3195 0 : serde_json::to_value(&config)
3196 0 : .context("serializing effective config")
3197 0 : .map_err(ApiError::InternalServerError)?,
3198 : ),
3199 : ]);
3200 :
3201 0 : Ok(response)
3202 0 : }
3203 :
3204 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3205 0 : &self,
3206 0 : time_travel_req: &TenantTimeTravelRequest,
3207 0 : tenant_id: TenantId,
3208 0 : timestamp: Cow<'_, str>,
3209 0 : done_if_after: Cow<'_, str>,
3210 0 : ) -> Result<(), ApiError> {
3211 0 : let _tenant_lock = trace_exclusive_lock(
3212 0 : &self.tenant_op_locks,
3213 0 : tenant_id,
3214 0 : TenantOperations::TimeTravelRemoteStorage,
3215 0 : )
3216 0 : .await;
3217 :
3218 0 : let node = {
3219 0 : let mut locked = self.inner.write().unwrap();
3220 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3221 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3222 : // but only at the start of the process, so it's really just to prevent operator
3223 : // mistakes.
3224 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3225 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3226 : {
3227 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3228 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3229 0 : )));
3230 0 : }
3231 0 : let maybe_attached = shard
3232 0 : .observed
3233 0 : .locations
3234 0 : .iter()
3235 0 : .filter_map(|(node_id, observed_location)| {
3236 0 : observed_location
3237 0 : .conf
3238 0 : .as_ref()
3239 0 : .map(|loc| (node_id, observed_location, loc.mode))
3240 0 : })
3241 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3242 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3243 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3244 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3245 0 : )));
3246 0 : }
3247 : }
3248 0 : let scheduler = &mut locked.scheduler;
3249 : // Right now we only perform the operation on a single node without parallelization
3250 : // TODO fan out the operation to multiple nodes for better performance
3251 0 : let node_id = scheduler.any_available_node()?;
3252 0 : let node = locked
3253 0 : .nodes
3254 0 : .get(&node_id)
3255 0 : .expect("Pageservers may not be deleted while lock is active");
3256 0 : node.clone()
3257 0 : };
3258 0 :
3259 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3260 0 : let mut counts = time_travel_req
3261 0 : .shard_counts
3262 0 : .iter()
3263 0 : .copied()
3264 0 : .collect::<HashSet<_>>()
3265 0 : .into_iter()
3266 0 : .collect::<Vec<_>>();
3267 0 : counts.sort_unstable();
3268 :
3269 0 : for count in counts {
3270 0 : let shard_ids = (0..count.count())
3271 0 : .map(|i| TenantShardId {
3272 0 : tenant_id,
3273 0 : shard_number: ShardNumber(i),
3274 0 : shard_count: count,
3275 0 : })
3276 0 : .collect::<Vec<_>>();
3277 0 : for tenant_shard_id in shard_ids {
3278 0 : let client = PageserverClient::new(
3279 0 : node.get_id(),
3280 0 : self.http_client.clone(),
3281 0 : node.base_url(),
3282 0 : self.config.pageserver_jwt_token.as_deref(),
3283 0 : );
3284 0 :
3285 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3286 :
3287 0 : client
3288 0 : .tenant_time_travel_remote_storage(
3289 0 : tenant_shard_id,
3290 0 : ×tamp,
3291 0 : &done_if_after,
3292 0 : )
3293 0 : .await
3294 0 : .map_err(|e| {
3295 0 : ApiError::InternalServerError(anyhow::anyhow!(
3296 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3297 0 : node
3298 0 : ))
3299 0 : })?;
3300 : }
3301 : }
3302 0 : Ok(())
3303 0 : }
3304 :
3305 0 : pub(crate) async fn tenant_secondary_download(
3306 0 : &self,
3307 0 : tenant_id: TenantId,
3308 0 : wait: Option<Duration>,
3309 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3310 0 : let _tenant_lock = trace_shared_lock(
3311 0 : &self.tenant_op_locks,
3312 0 : tenant_id,
3313 0 : TenantOperations::SecondaryDownload,
3314 0 : )
3315 0 : .await;
3316 :
3317 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3318 0 : let targets = {
3319 0 : let locked = self.inner.read().unwrap();
3320 0 : let mut targets = Vec::new();
3321 :
3322 0 : for (tenant_shard_id, shard) in
3323 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3324 : {
3325 0 : for node_id in shard.intent.get_secondary() {
3326 0 : let node = locked
3327 0 : .nodes
3328 0 : .get(node_id)
3329 0 : .expect("Pageservers may not be deleted while referenced");
3330 0 :
3331 0 : targets.push((*tenant_shard_id, node.clone()));
3332 0 : }
3333 : }
3334 0 : targets
3335 0 : };
3336 0 :
3337 0 : // Issue concurrent requests to all shards' locations
3338 0 : let mut futs = FuturesUnordered::new();
3339 0 : for (tenant_shard_id, node) in targets {
3340 0 : let client = PageserverClient::new(
3341 0 : node.get_id(),
3342 0 : self.http_client.clone(),
3343 0 : node.base_url(),
3344 0 : self.config.pageserver_jwt_token.as_deref(),
3345 0 : );
3346 0 : futs.push(async move {
3347 0 : let result = client
3348 0 : .tenant_secondary_download(tenant_shard_id, wait)
3349 0 : .await;
3350 0 : (result, node, tenant_shard_id)
3351 0 : })
3352 : }
3353 :
3354 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3355 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3356 : // well as more general cases like 503s, 500s, or timeouts.
3357 0 : let mut aggregate_progress = SecondaryProgress::default();
3358 0 : let mut aggregate_status: Option<StatusCode> = None;
3359 0 : let mut error: Option<mgmt_api::Error> = None;
3360 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3361 0 : match result {
3362 0 : Err(e) => {
3363 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3364 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3365 0 : // than they had hoped for.
3366 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3367 0 : error = Some(e)
3368 : }
3369 0 : Ok((status_code, progress)) => {
3370 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3371 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3372 0 : aggregate_progress.layers_total += progress.layers_total;
3373 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3374 0 : aggregate_progress.bytes_total += progress.bytes_total;
3375 0 : aggregate_progress.heatmap_mtime =
3376 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3377 0 : aggregate_status = match aggregate_status {
3378 0 : None => Some(status_code),
3379 0 : Some(StatusCode::OK) => Some(status_code),
3380 0 : Some(cur) => {
3381 0 : // Other status codes (e.g. 202) -- do not overwrite.
3382 0 : Some(cur)
3383 : }
3384 : };
3385 : }
3386 : }
3387 : }
3388 :
3389 : // If any of the shards return 202, indicate our result as 202.
3390 0 : match aggregate_status {
3391 : None => {
3392 0 : match error {
3393 0 : Some(e) => {
3394 0 : // No successes, and an error: surface it
3395 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3396 : }
3397 : None => {
3398 : // No shards found
3399 0 : Err(ApiError::NotFound(
3400 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3401 0 : ))
3402 : }
3403 : }
3404 : }
3405 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3406 : }
3407 0 : }
3408 :
3409 0 : pub(crate) async fn tenant_delete(
3410 0 : self: &Arc<Self>,
3411 0 : tenant_id: TenantId,
3412 0 : ) -> Result<StatusCode, ApiError> {
3413 0 : let _tenant_lock =
3414 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3415 :
3416 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3417 :
3418 : // Detach all shards. This also deletes local pageserver shard data.
3419 0 : let (detach_waiters, node) = {
3420 0 : let mut detach_waiters = Vec::new();
3421 0 : let mut locked = self.inner.write().unwrap();
3422 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3423 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3424 : // Update the tenant's intent to remove all attachments
3425 0 : shard.policy = PlacementPolicy::Detached;
3426 0 : shard
3427 0 : .schedule(scheduler, &mut ScheduleContext::default())
3428 0 : .expect("De-scheduling is infallible");
3429 0 : debug_assert!(shard.intent.get_attached().is_none());
3430 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3431 :
3432 0 : if let Some(waiter) =
3433 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3434 0 : {
3435 0 : detach_waiters.push(waiter);
3436 0 : }
3437 : }
3438 :
3439 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3440 : // was attached, just has to be able to see the S3 content)
3441 0 : let node_id = scheduler.any_available_node()?;
3442 0 : let node = nodes
3443 0 : .get(&node_id)
3444 0 : .expect("Pageservers may not be deleted while lock is active");
3445 0 : (detach_waiters, node.clone())
3446 0 : };
3447 0 :
3448 0 : // This reconcile wait can fail in a few ways:
3449 0 : // A there is a very long queue for the reconciler semaphore
3450 0 : // B some pageserver is failing to handle a detach promptly
3451 0 : // C some pageserver goes offline right at the moment we send it a request.
3452 0 : //
3453 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3454 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3455 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3456 0 : // deleting the underlying data).
3457 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3458 0 : .await?;
3459 :
3460 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3461 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3462 : // the tenant ID prefix, including all shards (even possibly stale ones).
3463 0 : match node
3464 0 : .with_client_retries(
3465 0 : |client| async move {
3466 0 : client
3467 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3468 0 : .await
3469 0 : },
3470 0 : &self.http_client,
3471 0 : &self.config.pageserver_jwt_token,
3472 0 : 1,
3473 0 : 3,
3474 0 : RECONCILE_TIMEOUT,
3475 0 : &self.cancel,
3476 0 : )
3477 0 : .await
3478 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3479 : {
3480 0 : Ok(_) => {}
3481 : Err(mgmt_api::Error::Cancelled) => {
3482 0 : return Err(ApiError::ShuttingDown);
3483 : }
3484 0 : Err(e) => {
3485 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3486 0 : // at large is unavailable.
3487 0 : tracing::error!("Error deleting via node {node}: {e}");
3488 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3489 : }
3490 : }
3491 :
3492 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3493 : // our in-memory state and database state.
3494 :
3495 : // Ordering: we delete persistent state first: if we then
3496 : // crash, we will drop the in-memory state.
3497 :
3498 : // Drop persistent state.
3499 0 : self.persistence.delete_tenant(tenant_id).await?;
3500 :
3501 : // Drop in-memory state
3502 : {
3503 0 : let mut locked = self.inner.write().unwrap();
3504 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3505 :
3506 : // Dereference Scheduler from shards before dropping them
3507 0 : for (_tenant_shard_id, shard) in
3508 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3509 0 : {
3510 0 : shard.intent.clear(scheduler);
3511 0 : }
3512 :
3513 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3514 0 : tracing::info!(
3515 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3516 0 : locked.tenants.len()
3517 : );
3518 : };
3519 :
3520 : // Delete the tenant from safekeepers (if needed)
3521 0 : self.tenant_delete_safekeepers(tenant_id)
3522 0 : .instrument(tracing::info_span!("tenant_delete_safekeepers", %tenant_id))
3523 0 : .await?;
3524 :
3525 : // Success is represented as 404, to imitate the existing pageserver deletion API
3526 0 : Ok(StatusCode::NOT_FOUND)
3527 0 : }
3528 :
3529 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3530 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3531 : /// the tenant's policies (configuration) within the storage controller
3532 0 : pub(crate) async fn tenant_update_policy(
3533 0 : &self,
3534 0 : tenant_id: TenantId,
3535 0 : req: TenantPolicyRequest,
3536 0 : ) -> Result<(), ApiError> {
3537 : // We require an exclusive lock, because we are updating persistent and in-memory state
3538 0 : let _tenant_lock = trace_exclusive_lock(
3539 0 : &self.tenant_op_locks,
3540 0 : tenant_id,
3541 0 : TenantOperations::UpdatePolicy,
3542 0 : )
3543 0 : .await;
3544 :
3545 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3546 :
3547 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3548 :
3549 : let TenantPolicyRequest {
3550 0 : placement,
3551 0 : mut scheduling,
3552 0 : } = req;
3553 :
3554 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3555 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3556 : // this to take effect.
3557 0 : if scheduling.is_none() {
3558 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3559 0 : }
3560 0 : }
3561 :
3562 0 : self.persistence
3563 0 : .update_tenant_shard(
3564 0 : TenantFilter::Tenant(tenant_id),
3565 0 : placement.clone(),
3566 0 : None,
3567 0 : None,
3568 0 : scheduling,
3569 0 : )
3570 0 : .await?;
3571 :
3572 0 : let mut schedule_context = ScheduleContext::default();
3573 0 : let mut locked = self.inner.write().unwrap();
3574 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3575 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3576 0 : if let Some(placement) = &placement {
3577 0 : shard.policy = placement.clone();
3578 0 :
3579 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3580 0 : "Updated placement policy to {placement:?}");
3581 0 : }
3582 :
3583 0 : if let Some(scheduling) = &scheduling {
3584 0 : shard.set_scheduling_policy(*scheduling);
3585 0 :
3586 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3587 0 : "Updated scheduling policy to {scheduling:?}");
3588 0 : }
3589 :
3590 : // In case scheduling is being switched back on, try it now.
3591 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3592 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3593 : }
3594 :
3595 0 : Ok(())
3596 0 : }
3597 :
3598 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3599 0 : &self,
3600 0 : tenant_id: TenantId,
3601 0 : mut create_req: TimelineCreateRequest,
3602 0 : ) -> Result<TimelineInfo, ApiError> {
3603 0 : tracing::info!(
3604 0 : "Creating timeline {}/{}",
3605 : tenant_id,
3606 : create_req.new_timeline_id,
3607 : );
3608 :
3609 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3610 0 : if targets.0.is_empty() {
3611 0 : return Err(ApiError::NotFound(
3612 0 : anyhow::anyhow!("Tenant not found").into(),
3613 0 : ));
3614 0 : };
3615 0 :
3616 0 : let (shard_zero_tid, shard_zero_locations) =
3617 0 : targets.0.pop_first().expect("Must have at least one shard");
3618 0 : assert!(shard_zero_tid.is_shard_zero());
3619 :
3620 0 : async fn create_one(
3621 0 : tenant_shard_id: TenantShardId,
3622 0 : locations: ShardMutationLocations,
3623 0 : http_client: reqwest::Client,
3624 0 : jwt: Option<String>,
3625 0 : create_req: TimelineCreateRequest,
3626 0 : ) -> Result<TimelineInfo, ApiError> {
3627 0 : let latest = locations.latest.node;
3628 0 :
3629 0 : tracing::info!(
3630 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3631 : tenant_shard_id,
3632 : create_req.new_timeline_id,
3633 : locations.latest.generation
3634 : );
3635 :
3636 0 : let client =
3637 0 : PageserverClient::new(latest.get_id(), http_client.clone(), latest.base_url(), jwt.as_deref());
3638 :
3639 0 : let timeline_info = client
3640 0 : .timeline_create(tenant_shard_id, &create_req)
3641 0 : .await
3642 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3643 :
3644 : // We propagate timeline creations to all attached locations such that a compute
3645 : // for the new timeline is able to start regardless of the current state of the
3646 : // tenant shard reconciliation.
3647 0 : for location in locations.other {
3648 0 : tracing::info!(
3649 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3650 : tenant_shard_id,
3651 : create_req.new_timeline_id,
3652 : location.node,
3653 : location.generation
3654 : );
3655 :
3656 0 : let client = PageserverClient::new(
3657 0 : location.node.get_id(),
3658 0 : http_client.clone(),
3659 0 : location.node.base_url(),
3660 0 : jwt.as_deref(),
3661 0 : );
3662 :
3663 0 : let res = client
3664 0 : .timeline_create(tenant_shard_id, &create_req)
3665 0 : .await;
3666 :
3667 0 : if let Err(e) = res {
3668 0 : match e {
3669 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3670 0 : // Tenant might have been detached from the stale location,
3671 0 : // so ignore 404s.
3672 0 : },
3673 : _ => {
3674 0 : return Err(passthrough_api_error(&location.node, e));
3675 : }
3676 : }
3677 0 : }
3678 : }
3679 :
3680 0 : Ok(timeline_info)
3681 0 : }
3682 :
3683 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3684 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3685 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3686 0 : let timeline_info = create_one(
3687 0 : shard_zero_tid,
3688 0 : shard_zero_locations,
3689 0 : self.http_client.clone(),
3690 0 : self.config.pageserver_jwt_token.clone(),
3691 0 : create_req.clone(),
3692 0 : )
3693 0 : .await?;
3694 :
3695 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3696 0 : match &mut create_req.mode {
3697 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3698 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3699 0 : },
3700 0 : _ => {}
3701 : }
3702 :
3703 : // Create timeline on remaining shards with number >0
3704 0 : if !targets.0.is_empty() {
3705 : // If we had multiple shards, issue requests for the remainder now.
3706 0 : let jwt = &self.config.pageserver_jwt_token;
3707 0 : self.tenant_for_shards(
3708 0 : targets
3709 0 : .0
3710 0 : .iter()
3711 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3712 0 : .collect(),
3713 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3714 0 : let create_req = create_req.clone();
3715 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3716 0 : Box::pin(create_one(
3717 0 : tenant_shard_id,
3718 0 : mutation_locations,
3719 0 : self.http_client.clone(),
3720 0 : jwt.clone(),
3721 0 : create_req,
3722 0 : ))
3723 0 : },
3724 0 : )
3725 0 : .await?;
3726 0 : }
3727 :
3728 0 : Ok(timeline_info)
3729 0 : })
3730 0 : .await?
3731 0 : }
3732 :
3733 0 : pub(crate) async fn tenant_timeline_create(
3734 0 : self: &Arc<Self>,
3735 0 : tenant_id: TenantId,
3736 0 : create_req: TimelineCreateRequest,
3737 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
3738 0 : let safekeepers = self.config.timelines_onto_safekeepers;
3739 0 : tracing::info!(
3740 : %safekeepers,
3741 0 : "Creating timeline {}/{}",
3742 : tenant_id,
3743 : create_req.new_timeline_id,
3744 : );
3745 :
3746 0 : let _tenant_lock = trace_shared_lock(
3747 0 : &self.tenant_op_locks,
3748 0 : tenant_id,
3749 0 : TenantOperations::TimelineCreate,
3750 0 : )
3751 0 : .await;
3752 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3753 0 : let create_mode = create_req.mode.clone();
3754 :
3755 0 : let timeline_info = self
3756 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
3757 0 : .await?;
3758 :
3759 0 : let safekeepers = if safekeepers {
3760 0 : let res = self
3761 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, create_mode)
3762 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
3763 0 : .await?;
3764 0 : Some(res)
3765 : } else {
3766 0 : None
3767 : };
3768 :
3769 0 : Ok(TimelineCreateResponseStorcon {
3770 0 : timeline_info,
3771 0 : safekeepers,
3772 0 : })
3773 0 : }
3774 :
3775 0 : pub(crate) async fn tenant_timeline_archival_config(
3776 0 : &self,
3777 0 : tenant_id: TenantId,
3778 0 : timeline_id: TimelineId,
3779 0 : req: TimelineArchivalConfigRequest,
3780 0 : ) -> Result<(), ApiError> {
3781 0 : tracing::info!(
3782 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3783 : req.state
3784 : );
3785 :
3786 0 : let _tenant_lock = trace_shared_lock(
3787 0 : &self.tenant_op_locks,
3788 0 : tenant_id,
3789 0 : TenantOperations::TimelineArchivalConfig,
3790 0 : )
3791 0 : .await;
3792 :
3793 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3794 0 : if targets.0.is_empty() {
3795 0 : return Err(ApiError::NotFound(
3796 0 : anyhow::anyhow!("Tenant not found").into(),
3797 0 : ));
3798 0 : }
3799 0 : async fn config_one(
3800 0 : tenant_shard_id: TenantShardId,
3801 0 : timeline_id: TimelineId,
3802 0 : node: Node,
3803 0 : http_client: reqwest::Client,
3804 0 : jwt: Option<String>,
3805 0 : req: TimelineArchivalConfigRequest,
3806 0 : ) -> Result<(), ApiError> {
3807 0 : tracing::info!(
3808 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3809 : );
3810 :
3811 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
3812 0 :
3813 0 : client
3814 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3815 0 : .await
3816 0 : .map_err(|e| match e {
3817 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3818 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3819 : }
3820 0 : _ => passthrough_api_error(&node, e),
3821 0 : })
3822 0 : }
3823 :
3824 : // no shard needs to go first/last; the operation should be idempotent
3825 : // TODO: it would be great to ensure that all shards return the same error
3826 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3827 0 : let results = self
3828 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3829 0 : futures::FutureExt::boxed(config_one(
3830 0 : tenant_shard_id,
3831 0 : timeline_id,
3832 0 : node,
3833 0 : self.http_client.clone(),
3834 0 : self.config.pageserver_jwt_token.clone(),
3835 0 : req.clone(),
3836 0 : ))
3837 0 : })
3838 0 : .await?;
3839 0 : assert!(!results.is_empty(), "must have at least one result");
3840 :
3841 0 : Ok(())
3842 0 : }).await?
3843 0 : }
3844 :
3845 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3846 0 : &self,
3847 0 : tenant_id: TenantId,
3848 0 : timeline_id: TimelineId,
3849 0 : behavior: Option<DetachBehavior>,
3850 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3851 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3852 :
3853 0 : let _tenant_lock = trace_shared_lock(
3854 0 : &self.tenant_op_locks,
3855 0 : tenant_id,
3856 0 : TenantOperations::TimelineDetachAncestor,
3857 0 : )
3858 0 : .await;
3859 :
3860 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3861 0 : if targets.0.is_empty() {
3862 0 : return Err(ApiError::NotFound(
3863 0 : anyhow::anyhow!("Tenant not found").into(),
3864 0 : ));
3865 0 : }
3866 :
3867 0 : async fn detach_one(
3868 0 : tenant_shard_id: TenantShardId,
3869 0 : timeline_id: TimelineId,
3870 0 : node: Node,
3871 0 : http_client: reqwest::Client,
3872 0 : jwt: Option<String>,
3873 0 : behavior: Option<DetachBehavior>,
3874 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3875 0 : tracing::info!(
3876 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3877 : );
3878 :
3879 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
3880 0 :
3881 0 : client
3882 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id, behavior)
3883 0 : .await
3884 0 : .map_err(|e| {
3885 : use mgmt_api::Error;
3886 :
3887 0 : match e {
3888 : // no ancestor (ever)
3889 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3890 0 : "{node}: {}",
3891 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3892 0 : )),
3893 : // too many ancestors
3894 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3895 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3896 : }
3897 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3898 0 : // avoid turning these into conflicts to remain compatible with
3899 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3900 0 : // detach
3901 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3902 : }
3903 : // rest can be mapped as usual
3904 0 : other => passthrough_api_error(&node, other),
3905 : }
3906 0 : })
3907 0 : .map(|res| (tenant_shard_id.shard_number, res))
3908 0 : }
3909 :
3910 : // no shard needs to go first/last; the operation should be idempotent
3911 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
3912 0 : let mut results = self
3913 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
3914 0 : futures::FutureExt::boxed(detach_one(
3915 0 : tenant_shard_id,
3916 0 : timeline_id,
3917 0 : node,
3918 0 : self.http_client.clone(),
3919 0 : self.config.pageserver_jwt_token.clone(),
3920 0 : behavior,
3921 0 : ))
3922 0 : })
3923 0 : .await?;
3924 :
3925 0 : let any = results.pop().expect("we must have at least one response");
3926 0 :
3927 0 : let mismatching = results
3928 0 : .iter()
3929 0 : .filter(|(_, res)| res != &any.1)
3930 0 : .collect::<Vec<_>>();
3931 0 : if !mismatching.is_empty() {
3932 : // this can be hit by races which should not happen because operation lock on cplane
3933 0 : let matching = results.len() - mismatching.len();
3934 0 : tracing::error!(
3935 : matching,
3936 : compared_against=?any,
3937 : ?mismatching,
3938 0 : "shards returned different results"
3939 : );
3940 :
3941 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3942 0 : }
3943 0 :
3944 0 : Ok(any.1)
3945 0 : }).await?
