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