Line data Source code
1 : pub mod chaos_injector;
2 : mod context_iterator;
3 : pub(crate) mod safekeeper_reconciler;
4 :
5 : use std::borrow::Cow;
6 : use std::cmp::Ordering;
7 : use std::collections::{BTreeMap, HashMap, HashSet};
8 : use std::error::Error;
9 : use std::num::NonZeroU32;
10 : use std::ops::{Deref, DerefMut};
11 : use std::path::PathBuf;
12 : use std::str::FromStr;
13 : use std::sync::Arc;
14 : use std::time::{Duration, Instant};
15 :
16 : use anyhow::Context;
17 : use context_iterator::TenantShardContextIterator;
18 : use control_plane::storage_controller::{
19 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
20 : };
21 : use diesel::result::DatabaseErrorKind;
22 : use futures::StreamExt;
23 : use futures::stream::FuturesUnordered;
24 : use http_utils::error::ApiError;
25 : use hyper::Uri;
26 : use itertools::Itertools;
27 : use pageserver_api::controller_api::{
28 : AvailabilityZone, MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability,
29 : NodeRegisterRequest, NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy,
30 : SafekeeperDescribeResponse, ShardSchedulingPolicy, ShardsPreferredAzsRequest,
31 : ShardsPreferredAzsResponse, SkSchedulingPolicy, TenantCreateRequest, TenantCreateResponse,
32 : TenantCreateResponseShard, TenantDescribeResponse, TenantDescribeResponseShard,
33 : TenantLocateResponse, TenantPolicyRequest, TenantShardMigrateRequest,
34 : TenantShardMigrateResponse,
35 : };
36 : use pageserver_api::models::{
37 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode, PageserverUtilization,
38 : SafekeeperInfo, SafekeepersInfo, 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::membership::{MemberSet, SafekeeperId};
55 : use safekeeper_api::models::SafekeeperUtilization;
56 : use safekeeper_reconciler::{SafekeeperReconcilers, ScheduleRequest};
57 : use tokio::sync::TryAcquireError;
58 : use tokio::sync::mpsc::error::TrySendError;
59 : use tokio::task::JoinSet;
60 : use tokio_util::sync::CancellationToken;
61 : use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
62 : use utils::completion::Barrier;
63 : use utils::generation::Generation;
64 : use utils::id::{NodeId, TenantId, TimelineId};
65 : use utils::logging::SecretString;
66 : use utils::sync::gate::Gate;
67 : use utils::{failpoint_support, pausable_failpoint};
68 :
69 : use crate::background_node_operations::{
70 : Drain, Fill, MAX_RECONCILES_PER_OPERATION, Operation, OperationError, OperationHandler,
71 : };
72 : use crate::compute_hook::{self, ComputeHook, NotifyError};
73 : use crate::drain_utils::{self, TenantShardDrain, TenantShardIterator};
74 : use crate::heartbeater::{Heartbeater, PageserverState, SafekeeperState};
75 : use crate::id_lock_map::{
76 : IdLockMap, TracingExclusiveGuard, trace_exclusive_lock, trace_shared_lock,
77 : };
78 : use crate::leadership::Leadership;
79 : use crate::metrics;
80 : use crate::node::{AvailabilityTransition, Node};
81 : use crate::pageserver_client::PageserverClient;
82 : use crate::peer_client::GlobalObservedState;
83 : use crate::persistence::split_state::SplitState;
84 : use crate::persistence::{
85 : AbortShardSplitStatus, ControllerPersistence, DatabaseError, DatabaseResult,
86 : MetadataHealthPersistence, Persistence, SafekeeperTimelineOpKind, ShardGenerationState,
87 : TenantFilter, TenantShardPersistence, TimelinePendingOpPersistence, TimelinePersistence,
88 : };
89 : use crate::reconciler::{
90 : ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder, ReconcilerPriority,
91 : attached_location_conf,
92 : };
93 : use crate::safekeeper::Safekeeper;
94 : use crate::scheduler::{
95 : AttachedShardTag, MaySchedule, ScheduleContext, ScheduleError, ScheduleMode, Scheduler,
96 : };
97 : use crate::tenant_shard::{
98 : IntentState, MigrateAttachment, ObservedState, ObservedStateDelta, ObservedStateLocation,
99 : ReconcileNeeded, ReconcileResult, ReconcileWaitError, ReconcilerStatus, ReconcilerWaiter,
100 : ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
101 : };
102 :
103 : const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
104 :
105 : // For operations that should be quick, like attaching a new tenant
106 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
107 :
108 : // For operations that might be slow, like migrating a tenant with
109 : // some data in it.
110 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
111 :
112 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
113 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
114 : // input generation from future requests as authoritative.
115 : const INITIAL_GENERATION: Generation = Generation::new(0);
116 :
117 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
118 : /// up on unresponsive pageservers and proceed.
119 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
120 :
121 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
122 : /// This must be long enough to cover node restarts as well as normal operations: in future
123 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
124 :
125 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
126 : /// offline.
127 : ///
128 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
129 : /// handling of the re-attach response may take a long time and blocks heartbeats from
130 : /// being handled on the pageserver side.
131 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
132 :
133 : /// How often to send heartbeats to registered nodes?
134 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
135 :
136 : /// How long is too long for a reconciliation?
137 : pub const LONG_RECONCILE_THRESHOLD_DEFAULT: Duration = Duration::from_secs(120);
138 :
139 : #[derive(Clone, strum_macros::Display)]
140 : enum TenantOperations {
141 : Create,
142 : LocationConfig,
143 : ConfigSet,
144 : ConfigPatch,
145 : TimeTravelRemoteStorage,
146 : Delete,
147 : UpdatePolicy,
148 : ShardSplit,
149 : SecondaryDownload,
150 : TimelineCreate,
151 : TimelineDelete,
152 : AttachHook,
153 : TimelineArchivalConfig,
154 : TimelineDetachAncestor,
155 : TimelineGcBlockUnblock,
156 : DropDetached,
157 : DownloadHeatmapLayers,
158 : }
159 :
160 : #[derive(Clone, strum_macros::Display)]
161 : enum NodeOperations {
162 : Register,
163 : Configure,
164 : Delete,
165 : }
166 :
167 : /// The leadership status for the storage controller process.
168 : /// Allowed transitions are:
169 : /// 1. Leader -> SteppedDown
170 : /// 2. Candidate -> Leader
171 : #[derive(
172 : Eq,
173 : PartialEq,
174 : Copy,
175 : Clone,
176 : strum_macros::Display,
177 0 : strum_macros::EnumIter,
178 : measured::FixedCardinalityLabel,
179 : )]
180 : #[strum(serialize_all = "snake_case")]
181 : pub(crate) enum LeadershipStatus {
182 : /// This is the steady state where the storage controller can produce
183 : /// side effects in the cluster.
184 : Leader,
185 : /// We've been notified to step down by another candidate. No reconciliations
186 : /// take place in this state.
187 : SteppedDown,
188 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
189 : #[allow(unused)]
190 : Candidate,
191 : }
192 :
193 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
194 : pub const PRIORITY_RECONCILER_CONCURRENCY_DEFAULT: usize = 256;
195 :
196 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
197 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
198 : // than they're being pushed onto the queue.
199 : const MAX_DELAYED_RECONCILES: usize = 10000;
200 :
201 : // Top level state available to all HTTP handlers
202 : struct ServiceState {
203 : leadership_status: LeadershipStatus,
204 :
205 : tenants: BTreeMap<TenantShardId, TenantShard>,
206 :
207 : nodes: Arc<HashMap<NodeId, Node>>,
208 :
209 : safekeepers: Arc<HashMap<NodeId, Safekeeper>>,
210 :
211 : safekeeper_reconcilers: SafekeeperReconcilers,
212 :
213 : scheduler: Scheduler,
214 :
215 : /// Ongoing background operation on the cluster if any is running.
216 : /// Note that only one such operation may run at any given time,
217 : /// hence the type choice.
218 : ongoing_operation: Option<OperationHandler>,
219 :
220 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
221 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
222 : }
223 :
224 : /// Transform an error from a pageserver into an error to return to callers of a storage
225 : /// controller API.
226 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
227 0 : match e {
228 0 : mgmt_api::Error::SendRequest(e) => {
229 0 : // Presume errors sending requests are connectivity/availability issues
230 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
231 : }
232 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
233 0 : // Presume errors receiving body are connectivity/availability issues
234 0 : ApiError::ResourceUnavailable(
235 0 : format!("{node} error receiving error body: {str}").into(),
236 0 : )
237 : }
238 0 : mgmt_api::Error::ReceiveBody(err) if err.is_decode() => {
239 0 : // Return 500 for decoding errors.
240 0 : ApiError::InternalServerError(anyhow::Error::from(err).context("error decoding body"))
241 : }
242 0 : mgmt_api::Error::ReceiveBody(err) => {
243 0 : // Presume errors receiving body are connectivity/availability issues except for decoding errors
244 0 : let src_str = err.source().map(|e| e.to_string()).unwrap_or_default();
245 0 : ApiError::ResourceUnavailable(
246 0 : format!("{node} error receiving error body: {err} {}", src_str).into(),
247 0 : )
248 : }
249 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
250 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
251 : }
252 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
253 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
254 : }
255 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
256 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
257 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
258 : // internal server errors, showing that something is wrong with the pageserver or
259 : // storage controller's auth configuration.
260 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
261 : }
262 0 : mgmt_api::Error::ApiError(status @ StatusCode::TOO_MANY_REQUESTS, msg) => {
263 0 : // Pass through 429 errors: if pageserver is asking us to wait + retry, we in
264 0 : // turn ask our clients to wait + retry
265 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
266 : }
267 0 : mgmt_api::Error::ApiError(status, msg) => {
268 0 : // Presume general case of pageserver API errors is that we tried to do something
269 0 : // that can't be done right now.
270 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
271 : }
272 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
273 0 : mgmt_api::Error::CreateClient(e) => ApiError::InternalServerError(anyhow::anyhow!(e)),
274 : }
275 0 : }
276 :
277 : impl ServiceState {
278 0 : fn new(
279 0 : nodes: HashMap<NodeId, Node>,
280 0 : safekeepers: HashMap<NodeId, Safekeeper>,
281 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
282 0 : scheduler: Scheduler,
283 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
284 0 : initial_leadership_status: LeadershipStatus,
285 0 : reconcilers_cancel: CancellationToken,
286 0 : ) -> Self {
287 0 : metrics::update_leadership_status(initial_leadership_status);
288 0 :
289 0 : Self {
290 0 : leadership_status: initial_leadership_status,
291 0 : tenants,
292 0 : nodes: Arc::new(nodes),
293 0 : safekeepers: Arc::new(safekeepers),
294 0 : safekeeper_reconcilers: SafekeeperReconcilers::new(reconcilers_cancel),
295 0 : scheduler,
296 0 : ongoing_operation: None,
297 0 : delayed_reconcile_rx,
298 0 : }
299 0 : }
300 :
301 0 : fn parts_mut(
302 0 : &mut self,
303 0 : ) -> (
304 0 : &mut Arc<HashMap<NodeId, Node>>,
305 0 : &mut BTreeMap<TenantShardId, TenantShard>,
306 0 : &mut Scheduler,
307 0 : ) {
308 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
309 0 : }
310 :
311 : #[allow(clippy::type_complexity)]
312 0 : fn parts_mut_sk(
313 0 : &mut self,
314 0 : ) -> (
315 0 : &mut Arc<HashMap<NodeId, Node>>,
316 0 : &mut Arc<HashMap<NodeId, Safekeeper>>,
317 0 : &mut BTreeMap<TenantShardId, TenantShard>,
318 0 : &mut Scheduler,
319 0 : ) {
320 0 : (
321 0 : &mut self.nodes,
322 0 : &mut self.safekeepers,
323 0 : &mut self.tenants,
324 0 : &mut self.scheduler,
325 0 : )
326 0 : }
327 :
328 0 : fn get_leadership_status(&self) -> LeadershipStatus {
329 0 : self.leadership_status
330 0 : }
331 :
332 0 : fn step_down(&mut self) {
333 0 : self.leadership_status = LeadershipStatus::SteppedDown;
334 0 : metrics::update_leadership_status(self.leadership_status);
335 0 : }
336 :
337 0 : fn become_leader(&mut self) {
338 0 : self.leadership_status = LeadershipStatus::Leader;
339 0 : metrics::update_leadership_status(self.leadership_status);
340 0 : }
341 : }
342 :
343 : #[derive(Clone)]
344 : pub struct Config {
345 : // All pageservers managed by one instance of this service must have
346 : // the same public key. This JWT token will be used to authenticate
347 : // this service to the pageservers it manages.
348 : pub pageserver_jwt_token: Option<String>,
349 :
350 : // All safekeepers managed by one instance of this service must have
351 : // the same public key. This JWT token will be used to authenticate
352 : // this service to the safekeepers it manages.
353 : pub safekeeper_jwt_token: Option<String>,
354 :
355 : // This JWT token will be used to authenticate this service to the control plane.
356 : pub control_plane_jwt_token: Option<String>,
357 :
358 : // This JWT token will be used to authenticate with other storage controller instances
359 : pub peer_jwt_token: Option<String>,
360 :
361 : /// Where the compute hook should send notifications of pageserver attachment locations
362 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
363 : /// assume it is running in a test environment and try to update neon_local.
364 : pub compute_hook_url: Option<String>,
365 :
366 : /// Prefix for storage API endpoints of the control plane. We use this prefix to compute
367 : /// URLs that we use to send pageserver and safekeeper attachment locations.
368 : /// If this is None, the compute hook will assume it is running in a test environment
369 : /// and try to invoke neon_local instead.
370 : ///
371 : /// For now, there is also `compute_hook_url` which allows configuration of the pageserver
372 : /// specific endpoint, but it is in the process of being phased out.
373 : pub control_plane_url: Option<String>,
374 :
375 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
376 : /// considered active. Once the grace period elapses, the next heartbeat failure will
377 : /// mark the pagseserver offline.
378 : pub max_offline_interval: Duration,
379 :
380 : /// Extended grace period within which pageserver may not respond to heartbeats.
381 : /// This extended grace period kicks in after the node has been drained for restart
382 : /// and/or upon handling the re-attach request from a node.
383 : pub max_warming_up_interval: Duration,
384 :
385 : /// How many normal-priority Reconcilers may be spawned concurrently
386 : pub reconciler_concurrency: usize,
387 :
388 : /// How many high-priority Reconcilers may be spawned concurrently
389 : pub priority_reconciler_concurrency: usize,
390 :
391 : /// How many API requests per second to allow per tenant, across all
392 : /// tenant-scoped API endpoints. Further API requests queue until ready.
393 : pub tenant_rate_limit: NonZeroU32,
394 :
395 : /// The size at which an unsharded tenant should be split (into 8 shards). This uses the logical
396 : /// size of the largest timeline in the shard (i.e. max_logical_size).
397 : ///
398 : /// None or 0 disables auto-splitting.
399 : ///
400 : /// TODO: consider using total logical size of all timelines instead.
401 : pub split_threshold: Option<u64>,
402 :
403 : // TODO: make this cfg(feature = "testing")
404 : pub neon_local_repo_dir: Option<PathBuf>,
405 :
406 : // Maximum acceptable download lag for the secondary location
407 : // while draining a node. If the secondary location is lagging
408 : // by more than the configured amount, then the secondary is not
409 : // upgraded to primary.
410 : pub max_secondary_lag_bytes: Option<u64>,
411 :
412 : pub heartbeat_interval: Duration,
413 :
414 : pub address_for_peers: Option<Uri>,
415 :
416 : pub start_as_candidate: bool,
417 :
418 : pub http_service_port: i32,
419 :
420 : pub long_reconcile_threshold: Duration,
421 :
422 : pub use_https_pageserver_api: bool,
423 :
424 : pub use_https_safekeeper_api: bool,
425 :
426 : pub ssl_ca_cert: Option<Certificate>,
427 :
428 : pub timelines_onto_safekeepers: bool,
429 : }
430 :
431 : impl From<DatabaseError> for ApiError {
432 0 : fn from(err: DatabaseError) -> ApiError {
433 0 : match err {
434 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
435 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
436 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
437 0 : ApiError::ShuttingDown
438 : }
439 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
440 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
441 : }
442 : }
443 0 : }
444 : }
445 :
446 : enum InitialShardScheduleOutcome {
447 : Scheduled(TenantCreateResponseShard),
448 : NotScheduled,
449 : ShardScheduleError(ScheduleError),
450 : }
451 :
452 : pub struct Service {
453 : inner: Arc<std::sync::RwLock<ServiceState>>,
454 : config: Config,
455 : persistence: Arc<Persistence>,
456 : compute_hook: Arc<ComputeHook>,
457 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
458 :
459 : heartbeater_ps: Heartbeater<Node, PageserverState>,
460 : heartbeater_sk: Heartbeater<Safekeeper, SafekeeperState>,
461 :
462 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
463 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
464 :
465 : // Locking on a tenant granularity (covers all shards in the tenant):
466 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
467 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
468 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
469 :
470 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
471 : // that transition it to/from Active.
472 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
473 :
474 : // Limit how many Reconcilers we will spawn concurrently for normal-priority tasks such as background reconciliations
475 : // and reconciliation on startup.
476 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
477 :
478 : // Limit how many Reconcilers we will spawn concurrently for high-priority tasks such as tenant/timeline CRUD, which
479 : // a human user might be waiting for.
480 : priority_reconciler_concurrency: Arc<tokio::sync::Semaphore>,
481 :
482 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
483 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
484 : ///
485 : /// Note that this state logically lives inside ServiceState, but carrying Sender here makes the code simpler
486 : /// by avoiding needing a &mut ref to something inside the ServiceState. This could be optimized to
487 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
488 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
489 :
490 : // Process shutdown will fire this token
491 : cancel: CancellationToken,
492 :
493 : // Child token of [`Service::cancel`] used by reconcilers
494 : reconcilers_cancel: CancellationToken,
495 :
496 : // Background tasks will hold this gate
497 : gate: Gate,
498 :
499 : // Reconcilers background tasks will hold this gate
500 : reconcilers_gate: Gate,
501 :
502 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
503 : /// passes, it isn't safe to do any actions that mutate tenants.
504 : pub(crate) startup_complete: Barrier,
505 : }
506 :
507 : impl From<ReconcileWaitError> for ApiError {
508 0 : fn from(value: ReconcileWaitError) -> Self {
509 0 : match value {
510 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
511 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
512 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
513 : }
514 0 : }
515 : }
516 :
517 : impl From<OperationError> for ApiError {
518 0 : fn from(value: OperationError) -> Self {
519 0 : match value {
520 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
521 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
522 : }
523 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
524 : }
525 0 : }
526 : }
527 :
528 : #[allow(clippy::large_enum_variant)]
529 : enum TenantCreateOrUpdate {
530 : Create(TenantCreateRequest),
531 : Update(Vec<ShardUpdate>),
532 : }
533 :
534 : struct ShardSplitParams {
535 : old_shard_count: ShardCount,
536 : new_shard_count: ShardCount,
537 : new_stripe_size: Option<ShardStripeSize>,
538 : targets: Vec<ShardSplitTarget>,
539 : policy: PlacementPolicy,
540 : config: TenantConfig,
541 : shard_ident: ShardIdentity,
542 : preferred_az_id: Option<AvailabilityZone>,
543 : }
544 :
545 : // When preparing for a shard split, we may either choose to proceed with the split,
546 : // or find that the work is already done and return NoOp.
547 : enum ShardSplitAction {
548 : Split(Box<ShardSplitParams>),
549 : NoOp(TenantShardSplitResponse),
550 : }
551 :
552 : // A parent shard which will be split
553 : struct ShardSplitTarget {
554 : parent_id: TenantShardId,
555 : node: Node,
556 : child_ids: Vec<TenantShardId>,
557 : }
558 :
559 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
560 : /// might not be available. We therefore use a queue of abort operations processed in the background.
561 : struct TenantShardSplitAbort {
562 : tenant_id: TenantId,
563 : /// The target values from the request that failed
564 : new_shard_count: ShardCount,
565 : new_stripe_size: Option<ShardStripeSize>,
566 : /// Until this abort op is complete, no other operations may be done on the tenant
567 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
568 : }
569 :
570 : #[derive(thiserror::Error, Debug)]
571 : enum TenantShardSplitAbortError {
572 : #[error(transparent)]
573 : Database(#[from] DatabaseError),
574 : #[error(transparent)]
575 : Remote(#[from] mgmt_api::Error),
576 : #[error("Unavailable")]
577 : Unavailable,
578 : }
579 :
580 : struct ShardUpdate {
581 : tenant_shard_id: TenantShardId,
582 : placement_policy: PlacementPolicy,
583 : tenant_config: TenantConfig,
584 :
585 : /// If this is None, generation is not updated.
586 : generation: Option<Generation>,
587 :
588 : /// If this is None, scheduling policy is not updated.
589 : scheduling_policy: Option<ShardSchedulingPolicy>,
590 : }
591 :
592 : enum StopReconciliationsReason {
593 : ShuttingDown,
594 : SteppingDown,
595 : }
596 :
597 : impl std::fmt::Display for StopReconciliationsReason {
598 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
599 0 : let s = match self {
600 0 : Self::ShuttingDown => "Shutting down",
601 0 : Self::SteppingDown => "Stepping down",
602 : };
603 0 : write!(writer, "{}", s)
604 0 : }
605 : }
606 :
607 : pub(crate) enum ReconcileResultRequest {
608 : ReconcileResult(ReconcileResult),
609 : Stop,
610 : }
611 :
612 : #[derive(Clone)]
613 : struct MutationLocation {
614 : node: Node,
615 : generation: Generation,
616 : }
617 :
618 : #[derive(Clone)]
619 : struct ShardMutationLocations {
620 : latest: MutationLocation,
621 : other: Vec<MutationLocation>,
622 : }
623 :
624 : #[derive(Default, Clone)]
625 : struct TenantMutationLocations(BTreeMap<TenantShardId, ShardMutationLocations>);
626 :
627 : impl Service {
628 0 : pub fn get_config(&self) -> &Config {
629 0 : &self.config
630 0 : }
631 :
632 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
633 : /// view of the world, and determine which pageservers are responsive.
634 : #[instrument(skip_all)]
635 : async fn startup_reconcile(
636 : self: &Arc<Service>,
637 : current_leader: Option<ControllerPersistence>,
638 : leader_step_down_state: Option<GlobalObservedState>,
639 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
640 : Result<(), (TenantShardId, NotifyError)>,
641 : >,
642 : ) {
643 : // Startup reconciliation does I/O to other services: whether they
644 : // are responsive or not, we should aim to finish within our deadline, because:
645 : // - If we don't, a k8s readiness hook watching /ready will kill us.
646 : // - While we're waiting for startup reconciliation, we are not fully
647 : // available for end user operations like creating/deleting tenants and timelines.
648 : //
649 : // We set multiple deadlines to break up the time available between the phases of work: this is
650 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
651 : let start_at = Instant::now();
652 : let node_scan_deadline = start_at
653 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
654 : .expect("Reconcile timeout is a modest constant");
655 :
656 : let observed = if let Some(state) = leader_step_down_state {
657 : tracing::info!(
658 : "Using observed state received from leader at {}",
659 : current_leader.as_ref().unwrap().address
660 : );
661 :
662 : state
663 : } else {
664 : self.build_global_observed_state(node_scan_deadline).await
665 : };
666 :
667 : // Accumulate a list of any tenant locations that ought to be detached
668 : let mut cleanup = Vec::new();
669 :
670 : // Send initial heartbeat requests to all nodes loaded from the database
671 : let all_nodes = {
672 : let locked = self.inner.read().unwrap();
673 : locked.nodes.clone()
674 : };
675 : let (mut nodes_online, mut sks_online) =
676 : self.initial_heartbeat_round(all_nodes.keys()).await;
677 :
678 : // List of tenants for which we will attempt to notify compute of their location at startup
679 : let mut compute_notifications = Vec::new();
680 :
681 : // Populate intent and observed states for all tenants, based on reported state on pageservers
682 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
683 : let shard_count = {
684 : let mut locked = self.inner.write().unwrap();
685 : let (nodes, safekeepers, tenants, scheduler) = locked.parts_mut_sk();
686 :
687 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
688 : let mut new_nodes = (**nodes).clone();
689 : for (node_id, node) in new_nodes.iter_mut() {
690 : if let Some(utilization) = nodes_online.remove(node_id) {
691 : node.set_availability(NodeAvailability::Active(utilization));
692 : scheduler.node_upsert(node);
693 : }
694 : }
695 : *nodes = Arc::new(new_nodes);
696 :
697 : let mut new_sks = (**safekeepers).clone();
698 : for (node_id, node) in new_sks.iter_mut() {
699 : if let Some((utilization, last_seen_at)) = sks_online.remove(node_id) {
700 : node.set_availability(SafekeeperState::Available {
701 : utilization,
702 : last_seen_at,
703 : });
704 : }
705 : }
706 : *safekeepers = Arc::new(new_sks);
707 :
708 : for (tenant_shard_id, observed_state) in observed.0 {
709 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
710 : for node_id in observed_state.locations.keys() {
711 : cleanup.push((tenant_shard_id, *node_id));
712 : }
713 :
714 : continue;
715 : };
716 :
717 : tenant_shard.observed = observed_state;
718 : }
719 :
720 : // Populate each tenant's intent state
721 : let mut schedule_context = ScheduleContext::default();
722 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
723 : if tenant_shard_id.shard_number == ShardNumber(0) {
724 : // Reset scheduling context each time we advance to the next Tenant
725 : schedule_context = ScheduleContext::default();
726 : }
727 :
728 : tenant_shard.intent_from_observed(scheduler);
729 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
730 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
731 : // not enough pageservers are available. The tenant may well still be available
732 : // to clients.
733 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
734 : } else {
735 : // If we're both intending and observed to be attached at a particular node, we will
736 : // emit a compute notification for this. In the case where our observed state does not
737 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
738 : if let Some(attached_at) = tenant_shard.stably_attached() {
739 : compute_notifications.push(compute_hook::ShardUpdate {
740 : tenant_shard_id: *tenant_shard_id,
741 : node_id: attached_at,
742 : stripe_size: tenant_shard.shard.stripe_size,
743 : preferred_az: tenant_shard
744 : .preferred_az()
745 0 : .map(|az| Cow::Owned(az.clone())),
746 : });
747 : }
748 : }
749 : }
750 :
751 : tenants.len()
752 : };
753 :
754 : // Before making any obeservable changes to the cluster, persist self
755 : // as leader in database and memory.
756 : let leadership = Leadership::new(
757 : self.persistence.clone(),
758 : self.config.clone(),
759 : self.cancel.child_token(),
760 : );
761 :
762 : if let Err(e) = leadership.become_leader(current_leader).await {
763 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
764 : std::process::exit(1);
765 : }
766 :
767 : let safekeepers = self.inner.read().unwrap().safekeepers.clone();
768 : let sk_schedule_requests =
769 : match safekeeper_reconciler::load_schedule_requests(self, &safekeepers).await {
770 : Ok(v) => v,
771 : Err(e) => {
772 : tracing::warn!(
773 : "Failed to load safekeeper pending ops at startup: {e}." // Don't abort for now: " Aborting start-up..."
774 : );
775 : // std::process::exit(1);
776 : Vec::new()
777 : }
778 : };
779 :
780 : {
781 : let mut locked = self.inner.write().unwrap();
782 : locked.become_leader();
783 :
784 : locked
785 : .safekeeper_reconcilers
786 : .schedule_request_vec(self, sk_schedule_requests);
787 : }
788 :
789 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
790 : // generation_pageserver in the database.
791 :
792 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
793 : // will emit compute hook notifications when they reconcile.
794 : //
795 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
796 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
797 : // calls will be correctly ordered wrt these.
798 : //
799 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
800 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
801 : // unit and start doing I/O.
802 : tracing::info!(
803 : "Sending {} compute notifications",
804 : compute_notifications.len()
805 : );
806 : self.compute_hook.notify_background(
807 : compute_notifications,
808 : bg_compute_notify_result_tx.clone(),
809 : &self.cancel,
810 : );
811 :
812 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
813 : // which require it: under normal circumstances this should only include tenants that were in some
814 : // transient state before we restarted, or any tenants whose compute hooks failed above.
815 : tracing::info!("Checking for shards in need of reconciliation...");
816 : let reconcile_tasks = self.reconcile_all();
817 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
818 : // normal operations may proceed.
819 :
820 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
821 : // background because it does not need to complete in order to proceed with other work.
822 : if !cleanup.is_empty() {
823 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
824 : tokio::task::spawn({
825 : let cleanup_self = self.clone();
826 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
827 : });
828 : }
829 :
830 : tracing::info!(
831 : "Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)"
832 : );
833 : }
834 :
835 0 : async fn initial_heartbeat_round<'a>(
836 0 : &self,
837 0 : node_ids: impl Iterator<Item = &'a NodeId>,
838 0 : ) -> (
839 0 : HashMap<NodeId, PageserverUtilization>,
840 0 : HashMap<NodeId, (SafekeeperUtilization, Instant)>,
841 0 : ) {
842 0 : assert!(!self.startup_complete.is_ready());
843 :
844 0 : let all_nodes = {
845 0 : let locked = self.inner.read().unwrap();
846 0 : locked.nodes.clone()
847 0 : };
848 0 :
849 0 : let mut nodes_to_heartbeat = HashMap::new();
850 0 : for node_id in node_ids {
851 0 : match all_nodes.get(node_id) {
852 0 : Some(node) => {
853 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
854 0 : }
855 : None => {
856 0 : tracing::warn!("Node {node_id} was removed during start-up");
857 : }
858 : }
859 : }
860 :
861 0 : let all_sks = {
862 0 : let locked = self.inner.read().unwrap();
863 0 : locked.safekeepers.clone()
864 0 : };
865 0 :
866 0 : tracing::info!("Sending initial heartbeats...");
867 : // Put a small, but reasonable timeout to get the initial heartbeats of the safekeepers to avoid a storage controller downtime
868 : const SK_TIMEOUT: Duration = Duration::from_secs(5);
869 0 : let (res_ps, res_sk) = tokio::join!(
870 0 : self.heartbeater_ps.heartbeat(Arc::new(nodes_to_heartbeat)),
871 0 : tokio::time::timeout(SK_TIMEOUT, self.heartbeater_sk.heartbeat(all_sks))
872 0 : );
873 :
874 0 : let mut online_nodes = HashMap::new();
875 0 : if let Ok(deltas) = res_ps {
876 0 : for (node_id, status) in deltas.0 {
877 0 : match status {
878 0 : PageserverState::Available { utilization, .. } => {
879 0 : online_nodes.insert(node_id, utilization);
880 0 : }
881 0 : PageserverState::Offline => {}
882 : PageserverState::WarmingUp { .. } => {
883 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
884 : }
885 : }
886 : }
887 0 : }
888 :
889 0 : let mut online_sks = HashMap::new();
890 0 : if let Ok(Ok(deltas)) = res_sk {
891 0 : for (node_id, status) in deltas.0 {
892 0 : match status {
893 : SafekeeperState::Available {
894 0 : utilization,
895 0 : last_seen_at,
896 0 : } => {
897 0 : online_sks.insert(node_id, (utilization, last_seen_at));
898 0 : }
899 0 : SafekeeperState::Offline => {}
900 : }
901 : }
902 0 : }
903 :
904 0 : (online_nodes, online_sks)
905 0 : }
906 :
907 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
908 : ///
909 : /// The result includes only nodes which responded within the deadline
910 0 : async fn scan_node_locations(
911 0 : &self,
912 0 : deadline: Instant,
913 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
914 0 : let nodes = {
915 0 : let locked = self.inner.read().unwrap();
916 0 : locked.nodes.clone()
917 0 : };
918 0 :
919 0 : let mut node_results = HashMap::new();
920 0 :
921 0 : let mut node_list_futs = FuturesUnordered::new();
922 0 :
923 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
924 0 : for node in nodes.values() {
925 0 : node_list_futs.push({
926 0 : async move {
927 0 : tracing::info!("Scanning shards on node {node}...");
928 0 : let timeout = Duration::from_secs(5);
929 0 : let response = node
930 0 : .with_client_retries(
931 0 : |client| async move { client.list_location_config().await },
932 0 : &self.config.pageserver_jwt_token,
933 0 : &self.config.ssl_ca_cert,
934 0 : 1,
935 0 : 5,
936 0 : timeout,
937 0 : &self.cancel,
938 0 : )
939 0 : .await;
940 0 : (node.get_id(), response)
941 0 : }
942 0 : });
943 0 : }
944 :
945 : loop {
946 0 : let (node_id, result) = tokio::select! {
947 0 : next = node_list_futs.next() => {
948 0 : match next {
949 0 : Some(result) => result,
950 : None =>{
951 : // We got results for all our nodes
952 0 : break;
953 : }
954 :
955 : }
956 : },
957 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
958 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
959 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
960 0 : break;
961 : }
962 : };
963 :
964 0 : let Some(list_response) = result else {
965 0 : tracing::info!("Shutdown during startup_reconcile");
966 0 : break;
967 : };
968 :
969 0 : match list_response {
970 0 : Err(e) => {
971 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
972 : }
973 0 : Ok(listing) => {
974 0 : node_results.insert(node_id, listing);
975 0 : }
976 : }
977 : }
978 :
979 0 : node_results
980 0 : }
981 :
982 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
983 0 : let node_listings = self.scan_node_locations(deadline).await;
984 0 : let mut observed = GlobalObservedState::default();
985 :
986 0 : for (node_id, location_confs) in node_listings {
987 0 : tracing::info!(
988 0 : "Received {} shard statuses from pageserver {}",
989 0 : location_confs.tenant_shards.len(),
990 : node_id
991 : );
992 :
993 0 : for (tid, location_conf) in location_confs.tenant_shards {
994 0 : let entry = observed.0.entry(tid).or_default();
995 0 : entry.locations.insert(
996 0 : node_id,
997 0 : ObservedStateLocation {
998 0 : conf: location_conf,
999 0 : },
1000 0 : );
1001 0 : }
1002 : }
1003 :
1004 0 : observed
1005 0 : }
1006 :
1007 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
1008 : ///
1009 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
1010 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
1011 : /// other task trying to attach it.
1012 : #[instrument(skip_all)]
1013 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
1014 : let nodes = self.inner.read().unwrap().nodes.clone();
1015 :
1016 : for (tenant_shard_id, node_id) in cleanup {
1017 : // A node reported a tenant_shard_id which is unknown to us: detach it.
1018 : let Some(node) = nodes.get(&node_id) else {
1019 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
1020 : // a location to clean up on a node that has since been removed.
1021 : tracing::info!(
1022 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
1023 : );
1024 : continue;
1025 : };
1026 :
1027 : if self.cancel.is_cancelled() {
1028 : break;
1029 : }
1030 :
1031 : let client = match PageserverClient::new(
1032 : node.get_id(),
1033 : node.base_url(),
1034 : self.config.pageserver_jwt_token.as_deref(),
1035 : self.config.ssl_ca_cert.clone(),
1036 : ) {
1037 : Ok(client) => client,
1038 : Err(e) => {
1039 : tracing::error!(
1040 : "Failed to create client to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1041 : );
1042 : continue;
1043 : }
1044 : };
1045 : match client
1046 : .location_config(
1047 : tenant_shard_id,
1048 : LocationConfig {
1049 : mode: LocationConfigMode::Detached,
1050 : generation: None,
1051 : secondary_conf: None,
1052 : shard_number: tenant_shard_id.shard_number.0,
1053 : shard_count: tenant_shard_id.shard_count.literal(),
1054 : shard_stripe_size: 0,
1055 : tenant_conf: models::TenantConfig::default(),
1056 : },
1057 : None,
1058 : false,
1059 : )
1060 : .await
1061 : {
1062 : Ok(()) => {
1063 : tracing::info!(
1064 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
1065 : );
1066 : }
1067 : Err(e) => {
1068 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
1069 : // break anything.
1070 : tracing::error!(
1071 : "Failed to detach unknown shard {tenant_shard_id} on pageserver {node_id}: {e}"
1072 : );
1073 : }
1074 : }
1075 : }
1076 : }
1077 :
1078 : /// Long running background task that periodically wakes up and looks for shards that need
1079 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
1080 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
1081 : /// for those retries.
1082 : #[instrument(skip_all)]
1083 : async fn background_reconcile(self: &Arc<Self>) {
1084 : self.startup_complete.clone().wait().await;
1085 :
1086 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
1087 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
1088 : while !self.reconcilers_cancel.is_cancelled() {
1089 : tokio::select! {
1090 : _ = interval.tick() => {
1091 : let reconciles_spawned = self.reconcile_all();
1092 : if reconciles_spawned == 0 {
1093 : // Run optimizer only when we didn't find any other work to do
1094 : self.optimize_all().await;
1095 : }
1096 : // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
1097 : // must be responsive when new projects begin ingesting and reach the threshold.
1098 : self.autosplit_tenants().await;
1099 : }
1100 : _ = self.reconcilers_cancel.cancelled() => return
1101 : }
1102 : }
1103 : }
1104 : /// Heartbeat all storage nodes once in a while.
1105 : #[instrument(skip_all)]
1106 : async fn spawn_heartbeat_driver(&self) {
1107 : self.startup_complete.clone().wait().await;
1108 :
1109 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
1110 : while !self.cancel.is_cancelled() {
1111 : tokio::select! {
1112 : _ = interval.tick() => { }
1113 : _ = self.cancel.cancelled() => return
1114 : };
1115 :
1116 : let nodes = {
1117 : let locked = self.inner.read().unwrap();
1118 : locked.nodes.clone()
1119 : };
1120 :
1121 : let safekeepers = {
1122 : let locked = self.inner.read().unwrap();
1123 : locked.safekeepers.clone()
1124 : };
1125 :
1126 : const SK_TIMEOUT: Duration = Duration::from_secs(3);
1127 : let (res_ps, res_sk) = tokio::join!(
1128 : self.heartbeater_ps.heartbeat(nodes),
1129 : tokio::time::timeout(SK_TIMEOUT, self.heartbeater_sk.heartbeat(safekeepers))
1130 : );
1131 :
1132 : if let Ok(deltas) = res_ps {
1133 : let mut to_handle = Vec::default();
1134 :
1135 : for (node_id, state) in deltas.0 {
1136 : let new_availability = match state {
1137 : PageserverState::Available { utilization, .. } => {
1138 : NodeAvailability::Active(utilization)
1139 : }
1140 : PageserverState::WarmingUp { started_at } => {
1141 : NodeAvailability::WarmingUp(started_at)
1142 : }
1143 : PageserverState::Offline => {
1144 : // The node might have been placed in the WarmingUp state
1145 : // while the heartbeat round was on-going. Hence, filter out
1146 : // offline transitions for WarmingUp nodes that are still within
1147 : // their grace period.
