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