3946 0 : }
3947 :
3948 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3949 0 : &self,
3950 0 : tenant_id: TenantId,
3951 0 : timeline_id: TimelineId,
3952 0 : dir: BlockUnblock,
3953 0 : ) -> Result<(), ApiError> {
3954 0 : let _tenant_lock = trace_shared_lock(
3955 0 : &self.tenant_op_locks,
3956 0 : tenant_id,
3957 0 : TenantOperations::TimelineGcBlockUnblock,
3958 0 : )
3959 0 : .await;
3960 :
3961 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3962 0 : if targets.0.is_empty() {
3963 0 : return Err(ApiError::NotFound(
3964 0 : anyhow::anyhow!("Tenant not found").into(),
3965 0 : ));
3966 0 : }
3967 :
3968 0 : async fn do_one(
3969 0 : tenant_shard_id: TenantShardId,
3970 0 : timeline_id: TimelineId,
3971 0 : node: Node,
3972 0 : http_client: reqwest::Client,
3973 0 : jwt: Option<String>,
3974 0 : dir: BlockUnblock,
3975 0 : ) -> Result<(), ApiError> {
3976 0 : let client = PageserverClient::new(
3977 0 : node.get_id(),
3978 0 : http_client,
3979 0 : node.base_url(),
3980 0 : jwt.as_deref(),
3981 0 : );
3982 0 :
3983 0 : client
3984 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3985 0 : .await
3986 0 : .map_err(|e| passthrough_api_error(&node, e))
3987 0 : }
3988 :
3989 : // no shard needs to go first/last; the operation should be idempotent
3990 0 : let locations = targets
3991 0 : .0
3992 0 : .iter()
3993 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3994 0 : .collect();
3995 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
3996 0 : futures::FutureExt::boxed(do_one(
3997 0 : tenant_shard_id,
3998 0 : timeline_id,
3999 0 : node,
4000 0 : self.http_client.clone(),
4001 0 : self.config.pageserver_jwt_token.clone(),
4002 0 : dir,
4003 0 : ))
4004 0 : })
4005 0 : .await
4006 0 : })
4007 0 : .await??;
4008 0 : Ok(())
4009 0 : }
4010 :
4011 0 : pub(crate) async fn tenant_timeline_lsn_lease(
4012 0 : &self,
4013 0 : tenant_id: TenantId,
4014 0 : timeline_id: TimelineId,
4015 0 : lsn: Lsn,
4016 0 : ) -> Result<LsnLease, ApiError> {
4017 0 : let _tenant_lock = trace_shared_lock(
4018 0 : &self.tenant_op_locks,
4019 0 : tenant_id,
4020 0 : TenantOperations::TimelineLsnLease,
4021 0 : )
4022 0 : .await;
4023 :
4024 0 : let targets = {
4025 0 : let locked = self.inner.read().unwrap();
4026 0 : let mut targets = Vec::new();
4027 0 :
4028 0 : // If the request got an unsharded tenant id, then apply
4029 0 : // the operation to all shards. Otherwise, apply it to a specific shard.
4030 0 : let shards_range = TenantShardId::tenant_range(tenant_id);
4031 :
4032 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4033 0 : if let Some(node_id) = shard.intent.get_attached() {
4034 0 : let node = locked
4035 0 : .nodes
4036 0 : .get(node_id)
4037 0 : .expect("Pageservers may not be deleted while referenced");
4038 0 :
4039 0 : targets.push((*tenant_shard_id, node.clone()));
4040 0 : }
4041 : }
4042 0 : targets
4043 : };
4044 :
4045 0 : let res = self
4046 0 : .tenant_for_shards_api(
4047 0 : targets,
4048 0 : |tenant_shard_id, client| async move {
4049 0 : client
4050 0 : .timeline_lease_lsn(tenant_shard_id, timeline_id, lsn)
4051 0 : .await
4052 0 : },
4053 0 : 1,
4054 0 : 1,
4055 0 : SHORT_RECONCILE_TIMEOUT,
4056 0 : &self.cancel,
4057 0 : )
4058 0 : .await;
4059 :
4060 0 : let mut valid_until = None;
4061 0 : for r in res {
4062 0 : match r {
4063 0 : Ok(lease) => {
4064 0 : if let Some(ref mut valid_until) = valid_until {
4065 0 : *valid_until = std::cmp::min(*valid_until, lease.valid_until);
4066 0 : } else {
4067 0 : valid_until = Some(lease.valid_until);
4068 0 : }
4069 : }
4070 0 : Err(e) => {
4071 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
4072 : }
4073 : }
4074 : }
4075 0 : Ok(LsnLease {
4076 0 : valid_until: valid_until.unwrap_or_else(SystemTime::now),
4077 0 : })
4078 0 : }
4079 :
4080 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4081 0 : &self,
4082 0 : tenant_shard_id: TenantShardId,
4083 0 : timeline_id: TimelineId,
4084 0 : concurrency: Option<usize>,
4085 0 : recurse: bool,
4086 0 : ) -> Result<(), ApiError> {
4087 0 : let _tenant_lock = trace_shared_lock(
4088 0 : &self.tenant_op_locks,
4089 0 : tenant_shard_id.tenant_id,
4090 0 : TenantOperations::DownloadHeatmapLayers,
4091 0 : )
4092 0 : .await;
4093 :
4094 0 : let targets = {
4095 0 : let locked = self.inner.read().unwrap();
4096 0 : let mut targets = Vec::new();
4097 :
4098 : // If the request got an unsharded tenant id, then apply
4099 : // the operation to all shards. Otherwise, apply it to a specific shard.
4100 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4101 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4102 : } else {
4103 0 : tenant_shard_id.range()
4104 : };
4105 :
4106 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4107 0 : if let Some(node_id) = shard.intent.get_attached() {
4108 0 : let node = locked
4109 0 : .nodes
4110 0 : .get(node_id)
4111 0 : .expect("Pageservers may not be deleted while referenced");
4112 0 :
4113 0 : targets.push((*tenant_shard_id, node.clone()));
4114 0 : }
4115 : }
4116 0 : targets
4117 0 : };
4118 0 :
4119 0 : self.tenant_for_shards_api(
4120 0 : targets,
4121 0 : |tenant_shard_id, client| async move {
4122 0 : client
4123 0 : .timeline_download_heatmap_layers(
4124 0 : tenant_shard_id,
4125 0 : timeline_id,
4126 0 : concurrency,
4127 0 : recurse,
4128 0 : )
4129 0 : .await
4130 0 : },
4131 0 : 1,
4132 0 : 1,
4133 0 : SHORT_RECONCILE_TIMEOUT,
4134 0 : &self.cancel,
4135 0 : )
4136 0 : .await;
4137 :
4138 0 : Ok(())
4139 0 : }
4140 :
4141 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4142 : ///
4143 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`
4144 : /// and returned element at index `i` is the result for `req_fn(op(locations[i])`.
4145 0 : async fn tenant_for_shards<F, R>(
4146 0 : &self,
4147 0 : locations: Vec<(TenantShardId, Node)>,
4148 0 : mut req_fn: F,
4149 0 : ) -> Result<Vec<R>, ApiError>
4150 0 : where
4151 0 : F: FnMut(
4152 0 : TenantShardId,
4153 0 : Node,
4154 0 : )
4155 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4156 0 : {
4157 0 : let mut futs = FuturesUnordered::new();
4158 0 : let mut results = Vec::with_capacity(locations.len());
4159 :
4160 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4161 0 : let fut = req_fn(tenant_shard_id, node);
4162 0 : futs.push(async move { (idx, fut.await) });
4163 0 : }
4164 :
4165 0 : while let Some((idx, r)) = futs.next().await {
4166 0 : results.push((idx, r?));
4167 : }
4168 :
4169 0 : results.sort_by_key(|(idx, _)| *idx);
4170 0 : Ok(results.into_iter().map(|(_, r)| r).collect())
4171 0 : }
4172 :
4173 : /// Concurrently invoke a pageserver API call on many shards at once.
4174 : ///
4175 : /// The returned Vec has the same length as the `locations` Vec,
4176 : /// and returned element at index `i` is the result for `op(locations[i])`.
4177 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
4178 0 : &self,
4179 0 : locations: Vec<(TenantShardId, Node)>,
4180 0 : op: O,
4181 0 : warn_threshold: u32,
4182 0 : max_retries: u32,
4183 0 : timeout: Duration,
4184 0 : cancel: &CancellationToken,
4185 0 : ) -> Vec<mgmt_api::Result<T>>
4186 0 : where
4187 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
4188 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
4189 0 : {
4190 0 : let mut futs = FuturesUnordered::new();
4191 0 : let mut results = Vec::with_capacity(locations.len());
4192 :
4193 0 : for (idx, (tenant_shard_id, node)) in locations.into_iter().enumerate() {
4194 0 : futs.push(async move {
4195 0 : let r = node
4196 0 : .with_client_retries(
4197 0 : |client| op(tenant_shard_id, client),
4198 0 : &self.http_client,
4199 0 : &self.config.pageserver_jwt_token,
4200 0 : warn_threshold,
4201 0 : max_retries,
4202 0 : timeout,
4203 0 : cancel,
4204 0 : )
4205 0 : .await;
4206 0 : (idx, r)
4207 0 : });
4208 0 : }
4209 :
4210 0 : while let Some((idx, r)) = futs.next().await {
4211 0 : results.push((idx, r.unwrap_or(Err(mgmt_api::Error::Cancelled))));
4212 0 : }
4213 :
4214 0 : results.sort_by_key(|(idx, _)| *idx);
4215 0 : results.into_iter().map(|(_, r)| r).collect()
4216 0 : }
4217 :
4218 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
4219 : /// when creating and deleting timelines:
4220 : /// - Makes sure shards are attached somewhere if they weren't already
4221 : /// - Looks up the shards and the nodes where they were most recently attached
4222 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
4223 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
4224 0 : async fn tenant_remote_mutation<R, O, F>(
4225 0 : &self,
4226 0 : tenant_id: TenantId,
4227 0 : op: O,
4228 0 : ) -> Result<R, ApiError>
4229 0 : where
4230 0 : O: FnOnce(TenantMutationLocations) -> F,
4231 0 : F: std::future::Future<Output = R>,
4232 0 : {
4233 0 : let mutation_locations = {
4234 0 : let mut locations = TenantMutationLocations::default();
4235 :
4236 : // Load the currently attached pageservers for the latest generation of each shard. This can
4237 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
4238 : // will still be the latest when we're done: we will check generations again at the end of
4239 : // this function to handle that.
4240 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
4241 :
4242 0 : if generations
4243 0 : .iter()
4244 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
4245 : {
4246 0 : let shard_generations = generations
4247 0 : .into_iter()
4248 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
4249 0 : .collect::<HashMap<_, _>>();
4250 0 :
4251 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
4252 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)
4253 0 : let locked = self.inner.read().unwrap();
4254 0 : for (shard_id, shard) in
4255 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4256 : {
4257 0 : match shard.policy {
4258 : PlacementPolicy::Attached(_) => {
4259 : // This shard is meant to be attached: the caller is not wrong to try and
4260 : // use this function, but we can't service the request right now.
4261 0 : let Some(generation) = shard_generations.get(shard_id) else {
4262 : // This can only happen if there is a split brain controller modifying the database. This should
4263 : // never happen when testing, and if it happens in production we can only log the issue.
4264 0 : debug_assert!(false);
4265 0 : tracing::error!(
4266 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
4267 : );
4268 0 : continue;
4269 : };
4270 0 : let (generation, generation_pageserver) = generation;
4271 0 : if let Some(generation) = generation {
4272 0 : if generation_pageserver.is_none() {
4273 : // This is legitimate only in a very narrow window where the shard was only just configured into
4274 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
4275 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
4276 0 : tracing::warn!(
4277 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
4278 : );
4279 0 : }
4280 : } else {
4281 : // This should never happen: a shard with no generation is only permitted when it was created in some state
4282 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
4283 0 : debug_assert!(false);
4284 0 : tracing::error!(
4285 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
4286 : );
4287 0 : continue;
4288 : }
4289 : }
4290 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
4291 0 : return Err(ApiError::Conflict(format!(
4292 0 : "Shard {shard_id} tenant has policy {:?}",
4293 0 : shard.policy
4294 0 : )));
4295 : }
4296 : }
4297 : }
4298 :
4299 0 : return Err(ApiError::ResourceUnavailable(
4300 0 : "One or more shards in tenant is not yet attached".into(),
4301 0 : ));
4302 0 : }
4303 0 :
4304 0 : let locked = self.inner.read().unwrap();
4305 : for ShardGenerationState {
4306 0 : tenant_shard_id,
4307 0 : generation,
4308 0 : generation_pageserver,
4309 0 : } in generations
4310 : {
4311 0 : let node_id = generation_pageserver.expect("We checked for None above");
4312 0 : let node = locked
4313 0 : .nodes
4314 0 : .get(&node_id)
4315 0 : .ok_or(ApiError::Conflict(format!(
4316 0 : "Raced with removal of node {node_id}"
4317 0 : )))?;
4318 0 : let generation = generation.expect("Checked above");
4319 0 :
4320 0 : let tenant = locked.tenants.get(&tenant_shard_id);
4321 :
4322 : // TODO(vlad): Abstract the logic that finds stale attached locations
4323 : // from observed state into a [`Service`] method.
4324 0 : let other_locations = match tenant {
4325 0 : Some(tenant) => {
4326 0 : let mut other = tenant.attached_locations();
4327 0 : let latest_location_index =
4328 0 : other.iter().position(|&l| l == (node.get_id(), generation));
4329 0 : if let Some(idx) = latest_location_index {
4330 0 : other.remove(idx);
4331 0 : }
4332 :
4333 0 : other
4334 : }
4335 0 : None => Vec::default(),
4336 : };
4337 :
4338 0 : let location = ShardMutationLocations {
4339 0 : latest: MutationLocation {
4340 0 : node: node.clone(),
4341 0 : generation,
4342 0 : },
4343 0 : other: other_locations
4344 0 : .into_iter()
4345 0 : .filter_map(|(node_id, generation)| {
4346 0 : let node = locked.nodes.get(&node_id)?;
4347 :
4348 0 : Some(MutationLocation {
4349 0 : node: node.clone(),
4350 0 : generation,
4351 0 : })
4352 0 : })
4353 0 : .collect(),
4354 0 : };
4355 0 : locations.0.insert(tenant_shard_id, location);
4356 0 : }
4357 :
4358 0 : locations
4359 : };
4360 :
4361 0 : let result = op(mutation_locations.clone()).await;
4362 :
4363 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
4364 : // our remote operation executed on the latest generation and is therefore persistent.
4365 : {
4366 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
4367 0 : if latest_generations
4368 0 : .into_iter()
4369 0 : .map(
4370 0 : |ShardGenerationState {
4371 : tenant_shard_id,
4372 : generation,
4373 : generation_pageserver: _,
4374 0 : }| (tenant_shard_id, generation),
4375 0 : )
4376 0 : .collect::<Vec<_>>()
4377 0 : != mutation_locations
4378 0 : .0
4379 0 : .into_iter()
4380 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
4381 0 : .collect::<Vec<_>>()
4382 : {
4383 : // We raced with something that incremented the generation, and therefore cannot be
4384 : // confident that our actions are persistent (they might have hit an old generation).
4385 : //
4386 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
4387 0 : return Err(ApiError::ResourceUnavailable(
4388 0 : "Tenant attachment changed, please retry".into(),
4389 0 : ));
4390 0 : }
4391 0 : }
4392 0 :
4393 0 : Ok(result)
4394 0 : }
4395 :
4396 0 : pub(crate) async fn tenant_timeline_delete(
4397 0 : self: &Arc<Self>,
4398 0 : tenant_id: TenantId,
4399 0 : timeline_id: TimelineId,
4400 0 : ) -> Result<StatusCode, ApiError> {
4401 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
4402 0 : let _tenant_lock = trace_shared_lock(
4403 0 : &self.tenant_op_locks,
4404 0 : tenant_id,
4405 0 : TenantOperations::TimelineDelete,
4406 0 : )
4407 0 : .await;
4408 :
4409 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
4410 0 : if targets.0.is_empty() {
4411 0 : return Err(ApiError::NotFound(
4412 0 : anyhow::anyhow!("Tenant not found").into(),
4413 0 : ));
4414 0 : }
4415 0 :
4416 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
4417 0 : assert!(shard_zero_tid.is_shard_zero());
4418 :
4419 0 : async fn delete_one(
4420 0 : tenant_shard_id: TenantShardId,
4421 0 : timeline_id: TimelineId,
4422 0 : node: Node,
4423 0 : http_client: reqwest::Client,
4424 0 : jwt: Option<String>,
4425 0 : ) -> Result<StatusCode, ApiError> {
4426 0 : tracing::info!(
4427 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4428 : );
4429 :
4430 0 : let client = PageserverClient::new(node.get_id(), http_client, node.base_url(), jwt.as_deref());
4431 0 : let res = client
4432 0 : .timeline_delete(tenant_shard_id, timeline_id)
4433 0 : .await;
4434 :
4435 0 : match res {
4436 0 : Ok(ok) => Ok(ok),
4437 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
4438 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
4439 0 : Err(e) => {
4440 0 : Err(
4441 0 : ApiError::InternalServerError(anyhow::anyhow!(
4442 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
4443 0 : ))
4444 0 : )
4445 : }
4446 : }
4447 0 : }
4448 :
4449 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4450 0 : let statuses = self
4451 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
4452 0 : Box::pin(delete_one(
4453 0 : tenant_shard_id,
4454 0 : timeline_id,
4455 0 : node,
4456 0 : self.http_client.clone(),
4457 0 : self.config.pageserver_jwt_token.clone(),
4458 0 : ))
4459 0 : })
4460 0 : .await?;
4461 :
4462 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
4463 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
4464 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
4465 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
4466 0 : return Ok(StatusCode::CONFLICT);
4467 0 : }
4468 0 :
4469 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
4470 0 : return Ok(StatusCode::ACCEPTED);
4471 0 : }
4472 :
4473 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
4474 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
4475 0 : let shard_zero_status = delete_one(
4476 0 : shard_zero_tid,
4477 0 : timeline_id,
4478 0 : shard_zero_locations.latest.node,
4479 0 : self.http_client.clone(),
4480 0 : self.config.pageserver_jwt_token.clone(),
4481 0 : )
4482 0 : .await?;
4483 0 : Ok(shard_zero_status)
4484 0 : }).await?;
4485 :
4486 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
4487 0 : .await?;
4488 :
4489 0 : status_code
4490 0 : }
4491 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
4492 0 : pub(crate) async fn tenant_shard0_node(
4493 0 : &self,
4494 0 : tenant_id: TenantId,
4495 0 : ) -> Result<(Node, TenantShardId), ApiError> {
4496 0 : let tenant_shard_id = {
4497 0 : let locked = self.inner.read().unwrap();
4498 0 : let Some((tenant_shard_id, _shard)) = locked
4499 0 : .tenants
4500 0 : .range(TenantShardId::tenant_range(tenant_id))
4501 0 : .next()
4502 : else {
4503 0 : return Err(ApiError::NotFound(
4504 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
4505 0 : ));
4506 : };
4507 :
4508 0 : *tenant_shard_id
4509 0 : };
4510 0 :
4511 0 : self.tenant_shard_node(tenant_shard_id)
4512 0 : .await
4513 0 : .map(|node| (node, tenant_shard_id))
4514 0 : }
4515 :
4516 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
4517 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
4518 0 : pub(crate) async fn tenant_shard_node(
4519 0 : &self,
4520 0 : tenant_shard_id: TenantShardId,
4521 0 : ) -> Result<Node, ApiError> {
4522 0 : // Look up in-memory state and maybe use the node from there.
4523 0 : {
4524 0 : let locked = self.inner.read().unwrap();
4525 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
4526 0 : return Err(ApiError::NotFound(
4527 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
4528 0 : ));
4529 : };
4530 :
4531 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
4532 0 : tracing::warn!(
4533 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
4534 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
4535 : shard.policy
4536 : );
4537 0 : return Err(ApiError::Conflict(
4538 0 : "Cannot call timeline API on non-attached tenant".to_string(),
4539 0 : ));
4540 : };
4541 :
4542 0 : if shard.reconciler.is_none() {
4543 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
4544 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
4545 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
4546 : // This should never happen
4547 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4548 0 : "Shard refers to nonexistent node"
4549 0 : )));
4550 : };
4551 0 : return Ok(node.clone());
4552 0 : }
4553 : };
4554 :
4555 : // Look up the latest attached pageserver location from the database
4556 : // generation state: this will reflect the progress of any ongoing migration.
4557 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4558 : // the case where they call through to the pageserver and get a 404.
4559 0 : let db_result = self
4560 0 : .persistence
4561 0 : .tenant_generations(tenant_shard_id.tenant_id)
4562 0 : .await?;
4563 : let Some(ShardGenerationState {
4564 : tenant_shard_id: _,
4565 : generation: _,
4566 0 : generation_pageserver: Some(node_id),
4567 0 : }) = db_result
4568 0 : .into_iter()
4569 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
4570 : else {
4571 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4572 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4573 : // or a conflict response (case where tenant was detached in background)
4574 0 : return Err(ApiError::ResourceUnavailable(
4575 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
4576 0 : ));
4577 : };
4578 0 : let locked = self.inner.read().unwrap();
4579 0 : let Some(node) = locked.nodes.get(&node_id) else {
4580 : // This should never happen
4581 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4582 0 : "Shard refers to nonexistent node"
4583 0 : )));
4584 : };
4585 :
4586 0 : Ok(node.clone())
4587 0 : }
4588 :
4589 0 : pub(crate) fn tenant_locate(
4590 0 : &self,
4591 0 : tenant_id: TenantId,
4592 0 : ) -> Result<TenantLocateResponse, ApiError> {
4593 0 : let locked = self.inner.read().unwrap();
4594 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4595 :
4596 0 : let mut result = Vec::new();
4597 0 : let mut shard_params: Option<ShardParameters> = None;
4598 :
4599 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4600 : {
4601 0 : let node_id =
4602 0 : shard
4603 0 : .intent
4604 0 : .get_attached()
4605 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4606 0 : "Cannot locate a tenant that is not attached"
4607 0 : )))?;
4608 :
4609 0 : let node = locked
4610 0 : .nodes
4611 0 : .get(&node_id)
4612 0 : .expect("Pageservers may not be deleted while referenced");
4613 0 :
4614 0 : result.push(node.shard_location(*tenant_shard_id));
4615 0 :
4616 0 : match &shard_params {
4617 0 : None => {
4618 0 : shard_params = Some(ShardParameters {
4619 0 : stripe_size: shard.shard.stripe_size,
4620 0 : count: shard.shard.count,
4621 0 : });
4622 0 : }
4623 0 : Some(params) => {
4624 0 : if params.stripe_size != shard.shard.stripe_size {
4625 : // This should never happen. We enforce at runtime because it's simpler than
4626 : // adding an extra per-tenant data structure to store the things that should be the same
4627 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4628 0 : "Inconsistent shard stripe size parameters!"
4629 0 : )));
4630 0 : }
4631 : }
4632 : }
4633 : }
4634 :
4635 0 : if result.is_empty() {
4636 0 : return Err(ApiError::NotFound(
4637 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4638 0 : ));
4639 0 : }
4640 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4641 0 : tracing::info!(
4642 0 : "Located tenant {} with params {:?} on shards {}",
4643 0 : tenant_id,
4644 0 : shard_params,
4645 0 : result
4646 0 : .iter()
4647 0 : .map(|s| format!("{:?}", s))
4648 0 : .collect::<Vec<_>>()
4649 0 : .join(",")
4650 : );
4651 :
4652 0 : Ok(TenantLocateResponse {
4653 0 : shards: result,
4654 0 : shard_params,
4655 0 : })
4656 0 : }
4657 :
4658 : /// Returns None if the input iterator of shards does not include a shard with number=0
4659 0 : fn tenant_describe_impl<'a>(
4660 0 : &self,
4661 0 : shards: impl Iterator<Item = &'a TenantShard>,
4662 0 : ) -> Option<TenantDescribeResponse> {
4663 0 : let mut shard_zero = None;
4664 0 : let mut describe_shards = Vec::new();
4665 :
4666 0 : for shard in shards {
4667 0 : if shard.tenant_shard_id.is_shard_zero() {
4668 0 : shard_zero = Some(shard);
4669 0 : }
4670 :
4671 0 : describe_shards.push(TenantDescribeResponseShard {
4672 0 : tenant_shard_id: shard.tenant_shard_id,
4673 0 : node_attached: *shard.intent.get_attached(),
4674 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4675 0 : last_error: shard
4676 0 : .last_error
4677 0 : .lock()
4678 0 : .unwrap()
4679 0 : .as_ref()
4680 0 : .map(|e| format!("{e}"))
4681 0 : .unwrap_or("".to_string())
4682 0 : .clone(),
4683 0 : is_reconciling: shard.reconciler.is_some(),
4684 0 : is_pending_compute_notification: shard.pending_compute_notification,
4685 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4686 0 : scheduling_policy: shard.get_scheduling_policy(),
4687 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4688 : })
4689 : }
4690 :
4691 0 : let shard_zero = shard_zero?;
4692 :
4693 0 : Some(TenantDescribeResponse {
4694 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
4695 0 : shards: describe_shards,
4696 0 : stripe_size: shard_zero.shard.stripe_size,
4697 0 : policy: shard_zero.policy.clone(),
4698 0 : config: shard_zero.config.clone(),
4699 0 : })
4700 0 : }
4701 :
4702 0 : pub(crate) fn tenant_describe(
4703 0 : &self,
4704 0 : tenant_id: TenantId,
4705 0 : ) -> Result<TenantDescribeResponse, ApiError> {
4706 0 : let locked = self.inner.read().unwrap();
4707 0 :
4708 0 : self.tenant_describe_impl(
4709 0 : locked
4710 0 : .tenants
4711 0 : .range(TenantShardId::tenant_range(tenant_id))
4712 0 : .map(|(_k, v)| v),
4713 0 : )
4714 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
4715 0 : }
4716 :
4717 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
4718 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
4719 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
4720 : /// in our external API.