1148 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
1149 : .get_node(node_id)
1150 : .await
1151 : .as_ref()
1152 0 : .map(|n| n.get_availability())
1153 : {
1154 : let now = Instant::now();
1155 : if now - *started_at >= self.config.max_warming_up_interval {
1156 : NodeAvailability::Offline
1157 : } else {
1158 : NodeAvailability::WarmingUp(*started_at)
1159 : }
1160 : } else {
1161 : NodeAvailability::Offline
1162 : }
1163 : }
1164 : };
1165 :
1166 : let node_lock = trace_exclusive_lock(
1167 : &self.node_op_locks,
1168 : node_id,
1169 : NodeOperations::Configure,
1170 : )
1171 : .await;
1172 :
1173 : pausable_failpoint!("heartbeat-pre-node-state-configure");
1174 :
1175 : // This is the code path for geniune availability transitions (i.e node
1176 : // goes unavailable and/or comes back online).
1177 : let res = self
1178 : .node_state_configure(node_id, Some(new_availability), None, &node_lock)
1179 : .await;
1180 :
1181 : match res {
1182 : Ok(transition) => {
1183 : // Keep hold of the lock until the availability transitions
1184 : // have been handled in
1185 : // [`Service::handle_node_availability_transitions`] in order avoid
1186 : // racing with [`Service::external_node_configure`].
1187 : to_handle.push((node_id, node_lock, transition));
1188 : }
1189 : Err(ApiError::NotFound(_)) => {
1190 : // This should be rare, but legitimate since the heartbeats are done
1191 : // on a snapshot of the nodes.
1192 : tracing::info!("Node {} was not found after heartbeat round", node_id);
1193 : }
1194 : Err(ApiError::ShuttingDown) => {
1195 : // No-op: we're shutting down, no need to try and update any nodes' statuses
1196 : }
1197 : Err(err) => {
1198 : // Transition to active involves reconciling: if a node responds to a heartbeat then
1199 : // becomes unavailable again, we may get an error here.
1200 : tracing::error!(
1201 : "Failed to update node state {} after heartbeat round: {}",
1202 : node_id,
1203 : err
1204 : );
1205 : }
1206 : }
1207 : }
1208 :
1209 : // We collected all the transitions above and now we handle them.
1210 : let res = self.handle_node_availability_transitions(to_handle).await;
1211 : if let Err(errs) = res {
1212 : for (node_id, err) in errs {
1213 : match err {
1214 : ApiError::NotFound(_) => {
1215 : // This should be rare, but legitimate since the heartbeats are done
1216 : // on a snapshot of the nodes.
1217 : tracing::info!(
1218 : "Node {} was not found after heartbeat round",
1219 : node_id
1220 : );
1221 : }
1222 : err => {
1223 : tracing::error!(
1224 : "Failed to handle availability transition for {} after heartbeat round: {}",
1225 : node_id,
1226 : err
1227 : );
1228 : }
1229 : }
1230 : }
1231 : }
1232 : }
1233 : if let Ok(Ok(deltas)) = res_sk {
1234 : let mut locked = self.inner.write().unwrap();
1235 : let mut safekeepers = (*locked.safekeepers).clone();
1236 : for (id, state) in deltas.0 {
1237 : let Some(sk) = safekeepers.get_mut(&id) else {
1238 : tracing::info!(
1239 : "Couldn't update safekeeper safekeeper state for id {id} from heartbeat={state:?}"
1240 : );
1241 : continue;
1242 : };
1243 : sk.set_availability(state);
1244 : }
1245 : locked.safekeepers = Arc::new(safekeepers);
1246 : }
1247 : }
1248 : }
1249 :
1250 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1251 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1252 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1253 : /// will indicate that reconciliation is not needed.
1254 : #[instrument(skip_all, fields(
1255 : seq=%result.sequence,
1256 : tenant_id=%result.tenant_shard_id.tenant_id,
1257 : shard_id=%result.tenant_shard_id.shard_slug(),
1258 : ))]
1259 : fn process_result(&self, result: ReconcileResult) {
1260 : let mut locked = self.inner.write().unwrap();
1261 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1262 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1263 : // A reconciliation result might race with removing a tenant: drop results for
1264 : // tenants that aren't in our map.
1265 : return;
1266 : };
1267 :
1268 : // Usually generation should only be updated via this path, so the max() isn't
1269 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1270 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1271 :
1272 : // If the reconciler signals that it failed to notify compute, set this state on
1273 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1274 : tenant.pending_compute_notification = result.pending_compute_notification;
1275 :
1276 : // Let the TenantShard know it is idle.
1277 : tenant.reconcile_complete(result.sequence);
1278 :
1279 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1280 : // make to the tenant
1281 0 : let deltas = result.observed_deltas.into_iter().flat_map(|delta| {
1282 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1283 : // make to the tenant
1284 0 : let node = nodes.get(delta.node_id())?;
1285 :
1286 0 : if node.is_available() {
1287 0 : return Some(delta);
1288 0 : }
1289 0 :
1290 0 : // In case a node became unavailable concurrently with the reconcile, observed
1291 0 : // locations on it are now uncertain. By convention, set them to None in order
1292 0 : // for them to get refreshed when the node comes back online.
1293 0 : Some(ObservedStateDelta::Upsert(Box::new((
1294 0 : node.get_id(),
1295 0 : ObservedStateLocation { conf: None },
1296 0 : ))))
1297 0 : });
1298 :
1299 : match result.result {
1300 : Ok(()) => {
1301 : tenant.apply_observed_deltas(deltas);
1302 : tenant.waiter.advance(result.sequence);
1303 : }
1304 : Err(e) => {
1305 : match e {
1306 : ReconcileError::Cancel => {
1307 : tracing::info!("Reconciler was cancelled");
1308 : }
1309 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1310 : // This might be due to the reconciler getting cancelled, or it might
1311 : // be due to the `Node` being marked offline.
1312 : tracing::info!("Reconciler cancelled during pageserver API call");
1313 : }
1314 : _ => {
1315 : tracing::warn!("Reconcile error: {}", e);
1316 : }
1317 : }
1318 :
1319 : // Ordering: populate last_error before advancing error_seq,
1320 : // so that waiters will see the correct error after waiting.
1321 : tenant.set_last_error(result.sequence, e);
1322 :
1323 : // Skip deletions on reconcile failures
1324 : let upsert_deltas =
1325 0 : deltas.filter(|delta| matches!(delta, ObservedStateDelta::Upsert(_)));
1326 : tenant.apply_observed_deltas(upsert_deltas);
1327 : }
1328 : }
1329 :
1330 : // If we just finished detaching all shards for a tenant, it might be time to drop it from memory.
1331 : if tenant.policy == PlacementPolicy::Detached {
1332 : // We may only drop a tenant from memory while holding the exclusive lock on the tenant ID: this protects us
1333 : // from concurrent execution wrt a request handler that might expect the tenant to remain in memory for the
1334 : // duration of the request.
1335 : let guard = self.tenant_op_locks.try_exclusive(
1336 : tenant.tenant_shard_id.tenant_id,
1337 : TenantOperations::DropDetached,
1338 : );
1339 : if let Some(guard) = guard {
1340 : self.maybe_drop_tenant(tenant.tenant_shard_id.tenant_id, &mut locked, &guard);
1341 : }
1342 : }
1343 :
1344 : // Maybe some other work can proceed now that this job finished.
1345 : //
1346 : // Only bother with this if we have some semaphore units available in the normal-priority semaphore (these
1347 : // reconciles are scheduled at `[ReconcilerPriority::Normal]`).
1348 : if self.reconciler_concurrency.available_permits() > 0 {
1349 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1350 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1351 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1352 : shard.delayed_reconcile = false;
1353 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
1354 : }
1355 :
1356 : if self.reconciler_concurrency.available_permits() == 0 {
1357 : break;
1358 : }
1359 : }
1360 : }
1361 : }
1362 :
1363 0 : async fn process_results(
1364 0 : &self,
1365 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1366 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1367 0 : Result<(), (TenantShardId, NotifyError)>,
1368 0 : >,
1369 0 : ) {
1370 : loop {
1371 : // Wait for the next result, or for cancellation
1372 0 : tokio::select! {
1373 0 : r = result_rx.recv() => {
1374 0 : match r {
1375 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1376 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1377 : }
1378 : }
1379 0 : _ = async{
1380 0 : match bg_compute_hook_result_rx.recv().await {
1381 0 : Some(result) => {
1382 0 : if let Err((tenant_shard_id, notify_error)) = result {
1383 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1384 0 : let mut locked = self.inner.write().unwrap();
1385 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1386 0 : shard.pending_compute_notification = true;
1387 0 : }
1388 :
1389 0 : }
1390 : },
1391 : None => {
1392 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1393 0 : self.cancel.cancelled().await;
1394 : }
1395 : }
1396 0 : } => {},
1397 0 : _ = self.cancel.cancelled() => {
1398 0 : break;
1399 : }
1400 : };
1401 : }
1402 0 : }
1403 :
1404 0 : async fn process_aborts(
1405 0 : &self,
1406 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1407 0 : ) {
1408 : loop {
1409 : // Wait for the next result, or for cancellation
1410 0 : let op = tokio::select! {
1411 0 : r = abort_rx.recv() => {
1412 0 : match r {
1413 0 : Some(op) => {op},
1414 0 : None => {break;}
1415 : }
1416 : }
1417 0 : _ = self.cancel.cancelled() => {
1418 0 : break;
1419 : }
1420 : };
1421 :
1422 : // Retry until shutdown: we must keep this request object alive until it is properly
1423 : // processed, as it holds a lock guard that prevents other operations trying to do things
1424 : // to the tenant while it is in a weird part-split state.
1425 0 : while !self.cancel.is_cancelled() {
1426 0 : match self.abort_tenant_shard_split(&op).await {
1427 0 : Ok(_) => break,
1428 0 : Err(e) => {
1429 0 : tracing::warn!(
1430 0 : "Failed to abort shard split on {}, will retry: {e}",
1431 : op.tenant_id
1432 : );
1433 :
1434 : // If a node is unavailable, we hope that it has been properly marked Offline
1435 : // when we retry, so that the abort op will succeed. If the abort op is failing
1436 : // for some other reason, we will keep retrying forever, or until a human notices
1437 : // and does something about it (either fixing a pageserver or restarting the controller).
1438 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1439 0 : .await
1440 0 : .ok();
1441 : }
1442 : }
1443 : }
1444 : }
1445 0 : }
1446 :
1447 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1448 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1449 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1450 0 :
1451 0 : let leadership_cancel = CancellationToken::new();
1452 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1453 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1454 :
1455 : // Apply the migrations **after** the current leader has stepped down
1456 : // (or we've given up waiting for it), but **before** reading from the
1457 : // database. The only exception is reading the current leader before
1458 : // migrating.
1459 0 : persistence.migration_run().await?;
1460 :
1461 0 : tracing::info!("Loading nodes from database...");
1462 0 : let nodes = persistence
1463 0 : .list_nodes()
1464 0 : .await?
1465 0 : .into_iter()
1466 0 : .map(|x| Node::from_persistent(x, config.use_https_pageserver_api))
1467 0 : .collect::<anyhow::Result<Vec<Node>>>()?;
1468 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1469 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1470 0 : metrics::METRICS_REGISTRY
1471 0 : .metrics_group
1472 0 : .storage_controller_pageserver_nodes
1473 0 : .set(nodes.len() as i64);
1474 0 :
1475 0 : tracing::info!("Loading safekeepers from database...");
1476 0 : let safekeepers = persistence
1477 0 : .list_safekeepers()
1478 0 : .await?
1479 0 : .into_iter()
1480 0 : .map(|skp| {
1481 0 : Safekeeper::from_persistence(
1482 0 : skp,
1483 0 : CancellationToken::new(),
1484 0 : config.use_https_safekeeper_api,
1485 0 : )
1486 0 : })
1487 0 : .collect::<anyhow::Result<Vec<_>>>()?;
1488 0 : let safekeepers: HashMap<NodeId, Safekeeper> =
1489 0 : safekeepers.into_iter().map(|n| (n.get_id(), n)).collect();
1490 0 : tracing::info!("Loaded {} safekeepers from database.", safekeepers.len());
1491 :
1492 0 : tracing::info!("Loading shards from database...");
1493 0 : let mut tenant_shard_persistence = persistence.load_active_tenant_shards().await?;
1494 0 : tracing::info!(
1495 0 : "Loaded {} shards from database.",
1496 0 : tenant_shard_persistence.len()
1497 : );
1498 :
1499 : // If any shard splits were in progress, reset the database state to abort them
1500 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1501 0 : HashMap::new();
1502 0 : for tsp in &mut tenant_shard_persistence {
1503 0 : let shard = tsp.get_shard_identity()?;
1504 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1505 0 : let entry = tenant_shard_count_min_max
1506 0 : .entry(tenant_shard_id.tenant_id)
1507 0 : .or_insert_with(|| (shard.count, shard.count));
1508 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1509 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1510 0 : }
1511 :
1512 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1513 0 : if count_min != count_max {
1514 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1515 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1516 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1517 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1518 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1519 :
1520 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1521 : // identified this tenant has having mismatching min/max counts.
1522 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1523 :
1524 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1525 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1526 0 : // Set idle split state on those shards that we will retain.
1527 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1528 0 : if tsp_tenant_id == tenant_id
1529 0 : && tsp.get_shard_identity().unwrap().count == count_min
1530 0 : {
1531 0 : tsp.splitting = SplitState::Idle;
1532 0 : } else if tsp_tenant_id == tenant_id {
1533 : // Leave the splitting state on the child shards: this will be used next to
1534 : // drop them.
1535 0 : tracing::info!(
1536 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1537 : );
1538 0 : }
1539 0 : });
1540 0 :
1541 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1542 0 : tenant_shard_persistence.retain(|tsp| {
1543 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1544 0 : || tsp.splitting == SplitState::Idle
1545 0 : });
1546 0 : }
1547 : }
1548 :
1549 0 : let mut tenants = BTreeMap::new();
1550 0 :
1551 0 : let mut scheduler = Scheduler::new(nodes.values());
1552 :
1553 : #[cfg(feature = "testing")]
1554 : {
1555 : use pageserver_api::controller_api::AvailabilityZone;
1556 :
1557 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1558 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1559 : // after when pageservers start up and register.
1560 0 : let mut node_ids = HashSet::new();
1561 0 : for tsp in &tenant_shard_persistence {
1562 0 : if let Some(node_id) = tsp.generation_pageserver {
1563 0 : node_ids.insert(node_id);
1564 0 : }
1565 : }
1566 0 : for node_id in node_ids {
1567 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1568 0 : let node = Node::new(
1569 0 : NodeId(node_id as u64),
1570 0 : "".to_string(),
1571 0 : 123,
1572 0 : None,
1573 0 : "".to_string(),
1574 0 : 123,
1575 0 : AvailabilityZone("test_az".to_string()),
1576 0 : false,
1577 0 : )
1578 0 : .unwrap();
1579 0 :
1580 0 : scheduler.node_upsert(&node);
1581 : }
1582 : }
1583 0 : for tsp in tenant_shard_persistence {
1584 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1585 :
1586 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1587 : // it with what we can infer: the node for which a generation was most recently issued.
1588 0 : let mut intent = IntentState::new(
1589 0 : tsp.preferred_az_id
1590 0 : .as_ref()
1591 0 : .map(|az| AvailabilityZone(az.clone())),
1592 0 : );
1593 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1594 : {
1595 0 : if nodes.contains_key(&generation_pageserver) {
1596 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1597 0 : } else {
1598 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1599 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1600 : // on different pageservers.
1601 0 : tracing::warn!(
1602 0 : "Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled"
1603 : );
1604 : }
1605 0 : }
1606 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1607 :
1608 0 : tenants.insert(tenant_shard_id, new_tenant);
1609 : }
1610 :
1611 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1612 0 :
1613 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1614 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1615 0 : tokio::sync::mpsc::channel(512);
1616 0 :
1617 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1618 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1619 0 :
1620 0 : let cancel = CancellationToken::new();
1621 0 : let reconcilers_cancel = cancel.child_token();
1622 0 :
1623 0 : let heartbeater_ps = Heartbeater::new(
1624 0 : config.pageserver_jwt_token.clone(),
1625 0 : config.ssl_ca_cert.clone(),
1626 0 : config.max_offline_interval,
1627 0 : config.max_warming_up_interval,
1628 0 : cancel.clone(),
1629 0 : );
1630 0 :
1631 0 : let heartbeater_sk = Heartbeater::new(
1632 0 : config.safekeeper_jwt_token.clone(),
1633 0 : config.ssl_ca_cert.clone(),
1634 0 : config.max_offline_interval,
1635 0 : config.max_warming_up_interval,
1636 0 : cancel.clone(),
1637 0 : );
1638 :
1639 0 : let initial_leadership_status = if config.start_as_candidate {
1640 0 : LeadershipStatus::Candidate
1641 : } else {
1642 0 : LeadershipStatus::Leader
1643 : };
1644 :
1645 0 : let this = Arc::new(Self {
1646 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1647 0 : nodes,
1648 0 : safekeepers,
1649 0 : tenants,
1650 0 : scheduler,
1651 0 : delayed_reconcile_rx,
1652 0 : initial_leadership_status,
1653 0 : reconcilers_cancel.clone(),
1654 0 : ))),
1655 0 : config: config.clone(),
1656 0 : persistence,
1657 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1658 0 : result_tx,
1659 0 : heartbeater_ps,
1660 0 : heartbeater_sk,
1661 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1662 0 : config.reconciler_concurrency,
1663 0 : )),
1664 0 : priority_reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1665 0 : config.priority_reconciler_concurrency,
1666 0 : )),
1667 0 : delayed_reconcile_tx,
1668 0 : abort_tx,
1669 0 : startup_complete: startup_complete.clone(),
1670 0 : cancel,
1671 0 : reconcilers_cancel,
1672 0 : gate: Gate::default(),
1673 0 : reconcilers_gate: Gate::default(),
1674 0 : tenant_op_locks: Default::default(),
1675 0 : node_op_locks: Default::default(),
1676 0 : });
1677 0 :
1678 0 : let result_task_this = this.clone();
1679 0 : tokio::task::spawn(async move {
1680 : // Block shutdown until we're done (we must respect self.cancel)
1681 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1682 0 : result_task_this
1683 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1684 0 : .await
1685 0 : }
1686 0 : });
1687 0 :
1688 0 : tokio::task::spawn({
1689 0 : let this = this.clone();
1690 0 : async move {
1691 : // Block shutdown until we're done (we must respect self.cancel)
1692 0 : if let Ok(_gate) = this.gate.enter() {
1693 0 : this.process_aborts(abort_rx).await
1694 0 : }
1695 0 : }
1696 0 : });
1697 0 :
1698 0 : tokio::task::spawn({
1699 0 : let this = this.clone();
1700 0 : async move {
1701 0 : if let Ok(_gate) = this.gate.enter() {
1702 : loop {
1703 0 : tokio::select! {
1704 0 : _ = this.cancel.cancelled() => {
1705 0 : break;
1706 : },
1707 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1708 0 : };
1709 0 : this.tenant_op_locks.housekeeping();
1710 : }
1711 0 : }
1712 0 : }
1713 0 : });
1714 0 :
1715 0 : tokio::task::spawn({
1716 0 : let this = this.clone();
1717 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1718 0 : // is done.
1719 0 : let startup_completion = startup_completion.clone();
1720 0 : async move {
1721 : // Block shutdown until we're done (we must respect self.cancel)
1722 0 : let Ok(_gate) = this.gate.enter() else {
1723 0 : return;
1724 : };
1725 :
1726 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1727 0 : .await;
1728 :
1729 0 : drop(startup_completion);
1730 0 : }
1731 0 : });
1732 0 :
1733 0 : tokio::task::spawn({
1734 0 : let this = this.clone();
1735 0 : let startup_complete = startup_complete.clone();
1736 0 : async move {
1737 0 : startup_complete.wait().await;
1738 0 : this.background_reconcile().await;
1739 0 : }
1740 0 : });
1741 0 :
1742 0 : tokio::task::spawn({
1743 0 : let this = this.clone();
1744 0 : let startup_complete = startup_complete.clone();
1745 0 : async move {
1746 0 : startup_complete.wait().await;
1747 0 : this.spawn_heartbeat_driver().await;
1748 0 : }
1749 0 : });
1750 0 :
1751 0 : Ok(this)
1752 0 : }
1753 :
1754 0 : pub(crate) async fn attach_hook(
1755 0 : &self,
1756 0 : attach_req: AttachHookRequest,
1757 0 : ) -> anyhow::Result<AttachHookResponse> {
1758 0 : let _tenant_lock = trace_exclusive_lock(
1759 0 : &self.tenant_op_locks,
1760 0 : attach_req.tenant_shard_id.tenant_id,
1761 0 : TenantOperations::AttachHook,
1762 0 : )
1763 0 : .await;
1764 :
1765 : // This is a test hook. To enable using it on tenants that were created directly with
1766 : // the pageserver API (not via this service), we will auto-create any missing tenant
1767 : // shards with default state.
1768 0 : let insert = {
1769 0 : match self
1770 0 : .maybe_load_tenant(attach_req.tenant_shard_id.tenant_id, &_tenant_lock)
1771 0 : .await
1772 : {
1773 0 : Ok(_) => false,
1774 0 : Err(ApiError::NotFound(_)) => true,
1775 0 : Err(e) => return Err(e.into()),
1776 : }
1777 : };
1778 :
1779 0 : if insert {
1780 0 : let tsp = TenantShardPersistence {
1781 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1782 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1783 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1784 0 : shard_stripe_size: 0,
1785 0 : generation: attach_req.generation_override.or(Some(0)),
1786 0 : generation_pageserver: None,
1787 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1788 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1789 0 : splitting: SplitState::default(),
1790 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1791 0 : .unwrap(),
1792 0 : preferred_az_id: None,
1793 0 : };
1794 0 :
1795 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1796 0 : Err(e) => match e {
1797 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1798 : DatabaseErrorKind::UniqueViolation,
1799 : _,
1800 : )) => {
1801 0 : tracing::info!(
1802 0 : "Raced with another request to insert tenant {}",
1803 : attach_req.tenant_shard_id
1804 : )
1805 : }
1806 0 : _ => return Err(e.into()),
1807 : },
1808 : Ok(()) => {
1809 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1810 :
1811 0 : let mut locked = self.inner.write().unwrap();
1812 0 : locked.tenants.insert(
1813 0 : attach_req.tenant_shard_id,
1814 0 : TenantShard::new(
1815 0 : attach_req.tenant_shard_id,
1816 0 : ShardIdentity::unsharded(),
1817 0 : PlacementPolicy::Attached(0),
1818 0 : None,
1819 0 : ),
1820 0 : );
1821 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1822 : }
1823 : }
1824 0 : }
1825 :
1826 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1827 0 : let maybe_tenant_conf = {
1828 0 : let locked = self.inner.write().unwrap();
1829 0 : locked
1830 0 : .tenants
1831 0 : .get(&attach_req.tenant_shard_id)
1832 0 : .map(|t| t.config.clone())
1833 0 : };
1834 0 :
1835 0 : match maybe_tenant_conf {
1836 0 : Some(conf) => {
1837 0 : let new_generation = self
1838 0 : .persistence
1839 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1840 0 : .await?;
1841 :
1842 : // Persist the placement policy update. This is required
1843 : // when we reattaching a detached tenant.
1844 0 : self.persistence
1845 0 : .update_tenant_shard(
1846 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1847 0 : Some(PlacementPolicy::Attached(0)),
1848 0 : Some(conf),
1849 0 : None,
1850 0 : None,
1851 0 : )
1852 0 : .await?;
1853 0 : Some(new_generation)
1854 : }
1855 : None => {
1856 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1857 : }
1858 : }
1859 : } else {
1860 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1861 0 : None
1862 : };
1863 :
1864 0 : let mut locked = self.inner.write().unwrap();
1865 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1866 0 :
1867 0 : let tenant_shard = tenants
1868 0 : .get_mut(&attach_req.tenant_shard_id)
1869 0 : .expect("Checked for existence above");
1870 :
1871 0 : if let Some(new_generation) = new_generation {
1872 0 : tenant_shard.generation = Some(new_generation);
1873 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1874 0 : } else {
1875 : // This is a detach notification. We must update placement policy to avoid re-attaching
1876 : // during background scheduling/reconciliation, or during storage controller restart.
1877 0 : assert!(attach_req.node_id.is_none());
1878 0 : tenant_shard.policy = PlacementPolicy::Detached;
1879 : }
1880 :
1881 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1882 0 : tracing::info!(
1883 : tenant_id = %attach_req.tenant_shard_id,
1884 : ps_id = %attaching_pageserver,
1885 : generation = ?tenant_shard.generation,
1886 0 : "issuing",
1887 : );
1888 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1889 0 : tracing::info!(
1890 : tenant_id = %attach_req.tenant_shard_id,
1891 : %ps_id,
1892 : generation = ?tenant_shard.generation,
1893 0 : "dropping",
1894 : );
1895 : } else {
1896 0 : tracing::info!(
1897 : tenant_id = %attach_req.tenant_shard_id,
1898 0 : "no-op: tenant already has no pageserver");
1899 : }
1900 0 : tenant_shard
1901 0 : .intent
1902 0 : .set_attached(scheduler, attach_req.node_id);
1903 0 :
1904 0 : tracing::info!(
1905 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1906 0 : attach_req.tenant_shard_id,
1907 0 : tenant_shard.generation,
1908 0 : // TODO: this is an odd number of 0xf's
1909 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1910 : );
1911 :
1912 : // Trick the reconciler into not doing anything for this tenant: this helps
1913 : // tests that manually configure a tenant on the pagesrever, and then call this
1914 : // attach hook: they don't want background reconciliation to modify what they
1915 : // did to the pageserver.
1916 : #[cfg(feature = "testing")]
1917 : {
1918 0 : if let Some(node_id) = attach_req.node_id {
1919 0 : tenant_shard.observed.locations = HashMap::from([(
1920 0 : node_id,
1921 0 : ObservedStateLocation {
1922 0 : conf: Some(attached_location_conf(
1923 0 : tenant_shard.generation.unwrap(),
1924 0 : &tenant_shard.shard,
1925 0 : &tenant_shard.config,
1926 0 : &PlacementPolicy::Attached(0),
1927 0 : )),
1928 0 : },
1929 0 : )]);
1930 0 : } else {
1931 0 : tenant_shard.observed.locations.clear();
1932 0 : }
1933 : }
1934 :
1935 0 : Ok(AttachHookResponse {
1936 0 : generation: attach_req
1937 0 : .node_id
1938 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1939 0 : })
1940 0 : }
1941 :
1942 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1943 0 : let locked = self.inner.read().unwrap();
1944 0 :
1945 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1946 0 :
1947 0 : InspectResponse {
1948 0 : attachment: tenant_shard.and_then(|s| {
1949 0 : s.intent
1950 0 : .get_attached()
1951 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1952 0 : }),
1953 0 : }
1954 0 : }
1955 :
1956 : // When the availability state of a node transitions to active, we must do a full reconciliation
1957 : // of LocationConfigs on that node. This is because while a node was offline:
1958 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1959 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1960 : //
1961 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1962 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1963 : // this function.
1964 : //
1965 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1966 : // for written for a single node rather than as a batch job for all nodes.
1967 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1968 : async fn node_activate_reconcile(
1969 : &self,
1970 : mut node: Node,
1971 : _lock: &TracingExclusiveGuard<NodeOperations>,
1972 : ) -> Result<(), ApiError> {
1973 : // This Node is a mutable local copy: we will set it active so that we can use its
1974 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1975 : // later.
1976 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1977 :
1978 : let configs = match node
1979 : .with_client_retries(
1980 0 : |client| async move { client.list_location_config().await },
1981 : &self.config.pageserver_jwt_token,
1982 : &self.config.ssl_ca_cert,
1983 : 1,
1984 : 5,
1985 : SHORT_RECONCILE_TIMEOUT,
1986 : &self.cancel,
1987 : )
1988 : .await
1989 : {
1990 : None => {
1991 : // We're shutting down (the Node's cancellation token can't have fired, because
1992 : // we're the only scope that has a reference to it, and we didn't fire it).
1993 : return Err(ApiError::ShuttingDown);
1994 : }
1995 : Some(Err(e)) => {
1996 : // This node didn't succeed listing its locations: it may not proceed to active state
1997 : // as it is apparently unavailable.
1998 : return Err(ApiError::PreconditionFailed(
1999 : format!("Failed to query node location configs, cannot activate ({e})").into(),
2000 : ));
2001 : }
2002 : Some(Ok(configs)) => configs,
2003 : };
2004 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
2005 :
2006 : let mut cleanup = Vec::new();
2007 : let mut mismatched_locations = 0;
2008 : {
2009 : let mut locked = self.inner.write().unwrap();
2010 :
2011 : for (tenant_shard_id, reported) in configs.tenant_shards {
2012 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
2013 : cleanup.push(tenant_shard_id);
2014 : continue;
2015 : };
2016 :
2017 : let on_record = &mut tenant_shard
2018 : .observed
2019 : .locations
2020 : .entry(node.get_id())
2021 0 : .or_insert_with(|| ObservedStateLocation { conf: None })
2022 : .conf;
2023 :
2024 : // If the location reported by the node does not match our observed state,
2025 : // then we mark it as uncertain and let the background reconciliation loop
2026 : // deal with it.
2027 : //
2028 : // Note that this also covers net new locations reported by the node.
2029 : if *on_record != reported {
2030 : mismatched_locations += 1;
2031 : *on_record = None;
2032 : }
2033 : }
2034 : }
2035 :
2036 : if mismatched_locations > 0 {
2037 : tracing::info!(
2038 : "Set observed state to None for {mismatched_locations} mismatched locations"
2039 : );
2040 : }
2041 :
2042 : for tenant_shard_id in cleanup {
2043 : tracing::info!("Detaching {tenant_shard_id}");
2044 : match node
2045 : .with_client_retries(
2046 0 : |client| async move {
2047 0 : let config = LocationConfig {
2048 0 : mode: LocationConfigMode::Detached,
2049 0 : generation: None,
2050 0 : secondary_conf: None,
2051 0 : shard_number: tenant_shard_id.shard_number.0,
2052 0 : shard_count: tenant_shard_id.shard_count.literal(),
2053 0 : shard_stripe_size: 0,
2054 0 : tenant_conf: models::TenantConfig::default(),
2055 0 : };
2056 0 : client
2057 0 : .location_config(tenant_shard_id, config, None, false)
2058 0 : .await
2059 0 : },
2060 : &self.config.pageserver_jwt_token,
2061 : &self.config.ssl_ca_cert,
2062 : 1,
2063 : 5,
2064 : SHORT_RECONCILE_TIMEOUT,
2065 : &self.cancel,
2066 : )
2067 : .await
2068 : {
2069 : None => {
2070 : // We're shutting down (the Node's cancellation token can't have fired, because
2071 : // we're the only scope that has a reference to it, and we didn't fire it).
2072 : return Err(ApiError::ShuttingDown);
2073 : }
2074 : Some(Err(e)) => {
2075 : // Do not let the node proceed to Active state if it is not responsive to requests
2076 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
2077 : // detach completing: we should not let this node back into the set of nodes considered
2078 : // okay for scheduling.
2079 : return Err(ApiError::Conflict(format!(
2080 : "Node {node} failed to detach {tenant_shard_id}: {e}"
2081 : )));
2082 : }
2083 : Some(Ok(_)) => {}
2084 : };
2085 : }
2086 :
2087 : Ok(())
2088 : }
2089 :
2090 0 : pub(crate) async fn re_attach(
2091 0 : &self,
2092 0 : reattach_req: ReAttachRequest,
2093 0 : ) -> Result<ReAttachResponse, ApiError> {
2094 0 : if let Some(register_req) = reattach_req.register {
2095 0 : self.node_register(register_req).await?;
2096 0 : }
2097 :
2098 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
2099 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
2100 :
2101 0 : tracing::info!(
2102 : node_id=%reattach_req.node_id,
2103 0 : "Incremented {} tenant shards' generations",
2104 0 : incremented_generations.len()
2105 : );
2106 :
2107 : // Apply the updated generation to our in-memory state, and
2108 : // gather discover secondary locations.
2109 0 : let mut locked = self.inner.write().unwrap();
2110 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2111 0 :
2112 0 : let mut response = ReAttachResponse {
2113 0 : tenants: Vec::new(),
2114 0 : };
2115 :
2116 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
2117 : // to call location_conf API with an old generation. Wait for cancellation to complete
2118 : // before responding to this request. Requires well implemented CancellationToken logic
2119 : // all the way to where we call location_conf. Even then, there can still be a location_conf
2120 : // request in flight over the network: TODO handle that by making location_conf API refuse
2121 : // to go backward in generations.
2122 :
2123 : // Scan through all shards, applying updates for ones where we updated generation
2124 : // and identifying shards that intend to have a secondary location on this node.
2125 0 : for (tenant_shard_id, shard) in tenants {
2126 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
2127 0 : let new_gen = *new_gen;
2128 0 : response.tenants.push(ReAttachResponseTenant {
2129 0 : id: *tenant_shard_id,
2130 0 : r#gen: Some(new_gen.into().unwrap()),
2131 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
2132 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
2133 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
2134 0 : // the stale/multi states at this point.
2135 0 : mode: LocationConfigMode::AttachedSingle,
2136 0 : });
2137 0 :
2138 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
2139 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
2140 : // Why can we update `observed` even though we're not sure our response will be received
2141 : // by the pageserver? Because the pageserver will not proceed with startup until
2142 : // it has processed response: if it loses it, we'll see another request and increment
2143 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
2144 0 : if let Some(conf) = observed.conf.as_mut() {
2145 0 : conf.generation = new_gen.into();
2146 0 : }
2147 0 : } else {
2148 0 : // This node has no observed state for the shard: perhaps it was offline
2149 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
2150 0 : // will be prompted to learn the location's state before it makes changes.
2151 0 : shard
2152 0 : .observed
2153 0 : .locations
2154 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
2155 0 : }
2156 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
2157 0 : // Ordering: pageserver will not accept /location_config requests until it has
2158 0 : // finished processing the response from re-attach. So we can update our in-memory state
2159 0 : // now, and be confident that we are not stamping on the result of some later location config.
2160 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
2161 0 : // so we might update observed state here, and then get over-written by some racing
2162 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
2163 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
2164 0 :
2165 0 : response.tenants.push(ReAttachResponseTenant {
2166 0 : id: *tenant_shard_id,
2167 0 : r#gen: None,
2168 0 : mode: LocationConfigMode::Secondary,
2169 0 : });
2170 0 :
2171 0 : // We must not update observed, because we have no guarantee that our
2172 0 : // response will be received by the pageserver. This could leave it
2173 0 : // falsely dirty, but the resulting reconcile should be idempotent.
2174 0 : }
2175 : }
2176 :
2177 : // We consider a node Active once we have composed a re-attach response, but we
2178 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
2179 : // implicitly synchronizes the LocationConfigs on the node.
2180 : //
2181 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
2182 : // but those requests will not be accepted by the node until it has finished processing
2183 : // the re-attach response.
2184 : //
2185 : // Additionally, reset the nodes scheduling policy to match the conditional update done
2186 : // in [`Persistence::re_attach`].
2187 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
2188 0 : let reset_scheduling = matches!(
2189 0 : node.get_scheduling(),
2190 : NodeSchedulingPolicy::PauseForRestart
2191 : | NodeSchedulingPolicy::Draining
2192 : | NodeSchedulingPolicy::Filling
2193 : );
2194 :
2195 0 : let mut new_nodes = (**nodes).clone();
2196 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
2197 0 : if reset_scheduling {
2198 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
2199 0 : }
2200 :
2201 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
2202 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
2203 0 :
2204 0 : scheduler.node_upsert(node);
2205 0 : let new_nodes = Arc::new(new_nodes);
2206 0 : *nodes = new_nodes;
2207 : } else {
2208 0 : tracing::error!(
2209 0 : "Reattaching node {} was removed while processing the request",
2210 : reattach_req.node_id
2211 : );
2212 : }
2213 0 : }
2214 :
2215 0 : Ok(response)
2216 0 : }
2217 :
2218 0 : pub(crate) async fn validate(
2219 0 : &self,
2220 0 : validate_req: ValidateRequest,
2221 0 : ) -> Result<ValidateResponse, DatabaseError> {
2222 : // Fast in-memory check: we may reject validation on anything that doesn't match our
2223 : // in-memory generation for a shard
2224 0 : let in_memory_result = {
2225 0 : let mut in_memory_result = Vec::new();
2226 0 : let locked = self.inner.read().unwrap();
2227 0 : for req_tenant in validate_req.tenants {
2228 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
2229 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.r#gen));
2230 0 : tracing::info!(
2231 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
2232 : req_tenant.id,
2233 : req_tenant.r#gen,
2234 : tenant_shard.generation
2235 : );
2236 :
2237 0 : in_memory_result.push((
2238 0 : req_tenant.id,
2239 0 : Generation::new(req_tenant.r#gen),
2240 0 : valid,
2241 0 : ));
2242 : } else {
2243 : // This is legal: for example during a shard split the pageserver may still
2244 : // have deletions in its queue from the old pre-split shard, or after deletion
2245 : // of a tenant that was busy with compaction/gc while being deleted.
2246 0 : tracing::info!(
2247 0 : "Refusing deletion validation for missing shard {}",
2248 : req_tenant.id
2249 : );
2250 : }
2251 : }
2252 :
2253 0 : in_memory_result
2254 : };
2255 :
2256 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
2257 : // in case of controller split-brain, where some other controller process might have incremented the generation.
2258 0 : let db_generations = self
2259 0 : .persistence
2260 0 : .shard_generations(
2261 0 : in_memory_result
2262 0 : .iter()
2263 0 : .filter_map(|i| if i.2 { Some(&i.0) } else { None }),
2264 0 : )
2265 0 : .await?;
2266 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
2267 0 :
2268 0 : let mut response = ValidateResponse {
2269 0 : tenants: Vec::new(),
2270 0 : };
2271 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
2272 0 : let valid = if valid {
2273 0 : let db_generation = db_generations.get(&tenant_shard_id);
2274 0 : db_generation == Some(&Some(validate_generation))
2275 : } else {
2276 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
2277 : // this prevents a pageserver from cleaning up an object in S3.
2278 0 : false
2279 : };
2280 :
2281 0 : response.tenants.push(ValidateResponseTenant {
2282 0 : id: tenant_shard_id,
2283 0 : valid,
2284 0 : })
2285 : }
2286 :
2287 0 : Ok(response)
2288 0 : }
2289 :
2290 0 : pub(crate) async fn tenant_create(
2291 0 : &self,
2292 0 : create_req: TenantCreateRequest,
2293 0 : ) -> Result<TenantCreateResponse, ApiError> {
2294 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
2295 :
2296 : // Exclude any concurrent attempts to create/access the same tenant ID
2297 0 : let _tenant_lock = trace_exclusive_lock(
2298 0 : &self.tenant_op_locks,
2299 0 : create_req.new_tenant_id.tenant_id,
2300 0 : TenantOperations::Create,
2301 0 : )
2302 0 : .await;
2303 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
2304 :
2305 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
2306 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
2307 : // accept compute notifications while it is in the process of creating. Reconciliation will
2308 : // be retried in the background.
