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