4721 0 : pub(crate) fn tenant_list(
4722 0 : &self,
4723 0 : limit: Option<usize>,
4724 0 : start_after: Option<TenantId>,
4725 0 : ) -> Vec<TenantDescribeResponse> {
4726 0 : let locked = self.inner.read().unwrap();
4727 :
4728 : // Apply start_from parameter
4729 0 : let shard_range = match start_after {
4730 0 : None => locked.tenants.range(..),
4731 0 : Some(tenant_id) => locked.tenants.range(
4732 0 : TenantShardId {
4733 0 : tenant_id,
4734 0 : shard_number: ShardNumber(u8::MAX),
4735 0 : shard_count: ShardCount(u8::MAX),
4736 0 : }..,
4737 0 : ),
4738 : };
4739 :
4740 0 : let mut result = Vec::new();
4741 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
4742 0 : result.push(
4743 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
4744 0 : .expect("Groups are always non-empty"),
4745 0 : );
4746 :
4747 : // Enforce `limit` parameter
4748 0 : if let Some(limit) = limit {
4749 0 : if result.len() >= limit {
4750 0 : break;
4751 0 : }
4752 0 : }
4753 : }
4754 :
4755 0 : result
4756 0 : }
4757 :
4758 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
4759 : async fn abort_tenant_shard_split(
4760 : &self,
4761 : op: &TenantShardSplitAbort,
4762 : ) -> Result<(), TenantShardSplitAbortError> {
4763 : // Cleaning up a split:
4764 : // - Parent shards are not destroyed during a split, just detached.
4765 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
4766 : // just the children attached, or both.
4767 : //
4768 : // Therefore our work to do is to:
4769 : // 1. Clean up storage controller's internal state to just refer to parents, no children
4770 : // 2. Call out to pageservers to ensure that children are detached
4771 : // 3. Call out to pageservers to ensure that parents are attached.
4772 : //
4773 : // Crash safety:
4774 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
4775 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
4776 : // and detach them.
4777 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
4778 : // from our database, then we will re-enter this cleanup routine on startup.
4779 :
4780 : let TenantShardSplitAbort {
4781 : tenant_id,
4782 : new_shard_count,
4783 : new_stripe_size,
4784 : ..
4785 : } = op;
4786 :
4787 : // First abort persistent state, if any exists.
4788 : match self
4789 : .persistence
4790 : .abort_shard_split(*tenant_id, *new_shard_count)
4791 : .await?
4792 : {
4793 : AbortShardSplitStatus::Aborted => {
4794 : // Proceed to roll back any child shards created on pageservers
4795 : }
4796 : AbortShardSplitStatus::Complete => {
4797 : // The split completed (we might hit that path if e.g. our database transaction
4798 : // to write the completion landed in the database, but we dropped connection
4799 : // before seeing the result).
4800 : //
4801 : // We must update in-memory state to reflect the successful split.
4802 : self.tenant_shard_split_commit_inmem(
4803 : *tenant_id,
4804 : *new_shard_count,
4805 : *new_stripe_size,
4806 : );
4807 : return Ok(());
4808 : }
4809 : }
4810 :
4811 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
4812 : let detach_locations: Vec<(Node, TenantShardId)> = {
4813 : let mut detach_locations = Vec::new();
4814 : let mut locked = self.inner.write().unwrap();
4815 : let (nodes, tenants, scheduler) = locked.parts_mut();
4816 :
4817 : for (tenant_shard_id, shard) in
4818 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
4819 : {
4820 : if shard.shard.count == op.new_shard_count {
4821 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
4822 : // is infallible, so if we got an error we shouldn't have got that far.
4823 : tracing::warn!(
4824 : "During split abort, child shard {tenant_shard_id} found in-memory"
4825 : );
4826 : continue;
4827 : }
4828 :
4829 : // Add the children of this shard to this list of things to detach
4830 : if let Some(node_id) = shard.intent.get_attached() {
4831 : for child_id in tenant_shard_id.split(*new_shard_count) {
4832 : detach_locations.push((
4833 : nodes
4834 : .get(node_id)
4835 : .expect("Intent references nonexistent node")
4836 : .clone(),
4837 : child_id,
4838 : ));
4839 : }
4840 : } else {
4841 : tracing::warn!(
4842 : "During split abort, shard {tenant_shard_id} has no attached location"
4843 : );
4844 : }
4845 :
4846 : tracing::info!("Restoring parent shard {tenant_shard_id}");
4847 :
4848 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
4849 : // if the original attachment location is offline.
4850 : if let Some(node_id) = shard.intent.get_attached() {
4851 : if !nodes.get(node_id).unwrap().is_available() {
4852 : tracing::info!(
4853 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
4854 : );
4855 : shard.intent.demote_attached(scheduler, *node_id);
4856 : }
4857 : }
4858 : for node_id in shard.intent.get_secondary().clone() {
4859 : if !nodes.get(&node_id).unwrap().is_available() {
4860 : tracing::info!(
4861 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
4862 : );
4863 : shard.intent.remove_secondary(scheduler, node_id);
4864 : }
4865 : }
4866 :
4867 : shard.splitting = SplitState::Idle;
4868 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
4869 : // If this shard can't be scheduled now (perhaps due to offline nodes or
4870 : // capacity issues), that must not prevent us rolling back a split. In this
4871 : // case it should be eventually scheduled in the background.
4872 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
4873 : }
4874 :
4875 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
4876 : }
4877 :
4878 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
4879 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
4880 :
4881 : detach_locations
4882 : };
4883 :
4884 : for (node, child_id) in detach_locations {
4885 : if !node.is_available() {
4886 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
4887 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
4888 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
4889 : // them from the node.
4890 : tracing::warn!(
4891 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
4892 : );
4893 : continue;
4894 : }
4895 :
4896 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
4897 : // a 503 and retry, up to our limit.
4898 : tracing::info!("Detaching {child_id} on {node}...");
4899 : match node
4900 : .with_client_retries(
4901 0 : |client| async move {
4902 0 : let config = LocationConfig {
4903 0 : mode: LocationConfigMode::Detached,
4904 0 : generation: None,
4905 0 : secondary_conf: None,
4906 0 : shard_number: child_id.shard_number.0,
4907 0 : shard_count: child_id.shard_count.literal(),
4908 0 : // Stripe size and tenant config don't matter when detaching
4909 0 : shard_stripe_size: 0,
4910 0 : tenant_conf: TenantConfig::default(),
4911 0 : };
4912 0 :
4913 0 : client.location_config(child_id, config, None, false).await
4914 0 : },
4915 : &self.http_client,
4916 : &self.config.pageserver_jwt_token,
4917 : 1,
4918 : 10,
4919 : Duration::from_secs(5),
4920 : &self.reconcilers_cancel,
4921 : )
4922 : .await
4923 : {
4924 : Some(Ok(_)) => {}
4925 : Some(Err(e)) => {
4926 : // We failed to communicate with the remote node. This is problematic: we may be
4927 : // leaving it with a rogue child shard.
4928 : tracing::warn!(
4929 : "Failed to detach child {child_id} from node {node} during abort"
4930 : );
4931 : return Err(e.into());
4932 : }
4933 : None => {
4934 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
4935 : // clean up on restart. The node going offline requires a retry.
4936 : return Err(TenantShardSplitAbortError::Unavailable);
4937 : }
4938 : };
4939 : }
4940 :
4941 : tracing::info!("Successfully aborted split");
4942 : Ok(())
4943 : }
4944 :
4945 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
4946 : /// of the tenant map to reflect the child shards that exist after the split.
4947 0 : fn tenant_shard_split_commit_inmem(
4948 0 : &self,
4949 0 : tenant_id: TenantId,
4950 0 : new_shard_count: ShardCount,
4951 0 : new_stripe_size: Option<ShardStripeSize>,
4952 0 : ) -> (
4953 0 : TenantShardSplitResponse,
4954 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
4955 0 : Vec<ReconcilerWaiter>,
4956 0 : ) {
4957 0 : let mut response = TenantShardSplitResponse {
4958 0 : new_shards: Vec::new(),
4959 0 : };
4960 0 : let mut child_locations = Vec::new();
4961 0 : let mut waiters = Vec::new();
4962 0 :
4963 0 : {
4964 0 : let mut locked = self.inner.write().unwrap();
4965 0 :
4966 0 : let parent_ids = locked
4967 0 : .tenants
4968 0 : .range(TenantShardId::tenant_range(tenant_id))
4969 0 : .map(|(shard_id, _)| *shard_id)
4970 0 : .collect::<Vec<_>>();
4971 0 :
4972 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4973 0 : for parent_id in parent_ids {
4974 0 : let child_ids = parent_id.split(new_shard_count);
4975 :
4976 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
4977 0 : let mut old_state = tenants
4978 0 : .remove(&parent_id)
4979 0 : .expect("It was present, we just split it");
4980 0 :
4981 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
4982 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
4983 0 : // nothing else can clear this.
4984 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
4985 :
4986 0 : let old_attached = old_state.intent.get_attached().unwrap();
4987 0 : old_state.intent.clear(scheduler);
4988 0 : let generation = old_state.generation.expect("Shard must have been attached");
4989 0 : (
4990 0 : old_attached,
4991 0 : generation,
4992 0 : old_state.policy.clone(),
4993 0 : old_state.shard,
4994 0 : old_state.config.clone(),
4995 0 : old_state.preferred_az().cloned(),
4996 0 : )
4997 0 : };
4998 0 :
4999 0 : let mut schedule_context = ScheduleContext::default();
5000 0 : for child in child_ids {
5001 0 : let mut child_shard = parent_ident;
5002 0 : child_shard.number = child.shard_number;
5003 0 : child_shard.count = child.shard_count;
5004 0 : if let Some(stripe_size) = new_stripe_size {
5005 0 : child_shard.stripe_size = stripe_size;
5006 0 : }
5007 :
5008 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5009 0 : child_observed.insert(
5010 0 : pageserver,
5011 0 : ObservedStateLocation {
5012 0 : conf: Some(attached_location_conf(
5013 0 : generation,
5014 0 : &child_shard,
5015 0 : &config,
5016 0 : &policy,
5017 0 : )),
5018 0 : },
5019 0 : );
5020 0 :
5021 0 : let mut child_state =
5022 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
5023 0 : child_state.intent =
5024 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
5025 0 : child_state.observed = ObservedState {
5026 0 : locations: child_observed,
5027 0 : };
5028 0 : child_state.generation = Some(generation);
5029 0 : child_state.config = config.clone();
5030 0 :
5031 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
5032 0 : // as at this point in the split process we have succeeded and this part is infallible:
5033 0 : // we will never need to do any special recovery from this state.
5034 0 :
5035 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
5036 :
5037 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
5038 : // This is not fatal, because we've implicitly already got an attached
5039 : // location for the child shard. Failure here just means we couldn't
5040 : // find a secondary (e.g. because cluster is overloaded).
5041 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
5042 0 : }
5043 : // In the background, attach secondary locations for the new shards
5044 0 : if let Some(waiter) = self.maybe_reconcile_shard(
5045 0 : &mut child_state,
5046 0 : nodes,
5047 0 : ReconcilerPriority::High,
5048 0 : ) {
5049 0 : waiters.push(waiter);
5050 0 : }
5051 :
5052 0 : tenants.insert(child, child_state);
5053 0 : response.new_shards.push(child);
5054 : }
5055 : }
5056 0 : (response, child_locations, waiters)
5057 0 : }
5058 0 : }
5059 :
5060 0 : async fn tenant_shard_split_start_secondaries(
5061 0 : &self,
5062 0 : tenant_id: TenantId,
5063 0 : waiters: Vec<ReconcilerWaiter>,
5064 0 : ) {
5065 : // Wait for initial reconcile of child shards, this creates the secondary locations
5066 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5067 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
5068 : // their secondaries couldn't be attached.
5069 0 : tracing::warn!("Failed to reconcile after split: {e}");
5070 0 : return;
5071 0 : }
5072 :
5073 : // Take the state lock to discover the attached & secondary intents for all shards
5074 0 : let (attached, secondary) = {
5075 0 : let locked = self.inner.read().unwrap();
5076 0 : let mut attached = Vec::new();
5077 0 : let mut secondary = Vec::new();
5078 :
5079 0 : for (tenant_shard_id, shard) in
5080 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5081 : {
5082 0 : let Some(node_id) = shard.intent.get_attached() else {
5083 : // Unexpected. Race with a PlacementPolicy change?
5084 0 : tracing::warn!(
5085 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
5086 : );
5087 0 : continue;
5088 : };
5089 :
5090 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
5091 : // No secondary location. Nothing for us to do.
5092 0 : continue;
5093 : };
5094 :
5095 0 : let attached_node = locked
5096 0 : .nodes
5097 0 : .get(node_id)
5098 0 : .expect("Pageservers may not be deleted while referenced");
5099 0 :
5100 0 : let secondary_node = locked
5101 0 : .nodes
5102 0 : .get(secondary_node_id)
5103 0 : .expect("Pageservers may not be deleted while referenced");
5104 0 :
5105 0 : attached.push((*tenant_shard_id, attached_node.clone()));
5106 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
5107 : }
5108 0 : (attached, secondary)
5109 0 : };
5110 0 :
5111 0 : if secondary.is_empty() {
5112 : // No secondary locations; nothing for us to do
5113 0 : return;
5114 0 : }
5115 :
5116 0 : for result in self
5117 0 : .tenant_for_shards_api(
5118 0 : attached,
5119 0 : |tenant_shard_id, client| async move {
5120 0 : client.tenant_heatmap_upload(tenant_shard_id).await
5121 0 : },
5122 0 : 1,
5123 0 : 1,
5124 0 : SHORT_RECONCILE_TIMEOUT,
5125 0 : &self.cancel,
5126 0 : )
5127 0 : .await
5128 : {
5129 0 : if let Err(e) = result {
5130 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
5131 0 : return;
5132 0 : }
5133 : }
5134 :
5135 0 : for result in self
5136 0 : .tenant_for_shards_api(
5137 0 : secondary,
5138 0 : |tenant_shard_id, client| async move {
5139 0 : client
5140 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
5141 0 : .await
5142 0 : },
5143 0 : 1,
5144 0 : 1,
5145 0 : SHORT_RECONCILE_TIMEOUT,
5146 0 : &self.cancel,
5147 0 : )
5148 0 : .await
5149 : {
5150 0 : if let Err(e) = result {
5151 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
5152 0 : return;
5153 0 : }
5154 : }
5155 0 : }
5156 :
5157 0 : pub(crate) async fn tenant_shard_split(
5158 0 : &self,
5159 0 : tenant_id: TenantId,
5160 0 : split_req: TenantShardSplitRequest,
5161 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
5162 : // TODO: return 503 if we get stuck waiting for this lock
5163 : // (issue https://github.com/neondatabase/neon/issues/7108)
5164 0 : let _tenant_lock = trace_exclusive_lock(
5165 0 : &self.tenant_op_locks,
5166 0 : tenant_id,
5167 0 : TenantOperations::ShardSplit,
5168 0 : )
5169 0 : .await;
5170 :
5171 0 : let _gate = self
5172 0 : .reconcilers_gate
5173 0 : .enter()
5174 0 : .map_err(|_| ApiError::ShuttingDown)?;
5175 :
5176 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
5177 0 : let new_stripe_size = split_req.new_stripe_size;
5178 :
5179 : // Validate the request and construct parameters. This phase is fallible, but does not require
5180 : // rollback on errors, as it does no I/O and mutates no state.
5181 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
5182 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
5183 0 : ShardSplitAction::Split(params) => params,
5184 : };
5185 :
5186 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
5187 : // we must roll back.
5188 0 : let r = self
5189 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
5190 0 : .await;
5191 :
5192 0 : let (response, waiters) = match r {
5193 0 : Ok(r) => r,
5194 0 : Err(e) => {
5195 0 : // Split might be part-done, we must do work to abort it.
5196 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
5197 0 : self.abort_tx
5198 0 : .send(TenantShardSplitAbort {
5199 0 : tenant_id,
5200 0 : new_shard_count,
5201 0 : new_stripe_size,
5202 0 : _tenant_lock,
5203 0 : _gate,
5204 0 : })
5205 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
5206 0 : .ok();
5207 0 : return Err(e);
5208 : }
5209 : };
5210 :
5211 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
5212 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
5213 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
5214 : // in [`Self::optimize_all`]
5215 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
5216 0 : .await;
5217 0 : Ok(response)
5218 0 : }
5219 :
5220 0 : fn prepare_tenant_shard_split(
5221 0 : &self,
5222 0 : tenant_id: TenantId,
5223 0 : split_req: TenantShardSplitRequest,
5224 0 : ) -> Result<ShardSplitAction, ApiError> {
5225 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
5226 0 : anyhow::anyhow!("failpoint")
5227 0 : )));
5228 :
5229 0 : let mut policy = None;
5230 0 : let mut config = None;
5231 0 : let mut shard_ident = None;
5232 0 : let mut preferred_az_id = None;
5233 : // Validate input, and calculate which shards we will create
5234 0 : let (old_shard_count, targets) =
5235 : {
5236 0 : let locked = self.inner.read().unwrap();
5237 0 :
5238 0 : let pageservers = locked.nodes.clone();
5239 0 :
5240 0 : let mut targets = Vec::new();
5241 0 :
5242 0 : // In case this is a retry, count how many already-split shards we found
5243 0 : let mut children_found = Vec::new();
5244 0 : let mut old_shard_count = None;
5245 :
5246 0 : for (tenant_shard_id, shard) in
5247 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5248 : {
5249 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
5250 : Ordering::Equal => {
5251 : // Already split this
5252 0 : children_found.push(*tenant_shard_id);
5253 0 : continue;
5254 : }
5255 : Ordering::Greater => {
5256 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5257 0 : "Requested count {} but already have shards at count {}",
5258 0 : split_req.new_shard_count,
5259 0 : shard.shard.count.count()
5260 0 : )));
5261 : }
5262 0 : Ordering::Less => {
5263 0 : // Fall through: this shard has lower count than requested,
5264 0 : // is a candidate for splitting.
5265 0 : }
5266 0 : }
5267 0 :
5268 0 : match old_shard_count {
5269 0 : None => old_shard_count = Some(shard.shard.count),
5270 0 : Some(old_shard_count) => {
5271 0 : if old_shard_count != shard.shard.count {
5272 : // We may hit this case if a caller asked for two splits to
5273 : // different sizes, before the first one is complete.
5274 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
5275 : // of shard_count=1 and shard_count=2 shards in the map.
5276 0 : return Err(ApiError::Conflict(
5277 0 : "Cannot split, currently mid-split".to_string(),
5278 0 : ));
5279 0 : }
5280 : }
5281 : }
5282 0 : if policy.is_none() {
5283 0 : policy = Some(shard.policy.clone());
5284 0 : }
5285 0 : if shard_ident.is_none() {
5286 0 : shard_ident = Some(shard.shard);
5287 0 : }
5288 0 : if config.is_none() {
5289 0 : config = Some(shard.config.clone());
5290 0 : }
5291 0 : if preferred_az_id.is_none() {
5292 0 : preferred_az_id = shard.preferred_az().cloned();
5293 0 : }
5294 :
5295 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
5296 0 : tracing::info!(
5297 0 : "Tenant shard {} already has shard count {}",
5298 : tenant_shard_id,
5299 : split_req.new_shard_count
5300 : );
5301 0 : continue;
5302 0 : }
5303 :
5304 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
5305 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
5306 0 : ))?;
5307 :
5308 0 : let node = pageservers
5309 0 : .get(&node_id)
5310 0 : .expect("Pageservers may not be deleted while referenced");
5311 0 :
5312 0 : targets.push(ShardSplitTarget {
5313 0 : parent_id: *tenant_shard_id,
5314 0 : node: node.clone(),
5315 0 : child_ids: tenant_shard_id
5316 0 : .split(ShardCount::new(split_req.new_shard_count)),
5317 0 : });
5318 : }
5319 :
5320 0 : if targets.is_empty() {
5321 0 : if children_found.len() == split_req.new_shard_count as usize {
5322 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
5323 0 : new_shards: children_found,
5324 0 : }));
5325 : } else {
5326 : // No shards found to split, and no existing children found: the
5327 : // tenant doesn't exist at all.
5328 0 : return Err(ApiError::NotFound(
5329 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
5330 0 : ));
5331 : }
5332 0 : }
5333 0 :
5334 0 : (old_shard_count, targets)
5335 0 : };
5336 0 :
5337 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
5338 0 : let old_shard_count = old_shard_count.unwrap();
5339 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
5340 : // This ShardIdentity will be used as the template for all children, so this implicitly
5341 : // applies the new stripe size to the children.
5342 0 : let mut shard_ident = shard_ident.unwrap();
5343 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
5344 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5345 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
5346 0 : shard_ident.stripe_size
5347 0 : )));
5348 0 : }
5349 0 :
5350 0 : shard_ident.stripe_size = new_stripe_size;
5351 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
5352 0 : shard_ident
5353 : } else {
5354 0 : shard_ident.unwrap()
5355 : };
5356 0 : let policy = policy.unwrap();
5357 0 : let config = config.unwrap();
5358 0 :
5359 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
5360 0 : old_shard_count,
5361 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
5362 0 : new_stripe_size: split_req.new_stripe_size,
5363 0 : targets,
5364 0 : policy,
5365 0 : config,
5366 0 : shard_ident,
5367 0 : preferred_az_id,
5368 0 : })))
5369 0 : }
5370 :
5371 0 : async fn do_tenant_shard_split(
5372 0 : &self,
5373 0 : tenant_id: TenantId,
5374 0 : params: Box<ShardSplitParams>,
5375 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
5376 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
5377 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
5378 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
5379 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
5380 0 : // (https://github.com/neondatabase/neon/issues/6676)
5381 0 :
5382 0 : let ShardSplitParams {
5383 0 : old_shard_count,
5384 0 : new_shard_count,
5385 0 : new_stripe_size,
5386 0 : mut targets,
5387 0 : policy,
5388 0 : config,
5389 0 : shard_ident,
5390 0 : preferred_az_id,
5391 0 : } = *params;
5392 :
5393 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
5394 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
5395 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
5396 : // at the time of split.
5397 0 : let waiters = {
5398 0 : let mut locked = self.inner.write().unwrap();
5399 0 : let mut waiters = Vec::new();
5400 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5401 0 : for target in &mut targets {
5402 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
5403 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5404 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5405 0 : "Shard {} not found",
5406 0 : target.parent_id
5407 0 : )));
5408 : };
5409 :
5410 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
5411 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5412 0 : return Err(ApiError::Conflict(format!(
5413 0 : "Shard {} unexpectedly rescheduled during split",
5414 0 : target.parent_id
5415 0 : )));
5416 0 : }
5417 0 :
5418 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
5419 0 : shard.intent.clear_secondary(scheduler);
5420 :
5421 : // Run Reconciler to execute detach fo secondary locations.
5422 0 : if let Some(waiter) =
5423 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5424 0 : {
5425 0 : waiters.push(waiter);
5426 0 : }
5427 : }
5428 0 : waiters
5429 0 : };
5430 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
5431 :
5432 : // Before creating any new child shards in memory or on the pageservers, persist them: this
5433 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
5434 : // acts as the protection against two concurrent attempts to split: one of them will get a database
5435 : // error trying to insert the child shards.