2309 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
2310 0 : }
2311 0 : Ok(response)
2312 0 : }
2313 :
2314 0 : pub(crate) async fn do_tenant_create(
2315 0 : &self,
2316 0 : create_req: TenantCreateRequest,
2317 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
2318 0 : let placement_policy = create_req
2319 0 : .placement_policy
2320 0 : .clone()
2321 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
2322 0 : .unwrap_or(PlacementPolicy::Attached(0));
2323 :
2324 : // This service expects to handle sharding itself: it is an error to try and directly create
2325 : // a particular shard here.
2326 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
2327 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2328 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
2329 0 : )));
2330 : } else {
2331 0 : create_req.new_tenant_id.tenant_id
2332 0 : };
2333 0 :
2334 0 : tracing::info!(
2335 0 : "Creating tenant {}, shard_count={:?}",
2336 : create_req.new_tenant_id,
2337 : create_req.shard_parameters.count,
2338 : );
2339 :
2340 0 : let create_ids = (0..create_req.shard_parameters.count.count())
2341 0 : .map(|i| TenantShardId {
2342 0 : tenant_id,
2343 0 : shard_number: ShardNumber(i),
2344 0 : shard_count: create_req.shard_parameters.count,
2345 0 : })
2346 0 : .collect::<Vec<_>>();
2347 :
2348 : // If the caller specifies a None generation, it means "start from default". This is different
2349 : // to [`Self::tenant_location_config`], where a None generation is used to represent
2350 : // an incompletely-onboarded tenant.
2351 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2352 0 : tracing::info!(
2353 0 : "tenant_create: secondary mode, generation is_some={}",
2354 0 : create_req.generation.is_some()
2355 : );
2356 0 : create_req.generation.map(Generation::new)
2357 : } else {
2358 0 : tracing::info!(
2359 0 : "tenant_create: not secondary mode, generation is_some={}",
2360 0 : create_req.generation.is_some()
2361 : );
2362 0 : Some(
2363 0 : create_req
2364 0 : .generation
2365 0 : .map(Generation::new)
2366 0 : .unwrap_or(INITIAL_GENERATION),
2367 0 : )
2368 : };
2369 :
2370 0 : let preferred_az_id = {
2371 0 : let locked = self.inner.read().unwrap();
2372 : // Idempotency: take the existing value if the tenant already exists
2373 0 : if let Some(shard) = locked.tenants.get(create_ids.first().unwrap()) {
2374 0 : shard.preferred_az().cloned()
2375 : } else {
2376 0 : locked.scheduler.get_az_for_new_tenant()
2377 : }
2378 : };
2379 :
2380 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2381 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2382 : // during the creation, rather than risking leaving orphan objects in S3.
2383 0 : let persist_tenant_shards = create_ids
2384 0 : .iter()
2385 0 : .map(|tenant_shard_id| TenantShardPersistence {
2386 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2387 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2388 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2389 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2390 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2391 0 : // The pageserver is not known until scheduling happens: we will set this column when
2392 0 : // incrementing the generation the first time we attach to a pageserver.
2393 0 : generation_pageserver: None,
2394 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2395 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2396 0 : splitting: SplitState::default(),
2397 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2398 0 : .unwrap(),
2399 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.to_string()),
2400 0 : })
2401 0 : .collect();
2402 0 :
2403 0 : match self
2404 0 : .persistence
2405 0 : .insert_tenant_shards(persist_tenant_shards)
2406 0 : .await
2407 : {
2408 0 : Ok(_) => {}
2409 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2410 : DatabaseErrorKind::UniqueViolation,
2411 : _,
2412 : ))) => {
2413 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2414 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2415 : // creation's shard count.
2416 0 : tracing::info!(
2417 0 : "Tenant shards already present in database, proceeding with idempotent creation..."
2418 : );
2419 : }
2420 : // Any other database error is unexpected and a bug.
2421 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2422 : };
2423 :
2424 0 : let mut schedule_context = ScheduleContext::default();
2425 0 : let mut schedule_error = None;
2426 0 : let mut response_shards = Vec::new();
2427 0 : for tenant_shard_id in create_ids {
2428 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2429 :
2430 0 : let outcome = self
2431 0 : .do_initial_shard_scheduling(
2432 0 : tenant_shard_id,
2433 0 : initial_generation,
2434 0 : &create_req.shard_parameters,
2435 0 : create_req.config.clone(),
2436 0 : placement_policy.clone(),
2437 0 : preferred_az_id.as_ref(),
2438 0 : &mut schedule_context,
2439 0 : )
2440 0 : .await;
2441 :
2442 0 : match outcome {
2443 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2444 0 : InitialShardScheduleOutcome::NotScheduled => {}
2445 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2446 0 : schedule_error = Some(err);
2447 0 : }
2448 : }
2449 : }
2450 :
2451 : // If we failed to schedule shards, then they are still created in the controller,
2452 : // but we return an error to the requester to avoid a silent failure when someone
2453 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2454 : // are present in the system. We do this here rather than in the above loop, to
2455 : // avoid situations where we only create a subset of shards in the tenant.
2456 0 : if let Some(e) = schedule_error {
2457 0 : return Err(ApiError::Conflict(format!(
2458 0 : "Failed to schedule shard(s): {e}"
2459 0 : )));
2460 0 : }
2461 0 :
2462 0 : let waiters = {
2463 0 : let mut locked = self.inner.write().unwrap();
2464 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2465 0 : let config = ReconcilerConfigBuilder::new(ReconcilerPriority::High)
2466 0 : .tenant_creation_hint(true)
2467 0 : .build();
2468 0 : tenants
2469 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2470 0 : .filter_map(|(_shard_id, shard)| {
2471 0 : self.maybe_configured_reconcile_shard(shard, nodes, config)
2472 0 : })
2473 0 : .collect::<Vec<_>>()
2474 0 : };
2475 0 :
2476 0 : Ok((
2477 0 : TenantCreateResponse {
2478 0 : shards: response_shards,
2479 0 : },
2480 0 : waiters,
2481 0 : ))
2482 0 : }
2483 :
2484 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2485 : /// case of a new tenant and a pre-existing one.
2486 : #[allow(clippy::too_many_arguments)]
2487 0 : async fn do_initial_shard_scheduling(
2488 0 : &self,
2489 0 : tenant_shard_id: TenantShardId,
2490 0 : initial_generation: Option<Generation>,
2491 0 : shard_params: &ShardParameters,
2492 0 : config: TenantConfig,
2493 0 : placement_policy: PlacementPolicy,
2494 0 : preferred_az_id: Option<&AvailabilityZone>,
2495 0 : schedule_context: &mut ScheduleContext,
2496 0 : ) -> InitialShardScheduleOutcome {
2497 0 : let mut locked = self.inner.write().unwrap();
2498 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2499 :
2500 : use std::collections::btree_map::Entry;
2501 0 : match tenants.entry(tenant_shard_id) {
2502 0 : Entry::Occupied(mut entry) => {
2503 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2504 :
2505 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2506 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2507 0 : }
2508 :
2509 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2510 0 : let generation = entry
2511 0 : .get()
2512 0 : .generation
2513 0 : .expect("Generation is set when in attached mode");
2514 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2515 0 : shard_id: tenant_shard_id,
2516 0 : node_id: *node_id,
2517 0 : generation: generation.into().unwrap(),
2518 0 : })
2519 : } else {
2520 0 : InitialShardScheduleOutcome::NotScheduled
2521 : }
2522 : }
2523 0 : Entry::Vacant(entry) => {
2524 0 : let state = entry.insert(TenantShard::new(
2525 0 : tenant_shard_id,
2526 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2527 0 : placement_policy,
2528 0 : preferred_az_id.cloned(),
2529 0 : ));
2530 0 :
2531 0 : state.generation = initial_generation;
2532 0 : state.config = config;
2533 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2534 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2535 0 : }
2536 :
2537 : // Only include shards in result if we are attaching: the purpose
2538 : // of the response is to tell the caller where the shards are attached.
2539 0 : if let Some(node_id) = state.intent.get_attached() {
2540 0 : let generation = state
2541 0 : .generation
2542 0 : .expect("Generation is set when in attached mode");
2543 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2544 0 : shard_id: tenant_shard_id,
2545 0 : node_id: *node_id,
2546 0 : generation: generation.into().unwrap(),
2547 0 : })
2548 : } else {
2549 0 : InitialShardScheduleOutcome::NotScheduled
2550 : }
2551 : }
2552 : }
2553 0 : }
2554 :
2555 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2556 : /// wait for reconciliation to complete before responding.
2557 0 : async fn await_waiters(
2558 0 : &self,
2559 0 : waiters: Vec<ReconcilerWaiter>,
2560 0 : timeout: Duration,
2561 0 : ) -> Result<(), ReconcileWaitError> {
2562 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2563 0 : for waiter in waiters {
2564 0 : let timeout = deadline.duration_since(Instant::now());
2565 0 : waiter.wait_timeout(timeout).await?;
2566 : }
2567 :
2568 0 : Ok(())
2569 0 : }
2570 :
2571 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2572 : /// in progress
2573 0 : async fn await_waiters_remainder(
2574 0 : &self,
2575 0 : waiters: Vec<ReconcilerWaiter>,
2576 0 : timeout: Duration,
2577 0 : ) -> Vec<ReconcilerWaiter> {
2578 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2579 0 : for waiter in waiters.iter() {
2580 0 : let timeout = deadline.duration_since(Instant::now());
2581 0 : let _ = waiter.wait_timeout(timeout).await;
2582 : }
2583 :
2584 0 : waiters
2585 0 : .into_iter()
2586 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2587 0 : .collect::<Vec<_>>()
2588 0 : }
2589 :
2590 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2591 : /// and transform it into either a tenant creation of a series of shard updates.
2592 : ///
2593 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2594 : /// still be returned.
2595 0 : fn tenant_location_config_prepare(
2596 0 : &self,
2597 0 : tenant_id: TenantId,
2598 0 : req: TenantLocationConfigRequest,
2599 0 : ) -> TenantCreateOrUpdate {
2600 0 : let mut updates = Vec::new();
2601 0 : let mut locked = self.inner.write().unwrap();
2602 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2603 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2604 :
2605 : // Use location config mode as an indicator of policy.
2606 0 : let placement_policy = match req.config.mode {
2607 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2608 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2609 : LocationConfigMode::AttachedMulti
2610 : | LocationConfigMode::AttachedSingle
2611 : | LocationConfigMode::AttachedStale => {
2612 0 : if nodes.len() > 1 {
2613 0 : PlacementPolicy::Attached(1)
2614 : } else {
2615 : // Convenience for dev/test: if we just have one pageserver, import
2616 : // tenants into non-HA mode so that scheduling will succeed.
2617 0 : PlacementPolicy::Attached(0)
2618 : }
2619 : }
2620 : };
2621 :
2622 : // Ordinarily we do not update scheduling policy, but when making major changes
2623 : // like detaching or demoting to secondary-only, we need to force the scheduling
2624 : // mode to Active, or the caller's expected outcome (detach it) will not happen.
2625 0 : let scheduling_policy = match req.config.mode {
2626 : LocationConfigMode::Detached | LocationConfigMode::Secondary => {
2627 : // Special case: when making major changes like detaching or demoting to secondary-only,
2628 : // we need to force the scheduling mode to Active, or nothing will happen.
2629 0 : Some(ShardSchedulingPolicy::Active)
2630 : }
2631 : LocationConfigMode::AttachedMulti
2632 : | LocationConfigMode::AttachedSingle
2633 : | LocationConfigMode::AttachedStale => {
2634 : // While attached, continue to respect whatever the existing scheduling mode is.
2635 0 : None
2636 : }
2637 : };
2638 :
2639 0 : let mut create = true;
2640 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2641 : // Saw an existing shard: this is not a creation
2642 0 : create = false;
2643 :
2644 : // Shards may have initially been created by a Secondary request, where we
2645 : // would have left generation as None.
2646 : //
2647 : // We only update generation the first time we see an attached-mode request,
2648 : // and if there is no existing generation set. The caller is responsible for
2649 : // ensuring that no non-storage-controller pageserver ever uses a higher
2650 : // generation than they passed in here.
2651 : use LocationConfigMode::*;
2652 0 : let set_generation = match req.config.mode {
2653 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2654 0 : req.config.generation.map(Generation::new)
2655 : }
2656 0 : _ => None,
2657 : };
2658 :
2659 0 : updates.push(ShardUpdate {
2660 0 : tenant_shard_id: *shard_id,
2661 0 : placement_policy: placement_policy.clone(),
2662 0 : tenant_config: req.config.tenant_conf.clone(),
2663 0 : generation: set_generation,
2664 0 : scheduling_policy,
2665 0 : });
2666 : }
2667 :
2668 0 : if create {
2669 : use LocationConfigMode::*;
2670 0 : let generation = match req.config.mode {
2671 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2672 : // If a caller provided a generation in a non-attached request, ignore it
2673 : // and leave our generation as None: this enables a subsequent update to set
2674 : // the generation when setting an attached mode for the first time.
2675 0 : _ => None,
2676 : };
2677 :
2678 0 : TenantCreateOrUpdate::Create(
2679 0 : // Synthesize a creation request
2680 0 : TenantCreateRequest {
2681 0 : new_tenant_id: tenant_shard_id,
2682 0 : generation,
2683 0 : shard_parameters: ShardParameters {
2684 0 : count: tenant_shard_id.shard_count,
2685 0 : // We only import un-sharded or single-sharded tenants, so stripe
2686 0 : // size can be made up arbitrarily here.
2687 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2688 0 : },
2689 0 : placement_policy: Some(placement_policy),
2690 0 : config: req.config.tenant_conf,
2691 0 : },
2692 0 : )
2693 : } else {
2694 0 : assert!(!updates.is_empty());
2695 0 : TenantCreateOrUpdate::Update(updates)
2696 : }
2697 0 : }
2698 :
2699 : /// For APIs that might act on tenants with [`PlacementPolicy::Detached`], first check if
2700 : /// the tenant is present in memory. If not, load it from the database. If it is found
2701 : /// in neither location, return a NotFound error.
2702 : ///
2703 : /// Caller must demonstrate they hold a lock guard, as otherwise two callers might try and load
2704 : /// it at the same time, or we might race with [`Self::maybe_drop_tenant`]
2705 0 : async fn maybe_load_tenant(
2706 0 : &self,
2707 0 : tenant_id: TenantId,
2708 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2709 0 : ) -> Result<(), ApiError> {
2710 : // Check if the tenant is present in memory, and select an AZ to use when loading
2711 : // if we will load it.
2712 0 : let load_in_az = {
2713 0 : let locked = self.inner.read().unwrap();
2714 0 : let existing = locked
2715 0 : .tenants
2716 0 : .range(TenantShardId::tenant_range(tenant_id))
2717 0 : .next();
2718 0 :
2719 0 : // If the tenant is not present in memory, we expect to load it from database,
2720 0 : // so let's figure out what AZ to load it into while we have self.inner locked.
2721 0 : if existing.is_none() {
2722 0 : locked
2723 0 : .scheduler
2724 0 : .get_az_for_new_tenant()
2725 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
2726 0 : "No AZ with nodes found to load tenant"
2727 0 : )))?
2728 : } else {
2729 : // We already have this tenant in memory
2730 0 : return Ok(());
2731 : }
2732 : };
2733 :
2734 0 : let tenant_shards = self.persistence.load_tenant(tenant_id).await?;
2735 0 : if tenant_shards.is_empty() {
2736 0 : return Err(ApiError::NotFound(
2737 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2738 0 : ));
2739 0 : }
2740 0 :
2741 0 : // Update the persistent shards with the AZ that we are about to apply to in-memory state
2742 0 : self.persistence
2743 0 : .set_tenant_shard_preferred_azs(
2744 0 : tenant_shards
2745 0 : .iter()
2746 0 : .map(|t| {
2747 0 : (
2748 0 : t.get_tenant_shard_id().expect("Corrupt shard in database"),
2749 0 : Some(load_in_az.clone()),
2750 0 : )
2751 0 : })
2752 0 : .collect(),
2753 0 : )
2754 0 : .await?;
2755 :
2756 0 : let mut locked = self.inner.write().unwrap();
2757 0 : tracing::info!(
2758 0 : "Loaded {} shards for tenant {}",
2759 0 : tenant_shards.len(),
2760 : tenant_id
2761 : );
2762 :
2763 0 : locked.tenants.extend(tenant_shards.into_iter().map(|p| {
2764 0 : let intent = IntentState::new(Some(load_in_az.clone()));
2765 0 : let shard =
2766 0 : TenantShard::from_persistent(p, intent).expect("Corrupt shard row in database");
2767 0 :
2768 0 : // Sanity check: when loading on-demand, we should always be loaded something Detached
2769 0 : debug_assert!(shard.policy == PlacementPolicy::Detached);
2770 0 : if shard.policy != PlacementPolicy::Detached {
2771 0 : tracing::error!(
2772 0 : "Tenant shard {} loaded on-demand, but has non-Detached policy {:?}",
2773 : shard.tenant_shard_id,
2774 : shard.policy
2775 : );
2776 0 : }
2777 :
2778 0 : (shard.tenant_shard_id, shard)
2779 0 : }));
2780 0 :
2781 0 : Ok(())
2782 0 : }
2783 :
2784 : /// If all shards for a tenant are detached, and in a fully quiescent state (no observed locations on pageservers),
2785 : /// and have no reconciler running, then we can drop the tenant from memory. It will be reloaded on-demand
2786 : /// if we are asked to attach it again (see [`Self::maybe_load_tenant`]).
2787 : ///
2788 : /// Caller must demonstrate they hold a lock guard, as otherwise it is unsafe to drop a tenant from
2789 : /// memory while some other function might assume it continues to exist while not holding the lock on Self::inner.
2790 0 : fn maybe_drop_tenant(
2791 0 : &self,
2792 0 : tenant_id: TenantId,
2793 0 : locked: &mut std::sync::RwLockWriteGuard<ServiceState>,
2794 0 : _guard: &TracingExclusiveGuard<TenantOperations>,
2795 0 : ) {
2796 0 : let mut tenant_shards = locked.tenants.range(TenantShardId::tenant_range(tenant_id));
2797 0 : if tenant_shards.all(|(_id, shard)| {
2798 0 : shard.policy == PlacementPolicy::Detached
2799 0 : && shard.reconciler.is_none()
2800 0 : && shard.observed.is_empty()
2801 0 : }) {
2802 0 : let keys = locked
2803 0 : .tenants
2804 0 : .range(TenantShardId::tenant_range(tenant_id))
2805 0 : .map(|(id, _)| id)
2806 0 : .copied()
2807 0 : .collect::<Vec<_>>();
2808 0 : for key in keys {
2809 0 : tracing::info!("Dropping detached tenant shard {} from memory", key);
2810 0 : locked.tenants.remove(&key);
2811 : }
2812 0 : }
2813 0 : }
2814 :
2815 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2816 : /// directly with pageservers into this service.
2817 : ///
2818 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2819 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2820 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2821 : /// tenant's source of generation numbers.
2822 : ///
2823 : /// The mode in this request coarse-grained control of tenants:
2824 : /// - Call with mode Attached* to upsert the tenant.
2825 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2826 : /// to set an existing tenant to PolicyMode::Secondary
2827 : /// - Call with mode Detached to switch to PolicyMode::Detached
2828 0 : pub(crate) async fn tenant_location_config(
2829 0 : &self,
2830 0 : tenant_shard_id: TenantShardId,
2831 0 : req: TenantLocationConfigRequest,
2832 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2833 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2834 0 : let _tenant_lock = trace_exclusive_lock(
2835 0 : &self.tenant_op_locks,
2836 0 : tenant_shard_id.tenant_id,
2837 0 : TenantOperations::LocationConfig,
2838 0 : )
2839 0 : .await;
2840 :
2841 0 : let tenant_id = if !tenant_shard_id.is_unsharded() {
2842 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2843 0 : "This API is for importing single-sharded or unsharded tenants"
2844 0 : )));
2845 : } else {
2846 0 : tenant_shard_id.tenant_id
2847 0 : };
2848 0 :
2849 0 : // In case we are waking up a Detached tenant
2850 0 : match self.maybe_load_tenant(tenant_id, &_tenant_lock).await {
2851 0 : Ok(()) | Err(ApiError::NotFound(_)) => {
2852 0 : // This is a creation or an update
2853 0 : }
2854 0 : Err(e) => {
2855 0 : return Err(e);
2856 : }
2857 : };
2858 :
2859 : // First check if this is a creation or an update
2860 0 : let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
2861 0 :
2862 0 : let mut result = TenantLocationConfigResponse {
2863 0 : shards: Vec::new(),
2864 0 : stripe_size: None,
2865 0 : };
2866 0 : let waiters = match create_or_update {
2867 0 : TenantCreateOrUpdate::Create(create_req) => {
2868 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2869 0 : result.shards = create_resp
2870 0 : .shards
2871 0 : .into_iter()
2872 0 : .map(|s| TenantShardLocation {
2873 0 : node_id: s.node_id,
2874 0 : shard_id: s.shard_id,
2875 0 : })
2876 0 : .collect();
2877 0 : waiters
2878 : }
2879 0 : TenantCreateOrUpdate::Update(updates) => {
2880 0 : // Persist updates
2881 0 : // Ordering: write to the database before applying changes in-memory, so that
2882 0 : // we will not appear time-travel backwards on a restart.
2883 0 :
2884 0 : let mut schedule_context = ScheduleContext::default();
2885 : for ShardUpdate {
2886 0 : tenant_shard_id,
2887 0 : placement_policy,
2888 0 : tenant_config,
2889 0 : generation,
2890 0 : scheduling_policy,
2891 0 : } in &updates
2892 : {
2893 0 : self.persistence
2894 0 : .update_tenant_shard(
2895 0 : TenantFilter::Shard(*tenant_shard_id),
2896 0 : Some(placement_policy.clone()),
2897 0 : Some(tenant_config.clone()),
2898 0 : *generation,
2899 0 : *scheduling_policy,
2900 0 : )
2901 0 : .await?;
2902 : }
2903 :
2904 : // Apply updates in-memory
2905 0 : let mut waiters = Vec::new();
2906 0 : {
2907 0 : let mut locked = self.inner.write().unwrap();
2908 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2909 :
2910 : for ShardUpdate {
2911 0 : tenant_shard_id,
2912 0 : placement_policy,
2913 0 : tenant_config,
2914 0 : generation: update_generation,
2915 0 : scheduling_policy,
2916 0 : } in updates
2917 : {
2918 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2919 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2920 0 : continue;
2921 : };
2922 :
2923 : // Update stripe size
2924 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2925 0 : result.stripe_size = Some(shard.shard.stripe_size);
2926 0 : }
2927 :
2928 0 : shard.policy = placement_policy;
2929 0 : shard.config = tenant_config;
2930 0 : if let Some(generation) = update_generation {
2931 0 : shard.generation = Some(generation);
2932 0 : }
2933 :
2934 0 : if let Some(scheduling_policy) = scheduling_policy {
2935 0 : shard.set_scheduling_policy(scheduling_policy);
2936 0 : }
2937 :
2938 0 : shard.schedule(scheduler, &mut schedule_context)?;
2939 :
2940 0 : let maybe_waiter =
2941 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
2942 0 : if let Some(waiter) = maybe_waiter {
2943 0 : waiters.push(waiter);
2944 0 : }
2945 :
2946 0 : if let Some(node_id) = shard.intent.get_attached() {
2947 0 : result.shards.push(TenantShardLocation {
2948 0 : shard_id: tenant_shard_id,
2949 0 : node_id: *node_id,
2950 0 : })
2951 0 : }
2952 : }
2953 : }
2954 0 : waiters
2955 : }
2956 : };
2957 :
2958 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2959 : // Do not treat a reconcile error as fatal: we have already applied any requested
2960 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2961 : // compute notification API. In these cases, it is important that we do not
2962 : // cause the cloud control plane to retry forever on this API.
2963 0 : tracing::warn!(
2964 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2965 : );
2966 0 : }
2967 :
2968 : // Logging the full result is useful because it lets us cross-check what the cloud control
2969 : // plane's tenant_shards table should contain.
2970 0 : tracing::info!("Complete, returning {result:?}");
2971 :
2972 0 : Ok(result)
2973 0 : }
2974 :
2975 0 : pub(crate) async fn tenant_config_patch(
2976 0 : &self,
2977 0 : req: TenantConfigPatchRequest,
2978 0 : ) -> Result<(), ApiError> {
2979 0 : let _tenant_lock = trace_exclusive_lock(
2980 0 : &self.tenant_op_locks,
2981 0 : req.tenant_id,
2982 0 : TenantOperations::ConfigPatch,
2983 0 : )
2984 0 : .await;
2985 :
2986 0 : let tenant_id = req.tenant_id;
2987 0 : let patch = req.config;
2988 0 :
2989 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
2990 :
2991 0 : let base = {
2992 0 : let locked = self.inner.read().unwrap();
2993 0 : let shards = locked
2994 0 : .tenants
2995 0 : .range(TenantShardId::tenant_range(req.tenant_id));
2996 0 :
2997 0 : let mut configs = shards.map(|(_sid, shard)| &shard.config).peekable();
2998 :
2999 0 : let first = match configs.peek() {
3000 0 : Some(first) => (*first).clone(),
3001 : None => {
3002 0 : return Err(ApiError::NotFound(
3003 0 : anyhow::anyhow!("Tenant {} not found", req.tenant_id).into(),
3004 0 : ));
3005 : }
3006 : };
3007 :
3008 0 : if !configs.all_equal() {
3009 0 : tracing::error!("Tenant configs for {} are mismatched. ", req.tenant_id);
3010 : // This can't happen because we atomically update the database records
3011 : // of all shards to the new value in [`Self::set_tenant_config_and_reconcile`].
3012 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3013 0 : "Tenant configs for {} are mismatched",
3014 0 : req.tenant_id
3015 0 : )));
3016 0 : }
3017 0 :
3018 0 : first
3019 : };
3020 :
3021 0 : let updated_config = base
3022 0 : .apply_patch(patch)
3023 0 : .map_err(|err| ApiError::BadRequest(anyhow::anyhow!(err)))?;
3024 0 : self.set_tenant_config_and_reconcile(tenant_id, updated_config)
3025 0 : .await
3026 0 : }
3027 :
3028 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
3029 : // We require an exclusive lock, because we are updating persistent and in-memory state
3030 0 : let _tenant_lock = trace_exclusive_lock(
3031 0 : &self.tenant_op_locks,
3032 0 : req.tenant_id,
3033 0 : TenantOperations::ConfigSet,
3034 0 : )
3035 0 : .await;
3036 :
3037 0 : self.maybe_load_tenant(req.tenant_id, &_tenant_lock).await?;
3038 :
3039 0 : self.set_tenant_config_and_reconcile(req.tenant_id, req.config)
3040 0 : .await
3041 0 : }
3042 :
3043 0 : async fn set_tenant_config_and_reconcile(
3044 0 : &self,
3045 0 : tenant_id: TenantId,
3046 0 : config: TenantConfig,
3047 0 : ) -> Result<(), ApiError> {
3048 0 : self.persistence
3049 0 : .update_tenant_shard(
3050 0 : TenantFilter::Tenant(tenant_id),
3051 0 : None,
3052 0 : Some(config.clone()),
3053 0 : None,
3054 0 : None,
3055 0 : )
3056 0 : .await?;
3057 :
3058 0 : let waiters = {
3059 0 : let mut waiters = Vec::new();
3060 0 : let mut locked = self.inner.write().unwrap();
3061 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
3062 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3063 0 : shard.config = config.clone();
3064 0 : if let Some(waiter) =
3065 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3066 0 : {
3067 0 : waiters.push(waiter);
3068 0 : }
3069 : }
3070 0 : waiters
3071 : };
3072 :
3073 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
3074 : // Treat this as success because we have stored the configuration. If e.g.
3075 : // a node was unavailable at this time, it should not stop us accepting a
3076 : // configuration change.
3077 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
3078 0 : }
3079 :
3080 0 : Ok(())
3081 0 : }
3082 :
3083 0 : pub(crate) fn tenant_config_get(
3084 0 : &self,
3085 0 : tenant_id: TenantId,
3086 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
3087 0 : let config = {
3088 0 : let locked = self.inner.read().unwrap();
3089 0 :
3090 0 : match locked
3091 0 : .tenants
3092 0 : .range(TenantShardId::tenant_range(tenant_id))
3093 0 : .next()
3094 : {
3095 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
3096 : None => {
3097 0 : return Err(ApiError::NotFound(
3098 0 : anyhow::anyhow!("Tenant not found").into(),
3099 0 : ));
3100 : }
3101 : }
3102 : };
3103 :
3104 : // Unlike the pageserver, we do not have a set of global defaults: the config is
3105 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
3106 : // and `effective_config` in the response is meaningless, but we retain that syntax
3107 : // in order to remain compatible with the pageserver API.
3108 :
3109 0 : let response = HashMap::from([
3110 : (
3111 : "tenant_specific_overrides",
3112 0 : serde_json::to_value(&config)
3113 0 : .context("serializing tenant specific overrides")
3114 0 : .map_err(ApiError::InternalServerError)?,
3115 : ),
3116 : (
3117 0 : "effective_config",
3118 0 : serde_json::to_value(&config)
3119 0 : .context("serializing effective config")
3120 0 : .map_err(ApiError::InternalServerError)?,
3121 : ),
3122 : ]);
3123 :
3124 0 : Ok(response)
3125 0 : }
3126 :
3127 0 : pub(crate) async fn tenant_time_travel_remote_storage(
3128 0 : &self,
3129 0 : time_travel_req: &TenantTimeTravelRequest,
3130 0 : tenant_id: TenantId,
3131 0 : timestamp: Cow<'_, str>,
3132 0 : done_if_after: Cow<'_, str>,
3133 0 : ) -> Result<(), ApiError> {
3134 0 : let _tenant_lock = trace_exclusive_lock(
3135 0 : &self.tenant_op_locks,
3136 0 : tenant_id,
3137 0 : TenantOperations::TimeTravelRemoteStorage,
3138 0 : )
3139 0 : .await;
3140 :
3141 0 : let node = {
3142 0 : let mut locked = self.inner.write().unwrap();
3143 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
3144 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
3145 : // but only at the start of the process, so it's really just to prevent operator
3146 : // mistakes.
3147 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
3148 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
3149 : {
3150 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3151 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
3152 0 : )));
3153 0 : }
3154 0 : let maybe_attached = shard
3155 0 : .observed
3156 0 : .locations
3157 0 : .iter()
3158 0 : .filter_map(|(node_id, observed_location)| {
3159 0 : observed_location
3160 0 : .conf
3161 0 : .as_ref()
3162 0 : .map(|loc| (node_id, observed_location, loc.mode))
3163 0 : })
3164 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
3165 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
3166 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3167 0 : "We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}"
3168 0 : )));
3169 0 : }
3170 : }
3171 0 : let scheduler = &mut locked.scheduler;
3172 : // Right now we only perform the operation on a single node without parallelization
3173 : // TODO fan out the operation to multiple nodes for better performance
3174 0 : let node_id = scheduler.any_available_node()?;
3175 0 : let node = locked
3176 0 : .nodes
3177 0 : .get(&node_id)
3178 0 : .expect("Pageservers may not be deleted while lock is active");
3179 0 : node.clone()
3180 0 : };
3181 0 :
3182 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
3183 0 : let mut counts = time_travel_req
3184 0 : .shard_counts
3185 0 : .iter()
3186 0 : .copied()
3187 0 : .collect::<HashSet<_>>()
3188 0 : .into_iter()
3189 0 : .collect::<Vec<_>>();
3190 0 : counts.sort_unstable();
3191 :
3192 0 : for count in counts {
3193 0 : let shard_ids = (0..count.count())
3194 0 : .map(|i| TenantShardId {
3195 0 : tenant_id,
3196 0 : shard_number: ShardNumber(i),
3197 0 : shard_count: count,
3198 0 : })
3199 0 : .collect::<Vec<_>>();
3200 0 : for tenant_shard_id in shard_ids {
3201 0 : let client = PageserverClient::new(
3202 0 : node.get_id(),
3203 0 : node.base_url(),
3204 0 : self.config.pageserver_jwt_token.as_deref(),
3205 0 : self.config.ssl_ca_cert.clone(),
3206 0 : )
3207 0 : .map_err(|e| passthrough_api_error(&node, e))?;
3208 :
3209 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
3210 :
3211 0 : client
3212 0 : .tenant_time_travel_remote_storage(
3213 0 : tenant_shard_id,
3214 0 : ×tamp,
3215 0 : &done_if_after,
3216 0 : )
3217 0 : .await
3218 0 : .map_err(|e| {
3219 0 : ApiError::InternalServerError(anyhow::anyhow!(
3220 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
3221 0 : node
3222 0 : ))
3223 0 : })?;
3224 : }
3225 : }
3226 0 : Ok(())
3227 0 : }
3228 :
3229 0 : pub(crate) async fn tenant_secondary_download(
3230 0 : &self,
3231 0 : tenant_id: TenantId,
3232 0 : wait: Option<Duration>,
3233 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
3234 0 : let _tenant_lock = trace_shared_lock(
3235 0 : &self.tenant_op_locks,
3236 0 : tenant_id,
3237 0 : TenantOperations::SecondaryDownload,
3238 0 : )
3239 0 : .await;
3240 :
3241 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
3242 0 : let targets = {
3243 0 : let locked = self.inner.read().unwrap();
3244 0 : let mut targets = Vec::new();
3245 :
3246 0 : for (tenant_shard_id, shard) in
3247 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3248 : {
3249 0 : for node_id in shard.intent.get_secondary() {
3250 0 : let node = locked
3251 0 : .nodes
3252 0 : .get(node_id)
3253 0 : .expect("Pageservers may not be deleted while referenced");
3254 0 :
3255 0 : targets.push((*tenant_shard_id, node.clone()));
3256 0 : }
3257 : }
3258 0 : targets
3259 0 : };
3260 0 :
3261 0 : // Issue concurrent requests to all shards' locations
3262 0 : let mut futs = FuturesUnordered::new();
3263 0 : for (tenant_shard_id, node) in targets {
3264 0 : let client = PageserverClient::new(
3265 0 : node.get_id(),
3266 0 : node.base_url(),
3267 0 : self.config.pageserver_jwt_token.as_deref(),
3268 0 : self.config.ssl_ca_cert.clone(),
3269 0 : )
3270 0 : .map_err(|e| passthrough_api_error(&node, e))?;
3271 0 : futs.push(async move {
3272 0 : let result = client
3273 0 : .tenant_secondary_download(tenant_shard_id, wait)
3274 0 : .await;
3275 0 : (result, node, tenant_shard_id)
3276 0 : })
3277 : }
3278 :
3279 : // Handle any errors returned by pageservers. This includes cases like this request racing with
3280 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
3281 : // well as more general cases like 503s, 500s, or timeouts.
3282 0 : let mut aggregate_progress = SecondaryProgress::default();
3283 0 : let mut aggregate_status: Option<StatusCode> = None;
3284 0 : let mut error: Option<mgmt_api::Error> = None;
3285 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
3286 0 : match result {
3287 0 : Err(e) => {
3288 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
3289 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
3290 0 : // than they had hoped for.
3291 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
3292 0 : error = Some(e)
3293 : }
3294 0 : Ok((status_code, progress)) => {
3295 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
3296 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
3297 0 : aggregate_progress.layers_total += progress.layers_total;
3298 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
3299 0 : aggregate_progress.bytes_total += progress.bytes_total;
3300 0 : aggregate_progress.heatmap_mtime =
3301 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
3302 0 : aggregate_status = match aggregate_status {
3303 0 : None => Some(status_code),
3304 0 : Some(StatusCode::OK) => Some(status_code),
3305 0 : Some(cur) => {
3306 0 : // Other status codes (e.g. 202) -- do not overwrite.
3307 0 : Some(cur)
3308 : }
3309 : };
3310 : }
3311 : }
3312 : }
3313 :
3314 : // If any of the shards return 202, indicate our result as 202.
3315 0 : match aggregate_status {
3316 : None => {
3317 0 : match error {
3318 0 : Some(e) => {
3319 0 : // No successes, and an error: surface it
3320 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
3321 : }
3322 : None => {
3323 : // No shards found
3324 0 : Err(ApiError::NotFound(
3325 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
3326 0 : ))
3327 : }
3328 : }
3329 : }
3330 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
3331 : }
3332 0 : }
3333 :
3334 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
3335 0 : let _tenant_lock =
3336 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
3337 :
3338 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3339 :
3340 : // Detach all shards. This also deletes local pageserver shard data.
3341 0 : let (detach_waiters, node) = {
3342 0 : let mut detach_waiters = Vec::new();
3343 0 : let mut locked = self.inner.write().unwrap();
3344 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3345 0 : for (_, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3346 : // Update the tenant's intent to remove all attachments
3347 0 : shard.policy = PlacementPolicy::Detached;
3348 0 : shard
3349 0 : .schedule(scheduler, &mut ScheduleContext::default())
3350 0 : .expect("De-scheduling is infallible");
3351 0 : debug_assert!(shard.intent.get_attached().is_none());
3352 0 : debug_assert!(shard.intent.get_secondary().is_empty());
3353 :
3354 0 : if let Some(waiter) =
3355 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
3356 0 : {
3357 0 : detach_waiters.push(waiter);
3358 0 : }
3359 : }
3360 :
3361 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
3362 : // was attached, just has to be able to see the S3 content)
3363 0 : let node_id = scheduler.any_available_node()?;
3364 0 : let node = nodes
3365 0 : .get(&node_id)
3366 0 : .expect("Pageservers may not be deleted while lock is active");
3367 0 : (detach_waiters, node.clone())
3368 0 : };
3369 0 :
3370 0 : // This reconcile wait can fail in a few ways:
3371 0 : // A there is a very long queue for the reconciler semaphore
3372 0 : // B some pageserver is failing to handle a detach promptly
3373 0 : // C some pageserver goes offline right at the moment we send it a request.
3374 0 : //
3375 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
3376 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
3377 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
3378 0 : // deleting the underlying data).
3379 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
3380 0 : .await?;
3381 :
3382 : // Delete the entire tenant (all shards) from remote storage via a random pageserver.
3383 : // Passing an unsharded tenant ID will cause the pageserver to remove all remote paths with
3384 : // the tenant ID prefix, including all shards (even possibly stale ones).
3385 0 : match node
3386 0 : .with_client_retries(
3387 0 : |client| async move {
3388 0 : client
3389 0 : .tenant_delete(TenantShardId::unsharded(tenant_id))
3390 0 : .await
3391 0 : },
3392 0 : &self.config.pageserver_jwt_token,
3393 0 : &self.config.ssl_ca_cert,
3394 0 : 1,
3395 0 : 3,
3396 0 : RECONCILE_TIMEOUT,
3397 0 : &self.cancel,
3398 0 : )
3399 0 : .await
3400 0 : .unwrap_or(Err(mgmt_api::Error::Cancelled))
3401 : {
3402 0 : Ok(_) => {}
3403 : Err(mgmt_api::Error::Cancelled) => {
3404 0 : return Err(ApiError::ShuttingDown);
3405 : }
3406 0 : Err(e) => {
3407 0 : // This is unexpected: remote deletion should be infallible, unless the object store
3408 0 : // at large is unavailable.
