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