5436 0 : let mut child_tsps = Vec::new();
5437 0 : for target in &targets {
5438 0 : let mut this_child_tsps = Vec::new();
5439 0 : for child in &target.child_ids {
5440 0 : let mut child_shard = shard_ident;
5441 0 : child_shard.number = child.shard_number;
5442 0 : child_shard.count = child.shard_count;
5443 0 :
5444 0 : tracing::info!(
5445 0 : "Create child shard persistence with stripe size {}",
5446 : shard_ident.stripe_size.0
5447 : );
5448 :
5449 0 : this_child_tsps.push(TenantShardPersistence {
5450 0 : tenant_id: child.tenant_id.to_string(),
5451 0 : shard_number: child.shard_number.0 as i32,
5452 0 : shard_count: child.shard_count.literal() as i32,
5453 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
5454 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
5455 0 : // populate the correct generation as part of its transaction, to protect us
5456 0 : // against racing with changes in the state of the parent.
5457 0 : generation: None,
5458 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
5459 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
5460 0 : config: serde_json::to_string(&config).unwrap(),
5461 0 : splitting: SplitState::Splitting,
5462 0 :
5463 0 : // Scheduling policies and preferred AZ do not carry through to children
5464 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
5465 0 : .unwrap(),
5466 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
5467 0 : });
5468 0 : }
5469 :
5470 0 : child_tsps.push((target.parent_id, this_child_tsps));
5471 : }
5472 :
5473 0 : if let Err(e) = self
5474 0 : .persistence
5475 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
5476 0 : .await
5477 : {
5478 0 : match e {
5479 : DatabaseError::Query(diesel::result::Error::DatabaseError(
5480 : DatabaseErrorKind::UniqueViolation,
5481 : _,
5482 : )) => {
5483 : // Inserting a child shard violated a unique constraint: we raced with another call to
5484 : // this function
5485 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
5486 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
5487 : }
5488 0 : _ => return Err(ApiError::InternalServerError(e.into())),
5489 : }
5490 0 : }
5491 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
5492 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5493 0 : ));
5494 :
5495 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
5496 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
5497 : // is not set in memory, then it was not persisted.
5498 : {
5499 0 : let mut locked = self.inner.write().unwrap();
5500 0 : for target in &targets {
5501 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
5502 0 : parent_shard.splitting = SplitState::Splitting;
5503 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
5504 0 : // split operation.
5505 0 : parent_shard
5506 0 : .observed
5507 0 : .locations
5508 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
5509 0 : }
5510 : }
5511 : }
5512 :
5513 : // TODO: issue split calls concurrently (this only matters once we're splitting
5514 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
5515 :
5516 0 : for target in &targets {
5517 : let ShardSplitTarget {
5518 0 : parent_id,
5519 0 : node,
5520 0 : child_ids,
5521 0 : } = target;
5522 0 : let client = PageserverClient::new(
5523 0 : node.get_id(),
5524 0 : self.http_client.clone(),
5525 0 : node.base_url(),
5526 0 : self.config.pageserver_jwt_token.as_deref(),
5527 0 : );
5528 0 : let response = client
5529 0 : .tenant_shard_split(
5530 0 : *parent_id,
5531 0 : TenantShardSplitRequest {
5532 0 : new_shard_count: new_shard_count.literal(),
5533 0 : new_stripe_size,
5534 0 : },
5535 0 : )
5536 0 : .await
5537 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
5538 :
5539 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
5540 0 : "failpoint".to_string()
5541 0 : )));
5542 :
5543 0 : failpoint_support::sleep_millis_async!(
5544 : "shard-split-post-remote-sleep",
5545 0 : &self.reconcilers_cancel
5546 : );
5547 :
5548 0 : tracing::info!(
5549 0 : "Split {} into {}",
5550 0 : parent_id,
5551 0 : response
5552 0 : .new_shards
5553 0 : .iter()
5554 0 : .map(|s| format!("{:?}", s))
5555 0 : .collect::<Vec<_>>()
5556 0 : .join(",")
5557 : );
5558 :
5559 0 : if &response.new_shards != child_ids {
5560 : // This should never happen: the pageserver should agree with us on how shard splits work.
5561 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5562 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
5563 0 : parent_id,
5564 0 : response.new_shards,
5565 0 : child_ids
5566 0 : )));
5567 0 : }
5568 : }
5569 :
5570 0 : pausable_failpoint!("shard-split-pre-complete");
5571 :
5572 : // TODO: if the pageserver restarted concurrently with our split API call,
5573 : // the actual generation of the child shard might differ from the generation
5574 : // we expect it to have. In order for our in-database generation to end up
5575 : // correct, we should carry the child generation back in the response and apply it here
5576 : // in complete_shard_split (and apply the correct generation in memory)
5577 : // (or, we can carry generation in the request and reject the request if
5578 : // it doesn't match, but that requires more retry logic on this side)
5579 :
5580 0 : self.persistence
5581 0 : .complete_shard_split(tenant_id, old_shard_count, new_shard_count)
5582 0 : .await?;
5583 :
5584 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5585 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5586 0 : ));
5587 :
5588 : // Replace all the shards we just split with their children: this phase is infallible.
5589 0 : let (response, child_locations, waiters) =
5590 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5591 0 :
5592 0 : // Send compute notifications for all the new shards
5593 0 : let mut failed_notifications = Vec::new();
5594 0 : for (child_id, child_ps, stripe_size) in child_locations {
5595 0 : if let Err(e) = self
5596 0 : .compute_hook
5597 0 : .notify(
5598 0 : compute_hook::ShardUpdate {
5599 0 : tenant_shard_id: child_id,
5600 0 : node_id: child_ps,
5601 0 : stripe_size,
5602 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5603 0 : },
5604 0 : &self.reconcilers_cancel,
5605 0 : )
5606 0 : .await
5607 : {
5608 0 : tracing::warn!(
5609 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5610 : child_id,
5611 : child_ps
5612 : );
5613 0 : failed_notifications.push(child_id);
5614 0 : }
5615 : }
5616 :
5617 : // If we failed any compute notifications, make a note to retry later.
5618 0 : if !failed_notifications.is_empty() {
5619 0 : let mut locked = self.inner.write().unwrap();
5620 0 : for failed in failed_notifications {
5621 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5622 0 : shard.pending_compute_notification = true;
5623 0 : }
5624 : }
5625 0 : }
5626 :
5627 0 : Ok((response, waiters))
5628 0 : }
5629 :
5630 : /// A graceful migration: update the preferred node and let optimisation handle the migration
5631 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
5632 : ///
5633 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
5634 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
5635 0 : fn tenant_shard_migrate_with_prewarm(
5636 0 : &self,
5637 0 : migrate_req: &TenantShardMigrateRequest,
5638 0 : shard: &mut TenantShard,
5639 0 : scheduler: &mut Scheduler,
5640 0 : schedule_context: ScheduleContext,
5641 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
5642 0 : shard.set_preferred_node(Some(migrate_req.node_id));
5643 0 :
5644 0 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
5645 0 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
5646 0 : // e.g. by checking secondary is warm enough.
5647 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
5648 0 : }
5649 :
5650 : /// Immediate migration: directly update the intent state and kick off a reconciler
5651 0 : fn tenant_shard_migrate_immediate(
5652 0 : &self,
5653 0 : migrate_req: &TenantShardMigrateRequest,
5654 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5655 0 : shard: &mut TenantShard,
5656 0 : scheduler: &mut Scheduler,
5657 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
5658 0 : // Non-graceful migration: update the intent state immediately
5659 0 : let old_attached = *shard.intent.get_attached();
5660 0 : match shard.policy {
5661 0 : PlacementPolicy::Attached(n) => {
5662 0 : // If our new attached node was a secondary, it no longer should be.
5663 0 : shard
5664 0 : .intent
5665 0 : .remove_secondary(scheduler, migrate_req.node_id);
5666 0 :
5667 0 : shard
5668 0 : .intent
5669 0 : .set_attached(scheduler, Some(migrate_req.node_id));
5670 :
5671 : // If we were already attached to something, demote that to a secondary
5672 0 : if let Some(old_attached) = old_attached {
5673 0 : if n > 0 {
5674 : // Remove other secondaries to make room for the location we'll demote
5675 0 : while shard.intent.get_secondary().len() >= n {
5676 0 : shard.intent.pop_secondary(scheduler);
5677 0 : }
5678 :
5679 0 : shard.intent.push_secondary(scheduler, old_attached);
5680 0 : }
5681 0 : }
5682 : }
5683 0 : PlacementPolicy::Secondary => {
5684 0 : shard.intent.clear(scheduler);
5685 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5686 0 : }
5687 : PlacementPolicy::Detached => {
5688 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5689 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
5690 0 : )));
5691 : }
5692 : }
5693 :
5694 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
5695 0 : shard.sequence = shard.sequence.next();
5696 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
5697 0 : Ok(self.maybe_configured_reconcile_shard(
5698 0 : shard,
5699 0 : nodes,
5700 0 : (&migrate_req.migration_config).into(),
5701 0 : ))
5702 0 : }
5703 :
5704 0 : pub(crate) async fn tenant_shard_migrate(
5705 0 : &self,
5706 0 : tenant_shard_id: TenantShardId,
5707 0 : migrate_req: TenantShardMigrateRequest,
5708 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5709 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
5710 : // get a different outcome to handle
5711 : enum MigrationOutcome {
5712 : Optimization(Option<ScheduleOptimization>),
5713 : Reconcile(Option<ReconcilerWaiter>),
5714 : }
5715 :
5716 0 : let outcome = {
5717 0 : let mut locked = self.inner.write().unwrap();
5718 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5719 :
5720 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5721 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5722 0 : "Node {} not found",
5723 0 : migrate_req.node_id
5724 0 : )));
5725 : };
5726 :
5727 : // Migration to unavavailable node requires force flag
5728 0 : if !node.is_available() {
5729 0 : if migrate_req.migration_config.override_scheduler {
5730 : // Warn but proceed: the caller may intend to manually adjust the placement of
5731 : // a shard even if the node is down, e.g. if intervening during an incident.
5732 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
5733 : } else {
5734 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
5735 0 : return Err(ApiError::PreconditionFailed(
5736 0 : format!("Node {node} is unavailable").into_boxed_str(),
5737 0 : ));
5738 : }
5739 0 : }
5740 :
5741 : // Calculate the ScheduleContext for this tenant
5742 0 : let mut schedule_context = ScheduleContext::default();
5743 0 : for (_shard_id, shard) in
5744 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
5745 0 : {
5746 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5747 0 : }
5748 :
5749 : // Look up the specific shard we will migrate
5750 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5751 0 : return Err(ApiError::NotFound(
5752 0 : anyhow::anyhow!("Tenant shard not found").into(),
5753 0 : ));
5754 : };
5755 :
5756 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
5757 : // be migrated back by the optimiser.
5758 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
5759 0 : scheduler,
5760 0 : &schedule_context,
5761 0 : migrate_req.node_id,
5762 0 : &[],
5763 0 : ) {
5764 0 : if !migrate_req.migration_config.override_scheduler {
5765 0 : return Err(ApiError::PreconditionFailed(
5766 0 : "Migration to a worse-scoring node".into(),
5767 0 : ));
5768 : } else {
5769 0 : tracing::info!(
5770 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
5771 : migrate_req.node_id
5772 : );
5773 : }
5774 0 : }
5775 :
5776 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
5777 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
5778 0 : return Err(ApiError::PreconditionFailed(
5779 0 : format!(
5780 0 : "Migration expected to originate from {} but shard is on {:?}",
5781 0 : origin_node_id,
5782 0 : shard.intent.get_attached()
5783 0 : )
5784 0 : .into(),
5785 0 : ));
5786 0 : }
5787 0 : }
5788 :
5789 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5790 : // No-op case: we will still proceed to wait for reconciliation in case it is
5791 : // incomplete from an earlier update to the intent.
5792 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
5793 :
5794 : // An instruction to migrate to the currently attached node should
5795 : // cancel any pending graceful migration
5796 0 : shard.set_preferred_node(None);
5797 0 :
5798 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
5799 0 : shard,
5800 0 : nodes,
5801 0 : (&migrate_req.migration_config).into(),
5802 0 : ))
5803 0 : } else if migrate_req.migration_config.prewarm {
5804 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
5805 0 : &migrate_req,
5806 0 : shard,
5807 0 : scheduler,
5808 0 : schedule_context,
5809 0 : )?)
5810 : } else {
5811 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
5812 0 : &migrate_req,
5813 0 : nodes,
5814 0 : shard,
5815 0 : scheduler,
5816 0 : )?)
5817 : }
5818 : };
5819 :
5820 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
5821 0 : let waiter = match outcome {
5822 0 : MigrationOutcome::Optimization(Some(optimization)) => {
5823 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
5824 : // we might as well do it more promptly as this is a direct external request.
5825 0 : let mut validated = self
5826 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
5827 0 : .await;
5828 0 : if let Some((_shard_id, optimization)) = validated.pop() {
5829 0 : let mut locked = self.inner.write().unwrap();
5830 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5831 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5832 : // Rare but possible: tenant is removed between generating optimisation and validating it.
5833 0 : return Err(ApiError::NotFound(
5834 0 : anyhow::anyhow!("Tenant shard not found").into(),
5835 0 : ));
5836 : };
5837 :
5838 0 : if !shard.apply_optimization(scheduler, optimization) {
5839 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
5840 : // and therefore not applied.
5841 0 : tracing::warn!(
5842 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
5843 : );
5844 0 : }
5845 0 : self.maybe_configured_reconcile_shard(
5846 0 : shard,
5847 0 : nodes,
5848 0 : (&migrate_req.migration_config).into(),
5849 0 : )
5850 : } else {
5851 0 : None
5852 : }
5853 : }
5854 0 : MigrationOutcome::Optimization(None) => None,
5855 0 : MigrationOutcome::Reconcile(waiter) => waiter,
5856 : };
5857 :
5858 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
5859 : // locations, this has a good chance of timing out.
5860 0 : if let Some(waiter) = waiter {
5861 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5862 : } else {
5863 0 : tracing::info!("Migration is a no-op");
5864 : }
5865 :
5866 0 : Ok(TenantShardMigrateResponse {})
5867 0 : }
5868 :
5869 0 : pub(crate) async fn tenant_shard_migrate_secondary(
5870 0 : &self,
5871 0 : tenant_shard_id: TenantShardId,
5872 0 : migrate_req: TenantShardMigrateRequest,
5873 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5874 0 : let waiter = {
5875 0 : let mut locked = self.inner.write().unwrap();
5876 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5877 :
5878 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5879 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5880 0 : "Node {} not found",
5881 0 : migrate_req.node_id
5882 0 : )));
5883 : };
5884 :
5885 0 : if !node.is_available() {
5886 : // Warn but proceed: the caller may intend to manually adjust the placement of
5887 : // a shard even if the node is down, e.g. if intervening during an incident.
5888 0 : tracing::warn!("Migrating to unavailable node {node}");
5889 0 : }
5890 :
5891 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5892 0 : return Err(ApiError::NotFound(
5893 0 : anyhow::anyhow!("Tenant shard not found").into(),
5894 0 : ));
5895 : };
5896 :
5897 0 : if shard.intent.get_secondary().len() == 1
5898 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
5899 : {
5900 0 : tracing::info!(
5901 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
5902 : shard.intent
5903 : );
5904 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5905 0 : tracing::info!(
5906 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
5907 : );
5908 : } else {
5909 0 : let old_secondaries = shard.intent.get_secondary().clone();
5910 0 : for secondary in old_secondaries {
5911 0 : shard.intent.remove_secondary(scheduler, secondary);
5912 0 : }
5913 :
5914 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5915 0 : shard.sequence = shard.sequence.next();
5916 0 : tracing::info!(
5917 0 : "Migrating secondary to {node}: new intent {:?}",
5918 : shard.intent
5919 : );
5920 : }
5921 :
5922 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5923 : };
5924 :
5925 0 : if let Some(waiter) = waiter {
5926 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
5927 : } else {
5928 0 : tracing::info!("Migration is a no-op");
5929 : }
5930 :
5931 0 : Ok(TenantShardMigrateResponse {})
5932 0 : }
5933 :
5934 : /// 'cancel' in this context means cancel any ongoing reconcile
5935 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
5936 0 : &self,
5937 0 : tenant_shard_id: TenantShardId,
5938 0 : ) -> Result<(), ApiError> {
5939 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
5940 0 : let waiter = {
5941 0 : let locked = self.inner.write().unwrap();
5942 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
5943 0 : return Err(ApiError::NotFound(
5944 0 : anyhow::anyhow!("Tenant shard not found").into(),
5945 0 : ));
5946 : };
5947 :
5948 0 : let waiter = shard.get_waiter();
5949 0 : match waiter {
5950 : None => {
5951 0 : tracing::info!("Shard does not have an ongoing Reconciler");
5952 0 : return Ok(());
5953 : }
5954 0 : Some(waiter) => {
5955 0 : tracing::info!("Cancelling Reconciler");
5956 0 : shard.cancel_reconciler();
5957 0 : waiter
5958 0 : }
5959 0 : }
5960 0 : };
5961 0 :
5962 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
5963 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
5964 0 : // the Reconciler is misbehaving and not respecting the cancellation token
5965 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
5966 0 : .await?;
5967 :
5968 0 : Ok(())
5969 0 : }
5970 :
5971 : /// This is for debug/support only: we simply drop all state for a tenant, without
5972 : /// detaching or deleting it on pageservers.
5973 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
5974 0 : self.persistence.delete_tenant(tenant_id).await?;
5975 :
5976 0 : let mut locked = self.inner.write().unwrap();
5977 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
5978 0 : let mut shards = Vec::new();
5979 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
5980 0 : shards.push(*tenant_shard_id);
5981 0 : }
5982 :
5983 0 : for shard_id in shards {
5984 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
5985 0 : shard.intent.clear(scheduler);
5986 0 : }
5987 : }
5988 :
5989 0 : Ok(())
5990 0 : }
5991 :
5992 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
5993 : /// tenant with a very high generation number so that it will see the existing data.
5994 0 : pub(crate) async fn tenant_import(
5995 0 : &self,
5996 0 : tenant_id: TenantId,
5997 0 : ) -> Result<TenantCreateResponse, ApiError> {
5998 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
5999 0 : let maybe_node = {
6000 0 : self.inner
6001 0 : .read()
6002 0 : .unwrap()
6003 0 : .nodes
6004 0 : .values()
6005 0 : .find(|n| n.is_available())
6006 0 : .cloned()
6007 : };
6008 0 : let Some(node) = maybe_node else {
6009 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
6010 : };
6011 :
6012 0 : let client = PageserverClient::new(
6013 0 : node.get_id(),
6014 0 : self.http_client.clone(),
6015 0 : node.base_url(),
6016 0 : self.config.pageserver_jwt_token.as_deref(),
6017 0 : );
6018 :
6019 0 : let scan_result = client
6020 0 : .tenant_scan_remote_storage(tenant_id)
6021 0 : .await
6022 0 : .map_err(|e| passthrough_api_error(&node, e))?;
6023 :
6024 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
6025 0 : let Some(shard_count) = scan_result
6026 0 : .shards
6027 0 : .iter()
6028 0 : .map(|s| s.tenant_shard_id.shard_count)
6029 0 : .max()
6030 : else {
6031 0 : return Err(ApiError::NotFound(
6032 0 : anyhow::anyhow!("No shards found").into(),
6033 0 : ));
6034 : };
6035 :
6036 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
6037 : // to
6038 0 : let generation = scan_result
6039 0 : .shards
6040 0 : .iter()
6041 0 : .map(|s| s.generation)
6042 0 : .max()
6043 0 : .expect("We already validated >0 shards");
6044 :
6045 : // Find the tenant's stripe size. This wasn't always persisted in the tenant manifest, so
6046 : // fall back to the original default stripe size of 32768 (256 MB) if it's not specified.
6047 : const ORIGINAL_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(32768);
6048 0 : let stripe_size = scan_result
6049 0 : .shards
6050 0 : .iter()
6051 0 : .find(|s| s.tenant_shard_id.shard_count == shard_count && s.generation == generation)
6052 0 : .expect("we validated >0 shards above")
6053 0 : .stripe_size
6054 0 : .unwrap_or_else(|| {
6055 0 : if shard_count.count() > 1 {
6056 0 : warn!("unknown stripe size, assuming {ORIGINAL_STRIPE_SIZE}");
6057 0 : }
6058 0 : ORIGINAL_STRIPE_SIZE
6059 0 : });
6060 :
6061 0 : let (response, waiters) = self
6062 0 : .do_tenant_create(TenantCreateRequest {
6063 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
6064 0 : generation,
6065 0 :
6066 0 : shard_parameters: ShardParameters {
6067 0 : count: shard_count,
6068 0 : stripe_size,
6069 0 : },
6070 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
6071 0 : config: TenantConfig::default(),
6072 0 : })
6073 0 : .await?;
6074 :
6075 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
6076 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
6077 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
6078 : // reconcile, as reconciliation includes notifying compute.
6079 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
6080 0 : }
6081 :
6082 0 : Ok(response)
6083 0 : }
6084 :
6085 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
6086 : /// we don't have to make TenantShard clonable in the return path.
6087 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
6088 0 : let serialized = {
6089 0 : let locked = self.inner.read().unwrap();
6090 0 : let result = locked.tenants.values().collect::<Vec<_>>();
6091 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
6092 : };
6093 :
6094 0 : hyper::Response::builder()
6095 0 : .status(hyper::StatusCode::OK)
6096 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
6097 0 : .body(hyper::Body::from(serialized))
6098 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
6099 0 : }
6100 :
6101 : /// Check the consistency of in-memory state vs. persistent state, and check that the
6102 : /// scheduler's statistics are up to date.
6103 : ///
6104 : /// These consistency checks expect an **idle** system. If changes are going on while
6105 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
6106 : /// checks, but not suitable for running continuously in the background in the field.
6107 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
6108 0 : let (mut expect_nodes, mut expect_shards) = {
6109 0 : let locked = self.inner.read().unwrap();
6110 0 :
6111 0 : locked
6112 0 : .scheduler
6113 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
6114 0 : .context("Scheduler checks")
6115 0 : .map_err(ApiError::InternalServerError)?;
6116 :
6117 0 : let expect_nodes = locked
6118 0 : .nodes
6119 0 : .values()
6120 0 : .map(|n| n.to_persistent())
6121 0 : .collect::<Vec<_>>();
6122 0 :
6123 0 : let expect_shards = locked
6124 0 : .tenants
6125 0 : .values()
6126 0 : .map(|t| t.to_persistent())
6127 0 : .collect::<Vec<_>>();
6128 :
6129 : // This method can only validate the state of an idle system: if a reconcile is in
6130 : // progress, fail out early to avoid giving false errors on state that won't match
6131 : // between database and memory under a ReconcileResult is processed.
6132 0 : for t in locked.tenants.values() {
6133 0 : if t.reconciler.is_some() {
6134 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6135 0 : "Shard {} reconciliation in progress",
6136 0 : t.tenant_shard_id
6137 0 : )));
6138 0 : }
6139 : }
6140 :
6141 0 : (expect_nodes, expect_shards)
6142 : };
6143 :
6144 0 : let mut nodes = self.persistence.list_nodes().await?;
6145 0 : expect_nodes.sort_by_key(|n| n.node_id);
6146 0 : nodes.sort_by_key(|n| n.node_id);
6147 :
6148 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
6149 0 : let node_result = if nodes != expect_nodes {
6150 0 : tracing::error!("Consistency check failed on nodes.");
6151 0 : tracing::error!(
6152 0 : "Nodes in memory: {}",
6153 0 : serde_json::to_string(&expect_nodes)
6154 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6155 : );
6156 0 : tracing::error!(
6157 0 : "Nodes in database: {}",
6158 0 : serde_json::to_string(&nodes)
6159 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6160 : );
6161 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
6162 0 : "Node consistency failure"
6163 0 : )))
6164 : } else {
6165 0 : Ok(())
6166 : };
6167 :
6168 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
6169 0 : persistent_shards
6170 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6171 0 :
6172 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6173 :
6174 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
6175 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
6176 : // fields are added, before doing a comparison.
6177 0 : for tsp in &mut persistent_shards {
6178 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
6179 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
6180 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
6181 : }
6182 :
6183 0 : if persistent_shards != expect_shards {
6184 0 : tracing::error!("Consistency check failed on shards.");
6185 :
6186 0 : tracing::error!(
6187 0 : "Shards in memory: {}",
6188 0 : serde_json::to_string(&expect_shards)
6189 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6190 : );
6191 0 : tracing::error!(
6192 0 : "Shards in database: {}",
6193 0 : serde_json::to_string(&persistent_shards)
6194 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6195 : );
6196 :
6197 : // The total dump log lines above are useful in testing but in the field grafana will
6198 : // usually just drop them because they're so large. So we also do some explicit logging
6199 : // of just the diffs.