3409 0 : tracing::error!("Error deleting via node {node}: {e}");
3410 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
3411 : }
3412 : }
3413 :
3414 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
3415 : // our in-memory state and database state.
3416 :
3417 : // Ordering: we delete persistent state first: if we then
3418 : // crash, we will drop the in-memory state.
3419 :
3420 : // Drop persistent state.
3421 0 : self.persistence.delete_tenant(tenant_id).await?;
3422 :
3423 : // Drop in-memory state
3424 : {
3425 0 : let mut locked = self.inner.write().unwrap();
3426 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
3427 :
3428 : // Dereference Scheduler from shards before dropping them
3429 0 : for (_tenant_shard_id, shard) in
3430 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
3431 0 : {
3432 0 : shard.intent.clear(scheduler);
3433 0 : }
3434 :
3435 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
3436 0 : tracing::info!(
3437 0 : "Deleted tenant {tenant_id}, now have {} tenants",
3438 0 : locked.tenants.len()
3439 : );
3440 : };
3441 :
3442 : // Success is represented as 404, to imitate the existing pageserver deletion API
3443 0 : Ok(StatusCode::NOT_FOUND)
3444 0 : }
3445 :
3446 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
3447 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
3448 : /// the tenant's policies (configuration) within the storage controller
3449 0 : pub(crate) async fn tenant_update_policy(
3450 0 : &self,
3451 0 : tenant_id: TenantId,
3452 0 : req: TenantPolicyRequest,
3453 0 : ) -> Result<(), ApiError> {
3454 : // We require an exclusive lock, because we are updating persistent and in-memory state
3455 0 : let _tenant_lock = trace_exclusive_lock(
3456 0 : &self.tenant_op_locks,
3457 0 : tenant_id,
3458 0 : TenantOperations::UpdatePolicy,
3459 0 : )
3460 0 : .await;
3461 :
3462 0 : self.maybe_load_tenant(tenant_id, &_tenant_lock).await?;
3463 :
3464 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
3465 :
3466 : let TenantPolicyRequest {
3467 0 : placement,
3468 0 : mut scheduling,
3469 0 : } = req;
3470 :
3471 0 : if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
3472 : // When someone configures a tenant to detach, we force the scheduling policy to enable
3473 : // this to take effect.
3474 0 : if scheduling.is_none() {
3475 0 : scheduling = Some(ShardSchedulingPolicy::Active);
3476 0 : }
3477 0 : }
3478 :
3479 0 : self.persistence
3480 0 : .update_tenant_shard(
3481 0 : TenantFilter::Tenant(tenant_id),
3482 0 : placement.clone(),
3483 0 : None,
3484 0 : None,
3485 0 : scheduling,
3486 0 : )
3487 0 : .await?;
3488 :
3489 0 : let mut schedule_context = ScheduleContext::default();
3490 0 : let mut locked = self.inner.write().unwrap();
3491 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3492 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
3493 0 : if let Some(placement) = &placement {
3494 0 : shard.policy = placement.clone();
3495 0 :
3496 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3497 0 : "Updated placement policy to {placement:?}");
3498 0 : }
3499 :
3500 0 : if let Some(scheduling) = &scheduling {
3501 0 : shard.set_scheduling_policy(*scheduling);
3502 0 :
3503 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
3504 0 : "Updated scheduling policy to {scheduling:?}");
3505 0 : }
3506 :
3507 : // In case scheduling is being switched back on, try it now.
3508 0 : shard.schedule(scheduler, &mut schedule_context).ok();
3509 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
3510 : }
3511 :
3512 0 : Ok(())
3513 0 : }
3514 :
3515 0 : pub(crate) async fn tenant_timeline_create_pageservers(
3516 0 : &self,
3517 0 : tenant_id: TenantId,
3518 0 : mut create_req: TimelineCreateRequest,
3519 0 : ) -> Result<TimelineInfo, ApiError> {
3520 0 : tracing::info!(
3521 0 : "Creating timeline {}/{}",
3522 : tenant_id,
3523 : create_req.new_timeline_id,
3524 : );
3525 :
3526 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3527 0 : if targets.0.is_empty() {
3528 0 : return Err(ApiError::NotFound(
3529 0 : anyhow::anyhow!("Tenant not found").into(),
3530 0 : ));
3531 0 : };
3532 0 :
3533 0 : let (shard_zero_tid, shard_zero_locations) =
3534 0 : targets.0.pop_first().expect("Must have at least one shard");
3535 0 : assert!(shard_zero_tid.is_shard_zero());
3536 :
3537 0 : async fn create_one(
3538 0 : tenant_shard_id: TenantShardId,
3539 0 : locations: ShardMutationLocations,
3540 0 : jwt: Option<String>,
3541 0 : ssl_ca_cert: Option<Certificate>,
3542 0 : create_req: TimelineCreateRequest,
3543 0 : ) -> Result<TimelineInfo, ApiError> {
3544 0 : let latest = locations.latest.node;
3545 0 :
3546 0 : tracing::info!(
3547 0 : "Creating timeline on shard {}/{}, attached to node {latest} in generation {:?}",
3548 : tenant_shard_id,
3549 : create_req.new_timeline_id,
3550 : locations.latest.generation
3551 : );
3552 :
3553 0 : let client =
3554 0 : PageserverClient::new(latest.get_id(), latest.base_url(), jwt.as_deref(), ssl_ca_cert.clone())
3555 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3556 :
3557 0 : let timeline_info = client
3558 0 : .timeline_create(tenant_shard_id, &create_req)
3559 0 : .await
3560 0 : .map_err(|e| passthrough_api_error(&latest, e))?;
3561 :
3562 : // We propagate timeline creations to all attached locations such that a compute
3563 : // for the new timeline is able to start regardless of the current state of the
3564 : // tenant shard reconciliation.
3565 0 : for location in locations.other {
3566 0 : tracing::info!(
3567 0 : "Creating timeline on shard {}/{}, stale attached to node {} in generation {:?}",
3568 : tenant_shard_id,
3569 : create_req.new_timeline_id,
3570 : location.node,
3571 : location.generation
3572 : );
3573 :
3574 0 : let client = PageserverClient::new(
3575 0 : location.node.get_id(),
3576 0 : location.node.base_url(),
3577 0 : jwt.as_deref(),
3578 0 : ssl_ca_cert.clone(),
3579 0 : )
3580 0 : .map_err(|e| passthrough_api_error(&location.node, e))?;
3581 :
3582 0 : let res = client
3583 0 : .timeline_create(tenant_shard_id, &create_req)
3584 0 : .await;
3585 :
3586 0 : if let Err(e) = res {
3587 0 : match e {
3588 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
3589 0 : // Tenant might have been detached from the stale location,
3590 0 : // so ignore 404s.
3591 0 : },
3592 : _ => {
3593 0 : return Err(passthrough_api_error(&location.node, e));
3594 : }
3595 : }
3596 0 : }
3597 : }
3598 :
3599 0 : Ok(timeline_info)
3600 0 : }
3601 :
3602 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3603 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3604 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3605 0 : let timeline_info = create_one(
3606 0 : shard_zero_tid,
3607 0 : shard_zero_locations,
3608 0 : self.config.pageserver_jwt_token.clone(),
3609 0 : self.config.ssl_ca_cert.clone(),
3610 0 : create_req.clone(),
3611 0 : )
3612 0 : .await?;
3613 :
3614 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3615 0 : match &mut create_req.mode {
3616 0 : models::TimelineCreateRequestMode::Branch { ancestor_start_lsn, .. } if ancestor_start_lsn.is_none() => {
3617 0 : *ancestor_start_lsn = timeline_info.ancestor_lsn;
3618 0 : },
3619 0 : _ => {}
3620 : }
3621 :
3622 : // Create timeline on remaining shards with number >0
3623 0 : if !targets.0.is_empty() {
3624 : // If we had multiple shards, issue requests for the remainder now.
3625 0 : let jwt = &self.config.pageserver_jwt_token;
3626 0 : self.tenant_for_shards(
3627 0 : targets
3628 0 : .0
3629 0 : .iter()
3630 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
3631 0 : .collect(),
3632 0 : |tenant_shard_id: TenantShardId, _node: Node| {
3633 0 : let create_req = create_req.clone();
3634 0 : let mutation_locations = targets.0.remove(&tenant_shard_id).unwrap();
3635 0 : Box::pin(create_one(
3636 0 : tenant_shard_id,
3637 0 : mutation_locations,
3638 0 : jwt.clone(),
3639 0 : self.config.ssl_ca_cert.clone(),
3640 0 : create_req,
3641 0 : ))
3642 0 : },
3643 0 : )
3644 0 : .await?;
3645 0 : }
3646 :
3647 0 : Ok(timeline_info)
3648 0 : })
3649 0 : .await?
3650 0 : }
3651 :
3652 : /// Timeline creation on safekeepers
3653 : ///
3654 : /// Returns `Ok(left)` if the timeline has been created on a quorum of safekeepers,
3655 : /// where `left` contains the list of safekeepers that didn't have a successful response.
3656 : /// Assumes tenant lock is held while calling this function.
3657 0 : async fn tenant_timeline_create_safekeepers_quorum(
3658 0 : &self,
3659 0 : tenant_id: TenantId,
3660 0 : timeline_id: TimelineId,
3661 0 : pg_version: u32,
3662 0 : timeline_persistence: &TimelinePersistence,
3663 0 : ) -> Result<Vec<NodeId>, ApiError> {
3664 0 : // If quorum is reached, return if we are outside of a specified timeout
3665 0 : let jwt = self
3666 0 : .config
3667 0 : .safekeeper_jwt_token
3668 0 : .clone()
3669 0 : .map(SecretString::from);
3670 0 : let mut joinset = JoinSet::new();
3671 0 :
3672 0 : let safekeepers = {
3673 0 : let locked = self.inner.read().unwrap();
3674 0 : locked.safekeepers.clone()
3675 0 : };
3676 0 :
3677 0 : let mut members = Vec::new();
3678 0 : for sk_id in timeline_persistence.sk_set.iter() {
3679 0 : let sk_id = NodeId(*sk_id as u64);
3680 0 : let Some(safekeeper) = safekeepers.get(&sk_id) else {
3681 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3682 0 : "couldn't find entry for safekeeper with id {sk_id}"
3683 0 : )))?;
3684 : };
3685 0 : members.push(SafekeeperId {
3686 0 : id: sk_id,
3687 0 : host: safekeeper.skp.host.clone(),
3688 0 : pg_port: safekeeper.skp.port as u16,
3689 0 : });
3690 : }
3691 0 : let mset = MemberSet::new(members).map_err(ApiError::InternalServerError)?;
3692 0 : let mconf = safekeeper_api::membership::Configuration::new(mset);
3693 0 :
3694 0 : let req = safekeeper_api::models::TimelineCreateRequest {
3695 0 : commit_lsn: None,
3696 0 : mconf,
3697 0 : pg_version,
3698 0 : start_lsn: timeline_persistence.start_lsn.0,
3699 0 : system_id: None,
3700 0 : tenant_id,
3701 0 : timeline_id,
3702 0 : wal_seg_size: None,
3703 0 : };
3704 : const SK_CREATE_TIMELINE_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
3705 0 : for sk in timeline_persistence.sk_set.iter() {
3706 0 : let sk_id = NodeId(*sk as u64);
3707 0 : let safekeepers = safekeepers.clone();
3708 0 : let jwt = jwt.clone();
3709 0 : let ssl_ca_cert = self.config.ssl_ca_cert.clone();
3710 0 : let req = req.clone();
3711 0 : joinset.spawn(async move {
3712 0 : // Unwrap is fine as we already would have returned error above
3713 0 : let sk_p = safekeepers.get(&sk_id).unwrap();
3714 0 : let res = sk_p
3715 0 : .with_client_retries(
3716 0 : |client| {
3717 0 : let req = req.clone();
3718 0 : async move { client.create_timeline(&req).await }
3719 0 : },
3720 0 : &jwt,
3721 0 : &ssl_ca_cert,
3722 0 : 3,
3723 0 : 3,
3724 0 : SK_CREATE_TIMELINE_RECONCILE_TIMEOUT,
3725 0 : &CancellationToken::new(),
3726 0 : )
3727 0 : .await;
3728 0 : (sk_id, sk_p.skp.host.clone(), res)
3729 0 : });
3730 0 : }
3731 : // After we have built the joinset, we now wait for the tasks to complete,
3732 : // but with a specified timeout to make sure we return swiftly, either with
3733 : // a failure or success.
3734 0 : let reconcile_deadline = tokio::time::Instant::now() + SK_CREATE_TIMELINE_RECONCILE_TIMEOUT;
3735 0 :
3736 0 : // Wait until all tasks finish or timeout is hit, whichever occurs
3737 0 : // first.
3738 0 : let mut reconcile_results = Vec::new();
3739 : loop {
3740 0 : if let Ok(res) = tokio::time::timeout_at(reconcile_deadline, joinset.join_next()).await
3741 : {
3742 0 : let Some(res) = res else { break };
3743 0 : match res {
3744 0 : Ok(res) => {
3745 0 : tracing::info!(
3746 0 : "response from safekeeper id:{} at {}: {:?}",
3747 : res.0,
3748 : res.1,
3749 : res.2
3750 : );
3751 0 : reconcile_results.push(res);
3752 : }
3753 0 : Err(join_err) => {
3754 0 : tracing::info!("join_err for task in joinset: {join_err}");
3755 : }
3756 : }
3757 : } else {
3758 0 : tracing::info!(
3759 0 : "timeout for creation call after {} responses",
3760 0 : reconcile_results.len()
3761 : );
3762 0 : break;
3763 : }
3764 : }
3765 :
3766 : // Now check now if quorum was reached in reconcile_results.
3767 0 : let total_result_count = reconcile_results.len();
3768 0 : let remaining = reconcile_results
3769 0 : .into_iter()
3770 0 : .filter_map(|res| res.2.is_err().then_some(res.0))
3771 0 : .collect::<Vec<_>>();
3772 0 : tracing::info!(
3773 0 : "Got {} non-successful responses from initial creation request of total {total_result_count} responses",
3774 0 : remaining.len()
3775 : );
3776 0 : if remaining.len() >= 2 {
3777 : // Failure
3778 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3779 0 : "not enough successful reconciliations to reach quorum, please retry: {} errored",
3780 0 : remaining.len()
3781 0 : )));
3782 0 : }
3783 0 :
3784 0 : Ok(remaining)
3785 0 : }
3786 :
3787 : /// Create timeline in controller database and on safekeepers.
3788 : /// `timeline_info` is result of timeline creation on pageserver.
3789 : ///
3790 : /// All actions must be idempotent as the call is retried until success. It
3791 : /// tries to create timeline in the db and on at least majority of
3792 : /// safekeepers + queue creation for safekeepers which missed it in the db
3793 : /// for infinite retries; after that, call returns Ok.
3794 : ///
3795 : /// The idea is that once this is reached as long as we have alive majority
3796 : /// of safekeepers it is expected to get eventually operational as storcon
3797 : /// will be able to seed timeline on nodes which missed creation by making
3798 : /// pull_timeline from peers. On the other hand we don't want to fail
3799 : /// timeline creation if one safekeeper is down.
3800 0 : async fn tenant_timeline_create_safekeepers(
3801 0 : self: &Arc<Self>,
3802 0 : tenant_id: TenantId,
3803 0 : timeline_info: &TimelineInfo,
3804 0 : create_mode: models::TimelineCreateRequestMode,
3805 0 : ) -> Result<SafekeepersInfo, ApiError> {
3806 0 : let timeline_id = timeline_info.timeline_id;
3807 0 : let pg_version = timeline_info.pg_version;
3808 : // Initially start_lsn is determined by last_record_lsn in pageserver
3809 : // response as it does initdb. However, later we persist it and in sk
3810 : // creation calls replace with the value from the timeline row if it
3811 : // previously existed as on retries in theory endpoint might have
3812 : // already written some data and advanced last_record_lsn, while we want
3813 : // safekeepers to have consistent start_lsn.
3814 0 : let start_lsn = match create_mode {
3815 0 : models::TimelineCreateRequestMode::Bootstrap { .. } => timeline_info.last_record_lsn,
3816 0 : models::TimelineCreateRequestMode::Branch { .. } => timeline_info.last_record_lsn,
3817 : models::TimelineCreateRequestMode::ImportPgdata { .. } => {
3818 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3819 0 : "import pgdata doesn't specify the start lsn, aborting creation on safekeepers"
3820 0 : )))?;
3821 : }
3822 : };
3823 : // Choose initial set of safekeepers respecting affinity
3824 0 : let sks = self.safekeepers_for_new_timeline().await?;
3825 0 : let sks_persistence = sks.iter().map(|sk| sk.id.0 as i64).collect::<Vec<_>>();
3826 0 : // Add timeline to db
3827 0 : let mut timeline_persist = TimelinePersistence {
3828 0 : tenant_id: tenant_id.to_string(),
3829 0 : timeline_id: timeline_id.to_string(),
3830 0 : start_lsn: start_lsn.into(),
3831 0 : generation: 0,
3832 0 : sk_set: sks_persistence.clone(),
3833 0 : new_sk_set: None,
3834 0 : cplane_notified_generation: 0,
3835 0 : deleted_at: None,
3836 0 : };
3837 0 : let inserted = self
3838 0 : .persistence
3839 0 : .insert_timeline(timeline_persist.clone())
3840 0 : .await?;
3841 0 : if !inserted {
3842 0 : if let Some(existent_persist) = self
3843 0 : .persistence
3844 0 : .get_timeline(tenant_id, timeline_id)
3845 0 : .await?
3846 0 : {
3847 0 : // Replace with what we have in the db, to get stuff like the generation right.
3848 0 : // We do still repeat the http calls to the safekeepers. After all, we could have
3849 0 : // crashed right after the wrote to the DB.
3850 0 : timeline_persist = existent_persist;
3851 0 : } else {
3852 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3853 0 : "insertion said timeline already in db, but looking it up, it was gone"
3854 0 : )));
3855 : }
3856 0 : }
3857 : // Create the timeline on a quorum of safekeepers
3858 0 : let remaining = self
3859 0 : .tenant_timeline_create_safekeepers_quorum(
3860 0 : tenant_id,
3861 0 : timeline_id,
3862 0 : pg_version,
3863 0 : &timeline_persist,
3864 0 : )
3865 0 : .await?;
3866 :
3867 : // For the remaining safekeepers, take care of their reconciliation asynchronously
3868 0 : for &remaining_id in remaining.iter() {
3869 0 : let pending_op = TimelinePendingOpPersistence {
3870 0 : tenant_id: tenant_id.to_string(),
3871 0 : timeline_id: timeline_id.to_string(),
3872 0 : generation: timeline_persist.generation,
3873 0 : op_kind: crate::persistence::SafekeeperTimelineOpKind::Pull,
3874 0 : sk_id: remaining_id.0 as i64,
3875 0 : };
3876 0 : tracing::info!("writing pending op for sk id {remaining_id}");
3877 0 : self.persistence.insert_pending_op(pending_op).await?;
3878 : }
3879 0 : if !remaining.is_empty() {
3880 0 : let mut locked = self.inner.write().unwrap();
3881 0 : for remaining_id in remaining {
3882 0 : let Some(sk) = locked.safekeepers.get(&remaining_id) else {
3883 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3884 0 : "Couldn't find safekeeper with id {remaining_id}"
3885 0 : )));
3886 : };
3887 0 : let Ok(host_list) = sks
3888 0 : .iter()
3889 0 : .map(|sk| {
3890 0 : Ok((
3891 0 : sk.id,
3892 0 : locked
3893 0 : .safekeepers
3894 0 : .get(&sk.id)
3895 0 : .ok_or_else(|| {
3896 0 : ApiError::InternalServerError(anyhow::anyhow!(
3897 0 : "Couldn't find safekeeper with id {remaining_id} to pull from"
3898 0 : ))
3899 0 : })?
3900 0 : .base_url(),
3901 : ))
3902 0 : })
3903 0 : .collect::<Result<_, ApiError>>()
3904 : else {
3905 0 : continue;
3906 : };
3907 0 : let req = ScheduleRequest {
3908 0 : safekeeper: Box::new(sk.clone()),
3909 0 : host_list,
3910 0 : tenant_id,
3911 0 : timeline_id,
3912 0 : generation: timeline_persist.generation as u32,
3913 0 : kind: crate::persistence::SafekeeperTimelineOpKind::Pull,
3914 0 : };
3915 0 : locked.safekeeper_reconcilers.schedule_request(self, req);
3916 : }
3917 0 : }
3918 :
3919 0 : Ok(SafekeepersInfo {
3920 0 : generation: timeline_persist.generation as u32,
3921 0 : safekeepers: sks,
3922 0 : tenant_id,
3923 0 : timeline_id,
3924 0 : })
3925 0 : }
3926 :
3927 0 : pub(crate) async fn tenant_timeline_create(
3928 0 : self: &Arc<Self>,
3929 0 : tenant_id: TenantId,
3930 0 : create_req: TimelineCreateRequest,
3931 0 : ) -> Result<TimelineCreateResponseStorcon, ApiError> {
3932 0 : let safekeepers = self.config.timelines_onto_safekeepers;
3933 0 : tracing::info!(
3934 : %safekeepers,
3935 0 : "Creating timeline {}/{}",
3936 : tenant_id,
3937 : create_req.new_timeline_id,
3938 : );
3939 :
3940 0 : let _tenant_lock = trace_shared_lock(
3941 0 : &self.tenant_op_locks,
3942 0 : tenant_id,
3943 0 : TenantOperations::TimelineCreate,
3944 0 : )
3945 0 : .await;
3946 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
3947 0 : let create_mode = create_req.mode.clone();
3948 :
3949 0 : let timeline_info = self
3950 0 : .tenant_timeline_create_pageservers(tenant_id, create_req)
3951 0 : .await?;
3952 :
3953 0 : let safekeepers = if safekeepers {
3954 0 : let res = self
3955 0 : .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, create_mode)
3956 0 : .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
3957 0 : .await?;
3958 0 : Some(res)
3959 : } else {
3960 0 : None
3961 : };
3962 :
3963 0 : Ok(TimelineCreateResponseStorcon {
3964 0 : timeline_info,
3965 0 : safekeepers,
3966 0 : })
3967 0 : }
3968 :
3969 0 : pub(crate) async fn tenant_timeline_archival_config(
3970 0 : &self,
3971 0 : tenant_id: TenantId,
3972 0 : timeline_id: TimelineId,
3973 0 : req: TimelineArchivalConfigRequest,
3974 0 : ) -> Result<(), ApiError> {
3975 0 : tracing::info!(
3976 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3977 : req.state
3978 : );
3979 :
3980 0 : let _tenant_lock = trace_shared_lock(
3981 0 : &self.tenant_op_locks,
3982 0 : tenant_id,
3983 0 : TenantOperations::TimelineArchivalConfig,
3984 0 : )
3985 0 : .await;
3986 :
3987 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3988 0 : if targets.0.is_empty() {
3989 0 : return Err(ApiError::NotFound(
3990 0 : anyhow::anyhow!("Tenant not found").into(),
3991 0 : ));
3992 0 : }
3993 0 : async fn config_one(
3994 0 : tenant_shard_id: TenantShardId,
3995 0 : timeline_id: TimelineId,
3996 0 : node: Node,
3997 0 : jwt: Option<String>,
3998 0 : ssl_ca_cert: Option<Certificate>,
3999 0 : req: TimelineArchivalConfigRequest,
4000 0 : ) -> Result<(), ApiError> {
4001 0 : tracing::info!(
4002 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4003 : );
4004 :
4005 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref(), ssl_ca_cert)
4006 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4007 :
4008 0 : client
4009 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
4010 0 : .await
4011 0 : .map_err(|e| match e {
4012 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
4013 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
4014 : }
4015 0 : _ => passthrough_api_error(&node, e),
4016 0 : })
4017 0 : }
4018 :
4019 : // no shard needs to go first/last; the operation should be idempotent
4020 : // TODO: it would be great to ensure that all shards return the same error
4021 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4022 0 : let results = self
4023 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4024 0 : futures::FutureExt::boxed(config_one(
4025 0 : tenant_shard_id,
4026 0 : timeline_id,
4027 0 : node,
4028 0 : self.config.pageserver_jwt_token.clone(),
4029 0 : self.config.ssl_ca_cert.clone(),
4030 0 : req.clone(),
4031 0 : ))
4032 0 : })
4033 0 : .await?;
4034 0 : assert!(!results.is_empty(), "must have at least one result");
4035 :
4036 0 : Ok(())
4037 0 : }).await?
4038 0 : }
4039 :
4040 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
4041 0 : &self,
4042 0 : tenant_id: TenantId,
4043 0 : timeline_id: TimelineId,
4044 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
4045 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
4046 :
4047 0 : let _tenant_lock = trace_shared_lock(
4048 0 : &self.tenant_op_locks,
4049 0 : tenant_id,
4050 0 : TenantOperations::TimelineDetachAncestor,
4051 0 : )
4052 0 : .await;
4053 :
4054 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4055 0 : if targets.0.is_empty() {
4056 0 : return Err(ApiError::NotFound(
4057 0 : anyhow::anyhow!("Tenant not found").into(),
4058 0 : ));
4059 0 : }
4060 :
4061 0 : async fn detach_one(
4062 0 : tenant_shard_id: TenantShardId,
4063 0 : timeline_id: TimelineId,
4064 0 : node: Node,
4065 0 : jwt: Option<String>,
4066 0 : ssl_ca_cert: Option<Certificate>,
4067 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
4068 0 : tracing::info!(
4069 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4070 : );
4071 :
4072 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref(), ssl_ca_cert)
4073 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4074 :
4075 0 : client
4076 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
4077 0 : .await
4078 0 : .map_err(|e| {
4079 : use mgmt_api::Error;
4080 :
4081 0 : match e {
4082 : // no ancestor (ever)
4083 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
4084 0 : "{node}: {}",
4085 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
4086 0 : )),
4087 : // too many ancestors
4088 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
4089 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
4090 : }
4091 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
4092 0 : // avoid turning these into conflicts to remain compatible with
4093 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
4094 0 : // detach
4095 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
4096 : }
4097 : // rest can be mapped as usual
4098 0 : other => passthrough_api_error(&node, other),
4099 : }
4100 0 : })
4101 0 : .map(|res| (tenant_shard_id.shard_number, res))
4102 0 : }
4103 :
4104 : // no shard needs to go first/last; the operation should be idempotent
4105 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4106 0 : let mut results = self
4107 0 : .tenant_for_shards(locations, |tenant_shard_id, node| {
4108 0 : futures::FutureExt::boxed(detach_one(
4109 0 : tenant_shard_id,
4110 0 : timeline_id,
4111 0 : node,
4112 0 : self.config.pageserver_jwt_token.clone(),
4113 0 : self.config.ssl_ca_cert.clone(),
4114 0 : ))
4115 0 : })
4116 0 : .await?;
4117 :
4118 0 : let any = results.pop().expect("we must have at least one response");
4119 0 :
4120 0 : let mismatching = results
4121 0 : .iter()
4122 0 : .filter(|(_, res)| res != &any.1)
4123 0 : .collect::<Vec<_>>();
4124 0 : if !mismatching.is_empty() {
4125 : // this can be hit by races which should not happen because operation lock on cplane
4126 0 : let matching = results.len() - mismatching.len();
4127 0 : tracing::error!(
4128 : matching,
4129 : compared_against=?any,
4130 : ?mismatching,
4131 0 : "shards returned different results"
4132 : );
4133 :
4134 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
4135 0 : }
4136 0 :
4137 0 : Ok(any.1)
4138 0 : }).await?
4139 0 : }
4140 :
4141 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
4142 0 : &self,
4143 0 : tenant_id: TenantId,
4144 0 : timeline_id: TimelineId,
4145 0 : dir: BlockUnblock,
4146 0 : ) -> Result<(), ApiError> {
4147 0 : let _tenant_lock = trace_shared_lock(
4148 0 : &self.tenant_op_locks,
4149 0 : tenant_id,
4150 0 : TenantOperations::TimelineGcBlockUnblock,
4151 0 : )
4152 0 : .await;
4153 :
4154 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
4155 0 : if targets.0.is_empty() {
4156 0 : return Err(ApiError::NotFound(
4157 0 : anyhow::anyhow!("Tenant not found").into(),
4158 0 : ));
4159 0 : }
4160 :
4161 0 : async fn do_one(
4162 0 : tenant_shard_id: TenantShardId,
4163 0 : timeline_id: TimelineId,
4164 0 : node: Node,
4165 0 : jwt: Option<String>,
4166 0 : ssl_ca_cert: Option<Certificate>,
4167 0 : dir: BlockUnblock,
4168 0 : ) -> Result<(), ApiError> {
4169 0 : let client = PageserverClient::new(
4170 0 : node.get_id(),
4171 0 : node.base_url(),
4172 0 : jwt.as_deref(),
4173 0 : ssl_ca_cert,
4174 0 : )
4175 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4176 :
4177 0 : client
4178 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
4179 0 : .await
4180 0 : .map_err(|e| passthrough_api_error(&node, e))
4181 0 : }
4182 :
4183 : // no shard needs to go first/last; the operation should be idempotent
4184 0 : let locations = targets
4185 0 : .0
4186 0 : .iter()
4187 0 : .map(|t| (*t.0, t.1.latest.node.clone()))
4188 0 : .collect();
4189 0 : self.tenant_for_shards(locations, |tenant_shard_id, node| {
4190 0 : futures::FutureExt::boxed(do_one(
4191 0 : tenant_shard_id,
4192 0 : timeline_id,
4193 0 : node,
4194 0 : self.config.pageserver_jwt_token.clone(),
4195 0 : self.config.ssl_ca_cert.clone(),
4196 0 : dir,
4197 0 : ))
4198 0 : })
4199 0 : .await
4200 0 : })
4201 0 : .await??;
4202 0 : Ok(())
4203 0 : }
4204 :
4205 0 : pub(crate) async fn tenant_timeline_download_heatmap_layers(
4206 0 : &self,
4207 0 : tenant_shard_id: TenantShardId,
4208 0 : timeline_id: TimelineId,
4209 0 : concurrency: Option<usize>,
4210 0 : recurse: bool,
4211 0 : ) -> Result<(), ApiError> {
4212 0 : let _tenant_lock = trace_shared_lock(
4213 0 : &self.tenant_op_locks,
4214 0 : tenant_shard_id.tenant_id,
4215 0 : TenantOperations::DownloadHeatmapLayers,
4216 0 : )
4217 0 : .await;
4218 :
4219 0 : let targets = {
4220 0 : let locked = self.inner.read().unwrap();
4221 0 : let mut targets = Vec::new();
4222 :
4223 : // If the request got an unsharded tenant id, then apply
4224 : // the operation to all shards. Otherwise, apply it to a specific shard.
4225 0 : let shards_range = if tenant_shard_id.is_unsharded() {
4226 0 : TenantShardId::tenant_range(tenant_shard_id.tenant_id)
4227 : } else {
4228 0 : tenant_shard_id.range()
4229 : };
4230 :
4231 0 : for (tenant_shard_id, shard) in locked.tenants.range(shards_range) {
4232 0 : if let Some(node_id) = shard.intent.get_attached() {
4233 0 : let node = locked
4234 0 : .nodes
4235 0 : .get(node_id)
4236 0 : .expect("Pageservers may not be deleted while referenced");
4237 0 :
4238 0 : targets.push((*tenant_shard_id, node.clone()));
4239 0 : }
4240 : }
4241 0 : targets
4242 0 : };
4243 0 :
4244 0 : self.tenant_for_shards_api(
4245 0 : targets,
4246 0 : |tenant_shard_id, client| async move {
4247 0 : client
4248 0 : .timeline_download_heatmap_layers(
4249 0 : tenant_shard_id,
4250 0 : timeline_id,
4251 0 : concurrency,
4252 0 : recurse,
4253 0 : )
4254 0 : .await
4255 0 : },
4256 0 : 1,
4257 0 : 1,
4258 0 : SHORT_RECONCILE_TIMEOUT,
4259 0 : &self.cancel,
4260 0 : )
4261 0 : .await;
4262 :
4263 0 : Ok(())
4264 0 : }
4265 :
4266 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
4267 : ///
4268 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
4269 0 : async fn tenant_for_shards<F, R>(
4270 0 : &self,
4271 0 : locations: Vec<(TenantShardId, Node)>,
4272 0 : mut req_fn: F,
4273 0 : ) -> Result<Vec<R>, ApiError>
4274 0 : where
4275 0 : F: FnMut(
4276 0 : TenantShardId,
4277 0 : Node,
4278 0 : )
4279 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
4280 0 : {
4281 0 : let mut futs = FuturesUnordered::new();
4282 0 : let mut results = Vec::with_capacity(locations.len());
4283 :
4284 0 : for (tenant_shard_id, node) in locations {
4285 0 : futs.push(req_fn(tenant_shard_id, node));
4286 0 : }
4287 :
4288 0 : while let Some(r) = futs.next().await {
4289 0 : results.push(r?);
4290 : }
4291 :
4292 0 : Ok(results)
4293 0 : }
4294 :
4295 : /// Concurrently invoke a pageserver API call on many shards at once
4296 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
4297 0 : &self,
4298 0 : locations: Vec<(TenantShardId, Node)>,
4299 0 : op: O,
4300 0 : warn_threshold: u32,
4301 0 : max_retries: u32,
4302 0 : timeout: Duration,
4303 0 : cancel: &CancellationToken,
4304 0 : ) -> Vec<mgmt_api::Result<T>>
4305 0 : where
4306 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
4307 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
4308 0 : {
4309 0 : let mut futs = FuturesUnordered::new();
4310 0 : let mut results = Vec::with_capacity(locations.len());
4311 :
4312 0 : for (tenant_shard_id, node) in locations {
4313 0 : futs.push(async move {
4314 0 : node.with_client_retries(
4315 0 : |client| op(tenant_shard_id, client),
4316 0 : &self.config.pageserver_jwt_token,
4317 0 : &self.config.ssl_ca_cert,
4318 0 : warn_threshold,
4319 0 : max_retries,
4320 0 : timeout,
4321 0 : cancel,
4322 0 : )
4323 0 : .await
4324 0 : });
4325 0 : }
4326 :
4327 0 : while let Some(r) = futs.next().await {
4328 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
4329 0 : results.push(r);
4330 0 : }
4331 :
4332 0 : results
4333 0 : }
4334 :
4335 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
4336 : /// when creating and deleting timelines:
4337 : /// - Makes sure shards are attached somewhere if they weren't already
4338 : /// - Looks up the shards and the nodes where they were most recently attached
4339 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
4340 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
4341 0 : async fn tenant_remote_mutation<R, O, F>(
4342 0 : &self,
4343 0 : tenant_id: TenantId,
4344 0 : op: O,
4345 0 : ) -> Result<R, ApiError>
4346 0 : where
4347 0 : O: FnOnce(TenantMutationLocations) -> F,
4348 0 : F: std::future::Future<Output = R>,
4349 0 : {
4350 0 : let mutation_locations = {
4351 0 : let mut locations = TenantMutationLocations::default();
4352 :
4353 : // Load the currently attached pageservers for the latest generation of each shard. This can
4354 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
4355 : // will still be the latest when we're done: we will check generations again at the end of
4356 : // this function to handle that.
4357 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
4358 :
4359 0 : if generations
4360 0 : .iter()
4361 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
4362 : {
4363 0 : let shard_generations = generations
4364 0 : .into_iter()
4365 0 : .map(|i| (i.tenant_shard_id, (i.generation, i.generation_pageserver)))
4366 0 : .collect::<HashMap<_, _>>();
4367 0 :
4368 0 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
4369 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)
4370 0 : let locked = self.inner.read().unwrap();
4371 0 : for (shard_id, shard) in
4372 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4373 : {
4374 0 : match shard.policy {
4375 : PlacementPolicy::Attached(_) => {
4376 : // This shard is meant to be attached: the caller is not wrong to try and
4377 : // use this function, but we can't service the request right now.
4378 0 : let Some(generation) = shard_generations.get(shard_id) else {
4379 : // This can only happen if there is a split brain controller modifying the database. This should
4380 : // never happen when testing, and if it happens in production we can only log the issue.
4381 0 : debug_assert!(false);
4382 0 : tracing::error!(
4383 0 : "Shard {shard_id} not found in generation state! Is another rogue controller running?"
4384 : );
4385 0 : continue;
4386 : };
4387 0 : let (generation, generation_pageserver) = generation;
4388 0 : if let Some(generation) = generation {
4389 0 : if generation_pageserver.is_none() {
4390 : // This is legitimate only in a very narrow window where the shard was only just configured into
4391 : // Attached mode after being created in Secondary or Detached mode, and it has had its generation
4392 : // set but not yet had a Reconciler run (reconciler is the only thing that sets generation_pageserver).
4393 0 : tracing::warn!(
4394 0 : "Shard {shard_id} generation is set ({generation:?}) but generation_pageserver is None, reconciler not run yet?"
4395 : );
4396 0 : }
4397 : } else {
4398 : // This should never happen: a shard with no generation is only permitted when it was created in some state
4399 : // other than PlacementPolicy::Attached (and generation is always written to DB before setting Attached in memory)
4400 0 : debug_assert!(false);
4401 0 : tracing::error!(
4402 0 : "Shard {shard_id} generation is None, but it is in PlacementPolicy::Attached mode!"
4403 : );
4404 0 : continue;
4405 : }
4406 : }
4407 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
4408 0 : return Err(ApiError::Conflict(format!(
4409 0 : "Shard {shard_id} tenant has policy {:?}",
4410 0 : shard.policy
4411 0 : )));
4412 : }
4413 : }
4414 : }
4415 :
4416 0 : return Err(ApiError::ResourceUnavailable(
4417 0 : "One or more shards in tenant is not yet attached".into(),
4418 0 : ));
4419 0 : }
4420 0 :
4421 0 : let locked = self.inner.read().unwrap();
4422 : for ShardGenerationState {
4423 0 : tenant_shard_id,
4424 0 : generation,
4425 0 : generation_pageserver,
4426 0 : } in generations
4427 : {
4428 0 : let node_id = generation_pageserver.expect("We checked for None above");
4429 0 : let node = locked
4430 0 : .nodes
4431 0 : .get(&node_id)
4432 0 : .ok_or(ApiError::Conflict(format!(
4433 0 : "Raced with removal of node {node_id}"
4434 0 : )))?;
4435 0 : let generation = generation.expect("Checked above");
4436 0 :
4437 0 : let tenant = locked.tenants.get(&tenant_shard_id);
4438 :
4439 : // TODO(vlad): Abstract the logic that finds stale attached locations
4440 : // from observed state into a [`Service`] method.
4441 0 : let other_locations = match tenant {
4442 0 : Some(tenant) => {
4443 0 : let mut other = tenant.attached_locations();
4444 0 : let latest_location_index =
4445 0 : other.iter().position(|&l| l == (node.get_id(), generation));
4446 0 : if let Some(idx) = latest_location_index {
4447 0 : other.remove(idx);
4448 0 : }
4449 :
4450 0 : other
4451 : }
4452 0 : None => Vec::default(),
4453 : };
4454 :
4455 0 : let location = ShardMutationLocations {
4456 0 : latest: MutationLocation {
4457 0 : node: node.clone(),
4458 0 : generation,
4459 0 : },
4460 0 : other: other_locations
4461 0 : .into_iter()
4462 0 : .filter_map(|(node_id, generation)| {
4463 0 : let node = locked.nodes.get(&node_id)?;
4464 :
4465 0 : Some(MutationLocation {
4466 0 : node: node.clone(),
4467 0 : generation,
4468 0 : })
4469 0 : })
4470 0 : .collect(),
4471 0 : };
4472 0 : locations.0.insert(tenant_shard_id, location);
4473 0 : }
4474 :
4475 0 : locations
4476 : };
4477 :
4478 0 : let result = op(mutation_locations.clone()).await;
4479 :
4480 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
4481 : // our remote operation executed on the latest generation and is therefore persistent.
4482 : {
4483 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
4484 0 : if latest_generations
4485 0 : .into_iter()
4486 0 : .map(
4487 0 : |ShardGenerationState {
4488 : tenant_shard_id,
4489 : generation,
4490 : generation_pageserver: _,
4491 0 : }| (tenant_shard_id, generation),
4492 0 : )
4493 0 : .collect::<Vec<_>>()
4494 0 : != mutation_locations
4495 0 : .0
4496 0 : .into_iter()
4497 0 : .map(|i| (i.0, Some(i.1.latest.generation)))
4498 0 : .collect::<Vec<_>>()
4499 : {
4500 : // We raced with something that incremented the generation, and therefore cannot be
4501 : // confident that our actions are persistent (they might have hit an old generation).
4502 : //
4503 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
4504 0 : return Err(ApiError::ResourceUnavailable(
4505 0 : "Tenant attachment changed, please retry".into(),
4506 0 : ));
4507 0 : }
4508 0 : }
4509 0 :
4510 0 : Ok(result)
4511 0 : }
4512 :
4513 0 : pub(crate) async fn tenant_timeline_delete(
4514 0 : self: &Arc<Self>,
4515 0 : tenant_id: TenantId,
4516 0 : timeline_id: TimelineId,
4517 0 : ) -> Result<StatusCode, ApiError> {
4518 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
4519 0 : let _tenant_lock = trace_shared_lock(
4520 0 : &self.tenant_op_locks,
4521 0 : tenant_id,
4522 0 : TenantOperations::TimelineDelete,
4523 0 : )
4524 0 : .await;
4525 :
4526 0 : let status_code = self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
4527 0 : if targets.0.is_empty() {
4528 0 : return Err(ApiError::NotFound(
4529 0 : anyhow::anyhow!("Tenant not found").into(),
4530 0 : ));
4531 0 : }
4532 0 :
4533 0 : let (shard_zero_tid, shard_zero_locations) = targets.0.pop_first().expect("Must have at least one shard");
4534 0 : assert!(shard_zero_tid.is_shard_zero());
4535 :
4536 0 : async fn delete_one(
4537 0 : tenant_shard_id: TenantShardId,
4538 0 : timeline_id: TimelineId,
4539 0 : node: Node,
4540 0 : jwt: Option<String>,
4541 0 : ssl_ca_cert: Option<Certificate>,
4542 0 : ) -> Result<StatusCode, ApiError> {
4543 0 : tracing::info!(
4544 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
4545 : );
4546 :
4547 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref(), ssl_ca_cert)
4548 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4549 0 : let res = client
4550 0 : .timeline_delete(tenant_shard_id, timeline_id)
4551 0 : .await;
4552 :
4553 0 : match res {
4554 0 : Ok(ok) => Ok(ok),
4555 0 : Err(mgmt_api::Error::ApiError(StatusCode::CONFLICT, _)) => Ok(StatusCode::CONFLICT),
4556 0 : Err(mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg)) => Err(ApiError::ResourceUnavailable(msg.into())),
4557 0 : Err(e) => {
4558 0 : Err(
4559 0 : ApiError::InternalServerError(anyhow::anyhow!(
4560 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
4561 0 : ))
4562 0 : )
4563 : }
4564 : }
4565 0 : }
4566 :
4567 0 : let locations = targets.0.iter().map(|t| (*t.0, t.1.latest.node.clone())).collect();
4568 0 : let statuses = self
4569 0 : .tenant_for_shards(locations, |tenant_shard_id: TenantShardId, node: Node| {
4570 0 : Box::pin(delete_one(
4571 0 : tenant_shard_id,
4572 0 : timeline_id,
4573 0 : node,
4574 0 : self.config.pageserver_jwt_token.clone(),
4575 0 : self.config.ssl_ca_cert.clone(),
4576 0 : ))
4577 0 : })
4578 0 : .await?;
4579 :
4580 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero.
4581 : // We return 409 (Conflict) if deletion was already in progress on any of the shards
4582 : // and 202 (Accepted) if deletion was not already in progress on any of the shards.
4583 0 : if statuses.iter().any(|s| s == &StatusCode::CONFLICT) {
4584 0 : return Ok(StatusCode::CONFLICT);
4585 0 : }
4586 0 :
4587 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
4588 0 : return Ok(StatusCode::ACCEPTED);
4589 0 : }
4590 :
4591 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
4592 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
4593 0 : let shard_zero_status = delete_one(
4594 0 : shard_zero_tid,
4595 0 : timeline_id,
4596 0 : shard_zero_locations.latest.node,
4597 0 : self.config.pageserver_jwt_token.clone(),
4598 0 : self.config.ssl_ca_cert.clone(),
4599 0 : )
4600 0 : .await?;
4601 0 : Ok(shard_zero_status)
4602 0 : }).await?;
4603 :
4604 0 : self.tenant_timeline_delete_safekeepers(tenant_id, timeline_id)
4605 0 : .await?;
4606 :
4607 0 : status_code
4608 0 : }
4609 : /// Perform timeline deletion on safekeepers. Will return success: we persist the deletion into the reconciler.
4610 0 : async fn tenant_timeline_delete_safekeepers(
4611 0 : self: &Arc<Self>,
4612 0 : tenant_id: TenantId,
4613 0 : timeline_id: TimelineId,
4614 0 : ) -> Result<(), ApiError> {
4615 0 : let tl = self
4616 0 : .persistence
4617 0 : .get_timeline(tenant_id, timeline_id)
4618 0 : .await?;
4619 0 : let Some(tl) = tl else {
4620 0 : tracing::info!(
4621 0 : "timeline {tenant_id}/{timeline_id} doesn't exist in timelines table, no deletions on safekeepers needed"
4622 : );
4623 0 : return Ok(());
4624 : };
4625 0 : let all_sks = tl
4626 0 : .new_sk_set
4627 0 : .iter()
4628 0 : .flat_map(|sks| {
4629 0 : sks.iter()
4630 0 : .map(|sk| (*sk, SafekeeperTimelineOpKind::Exclude))
4631 0 : })
4632 0 : .chain(
4633 0 : tl.sk_set
4634 0 : .iter()
4635 0 : .map(|v| (*v, SafekeeperTimelineOpKind::Delete)),
4636 0 : )
4637 0 : .collect::<HashMap<_, _>>();
4638 0 :
4639 0 : // Schedule reconciliations
4640 0 : {
4641 0 : let mut locked = self.inner.write().unwrap();
4642 0 : for (sk_id, kind) in all_sks {
4643 0 : let sk_id = NodeId(sk_id as u64);
4644 0 : let Some(sk) = locked.safekeepers.get(&sk_id) else {
4645 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4646 0 : "Couldn't find safekeeper with id {sk_id}"
4647 0 : )));
4648 : };
4649 :
4650 0 : let req = ScheduleRequest {
4651 0 : safekeeper: Box::new(sk.clone()),
4652 0 : // we don't use this for this kind, put a dummy value
4653 0 : host_list: Vec::new(),
4654 0 : tenant_id,
4655 0 : timeline_id,
4656 0 : generation: tl.generation as u32,
4657 0 : kind,
4658 0 : };
4659 0 : locked.safekeeper_reconcilers.schedule_request(self, req);
4660 : }
4661 : }
4662 0 : Ok(())
4663 0 : }
4664 :
4665 : /// When you know the TenantId but not a specific shard, and would like to get the node holding shard 0.
4666 0 : pub(crate) async fn tenant_shard0_node(
4667 0 : &self,
4668 0 : tenant_id: TenantId,
4669 0 : ) -> Result<(Node, TenantShardId), ApiError> {
4670 0 : let tenant_shard_id = {
4671 0 : let locked = self.inner.read().unwrap();
4672 0 : let Some((tenant_shard_id, _shard)) = locked
4673 0 : .tenants
4674 0 : .range(TenantShardId::tenant_range(tenant_id))
4675 0 : .next()
4676 : else {
4677 0 : return Err(ApiError::NotFound(
4678 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
4679 0 : ));
4680 : };
4681 :
4682 0 : *tenant_shard_id
4683 0 : };
4684 0 :
4685 0 : self.tenant_shard_node(tenant_shard_id)
4686 0 : .await
4687 0 : .map(|node| (node, tenant_shard_id))
4688 0 : }
4689 :
4690 : /// When you need to send an HTTP request to the pageserver that holds a shard of a tenant, this
4691 : /// function looks up and returns node. If the shard isn't found, returns Err(ApiError::NotFound)
4692 0 : pub(crate) async fn tenant_shard_node(
4693 0 : &self,
4694 0 : tenant_shard_id: TenantShardId,
4695 0 : ) -> Result<Node, ApiError> {
4696 0 : // Look up in-memory state and maybe use the node from there.
4697 0 : {
4698 0 : let locked = self.inner.read().unwrap();
4699 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
4700 0 : return Err(ApiError::NotFound(
4701 0 : anyhow::anyhow!("Tenant shard {tenant_shard_id} not found").into(),
4702 0 : ));
4703 : };
4704 :
4705 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
4706 0 : tracing::warn!(
4707 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
4708 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
4709 : shard.policy
4710 : );
4711 0 : return Err(ApiError::Conflict(
4712 0 : "Cannot call timeline API on non-attached tenant".to_string(),
4713 0 : ));
4714 : };
4715 :
4716 0 : if shard.reconciler.is_none() {
4717 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
4718 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
4719 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
4720 : // This should never happen
4721 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4722 0 : "Shard refers to nonexistent node"
4723 0 : )));
4724 : };
4725 0 : return Ok(node.clone());
4726 0 : }
4727 : };
4728 :
4729 : // Look up the latest attached pageserver location from the database
4730 : // generation state: this will reflect the progress of any ongoing migration.
4731 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
4732 : // the case where they call through to the pageserver and get a 404.
4733 0 : let db_result = self
4734 0 : .persistence
4735 0 : .tenant_generations(tenant_shard_id.tenant_id)
4736 0 : .await?;
4737 : let Some(ShardGenerationState {
4738 : tenant_shard_id: _,
4739 : generation: _,
4740 0 : generation_pageserver: Some(node_id),
4741 0 : }) = db_result
4742 0 : .into_iter()
4743 0 : .find(|s| s.tenant_shard_id == tenant_shard_id)
4744 : else {
4745 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
4746 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
4747 : // or a conflict response (case where tenant was detached in background)
4748 0 : return Err(ApiError::ResourceUnavailable(
4749 0 : format!("Shard {tenant_shard_id} not found in database, or is not attached").into(),
4750 0 : ));
4751 : };
4752 0 : let locked = self.inner.read().unwrap();
4753 0 : let Some(node) = locked.nodes.get(&node_id) else {
4754 : // This should never happen
4755 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4756 0 : "Shard refers to nonexistent node"
4757 0 : )));
4758 : };
4759 :
4760 0 : Ok(node.clone())
4761 0 : }
4762 :
4763 0 : pub(crate) fn tenant_locate(
4764 0 : &self,
4765 0 : tenant_id: TenantId,
4766 0 : ) -> Result<TenantLocateResponse, ApiError> {
4767 0 : let locked = self.inner.read().unwrap();
4768 0 : tracing::info!("Locating shards for tenant {tenant_id}");
4769 :
4770 0 : let mut result = Vec::new();
4771 0 : let mut shard_params: Option<ShardParameters> = None;
4772 :
4773 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4774 : {
4775 0 : let node_id =
4776 0 : shard
4777 0 : .intent
4778 0 : .get_attached()
4779 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
4780 0 : "Cannot locate a tenant that is not attached"
4781 0 : )))?;
4782 :
4783 0 : let node = locked
4784 0 : .nodes
4785 0 : .get(&node_id)
4786 0 : .expect("Pageservers may not be deleted while referenced");
4787 0 :
4788 0 : result.push(node.shard_location(*tenant_shard_id));
4789 0 :
4790 0 : match &shard_params {
4791 0 : None => {
4792 0 : shard_params = Some(ShardParameters {
4793 0 : stripe_size: shard.shard.stripe_size,
4794 0 : count: shard.shard.count,
4795 0 : });
4796 0 : }
4797 0 : Some(params) => {
4798 0 : if params.stripe_size != shard.shard.stripe_size {
4799 : // This should never happen. We enforce at runtime because it's simpler than
4800 : // adding an extra per-tenant data structure to store the things that should be the same
4801 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4802 0 : "Inconsistent shard stripe size parameters!"
4803 0 : )));
4804 0 : }
4805 : }
4806 : }
4807 : }
4808 :
4809 0 : if result.is_empty() {
4810 0 : return Err(ApiError::NotFound(
4811 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
4812 0 : ));
4813 0 : }
4814 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
4815 0 : tracing::info!(
4816 0 : "Located tenant {} with params {:?} on shards {}",
4817 0 : tenant_id,
4818 0 : shard_params,
4819 0 : result
4820 0 : .iter()
4821 0 : .map(|s| format!("{:?}", s))
4822 0 : .collect::<Vec<_>>()
4823 0 : .join(",")
4824 : );
4825 :
4826 0 : Ok(TenantLocateResponse {
4827 0 : shards: result,
4828 0 : shard_params,
4829 0 : })
4830 0 : }
4831 :
4832 : /// Returns None if the input iterator of shards does not include a shard with number=0
4833 0 : fn tenant_describe_impl<'a>(
4834 0 : &self,
4835 0 : shards: impl Iterator<Item = &'a TenantShard>,
4836 0 : ) -> Option<TenantDescribeResponse> {
4837 0 : let mut shard_zero = None;
4838 0 : let mut describe_shards = Vec::new();
4839 :
4840 0 : for shard in shards {
4841 0 : if shard.tenant_shard_id.is_shard_zero() {
4842 0 : shard_zero = Some(shard);
4843 0 : }
4844 :
4845 0 : describe_shards.push(TenantDescribeResponseShard {
4846 0 : tenant_shard_id: shard.tenant_shard_id,
4847 0 : node_attached: *shard.intent.get_attached(),
4848 0 : node_secondary: shard.intent.get_secondary().to_vec(),
4849 0 : last_error: shard
4850 0 : .last_error
4851 0 : .lock()
4852 0 : .unwrap()
4853 0 : .as_ref()
4854 0 : .map(|e| format!("{e}"))
4855 0 : .unwrap_or("".to_string())
4856 0 : .clone(),
4857 0 : is_reconciling: shard.reconciler.is_some(),
4858 0 : is_pending_compute_notification: shard.pending_compute_notification,
4859 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
4860 0 : scheduling_policy: shard.get_scheduling_policy(),
4861 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
4862 : })
4863 : }
4864 :
4865 0 : let shard_zero = shard_zero?;
4866 :
4867 0 : Some(TenantDescribeResponse {
4868 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
4869 0 : shards: describe_shards,
4870 0 : stripe_size: shard_zero.shard.stripe_size,
4871 0 : policy: shard_zero.policy.clone(),
4872 0 : config: shard_zero.config.clone(),
4873 0 : })
4874 0 : }
4875 :
4876 0 : pub(crate) fn tenant_describe(
4877 0 : &self,
4878 0 : tenant_id: TenantId,
4879 0 : ) -> Result<TenantDescribeResponse, ApiError> {
4880 0 : let locked = self.inner.read().unwrap();
4881 0 :
4882 0 : self.tenant_describe_impl(
4883 0 : locked
4884 0 : .tenants
4885 0 : .range(TenantShardId::tenant_range(tenant_id))
4886 0 : .map(|(_k, v)| v),
4887 0 : )
4888 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
4889 0 : }
4890 :
4891 : /// limit & offset are pagination parameters. Since we are walking an in-memory HashMap, `offset` does not
4892 : /// avoid traversing data, it just avoid returning it. This is suitable for our purposes, since our in memory
4893 : /// maps are small enough to traverse fast, our pagination is just to avoid serializing huge JSON responses
4894 : /// in our external API.
4895 0 : pub(crate) fn tenant_list(
4896 0 : &self,
4897 0 : limit: Option<usize>,
4898 0 : start_after: Option<TenantId>,
4899 0 : ) -> Vec<TenantDescribeResponse> {
4900 0 : let locked = self.inner.read().unwrap();
4901 :
4902 : // Apply start_from parameter
4903 0 : let shard_range = match start_after {
4904 0 : None => locked.tenants.range(..),
4905 0 : Some(tenant_id) => locked.tenants.range(
4906 0 : TenantShardId {
4907 0 : tenant_id,
4908 0 : shard_number: ShardNumber(u8::MAX),
4909 0 : shard_count: ShardCount(u8::MAX),
4910 0 : }..,
4911 0 : ),
4912 : };
4913 :
4914 0 : let mut result = Vec::new();
4915 0 : for (_tenant_id, tenant_shards) in &shard_range.group_by(|(id, _shard)| id.tenant_id) {
4916 0 : result.push(
4917 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
4918 0 : .expect("Groups are always non-empty"),
4919 0 : );
4920 :
4921 : // Enforce `limit` parameter
4922 0 : if let Some(limit) = limit {
4923 0 : if result.len() >= limit {
4924 0 : break;
4925 0 : }
4926 0 : }
4927 : }
4928 :
4929 0 : result
4930 0 : }
4931 :
4932 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
4933 : async fn abort_tenant_shard_split(
4934 : &self,
4935 : op: &TenantShardSplitAbort,
4936 : ) -> Result<(), TenantShardSplitAbortError> {
4937 : // Cleaning up a split:
4938 : // - Parent shards are not destroyed during a split, just detached.
4939 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
4940 : // just the children attached, or both.
4941 : //
4942 : // Therefore our work to do is to:
4943 : // 1. Clean up storage controller's internal state to just refer to parents, no children
4944 : // 2. Call out to pageservers to ensure that children are detached
4945 : // 3. Call out to pageservers to ensure that parents are attached.
4946 : //
4947 : // Crash safety:
4948 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
4949 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
4950 : // and detach them.
4951 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
4952 : // from our database, then we will re-enter this cleanup routine on startup.
4953 :
4954 : let TenantShardSplitAbort {
4955 : tenant_id,
4956 : new_shard_count,
4957 : new_stripe_size,
4958 : ..
4959 : } = op;
4960 :
4961 : // First abort persistent state, if any exists.
4962 : match self
4963 : .persistence
4964 : .abort_shard_split(*tenant_id, *new_shard_count)
4965 : .await?
4966 : {
4967 : AbortShardSplitStatus::Aborted => {
4968 : // Proceed to roll back any child shards created on pageservers
4969 : }
4970 : AbortShardSplitStatus::Complete => {
4971 : // The split completed (we might hit that path if e.g. our database transaction
4972 : // to write the completion landed in the database, but we dropped connection
4973 : // before seeing the result).
4974 : //
4975 : // We must update in-memory state to reflect the successful split.
4976 : self.tenant_shard_split_commit_inmem(
4977 : *tenant_id,
4978 : *new_shard_count,
4979 : *new_stripe_size,
4980 : );
4981 : return Ok(());
4982 : }
4983 : }
4984 :
4985 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
4986 : let detach_locations: Vec<(Node, TenantShardId)> = {
4987 : let mut detach_locations = Vec::new();
4988 : let mut locked = self.inner.write().unwrap();
4989 : let (nodes, tenants, scheduler) = locked.parts_mut();
4990 :
4991 : for (tenant_shard_id, shard) in
4992 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
4993 : {
4994 : if shard.shard.count == op.new_shard_count {
4995 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
4996 : // is infallible, so if we got an error we shouldn't have got that far.
4997 : tracing::warn!(
4998 : "During split abort, child shard {tenant_shard_id} found in-memory"
4999 : );
5000 : continue;
5001 : }
5002 :
5003 : // Add the children of this shard to this list of things to detach
5004 : if let Some(node_id) = shard.intent.get_attached() {
5005 : for child_id in tenant_shard_id.split(*new_shard_count) {
5006 : detach_locations.push((
5007 : nodes
5008 : .get(node_id)
5009 : .expect("Intent references nonexistent node")
5010 : .clone(),
5011 : child_id,
5012 : ));
5013 : }
5014 : } else {
5015 : tracing::warn!(
5016 : "During split abort, shard {tenant_shard_id} has no attached location"
5017 : );
5018 : }
5019 :
5020 : tracing::info!("Restoring parent shard {tenant_shard_id}");
5021 :
5022 : // Drop any intents that refer to unavailable nodes, to enable this abort to proceed even
5023 : // if the original attachment location is offline.
5024 : if let Some(node_id) = shard.intent.get_attached() {
5025 : if !nodes.get(node_id).unwrap().is_available() {
5026 : tracing::info!(
5027 : "Demoting attached intent for {tenant_shard_id} on unavailable node {node_id}"
5028 : );
5029 : shard.intent.demote_attached(scheduler, *node_id);
5030 : }
5031 : }
5032 : for node_id in shard.intent.get_secondary().clone() {
5033 : if !nodes.get(&node_id).unwrap().is_available() {
5034 : tracing::info!(
5035 : "Dropping secondary intent for {tenant_shard_id} on unavailable node {node_id}"
5036 : );
5037 : shard.intent.remove_secondary(scheduler, node_id);
5038 : }
5039 : }
5040 :
5041 : shard.splitting = SplitState::Idle;
5042 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
5043 : // If this shard can't be scheduled now (perhaps due to offline nodes or
5044 : // capacity issues), that must not prevent us rolling back a split. In this
5045 : // case it should be eventually scheduled in the background.
5046 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
5047 : }
5048 :
5049 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High);
5050 : }
5051 :
5052 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
5053 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
5054 :
5055 : detach_locations
5056 : };
5057 :
5058 : for (node, child_id) in detach_locations {
5059 : if !node.is_available() {
5060 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
5061 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
5062 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
5063 : // them from the node.
5064 : tracing::warn!(
5065 : "Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated."
5066 : );
5067 : continue;
5068 : }
5069 :
5070 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
5071 : // a 503 and retry, up to our limit.
5072 : tracing::info!("Detaching {child_id} on {node}...");
5073 : match node
5074 : .with_client_retries(
5075 0 : |client| async move {
5076 0 : let config = LocationConfig {
5077 0 : mode: LocationConfigMode::Detached,
5078 0 : generation: None,
5079 0 : secondary_conf: None,
5080 0 : shard_number: child_id.shard_number.0,
5081 0 : shard_count: child_id.shard_count.literal(),
5082 0 : // Stripe size and tenant config don't matter when detaching
5083 0 : shard_stripe_size: 0,
5084 0 : tenant_conf: TenantConfig::default(),
5085 0 : };
5086 0 :
5087 0 : client.location_config(child_id, config, None, false).await
5088 0 : },
5089 : &self.config.pageserver_jwt_token,
5090 : &self.config.ssl_ca_cert,
5091 : 1,
5092 : 10,
5093 : Duration::from_secs(5),
5094 : &self.cancel,
5095 : )
5096 : .await
5097 : {
5098 : Some(Ok(_)) => {}
5099 : Some(Err(e)) => {
5100 : // We failed to communicate with the remote node. This is problematic: we may be
5101 : // leaving it with a rogue child shard.
5102 : tracing::warn!(
5103 : "Failed to detach child {child_id} from node {node} during abort"
5104 : );
5105 : return Err(e.into());
5106 : }
5107 : None => {
5108 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
5109 : // clean up on restart. The node going offline requires a retry.
5110 : return Err(TenantShardSplitAbortError::Unavailable);
5111 : }
5112 : };
5113 : }
5114 :
5115 : tracing::info!("Successfully aborted split");
5116 : Ok(())
5117 : }
5118 :
5119 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
5120 : /// of the tenant map to reflect the child shards that exist after the split.
5121 0 : fn tenant_shard_split_commit_inmem(
5122 0 : &self,
5123 0 : tenant_id: TenantId,
5124 0 : new_shard_count: ShardCount,
5125 0 : new_stripe_size: Option<ShardStripeSize>,
5126 0 : ) -> (
5127 0 : TenantShardSplitResponse,
5128 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
5129 0 : Vec<ReconcilerWaiter>,
5130 0 : ) {
5131 0 : let mut response = TenantShardSplitResponse {
5132 0 : new_shards: Vec::new(),
5133 0 : };
5134 0 : let mut child_locations = Vec::new();
5135 0 : let mut waiters = Vec::new();
5136 0 :
5137 0 : {
5138 0 : let mut locked = self.inner.write().unwrap();
5139 0 :
5140 0 : let parent_ids = locked
5141 0 : .tenants
5142 0 : .range(TenantShardId::tenant_range(tenant_id))
5143 0 : .map(|(shard_id, _)| *shard_id)
5144 0 : .collect::<Vec<_>>();
5145 0 :
5146 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5147 0 : for parent_id in parent_ids {
5148 0 : let child_ids = parent_id.split(new_shard_count);
5149 :
5150 0 : let (pageserver, generation, policy, parent_ident, config, preferred_az) = {
5151 0 : let mut old_state = tenants
5152 0 : .remove(&parent_id)
5153 0 : .expect("It was present, we just split it");
5154 0 :
5155 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
5156 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
5157 0 : // nothing else can clear this.
5158 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
5159 :
5160 0 : let old_attached = old_state.intent.get_attached().unwrap();
5161 0 : old_state.intent.clear(scheduler);
5162 0 : let generation = old_state.generation.expect("Shard must have been attached");
5163 0 : (
5164 0 : old_attached,
5165 0 : generation,
5166 0 : old_state.policy.clone(),
5167 0 : old_state.shard,
5168 0 : old_state.config.clone(),
5169 0 : old_state.preferred_az().cloned(),
5170 0 : )
5171 0 : };
5172 0 :
5173 0 : let mut schedule_context = ScheduleContext::default();
5174 0 : for child in child_ids {
5175 0 : let mut child_shard = parent_ident;
5176 0 : child_shard.number = child.shard_number;
5177 0 : child_shard.count = child.shard_count;
5178 0 : if let Some(stripe_size) = new_stripe_size {
5179 0 : child_shard.stripe_size = stripe_size;
5180 0 : }
5181 :
5182 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
5183 0 : child_observed.insert(
5184 0 : pageserver,
5185 0 : ObservedStateLocation {
5186 0 : conf: Some(attached_location_conf(
5187 0 : generation,
5188 0 : &child_shard,
5189 0 : &config,
5190 0 : &policy,
5191 0 : )),
5192 0 : },
5193 0 : );
5194 0 :
5195 0 : let mut child_state =
5196 0 : TenantShard::new(child, child_shard, policy.clone(), preferred_az.clone());
5197 0 : child_state.intent =
5198 0 : IntentState::single(scheduler, Some(pageserver), preferred_az.clone());
5199 0 : child_state.observed = ObservedState {
5200 0 : locations: child_observed,
5201 0 : };
5202 0 : child_state.generation = Some(generation);
5203 0 : child_state.config = config.clone();
5204 0 :
5205 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
5206 0 : // as at this point in the split process we have succeeded and this part is infallible:
5207 0 : // we will never need to do any special recovery from this state.
5208 0 :
5209 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
5210 :
5211 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
5212 : // This is not fatal, because we've implicitly already got an attached
5213 : // location for the child shard. Failure here just means we couldn't
5214 : // find a secondary (e.g. because cluster is overloaded).
5215 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
5216 0 : }
5217 : // In the background, attach secondary locations for the new shards
5218 0 : if let Some(waiter) = self.maybe_reconcile_shard(
5219 0 : &mut child_state,
5220 0 : nodes,
5221 0 : ReconcilerPriority::High,
5222 0 : ) {
5223 0 : waiters.push(waiter);
5224 0 : }
5225 :
5226 0 : tenants.insert(child, child_state);
5227 0 : response.new_shards.push(child);
5228 : }
5229 : }
5230 0 : (response, child_locations, waiters)
5231 0 : }
5232 0 : }
5233 :
5234 0 : async fn tenant_shard_split_start_secondaries(
5235 0 : &self,
5236 0 : tenant_id: TenantId,
5237 0 : waiters: Vec<ReconcilerWaiter>,
5238 0 : ) {
5239 : // Wait for initial reconcile of child shards, this creates the secondary locations
5240 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
5241 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
5242 : // their secondaries couldn't be attached.
5243 0 : tracing::warn!("Failed to reconcile after split: {e}");
5244 0 : return;
5245 0 : }
5246 :
5247 : // Take the state lock to discover the attached & secondary intents for all shards
5248 0 : let (attached, secondary) = {
5249 0 : let locked = self.inner.read().unwrap();
5250 0 : let mut attached = Vec::new();
5251 0 : let mut secondary = Vec::new();
5252 :
5253 0 : for (tenant_shard_id, shard) in
5254 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5255 : {
5256 0 : let Some(node_id) = shard.intent.get_attached() else {
5257 : // Unexpected. Race with a PlacementPolicy change?
5258 0 : tracing::warn!(
5259 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
5260 : );
5261 0 : continue;
5262 : };
5263 :
5264 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
5265 : // No secondary location. Nothing for us to do.
5266 0 : continue;
5267 : };
5268 :
5269 0 : let attached_node = locked
5270 0 : .nodes
5271 0 : .get(node_id)
5272 0 : .expect("Pageservers may not be deleted while referenced");
5273 0 :
5274 0 : let secondary_node = locked
5275 0 : .nodes
5276 0 : .get(secondary_node_id)
5277 0 : .expect("Pageservers may not be deleted while referenced");
5278 0 :
5279 0 : attached.push((*tenant_shard_id, attached_node.clone()));
5280 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
5281 : }
5282 0 : (attached, secondary)
5283 0 : };
5284 0 :
5285 0 : if secondary.is_empty() {
5286 : // No secondary locations; nothing for us to do
5287 0 : return;
5288 0 : }
5289 :
5290 0 : for result in self
5291 0 : .tenant_for_shards_api(
5292 0 : attached,
5293 0 : |tenant_shard_id, client| async move {
5294 0 : client.tenant_heatmap_upload(tenant_shard_id).await
5295 0 : },
5296 0 : 1,
5297 0 : 1,
5298 0 : SHORT_RECONCILE_TIMEOUT,
5299 0 : &self.cancel,
5300 0 : )
5301 0 : .await
5302 : {
5303 0 : if let Err(e) = result {
5304 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
5305 0 : return;
5306 0 : }
5307 : }
5308 :
5309 0 : for result in self
5310 0 : .tenant_for_shards_api(
5311 0 : secondary,
5312 0 : |tenant_shard_id, client| async move {
5313 0 : client
5314 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
5315 0 : .await
5316 0 : },
5317 0 : 1,
5318 0 : 1,
5319 0 : SHORT_RECONCILE_TIMEOUT,
5320 0 : &self.cancel,
5321 0 : )
5322 0 : .await
5323 : {
5324 0 : if let Err(e) = result {
5325 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
5326 0 : return;
5327 0 : }
5328 : }
5329 0 : }
5330 :
5331 0 : pub(crate) async fn tenant_shard_split(
5332 0 : &self,
5333 0 : tenant_id: TenantId,
5334 0 : split_req: TenantShardSplitRequest,
5335 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
5336 : // TODO: return 503 if we get stuck waiting for this lock
5337 : // (issue https://github.com/neondatabase/neon/issues/7108)
5338 0 : let _tenant_lock = trace_exclusive_lock(
5339 0 : &self.tenant_op_locks,
5340 0 : tenant_id,
5341 0 : TenantOperations::ShardSplit,
5342 0 : )
5343 0 : .await;
5344 :
5345 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
5346 0 : let new_stripe_size = split_req.new_stripe_size;
5347 :
5348 : // Validate the request and construct parameters. This phase is fallible, but does not require
5349 : // rollback on errors, as it does no I/O and mutates no state.
5350 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
5351 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
5352 0 : ShardSplitAction::Split(params) => params,
5353 : };
5354 :
5355 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
5356 : // we must roll back.
5357 0 : let r = self
5358 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
5359 0 : .await;
5360 :
5361 0 : let (response, waiters) = match r {
5362 0 : Ok(r) => r,
5363 0 : Err(e) => {
5364 0 : // Split might be part-done, we must do work to abort it.
5365 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
5366 0 : self.abort_tx
5367 0 : .send(TenantShardSplitAbort {
5368 0 : tenant_id,
5369 0 : new_shard_count,
5370 0 : new_stripe_size,
5371 0 : _tenant_lock,
5372 0 : })
5373 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
5374 0 : .ok();
5375 0 : return Err(e);
5376 : }
5377 : };
5378 :
5379 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
5380 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
5381 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
5382 : // in [`Self::optimize_all`]
5383 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
5384 0 : .await;
5385 0 : Ok(response)
5386 0 : }
5387 :
5388 0 : fn prepare_tenant_shard_split(
5389 0 : &self,
5390 0 : tenant_id: TenantId,
5391 0 : split_req: TenantShardSplitRequest,
5392 0 : ) -> Result<ShardSplitAction, ApiError> {
5393 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
5394 0 : anyhow::anyhow!("failpoint")
5395 0 : )));
5396 :
5397 0 : let mut policy = None;
5398 0 : let mut config = None;
5399 0 : let mut shard_ident = None;
5400 0 : let mut preferred_az_id = None;
5401 : // Validate input, and calculate which shards we will create
5402 0 : let (old_shard_count, targets) =
5403 : {
5404 0 : let locked = self.inner.read().unwrap();
5405 0 :
5406 0 : let pageservers = locked.nodes.clone();
5407 0 :
5408 0 : let mut targets = Vec::new();
5409 0 :
5410 0 : // In case this is a retry, count how many already-split shards we found
5411 0 : let mut children_found = Vec::new();
5412 0 : let mut old_shard_count = None;
5413 :
5414 0 : for (tenant_shard_id, shard) in
5415 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
5416 : {
5417 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
5418 : Ordering::Equal => {
5419 : // Already split this
5420 0 : children_found.push(*tenant_shard_id);
5421 0 : continue;
5422 : }
5423 : Ordering::Greater => {
5424 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5425 0 : "Requested count {} but already have shards at count {}",
5426 0 : split_req.new_shard_count,
5427 0 : shard.shard.count.count()
5428 0 : )));
5429 : }
5430 0 : Ordering::Less => {
5431 0 : // Fall through: this shard has lower count than requested,
5432 0 : // is a candidate for splitting.
5433 0 : }
5434 0 : }
5435 0 :
5436 0 : match old_shard_count {
5437 0 : None => old_shard_count = Some(shard.shard.count),
5438 0 : Some(old_shard_count) => {
5439 0 : if old_shard_count != shard.shard.count {
5440 : // We may hit this case if a caller asked for two splits to
5441 : // different sizes, before the first one is complete.
5442 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
5443 : // of shard_count=1 and shard_count=2 shards in the map.
5444 0 : return Err(ApiError::Conflict(
5445 0 : "Cannot split, currently mid-split".to_string(),
5446 0 : ));
5447 0 : }
5448 : }
5449 : }
5450 0 : if policy.is_none() {
5451 0 : policy = Some(shard.policy.clone());
5452 0 : }
5453 0 : if shard_ident.is_none() {
5454 0 : shard_ident = Some(shard.shard);
5455 0 : }
5456 0 : if config.is_none() {
5457 0 : config = Some(shard.config.clone());
5458 0 : }
5459 0 : if preferred_az_id.is_none() {
5460 0 : preferred_az_id = shard.preferred_az().cloned();
5461 0 : }
5462 :
5463 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
5464 0 : tracing::info!(
5465 0 : "Tenant shard {} already has shard count {}",
5466 : tenant_shard_id,
5467 : split_req.new_shard_count
5468 : );
5469 0 : continue;
5470 0 : }
5471 :
5472 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
5473 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
5474 0 : ))?;
5475 :
5476 0 : let node = pageservers
5477 0 : .get(&node_id)
5478 0 : .expect("Pageservers may not be deleted while referenced");
5479 0 :
5480 0 : targets.push(ShardSplitTarget {
5481 0 : parent_id: *tenant_shard_id,
5482 0 : node: node.clone(),
5483 0 : child_ids: tenant_shard_id
5484 0 : .split(ShardCount::new(split_req.new_shard_count)),
5485 0 : });
5486 : }
5487 :
5488 0 : if targets.is_empty() {
5489 0 : if children_found.len() == split_req.new_shard_count as usize {
5490 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
5491 0 : new_shards: children_found,
5492 0 : }));
5493 : } else {
5494 : // No shards found to split, and no existing children found: the
5495 : // tenant doesn't exist at all.
5496 0 : return Err(ApiError::NotFound(
5497 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
5498 0 : ));
5499 : }
5500 0 : }
5501 0 :
5502 0 : (old_shard_count, targets)
5503 0 : };
5504 0 :
5505 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
5506 0 : let old_shard_count = old_shard_count.unwrap();
5507 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
5508 : // This ShardIdentity will be used as the template for all children, so this implicitly
5509 : // applies the new stripe size to the children.
5510 0 : let mut shard_ident = shard_ident.unwrap();
5511 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
5512 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5513 0 : "Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards",
5514 0 : shard_ident.stripe_size
5515 0 : )));
5516 0 : }
5517 0 :
5518 0 : shard_ident.stripe_size = new_stripe_size;
5519 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
5520 0 : shard_ident
5521 : } else {
5522 0 : shard_ident.unwrap()
5523 : };
5524 0 : let policy = policy.unwrap();
5525 0 : let config = config.unwrap();
5526 0 :
5527 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
5528 0 : old_shard_count,
5529 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
5530 0 : new_stripe_size: split_req.new_stripe_size,
5531 0 : targets,
5532 0 : policy,
5533 0 : config,
5534 0 : shard_ident,
5535 0 : preferred_az_id,
5536 0 : })))
5537 0 : }
5538 :
5539 0 : async fn do_tenant_shard_split(
5540 0 : &self,
5541 0 : tenant_id: TenantId,
5542 0 : params: Box<ShardSplitParams>,
5543 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
5544 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
5545 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
5546 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
5547 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
5548 0 : // (https://github.com/neondatabase/neon/issues/6676)
5549 0 :
5550 0 : let ShardSplitParams {
5551 0 : old_shard_count,
5552 0 : new_shard_count,
5553 0 : new_stripe_size,
5554 0 : mut targets,
5555 0 : policy,
5556 0 : config,
5557 0 : shard_ident,
5558 0 : preferred_az_id,
5559 0 : } = *params;
5560 :
5561 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
5562 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
5563 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
5564 : // at the time of split.