6200 0 : let persistent_shards = persistent_shards
6201 0 : .into_iter()
6202 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6203 0 : .collect::<HashMap<_, _>>();
6204 0 : let expect_shards = expect_shards
6205 0 : .into_iter()
6206 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6207 0 : .collect::<HashMap<_, _>>();
6208 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
6209 0 : match expect_shards.get(tenant_shard_id) {
6210 : None => {
6211 0 : tracing::error!(
6212 0 : "Shard {} found in database but not in memory",
6213 : tenant_shard_id
6214 : );
6215 : }
6216 0 : Some(expect_tsp) => {
6217 0 : if expect_tsp != persistent_tsp {
6218 0 : tracing::error!(
6219 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
6220 0 : tenant_shard_id,
6221 0 : serde_json::to_string(expect_tsp).unwrap(),
6222 0 : serde_json::to_string(&persistent_tsp).unwrap()
6223 : );
6224 0 : }
6225 : }
6226 : }
6227 : }
6228 :
6229 : // Having already logged any differences, log any shards that simply aren't present in the database
6230 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
6231 0 : if !persistent_shards.contains_key(tenant_shard_id) {
6232 0 : tracing::error!(
6233 0 : "Shard {} found in memory but not in database: {}",
6234 0 : tenant_shard_id,
6235 0 : serde_json::to_string(memory_tsp)
6236 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6237 : );
6238 0 : }
6239 : }
6240 :
6241 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6242 0 : "Shard consistency failure"
6243 0 : )));
6244 0 : }
6245 0 :
6246 0 : node_result
6247 0 : }
6248 :
6249 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
6250 : /// we don't have to make TenantShard clonable in the return path.
6251 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
6252 0 : let serialized = {
6253 0 : let locked = self.inner.read().unwrap();
6254 0 : serde_json::to_string(&locked.scheduler)
6255 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6256 : };
6257 :
6258 0 : hyper::Response::builder()
6259 0 : .status(hyper::StatusCode::OK)
6260 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
6261 0 : .body(hyper::Body::from(serialized))
6262 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
6263 0 : }
6264 :
6265 : /// This is for debug/support only: we simply drop all state for a tenant, without
6266 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
6267 : /// tenants that were on this node.
6268 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
6269 0 : self.persistence.delete_node(node_id).await?;
6270 :
6271 0 : let mut locked = self.inner.write().unwrap();
6272 :
6273 0 : for shard in locked.tenants.values_mut() {
6274 0 : shard.deref_node(node_id);
6275 0 : shard.observed.locations.remove(&node_id);
6276 0 : }
6277 :
6278 0 : let mut nodes = (*locked.nodes).clone();
6279 0 : nodes.remove(&node_id);
6280 0 : locked.nodes = Arc::new(nodes);
6281 0 : metrics::METRICS_REGISTRY
6282 0 : .metrics_group
6283 0 : .storage_controller_pageserver_nodes
6284 0 : .set(locked.nodes.len() as i64);
6285 0 : metrics::METRICS_REGISTRY
6286 0 : .metrics_group
6287 0 : .storage_controller_https_pageserver_nodes
6288 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
6289 0 :
6290 0 : locked.scheduler.node_remove(node_id);
6291 0 :
6292 0 : Ok(())
6293 0 : }
6294 :
6295 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
6296 : /// that we don't leave any bad state behind in the storage controller, but unclean
6297 : /// in the sense that we are not carefully draining the node.
6298 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
6299 0 : let _node_lock =
6300 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
6301 :
6302 : // 1. Atomically update in-memory state:
6303 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
6304 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
6305 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
6306 : // re-insert references to this node into the ObservedState of shards
6307 : // - drop the node from the scheduler
6308 : {
6309 0 : let mut locked = self.inner.write().unwrap();
6310 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6311 0 :
6312 0 : {
6313 0 : let mut nodes_mut = (*nodes).deref().clone();
6314 0 : match nodes_mut.get_mut(&node_id) {
6315 0 : Some(node) => {
6316 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
6317 0 : // if we crash it would not be desirable to leave the node paused after a restart.
6318 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
6319 0 : }
6320 : None => {
6321 0 : tracing::info!(
6322 0 : "Node not found: presuming this is a retry and returning success"
6323 : );
6324 0 : return Ok(());
6325 : }
6326 : }
6327 :
6328 0 : *nodes = Arc::new(nodes_mut);
6329 : }
6330 :
6331 0 : for (_tenant_id, mut schedule_context, shards) in
6332 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6333 : {
6334 0 : for shard in shards {
6335 0 : if shard.deref_node(node_id) {
6336 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
6337 : // TODO: implement force flag to remove a node even if we can't reschedule
6338 : // a tenant
6339 0 : tracing::error!(
6340 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
6341 : shard.tenant_shard_id
6342 : );
6343 0 : return Err(e.into());
6344 : } else {
6345 0 : tracing::info!(
6346 0 : "Rescheduled shard {} away from node during deletion",
6347 : shard.tenant_shard_id
6348 : )
6349 : }
6350 :
6351 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
6352 0 : }
6353 :
6354 : // Here we remove an existing observed location for the node we're removing, and it will
6355 : // not be re-added by a reconciler's completion because we filter out removed nodes in
6356 : // process_result.
6357 : //
6358 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
6359 : // means any reconciles we spawned will know about the node we're deleting, enabling them
6360 : // to do live migrations if it's still online.
6361 0 : shard.observed.locations.remove(&node_id);
6362 : }
6363 : }
6364 :
6365 0 : scheduler.node_remove(node_id);
6366 0 :
6367 0 : {
6368 0 : let mut nodes_mut = (**nodes).clone();
6369 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
6370 0 : // Ensure that any reconciler holding an Arc<> to this node will
6371 0 : // drop out when trying to RPC to it (setting Offline state sets the
6372 0 : // cancellation token on the Node object).
6373 0 : removed_node.set_availability(NodeAvailability::Offline);
6374 0 : }
6375 0 : *nodes = Arc::new(nodes_mut);
6376 0 : metrics::METRICS_REGISTRY
6377 0 : .metrics_group
6378 0 : .storage_controller_pageserver_nodes
6379 0 : .set(nodes.len() as i64);
6380 0 : metrics::METRICS_REGISTRY
6381 0 : .metrics_group
6382 0 : .storage_controller_https_pageserver_nodes
6383 0 : .set(nodes.values().filter(|n| n.has_https_port()).count() as i64);
6384 0 : }
6385 0 : }
6386 0 :
6387 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
6388 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
6389 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
6390 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
6391 0 : // that exists.
6392 0 :
6393 0 : // 2. Actually delete the node from the database and from in-memory state
6394 0 : tracing::info!("Deleting node from database");
6395 0 : self.persistence.delete_node(node_id).await?;
6396 :
6397 0 : Ok(())
6398 0 : }
6399 :
6400 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
6401 0 : let nodes = {
6402 0 : self.inner
6403 0 : .read()
6404 0 : .unwrap()
6405 0 : .nodes
6406 0 : .values()
6407 0 : .cloned()
6408 0 : .collect::<Vec<_>>()
6409 0 : };
6410 0 :
6411 0 : Ok(nodes)
6412 0 : }
6413 :
6414 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
6415 0 : self.inner
6416 0 : .read()
6417 0 : .unwrap()
6418 0 : .nodes
6419 0 : .get(&node_id)
6420 0 : .cloned()
6421 0 : .ok_or(ApiError::NotFound(
6422 0 : format!("Node {node_id} not registered").into(),
6423 0 : ))
6424 0 : }
6425 :
6426 0 : pub(crate) async fn get_node_shards(
6427 0 : &self,
6428 0 : node_id: NodeId,
6429 0 : ) -> Result<NodeShardResponse, ApiError> {
6430 0 : let locked = self.inner.read().unwrap();
6431 0 : let mut shards = Vec::new();
6432 0 : for (tid, tenant) in locked.tenants.iter() {
6433 0 : let is_intended_secondary = match (
6434 0 : tenant.intent.get_attached() == &Some(node_id),
6435 0 : tenant.intent.get_secondary().contains(&node_id),
6436 0 : ) {
6437 : (true, true) => {
6438 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6439 0 : "{} attached as primary+secondary on the same node",
6440 0 : tid
6441 0 : )));
6442 : }
6443 0 : (true, false) => Some(false),
6444 0 : (false, true) => Some(true),
6445 0 : (false, false) => None,
6446 : };
6447 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
6448 0 : tenant.observed.locations.get(&node_id)
6449 : {
6450 0 : Some(conf.secondary_conf.is_some())
6451 : } else {
6452 0 : None
6453 : };
6454 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
6455 0 : shards.push(NodeShard {
6456 0 : tenant_shard_id: *tid,
6457 0 : is_intended_secondary,
6458 0 : is_observed_secondary,
6459 0 : });
6460 0 : }
6461 : }
6462 0 : Ok(NodeShardResponse { node_id, shards })
6463 0 : }
6464 :
6465 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
6466 0 : self.persistence.get_leader().await
6467 0 : }
6468 :
6469 0 : pub(crate) async fn node_register(
6470 0 : &self,
6471 0 : register_req: NodeRegisterRequest,
6472 0 : ) -> Result<(), ApiError> {
6473 0 : let _node_lock = trace_exclusive_lock(
6474 0 : &self.node_op_locks,
6475 0 : register_req.node_id,
6476 0 : NodeOperations::Register,
6477 0 : )
6478 0 : .await;
6479 :
6480 : #[derive(PartialEq)]
6481 : enum RegistrationStatus {
6482 : UpToDate,
6483 : NeedUpdate,
6484 : Mismatched,
6485 : New,
6486 : }
6487 :
6488 0 : let registration_status = {
6489 0 : let locked = self.inner.read().unwrap();
6490 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
6491 0 : if node.registration_match(®ister_req) {
6492 0 : if node.need_update(®ister_req) {
6493 0 : RegistrationStatus::NeedUpdate
6494 : } else {
6495 0 : RegistrationStatus::UpToDate
6496 : }
6497 : } else {
6498 0 : RegistrationStatus::Mismatched
6499 : }
6500 : } else {
6501 0 : RegistrationStatus::New
6502 : }
6503 : };
6504 :
6505 0 : match registration_status {
6506 : RegistrationStatus::UpToDate => {
6507 0 : tracing::info!(
6508 0 : "Node {} re-registered with matching address and is up to date",
6509 : register_req.node_id
6510 : );
6511 :
6512 0 : return Ok(());
6513 : }
6514 : RegistrationStatus::Mismatched => {
6515 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
6516 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
6517 : // a fixed address through the lifetime of a node.
6518 0 : tracing::warn!(
6519 0 : "Node {} tried to register with different address",
6520 : register_req.node_id
6521 : );
6522 0 : return Err(ApiError::Conflict(
6523 0 : "Node is already registered with different address".to_string(),
6524 0 : ));
6525 : }
6526 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
6527 0 : // fallthrough
6528 0 : }
6529 0 : }
6530 0 :
6531 0 : // We do not require that a node is actually online when registered (it will start life
6532 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
6533 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
6534 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
6535 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
6536 0 : if tokio::net::lookup_host(format!(
6537 0 : "{}:{}",
6538 0 : register_req.listen_http_addr, register_req.listen_http_port
6539 0 : ))
6540 0 : .await
6541 0 : .is_err()
6542 : {
6543 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
6544 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
6545 : // we return a soft 503 error, to encourage callers to retry past transient issues.
6546 0 : return Err(ApiError::ResourceUnavailable(
6547 0 : format!(
6548 0 : "Node {} tried to register with unknown DNS name '{}'",
6549 0 : register_req.node_id, register_req.listen_http_addr
6550 0 : )
6551 0 : .into(),
6552 0 : ));
6553 0 : }
6554 0 :
6555 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
6556 0 : return Err(ApiError::PreconditionFailed(
6557 0 : format!(
6558 0 : "Node {} has no https port, but use_https is enabled",
6559 0 : register_req.node_id
6560 0 : )
6561 0 : .into(),
6562 0 : ));
6563 0 : }
6564 0 :
6565 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
6566 0 : // This ensures that before we use it for anything or expose it via any external
6567 0 : // API, it is guaranteed to be available after a restart.
6568 0 : let new_node = Node::new(
6569 0 : register_req.node_id,
6570 0 : register_req.listen_http_addr,
6571 0 : register_req.listen_http_port,
6572 0 : register_req.listen_https_port,
6573 0 : register_req.listen_pg_addr,
6574 0 : register_req.listen_pg_port,
6575 0 : register_req.availability_zone_id.clone(),
6576 0 : self.config.use_https_pageserver_api,
6577 0 : );
6578 0 : let new_node = match new_node {
6579 0 : Ok(new_node) => new_node,
6580 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
6581 : };
6582 :
6583 0 : match registration_status {
6584 0 : RegistrationStatus::New => self.persistence.insert_node(&new_node).await?,
6585 : RegistrationStatus::NeedUpdate => {
6586 0 : self.persistence
6587 0 : .update_node_on_registration(
6588 0 : register_req.node_id,
6589 0 : register_req.listen_https_port,
6590 0 : )
6591 0 : .await?
6592 : }
6593 0 : _ => unreachable!("Other statuses have been processed earlier"),
6594 : }
6595 :
6596 0 : let mut locked = self.inner.write().unwrap();
6597 0 : let mut new_nodes = (*locked.nodes).clone();
6598 0 :
6599 0 : locked.scheduler.node_upsert(&new_node);
6600 0 : new_nodes.insert(register_req.node_id, new_node);
6601 0 :
6602 0 : locked.nodes = Arc::new(new_nodes);
6603 0 :
6604 0 : metrics::METRICS_REGISTRY
6605 0 : .metrics_group
6606 0 : .storage_controller_pageserver_nodes
6607 0 : .set(locked.nodes.len() as i64);
6608 0 : metrics::METRICS_REGISTRY
6609 0 : .metrics_group
6610 0 : .storage_controller_https_pageserver_nodes
6611 0 : .set(locked.nodes.values().filter(|n| n.has_https_port()).count() as i64);
6612 0 :
6613 0 : match registration_status {
6614 : RegistrationStatus::New => {
6615 0 : tracing::info!(
6616 0 : "Registered pageserver {} ({}), now have {} pageservers",
6617 0 : register_req.node_id,
6618 0 : register_req.availability_zone_id,
6619 0 : locked.nodes.len()
6620 : );
6621 : }
6622 : RegistrationStatus::NeedUpdate => {
6623 0 : tracing::info!(
6624 0 : "Re-registered and updated node {} ({})",
6625 : register_req.node_id,
6626 : register_req.availability_zone_id,
6627 : );
6628 : }
6629 0 : _ => unreachable!("Other statuses have been processed earlier"),
6630 : }
6631 0 : Ok(())
6632 0 : }
6633 :
6634 : /// Configure in-memory and persistent state of a node as requested
6635 : ///
6636 : /// Note that this function does not trigger any immediate side effects in response
6637 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
6638 0 : async fn node_state_configure(
6639 0 : &self,
6640 0 : node_id: NodeId,
6641 0 : availability: Option<NodeAvailability>,
6642 0 : scheduling: Option<NodeSchedulingPolicy>,
6643 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
6644 0 : ) -> Result<AvailabilityTransition, ApiError> {
6645 0 : if let Some(scheduling) = scheduling {
6646 : // Scheduling is a persistent part of Node: we must write updates to the database before
6647 : // applying them in memory
6648 0 : self.persistence
6649 0 : .update_node_scheduling_policy(node_id, scheduling)
6650 0 : .await?;
6651 0 : }
6652 :
6653 : // If we're activating a node, then before setting it active we must reconcile any shard locations
6654 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
6655 : // by calling [`Self::node_activate_reconcile`]
6656 : //
6657 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
6658 : // nothing else can mutate its availability while we run.
6659 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
6660 0 : let (activate_node, availability_transition) = {
6661 0 : let locked = self.inner.read().unwrap();
6662 0 : let Some(node) = locked.nodes.get(&node_id) else {
6663 0 : return Err(ApiError::NotFound(
6664 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6665 0 : ));
6666 : };
6667 :
6668 0 : (
6669 0 : node.clone(),
6670 0 : node.get_availability_transition(input_availability),
6671 0 : )
6672 : };
6673 :
6674 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
6675 0 : self.node_activate_reconcile(activate_node, node_lock)
6676 0 : .await?;
6677 0 : }
6678 0 : availability_transition
6679 : } else {
6680 0 : AvailabilityTransition::Unchanged
6681 : };
6682 :
6683 : // Apply changes from the request to our in-memory state for the Node
6684 0 : let mut locked = self.inner.write().unwrap();
6685 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
6686 0 :
6687 0 : let mut new_nodes = (**nodes).clone();
6688 :
6689 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
6690 0 : return Err(ApiError::NotFound(
6691 0 : anyhow::anyhow!("Node not registered").into(),
6692 0 : ));
6693 : };
6694 :
6695 0 : if let Some(availability) = availability {
6696 0 : node.set_availability(availability);
6697 0 : }
6698 :
6699 0 : if let Some(scheduling) = scheduling {
6700 0 : node.set_scheduling(scheduling);
6701 0 : }
6702 :
6703 : // Update the scheduler, in case the elegibility of the node for new shards has changed
6704 0 : scheduler.node_upsert(node);
6705 0 :
6706 0 : let new_nodes = Arc::new(new_nodes);
6707 0 : locked.nodes = new_nodes;
6708 0 :
6709 0 : Ok(availability_transition)
6710 0 : }
6711 :
6712 : /// Handle availability transition of one node
6713 : ///
6714 : /// Note that you should first call [`Self::node_state_configure`] to update
6715 : /// the in-memory state referencing that node. If you need to handle more than one transition
6716 : /// consider using [`Self::handle_node_availability_transitions`].
6717 0 : async fn handle_node_availability_transition(
6718 0 : &self,
6719 0 : node_id: NodeId,
6720 0 : transition: AvailabilityTransition,
6721 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
6722 0 : ) -> Result<(), ApiError> {
6723 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
6724 0 : match transition {
6725 : AvailabilityTransition::ToOffline => {
6726 0 : tracing::info!("Node {} transition to offline", node_id);
6727 :
6728 0 : let mut locked = self.inner.write().unwrap();
6729 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6730 0 :
6731 0 : let mut tenants_affected: usize = 0;
6732 :
6733 0 : for (_tenant_id, mut schedule_context, shards) in
6734 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6735 : {
6736 0 : for tenant_shard in shards {
6737 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
6738 0 : if let Some(observed_loc) =
6739 0 : tenant_shard.observed.locations.get_mut(&node_id)
6740 0 : {
6741 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
6742 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
6743 0 : observed_loc.conf = None;
6744 0 : }
6745 :
6746 0 : if nodes.len() == 1 {
6747 : // Special case for single-node cluster: there is no point trying to reschedule
6748 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
6749 : // failures to schedule them.
6750 0 : continue;
6751 0 : }
6752 0 :
6753 0 : if !nodes
6754 0 : .values()
6755 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6756 : {
6757 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
6758 : // trying to reschedule since there's nowhere else to go. Without this
6759 : // branch we incorrectly detach tenants in response to node unavailability.
6760 0 : continue;
6761 0 : }
6762 0 :
6763 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
6764 0 : tenant_shard.sequence = tenant_shard.sequence.next();
6765 0 :
6766 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
6767 0 : Err(e) => {
6768 0 : // It is possible that some tenants will become unschedulable when too many pageservers
6769 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
6770 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
6771 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
6772 : }
6773 : Ok(()) => {
6774 0 : if self
6775 0 : .maybe_reconcile_shard(
6776 0 : tenant_shard,
6777 0 : nodes,
6778 0 : ReconcilerPriority::Normal,
6779 0 : )
6780 0 : .is_some()
6781 0 : {
6782 0 : tenants_affected += 1;
6783 0 : };
6784 : }
6785 : }
6786 0 : }
6787 : }
6788 : }
6789 0 : tracing::info!(
6790 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
6791 : tenants_affected,
6792 : node_id
6793 : )
6794 : }
6795 : AvailabilityTransition::ToActive => {
6796 0 : tracing::info!("Node {} transition to active", node_id);
6797 :
6798 0 : let mut locked = self.inner.write().unwrap();
6799 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
6800 :
6801 : // When a node comes back online, we must reconcile any tenant that has a None observed
6802 : // location on the node.
6803 0 : for tenant_shard in tenants.values_mut() {
6804 : // If a reconciliation is already in progress, rely on the previous scheduling
6805 : // decision and skip triggering a new reconciliation.
6806 0 : if tenant_shard.reconciler.is_some() {
6807 0 : continue;
6808 0 : }
6809 :
6810 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
6811 0 : if observed_loc.conf.is_none() {
6812 0 : self.maybe_reconcile_shard(
6813 0 : tenant_shard,
6814 0 : nodes,
6815 0 : ReconcilerPriority::Normal,
6816 0 : );
6817 0 : }
6818 0 : }
6819 : }
6820 :
6821 : // TODO: in the background, we should balance work back onto this pageserver
6822 : }
6823 : // No action required for the intermediate unavailable state.
6824 : // When we transition into active or offline from the unavailable state,
6825 : // the correct handling above will kick in.
6826 : AvailabilityTransition::ToWarmingUpFromActive => {
6827 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
6828 : }
6829 : AvailabilityTransition::ToWarmingUpFromOffline => {
6830 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
6831 : }
6832 : AvailabilityTransition::Unchanged => {
6833 0 : tracing::debug!("Node {} no availability change during config", node_id);
6834 : }
6835 : }
6836 :
6837 0 : Ok(())
6838 0 : }
6839 :
6840 : /// Handle availability transition for multiple nodes
6841 : ///
6842 : /// Note that you should first call [`Self::node_state_configure`] for
6843 : /// all nodes being handled here for the handling to use fresh in-memory state.
6844 0 : async fn handle_node_availability_transitions(
6845 0 : &self,
6846 0 : transitions: Vec<(
6847 0 : NodeId,
6848 0 : TracingExclusiveGuard<NodeOperations>,
6849 0 : AvailabilityTransition,
6850 0 : )>,
6851 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
6852 0 : let mut errors = Vec::default();
6853 0 : for (node_id, node_lock, transition) in transitions {
6854 0 : let res = self
6855 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
6856 0 : .await;
6857 0 : if let Err(err) = res {
6858 0 : errors.push((node_id, err));
6859 0 : }
6860 : }
6861 :
6862 0 : if errors.is_empty() {
6863 0 : Ok(())
6864 : } else {
6865 0 : Err(errors)
6866 : }
6867 0 : }
6868 :
6869 0 : pub(crate) async fn node_configure(
6870 0 : &self,
6871 0 : node_id: NodeId,
6872 0 : availability: Option<NodeAvailability>,
6873 0 : scheduling: Option<NodeSchedulingPolicy>,
6874 0 : ) -> Result<(), ApiError> {
6875 0 : let node_lock =
6876 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
6877 :
6878 0 : let transition = self
6879 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
6880 0 : .await?;
6881 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
6882 0 : .await
6883 0 : }
6884 :
6885 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
6886 : /// operation for HTTP api.