5565 0 : let waiters = {
5566 0 : let mut locked = self.inner.write().unwrap();
5567 0 : let mut waiters = Vec::new();
5568 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5569 0 : for target in &mut targets {
5570 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
5571 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5572 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5573 0 : "Shard {} not found",
5574 0 : target.parent_id
5575 0 : )));
5576 : };
5577 :
5578 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
5579 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
5580 0 : return Err(ApiError::Conflict(format!(
5581 0 : "Shard {} unexpectedly rescheduled during split",
5582 0 : target.parent_id
5583 0 : )));
5584 0 : }
5585 0 :
5586 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
5587 0 : shard.intent.clear_secondary(scheduler);
5588 :
5589 : // Run Reconciler to execute detach fo secondary locations.
5590 0 : if let Some(waiter) =
5591 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
5592 0 : {
5593 0 : waiters.push(waiter);
5594 0 : }
5595 : }
5596 0 : waiters
5597 0 : };
5598 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
5599 :
5600 : // Before creating any new child shards in memory or on the pageservers, persist them: this
5601 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
5602 : // acts as the protection against two concurrent attempts to split: one of them will get a database
5603 : // error trying to insert the child shards.
5604 0 : let mut child_tsps = Vec::new();
5605 0 : for target in &targets {
5606 0 : let mut this_child_tsps = Vec::new();
5607 0 : for child in &target.child_ids {
5608 0 : let mut child_shard = shard_ident;
5609 0 : child_shard.number = child.shard_number;
5610 0 : child_shard.count = child.shard_count;
5611 0 :
5612 0 : tracing::info!(
5613 0 : "Create child shard persistence with stripe size {}",
5614 : shard_ident.stripe_size.0
5615 : );
5616 :
5617 0 : this_child_tsps.push(TenantShardPersistence {
5618 0 : tenant_id: child.tenant_id.to_string(),
5619 0 : shard_number: child.shard_number.0 as i32,
5620 0 : shard_count: child.shard_count.literal() as i32,
5621 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
5622 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
5623 0 : // populate the correct generation as part of its transaction, to protect us
5624 0 : // against racing with changes in the state of the parent.
5625 0 : generation: None,
5626 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
5627 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
5628 0 : config: serde_json::to_string(&config).unwrap(),
5629 0 : splitting: SplitState::Splitting,
5630 0 :
5631 0 : // Scheduling policies and preferred AZ do not carry through to children
5632 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
5633 0 : .unwrap(),
5634 0 : preferred_az_id: preferred_az_id.as_ref().map(|az| az.0.clone()),
5635 0 : });
5636 0 : }
5637 :
5638 0 : child_tsps.push((target.parent_id, this_child_tsps));
5639 : }
5640 :
5641 0 : if let Err(e) = self
5642 0 : .persistence
5643 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
5644 0 : .await
5645 : {
5646 0 : match e {
5647 : DatabaseError::Query(diesel::result::Error::DatabaseError(
5648 : DatabaseErrorKind::UniqueViolation,
5649 : _,
5650 : )) => {
5651 : // Inserting a child shard violated a unique constraint: we raced with another call to
5652 : // this function
5653 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
5654 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
5655 : }
5656 0 : _ => return Err(ApiError::InternalServerError(e.into())),
5657 : }
5658 0 : }
5659 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
5660 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5661 0 : ));
5662 :
5663 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
5664 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
5665 : // is not set in memory, then it was not persisted.
5666 : {
5667 0 : let mut locked = self.inner.write().unwrap();
5668 0 : for target in &targets {
5669 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
5670 0 : parent_shard.splitting = SplitState::Splitting;
5671 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
5672 0 : // split operation.
5673 0 : parent_shard
5674 0 : .observed
5675 0 : .locations
5676 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
5677 0 : }
5678 : }
5679 : }
5680 :
5681 : // TODO: issue split calls concurrently (this only matters once we're splitting
5682 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
5683 :
5684 0 : for target in &targets {
5685 : let ShardSplitTarget {
5686 0 : parent_id,
5687 0 : node,
5688 0 : child_ids,
5689 0 : } = target;
5690 0 : let client = PageserverClient::new(
5691 0 : node.get_id(),
5692 0 : node.base_url(),
5693 0 : self.config.pageserver_jwt_token.as_deref(),
5694 0 : self.config.ssl_ca_cert.clone(),
5695 0 : )
5696 0 : .map_err(|e| passthrough_api_error(node, e))?;
5697 0 : let response = client
5698 0 : .tenant_shard_split(
5699 0 : *parent_id,
5700 0 : TenantShardSplitRequest {
5701 0 : new_shard_count: new_shard_count.literal(),
5702 0 : new_stripe_size,
5703 0 : },
5704 0 : )
5705 0 : .await
5706 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
5707 :
5708 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
5709 0 : "failpoint".to_string()
5710 0 : )));
5711 :
5712 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
5713 :
5714 0 : tracing::info!(
5715 0 : "Split {} into {}",
5716 0 : parent_id,
5717 0 : response
5718 0 : .new_shards
5719 0 : .iter()
5720 0 : .map(|s| format!("{:?}", s))
5721 0 : .collect::<Vec<_>>()
5722 0 : .join(",")
5723 : );
5724 :
5725 0 : if &response.new_shards != child_ids {
5726 : // This should never happen: the pageserver should agree with us on how shard splits work.
5727 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5728 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
5729 0 : parent_id,
5730 0 : response.new_shards,
5731 0 : child_ids
5732 0 : )));
5733 0 : }
5734 : }
5735 :
5736 : // TODO: if the pageserver restarted concurrently with our split API call,
5737 : // the actual generation of the child shard might differ from the generation
5738 : // we expect it to have. In order for our in-database generation to end up
5739 : // correct, we should carry the child generation back in the response and apply it here
5740 : // in complete_shard_split (and apply the correct generation in memory)
5741 : // (or, we can carry generation in the request and reject the request if
5742 : // it doesn't match, but that requires more retry logic on this side)
5743 :
5744 0 : self.persistence
5745 0 : .complete_shard_split(tenant_id, old_shard_count)
5746 0 : .await?;
5747 :
5748 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
5749 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
5750 0 : ));
5751 :
5752 : // Replace all the shards we just split with their children: this phase is infallible.
5753 0 : let (response, child_locations, waiters) =
5754 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
5755 0 :
5756 0 : // Send compute notifications for all the new shards
5757 0 : let mut failed_notifications = Vec::new();
5758 0 : for (child_id, child_ps, stripe_size) in child_locations {
5759 0 : if let Err(e) = self
5760 0 : .compute_hook
5761 0 : .notify(
5762 0 : compute_hook::ShardUpdate {
5763 0 : tenant_shard_id: child_id,
5764 0 : node_id: child_ps,
5765 0 : stripe_size,
5766 0 : preferred_az: preferred_az_id.as_ref().map(Cow::Borrowed),
5767 0 : },
5768 0 : &self.cancel,
5769 0 : )
5770 0 : .await
5771 : {
5772 0 : tracing::warn!(
5773 0 : "Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
5774 : child_id,
5775 : child_ps
5776 : );
5777 0 : failed_notifications.push(child_id);
5778 0 : }
5779 : }
5780 :
5781 : // If we failed any compute notifications, make a note to retry later.
5782 0 : if !failed_notifications.is_empty() {
5783 0 : let mut locked = self.inner.write().unwrap();
5784 0 : for failed in failed_notifications {
5785 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
5786 0 : shard.pending_compute_notification = true;
5787 0 : }
5788 : }
5789 0 : }
5790 :
5791 0 : Ok((response, waiters))
5792 0 : }
5793 :
5794 : /// A graceful migration: update the preferred node and let optimisation handle the migration
5795 : /// in the background (may take a long time as it will fully warm up a location before cutting over)
5796 : ///
5797 : /// Our external API calls this a 'prewarm=true' migration, but internally it isn't a special prewarm step: it's
5798 : /// just a migration that uses the same graceful procedure as our background scheduling optimisations would use.
5799 0 : fn tenant_shard_migrate_with_prewarm(
5800 0 : &self,
5801 0 : migrate_req: &TenantShardMigrateRequest,
5802 0 : shard: &mut TenantShard,
5803 0 : scheduler: &mut Scheduler,
5804 0 : schedule_context: ScheduleContext,
5805 0 : ) -> Result<Option<ScheduleOptimization>, ApiError> {
5806 0 : shard.set_preferred_node(Some(migrate_req.node_id));
5807 0 :
5808 0 : // Generate whatever the initial change to the intent is: this could be creation of a secondary, or
5809 0 : // cutting over to an existing secondary. Caller is responsible for validating this before applying it,
5810 0 : // e.g. by checking secondary is warm enough.
5811 0 : Ok(shard.optimize_attachment(scheduler, &schedule_context))
5812 0 : }
5813 :
5814 : /// Immediate migration: directly update the intent state and kick off a reconciler
5815 0 : fn tenant_shard_migrate_immediate(
5816 0 : &self,
5817 0 : migrate_req: &TenantShardMigrateRequest,
5818 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5819 0 : shard: &mut TenantShard,
5820 0 : scheduler: &mut Scheduler,
5821 0 : ) -> Result<Option<ReconcilerWaiter>, ApiError> {
5822 0 : // Non-graceful migration: update the intent state immediately
5823 0 : let old_attached = *shard.intent.get_attached();
5824 0 : match shard.policy {
5825 0 : PlacementPolicy::Attached(n) => {
5826 0 : // If our new attached node was a secondary, it no longer should be.
5827 0 : shard
5828 0 : .intent
5829 0 : .remove_secondary(scheduler, migrate_req.node_id);
5830 0 :
5831 0 : shard
5832 0 : .intent
5833 0 : .set_attached(scheduler, Some(migrate_req.node_id));
5834 :
5835 : // If we were already attached to something, demote that to a secondary
5836 0 : if let Some(old_attached) = old_attached {
5837 0 : if n > 0 {
5838 : // Remove other secondaries to make room for the location we'll demote
5839 0 : while shard.intent.get_secondary().len() >= n {
5840 0 : shard.intent.pop_secondary(scheduler);
5841 0 : }
5842 :
5843 0 : shard.intent.push_secondary(scheduler, old_attached);
5844 0 : }
5845 0 : }
5846 : }
5847 0 : PlacementPolicy::Secondary => {
5848 0 : shard.intent.clear(scheduler);
5849 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
5850 0 : }
5851 : PlacementPolicy::Detached => {
5852 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5853 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
5854 0 : )));
5855 : }
5856 : }
5857 :
5858 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
5859 0 : shard.sequence = shard.sequence.next();
5860 0 : shard.set_preferred_node(None); // Abort any in-flight graceful migration
5861 0 : Ok(self.maybe_configured_reconcile_shard(
5862 0 : shard,
5863 0 : nodes,
5864 0 : (&migrate_req.migration_config).into(),
5865 0 : ))
5866 0 : }
5867 :
5868 0 : pub(crate) async fn tenant_shard_migrate(
5869 0 : &self,
5870 0 : tenant_shard_id: TenantShardId,
5871 0 : migrate_req: TenantShardMigrateRequest,
5872 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
5873 : // Depending on whether the migration is a change and whether it's graceful or immediate, we might
5874 : // get a different outcome to handle
5875 : enum MigrationOutcome {
5876 : Optimization(Option<ScheduleOptimization>),
5877 : Reconcile(Option<ReconcilerWaiter>),
5878 : }
5879 :
5880 0 : let outcome = {
5881 0 : let mut locked = self.inner.write().unwrap();
5882 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5883 :
5884 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
5885 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
5886 0 : "Node {} not found",
5887 0 : migrate_req.node_id
5888 0 : )));
5889 : };
5890 :
5891 : // Migration to unavavailable node requires force flag
5892 0 : if !node.is_available() {
5893 0 : if migrate_req.migration_config.override_scheduler {
5894 : // Warn but proceed: the caller may intend to manually adjust the placement of
5895 : // a shard even if the node is down, e.g. if intervening during an incident.
5896 0 : tracing::warn!("Forcibly migrating to unavailable node {node}");
5897 : } else {
5898 0 : tracing::warn!("Node {node} is unavailable, refusing migration");
5899 0 : return Err(ApiError::PreconditionFailed(
5900 0 : format!("Node {node} is unavailable").into_boxed_str(),
5901 0 : ));
5902 : }
5903 0 : }
5904 :
5905 : // Calculate the ScheduleContext for this tenant
5906 0 : let mut schedule_context = ScheduleContext::default();
5907 0 : for (_shard_id, shard) in
5908 0 : tenants.range(TenantShardId::tenant_range(tenant_shard_id.tenant_id))
5909 0 : {
5910 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5911 0 : }
5912 :
5913 : // Look up the specific shard we will migrate
5914 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5915 0 : return Err(ApiError::NotFound(
5916 0 : anyhow::anyhow!("Tenant shard not found").into(),
5917 0 : ));
5918 : };
5919 :
5920 : // Migration to a node with unfavorable scheduling score requires a force flag, because it might just
5921 : // be migrated back by the optimiser.
5922 0 : if let Some(better_node) = shard.find_better_location::<AttachedShardTag>(
5923 0 : scheduler,
5924 0 : &schedule_context,
5925 0 : migrate_req.node_id,
5926 0 : &[],
5927 0 : ) {
5928 0 : if !migrate_req.migration_config.override_scheduler {
5929 0 : return Err(ApiError::PreconditionFailed(
5930 0 : "Migration to a worse-scoring node".into(),
5931 0 : ));
5932 : } else {
5933 0 : tracing::info!(
5934 0 : "Migrating to a worse-scoring node {} (optimiser would prefer {better_node})",
5935 : migrate_req.node_id
5936 : );
5937 : }
5938 0 : }
5939 :
5940 0 : if let Some(origin_node_id) = migrate_req.origin_node_id {
5941 0 : if shard.intent.get_attached() != &Some(origin_node_id) {
5942 0 : return Err(ApiError::PreconditionFailed(
5943 0 : format!(
5944 0 : "Migration expected to originate from {} but shard is on {:?}",
5945 0 : origin_node_id,
5946 0 : shard.intent.get_attached()
5947 0 : )
5948 0 : .into(),
5949 0 : ));
5950 0 : }
5951 0 : }
5952 :
5953 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
5954 : // No-op case: we will still proceed to wait for reconciliation in case it is
5955 : // incomplete from an earlier update to the intent.
5956 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
5957 :
5958 : // An instruction to migrate to the currently attached node should
5959 : // cancel any pending graceful migration
5960 0 : shard.set_preferred_node(None);
5961 0 :
5962 0 : MigrationOutcome::Reconcile(self.maybe_configured_reconcile_shard(
5963 0 : shard,
5964 0 : nodes,
5965 0 : (&migrate_req.migration_config).into(),
5966 0 : ))
5967 0 : } else if migrate_req.migration_config.prewarm {
5968 0 : MigrationOutcome::Optimization(self.tenant_shard_migrate_with_prewarm(
5969 0 : &migrate_req,
5970 0 : shard,
5971 0 : scheduler,
5972 0 : schedule_context,
5973 0 : )?)
5974 : } else {
5975 0 : MigrationOutcome::Reconcile(self.tenant_shard_migrate_immediate(
5976 0 : &migrate_req,
5977 0 : nodes,
5978 0 : shard,
5979 0 : scheduler,
5980 0 : )?)
5981 : }
5982 : };
5983 :
5984 : // We may need to validate + apply an optimisation, or we may need to just retrive a reconcile waiter
5985 0 : let waiter = match outcome {
5986 0 : MigrationOutcome::Optimization(Some(optimization)) => {
5987 : // Validate and apply the optimization -- this would happen anyway in background reconcile loop, but
5988 : // we might as well do it more promptly as this is a direct external request.
5989 0 : let mut validated = self
5990 0 : .optimize_all_validate(vec![(tenant_shard_id, optimization)])
5991 0 : .await;
5992 0 : if let Some((_shard_id, optimization)) = validated.pop() {
5993 0 : let mut locked = self.inner.write().unwrap();
5994 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5995 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5996 : // Rare but possible: tenant is removed between generating optimisation and validating it.
5997 0 : return Err(ApiError::NotFound(
5998 0 : anyhow::anyhow!("Tenant shard not found").into(),
5999 0 : ));
6000 : };
6001 :
6002 0 : if !shard.apply_optimization(scheduler, optimization) {
6003 : // This can happen but is unusual enough to warn on: something else changed in the shard that made the optimisation stale
6004 : // and therefore not applied.
6005 0 : tracing::warn!(
6006 0 : "Schedule optimisation generated during graceful migration was not applied, shard changed?"
6007 : );
6008 0 : }
6009 0 : self.maybe_configured_reconcile_shard(
6010 0 : shard,
6011 0 : nodes,
6012 0 : (&migrate_req.migration_config).into(),
6013 0 : )
6014 : } else {
6015 0 : None
6016 : }
6017 : }
6018 0 : MigrationOutcome::Optimization(None) => None,
6019 0 : MigrationOutcome::Reconcile(waiter) => waiter,
6020 : };
6021 :
6022 : // Finally, wait for any reconcile we started to complete. In the case of immediate-mode migrations to cold
6023 : // locations, this has a good chance of timing out.
6024 0 : if let Some(waiter) = waiter {
6025 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6026 : } else {
6027 0 : tracing::info!("Migration is a no-op");
6028 : }
6029 :
6030 0 : Ok(TenantShardMigrateResponse {})
6031 0 : }
6032 :
6033 0 : pub(crate) async fn tenant_shard_migrate_secondary(
6034 0 : &self,
6035 0 : tenant_shard_id: TenantShardId,
6036 0 : migrate_req: TenantShardMigrateRequest,
6037 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
6038 0 : let waiter = {
6039 0 : let mut locked = self.inner.write().unwrap();
6040 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6041 :
6042 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
6043 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
6044 0 : "Node {} not found",
6045 0 : migrate_req.node_id
6046 0 : )));
6047 : };
6048 :
6049 0 : if !node.is_available() {
6050 : // Warn but proceed: the caller may intend to manually adjust the placement of
6051 : // a shard even if the node is down, e.g. if intervening during an incident.
6052 0 : tracing::warn!("Migrating to unavailable node {node}");
6053 0 : }
6054 :
6055 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
6056 0 : return Err(ApiError::NotFound(
6057 0 : anyhow::anyhow!("Tenant shard not found").into(),
6058 0 : ));
6059 : };
6060 :
6061 0 : if shard.intent.get_secondary().len() == 1
6062 0 : && shard.intent.get_secondary()[0] == migrate_req.node_id
6063 : {
6064 0 : tracing::info!(
6065 0 : "Migrating secondary to {node}: intent is unchanged {:?}",
6066 : shard.intent
6067 : );
6068 0 : } else if shard.intent.get_attached() == &Some(migrate_req.node_id) {
6069 0 : tracing::info!(
6070 0 : "Migrating secondary to {node}: already attached where we were asked to create a secondary"
6071 : );
6072 : } else {
6073 0 : let old_secondaries = shard.intent.get_secondary().clone();
6074 0 : for secondary in old_secondaries {
6075 0 : shard.intent.remove_secondary(scheduler, secondary);
6076 0 : }
6077 :
6078 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
6079 0 : shard.sequence = shard.sequence.next();
6080 0 : tracing::info!(
6081 0 : "Migrating secondary to {node}: new intent {:?}",
6082 : shard.intent
6083 : );
6084 : }
6085 :
6086 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::High)
6087 : };
6088 :
6089 0 : if let Some(waiter) = waiter {
6090 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
6091 : } else {
6092 0 : tracing::info!("Migration is a no-op");
6093 : }
6094 :
6095 0 : Ok(TenantShardMigrateResponse {})
6096 0 : }
6097 :
6098 : /// 'cancel' in this context means cancel any ongoing reconcile
6099 0 : pub(crate) async fn tenant_shard_cancel_reconcile(
6100 0 : &self,
6101 0 : tenant_shard_id: TenantShardId,
6102 0 : ) -> Result<(), ApiError> {
6103 : // Take state lock and fire the cancellation token, after which we drop lock and wait for any ongoing reconcile to complete
6104 0 : let waiter = {
6105 0 : let locked = self.inner.write().unwrap();
6106 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
6107 0 : return Err(ApiError::NotFound(
6108 0 : anyhow::anyhow!("Tenant shard not found").into(),
6109 0 : ));
6110 : };
6111 :
6112 0 : let waiter = shard.get_waiter();
6113 0 : match waiter {
6114 : None => {
6115 0 : tracing::info!("Shard does not have an ongoing Reconciler");
6116 0 : return Ok(());
6117 : }
6118 0 : Some(waiter) => {
6119 0 : tracing::info!("Cancelling Reconciler");
6120 0 : shard.cancel_reconciler();
6121 0 : waiter
6122 0 : }
6123 0 : }
6124 0 : };
6125 0 :
6126 0 : // Cancellation should be prompt. If this fails we have still done our job of firing the
6127 0 : // cancellation token, but by returning an ApiError we will indicate to the caller that
6128 0 : // the Reconciler is misbehaving and not respecting the cancellation token
6129 0 : self.await_waiters(vec![waiter], SHORT_RECONCILE_TIMEOUT)
6130 0 : .await?;
6131 :
6132 0 : Ok(())
6133 0 : }
6134 :
6135 : /// This is for debug/support only: we simply drop all state for a tenant, without
6136 : /// detaching or deleting it on pageservers.
6137 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
6138 0 : self.persistence.delete_tenant(tenant_id).await?;
6139 :
6140 0 : let mut locked = self.inner.write().unwrap();
6141 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
6142 0 : let mut shards = Vec::new();
6143 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
6144 0 : shards.push(*tenant_shard_id);
6145 0 : }
6146 :
6147 0 : for shard_id in shards {
6148 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
6149 0 : shard.intent.clear(scheduler);
6150 0 : }
6151 : }
6152 :
6153 0 : Ok(())
6154 0 : }
6155 :
6156 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
6157 : /// tenant with a very high generation number so that it will see the existing data.
6158 0 : pub(crate) async fn tenant_import(
6159 0 : &self,
6160 0 : tenant_id: TenantId,
6161 0 : ) -> Result<TenantCreateResponse, ApiError> {
6162 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
6163 0 : let maybe_node = {
6164 0 : self.inner
6165 0 : .read()
6166 0 : .unwrap()
6167 0 : .nodes
6168 0 : .values()
6169 0 : .find(|n| n.is_available())
6170 0 : .cloned()
6171 : };
6172 0 : let Some(node) = maybe_node else {
6173 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
6174 : };
6175 :
6176 0 : let client = PageserverClient::new(
6177 0 : node.get_id(),
6178 0 : node.base_url(),
6179 0 : self.config.pageserver_jwt_token.as_deref(),
6180 0 : self.config.ssl_ca_cert.clone(),
6181 0 : )
6182 0 : .map_err(|e| passthrough_api_error(&node, e))?;
6183 :
6184 0 : let scan_result = client
6185 0 : .tenant_scan_remote_storage(tenant_id)
6186 0 : .await
6187 0 : .map_err(|e| passthrough_api_error(&node, e))?;
6188 :
6189 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
6190 0 : let Some(shard_count) = scan_result
6191 0 : .shards
6192 0 : .iter()
6193 0 : .map(|s| s.tenant_shard_id.shard_count)
6194 0 : .max()
6195 : else {
6196 0 : return Err(ApiError::NotFound(
6197 0 : anyhow::anyhow!("No shards found").into(),
6198 0 : ));
6199 : };
6200 :
6201 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
6202 : // to
6203 0 : let generation = scan_result
6204 0 : .shards
6205 0 : .iter()
6206 0 : .map(|s| s.generation)
6207 0 : .max()
6208 0 : .expect("We already validated >0 shards");
6209 0 :
6210 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
6211 0 : // only work if they were using the default stripe size.
6212 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
6213 :
6214 0 : let (response, waiters) = self
6215 0 : .do_tenant_create(TenantCreateRequest {
6216 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
6217 0 : generation,
6218 0 :
6219 0 : shard_parameters: ShardParameters {
6220 0 : count: shard_count,
6221 0 : stripe_size,
6222 0 : },
6223 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
6224 0 : config: TenantConfig::default(),
6225 0 : })
6226 0 : .await?;
6227 :
6228 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
6229 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
6230 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
6231 : // reconcile, as reconciliation includes notifying compute.
6232 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
6233 0 : }
6234 :
6235 0 : Ok(response)
6236 0 : }
6237 :
6238 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
6239 : /// we don't have to make TenantShard clonable in the return path.
6240 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
6241 0 : let serialized = {
6242 0 : let locked = self.inner.read().unwrap();
6243 0 : let result = locked.tenants.values().collect::<Vec<_>>();
6244 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
6245 : };
6246 :
6247 0 : hyper::Response::builder()
6248 0 : .status(hyper::StatusCode::OK)
6249 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
6250 0 : .body(hyper::Body::from(serialized))
6251 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
6252 0 : }
6253 :
6254 : /// Check the consistency of in-memory state vs. persistent state, and check that the
6255 : /// scheduler's statistics are up to date.
6256 : ///
6257 : /// These consistency checks expect an **idle** system. If changes are going on while
6258 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
6259 : /// checks, but not suitable for running continuously in the background in the field.
6260 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
6261 0 : let (mut expect_nodes, mut expect_shards) = {
6262 0 : let locked = self.inner.read().unwrap();
6263 0 :
6264 0 : locked
6265 0 : .scheduler
6266 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
6267 0 : .context("Scheduler checks")
6268 0 : .map_err(ApiError::InternalServerError)?;
6269 :
6270 0 : let expect_nodes = locked
6271 0 : .nodes
6272 0 : .values()
6273 0 : .map(|n| n.to_persistent())
6274 0 : .collect::<Vec<_>>();
6275 0 :
6276 0 : let expect_shards = locked
6277 0 : .tenants
6278 0 : .values()
6279 0 : .map(|t| t.to_persistent())
6280 0 : .collect::<Vec<_>>();
6281 :
6282 : // This method can only validate the state of an idle system: if a reconcile is in
6283 : // progress, fail out early to avoid giving false errors on state that won't match
6284 : // between database and memory under a ReconcileResult is processed.
6285 0 : for t in locked.tenants.values() {
6286 0 : if t.reconciler.is_some() {
6287 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6288 0 : "Shard {} reconciliation in progress",
6289 0 : t.tenant_shard_id
6290 0 : )));
6291 0 : }
6292 : }
6293 :
6294 0 : (expect_nodes, expect_shards)
6295 : };
6296 :
6297 0 : let mut nodes = self.persistence.list_nodes().await?;
6298 0 : expect_nodes.sort_by_key(|n| n.node_id);
6299 0 : nodes.sort_by_key(|n| n.node_id);
6300 :
6301 : // Errors relating to nodes are deferred so that we don't skip the shard checks below if we have a node error
6302 0 : let node_result = if nodes != expect_nodes {
6303 0 : tracing::error!("Consistency check failed on nodes.");
6304 0 : tracing::error!(
6305 0 : "Nodes in memory: {}",
6306 0 : serde_json::to_string(&expect_nodes)
6307 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6308 : );
6309 0 : tracing::error!(
6310 0 : "Nodes in database: {}",
6311 0 : serde_json::to_string(&nodes)
6312 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6313 : );
6314 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
6315 0 : "Node consistency failure"
6316 0 : )))
6317 : } else {
6318 0 : Ok(())
6319 : };
6320 :
6321 0 : let mut persistent_shards = self.persistence.load_active_tenant_shards().await?;
6322 0 : persistent_shards
6323 0 : .sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6324 0 :
6325 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
6326 :
6327 : // Because JSON contents of persistent tenants might disagree with the fields in current `TenantConfig`
6328 : // definition, we will do an encode/decode cycle to ensure any legacy fields are dropped and any new
6329 : // fields are added, before doing a comparison.
6330 0 : for tsp in &mut persistent_shards {
6331 0 : let config: TenantConfig = serde_json::from_str(&tsp.config)
6332 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
6333 0 : tsp.config = serde_json::to_string(&config).expect("Encoding config is infallible");
6334 : }
6335 :
6336 0 : if persistent_shards != expect_shards {
6337 0 : tracing::error!("Consistency check failed on shards.");
6338 :
6339 0 : tracing::error!(
6340 0 : "Shards in memory: {}",
6341 0 : serde_json::to_string(&expect_shards)
6342 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6343 : );
6344 0 : tracing::error!(
6345 0 : "Shards in database: {}",
6346 0 : serde_json::to_string(&persistent_shards)
6347 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6348 : );
6349 :
6350 : // The total dump log lines above are useful in testing but in the field grafana will
6351 : // usually just drop them because they're so large. So we also do some explicit logging
6352 : // of just the diffs.
6353 0 : let persistent_shards = persistent_shards
6354 0 : .into_iter()
6355 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6356 0 : .collect::<HashMap<_, _>>();
6357 0 : let expect_shards = expect_shards
6358 0 : .into_iter()
6359 0 : .map(|tsp| (tsp.get_tenant_shard_id().unwrap(), tsp))
6360 0 : .collect::<HashMap<_, _>>();
6361 0 : for (tenant_shard_id, persistent_tsp) in &persistent_shards {
6362 0 : match expect_shards.get(tenant_shard_id) {
6363 : None => {
6364 0 : tracing::error!(
6365 0 : "Shard {} found in database but not in memory",
6366 : tenant_shard_id
6367 : );
6368 : }
6369 0 : Some(expect_tsp) => {
6370 0 : if expect_tsp != persistent_tsp {
6371 0 : tracing::error!(
6372 0 : "Shard {} is inconsistent. In memory: {}, database has: {}",
6373 0 : tenant_shard_id,
6374 0 : serde_json::to_string(expect_tsp).unwrap(),
6375 0 : serde_json::to_string(&persistent_tsp).unwrap()
6376 : );
6377 0 : }
6378 : }
6379 : }
6380 : }
6381 :
6382 : // Having already logged any differences, log any shards that simply aren't present in the database
6383 0 : for (tenant_shard_id, memory_tsp) in &expect_shards {
6384 0 : if !persistent_shards.contains_key(tenant_shard_id) {
6385 0 : tracing::error!(
6386 0 : "Shard {} found in memory but not in database: {}",
6387 0 : tenant_shard_id,
6388 0 : serde_json::to_string(memory_tsp)
6389 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6390 : );
6391 0 : }
6392 : }
6393 :
6394 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6395 0 : "Shard consistency failure"
6396 0 : )));
6397 0 : }
6398 0 :
6399 0 : node_result
6400 0 : }
6401 :
6402 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
6403 : /// we don't have to make TenantShard clonable in the return path.
6404 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
6405 0 : let serialized = {
6406 0 : let locked = self.inner.read().unwrap();
6407 0 : serde_json::to_string(&locked.scheduler)
6408 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
6409 : };
6410 :
6411 0 : hyper::Response::builder()
6412 0 : .status(hyper::StatusCode::OK)
6413 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
6414 0 : .body(hyper::Body::from(serialized))
6415 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
6416 0 : }
6417 :
6418 : /// This is for debug/support only: we simply drop all state for a tenant, without
6419 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
6420 : /// tenants that were on this node.
6421 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
6422 0 : self.persistence.delete_node(node_id).await?;
6423 :
6424 0 : let mut locked = self.inner.write().unwrap();
6425 :
6426 0 : for shard in locked.tenants.values_mut() {
6427 0 : shard.deref_node(node_id);
6428 0 : shard.observed.locations.remove(&node_id);
6429 0 : }
6430 :
6431 0 : let mut nodes = (*locked.nodes).clone();
6432 0 : nodes.remove(&node_id);
6433 0 : locked.nodes = Arc::new(nodes);
6434 0 : metrics::METRICS_REGISTRY
6435 0 : .metrics_group
6436 0 : .storage_controller_pageserver_nodes
6437 0 : .set(locked.nodes.len() as i64);
6438 0 :
6439 0 : locked.scheduler.node_remove(node_id);
6440 0 :
6441 0 : Ok(())
6442 0 : }
6443 :
6444 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
6445 : /// that we don't leave any bad state behind in the storage controller, but unclean
6446 : /// in the sense that we are not carefully draining the node.
6447 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
6448 0 : let _node_lock =
6449 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
6450 :
6451 : // 1. Atomically update in-memory state:
6452 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
6453 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
6454 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
6455 : // re-insert references to this node into the ObservedState of shards
6456 : // - drop the node from the scheduler
6457 : {
6458 0 : let mut locked = self.inner.write().unwrap();
6459 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6460 0 :
6461 0 : {
6462 0 : let mut nodes_mut = (*nodes).deref().clone();
6463 0 : match nodes_mut.get_mut(&node_id) {
6464 0 : Some(node) => {
6465 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
6466 0 : // if we crash it would not be desirable to leave the node paused after a restart.
6467 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
6468 0 : }
6469 : None => {
6470 0 : tracing::info!(
6471 0 : "Node not found: presuming this is a retry and returning success"
6472 : );
6473 0 : return Ok(());
6474 : }
6475 : }
6476 :
6477 0 : *nodes = Arc::new(nodes_mut);
6478 : }
6479 :
6480 0 : for (_tenant_id, mut schedule_context, shards) in
6481 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6482 : {
6483 0 : for shard in shards {
6484 0 : if shard.deref_node(node_id) {
6485 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
6486 : // TODO: implement force flag to remove a node even if we can't reschedule
6487 : // a tenant
6488 0 : tracing::error!(
6489 0 : "Refusing to delete node, shard {} can't be rescheduled: {e}",
6490 : shard.tenant_shard_id
6491 : );
6492 0 : return Err(e.into());
6493 : } else {
6494 0 : tracing::info!(
6495 0 : "Rescheduled shard {} away from node during deletion",
6496 : shard.tenant_shard_id
6497 : )
6498 : }
6499 :
6500 0 : self.maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal);
6501 0 : }
6502 :
6503 : // Here we remove an existing observed location for the node we're removing, and it will
6504 : // not be re-added by a reconciler's completion because we filter out removed nodes in
6505 : // process_result.
6506 : //
6507 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
6508 : // means any reconciles we spawned will know about the node we're deleting, enabling them
6509 : // to do live migrations if it's still online.
6510 0 : shard.observed.locations.remove(&node_id);
6511 : }
6512 : }
6513 :
6514 0 : scheduler.node_remove(node_id);
6515 0 :
6516 0 : {
6517 0 : let mut nodes_mut = (**nodes).clone();
6518 0 : if let Some(mut removed_node) = nodes_mut.remove(&node_id) {
6519 0 : // Ensure that any reconciler holding an Arc<> to this node will
6520 0 : // drop out when trying to RPC to it (setting Offline state sets the
6521 0 : // cancellation token on the Node object).
6522 0 : removed_node.set_availability(NodeAvailability::Offline);
6523 0 : }
6524 0 : *nodes = Arc::new(nodes_mut);
6525 0 : metrics::METRICS_REGISTRY
6526 0 : .metrics_group
6527 0 : .storage_controller_pageserver_nodes
6528 0 : .set(nodes.len() as i64);
6529 0 : }
6530 0 : }
6531 0 :
6532 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
6533 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
6534 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
6535 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
6536 0 : // that exists.
6537 0 :
6538 0 : // 2. Actually delete the node from the database and from in-memory state
6539 0 : tracing::info!("Deleting node from database");
6540 0 : self.persistence.delete_node(node_id).await?;
6541 :
6542 0 : Ok(())
6543 0 : }
6544 :
6545 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
6546 0 : let nodes = {
6547 0 : self.inner
6548 0 : .read()
6549 0 : .unwrap()
6550 0 : .nodes
6551 0 : .values()
6552 0 : .cloned()
6553 0 : .collect::<Vec<_>>()
6554 0 : };
6555 0 :
6556 0 : Ok(nodes)
6557 0 : }
6558 :
6559 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
6560 0 : self.inner
6561 0 : .read()
6562 0 : .unwrap()
6563 0 : .nodes
6564 0 : .get(&node_id)
6565 0 : .cloned()
6566 0 : .ok_or(ApiError::NotFound(
6567 0 : format!("Node {node_id} not registered").into(),
6568 0 : ))
6569 0 : }
6570 :
6571 0 : pub(crate) async fn get_node_shards(
6572 0 : &self,
6573 0 : node_id: NodeId,
6574 0 : ) -> Result<NodeShardResponse, ApiError> {
6575 0 : let locked = self.inner.read().unwrap();
6576 0 : let mut shards = Vec::new();
6577 0 : for (tid, tenant) in locked.tenants.iter() {
6578 0 : let is_intended_secondary = match (
6579 0 : tenant.intent.get_attached() == &Some(node_id),
6580 0 : tenant.intent.get_secondary().contains(&node_id),
6581 0 : ) {
6582 : (true, true) => {
6583 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
6584 0 : "{} attached as primary+secondary on the same node",
6585 0 : tid
6586 0 : )));
6587 : }
6588 0 : (true, false) => Some(false),
6589 0 : (false, true) => Some(true),
6590 0 : (false, false) => None,
6591 : };
6592 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
6593 0 : tenant.observed.locations.get(&node_id)
6594 : {
6595 0 : Some(conf.secondary_conf.is_some())
6596 : } else {
6597 0 : None
6598 : };
6599 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
6600 0 : shards.push(NodeShard {
6601 0 : tenant_shard_id: *tid,
6602 0 : is_intended_secondary,
6603 0 : is_observed_secondary,
6604 0 : });
6605 0 : }
6606 : }
6607 0 : Ok(NodeShardResponse { node_id, shards })
6608 0 : }
6609 :
6610 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
6611 0 : self.persistence.get_leader().await
6612 0 : }
6613 :
6614 0 : pub(crate) async fn node_register(
6615 0 : &self,
6616 0 : register_req: NodeRegisterRequest,
6617 0 : ) -> Result<(), ApiError> {
6618 0 : let _node_lock = trace_exclusive_lock(
6619 0 : &self.node_op_locks,
6620 0 : register_req.node_id,
6621 0 : NodeOperations::Register,
6622 0 : )
6623 0 : .await;
6624 :
6625 : #[derive(PartialEq)]
6626 : enum RegistrationStatus {
6627 : UpToDate,
6628 : NeedUpdate,
6629 : Mismatched,
6630 : New,
6631 : }
6632 :
6633 0 : let registration_status = {
6634 0 : let locked = self.inner.read().unwrap();
6635 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
6636 0 : if node.registration_match(®ister_req) {
6637 0 : if node.need_update(®ister_req) {
6638 0 : RegistrationStatus::NeedUpdate
6639 : } else {
6640 0 : RegistrationStatus::UpToDate
6641 : }
6642 : } else {
6643 0 : RegistrationStatus::Mismatched
6644 : }
6645 : } else {
6646 0 : RegistrationStatus::New
6647 : }
6648 : };
6649 :
6650 0 : match registration_status {
6651 : RegistrationStatus::UpToDate => {
6652 0 : tracing::info!(
6653 0 : "Node {} re-registered with matching address and is up to date",
6654 : register_req.node_id
6655 : );
6656 :
6657 0 : return Ok(());
6658 : }
6659 : RegistrationStatus::Mismatched => {
6660 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
6661 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
6662 : // a fixed address through the lifetime of a node.