6887 0 : pub(crate) async fn external_node_configure(
6888 0 : &self,
6889 0 : node_id: NodeId,
6890 0 : availability: Option<NodeAvailability>,
6891 0 : scheduling: Option<NodeSchedulingPolicy>,
6892 0 : ) -> Result<(), ApiError> {
6893 0 : {
6894 0 : let locked = self.inner.read().unwrap();
6895 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
6896 0 : return Err(ApiError::PreconditionFailed(
6897 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
6898 0 : ));
6899 0 : }
6900 0 : }
6901 0 :
6902 0 : self.node_configure(node_id, availability, scheduling).await
6903 0 : }
6904 :
6905 0 : pub(crate) async fn start_node_drain(
6906 0 : self: &Arc<Self>,
6907 0 : node_id: NodeId,
6908 0 : ) -> Result<(), ApiError> {
6909 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
6910 0 : let locked = self.inner.read().unwrap();
6911 0 : let nodes = &locked.nodes;
6912 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
6913 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6914 0 : ))?;
6915 0 : let schedulable_nodes_count = nodes
6916 0 : .iter()
6917 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6918 0 : .count();
6919 0 :
6920 0 : (
6921 0 : locked
6922 0 : .ongoing_operation
6923 0 : .as_ref()
6924 0 : .map(|ongoing| ongoing.operation),
6925 0 : node.is_available(),
6926 0 : node.get_scheduling(),
6927 0 : schedulable_nodes_count,
6928 0 : )
6929 0 : };
6930 :
6931 0 : if let Some(ongoing) = ongoing_op {
6932 0 : return Err(ApiError::PreconditionFailed(
6933 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
6934 0 : ));
6935 0 : }
6936 0 :
6937 0 : if !node_available {
6938 0 : return Err(ApiError::ResourceUnavailable(
6939 0 : format!("Node {node_id} is currently unavailable").into(),
6940 0 : ));
6941 0 : }
6942 0 :
6943 0 : if schedulable_nodes_count == 0 {
6944 0 : return Err(ApiError::PreconditionFailed(
6945 0 : "No other schedulable nodes to drain to".into(),
6946 0 : ));
6947 0 : }
6948 0 :
6949 0 : match node_policy {
6950 : NodeSchedulingPolicy::Active => {
6951 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
6952 0 : .await?;
6953 :
6954 0 : let cancel = self.cancel.child_token();
6955 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
6956 :
6957 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
6958 0 : operation: Operation::Drain(Drain { node_id }),
6959 0 : cancel: cancel.clone(),
6960 0 : });
6961 :
6962 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
6963 :
6964 0 : tokio::task::spawn({
6965 0 : let service = self.clone();
6966 0 : let cancel = cancel.clone();
6967 0 : async move {
6968 0 : let _gate_guard = gate_guard;
6969 0 :
6970 0 : scopeguard::defer! {
6971 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
6972 0 :
6973 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
6974 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
6975 0 : } else {
6976 0 : panic!("We always remove the same operation")
6977 0 : }
6978 0 : }
6979 0 :
6980 0 : tracing::info!("Drain background operation starting");
6981 0 : let res = service.drain_node(node_id, cancel).await;
6982 0 : match res {
6983 : Ok(()) => {
6984 0 : tracing::info!("Drain background operation completed successfully");
6985 : }
6986 : Err(OperationError::Cancelled) => {
6987 0 : tracing::info!("Drain background operation was cancelled");
6988 : }
6989 0 : Err(err) => {
6990 0 : tracing::error!("Drain background operation encountered: {err}")
6991 : }
6992 : }
6993 0 : }
6994 0 : }.instrument(span));
6995 0 : }
6996 : NodeSchedulingPolicy::Draining => {
6997 0 : return Err(ApiError::Conflict(format!(
6998 0 : "Node {node_id} has drain in progress"
6999 0 : )));
7000 : }
7001 0 : policy => {
7002 0 : return Err(ApiError::PreconditionFailed(
7003 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
7004 0 : ));
7005 : }
7006 : }
7007 :
7008 0 : Ok(())
7009 0 : }
7010 :
7011 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
7012 0 : let node_available = {
7013 0 : let locked = self.inner.read().unwrap();
7014 0 : let nodes = &locked.nodes;
7015 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7016 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7017 0 : ))?;
7018 :
7019 0 : node.is_available()
7020 0 : };
7021 0 :
7022 0 : if !node_available {
7023 0 : return Err(ApiError::ResourceUnavailable(
7024 0 : format!("Node {node_id} is currently unavailable").into(),
7025 0 : ));
7026 0 : }
7027 :
7028 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
7029 0 : if let Operation::Drain(drain) = op_handler.operation {
7030 0 : if drain.node_id == node_id {
7031 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
7032 0 : op_handler.cancel.cancel();
7033 0 : return Ok(());
7034 0 : }
7035 0 : }
7036 0 : }
7037 :
7038 0 : Err(ApiError::PreconditionFailed(
7039 0 : format!("Node {node_id} has no drain in progress").into(),
7040 0 : ))
7041 0 : }
7042 :
7043 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
7044 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
7045 0 : let locked = self.inner.read().unwrap();
7046 0 : let nodes = &locked.nodes;
7047 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7048 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7049 0 : ))?;
7050 :
7051 0 : (
7052 0 : locked
7053 0 : .ongoing_operation
7054 0 : .as_ref()
7055 0 : .map(|ongoing| ongoing.operation),
7056 0 : node.is_available(),
7057 0 : node.get_scheduling(),
7058 0 : nodes.len(),
7059 0 : )
7060 0 : };
7061 :
7062 0 : if let Some(ongoing) = ongoing_op {
7063 0 : return Err(ApiError::PreconditionFailed(
7064 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
7065 0 : ));
7066 0 : }
7067 0 :
7068 0 : if !node_available {
7069 0 : return Err(ApiError::ResourceUnavailable(
7070 0 : format!("Node {node_id} is currently unavailable").into(),
7071 0 : ));
7072 0 : }
7073 0 :
7074 0 : if total_nodes_count <= 1 {
7075 0 : return Err(ApiError::PreconditionFailed(
7076 0 : "No other nodes to fill from".into(),
7077 0 : ));
7078 0 : }
7079 0 :
7080 0 : match node_policy {
7081 : NodeSchedulingPolicy::Active => {
7082 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
7083 0 : .await?;
7084 :
7085 0 : let cancel = self.cancel.child_token();
7086 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
7087 :
7088 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
7089 0 : operation: Operation::Fill(Fill { node_id }),
7090 0 : cancel: cancel.clone(),
7091 0 : });
7092 :
7093 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
7094 :
7095 0 : tokio::task::spawn({
7096 0 : let service = self.clone();
7097 0 : let cancel = cancel.clone();
7098 0 : async move {
7099 0 : let _gate_guard = gate_guard;
7100 0 :
7101 0 : scopeguard::defer! {
7102 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
7103 0 :
7104 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
7105 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
7106 0 : } else {
7107 0 : panic!("We always remove the same operation")
7108 0 : }
7109 0 : }
7110 0 :
7111 0 : tracing::info!("Fill background operation starting");
7112 0 : let res = service.fill_node(node_id, cancel).await;
7113 0 : match res {
7114 : Ok(()) => {
7115 0 : tracing::info!("Fill background operation completed successfully");
7116 : }
7117 : Err(OperationError::Cancelled) => {
7118 0 : tracing::info!("Fill background operation was cancelled");
7119 : }
7120 0 : Err(err) => {
7121 0 : tracing::error!("Fill background operation encountered: {err}")
7122 : }
7123 : }
7124 0 : }
7125 0 : }.instrument(span));
7126 0 : }
7127 : NodeSchedulingPolicy::Filling => {
7128 0 : return Err(ApiError::Conflict(format!(
7129 0 : "Node {node_id} has fill in progress"
7130 0 : )));
7131 : }
7132 0 : policy => {
7133 0 : return Err(ApiError::PreconditionFailed(
7134 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
7135 0 : ));
7136 : }
7137 : }
7138 :
7139 0 : Ok(())
7140 0 : }
7141 :
7142 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
7143 0 : let node_available = {
7144 0 : let locked = self.inner.read().unwrap();
7145 0 : let nodes = &locked.nodes;
7146 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7147 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7148 0 : ))?;
7149 :
7150 0 : node.is_available()
7151 0 : };
7152 0 :
7153 0 : if !node_available {
7154 0 : return Err(ApiError::ResourceUnavailable(
7155 0 : format!("Node {node_id} is currently unavailable").into(),
7156 0 : ));
7157 0 : }
7158 :
7159 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
7160 0 : if let Operation::Fill(fill) = op_handler.operation {
7161 0 : if fill.node_id == node_id {
7162 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
7163 0 : op_handler.cancel.cancel();
7164 0 : return Ok(());
7165 0 : }
7166 0 : }
7167 0 : }
7168 :
7169 0 : Err(ApiError::PreconditionFailed(
7170 0 : format!("Node {node_id} has no fill in progress").into(),
7171 0 : ))
7172 0 : }
7173 :
7174 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
7175 : /// configuration
7176 0 : fn maybe_reconcile_shard(
7177 0 : &self,
7178 0 : shard: &mut TenantShard,
7179 0 : nodes: &Arc<HashMap<NodeId, Node>>,
7180 0 : priority: ReconcilerPriority,
7181 0 : ) -> Option<ReconcilerWaiter> {
7182 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
7183 0 : }
7184 :
7185 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
7186 0 : fn get_reconciler_units(
7187 0 : &self,
7188 0 : priority: ReconcilerPriority,
7189 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
7190 0 : let units = match priority {
7191 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
7192 : ReconcilerPriority::High => {
7193 0 : match self
7194 0 : .priority_reconciler_concurrency
7195 0 : .clone()
7196 0 : .try_acquire_owned()
7197 : {
7198 0 : Ok(u) => Ok(u),
7199 : Err(TryAcquireError::NoPermits) => {
7200 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
7201 : // the normal priority semaphore.
7202 0 : self.reconciler_concurrency.clone().try_acquire_owned()
7203 : }
7204 0 : Err(e) => Err(e),
7205 : }
7206 : }
7207 : };
7208 :
7209 0 : units.map(ReconcileUnits::new)
7210 0 : }
7211 :
7212 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
7213 0 : fn maybe_configured_reconcile_shard(
7214 0 : &self,
7215 0 : shard: &mut TenantShard,
7216 0 : nodes: &Arc<HashMap<NodeId, Node>>,
7217 0 : reconciler_config: ReconcilerConfig,
7218 0 : ) -> Option<ReconcilerWaiter> {
7219 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
7220 :
7221 0 : let reconcile_reason = match reconcile_needed {
7222 0 : ReconcileNeeded::No => return None,
7223 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
7224 0 : ReconcileNeeded::Yes(reason) => {
7225 0 : // Fall through to try and acquire units for spawning reconciler
7226 0 : reason
7227 : }
7228 : };
7229 :
7230 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
7231 0 : Ok(u) => u,
7232 : Err(_) => {
7233 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
7234 0 : "Concurrency limited: enqueued for reconcile later");
7235 0 : if !shard.delayed_reconcile {
7236 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
7237 0 : Err(TrySendError::Closed(_)) => {
7238 0 : // Weird mid-shutdown case?
7239 0 : }
7240 : Err(TrySendError::Full(_)) => {
7241 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
7242 0 : tracing::warn!(
7243 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
7244 : );
7245 : }
7246 0 : Ok(()) => {
7247 0 : shard.delayed_reconcile = true;
7248 0 : }
7249 : }
7250 0 : }
7251 :
7252 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
7253 : // number to advance. When this function is eventually called again and succeeds in getting units,
7254 : // it will spawn a reconciler that makes this waiter complete.
7255 0 : return Some(shard.future_reconcile_waiter());
7256 : }
7257 : };
7258 :
7259 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
7260 : // Gate closed: we're shutting down, drop out.
7261 0 : return None;
7262 : };
7263 :
7264 0 : shard.spawn_reconciler(
7265 0 : reconcile_reason,
7266 0 : &self.result_tx,
7267 0 : nodes,
7268 0 : &self.compute_hook,
7269 0 : reconciler_config,
7270 0 : &self.config,
7271 0 : &self.persistence,
7272 0 : units,
7273 0 : gate_guard,
7274 0 : &self.reconcilers_cancel,
7275 0 : self.http_client.clone(),
7276 0 : )
7277 0 : }
7278 :
7279 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
7280 : /// Additionally, reschedule tenants that require it.
7281 : ///
7282 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
7283 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
7284 : /// available. A return value of 0 indicates that everything is fully reconciled already.
7285 0 : fn reconcile_all(&self) -> usize {
7286 0 : let mut locked = self.inner.write().unwrap();
7287 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7288 0 : let pageservers = nodes.clone();
7289 0 :
7290 0 : // This function is an efficient place to update lazy statistics, since we are walking
7291 0 : // all tenants.
7292 0 : let mut pending_reconciles = 0;
7293 0 : let mut az_violations = 0;
7294 0 :
7295 0 : // If we find any tenants to drop from memory, stash them to offload after
7296 0 : // we're done traversing the map of tenants.
7297 0 : let mut drop_detached_tenants = Vec::new();
7298 0 :
7299 0 : let mut reconciles_spawned = 0;
7300 0 : for shard in tenants.values_mut() {
7301 : // Accumulate scheduling statistics
7302 0 : if let (Some(attached), Some(preferred)) =
7303 0 : (shard.intent.get_attached(), shard.preferred_az())
7304 : {
7305 0 : let node_az = nodes
7306 0 : .get(attached)
7307 0 : .expect("Nodes exist if referenced")
7308 0 : .get_availability_zone_id();
7309 0 : if node_az != preferred {
7310 0 : az_violations += 1;
7311 0 : }
7312 0 : }
7313 :
7314 : // Skip checking if this shard is already enqueued for reconciliation
7315 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
7316 : // If there is something delayed, then return a nonzero count so that
7317 : // callers like reconcile_all_now do not incorrectly get the impression
7318 : // that the system is in a quiescent state.
7319 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
7320 0 : pending_reconciles += 1;
7321 0 : continue;
7322 0 : }
7323 0 :
7324 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
7325 0 : // dirty, spawn another one
7326 0 : if self
7327 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
7328 0 : .is_some()
7329 0 : {
7330 0 : reconciles_spawned += 1;
7331 0 : } else if shard.delayed_reconcile {
7332 0 : // Shard wanted to reconcile but for some reason couldn't.
7333 0 : pending_reconciles += 1;
7334 0 : }
7335 :
7336 : // If this tenant is detached, try dropping it from memory. This is usually done
7337 : // proactively in [`Self::process_results`], but we do it here to handle the edge
7338 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
7339 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
7340 0 : && shard.policy == PlacementPolicy::Detached
7341 : {
7342 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
7343 0 : shard.tenant_shard_id.tenant_id,
7344 0 : TenantOperations::DropDetached,
7345 0 : ) {
7346 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
7347 0 : }
7348 0 : }
7349 : }
7350 :
7351 : // Some metrics are calculated from SchedulerNode state, update these periodically
7352 0 : scheduler.update_metrics();
7353 :
7354 : // Process any deferred tenant drops
7355 0 : for (tenant_id, guard) in drop_detached_tenants {
7356 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
7357 0 : }
7358 :
7359 0 : metrics::METRICS_REGISTRY
7360 0 : .metrics_group
7361 0 : .storage_controller_schedule_az_violation
7362 0 : .set(az_violations as i64);
7363 0 :
7364 0 : metrics::METRICS_REGISTRY
7365 0 : .metrics_group
7366 0 : .storage_controller_pending_reconciles
7367 0 : .set(pending_reconciles as i64);
7368 0 :
7369 0 : reconciles_spawned
7370 0 : }
7371 :
7372 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
7373 : /// could be scheduled somewhere better:
7374 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
7375 : /// * e.g. after a node fails then recovers, to move some work back to it
7376 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
7377 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
7378 : /// we did the split, but are probably better placed elsewhere.
7379 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
7380 : /// * e.g. after a shard split, some locations will be on the same node (where the split
7381 : /// happened), and will probably be better placed elsewhere.
7382 : ///
7383 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
7384 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
7385 : /// according to those same soft constraints.
7386 0 : async fn optimize_all(&self) -> usize {
7387 : // Limit on how many shards' optmizations each call to this function will execute. Combined
7388 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
7389 : // trickle of optimizations in the background, rather than executing a large number in parallel
7390 : // when a change occurs.
7391 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
7392 :
7393 : // Synchronous prepare: scan shards for possible scheduling optimizations
7394 0 : let candidate_work = self.optimize_all_plan();
7395 0 : let candidate_work_len = candidate_work.len();
7396 :
7397 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
7398 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
7399 :
7400 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
7401 0 :
7402 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
7403 0 : let mut reconciles_spawned = 0;
7404 0 : let mut optimizations_applied = 0;
7405 0 : let mut locked = self.inner.write().unwrap();
7406 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7407 0 : for (tenant_shard_id, optimization) in validated_work {
7408 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
7409 : // Shard was dropped between planning and execution;
7410 0 : continue;
7411 : };
7412 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
7413 0 : if shard.apply_optimization(scheduler, optimization) {
7414 0 : optimizations_applied += 1;
7415 0 : if self
7416 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
7417 0 : .is_some()
7418 0 : {
7419 0 : reconciles_spawned += 1;
7420 0 : }
7421 0 : }
7422 :
7423 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
7424 0 : break;
7425 0 : }
7426 : }
7427 :
7428 0 : if was_work_filtered {
7429 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
7430 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
7431 0 : // as these validations start passing.
7432 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
7433 0 : }
7434 :
7435 0 : reconciles_spawned
7436 0 : }
7437 :
7438 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
7439 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
7440 : // this higher than the execution limit gives us a chance to execute some work even if the first
7441 : // few optimizations we find are not ready.
7442 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
7443 :
7444 0 : let mut work = Vec::new();
7445 0 : let mut locked = self.inner.write().unwrap();
7446 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
7447 :
7448 : // We are going to plan a bunch of optimisations before applying any of them, so the
7449 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
7450 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
7451 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
7452 : // to ignore the utilisation component of the score.
7453 :
7454 0 : for (_tenant_id, schedule_context, shards) in
7455 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
7456 : {
7457 0 : for shard in shards {
7458 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
7459 0 : break;
7460 0 : }
7461 0 : match shard.get_scheduling_policy() {
7462 0 : ShardSchedulingPolicy::Active => {
7463 0 : // Ok to do optimization
7464 0 : }
7465 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
7466 0 : // Ok to do optimization: we are executing a graceful migration that
7467 0 : // has set preferred_node
7468 0 : }
7469 : ShardSchedulingPolicy::Essential
7470 : | ShardSchedulingPolicy::Pause
7471 : | ShardSchedulingPolicy::Stop => {
7472 : // Policy prevents optimizing this shard.
7473 0 : continue;
7474 : }
7475 : }
7476 :
7477 0 : if !matches!(shard.splitting, SplitState::Idle)
7478 0 : || matches!(shard.policy, PlacementPolicy::Detached)
7479 0 : || shard.reconciler.is_some()
7480 : {
7481 : // Do not start any optimizations while another change to the tenant is ongoing: this
7482 : // is not necessary for correctness, but simplifies operations and implicitly throttles
7483 : // optimization changes to happen in a "trickle" over time.
7484 0 : continue;
7485 0 : }
7486 0 :
7487 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
7488 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
7489 0 : if cfg!(feature = "testing") {
7490 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
7491 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
7492 : // panic in prod if we hit this, or spend cycles on it in prod.
7493 0 : assert!(
7494 0 : shard
7495 0 : .optimize_attachment(scheduler, &schedule_context)
7496 0 : .is_none()
7497 0 : );
7498 0 : assert!(
7499 0 : shard
7500 0 : .optimize_secondary(scheduler, &schedule_context)
7501 0 : .is_none()
7502 0 : );
7503 0 : }
7504 0 : continue;
7505 0 : }
7506 :
7507 0 : if let Some(optimization) =
7508 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
7509 : // its primary location based on soft constraints, cut it over.
7510 0 : shard.optimize_attachment(scheduler, &schedule_context)
7511 : {
7512 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
7513 0 : work.push((shard.tenant_shard_id, optimization));
7514 0 : break;
7515 0 : } else if let Some(optimization) =
7516 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
7517 : // better placed on another node, based on ScheduleContext, then adjust it. This
7518 : // covers cases like after a shard split, where we might have too many shards
7519 : // in the same tenant with secondary locations on the node where they originally split.
7520 0 : shard.optimize_secondary(scheduler, &schedule_context)
7521 : {
7522 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
7523 0 : work.push((shard.tenant_shard_id, optimization));
7524 0 : break;
7525 0 : }
7526 : }
7527 : }
7528 :
7529 0 : work
7530 0 : }
7531 :
7532 0 : async fn optimize_all_validate(
7533 0 : &self,
7534 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
7535 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
7536 0 : // Take a clone of the node map to use outside the lock in async validation phase
7537 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
7538 0 :
7539 0 : let mut want_secondary_status = Vec::new();
7540 0 :
7541 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
7542 0 : // check that the state of locations is acceptable to run the optimization, such as
7543 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
7544 0 : // in a live migration.
7545 0 : let mut validated_work = Vec::new();
7546 0 : for (tenant_shard_id, optimization) in candidate_work {
7547 0 : match optimization.action {
7548 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
7549 : old_attached_node_id: _,
7550 0 : new_attached_node_id,
7551 0 : }) => {
7552 0 : match validation_nodes.get(&new_attached_node_id) {
7553 0 : None => {
7554 0 : // Node was dropped between planning and validation
7555 0 : }
7556 0 : Some(node) => {
7557 0 : if !node.is_available() {
7558 0 : tracing::info!(
7559 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
7560 : );
7561 0 : } else {
7562 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
7563 0 : // remote API requests concurrently.
7564 0 : want_secondary_status.push((
7565 0 : tenant_shard_id,
7566 0 : node.clone(),
7567 0 : optimization,
7568 0 : ));
7569 0 : }
7570 : }
7571 : }
7572 : }
7573 : ScheduleOptimizationAction::ReplaceSecondary(_)
7574 : | ScheduleOptimizationAction::CreateSecondary(_)
7575 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
7576 : // No extra checks needed to manage secondaries: this does not interrupt client access
7577 0 : validated_work.push((tenant_shard_id, optimization))
7578 : }
7579 : };
7580 : }
7581 :
7582 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
7583 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
7584 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
7585 0 : let results = self
7586 0 : .tenant_for_shards_api(
7587 0 : want_secondary_status
7588 0 : .iter()
7589 0 : .map(|i| (i.0, i.1.clone()))
7590 0 : .collect(),
7591 0 : |tenant_shard_id, client| async move {
7592 0 : client.tenant_secondary_status(tenant_shard_id).await
7593 0 : },
7594 0 : 1,
7595 0 : 1,
7596 0 : SHORT_RECONCILE_TIMEOUT,
7597 0 : &self.cancel,
7598 0 : )
7599 0 : .await;
7600 :
7601 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
7602 0 : want_secondary_status.into_iter().zip(results.into_iter())
7603 : {
7604 0 : match secondary_status {
7605 0 : Err(e) => {
7606 0 : tracing::info!(
7607 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
7608 : );
7609 : }
7610 0 : Ok(progress) => {
7611 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
7612 : // them in an optimization
7613 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
7614 :
7615 0 : if progress.heatmap_mtime.is_none()
7616 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
7617 0 : && progress.bytes_downloaded != progress.bytes_total
7618 0 : || progress.bytes_total - progress.bytes_downloaded
7619 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
7620 : {
7621 0 : tracing::info!(
7622 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
7623 : );
7624 :
7625 : #[cfg(feature = "testing")]
7626 0 : if progress.heatmap_mtime.is_none() {
7627 : // No heatmap might mean the attached location has never uploaded one, or that
7628 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
7629 : // but fairly common in tests.
7630 0 : self.kick_secondary_download(tenant_shard_id).await;
7631 0 : }
7632 : } else {
7633 : // Location looks ready: proceed
7634 0 : tracing::info!(
7635 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
7636 : );
7637 0 : validated_work.push((tenant_shard_id, optimization))
7638 : }
7639 : }
7640 : }
7641 : }
7642 :
7643 0 : validated_work
7644 0 : }
7645 :
7646 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
7647 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
7648 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
7649 : /// we have this helper to move things along faster.
7650 : #[cfg(feature = "testing")]
7651 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
7652 0 : let (attached_node, secondaries) = {
7653 0 : let locked = self.inner.read().unwrap();
7654 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
7655 0 : tracing::warn!(
7656 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
7657 : );
7658 0 : return;
7659 : };
7660 :
7661 0 : let Some(attached) = shard.intent.get_attached() else {
7662 0 : tracing::warn!(
7663 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
7664 : );
7665 0 : return;
7666 : };
7667 :
7668 0 : let secondaries = shard
7669 0 : .intent
7670 0 : .get_secondary()
7671 0 : .iter()
7672 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
7673 0 : .collect::<Vec<_>>();
7674 0 :
7675 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
7676 0 : };
7677 0 :
7678 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
7679 0 : // a 'kick' to let scheduling optimisation run more promptly.