6663 0 : tracing::warn!(
6664 0 : "Node {} tried to register with different address",
6665 : register_req.node_id
6666 : );
6667 0 : return Err(ApiError::Conflict(
6668 0 : "Node is already registered with different address".to_string(),
6669 0 : ));
6670 : }
6671 0 : RegistrationStatus::New | RegistrationStatus::NeedUpdate => {
6672 0 : // fallthrough
6673 0 : }
6674 0 : }
6675 0 :
6676 0 : // We do not require that a node is actually online when registered (it will start life
6677 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
6678 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
6679 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
6680 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
6681 0 : if tokio::net::lookup_host(format!(
6682 0 : "{}:{}",
6683 0 : register_req.listen_http_addr, register_req.listen_http_port
6684 0 : ))
6685 0 : .await
6686 0 : .is_err()
6687 : {
6688 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
6689 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
6690 : // we return a soft 503 error, to encourage callers to retry past transient issues.
6691 0 : return Err(ApiError::ResourceUnavailable(
6692 0 : format!(
6693 0 : "Node {} tried to register with unknown DNS name '{}'",
6694 0 : register_req.node_id, register_req.listen_http_addr
6695 0 : )
6696 0 : .into(),
6697 0 : ));
6698 0 : }
6699 0 :
6700 0 : if self.config.use_https_pageserver_api && register_req.listen_https_port.is_none() {
6701 0 : return Err(ApiError::PreconditionFailed(
6702 0 : format!(
6703 0 : "Node {} has no https port, but use_https is enabled",
6704 0 : register_req.node_id
6705 0 : )
6706 0 : .into(),
6707 0 : ));
6708 0 : }
6709 0 :
6710 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
6711 0 : // This ensures that before we use it for anything or expose it via any external
6712 0 : // API, it is guaranteed to be available after a restart.
6713 0 : let new_node = Node::new(
6714 0 : register_req.node_id,
6715 0 : register_req.listen_http_addr,
6716 0 : register_req.listen_http_port,
6717 0 : register_req.listen_https_port,
6718 0 : register_req.listen_pg_addr,
6719 0 : register_req.listen_pg_port,
6720 0 : register_req.availability_zone_id.clone(),
6721 0 : self.config.use_https_pageserver_api,
6722 0 : );
6723 0 : let new_node = match new_node {
6724 0 : Ok(new_node) => new_node,
6725 0 : Err(error) => return Err(ApiError::InternalServerError(error)),
6726 : };
6727 :
6728 0 : match registration_status {
6729 0 : RegistrationStatus::New => self.persistence.insert_node(&new_node).await?,
6730 : RegistrationStatus::NeedUpdate => {
6731 0 : self.persistence
6732 0 : .update_node_on_registration(
6733 0 : register_req.node_id,
6734 0 : register_req.listen_https_port,
6735 0 : )
6736 0 : .await?
6737 : }
6738 0 : _ => unreachable!("Other statuses have been processed earlier"),
6739 : }
6740 :
6741 0 : let mut locked = self.inner.write().unwrap();
6742 0 : let mut new_nodes = (*locked.nodes).clone();
6743 0 :
6744 0 : locked.scheduler.node_upsert(&new_node);
6745 0 : new_nodes.insert(register_req.node_id, new_node);
6746 0 :
6747 0 : locked.nodes = Arc::new(new_nodes);
6748 0 :
6749 0 : metrics::METRICS_REGISTRY
6750 0 : .metrics_group
6751 0 : .storage_controller_pageserver_nodes
6752 0 : .set(locked.nodes.len() as i64);
6753 0 :
6754 0 : match registration_status {
6755 : RegistrationStatus::New => {
6756 0 : tracing::info!(
6757 0 : "Registered pageserver {} ({}), now have {} pageservers",
6758 0 : register_req.node_id,
6759 0 : register_req.availability_zone_id,
6760 0 : locked.nodes.len()
6761 : );
6762 : }
6763 : RegistrationStatus::NeedUpdate => {
6764 0 : tracing::info!(
6765 0 : "Re-registered and updated node {} ({})",
6766 : register_req.node_id,
6767 : register_req.availability_zone_id,
6768 : );
6769 : }
6770 0 : _ => unreachable!("Other statuses have been processed earlier"),
6771 : }
6772 0 : Ok(())
6773 0 : }
6774 :
6775 : /// Configure in-memory and persistent state of a node as requested
6776 : ///
6777 : /// Note that this function does not trigger any immediate side effects in response
6778 : /// to the changes. That part is handled by [`Self::handle_node_availability_transition`].
6779 0 : async fn node_state_configure(
6780 0 : &self,
6781 0 : node_id: NodeId,
6782 0 : availability: Option<NodeAvailability>,
6783 0 : scheduling: Option<NodeSchedulingPolicy>,
6784 0 : node_lock: &TracingExclusiveGuard<NodeOperations>,
6785 0 : ) -> Result<AvailabilityTransition, ApiError> {
6786 0 : if let Some(scheduling) = scheduling {
6787 : // Scheduling is a persistent part of Node: we must write updates to the database before
6788 : // applying them in memory
6789 0 : self.persistence
6790 0 : .update_node_scheduling_policy(node_id, scheduling)
6791 0 : .await?;
6792 0 : }
6793 :
6794 : // If we're activating a node, then before setting it active we must reconcile any shard locations
6795 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
6796 : // by calling [`Self::node_activate_reconcile`]
6797 : //
6798 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
6799 : // nothing else can mutate its availability while we run.
6800 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
6801 0 : let (activate_node, availability_transition) = {
6802 0 : let locked = self.inner.read().unwrap();
6803 0 : let Some(node) = locked.nodes.get(&node_id) else {
6804 0 : return Err(ApiError::NotFound(
6805 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
6806 0 : ));
6807 : };
6808 :
6809 0 : (
6810 0 : node.clone(),
6811 0 : node.get_availability_transition(input_availability),
6812 0 : )
6813 : };
6814 :
6815 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
6816 0 : self.node_activate_reconcile(activate_node, node_lock)
6817 0 : .await?;
6818 0 : }
6819 0 : availability_transition
6820 : } else {
6821 0 : AvailabilityTransition::Unchanged
6822 : };
6823 :
6824 : // Apply changes from the request to our in-memory state for the Node
6825 0 : let mut locked = self.inner.write().unwrap();
6826 0 : let (nodes, _tenants, scheduler) = locked.parts_mut();
6827 0 :
6828 0 : let mut new_nodes = (**nodes).clone();
6829 :
6830 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
6831 0 : return Err(ApiError::NotFound(
6832 0 : anyhow::anyhow!("Node not registered").into(),
6833 0 : ));
6834 : };
6835 :
6836 0 : if let Some(availability) = availability {
6837 0 : node.set_availability(availability);
6838 0 : }
6839 :
6840 0 : if let Some(scheduling) = scheduling {
6841 0 : node.set_scheduling(scheduling);
6842 0 : }
6843 :
6844 : // Update the scheduler, in case the elegibility of the node for new shards has changed
6845 0 : scheduler.node_upsert(node);
6846 0 :
6847 0 : let new_nodes = Arc::new(new_nodes);
6848 0 : locked.nodes = new_nodes;
6849 0 :
6850 0 : Ok(availability_transition)
6851 0 : }
6852 :
6853 : /// Handle availability transition of one node
6854 : ///
6855 : /// Note that you should first call [`Self::node_state_configure`] to update
6856 : /// the in-memory state referencing that node. If you need to handle more than one transition
6857 : /// consider using [`Self::handle_node_availability_transitions`].
6858 0 : async fn handle_node_availability_transition(
6859 0 : &self,
6860 0 : node_id: NodeId,
6861 0 : transition: AvailabilityTransition,
6862 0 : _node_lock: &TracingExclusiveGuard<NodeOperations>,
6863 0 : ) -> Result<(), ApiError> {
6864 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
6865 0 : match transition {
6866 : AvailabilityTransition::ToOffline => {
6867 0 : tracing::info!("Node {} transition to offline", node_id);
6868 :
6869 0 : let mut locked = self.inner.write().unwrap();
6870 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6871 0 :
6872 0 : let mut tenants_affected: usize = 0;
6873 :
6874 0 : for (_tenant_id, mut schedule_context, shards) in
6875 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Normal)
6876 : {
6877 0 : for tenant_shard in shards {
6878 0 : let tenant_shard_id = tenant_shard.tenant_shard_id;
6879 0 : if let Some(observed_loc) =
6880 0 : tenant_shard.observed.locations.get_mut(&node_id)
6881 0 : {
6882 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
6883 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
6884 0 : observed_loc.conf = None;
6885 0 : }
6886 :
6887 0 : if nodes.len() == 1 {
6888 : // Special case for single-node cluster: there is no point trying to reschedule
6889 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
6890 : // failures to schedule them.
6891 0 : continue;
6892 0 : }
6893 0 :
6894 0 : if !nodes
6895 0 : .values()
6896 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
6897 : {
6898 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
6899 : // trying to reschedule since there's nowhere else to go. Without this
6900 : // branch we incorrectly detach tenants in response to node unavailability.
6901 0 : continue;
6902 0 : }
6903 0 :
6904 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
6905 0 : tenant_shard.sequence = tenant_shard.sequence.next();
6906 0 :
6907 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
6908 0 : Err(e) => {
6909 0 : // It is possible that some tenants will become unschedulable when too many pageservers
6910 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
6911 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
6912 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
6913 : }
6914 : Ok(()) => {
6915 0 : if self
6916 0 : .maybe_reconcile_shard(
6917 0 : tenant_shard,
6918 0 : nodes,
6919 0 : ReconcilerPriority::Normal,
6920 0 : )
6921 0 : .is_some()
6922 0 : {
6923 0 : tenants_affected += 1;
6924 0 : };
6925 : }
6926 : }
6927 0 : }
6928 : }
6929 : }
6930 0 : tracing::info!(
6931 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
6932 : tenants_affected,
6933 : node_id
6934 : )
6935 : }
6936 : AvailabilityTransition::ToActive => {
6937 0 : tracing::info!("Node {} transition to active", node_id);
6938 :
6939 0 : let mut locked = self.inner.write().unwrap();
6940 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
6941 :
6942 : // When a node comes back online, we must reconcile any tenant that has a None observed
6943 : // location on the node.
6944 0 : for tenant_shard in tenants.values_mut() {
6945 : // If a reconciliation is already in progress, rely on the previous scheduling
6946 : // decision and skip triggering a new reconciliation.
6947 0 : if tenant_shard.reconciler.is_some() {
6948 0 : continue;
6949 0 : }
6950 :
6951 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
6952 0 : if observed_loc.conf.is_none() {
6953 0 : self.maybe_reconcile_shard(
6954 0 : tenant_shard,
6955 0 : nodes,
6956 0 : ReconcilerPriority::Normal,
6957 0 : );
6958 0 : }
6959 0 : }
6960 : }
6961 :
6962 : // TODO: in the background, we should balance work back onto this pageserver
6963 : }
6964 : // No action required for the intermediate unavailable state.
6965 : // When we transition into active or offline from the unavailable state,
6966 : // the correct handling above will kick in.
6967 : AvailabilityTransition::ToWarmingUpFromActive => {
6968 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
6969 : }
6970 : AvailabilityTransition::ToWarmingUpFromOffline => {
6971 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
6972 : }
6973 : AvailabilityTransition::Unchanged => {
6974 0 : tracing::debug!("Node {} no availability change during config", node_id);
6975 : }
6976 : }
6977 :
6978 0 : Ok(())
6979 0 : }
6980 :
6981 : /// Handle availability transition for multiple nodes
6982 : ///
6983 : /// Note that you should first call [`Self::node_state_configure`] for
6984 : /// all nodes being handled here for the handling to use fresh in-memory state.
6985 0 : async fn handle_node_availability_transitions(
6986 0 : &self,
6987 0 : transitions: Vec<(
6988 0 : NodeId,
6989 0 : TracingExclusiveGuard<NodeOperations>,
6990 0 : AvailabilityTransition,
6991 0 : )>,
6992 0 : ) -> Result<(), Vec<(NodeId, ApiError)>> {
6993 0 : let mut errors = Vec::default();
6994 0 : for (node_id, node_lock, transition) in transitions {
6995 0 : let res = self
6996 0 : .handle_node_availability_transition(node_id, transition, &node_lock)
6997 0 : .await;
6998 0 : if let Err(err) = res {
6999 0 : errors.push((node_id, err));
7000 0 : }
7001 : }
7002 :
7003 0 : if errors.is_empty() {
7004 0 : Ok(())
7005 : } else {
7006 0 : Err(errors)
7007 : }
7008 0 : }
7009 :
7010 0 : pub(crate) async fn node_configure(
7011 0 : &self,
7012 0 : node_id: NodeId,
7013 0 : availability: Option<NodeAvailability>,
7014 0 : scheduling: Option<NodeSchedulingPolicy>,
7015 0 : ) -> Result<(), ApiError> {
7016 0 : let node_lock =
7017 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
7018 :
7019 0 : let transition = self
7020 0 : .node_state_configure(node_id, availability, scheduling, &node_lock)
7021 0 : .await?;
7022 0 : self.handle_node_availability_transition(node_id, transition, &node_lock)
7023 0 : .await
7024 0 : }
7025 :
7026 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
7027 : /// operation for HTTP api.
7028 0 : pub(crate) async fn external_node_configure(
7029 0 : &self,
7030 0 : node_id: NodeId,
7031 0 : availability: Option<NodeAvailability>,
7032 0 : scheduling: Option<NodeSchedulingPolicy>,
7033 0 : ) -> Result<(), ApiError> {
7034 0 : {
7035 0 : let locked = self.inner.read().unwrap();
7036 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
7037 0 : return Err(ApiError::PreconditionFailed(
7038 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
7039 0 : ));
7040 0 : }
7041 0 : }
7042 0 :
7043 0 : self.node_configure(node_id, availability, scheduling).await
7044 0 : }
7045 :
7046 0 : pub(crate) async fn start_node_drain(
7047 0 : self: &Arc<Self>,
7048 0 : node_id: NodeId,
7049 0 : ) -> Result<(), ApiError> {
7050 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
7051 0 : let locked = self.inner.read().unwrap();
7052 0 : let nodes = &locked.nodes;
7053 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7054 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7055 0 : ))?;
7056 0 : let schedulable_nodes_count = nodes
7057 0 : .iter()
7058 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
7059 0 : .count();
7060 0 :
7061 0 : (
7062 0 : locked
7063 0 : .ongoing_operation
7064 0 : .as_ref()
7065 0 : .map(|ongoing| ongoing.operation),
7066 0 : node.is_available(),
7067 0 : node.get_scheduling(),
7068 0 : schedulable_nodes_count,
7069 0 : )
7070 0 : };
7071 :
7072 0 : if let Some(ongoing) = ongoing_op {
7073 0 : return Err(ApiError::PreconditionFailed(
7074 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
7075 0 : ));
7076 0 : }
7077 0 :
7078 0 : if !node_available {
7079 0 : return Err(ApiError::ResourceUnavailable(
7080 0 : format!("Node {node_id} is currently unavailable").into(),
7081 0 : ));
7082 0 : }
7083 0 :
7084 0 : if schedulable_nodes_count == 0 {
7085 0 : return Err(ApiError::PreconditionFailed(
7086 0 : "No other schedulable nodes to drain to".into(),
7087 0 : ));
7088 0 : }
7089 0 :
7090 0 : match node_policy {
7091 : NodeSchedulingPolicy::Active => {
7092 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
7093 0 : .await?;
7094 :
7095 0 : let cancel = self.cancel.child_token();
7096 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
7097 :
7098 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
7099 0 : operation: Operation::Drain(Drain { node_id }),
7100 0 : cancel: cancel.clone(),
7101 0 : });
7102 :
7103 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
7104 :
7105 0 : tokio::task::spawn({
7106 0 : let service = self.clone();
7107 0 : let cancel = cancel.clone();
7108 0 : async move {
7109 0 : let _gate_guard = gate_guard;
7110 0 :
7111 0 : scopeguard::defer! {
7112 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
7113 0 :
7114 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
7115 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
7116 0 : } else {
7117 0 : panic!("We always remove the same operation")
7118 0 : }
7119 0 : }
7120 0 :
7121 0 : tracing::info!("Drain background operation starting");
7122 0 : let res = service.drain_node(node_id, cancel).await;
7123 0 : match res {
7124 : Ok(()) => {
7125 0 : tracing::info!("Drain background operation completed successfully");
7126 : }
7127 : Err(OperationError::Cancelled) => {
7128 0 : tracing::info!("Drain background operation was cancelled");
7129 : }
7130 0 : Err(err) => {
7131 0 : tracing::error!("Drain background operation encountered: {err}")
7132 : }
7133 : }
7134 0 : }
7135 0 : }.instrument(span));
7136 0 : }
7137 : NodeSchedulingPolicy::Draining => {
7138 0 : return Err(ApiError::Conflict(format!(
7139 0 : "Node {node_id} has drain in progress"
7140 0 : )));
7141 : }
7142 0 : policy => {
7143 0 : return Err(ApiError::PreconditionFailed(
7144 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
7145 0 : ));
7146 : }
7147 : }
7148 :
7149 0 : Ok(())
7150 0 : }
7151 :
7152 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
7153 0 : let node_available = {
7154 0 : let locked = self.inner.read().unwrap();
7155 0 : let nodes = &locked.nodes;
7156 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7157 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7158 0 : ))?;
7159 :
7160 0 : node.is_available()
7161 0 : };
7162 0 :
7163 0 : if !node_available {
7164 0 : return Err(ApiError::ResourceUnavailable(
7165 0 : format!("Node {node_id} is currently unavailable").into(),
7166 0 : ));
7167 0 : }
7168 :
7169 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
7170 0 : if let Operation::Drain(drain) = op_handler.operation {
7171 0 : if drain.node_id == node_id {
7172 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
7173 0 : op_handler.cancel.cancel();
7174 0 : return Ok(());
7175 0 : }
7176 0 : }
7177 0 : }
7178 :
7179 0 : Err(ApiError::PreconditionFailed(
7180 0 : format!("Node {node_id} has no drain in progress").into(),
7181 0 : ))
7182 0 : }
7183 :
7184 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
7185 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
7186 0 : let locked = self.inner.read().unwrap();
7187 0 : let nodes = &locked.nodes;
7188 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7189 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7190 0 : ))?;
7191 :
7192 0 : (
7193 0 : locked
7194 0 : .ongoing_operation
7195 0 : .as_ref()
7196 0 : .map(|ongoing| ongoing.operation),
7197 0 : node.is_available(),
7198 0 : node.get_scheduling(),
7199 0 : nodes.len(),
7200 0 : )
7201 0 : };
7202 :
7203 0 : if let Some(ongoing) = ongoing_op {
7204 0 : return Err(ApiError::PreconditionFailed(
7205 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
7206 0 : ));
7207 0 : }
7208 0 :
7209 0 : if !node_available {
7210 0 : return Err(ApiError::ResourceUnavailable(
7211 0 : format!("Node {node_id} is currently unavailable").into(),
7212 0 : ));
7213 0 : }
7214 0 :
7215 0 : if total_nodes_count <= 1 {
7216 0 : return Err(ApiError::PreconditionFailed(
7217 0 : "No other nodes to fill from".into(),
7218 0 : ));
7219 0 : }
7220 0 :
7221 0 : match node_policy {
7222 : NodeSchedulingPolicy::Active => {
7223 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
7224 0 : .await?;
7225 :
7226 0 : let cancel = self.cancel.child_token();
7227 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
7228 :
7229 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
7230 0 : operation: Operation::Fill(Fill { node_id }),
7231 0 : cancel: cancel.clone(),
7232 0 : });
7233 :
7234 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
7235 :
7236 0 : tokio::task::spawn({
7237 0 : let service = self.clone();
7238 0 : let cancel = cancel.clone();
7239 0 : async move {
7240 0 : let _gate_guard = gate_guard;
7241 0 :
7242 0 : scopeguard::defer! {
7243 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
7244 0 :
7245 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
7246 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
7247 0 : } else {
7248 0 : panic!("We always remove the same operation")
7249 0 : }
7250 0 : }
7251 0 :
7252 0 : tracing::info!("Fill background operation starting");
7253 0 : let res = service.fill_node(node_id, cancel).await;
7254 0 : match res {
7255 : Ok(()) => {
7256 0 : tracing::info!("Fill background operation completed successfully");
7257 : }
7258 : Err(OperationError::Cancelled) => {
7259 0 : tracing::info!("Fill background operation was cancelled");
7260 : }
7261 0 : Err(err) => {
7262 0 : tracing::error!("Fill background operation encountered: {err}")
7263 : }
7264 : }
7265 0 : }
7266 0 : }.instrument(span));
7267 0 : }
7268 : NodeSchedulingPolicy::Filling => {
7269 0 : return Err(ApiError::Conflict(format!(
7270 0 : "Node {node_id} has fill in progress"
7271 0 : )));
7272 : }
7273 0 : policy => {
7274 0 : return Err(ApiError::PreconditionFailed(
7275 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
7276 0 : ));
7277 : }
7278 : }
7279 :
7280 0 : Ok(())
7281 0 : }
7282 :
7283 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
7284 0 : let node_available = {
7285 0 : let locked = self.inner.read().unwrap();
7286 0 : let nodes = &locked.nodes;
7287 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
7288 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
7289 0 : ))?;
7290 :
7291 0 : node.is_available()
7292 0 : };
7293 0 :
7294 0 : if !node_available {
7295 0 : return Err(ApiError::ResourceUnavailable(
7296 0 : format!("Node {node_id} is currently unavailable").into(),
7297 0 : ));
7298 0 : }
7299 :
7300 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
7301 0 : if let Operation::Fill(fill) = op_handler.operation {
7302 0 : if fill.node_id == node_id {
7303 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
7304 0 : op_handler.cancel.cancel();
7305 0 : return Ok(());
7306 0 : }
7307 0 : }
7308 0 : }
7309 :
7310 0 : Err(ApiError::PreconditionFailed(
7311 0 : format!("Node {node_id} has no fill in progress").into(),
7312 0 : ))
7313 0 : }
7314 :
7315 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
7316 : /// configuration
7317 0 : fn maybe_reconcile_shard(
7318 0 : &self,
7319 0 : shard: &mut TenantShard,
7320 0 : nodes: &Arc<HashMap<NodeId, Node>>,
7321 0 : priority: ReconcilerPriority,
7322 0 : ) -> Option<ReconcilerWaiter> {
7323 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::new(priority))
7324 0 : }
7325 :
7326 : /// Before constructing a Reconciler, acquire semaphore units from the appropriate concurrency limit (depends on priority)
7327 0 : fn get_reconciler_units(
7328 0 : &self,
7329 0 : priority: ReconcilerPriority,
7330 0 : ) -> Result<ReconcileUnits, TryAcquireError> {
7331 0 : let units = match priority {
7332 0 : ReconcilerPriority::Normal => self.reconciler_concurrency.clone().try_acquire_owned(),
7333 : ReconcilerPriority::High => {
7334 0 : match self
7335 0 : .priority_reconciler_concurrency
7336 0 : .clone()
7337 0 : .try_acquire_owned()
7338 : {
7339 0 : Ok(u) => Ok(u),
7340 : Err(TryAcquireError::NoPermits) => {
7341 : // If the high priority semaphore is exhausted, then high priority tasks may steal units from
7342 : // the normal priority semaphore.
7343 0 : self.reconciler_concurrency.clone().try_acquire_owned()
7344 : }
7345 0 : Err(e) => Err(e),
7346 : }
7347 : }
7348 : };
7349 :
7350 0 : units.map(ReconcileUnits::new)
7351 0 : }
7352 :
7353 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
7354 0 : fn maybe_configured_reconcile_shard(
7355 0 : &self,
7356 0 : shard: &mut TenantShard,
7357 0 : nodes: &Arc<HashMap<NodeId, Node>>,
7358 0 : reconciler_config: ReconcilerConfig,
7359 0 : ) -> Option<ReconcilerWaiter> {
7360 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
7361 :
7362 0 : let reconcile_reason = match reconcile_needed {
7363 0 : ReconcileNeeded::No => return None,
7364 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
7365 0 : ReconcileNeeded::Yes(reason) => {
7366 0 : // Fall through to try and acquire units for spawning reconciler
7367 0 : reason
7368 : }
7369 : };
7370 :
7371 0 : let units = match self.get_reconciler_units(reconciler_config.priority) {
7372 0 : Ok(u) => u,
7373 : Err(_) => {
7374 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
7375 0 : "Concurrency limited: enqueued for reconcile later");
7376 0 : if !shard.delayed_reconcile {
7377 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
7378 0 : Err(TrySendError::Closed(_)) => {
7379 0 : // Weird mid-shutdown case?
7380 0 : }
7381 : Err(TrySendError::Full(_)) => {
7382 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
7383 0 : tracing::warn!(
7384 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
7385 : );
7386 : }
7387 0 : Ok(()) => {
7388 0 : shard.delayed_reconcile = true;
7389 0 : }
7390 : }
7391 0 : }
7392 :
7393 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
7394 : // number to advance. When this function is eventually called again and succeeds in getting units,
7395 : // it will spawn a reconciler that makes this waiter complete.
7396 0 : return Some(shard.future_reconcile_waiter());
7397 : }
7398 : };
7399 :
7400 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
7401 : // Gate closed: we're shutting down, drop out.
7402 0 : return None;
7403 : };
7404 :
7405 0 : shard.spawn_reconciler(
7406 0 : reconcile_reason,
7407 0 : &self.result_tx,
7408 0 : nodes,
7409 0 : &self.compute_hook,
7410 0 : reconciler_config,
7411 0 : &self.config,
7412 0 : &self.persistence,
7413 0 : units,
7414 0 : gate_guard,
7415 0 : &self.reconcilers_cancel,
7416 0 : )
7417 0 : }
7418 :
7419 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
7420 : /// Additionally, reschedule tenants that require it.
7421 : ///
7422 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
7423 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
7424 : /// available. A return value of 0 indicates that everything is fully reconciled already.
7425 0 : fn reconcile_all(&self) -> usize {
7426 0 : let mut locked = self.inner.write().unwrap();
7427 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7428 0 : let pageservers = nodes.clone();
7429 0 :
7430 0 : // This function is an efficient place to update lazy statistics, since we are walking
7431 0 : // all tenants.
7432 0 : let mut pending_reconciles = 0;
7433 0 : let mut az_violations = 0;
7434 0 :
7435 0 : // If we find any tenants to drop from memory, stash them to offload after
7436 0 : // we're done traversing the map of tenants.
7437 0 : let mut drop_detached_tenants = Vec::new();
7438 0 :
7439 0 : let mut reconciles_spawned = 0;
7440 0 : for shard in tenants.values_mut() {
7441 : // Accumulate scheduling statistics
7442 0 : if let (Some(attached), Some(preferred)) =
7443 0 : (shard.intent.get_attached(), shard.preferred_az())
7444 : {
7445 0 : let node_az = nodes
7446 0 : .get(attached)
7447 0 : .expect("Nodes exist if referenced")
7448 0 : .get_availability_zone_id();
7449 0 : if node_az != preferred {
7450 0 : az_violations += 1;
7451 0 : }
7452 0 : }
7453 :
7454 : // Skip checking if this shard is already enqueued for reconciliation
7455 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
7456 : // If there is something delayed, then return a nonzero count so that
7457 : // callers like reconcile_all_now do not incorrectly get the impression
7458 : // that the system is in a quiescent state.
7459 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
7460 0 : pending_reconciles += 1;
7461 0 : continue;
7462 0 : }
7463 0 :
7464 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
7465 0 : // dirty, spawn another rone
7466 0 : if self
7467 0 : .maybe_reconcile_shard(shard, &pageservers, ReconcilerPriority::Normal)
7468 0 : .is_some()
7469 0 : {
7470 0 : reconciles_spawned += 1;
7471 0 : } else if shard.delayed_reconcile {
7472 0 : // Shard wanted to reconcile but for some reason couldn't.
7473 0 : pending_reconciles += 1;
7474 0 : }
7475 :
7476 : // If this tenant is detached, try dropping it from memory. This is usually done
7477 : // proactively in [`Self::process_results`], but we do it here to handle the edge
7478 : // case where a reconcile completes while someone else is holding an op lock for the tenant.
7479 0 : if shard.tenant_shard_id.shard_number == ShardNumber(0)
7480 0 : && shard.policy == PlacementPolicy::Detached
7481 : {
7482 0 : if let Some(guard) = self.tenant_op_locks.try_exclusive(
7483 0 : shard.tenant_shard_id.tenant_id,
7484 0 : TenantOperations::DropDetached,
7485 0 : ) {
7486 0 : drop_detached_tenants.push((shard.tenant_shard_id.tenant_id, guard));
7487 0 : }
7488 0 : }
7489 : }
7490 :
7491 : // Some metrics are calculated from SchedulerNode state, update these periodically
7492 0 : scheduler.update_metrics();
7493 :
7494 : // Process any deferred tenant drops
7495 0 : for (tenant_id, guard) in drop_detached_tenants {
7496 0 : self.maybe_drop_tenant(tenant_id, &mut locked, &guard);
7497 0 : }
7498 :
7499 0 : metrics::METRICS_REGISTRY
7500 0 : .metrics_group
7501 0 : .storage_controller_schedule_az_violation
7502 0 : .set(az_violations as i64);
7503 0 :
7504 0 : metrics::METRICS_REGISTRY
7505 0 : .metrics_group
7506 0 : .storage_controller_pending_reconciles
7507 0 : .set(pending_reconciles as i64);
7508 0 :
7509 0 : reconciles_spawned
7510 0 : }
7511 :
7512 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
7513 : /// could be scheduled somewhere better:
7514 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
7515 : /// * e.g. after a node fails then recovers, to move some work back to it
7516 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
7517 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
7518 : /// we did the split, but are probably better placed elsewhere.
7519 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
7520 : /// * e.g. after a shard split, some locations will be on the same node (where the split
7521 : /// happened), and will probably be better placed elsewhere.
7522 : ///
7523 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
7524 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
7525 : /// according to those same soft constraints.
7526 0 : async fn optimize_all(&self) -> usize {
7527 : // Limit on how many shards' optmizations each call to this function will execute. Combined
7528 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
7529 : // trickle of optimizations in the background, rather than executing a large number in parallel
7530 : // when a change occurs.
7531 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 16;
7532 :
7533 : // Synchronous prepare: scan shards for possible scheduling optimizations
7534 0 : let candidate_work = self.optimize_all_plan();
7535 0 : let candidate_work_len = candidate_work.len();
7536 :
7537 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
7538 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
7539 :
7540 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
7541 0 :
7542 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
7543 0 : let mut reconciles_spawned = 0;
7544 0 : let mut optimizations_applied = 0;
7545 0 : let mut locked = self.inner.write().unwrap();
7546 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
7547 0 : for (tenant_shard_id, optimization) in validated_work {
7548 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
7549 : // Shard was dropped between planning and execution;
7550 0 : continue;
7551 : };
7552 0 : tracing::info!(tenant_shard_id=%tenant_shard_id, "Applying optimization: {optimization:?}");
7553 0 : if shard.apply_optimization(scheduler, optimization) {
7554 0 : optimizations_applied += 1;
7555 0 : if self
7556 0 : .maybe_reconcile_shard(shard, nodes, ReconcilerPriority::Normal)
7557 0 : .is_some()
7558 0 : {
7559 0 : reconciles_spawned += 1;
7560 0 : }
7561 0 : }
7562 :
7563 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
7564 0 : break;
7565 0 : }
7566 : }
7567 :
7568 0 : if was_work_filtered {
7569 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
7570 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
7571 0 : // as these validations start passing.
7572 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
7573 0 : }
7574 :
7575 0 : reconciles_spawned
7576 0 : }
7577 :
7578 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
7579 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
7580 : // this higher than the execution limit gives us a chance to execute some work even if the first
7581 : // few optimizations we find are not ready.
7582 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 64;
7583 :
7584 0 : let mut work = Vec::new();
7585 0 : let mut locked = self.inner.write().unwrap();
7586 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
7587 :
7588 : // We are going to plan a bunch of optimisations before applying any of them, so the
7589 : // utilisation stats on nodes will be effectively stale for the >1st optimisation we
7590 : // generate. To avoid this causing unstable migrations/flapping, it's important that the
7591 : // code in TenantShard for finding optimisations uses [`NodeAttachmentSchedulingScore::disregard_utilization`]
7592 : // to ignore the utilisation component of the score.
7593 :
7594 0 : for (_tenant_id, schedule_context, shards) in
7595 0 : TenantShardContextIterator::new(tenants, ScheduleMode::Speculative)
7596 : {
7597 0 : for shard in shards {
7598 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
7599 0 : break;
7600 0 : }
7601 0 : match shard.get_scheduling_policy() {
7602 0 : ShardSchedulingPolicy::Active => {
7603 0 : // Ok to do optimization
7604 0 : }
7605 0 : ShardSchedulingPolicy::Essential if shard.get_preferred_node().is_some() => {
7606 0 : // Ok to do optimization: we are executing a graceful migration that
7607 0 : // has set preferred_node
7608 0 : }
7609 : ShardSchedulingPolicy::Essential
7610 : | ShardSchedulingPolicy::Pause
7611 : | ShardSchedulingPolicy::Stop => {
7612 : // Policy prevents optimizing this shard.
7613 0 : continue;
7614 : }
7615 : }
7616 :
7617 0 : if !matches!(shard.splitting, SplitState::Idle)
7618 0 : || matches!(shard.policy, PlacementPolicy::Detached)
7619 0 : || shard.reconciler.is_some()
7620 : {
7621 : // Do not start any optimizations while another change to the tenant is ongoing: this
7622 : // is not necessary for correctness, but simplifies operations and implicitly throttles
7623 : // optimization changes to happen in a "trickle" over time.
7624 0 : continue;
7625 0 : }
7626 0 :
7627 0 : // Fast path: we may quickly identify shards that don't have any possible optimisations
7628 0 : if !shard.maybe_optimizable(scheduler, &schedule_context) {
7629 0 : if cfg!(feature = "testing") {
7630 : // Check that maybe_optimizable doesn't disagree with the actual optimization functions.
7631 : // Only do this in testing builds because it is not a correctness-critical check, so we shouldn't
7632 : // panic in prod if we hit this, or spend cycles on it in prod.
7633 0 : assert!(
7634 0 : shard
7635 0 : .optimize_attachment(scheduler, &schedule_context)
7636 0 : .is_none()
7637 0 : );
7638 0 : assert!(
7639 0 : shard
7640 0 : .optimize_secondary(scheduler, &schedule_context)
7641 0 : .is_none()
7642 0 : );
7643 0 : }
7644 0 : continue;
7645 0 : }
7646 :
7647 0 : if let Some(optimization) =
7648 : // If idle, maybe optimize attachments: if a shard has a secondary location that is preferable to
7649 : // its primary location based on soft constraints, cut it over.
7650 0 : shard.optimize_attachment(scheduler, &schedule_context)
7651 : {
7652 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for attachment: {optimization:?}");
7653 0 : work.push((shard.tenant_shard_id, optimization));
7654 0 : break;
7655 0 : } else if let Some(optimization) =
7656 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
7657 : // better placed on another node, based on ScheduleContext, then adjust it. This
7658 : // covers cases like after a shard split, where we might have too many shards
7659 : // in the same tenant with secondary locations on the node where they originally split.
7660 0 : shard.optimize_secondary(scheduler, &schedule_context)
7661 : {
7662 0 : tracing::info!(tenant_shard_id=%shard.tenant_shard_id, "Identified optimization for secondary: {optimization:?}");
7663 0 : work.push((shard.tenant_shard_id, optimization));
7664 0 : break;
7665 0 : }
7666 : }
7667 : }
7668 :
7669 0 : work
7670 0 : }
7671 :
7672 0 : async fn optimize_all_validate(
7673 0 : &self,
7674 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
7675 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
7676 0 : // Take a clone of the node map to use outside the lock in async validation phase
7677 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
7678 0 :
7679 0 : let mut want_secondary_status = Vec::new();
7680 0 :
7681 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
7682 0 : // check that the state of locations is acceptable to run the optimization, such as
7683 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
7684 0 : // in a live migration.
7685 0 : let mut validated_work = Vec::new();
7686 0 : for (tenant_shard_id, optimization) in candidate_work {
7687 0 : match optimization.action {
7688 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
7689 : old_attached_node_id: _,
7690 0 : new_attached_node_id,
7691 0 : }) => {
7692 0 : match validation_nodes.get(&new_attached_node_id) {
7693 0 : None => {
7694 0 : // Node was dropped between planning and validation
7695 0 : }
7696 0 : Some(node) => {
7697 0 : if !node.is_available() {
7698 0 : tracing::info!(
7699 0 : "Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable"
7700 : );
7701 0 : } else {
7702 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
7703 0 : // remote API requests concurrently.
7704 0 : want_secondary_status.push((
7705 0 : tenant_shard_id,
7706 0 : node.clone(),
7707 0 : optimization,
7708 0 : ));
7709 0 : }
7710 : }
7711 : }
7712 : }
7713 : ScheduleOptimizationAction::ReplaceSecondary(_)
7714 : | ScheduleOptimizationAction::CreateSecondary(_)
7715 : | ScheduleOptimizationAction::RemoveSecondary(_) => {
7716 : // No extra checks needed to manage secondaries: this does not interrupt client access
7717 0 : validated_work.push((tenant_shard_id, optimization))
7718 : }
7719 : };
7720 : }
7721 :
7722 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
7723 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
7724 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
7725 0 : let results = self
7726 0 : .tenant_for_shards_api(
7727 0 : want_secondary_status
7728 0 : .iter()
7729 0 : .map(|i| (i.0, i.1.clone()))
7730 0 : .collect(),
7731 0 : |tenant_shard_id, client| async move {
7732 0 : client.tenant_secondary_status(tenant_shard_id).await
7733 0 : },
7734 0 : 1,
7735 0 : 1,
7736 0 : SHORT_RECONCILE_TIMEOUT,
7737 0 : &self.cancel,
7738 0 : )
7739 0 : .await;
7740 :
7741 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
7742 0 : want_secondary_status.into_iter().zip(results.into_iter())
7743 : {
7744 0 : match secondary_status {
7745 0 : Err(e) => {
7746 0 : tracing::info!(
7747 0 : "Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}"
7748 : );
7749 : }
7750 0 : Ok(progress) => {
7751 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
7752 : // them in an optimization
7753 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
7754 :
7755 0 : if progress.heatmap_mtime.is_none()
7756 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
7757 0 : && progress.bytes_downloaded != progress.bytes_total
7758 0 : || progress.bytes_total - progress.bytes_downloaded
7759 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
7760 : {
7761 0 : tracing::info!(
7762 0 : "Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}"
7763 : );
7764 :
7765 : #[cfg(feature = "testing")]
7766 0 : if progress.heatmap_mtime.is_none() {
7767 : // No heatmap might mean the attached location has never uploaded one, or that
7768 : // the secondary download hasn't happened yet. This is relatively unusual in the field,
7769 : // but fairly common in tests.
7770 0 : self.kick_secondary_download(tenant_shard_id).await;
7771 0 : }
7772 : } else {
7773 : // Location looks ready: proceed
7774 0 : tracing::info!(
7775 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
7776 : );
7777 0 : validated_work.push((tenant_shard_id, optimization))
7778 : }
7779 : }
7780 : }
7781 : }
7782 :
7783 0 : validated_work
7784 0 : }
7785 :
7786 : /// Some aspects of scheduling optimisation wait for secondary locations to be warm. This
7787 : /// happens on multi-minute timescales in the field, which is fine because optimisation is meant
7788 : /// to be a lazy background thing. However, when testing, it is not practical to wait around, so
7789 : /// we have this helper to move things along faster.
7790 : #[cfg(feature = "testing")]
7791 0 : async fn kick_secondary_download(&self, tenant_shard_id: TenantShardId) {
7792 0 : let (attached_node, secondaries) = {
7793 0 : let locked = self.inner.read().unwrap();
7794 0 : let Some(shard) = locked.tenants.get(&tenant_shard_id) else {
7795 0 : tracing::warn!(
7796 0 : "Skipping kick of secondary download for {tenant_shard_id}: not found"
7797 : );
7798 0 : return;
7799 : };
7800 :
7801 0 : let Some(attached) = shard.intent.get_attached() else {
7802 0 : tracing::warn!(
7803 0 : "Skipping kick of secondary download for {tenant_shard_id}: no attached"
7804 : );
7805 0 : return;
7806 : };
7807 :
7808 0 : let secondaries = shard
7809 0 : .intent
7810 0 : .get_secondary()
7811 0 : .iter()
7812 0 : .map(|n| locked.nodes.get(n).unwrap().clone())
7813 0 : .collect::<Vec<_>>();
7814 0 :
7815 0 : (locked.nodes.get(attached).unwrap().clone(), secondaries)
7816 0 : };
7817 0 :
7818 0 : // Make remote API calls to upload + download heatmaps: we ignore errors because this is just
7819 0 : // a 'kick' to let scheduling optimisation run more promptly.
7820 0 : match attached_node
7821 0 : .with_client_retries(
7822 0 : |client| async move { client.tenant_heatmap_upload(tenant_shard_id).await },
7823 0 : &self.config.pageserver_jwt_token,
7824 0 : &self.config.ssl_ca_cert,
7825 0 : 3,
7826 0 : 10,
7827 0 : SHORT_RECONCILE_TIMEOUT,
7828 0 : &self.cancel,
7829 0 : )
7830 0 : .await
7831 : {
7832 0 : Some(Err(e)) => {
7833 0 : tracing::info!(
7834 0 : "Failed to upload heatmap from {attached_node} for {tenant_shard_id}: {e}"
7835 : );
7836 : }
7837 : None => {
7838 0 : tracing::info!(
7839 0 : "Cancelled while uploading heatmap from {attached_node} for {tenant_shard_id}"
7840 : );
7841 : }
7842 : Some(Ok(_)) => {
7843 0 : tracing::info!(
7844 0 : "Successfully uploaded heatmap from {attached_node} for {tenant_shard_id}"
7845 : );
7846 : }
7847 : }
7848 :
7849 0 : for secondary_node in secondaries {
7850 0 : match secondary_node
7851 0 : .with_client_retries(
7852 0 : |client| async move {
7853 0 : client
7854 0 : .tenant_secondary_download(
7855 0 : tenant_shard_id,
7856 0 : Some(Duration::from_secs(1)),
7857 0 : )
7858 0 : .await
7859 0 : },
7860 0 : &self.config.pageserver_jwt_token,
7861 0 : &self.config.ssl_ca_cert,
7862 0 : 3,
7863 0 : 10,
7864 0 : SHORT_RECONCILE_TIMEOUT,
7865 0 : &self.cancel,
7866 0 : )
7867 0 : .await
7868 : {
7869 0 : Some(Err(e)) => {
7870 0 : tracing::info!(
7871 0 : "Failed to download heatmap from {secondary_node} for {tenant_shard_id}: {e}"
7872 : );
7873 : }
7874 : None => {
7875 0 : tracing::info!(
7876 0 : "Cancelled while downloading heatmap from {secondary_node} for {tenant_shard_id}"
7877 : );
7878 : }
7879 0 : Some(Ok(progress)) => {
7880 0 : tracing::info!(
7881 0 : "Successfully downloaded heatmap from {secondary_node} for {tenant_shard_id}: {progress:?}"
7882 : );
7883 : }
7884 : }
7885 : }
7886 0 : }
7887 :
7888 : /// Asynchronously split a tenant that's eligible for automatic splits:
7889 : ///
7890 : /// * The tenant is unsharded.
7891 : /// * The logical size of its largest timeline exceeds split_threshold.
7892 : /// * The tenant's scheduling policy is active.
7893 : ///
7894 : /// At most one tenant will be split per call: the one with the largest max logical size. It
7895 : /// will split 1 → 8 shards.
7896 : ///
7897 : /// TODO: consider splitting based on total logical size rather than max logical size.
7898 : ///
7899 : /// TODO: consider spawning multiple splits in parallel: this is only called once every 20
7900 : /// seconds, so a large backlog can take a long time, and if a tenant fails to split it will
7901 : /// block all other splits.
7902 0 : async fn autosplit_tenants(self: &Arc<Self>) {
7903 0 : let Some(split_threshold) = self.config.split_threshold else {
7904 0 : return; // auto-splits are disabled
7905 : };
7906 0 : if split_threshold == 0 {
7907 0 : return;
7908 0 : }
7909 :
7910 : // Fetch the largest eligible shards by logical size.
7911 : const MAX_SHARDS: ShardCount = ShardCount::new(8);
7912 :
7913 0 : let mut top_n = self
7914 0 : .get_top_tenant_shards(&TopTenantShardsRequest {
7915 0 : order_by: TenantSorting::MaxLogicalSize,
7916 0 : limit: 10,
7917 0 : where_shards_lt: Some(MAX_SHARDS),
7918 0 : where_gt: Some(split_threshold),
7919 0 : })
7920 0 : .await;
7921 :
7922 : // Filter out tenants in a prohibiting scheduling mode.
7923 0 : {
7924 0 : let state = self.inner.read().unwrap();
7925 0 : top_n.retain(|i| {
7926 0 : let policy = state.tenants.get(&i.id).map(|s| s.get_scheduling_policy());
7927 0 : policy == Some(ShardSchedulingPolicy::Active)
7928 0 : });
7929 0 : }
7930 :
7931 0 : let Some(split_candidate) = top_n.into_iter().next() else {
7932 0 : debug!("No split-elegible shards found");
7933 0 : return;
7934 : };
7935 :
7936 : // We spawn a task to run this, so it's exactly like some external API client requesting it.
7937 : // We don't want to block the background reconcile loop on this.
7938 0 : info!(
7939 0 : "Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}"
7940 : );
7941 :
7942 0 : let this = self.clone();
7943 0 : tokio::spawn(
7944 0 : async move {
7945 0 : match this
7946 0 : .tenant_shard_split(
7947 0 : split_candidate.id.tenant_id,
7948 0 : TenantShardSplitRequest {
7949 0 : // Always split to the max number of shards: this avoids stepping
7950 0 : // through intervening shard counts and encountering the overhead of a
7951 0 : // split+cleanup each time as a tenant grows, and is not too expensive
7952 0 : // because our max shard count is relatively low anyway. This policy
7953 0 : // will be adjusted in future once we support higher shard count.
7954 0 : new_shard_count: MAX_SHARDS.literal(),
7955 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
7956 0 : },
7957 0 : )
7958 0 : .await
7959 : {
7960 0 : Ok(_) => info!("Successful auto-split"),
7961 0 : Err(err) => error!("Auto-split failed: {err}"),
7962 : }
7963 0 : }
7964 0 : .instrument(info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
7965 : );
7966 0 : }
7967 :
7968 : /// Fetches the top tenant shards from every node, in descending order of
7969 : /// max logical size. Any node errors will be logged and ignored.
7970 0 : async fn get_top_tenant_shards(
7971 0 : &self,
7972 0 : request: &TopTenantShardsRequest,
7973 0 : ) -> Vec<TopTenantShardItem> {
7974 0 : let nodes = self
7975 0 : .inner
7976 0 : .read()
7977 0 : .unwrap()
7978 0 : .nodes
7979 0 : .values()
7980 0 : .cloned()
7981 0 : .collect_vec();
7982 0 :
7983 0 : let mut futures = FuturesUnordered::new();
7984 0 : for node in nodes {
7985 0 : futures.push(async move {
7986 0 : node.with_client_retries(
7987 0 : |client| async move { client.top_tenant_shards(request.clone()).await },
7988 0 : &self.config.pageserver_jwt_token,
7989 0 : &self.config.ssl_ca_cert,
7990 0 : 3,
7991 0 : 3,
7992 0 : Duration::from_secs(5),
7993 0 : &self.cancel,
7994 0 : )
7995 0 : .await
7996 0 : });
7997 0 : }
7998 :
7999 0 : let mut top = Vec::new();
8000 0 : while let Some(output) = futures.next().await {
8001 0 : match output {
8002 0 : Some(Ok(response)) => top.extend(response.shards),
8003 0 : Some(Err(mgmt_api::Error::Cancelled)) => {}
8004 0 : Some(Err(err)) => warn!("failed to fetch top tenants: {err}"),
8005 0 : None => {} // node is shutting down
8006 : }
8007 : }
8008 :
8009 0 : top.sort_by_key(|i| i.max_logical_size);
8010 0 : top.reverse();
8011 0 : top
8012 0 : }
8013 :
8014 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
8015 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
8016 : /// put the system into a quiescent state where future background reconciliations won't do anything.
8017 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
8018 0 : let reconciles_spawned = self.reconcile_all();
8019 0 : let reconciles_spawned = if reconciles_spawned == 0 {
8020 : // Only optimize when we are otherwise idle
8021 0 : self.optimize_all().await
8022 : } else {
8023 0 : reconciles_spawned
8024 : };
8025 :
8026 0 : let waiters = {
8027 0 : let mut waiters = Vec::new();
8028 0 : let locked = self.inner.read().unwrap();
8029 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
8030 0 : if let Some(waiter) = shard.get_waiter() {
8031 0 : waiters.push(waiter);
8032 0 : }
8033 : }
8034 0 : waiters
8035 0 : };
8036 0 :
8037 0 : let waiter_count = waiters.len();
8038 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
8039 0 : Ok(()) => {}
8040 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
8041 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
8042 0 : {
8043 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
8044 0 : // because some other change superceded it. We will return a nonzero number,
8045 0 : // so the caller knows they might have to call again to quiesce the system.
8046 0 : }
8047 0 : Err(e) => {
8048 0 : return Err(e);
8049 : }
8050 : };
8051 :
8052 0 : tracing::info!(
8053 0 : "{} reconciles in reconcile_all, {} waiters",
8054 : reconciles_spawned,
8055 : waiter_count
8056 : );
8057 :
8058 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
8059 0 : }
8060 :
8061 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
8062 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
8063 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
8064 0 : self.reconcilers_cancel.cancel();
8065 0 : self.reconcilers_gate.close().await;
8066 :
8067 : // Signal the background loop in [`Service::process_results`] to exit once
8068 : // it has proccessed the results from all the reconciles we cancelled earlier.
8069 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
8070 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
8071 0 : self.result_tx.closed().await;
8072 0 : }
8073 :
8074 0 : pub async fn shutdown(&self) {
8075 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
8076 0 : .await;
8077 :
8078 : // Background tasks hold gate guards: this notifies them of the cancellation and
8079 : // waits for them all to complete.
8080 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
8081 0 : self.cancel.cancel();
8082 0 : self.gate.close().await;
8083 0 : }
8084 :
8085 : /// Spot check the download lag for a secondary location of a shard.
8086 : /// Should be used as a heuristic, since it's not always precise: the
8087 : /// secondary might have not downloaded the new heat map yet and, hence,
8088 : /// is not aware of the lag.
8089 : ///
8090 : /// Returns:
8091 : /// * Ok(None) if the lag could not be determined from the status,
8092 : /// * Ok(Some(_)) if the lag could be determind
8093 : /// * Err on failures to query the pageserver.
8094 0 : async fn secondary_lag(
8095 0 : &self,
8096 0 : secondary: &NodeId,
8097 0 : tenant_shard_id: TenantShardId,
8098 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
8099 0 : let nodes = self.inner.read().unwrap().nodes.clone();
8100 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
8101 0 : StatusCode::NOT_FOUND,
8102 0 : format!("Node with id {} not found", secondary),
8103 0 : ))?;
8104 :
8105 0 : match node
8106 0 : .with_client_retries(
8107 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
8108 0 : &self.config.pageserver_jwt_token,
8109 0 : &self.config.ssl_ca_cert,
8110 0 : 1,
8111 0 : 3,
8112 0 : Duration::from_millis(250),
8113 0 : &self.cancel,
8114 0 : )
8115 0 : .await
8116 : {
8117 0 : Some(Ok(status)) => match status.heatmap_mtime {
8118 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
8119 0 : None => Ok(None),
8120 : },
8121 0 : Some(Err(e)) => Err(e),
8122 0 : None => Err(mgmt_api::Error::Cancelled),
8123 : }
8124 0 : }
8125 :
8126 : /// Drain a node by moving the shards attached to it as primaries.
8127 : /// This is a long running operation and it should run as a separate Tokio task.
8128 0 : pub(crate) async fn drain_node(
8129 0 : self: &Arc<Self>,
8130 0 : node_id: NodeId,
8131 0 : cancel: CancellationToken,
8132 0 : ) -> Result<(), OperationError> {
8133 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
8134 0 : let max_secondary_lag_bytes = self
8135 0 : .config
8136 0 : .max_secondary_lag_bytes
8137 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
8138 :
8139 : // By default, live migrations are generous about the wait time for getting
8140 : // the secondary location up to speed. When draining, give up earlier in order
8141 : // to not stall the operation when a cold secondary is encountered.
8142 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
8143 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
8144 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
8145 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
8146 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
8147 0 : .build();
8148 0 :
8149 0 : let mut waiters = Vec::new();
8150 0 :
8151 0 : let mut tid_iter = TenantShardIterator::new({
8152 0 : let service = self.clone();
8153 0 : move |last_inspected_shard: Option<TenantShardId>| {
8154 0 : let locked = &service.inner.read().unwrap();
8155 0 : let tenants = &locked.tenants;
8156 0 : let entry = match last_inspected_shard {
8157 0 : Some(skip_past) => {
8158 0 : // Skip to the last seen tenant shard id
8159 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
8160 0 :
8161 0 : // Skip past the last seen
8162 0 : cursor.nth(1)
8163 : }
8164 0 : None => tenants.first_key_value(),
8165 : };
8166 :
8167 0 : entry.map(|(tid, _)| tid).copied()
8168 0 : }
8169 0 : });
8170 :
8171 0 : while !tid_iter.finished() {
8172 0 : if cancel.is_cancelled() {
8173 0 : match self
8174 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8175 0 : .await
8176 : {
8177 0 : Ok(()) => return Err(OperationError::Cancelled),
8178 0 : Err(err) => {
8179 0 : return Err(OperationError::FinalizeError(
8180 0 : format!(
8181 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8182 0 : node_id, err
8183 0 : )
8184 0 : .into(),
8185 0 : ));
8186 : }
8187 : }
8188 0 : }
8189 0 :
8190 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
8191 :
8192 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
8193 0 : let tid = match tid_iter.next() {
8194 0 : Some(tid) => tid,
8195 : None => {
8196 0 : break;
8197 : }
8198 : };
8199 :
8200 0 : let tid_drain = TenantShardDrain {
8201 0 : drained_node: node_id,
8202 0 : tenant_shard_id: tid,
8203 0 : };
8204 :
8205 0 : let dest_node_id = {
8206 0 : let locked = self.inner.read().unwrap();
8207 0 :
8208 0 : match tid_drain
8209 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
8210 : {
8211 0 : Some(node_id) => node_id,
8212 : None => {
8213 0 : continue;
8214 : }
8215 : }
8216 : };
8217 :
8218 0 : match self.secondary_lag(&dest_node_id, tid).await {
8219 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
8220 0 : // The secondary is reasonably up to date.
8221 0 : // Migrate to it
8222 0 : }
8223 0 : Ok(Some(lag)) => {
8224 0 : tracing::info!(
8225 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8226 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
8227 : );
8228 0 : continue;
8229 : }
8230 : Ok(None) => {
8231 0 : tracing::info!(
8232 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8233 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
8234 : );
8235 0 : continue;
8236 : }
8237 0 : Err(err) => {
8238 0 : tracing::warn!(
8239 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8240 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
8241 : );
8242 0 : continue;
8243 : }
8244 : }
8245 :
8246 : {
8247 0 : let mut locked = self.inner.write().unwrap();
8248 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8249 0 : let rescheduled = tid_drain.reschedule_to_secondary(
8250 0 : dest_node_id,
8251 0 : tenants,
8252 0 : scheduler,
8253 0 : nodes,
8254 0 : )?;
8255 :
8256 0 : if let Some(tenant_shard) = rescheduled {
8257 0 : let waiter = self.maybe_configured_reconcile_shard(
8258 0 : tenant_shard,
8259 0 : nodes,
8260 0 : reconciler_config,
8261 0 : );
8262 0 : if let Some(some) = waiter {
8263 0 : waiters.push(some);
8264 0 : }
8265 0 : }
8266 : }
8267 : }
8268 :
8269 0 : waiters = self
8270 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
8271 0 : .await;
8272 :
8273 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
8274 : }
8275 :
8276 0 : while !waiters.is_empty() {
8277 0 : if cancel.is_cancelled() {
8278 0 : match self
8279 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8280 0 : .await
8281 : {
8282 0 : Ok(()) => return Err(OperationError::Cancelled),
8283 0 : Err(err) => {
8284 0 : return Err(OperationError::FinalizeError(
8285 0 : format!(
8286 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8287 0 : node_id, err
8288 0 : )
8289 0 : .into(),
8290 0 : ));
8291 : }
8292 : }
8293 0 : }
8294 0 :
8295 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
8296 :
8297 0 : waiters = self
8298 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
8299 0 : .await;
8300 : }
8301 :
8302 : // At this point we have done the best we could to drain shards from this node.
8303 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
8304 : // to complete the drain.
8305 0 : if let Err(err) = self
8306 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
8307 0 : .await
8308 : {
8309 : // This is not fatal. Anything that is polling the node scheduling policy to detect
8310 : // the end of the drain operations will hang, but all such places should enforce an
8311 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
8312 : // by the counterpart fill operation.
8313 0 : return Err(OperationError::FinalizeError(
8314 0 : format!(
8315 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
8316 0 : )
8317 0 : .into(),
8318 0 : ));
8319 0 : }
8320 0 :
8321 0 : Ok(())
8322 0 : }
8323 :
8324 : /// Create a node fill plan (pick secondaries to promote), based on:
8325 : /// 1. Shards which have a secondary on this node, and this node is in their home AZ, and are currently attached to a node
8326 : /// outside their home AZ, should be migrated back here.
8327 : /// 2. If after step 1 we have not migrated enough shards for this node to have its fair share of
8328 : /// attached shards, we will promote more shards from the nodes with the most attached shards, unless
8329 : /// those shards have a home AZ that doesn't match the node we're filling.
8330 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
8331 0 : let mut locked = self.inner.write().unwrap();
8332 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
8333 0 :
8334 0 : let node_az = nodes
8335 0 : .get(&node_id)
8336 0 : .expect("Node must exist")
8337 0 : .get_availability_zone_id()
8338 0 : .clone();
8339 0 :
8340 0 : // The tenant shard IDs that we plan to promote from secondary to attached on this node
8341 0 : let mut plan = Vec::new();
8342 0 :
8343 0 : // Collect shards which do not have a preferred AZ & are elegible for moving in stage 2
8344 0 : let mut free_tids_by_node: HashMap<NodeId, Vec<TenantShardId>> = HashMap::new();
8345 0 :
8346 0 : // Don't respect AZ preferences if there is only one AZ. This comes up in tests, but it could
8347 0 : // conceivably come up in real life if deploying a single-AZ region intentionally.
8348 0 : let respect_azs = nodes
8349 0 : .values()
8350 0 : .map(|n| n.get_availability_zone_id())
8351 0 : .unique()
8352 0 : .count()
8353 0 : > 1;
8354 :
8355 : // Step 1: collect all shards that we are required to migrate back to this node because their AZ preference
8356 : // requires it.
8357 0 : for (tsid, tenant_shard) in tenants {
8358 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
8359 : // Shard doesn't have a secondary on this node, ignore it.
8360 0 : continue;
8361 0 : }
8362 0 :
8363 0 : // AZ check: when filling nodes after a restart, our intent is to move _back_ the
8364 0 : // shards which belong on this node, not to promote shards whose scheduling preference
8365 0 : // would be on their currently attached node. So will avoid promoting shards whose
8366 0 : // home AZ doesn't match the AZ of the node we're filling.
8367 0 : match tenant_shard.preferred_az() {
8368 0 : _ if !respect_azs => {
8369 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8370 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
8371 0 : }
8372 : }
8373 : None => {
8374 : // Shard doesn't have an AZ preference: it is elegible to be moved, but we
8375 : // will only do so if our target shard count requires it.
8376 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8377 0 : free_tids_by_node.entry(*primary).or_default().push(*tsid);
8378 0 : }
8379 : }
8380 0 : Some(az) if az == &node_az => {
8381 : // This shard's home AZ is equal to the node we're filling: it should
8382 : // be moved back to this node as part of filling, unless its currently
8383 : // attached location is also in its home AZ.
8384 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
8385 0 : if nodes
8386 0 : .get(primary)
8387 0 : .expect("referenced node must exist")
8388 0 : .get_availability_zone_id()
8389 0 : != tenant_shard
8390 0 : .preferred_az()
8391 0 : .expect("tenant must have an AZ preference")
8392 : {
8393 0 : plan.push(*tsid)
8394 0 : }
8395 : } else {
8396 0 : plan.push(*tsid)
8397 : }
8398 : }
8399 0 : Some(_) => {
8400 0 : // This shard's home AZ is somewhere other than the node we're filling,
8401 0 : // it may not be moved back to this node as part of filling. Ignore it
8402 0 : }
8403 : }
8404 : }
8405 :
8406 : // Step 2: also promote any AZ-agnostic shards as required to achieve the target number of attachments
8407 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
8408 0 :
8409 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
8410 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
8411 0 :
8412 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
8413 :
8414 0 : for (node_id, attached) in nodes_by_load {
8415 0 : let available = locked.nodes.get(&node_id).is_some_and(|n| n.is_available());
8416 0 : if !available {
8417 0 : continue;
8418 0 : }
8419 0 :
8420 0 : if plan.len() >= fill_requirement
8421 0 : || free_tids_by_node.is_empty()
8422 0 : || attached <= expected_attached
8423 : {
8424 0 : break;
8425 0 : }
8426 0 :
8427 0 : let can_take = attached - expected_attached;
8428 0 : let needed = fill_requirement - plan.len();
8429 0 : let mut take = std::cmp::min(can_take, needed);
8430 0 :
8431 0 : let mut remove_node = false;
8432 0 : while take > 0 {
8433 0 : match free_tids_by_node.get_mut(&node_id) {
8434 0 : Some(tids) => match tids.pop() {
8435 0 : Some(tid) => {
8436 0 : let max_promote_for_tenant = std::cmp::max(
8437 0 : tid.shard_count.count() as usize / locked.nodes.len(),
8438 0 : 1,
8439 0 : );
8440 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
8441 0 : if *promoted < max_promote_for_tenant {
8442 0 : plan.push(tid);
8443 0 : *promoted += 1;
8444 0 : take -= 1;
8445 0 : }
8446 : }
8447 : None => {
8448 0 : remove_node = true;
8449 0 : break;
8450 : }
8451 : },
8452 : None => {
8453 0 : break;
8454 : }
8455 : }
8456 : }
8457 :
8458 0 : if remove_node {
8459 0 : free_tids_by_node.remove(&node_id);
8460 0 : }
8461 : }
8462 :
8463 0 : plan
8464 0 : }
8465 :
8466 : /// Fill a node by promoting its secondaries until the cluster is balanced
8467 : /// with regards to attached shard counts. Note that this operation only
8468 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
8469 : /// This is a long running operation and it should run as a separate Tokio task.
8470 0 : pub(crate) async fn fill_node(
8471 0 : &self,
8472 0 : node_id: NodeId,
8473 0 : cancel: CancellationToken,
8474 0 : ) -> Result<(), OperationError> {
8475 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
8476 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
8477 0 : let reconciler_config = ReconcilerConfigBuilder::new(ReconcilerPriority::Normal)
8478 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
8479 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
8480 0 : .build();
8481 0 :
8482 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
8483 0 : let mut waiters = Vec::new();
8484 :
8485 : // Execute the plan we've composed above. Before aplying each move from the plan,
8486 : // we validate to ensure that it has not gone stale in the meantime.
8487 0 : while !tids_to_promote.is_empty() {
8488 0 : if cancel.is_cancelled() {
8489 0 : match self
8490 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8491 0 : .await
8492 : {
8493 0 : Ok(()) => return Err(OperationError::Cancelled),
8494 0 : Err(err) => {
8495 0 : return Err(OperationError::FinalizeError(
8496 0 : format!(
8497 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8498 0 : node_id, err
8499 0 : )
8500 0 : .into(),
8501 0 : ));
8502 : }
8503 : }
8504 0 : }
8505 0 :
8506 0 : {
8507 0 : let mut locked = self.inner.write().unwrap();
8508 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
8509 :
8510 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
8511 0 : format!("node {node_id} was removed").into(),
8512 0 : ))?;
8513 :
8514 0 : let current_policy = node.get_scheduling();
8515 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
8516 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
8517 : // about it
8518 0 : return Err(OperationError::NodeStateChanged(
8519 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
8520 0 : ));
8521 0 : }
8522 :
8523 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
8524 0 : if let Some(tid) = tids_to_promote.pop() {
8525 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
8526 : // If the node being filled is not a secondary anymore,
8527 : // skip the promotion.
8528 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
8529 0 : continue;
8530 0 : }
8531 0 :
8532 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
8533 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
8534 0 : Err(e) => {
8535 0 : tracing::warn!(
8536 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8537 0 : "Scheduling error when filling pageserver {} : {e}", node_id
8538 : );
8539 : }
8540 : Ok(()) => {
8541 0 : tracing::info!(
8542 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
8543 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
8544 : node_id,
8545 : previously_attached_to,
8546 : node_id
8547 : );
8548 :
8549 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
8550 0 : tenant_shard,
8551 0 : nodes,
8552 0 : reconciler_config,
8553 0 : ) {
8554 0 : waiters.push(waiter);
8555 0 : }
8556 : }
8557 : }
8558 0 : }
8559 : } else {
8560 0 : break;
8561 : }
8562 : }
8563 : }
8564 :
8565 0 : waiters = self
8566 0 : .await_waiters_remainder(waiters, WAITER_FILL_DRAIN_POLL_TIMEOUT)
8567 0 : .await;
8568 : }
8569 :
8570 0 : while !waiters.is_empty() {
8571 0 : if cancel.is_cancelled() {
8572 0 : match self
8573 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8574 0 : .await
8575 : {
8576 0 : Ok(()) => return Err(OperationError::Cancelled),
8577 0 : Err(err) => {
8578 0 : return Err(OperationError::FinalizeError(
8579 0 : format!(
8580 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
8581 0 : node_id, err
8582 0 : )
8583 0 : .into(),
8584 0 : ));
8585 : }
8586 : }
8587 0 : }
8588 0 :
8589 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
8590 :
8591 0 : waiters = self
8592 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
8593 0 : .await;
8594 : }
8595 :
8596 0 : if let Err(err) = self
8597 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
8598 0 : .await
8599 : {
8600 : // This isn't a huge issue since the filling process starts upon request. However, it
8601 : // will prevent the next drain from starting. The only case in which this can fail
8602 : // is database unavailability. Such a case will require manual intervention.
8603 0 : return Err(OperationError::FinalizeError(
8604 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
8605 0 : .into(),
8606 0 : ));
8607 0 : }
8608 0 :
8609 0 : Ok(())
8610 0 : }
8611 :
8612 : /// Updates scrubber metadata health check results.
8613 0 : pub(crate) async fn metadata_health_update(
8614 0 : &self,
8615 0 : update_req: MetadataHealthUpdateRequest,
8616 0 : ) -> Result<(), ApiError> {
8617 0 : let now = chrono::offset::Utc::now();
8618 0 : let (healthy_records, unhealthy_records) = {
8619 0 : let locked = self.inner.read().unwrap();
8620 0 : let healthy_records = update_req
8621 0 : .healthy_tenant_shards
8622 0 : .into_iter()
8623 0 : // Retain only health records associated with tenant shards managed by storage controller.
8624 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
8625 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
8626 0 : .collect();
8627 0 : let unhealthy_records = update_req
8628 0 : .unhealthy_tenant_shards
8629 0 : .into_iter()
8630 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
8631 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
8632 0 : .collect();
8633 0 :
8634 0 : (healthy_records, unhealthy_records)
8635 0 : };
8636 0 :
8637 0 : self.persistence
8638 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
8639 0 : .await?;
8640 0 : Ok(())
8641 0 : }
8642 :
8643 : /// Lists the tenant shards that has unhealthy metadata status.
8644 0 : pub(crate) async fn metadata_health_list_unhealthy(
8645 0 : &self,
8646 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
8647 0 : let result = self
8648 0 : .persistence
8649 0 : .list_unhealthy_metadata_health_records()
8650 0 : .await?
8651 0 : .iter()
8652 0 : .map(|p| p.get_tenant_shard_id().unwrap())
8653 0 : .collect();
8654 0 :
8655 0 : Ok(result)
8656 0 : }
8657 :
8658 : /// Lists the tenant shards that have not been scrubbed for some duration.
8659 0 : pub(crate) async fn metadata_health_list_outdated(
8660 0 : &self,
8661 0 : not_scrubbed_for: Duration,
8662 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
8663 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
8664 0 : let result = self
8665 0 : .persistence
8666 0 : .list_outdated_metadata_health_records(earlier)
8667 0 : .await?
8668 0 : .into_iter()
8669 0 : .map(|record| record.into())
8670 0 : .collect();
8671 0 : Ok(result)
8672 0 : }
8673 :
8674 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
8675 0 : self.inner.read().unwrap().get_leadership_status()
8676 0 : }
8677 :
8678 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
8679 0 : tracing::info!("Received step down request from peer");
8680 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
8681 :
8682 0 : self.inner.write().unwrap().step_down();
8683 0 : // TODO: would it make sense to have a time-out for this?
8684 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
8685 0 : .await;
8686 :
8687 0 : let mut global_observed = GlobalObservedState::default();
8688 0 : let locked = self.inner.read().unwrap();
8689 0 : for (tid, tenant_shard) in locked.tenants.iter() {
8690 0 : global_observed
8691 0 : .0
8692 0 : .insert(*tid, tenant_shard.observed.clone());
8693 0 : }
8694 :
8695 0 : global_observed
8696 0 : }
8697 :
8698 : /// Choose safekeepers for the new timeline: 3 in different azs.
8699 0 : pub(crate) async fn safekeepers_for_new_timeline(
8700 0 : &self,
8701 0 : ) -> Result<Vec<SafekeeperInfo>, ApiError> {
8702 0 : let mut all_safekeepers = {
8703 0 : let locked = self.inner.read().unwrap();
8704 0 : locked
8705 0 : .safekeepers
8706 0 : .iter()
8707 0 : .filter_map(|sk| {
8708 0 : if sk.1.scheduling_policy() != SkSchedulingPolicy::Active {
8709 : // If we don't want to schedule stuff onto the safekeeper, respect that.
8710 0 : return None;
8711 0 : }
8712 0 : let utilization_opt = if let SafekeeperState::Available {
8713 : last_seen_at: _,
8714 0 : utilization,
8715 0 : } = sk.1.availability()
8716 : {
8717 0 : Some(utilization)
8718 : } else {
8719 : // non-available safekeepers still get a chance for new timelines,
8720 : // but put them last in the list.
8721 0 : None
8722 : };
8723 0 : let info = SafekeeperInfo {
8724 0 : hostname: sk.1.skp.host.clone(),
8725 0 : id: NodeId(sk.1.skp.id as u64),
8726 0 : };
8727 0 : Some((utilization_opt, info, sk.1.skp.availability_zone_id.clone()))
8728 0 : })
8729 0 : .collect::<Vec<_>>()
8730 0 : };
8731 0 : all_safekeepers.sort_by_key(|sk| {
8732 0 : (
8733 0 : sk.0.as_ref()
8734 0 : .map(|ut| ut.timeline_count)
8735 0 : .unwrap_or(u64::MAX),
8736 0 : // Use the id to decide on equal scores for reliability
8737 0 : sk.1.id.0,
8738 0 : )
8739 0 : });
8740 0 : let mut sks = Vec::new();
8741 0 : let mut azs = HashSet::new();
8742 0 : for (_sk_util, sk_info, az_id) in all_safekeepers.iter() {
8743 0 : if !azs.insert(az_id) {
8744 0 : continue;
8745 0 : }
8746 0 : sks.push(sk_info.clone());
8747 0 : if sks.len() == 3 {
8748 0 : break;
8749 0 : }
8750 : }
8751 0 : if sks.len() == 3 {
8752 0 : Ok(sks)
8753 : } else {
8754 0 : Err(ApiError::InternalServerError(anyhow::anyhow!(
8755 0 : "couldn't find three safekeepers in different AZs for new timeline"
8756 0 : )))
8757 : }
8758 0 : }
8759 :
8760 0 : pub(crate) async fn safekeepers_list(
8761 0 : &self,
8762 0 : ) -> Result<Vec<SafekeeperDescribeResponse>, DatabaseError> {
8763 0 : let locked = self.inner.read().unwrap();
8764 0 : let mut list = locked
8765 0 : .safekeepers
8766 0 : .iter()
8767 0 : .map(|sk| sk.1.describe_response())
8768 0 : .collect::<Result<Vec<_>, _>>()?;
8769 0 : list.sort_by_key(|v| v.id);
8770 0 : Ok(list)
8771 0 : }
8772 :
8773 0 : pub(crate) async fn get_safekeeper(
8774 0 : &self,
8775 0 : id: i64,
8776 0 : ) -> Result<SafekeeperDescribeResponse, DatabaseError> {
8777 0 : let locked = self.inner.read().unwrap();
8778 0 : let sk = locked
8779 0 : .safekeepers
8780 0 : .get(&NodeId(id as u64))
8781 0 : .ok_or(diesel::result::Error::NotFound)?;
8782 0 : sk.describe_response()
8783 0 : }
8784 :
8785 0 : pub(crate) async fn upsert_safekeeper(
8786 0 : &self,
8787 0 : record: crate::persistence::SafekeeperUpsert,
8788 0 : ) -> Result<(), ApiError> {
8789 0 : let node_id = NodeId(record.id as u64);
8790 0 : let use_https = self.config.use_https_safekeeper_api;
8791 0 :
8792 0 : if use_https && record.https_port.is_none() {
8793 0 : return Err(ApiError::PreconditionFailed(
8794 0 : format!(
8795 0 : "cannot upsert safekeeper {node_id}: \
8796 0 : https is enabled, but https port is not specified"
8797 0 : )
8798 0 : .into(),
8799 0 : ));
8800 0 : }
8801 0 :
8802 0 : self.persistence.safekeeper_upsert(record.clone()).await?;
8803 : {
8804 0 : let mut locked = self.inner.write().unwrap();
8805 0 : let mut safekeepers = (*locked.safekeepers).clone();
8806 0 : match safekeepers.entry(node_id) {
8807 0 : std::collections::hash_map::Entry::Occupied(mut entry) => entry
8808 0 : .get_mut()
8809 0 : .update_from_record(record)
8810 0 : .expect("all preconditions should be checked before upsert to database"),
8811 0 : std::collections::hash_map::Entry::Vacant(entry) => {
8812 0 : entry.insert(
8813 0 : Safekeeper::from_persistence(
8814 0 : crate::persistence::SafekeeperPersistence::from_upsert(
8815 0 : record,
8816 0 : SkSchedulingPolicy::Pause,
8817 0 : ),
8818 0 : CancellationToken::new(),
8819 0 : use_https,
8820 0 : )
8821 0 : .expect("all preconditions should be checked before upsert to database"),
8822 0 : );
8823 0 : }
8824 : }
8825 0 : locked.safekeepers = Arc::new(safekeepers);
8826 0 : }
8827 0 : Ok(())
8828 0 : }
8829 :
8830 0 : pub(crate) async fn set_safekeeper_scheduling_policy(
8831 0 : &self,
8832 0 : id: i64,
8833 0 : scheduling_policy: SkSchedulingPolicy,
8834 0 : ) -> Result<(), DatabaseError> {
8835 0 : self.persistence
8836 0 : .set_safekeeper_scheduling_policy(id, scheduling_policy)
8837 0 : .await?;
8838 0 : let node_id = NodeId(id as u64);
8839 0 : // After the change has been persisted successfully, update the in-memory state
8840 0 : {
8841 0 : let mut locked = self.inner.write().unwrap();
8842 0 : let mut safekeepers = (*locked.safekeepers).clone();
8843 0 : let sk = safekeepers
8844 0 : .get_mut(&node_id)
8845 0 : .ok_or(DatabaseError::Logical("Not found".to_string()))?;
8846 0 : sk.set_scheduling_policy(scheduling_policy);
8847 0 :
8848 0 : match scheduling_policy {
8849 0 : SkSchedulingPolicy::Active => (),
8850 0 : SkSchedulingPolicy::Decomissioned | SkSchedulingPolicy::Pause => {
8851 0 : locked.safekeeper_reconcilers.cancel_safekeeper(node_id);
8852 0 : }
8853 : }
8854 :
8855 0 : locked.safekeepers = Arc::new(safekeepers);
8856 0 : }
8857 0 : Ok(())
8858 0 : }
8859 :
8860 0 : pub(crate) async fn update_shards_preferred_azs(
8861 0 : &self,
8862 0 : req: ShardsPreferredAzsRequest,
8863 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
8864 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
8865 0 : let updated = self
8866 0 : .persistence
8867 0 : .set_tenant_shard_preferred_azs(preferred_azs)
8868 0 : .await
8869 0 : .map_err(|err| {
8870 0 : ApiError::InternalServerError(anyhow::anyhow!(
8871 0 : "Failed to persist preferred AZs: {err}"
8872 0 : ))
8873 0 : })?;
8874 :
8875 0 : let mut updated_in_mem_and_db = Vec::default();
8876 0 :
8877 0 : let mut locked = self.inner.write().unwrap();
8878 0 : let state = locked.deref_mut();
8879 0 : for (tid, az_id) in updated {
8880 0 : let shard = state.tenants.get_mut(&tid);
8881 0 : if let Some(shard) = shard {
8882 0 : shard.set_preferred_az(&mut state.scheduler, az_id);
8883 0 : updated_in_mem_and_db.push(tid);
8884 0 : }
8885 : }
8886 :
8887 0 : Ok(ShardsPreferredAzsResponse {
8888 0 : updated: updated_in_mem_and_db,
8889 0 : })
8890 0 : }
8891 : }
|