7680 0 : match attached_node
7681 0 : .with_client_retries(
7682 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
7683 0 : &self.http_client,
7684 0 : &self.config.pageserver_jwt_token,
7685 0 : 3,
7686 0 : 10,
7687 0 : SHORT_RECONCILE_TIMEOUT,
7688 0 : &self.cancel,
7689 0 : )
7690 0 : .await
7691 : {
7692 0 : Some(Err(e)) => {
7693 0 : tracing::info!(
7694 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
7695 : );
7696 : }
7697 : None => {
7698 0 : tracing::info!(
7699 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
7700 : );
7701 : }
7702 : Some(Ok(_)) => {
7703 0 : tracing::info!(
7704 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
7705 : );
7706 : }
7707 : }
7708 :
7709 0 : for secondary_node in secondaries {
7710 0 : match secondary_node
7711 0 : .with_client_retries(
7712 0 : |client| async move {
7713 0 : client
7714 0 : .tenant_secondary_download(
7715 0 : tenant_shard_id,
7716 0 : Some(Duration::from_secs(1)),
7717 0 : )
7718 0 : .await
7719 0 : },
7720 0 : &self.http_client,
7721 0 : &self.config.pageserver_jwt_token,
7722 0 : 3,
7723 0 : 10,
7724 0 : SHORT_RECONCILE_TIMEOUT,
7725 0 : &self.cancel,
7726 0 : )
7727 0 : .await
7728 : {
7729 0 : Some(Err(e)) => {
7730 0 : tracing::info!(
7731 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
7732 : );
7733 : }
7734 : None => {
7735 0 : tracing::info!(
7736 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
7737 : );
7738 : }
7739 0 : Some(Ok(progress)) => {
7740 0 : tracing::info!(
7741 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
7742 : );
7743 : }
7744 : }
7745 : }
7746 0 : }
7747 :
7748 : /// Asynchronously split a tenant that's eligible for automatic splits. At most one tenant will
7749 : /// be split per call.
7750 : ///
7751 : /// Two sets of criteria are used: initial splits and size-based splits (in that order).
7752 : /// Initial splits are used to eagerly split unsharded tenants that may be performing initial
7753 : /// ingestion, since sharded tenants have significantly better ingestion throughput. Size-based
7754 : /// splits are used to bound the maximum shard size and balance out load.
7755 : ///
7756 : /// Splits are based on max_logical_size, i.e. the logical size of the largest timeline in a
7757 : /// tenant. We use this instead of the total logical size because branches will duplicate
7758 : /// logical size without actually using more storage. We could also use visible physical size,
7759 : /// but this might overestimate tenants that frequently churn branches.
7760 : ///
7761 : /// Initial splits (initial_split_threshold):
7762 : /// * Applies to tenants with 1 shard.
7763 : /// * The largest timeline (max_logical_size) exceeds initial_split_threshold.
7764 : /// * Splits into initial_split_shards.
7765 : ///
7766 : /// Size-based splits (split_threshold):
7767 : /// * Applies to all tenants.
7768 : /// * The largest timeline (max_logical_size) divided by shard count exceeds split_threshold.
7769 : /// * Splits such that max_logical_size / shard_count <= split_threshold, in powers of 2.
7770 : ///
7771 : /// Tenant shards are ordered by descending max_logical_size, first initial split candidates
7772 : /// then size-based split candidates. The first matching candidate is split.
7773 : ///
7774 : /// The shard count is clamped to max_split_shards. If a candidate is eligible for both initial
7775 : /// and size-based splits, the largest shard count will be used.
7776 : ///
7777 : /// An unsharded tenant will get DEFAULT_STRIPE_SIZE, regardless of what its ShardIdentity says.
7778 : /// A sharded tenant will retain its stripe size, as splits do not allow changing it.
7779 : ///
7780 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
7781 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
7782 : /// block all other splits.
7783 0 : async fn autosplit_tenants(self: &Arc<Self>) {
7784 0 : // If max_split_shards is set to 0 or 1, we can't split.
7785 0 : let max_split_shards = self.config.max_split_shards;
7786 0 : if max_split_shards <= 1 {
7787 0 : return;
7788 0 : }
7789 0 :
7790 0 : // If initial_split_shards is set to 0 or 1, disable initial splits.
7791 0 : let mut initial_split_threshold = self.config.initial_split_threshold.unwrap_or(0);
7792 0 : let initial_split_shards = self.config.initial_split_shards;
7793 0 : if initial_split_shards <= 1 {
7794 0 : initial_split_threshold = 0;
7795 0 : }
7796 :
7797 : // If no split_threshold nor initial_split_threshold, disable autosplits.
7798 0 : let split_threshold = self.config.split_threshold.unwrap_or(0);
7799 0 : if split_threshold == 0 && initial_split_threshold == 0 {
7800 0 : return;
7801 0 : }
7802 0 :
7803 0 : // Fetch split candidates in prioritized order.
7804 0 : //
7805 0 : // If initial splits are enabled, fetch eligible tenants first. We prioritize initial splits
7806 0 : // over size-based splits, since these are often performing initial ingestion and rely on
7807 0 : // splits to improve ingest throughput.
7808 0 : let mut candidates = Vec::new();
7809 0 :
7810 0 : if initial_split_threshold > 0 {
7811 : // Initial splits: fetch tenants with 1 shard where the logical size of the largest
7812 : // timeline exceeds the initial split threshold.
7813 0 : let initial_candidates = self
7814 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
7815 0 : order_by: TenantSorting::MaxLogicalSize,
7816 0 : limit: 10,
7817 0 : where_shards_lt: Some(ShardCount(2)),
7818 0 : where_gt: Some(initial_split_threshold),
7819 0 : })
7820 0 : .await;
7821 0 : candidates.extend(initial_candidates);
7822 0 : }
7823 :
7824 0 : if split_threshold > 0 {
7825 : // Size-based splits: fetch tenants where the logical size of the largest timeline
7826 : // divided by shard count exceeds the split threshold.
7827 : //
7828 : // max_logical_size is only tracked on shard 0, and contains the total logical size
7829 : // across all shards. We have to order and filter by MaxLogicalSizePerShard, i.e.
7830 : // max_logical_size / shard_count, such that we only receive tenants that are actually
7831 : // eligible for splits. But we still use max_logical_size for later split calculations.
7832 0 : let size_candidates = self
7833 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
7834 0 : order_by: TenantSorting::MaxLogicalSizePerShard,
7835 0 : limit: 10,
7836 0 : where_shards_lt: Some(ShardCount(max_split_shards)),
7837 0 : where_gt: Some(split_threshold),
7838 0 : })
7839 0 : .await;
7840 : #[cfg(feature = "testing")]
7841 0 : assert!(
7842 0 : size_candidates.iter().all(|c| c.id.is_shard_zero()),
7843 0 : "MaxLogicalSizePerShard returned non-zero shard: {size_candidates:?}",
7844 : );
7845 0 : candidates.extend(size_candidates);
7846 0 : }
7847 :
7848 : // Filter out tenants in a prohibiting scheduling mode.
7849 0 : {
7850 0 : let state = self.inner.read().unwrap();
7851 0 : candidates.retain(|i| {
7852 0 : let policy = state.tenants.get(&i.id).map(|s| s.get_scheduling_policy());
7853 0 : policy == Some(ShardSchedulingPolicy::Active)
7854 0 : });
7855 0 : }
7856 :
7857 : // Pick the first candidate to split. This will generally always be the first one in
7858 : // candidates, but we defensively skip candidates that end up not actually splitting.
7859 0 : let Some((candidate, new_shard_count)) = candidates
7860 0 : .into_iter()
7861 0 : .filter_map(|candidate| {
7862 0 : let new_shard_count = Self::compute_split_shards(ShardSplitInputs {
7863 0 : shard_count: candidate.id.shard_count,
7864 0 : max_logical_size: candidate.max_logical_size,
7865 0 : split_threshold,
7866 0 : max_split_shards,
7867 0 : initial_split_threshold,
7868 0 : initial_split_shards,
7869 0 : });
7870 0 : new_shard_count.map(|shards| (candidate, shards.count()))
7871 0 : })
7872 0 : .next()
7873 : else {
7874 0 : debug!("no split-eligible tenants found");
7875 0 : return;
7876 : };
7877 :
7878 : // Retain the stripe size of sharded tenants, as splits don't allow changing it. Otherwise,
7879 : // use DEFAULT_STRIPE_SIZE for unsharded tenants -- their stripe size doesn't really matter,
7880 : // and if we change the default stripe size we want to use the new default rather than an
7881 : // old, persisted stripe size.
7882 0 : let new_stripe_size = match candidate.id.shard_count.count() {
7883 0 : 0 => panic!("invalid shard count 0"),
7884 0 : 1 => Some(DEFAULT_STRIPE_SIZE),
7885 0 : 2.. => None,
7886 : };
7887 :
7888 : // We spawn a task to run this, so it's exactly like some external API client requesting
7889 : // it. We don't want to block the background reconcile loop on this.
7890 0 : let old_shard_count = candidate.id.shard_count.count();
7891 0 : info!(
7892 0 : "auto-splitting tenant {old_shard_count} → {new_shard_count} shards, \
7893 0 : current size {candidate:?} (split_threshold={split_threshold} \
7894 0 : initial_split_threshold={initial_split_threshold})"
7895 : );
7896 :
7897 0 : let this = self.clone();
7898 0 : tokio::spawn(
7899 0 : async move {
7900 0 : match this
7901 0 : .tenant_shard_split(
7902 0 : candidate.id.tenant_id,
7903 0 : TenantShardSplitRequest {
7904 0 : new_shard_count,
7905 0 : new_stripe_size,
7906 0 : },
7907 0 : )
7908 0 : .await
7909 : {
7910 : Ok(_) => {
7911 0 : info!("successful auto-split {old_shard_count} → {new_shard_count} shards")
7912 : }
7913 0 : Err(err) => error!("auto-split failed: {err}"),
7914 : }
7915 0 : }
7916 0 : .instrument(info_span!("auto_split", tenant_id=%candidate.id.tenant_id)),
7917 : );
7918 0 : }
7919 :
7920 : /// Returns the number of shards to split a tenant into, or None if the tenant shouldn't split,
7921 : /// based on the total logical size of the largest timeline summed across all shards. Uses the
7922 : /// larger of size-based and initial splits, clamped to max_split_shards.
7923 : ///
7924 : /// NB: the thresholds are exclusive, since TopTenantShardsRequest uses where_gt.
7925 25 : fn compute_split_shards(inputs: ShardSplitInputs) -> Option<ShardCount> {
7926 25 : let ShardSplitInputs {
7927 25 : shard_count,
7928 25 : max_logical_size,
7929 25 : split_threshold,
7930 25 : max_split_shards,
7931 25 : initial_split_threshold,
7932 25 : initial_split_shards,
7933 25 : } = inputs;
7934 25 :
7935 25 : let mut new_shard_count: u8 = shard_count.count();
7936 25 :
7937 25 : // Size-based splits. Ensures max_logical_size / new_shard_count <= split_threshold, using
7938 25 : // power-of-two shard counts.
7939 25 : //
7940 25 : // If the current shard count is not a power of two, and does not exceed split_threshold,
7941 25 : // then we leave it alone rather than forcing a power-of-two split.
7942 25 : if split_threshold > 0
7943 18 : && max_logical_size.div_ceil(split_threshold) > shard_count.count() as u64
7944 12 : {
7945 12 : new_shard_count = max_logical_size
7946 12 : .div_ceil(split_threshold)
7947 12 : .checked_next_power_of_two()
7948 12 : .unwrap_or(u8::MAX as u64)
7949 12 : .try_into()
7950 12 : .unwrap_or(u8::MAX);
7951 13 : }
7952 :
7953 : // Initial splits. Use the larger of size-based and initial split shard counts. This only
7954 : // applies to unsharded tenants, i.e. changes to initial_split_threshold or
7955 : // initial_split_shards are not retroactive for sharded tenants.
7956 25 : if initial_split_threshold > 0
7957 14 : && shard_count.count() <= 1
7958 11 : && max_logical_size > initial_split_threshold
7959 8 : {
7960 8 : new_shard_count = new_shard_count.max(initial_split_shards);
7961 17 : }
7962 :
7963 : // Clamp to max shards.
7964 25 : new_shard_count = new_shard_count.min(max_split_shards);
7965 25 :
7966 25 : // Don't split if we're not increasing the shard count.
7967 25 : if new_shard_count <= shard_count.count() {
7968 10 : return None;
7969 15 : }
7970 15 :
7971 15 : Some(ShardCount(new_shard_count))
7972 25 : }
7973 :
7974 : /// Fetches the top tenant shards from every node, in descending order of
7975 : /// max logical size. Any node errors will be logged and ignored.
7976 0 : async fn get_top_tenant_shards(
7977 0 : &self,
7978 0 : request: &TopTenantShardsRequest,
7979 0 : ) -> Vec<TopTenantShardItem> {
7980 0 : let nodes = self
7981 0 : .inner
7982 0 : .read()
7983 0 : .unwrap()
7984 0 : .nodes
7985 0 : .values()
7986 0 : .cloned()
7987 0 : .collect_vec();
7988 0 :
7989 0 : let mut futures = FuturesUnordered::new();
7990 0 : for node in nodes {
7991 0 : futures.push(async move {
7992 0 : node.with_client_retries(
7993 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
7994 0 : &self.http_client,
7995 0 : &self.config.pageserver_jwt_token,
7996 0 : 3,
7997 0 : 3,
7998 0 : Duration::from_secs(5),
7999 0 : &self.cancel,
8000 0 : )
8001 0 : .await
8002 0 : });
8003 0 : }
8004 :
8005 0 : let mut top = Vec::new();
8006 0 : while let Some(output) = futures.next().await {
8007 0 : match output {
8008 0 : Some(Ok(response)) => top.extend(response.shards),
8009 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
8010 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
8011 0 : None => {} // node is shutting down
8012 : }
8013 : }
8014 :
8015 0 : top.sort_by_key(|i| i.max_logical_size);
8016 0 : top.reverse();
8017 0 : top
8018 0 : }
8019 :
8020 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
8021 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
8022 : /// put the system into a quiescent state where future background reconciliations won't do anything.
8023 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
8024 0 : let reconciles_spawned = self.reconcile_all();
8025 0 : let reconciles_spawned = if reconciles_spawned == 0 {
8026 : // Only optimize when we are otherwise idle
8027 0 : self.optimize_all().await
8028 : } else {
8029 0 : reconciles_spawned
8030 : };
8031 :
8032 0 : let waiters = {
8033 0 : let mut waiters = Vec::new();
8034 0 : let locked = self.inner.read().unwrap();
8035 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
8036 0 : if let Some(waiter) = shard.get_waiter() {
8037 0 : waiters.push(waiter);
8038 0 : }
8039 : }
8040 0 : waiters
8041 0 : };
8042 0 :
8043 0 : let waiter_count = waiters.len();
8044 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
8045 0 : Ok(()) => {}
8046 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
8047 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
8048 0 : {
8049 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
8050 0 : // because some other change superceded it. We will return a nonzero number,
8051 0 : // so the caller knows they might have to call again to quiesce the system.
8052 0 : }
8053 0 : Err(e) => {
8054 0 : return Err(e);
8055 : }
8056 : };
8057 :
8058 0 : tracing::info!(
8059 0 : "{} reconciles in reconcile_all, {} waiters",
8060 : reconciles_spawned,
8061 : waiter_count
8062 : );
8063 :
8064 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
8065 0 : }
8066 :
8067 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
8068 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
8069 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
8070 0 : self.reconcilers_cancel.cancel();
8071 0 : self.reconcilers_gate.close().await;
8072 :
8073 : // Signal the background loop in [`Service::process_results`] to exit once
8074 : // it has proccessed the results from all the reconciles we cancelled earlier.
8075 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
8076 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
8077 0 : self.result_tx.closed().await;
8078 0 : }
8079 :
8080 0 : pub async fn shutdown(&self) {
8081 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
8082 0 : .await;
8083 :
8084 : // Background tasks hold gate guards: this notifies them of the cancellation and
8085 : // waits for them all to complete.
8086 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
8087 0 : self.cancel.cancel();
8088 0 : self.gate.close().await;
8089 0 : }
8090 :
8091 : /// Spot check the download lag for a secondary location of a shard.
8092 : /// Should be used as a heuristic, since it's not always precise: the
8093 : /// secondary might have not downloaded the new heat map yet and, hence,
8094 : /// is not aware of the lag.
8095 : ///
8096 : /// Returns:
8097 : /// * Ok(None) if the lag could not be determined from the status,
8098 : /// * Ok(Some(_)) if the lag could be determind
8099 : /// * Err on failures to query the pageserver.
8100 0 : async fn secondary_lag(
8101 0 : &self,
8102 0 : secondary: &NodeId,
8103 0 : tenant_shard_id: TenantShardId,
8104 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
8105 0 : let nodes = self.inner.read().unwrap().nodes.clone();
8106 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
8107 0 : StatusCode::NOT_FOUND,
8108 0 : format!("Node with id {} not found", secondary),
8109 0 : ))?;
8110 :
8111 0 : match node
8112 0 : .with_client_retries(
8113 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
8114 0 : &self.http_client,
8115 0 : &self.config.pageserver_jwt_token,
8116 0 : 1,
8117 0 : 3,
8118 0 : Duration::from_millis(250),
8119 0 : &self.cancel,
8120 0 : )
8121 0 : .await
8122 : {
8123 0 : Some(Ok(status)) => match status.heatmap_mtime {
8124 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
8125 0 : None => Ok(None),
8126 : },
8127 0 : Some(Err(e)) => Err(e),
8128 0 : None => Err(mgmt_api::Error::Cancelled),
8129 : }
8130 0 : }
8131 :
8132 : /// Drain a node by moving the shards attached to it as primaries.
8133 : /// This is a long running operation and it should run as a separate Tokio task.
8134 0 : pub(crate) async fn drain_node(
8135 0 : self: &Arc<Self>,
8136 0 : node_id: NodeId,
8137 0 : cancel: CancellationToken,
8138 0 : ) -> Result<(), OperationError> {
8139 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
8140 0 : let max_secondary_lag_bytes = self
8141 0 : .config
8142 0 : .max_secondary_lag_bytes
8143 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
8144 :
8145 : // By default, live migrations are generous about the wait time for getting
8146 : // the secondary location up to speed. When draining, give up earlier in order
8147 : // to not stall the operation when a cold secondary is encountered.
8148 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
8149 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
8150 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
8151 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
8152 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
8153 0 : .build();
8154 0 :
8155 0 : let mut waiters = Vec::new();
8156 0 :
8157 0 : let mut tid_iter = TenantShardIterator::new({
8158 0 : let service = self.clone();
8159 0 : move |last_inspected_shard: Option<TenantShardId>| {
8160 0 : let locked = &service.inner.read().unwrap();
8161 0 : let tenants = &locked.tenants;
8162 0 : let entry = match last_inspected_shard {
8163 0 : Some(skip_past) => {
8164 0 : // Skip to the last seen tenant shard id
8165 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
8166 0 :
8167 0 : // Skip past the last seen
8168 0 : cursor.nth(1)
8169 : }
8170 0 : None => tenants.first_key_value(),
8171 : };
8172 :
8173 0 : entry.map(|(tid, _)| tid).copied()
8174 0 : }
8175 0 : });
8176 :
8177 0 : while !tid_iter.finished() {
8178 0 : if cancel.is_cancelled() {
8179 0 : match self
8180 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8181 0 : .await
8182 : {
8183 0 : Ok(()) => return Err(OperationError::Cancelled),
8184 0 : Err(err) => {
8185 0 : return Err(OperationError::FinalizeError(
8186 0 : format!(
8187 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8188 0 : node_id, err
8189 0 : )
8190 0 : .into(),
8191 0 : ));
8192 : }
8193 : }
8194 0 : }
8195 0 :
8196 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
8197 :
8198 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
8199 0 : let tid = match tid_iter.next() {
8200 0 : Some(tid) => tid,
8201 : None => {
8202 0 : break;
8203 : }
8204 : };
8205 :
8206 0 : let tid_drain = TenantShardDrain {
8207 0 : drained_node: node_id,
8208 0 : tenant_shard_id: tid,
8209 0 : };
8210 :
8211 0 : let dest_node_id = {
8212 0 : let locked = self.inner.read().unwrap();
8213 0 :
8214 0 : match tid_drain
8215 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
8216 : {
8217 0 : Some(node_id) => node_id,
8218 : None => {
8219 0 : continue;
8220 : }
8221 : }
8222 : };
8223 :
8224 0 : match self.secondary_lag(&dest_node_id, tid).await {
8225 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
8226 0 : // The secondary is reasonably up to date.
8227 0 : // Migrate to it
8228 0 : }
8229 0 : Ok(Some(lag)) => {
8230 0 : tracing::info!(
8231 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8232 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
8233 : );
8234 0 : continue;
8235 : }
8236 : Ok(None) => {
8237 0 : tracing::info!(
8238 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8239 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
8240 : );
8241 0 : continue;
8242 : }
8243 0 : Err(err) => {
8244 0 : tracing::warn!(
8245 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8246 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
8247 : );
8248 0 : continue;
8249 : }
8250 : }
8251 :
8252 : {
8253 0 : let mut locked = self.inner.write().unwrap();
8254 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8255 0 : let rescheduled = tid_drain.reschedule_to_secondary(
8256 0 : dest_node_id,
8257 0 : tenants,
8258 0 : scheduler,
8259 0 : nodes,
8260 0 : )?;
8261 :
8262 0 : if let Some(tenant_shard) = rescheduled {
8263 0 : let waiter = self.maybe_configured_reconcile_shard(
8264 0 : tenant_shard,
8265 0 : nodes,
8266 0 : reconciler_config,
8267 0 : );
8268 0 : if let Some(some) = waiter {
8269 0 : waiters.push(some);
8270 0 : }
8271 0 : }
8272 : }
8273 : }
8274 :
8275 0 : waiters = self
8276 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
8277 0 : .await;
8278 :
8279 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
8280 : }
8281 :
8282 0 : while !waiters.is_empty() {
8283 0 : if cancel.is_cancelled() {
8284 0 : match self
8285 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8286 0 : .await
8287 : {
8288 0 : Ok(()) => return Err(OperationError::Cancelled),
8289 0 : Err(err) => {
8290 0 : return Err(OperationError::FinalizeError(
8291 0 : format!(
8292 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8293 0 : node_id, err
8294 0 : )
8295 0 : .into(),
8296 0 : ));
8297 : }
8298 : }
8299 0 : }
8300 0 :
8301 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
8302 :
8303 0 : waiters = self
8304 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
8305 0 : .await;
8306 : }
8307 :
8308 : // At this point we have done the best we could to drain shards from this node.
8309 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
8310 : // to complete the drain.
8311 0 : if let Err(err) = self
8312 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
8313 0 : .await
8314 : {
8315 : // This is not fatal. Anything that is polling the node scheduling policy to detect
8316 : // the end of the drain operations will hang, but all such places should enforce an
8317 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
8318 : // by the counterpart fill operation.
8319 0 : return Err(OperationError::FinalizeError(
8320 0 : format!(
8321 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
8322 0 : )
8323 0 : .into(),
8324 0 : ));
8325 0 : }
8326 0 :
8327 0 : Ok(())
8328 0 : }
8329 :
8330 : /// Create a node fill plan (pick secondaries to promote), based on:
8331 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
8332 : /// outside their home AZ, should be migrated back here.
8333 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
8334 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
8335 : /// those shards have a home AZ that doesn't match the node we're filling.
8336 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
8337 0 : let mut locked = self.inner.write().unwrap();
8338 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
8339 0 :
8340 0 : let node_az = nodes
8341 0 : .get(&node_id)
8342 0 : .expect("Node must exist")
8343 0 : .get_availability_zone_id()
8344 0 : .clone();
8345 0 :
8346 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
8347 0 : let mut plan = Vec::new();
8348 0 :
8349 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
8350 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
8351 0 :
8352 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
8353 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
8354 0 : let respect_azs = nodes
8355 0 : .values()
8356 0 : .map(|n| n.get_availability_zone_id())
8357 0 : .unique()
8358 0 : .count()
8359 0 : > 1;
8360 :
8361 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
8362 : // requires it.
8363 0 : for (tsid, tenant_shard) in tenants {
8364 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
8365 : // Shard doesn't have a secondary on this node, ignore it.
8366 0 : continue;
8367 0 : }
8368 0 :
8369 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
8370 0 : // shards which belong on this node, not to promote shards whose scheduling preference
8371 0 : // would be on their currently attached node. So will avoid promoting shards whose
8372 0 : // home AZ doesn't match the AZ of the node we're filling.
8373 0 : match tenant_shard.preferred_az() {
8374 0 : _ if !respect_azs => {
8375 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8376 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
8377 0 : }
8378 : }
8379 : None => {
8380 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
8381 : // will only do so if our target shard count requires it.
8382 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8383 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
8384 0 : }
8385 : }
8386 0 : Some(az) if az == &node_az => {
8387 : // This shard's home AZ is equal to the node we're filling: it should
8388 : // be moved back to this node as part of filling, unless its currently
8389 : // attached location is also in its home AZ.
8390 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8391 0 : if nodes
8392 0 : .get(primary)
8393 0 : .expect("referenced node must exist")
8394 0 : .get_availability_zone_id()
8395 0 : != tenant_shard
8396 0 : .preferred_az()
8397 0 : .expect("tenant must have an AZ preference")
8398 : {
8399 0 : plan.push(*tsid)
8400 0 : }
8401 : } else {
8402 0 : plan.push(*tsid)
8403 : }
8404 : }
8405 0 : Some(_) => {
8406 0 : // This shard's home AZ is somewhere other than the node we're filling,
8407 0 : // it may not be moved back to this node as part of filling. Ignore it
8408 0 : }
8409 : }
8410 : }
8411 :
8412 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
8413 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
8414 0 :
8415 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
8416 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
8417 0 :
8418 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
8419 :
8420 0 : for (node_id, attached) in nodes_by_load {
8421 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
8422 0 : if !available {
8423 0 : continue;
8424 0 : }
8425 0 :
8426 0 : if plan.len() >= fill_requirement
8427 0 : || free_tids_by_node.is_empty()
8428 0 : || attached <= expected_attached
8429 : {
8430 0 : break;
8431 0 : }
8432 0 :
8433 0 : let can_take = attached - expected_attached;
8434 0 : let needed = fill_requirement - plan.len();
8435 0 : let mut take = std::cmp::min(can_take, needed);
8436 0 :
8437 0 : let mut remove_node = false;
8438 0 : while take > 0 {
8439 0 : match free_tids_by_node.get_mut(&node_id) {
8440 0 : Some(tids) => match tids.pop() {
8441 0 : Some(tid) => {
8442 0 : let max_promote_for_tenant = std::cmp::max(
8443 0 : tid.shard_count.count() as usize / locked.nodes.len(),
8444 0 : 1,
8445 0 : );
8446 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
8447 0 : if *promoted < max_promote_for_tenant {
8448 0 : plan.push(tid);
8449 0 : *promoted += 1;
8450 0 : take -= 1;
8451 0 : }
8452 : }
8453 : None => {
8454 0 : remove_node = true;
8455 0 : break;
8456 : }
8457 : },
8458 : None => {
8459 0 : break;
8460 : }
8461 : }
8462 : }
8463 :
8464 0 : if remove_node {
8465 0 : free_tids_by_node.remove(&node_id);
8466 0 : }
8467 : }
8468 :
8469 0 : plan
8470 0 : }
8471 :
8472 : /// Fill a node by promoting its secondaries until the cluster is balanced
8473 : /// with regards to attached shard counts. Note that this operation only
8474 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
8475 : /// This is a long running operation and it should run as a separate Tokio task.
8476 0 : pub(crate) async fn fill_node(
8477 0 : &self,
8478 0 : node_id: NodeId,
8479 0 : cancel: CancellationToken,
8480 0 : ) -> Result<(), OperationError> {
8481 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
8482 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
8483 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
8484 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
8485 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
8486 0 : .build();
8487 0 :
8488 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
8489 0 : let mut waiters = Vec::new();
8490 :
8491 : // Execute the plan we've composed above. Before aplying each move from the plan,
8492 : // we validate to ensure that it has not gone stale in the meantime.
8493 0 : while !tids_to_promote.is_empty() {
8494 0 : if cancel.is_cancelled() {
8495 0 : match self
8496 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8497 0 : .await
8498 : {
8499 0 : Ok(()) => return Err(OperationError::Cancelled),
8500 0 : Err(err) => {
8501 0 : return Err(OperationError::FinalizeError(
8502 0 : format!(
8503 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8504 0 : node_id, err
8505 0 : )
8506 0 : .into(),
8507 0 : ));
8508 : }
8509 : }
8510 0 : }
8511 0 :
8512 0 : {
8513 0 : let mut locked = self.inner.write().unwrap();
8514 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8515 :
8516 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
8517 0 : format!("node {node_id} was removed").into(),
8518 0 : ))?;
8519 :
8520 0 : let current_policy = node.get_scheduling();
8521 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
8522 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
8523 : // about it
8524 0 : return Err(OperationError::NodeStateChanged(
8525 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
8526 0 : ));
8527 0 : }
8528 :
8529 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
8530 0 : if let Some(tid) = tids_to_promote.pop() {
8531 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
8532 : // If the node being filled is not a secondary anymore,
8533 : // skip the promotion.
8534 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
8535 0 : continue;
8536 0 : }
8537 0 :
8538 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
8539 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
8540 0 : Err(e) => {
8541 0 : tracing::warn!(
8542 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8543 0 : "Scheduling error when filling pageserver {} : {e}", node_id
8544 : );
8545 : }
8546 : Ok(()) => {
8547 0 : tracing::info!(
8548 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8549 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
8550 : node_id,
8551 : previously_attached_to,
8552 : node_id
8553 : );
8554 :
8555 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
8556 0 : tenant_shard,
8557 0 : nodes,
8558 0 : reconciler_config,
8559 0 : ) {
8560 0 : waiters.push(waiter);
8561 0 : }
8562 : }
8563 : }
8564 0 : }
8565 : } else {
8566 0 : break;
8567 : }
8568 : }
8569 : }
8570 :
8571 0 : waiters = self
8572 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
8573 0 : .await;
8574 : }
8575 :
8576 0 : while !waiters.is_empty() {
8577 0 : if cancel.is_cancelled() {
8578 0 : match self
8579 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8580 0 : .await
8581 : {
8582 0 : Ok(()) => return Err(OperationError::Cancelled),
8583 0 : Err(err) => {
8584 0 : return Err(OperationError::FinalizeError(
8585 0 : format!(
8586 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8587 0 : node_id, err
8588 0 : )
8589 0 : .into(),
8590 0 : ));
8591 : }
8592 : }
8593 0 : }
8594 0 :
8595 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
8596 :
8597 0 : waiters = self
8598 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
8599 0 : .await;
8600 : }
8601 :
8602 0 : if let Err(err) = self
8603 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8604 0 : .await
8605 : {
8606 : // This isn't a huge issue since the filling process starts upon request. However, it
8607 : // will prevent the next drain from starting. The only case in which this can fail
8608 : // is database unavailability. Such a case will require manual intervention.
8609 0 : return Err(OperationError::FinalizeError(
8610 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
8611 0 : .into(),
8612 0 : ));
8613 0 : }
8614 0 :
8615 0 : Ok(())
8616 0 : }
8617 :
8618 : /// Updates scrubber metadata health check results.
8619 0 : pub(crate) async fn metadata_health_update(
8620 0 : &self,
8621 0 : update_req: MetadataHealthUpdateRequest,
8622 0 : ) -> Result<(), ApiError> {
8623 0 : let now = chrono::offset::Utc::now();
8624 0 : let (healthy_records, unhealthy_records) = {
8625 0 : let locked = self.inner.read().unwrap();
8626 0 : let healthy_records = update_req
8627 0 : .healthy_tenant_shards
8628 0 : .into_iter()
8629 0 : // Retain only health records associated with tenant shards managed by storage controller.
8630 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
8631 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
8632 0 : .collect();
8633 0 : let unhealthy_records = update_req
8634 0 : .unhealthy_tenant_shards
8635 0 : .into_iter()
8636 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
8637 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
8638 0 : .collect();
8639 0 :
8640 0 : (healthy_records, unhealthy_records)
8641 0 : };
8642 0 :
8643 0 : self.persistence
8644 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
8645 0 : .await?;
8646 0 : Ok(())
8647 0 : }
8648 :
8649 : /// Lists the tenant shards that has unhealthy metadata status.
8650 0 : pub(crate) async fn metadata_health_list_unhealthy(
8651 0 : &self,
8652 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
8653 0 : let result = self
8654 0 : .persistence
8655 0 : .list_unhealthy_metadata_health_records()
8656 0 : .await?
8657 0 : .iter()
8658 0 : .map(|p| p.get_tenant_shard_id().unwrap())
8659 0 : .collect();
8660 0 :
8661 0 : Ok(result)
8662 0 : }
8663 :
8664 : /// Lists the tenant shards that have not been scrubbed for some duration.
8665 0 : pub(crate) async fn metadata_health_list_outdated(
8666 0 : &self,
8667 0 : not_scrubbed_for: Duration,
8668 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
8669 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
8670 0 : let result = self
8671 0 : .persistence
8672 0 : .list_outdated_metadata_health_records(earlier)
8673 0 : .await?
8674 0 : .into_iter()
8675 0 : .map(|record| record.into())
8676 0 : .collect();
8677 0 : Ok(result)
8678 0 : }
8679 :
8680 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
8681 0 : self.inner.read().unwrap().get_leadership_status()
8682 0 : }
8683 :
8684 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
8685 0 : tracing::info!("Received step down request from peer");
8686 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
8687 :
8688 0 : self.inner.write().unwrap().step_down();
8689 0 :
8690 0 : // Wait for reconciliations to stop, or terminate this process if they
8691 0 : // fail to stop in time (this indicates a bug in shutdown)
8692 0 : tokio::select! {
8693 0 : _ = self.stop_reconciliations(StopReconciliationsReason::SteppingDown) => {
8694 0 : tracing::info!("Reconciliations stopped, proceeding with step down");
8695 : }
8696 0 : _ = async {
8697 0 : failpoint_support::sleep_millis_async!("step-down-delay-timeout");
8698 0 : tokio::time::sleep(Duration::from_secs(10)).await
8699 0 : } => {
8700 0 : tracing::warn!("Step down timed out while waiting for reconciliation gate, terminating process");
8701 :
8702 : // The caller may proceed to act as leader when it sees this request fail: reduce the chance
8703 : // of a split-brain situation by terminating this controller instead of leaving it up in a partially-shut-down state.
8704 0 : std::process::exit(1);
8705 : }
8706 : }
8707 :
8708 0 : let mut global_observed = GlobalObservedState::default();
8709 0 : let locked = self.inner.read().unwrap();
8710 0 : for (tid, tenant_shard) in locked.tenants.iter() {
8711 0 : global_observed
8712 0 : .0
8713 0 : .insert(*tid, tenant_shard.observed.clone());
8714 0 : }
8715 :
8716 0 : global_observed
8717 0 : }
8718 :
8719 0 : pub(crate) async fn update_shards_preferred_azs(
8720 0 : &self,
8721 0 : req: ShardsPreferredAzsRequest,
8722 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
8723 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
8724 0 : let updated = self
8725 0 : .persistence
8726 0 : .set_tenant_shard_preferred_azs(preferred_azs)
8727 0 : .await
8728 0 : .map_err(|err| {
8729 0 : ApiError::InternalServerError(anyhow::anyhow!(
8730 0 : "Failed to persist preferred AZs: {err}"
8731 0 : ))
8732 0 : })?;
8733 :
8734 0 : let mut updated_in_mem_and_db = Vec::default();
8735 0 :
8736 0 : let mut locked = self.inner.write().unwrap();
8737 0 : let state = locked.deref_mut();
8738 0 : for (tid, az_id) in updated {
8739 0 : let shard = state.tenants.get_mut(&tid);
8740 0 : if let Some(shard) = shard {
8741 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
8742 0 : updated_in_mem_and_db.push(tid);
8743 0 : }
8744 : }
8745 :
8746 0 : Ok(ShardsPreferredAzsResponse {
8747 0 : updated: updated_in_mem_and_db,
8748 0 : })
8749 0 : }
8750 : }
8751 :
8752 : #[cfg(test)]
8753 : mod tests {
8754 : use super::*;
8755 :
8756 : /// Tests Service::compute_split_shards. For readability, this specifies sizes in GBs rather
8757 : /// than bytes. Note that max_logical_size is the total logical size of the largest timeline
8758 : /// summed across all shards.
8759 : #[test]
8760 1 : fn compute_split_shards() {
8761 1 : // Size-based split: two shards have a 500 GB timeline, which need to split into 8 shards
8762 1 : // that are <= 64 GB,
8763 1 : assert_eq!(
8764 1 : Service::compute_split_shards(ShardSplitInputs {
8765 1 : shard_count: ShardCount(2),
8766 1 : max_logical_size: 500,
8767 1 : split_threshold: 64,
8768 1 : max_split_shards: 16,
8769 1 : initial_split_threshold: 0,
8770 1 : initial_split_shards: 0,
8771 1 : }),
8772 1 : Some(ShardCount(8))
8773 1 : );
8774 :
8775 : // Size-based split: noop at or below threshold, fires above.
8776 1 : assert_eq!(
8777 1 : Service::compute_split_shards(ShardSplitInputs {
8778 1 : shard_count: ShardCount(2),
8779 1 : max_logical_size: 127,
8780 1 : split_threshold: 64,
8781 1 : max_split_shards: 16,
8782 1 : initial_split_threshold: 0,
8783 1 : initial_split_shards: 0,
8784 1 : }),
8785 1 : None,
8786 1 : );
8787 1 : assert_eq!(
8788 1 : Service::compute_split_shards(ShardSplitInputs {
8789 1 : shard_count: ShardCount(2),
8790 1 : max_logical_size: 128,
8791 1 : split_threshold: 64,
8792 1 : max_split_shards: 16,
8793 1 : initial_split_threshold: 0,
8794 1 : initial_split_shards: 0,
8795 1 : }),
8796 1 : None,
8797 1 : );
8798 1 : assert_eq!(
8799 1 : Service::compute_split_shards(ShardSplitInputs {
8800 1 : shard_count: ShardCount(2),
8801 1 : max_logical_size: 129,
8802 1 : split_threshold: 64,
8803 1 : max_split_shards: 16,
8804 1 : initial_split_threshold: 0,
8805 1 : initial_split_shards: 0,
8806 1 : }),
8807 1 : Some(ShardCount(4)),
8808 1 : );
8809 :
8810 : // Size-based split: clamped to max_split_shards.
8811 1 : assert_eq!(
8812 1 : Service::compute_split_shards(ShardSplitInputs {
8813 1 : shard_count: ShardCount(2),
8814 1 : max_logical_size: 10000,
8815 1 : split_threshold: 64,
8816 1 : max_split_shards: 16,
8817 1 : initial_split_threshold: 0,
8818 1 : initial_split_shards: 0,
8819 1 : }),
8820 1 : Some(ShardCount(16))
8821 1 : );
8822 :
8823 : // Size-based split: tenant already at or beyond max_split_shards is not split.
8824 1 : assert_eq!(
8825 1 : Service::compute_split_shards(ShardSplitInputs {
8826 1 : shard_count: ShardCount(16),
8827 1 : max_logical_size: 10000,
8828 1 : split_threshold: 64,
8829 1 : max_split_shards: 16,
8830 1 : initial_split_threshold: 0,
8831 1 : initial_split_shards: 0,
8832 1 : }),
8833 1 : None
8834 1 : );
8835 :
8836 1 : assert_eq!(
8837 1 : Service::compute_split_shards(ShardSplitInputs {
8838 1 : shard_count: ShardCount(32),
8839 1 : max_logical_size: 10000,
8840 1 : split_threshold: 64,
8841 1 : max_split_shards: 16,
8842 1 : initial_split_threshold: 0,
8843 1 : initial_split_shards: 0,
8844 1 : }),
8845 1 : None
8846 1 : );
8847 :
8848 : // Size-based split: a non-power-of-2 shard count is normalized to power-of-2 if it
8849 : // exceeds split_threshold (i.e. a 3-shard tenant splits into 8, not 6).
8850 1 : assert_eq!(
8851 1 : Service::compute_split_shards(ShardSplitInputs {
8852 1 : shard_count: ShardCount(3),
8853 1 : max_logical_size: 320,
8854 1 : split_threshold: 64,
8855 1 : max_split_shards: 16,
8856 1 : initial_split_threshold: 0,
8857 1 : initial_split_shards: 0,
8858 1 : }),
8859 1 : Some(ShardCount(8))
8860 1 : );
8861 :
8862 : // Size-based split: a non-power-of-2 shard count is not normalized to power-of-2 if the
8863 : // existing shards are below or at split_threshold, but splits into 4 if it exceeds it.
8864 1 : assert_eq!(
8865 1 : Service::compute_split_shards(ShardSplitInputs {
8866 1 : shard_count: ShardCount(3),
8867 1 : max_logical_size: 191,
8868 1 : split_threshold: 64,
8869 1 : max_split_shards: 16,
8870 1 : initial_split_threshold: 0,
8871 1 : initial_split_shards: 0,
8872 1 : }),
8873 1 : None
8874 1 : );
8875 1 : assert_eq!(
8876 1 : Service::compute_split_shards(ShardSplitInputs {
8877 1 : shard_count: ShardCount(3),
8878 1 : max_logical_size: 192,
8879 1 : split_threshold: 64,
8880 1 : max_split_shards: 16,
8881 1 : initial_split_threshold: 0,
8882 1 : initial_split_shards: 0,
8883 1 : }),
8884 1 : None
8885 1 : );
8886 1 : assert_eq!(
8887 1 : Service::compute_split_shards(ShardSplitInputs {
8888 1 : shard_count: ShardCount(3),
8889 1 : max_logical_size: 193,
8890 1 : split_threshold: 64,
8891 1 : max_split_shards: 16,
8892 1 : initial_split_threshold: 0,
8893 1 : initial_split_shards: 0,
8894 1 : }),
8895 1 : Some(ShardCount(4))
8896 1 : );
8897 :
8898 : // Initial split: tenant has a 10 GB timeline, split into 4 shards.
8899 1 : assert_eq!(
8900 1 : Service::compute_split_shards(ShardSplitInputs {
8901 1 : shard_count: ShardCount(1),
8902 1 : max_logical_size: 10,
8903 1 : split_threshold: 0,
8904 1 : max_split_shards: 16,
8905 1 : initial_split_threshold: 8,
8906 1 : initial_split_shards: 4,
8907 1 : }),
8908 1 : Some(ShardCount(4))
8909 1 : );
8910 :
8911 : // Initial split: 0 ShardCount is equivalent to 1.
8912 1 : assert_eq!(
8913 1 : Service::compute_split_shards(ShardSplitInputs {
8914 1 : shard_count: ShardCount(0),
8915 1 : max_logical_size: 10,
8916 1 : split_threshold: 0,
8917 1 : max_split_shards: 16,
8918 1 : initial_split_threshold: 8,
8919 1 : initial_split_shards: 4,
8920 1 : }),
8921 1 : Some(ShardCount(4))
8922 1 : );
8923 :
8924 : // Initial split: at or below threshold is noop.
8925 1 : assert_eq!(
8926 1 : Service::compute_split_shards(ShardSplitInputs {
8927 1 : shard_count: ShardCount(1),
8928 1 : max_logical_size: 7,
8929 1 : split_threshold: 0,
8930 1 : max_split_shards: 16,
8931 1 : initial_split_threshold: 8,
8932 1 : initial_split_shards: 4,
8933 1 : }),
8934 1 : None,
8935 1 : );
8936 1 : assert_eq!(
8937 1 : Service::compute_split_shards(ShardSplitInputs {
8938 1 : shard_count: ShardCount(1),
8939 1 : max_logical_size: 8,
8940 1 : split_threshold: 0,
8941 1 : max_split_shards: 16,
8942 1 : initial_split_threshold: 8,
8943 1 : initial_split_shards: 4,
8944 1 : }),
8945 1 : None,
8946 1 : );
8947 1 : assert_eq!(
8948 1 : Service::compute_split_shards(ShardSplitInputs {
8949 1 : shard_count: ShardCount(1),
8950 1 : max_logical_size: 9,
8951 1 : split_threshold: 0,
8952 1 : max_split_shards: 16,
8953 1 : initial_split_threshold: 8,
8954 1 : initial_split_shards: 4,
8955 1 : }),
8956 1 : Some(ShardCount(4))
8957 1 : );
8958 :
8959 : // Initial split: already sharded tenant is not affected, even if above threshold and below
8960 : // shard count.
8961 1 : assert_eq!(
8962 1 : Service::compute_split_shards(ShardSplitInputs {
8963 1 : shard_count: ShardCount(2),
8964 1 : max_logical_size: 20,
8965 1 : split_threshold: 0,
8966 1 : max_split_shards: 16,
8967 1 : initial_split_threshold: 8,
8968 1 : initial_split_shards: 4,
8969 1 : }),
8970 1 : None,
8971 1 : );
8972 :
8973 : // Initial split: clamped to max_shards.
8974 1 : assert_eq!(
8975 1 : Service::compute_split_shards(ShardSplitInputs {
8976 1 : shard_count: ShardCount(1),
8977 1 : max_logical_size: 10,
8978 1 : split_threshold: 0,
8979 1 : max_split_shards: 3,
8980 1 : initial_split_threshold: 8,
8981 1 : initial_split_shards: 4,
8982 1 : }),
8983 1 : Some(ShardCount(3)),
8984 1 : );
8985 :
8986 : // Initial+size split: tenant eligible for both will use the larger shard count.
8987 1 : assert_eq!(
8988 1 : Service::compute_split_shards(ShardSplitInputs {
8989 1 : shard_count: ShardCount(1),
8990 1 : max_logical_size: 10,
8991 1 : split_threshold: 64,
8992 1 : max_split_shards: 16,
8993 1 : initial_split_threshold: 8,
8994 1 : initial_split_shards: 4,
8995 1 : }),
8996 1 : Some(ShardCount(4)),
8997 1 : );
8998 1 : assert_eq!(
8999 1 : Service::compute_split_shards(ShardSplitInputs {
9000 1 : shard_count: ShardCount(1),
9001 1 : max_logical_size: 500,
9002 1 : split_threshold: 64,
9003 1 : max_split_shards: 16,
9004 1 : initial_split_threshold: 8,
9005 1 : initial_split_shards: 4,
9006 1 : }),
9007 1 : Some(ShardCount(8)),
9008 1 : );
9009 :
9010 : // Initial+size split: sharded tenant is only eligible for size-based split.
9011 1 : assert_eq!(
9012 1 : Service::compute_split_shards(ShardSplitInputs {
9013 1 : shard_count: ShardCount(2),
9014 1 : max_logical_size: 200,
9015 1 : split_threshold: 64,
9016 1 : max_split_shards: 16,
9017 1 : initial_split_threshold: 8,
9018 1 : initial_split_shards: 8,
9019 1 : }),
9020 1 : Some(ShardCount(4)),
9021 1 : );
9022 :
9023 : // Initial+size split: uses the larger shard count even with initial_split_threshold above
9024 : // split_threshold.
9025 1 : assert_eq!(
9026 1 : Service::compute_split_shards(ShardSplitInputs {
9027 1 : shard_count: ShardCount(1),
9028 1 : max_logical_size: 10,
9029 1 : split_threshold: 4,
9030 1 : max_split_shards: 16,
9031 1 : initial_split_threshold: 8,
9032 1 : initial_split_shards: 8,
9033 1 : }),
9034 1 : Some(ShardCount(8)),
9035 1 : );
9036 :
9037 : // Test backwards compatibility with production settings when initial/size-based splits were
9038 : // rolled out: a single split into 8 shards at 64 GB. Any already sharded tenants with <8
9039 : // shards will split according to split_threshold.
9040 1 : assert_eq!(
9041 1 : Service::compute_split_shards(ShardSplitInputs {
9042 1 : shard_count: ShardCount(1),
9043 1 : max_logical_size: 65,
9044 1 : split_threshold: 64,
9045 1 : max_split_shards: 8,
9046 1 : initial_split_threshold: 64,
9047 1 : initial_split_shards: 8,
9048 1 : }),
9049 1 : Some(ShardCount(8)),
9050 1 : );
9051 :
9052 1 : assert_eq!(
9053 1 : Service::compute_split_shards(ShardSplitInputs {
9054 1 : shard_count: ShardCount(1),
9055 1 : max_logical_size: 64,
9056 1 : split_threshold: 64,
9057 1 : max_split_shards: 8,
9058 1 : initial_split_threshold: 64,
9059 1 : initial_split_shards: 8,
9060 1 : }),
9061 1 : None,
9062 1 : );
9063 :
9064 1 : assert_eq!(
9065 1 : Service::compute_split_shards(ShardSplitInputs {
9066 1 : shard_count: ShardCount(2),
9067 1 : max_logical_size: 129,
9068 1 : split_threshold: 64,
9069 1 : max_split_shards: 8,
9070 1 : initial_split_threshold: 64,
9071 1 : initial_split_shards: 8,
9072 1 : }),
9073 1 : Some(ShardCount(4)),
9074 1 : );
9075 1 : }
9076 : }
|
