Line data Source code
1 : use hyper::Uri;
2 : use std::{
3 : borrow::Cow,
4 : cmp::Ordering,
5 : collections::{BTreeMap, HashMap, HashSet},
6 : ops::Deref,
7 : path::PathBuf,
8 : str::FromStr,
9 : sync::Arc,
10 : time::{Duration, Instant},
11 : };
12 :
13 : use crate::{
14 : background_node_operations::{
15 : Drain, Fill, Operation, OperationError, OperationHandler, MAX_RECONCILES_PER_OPERATION,
16 : },
17 : compute_hook::NotifyError,
18 : drain_utils::{self, TenantShardDrain, TenantShardIterator},
19 : id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, TracingExclusiveGuard},
20 : leadership::Leadership,
21 : metrics,
22 : peer_client::GlobalObservedState,
23 : persistence::{
24 : AbortShardSplitStatus, ControllerPersistence, DatabaseResult, MetadataHealthPersistence,
25 : ShardGenerationState, TenantFilter,
26 : },
27 : reconciler::{ReconcileError, ReconcileUnits, ReconcilerConfig, ReconcilerConfigBuilder},
28 : scheduler::{MaySchedule, ScheduleContext, ScheduleError, ScheduleMode},
29 : tenant_shard::{
30 : MigrateAttachment, ReconcileNeeded, ReconcilerStatus, ScheduleOptimization,
31 : ScheduleOptimizationAction,
32 : },
33 : };
34 : use anyhow::Context;
35 : use control_plane::storage_controller::{
36 : AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
37 : };
38 : use diesel::result::DatabaseErrorKind;
39 : use futures::{stream::FuturesUnordered, StreamExt};
40 : use itertools::Itertools;
41 : use pageserver_api::{
42 : controller_api::{
43 : MetadataHealthRecord, MetadataHealthUpdateRequest, NodeAvailability, NodeRegisterRequest,
44 : NodeSchedulingPolicy, NodeShard, NodeShardResponse, PlacementPolicy, ShardSchedulingPolicy,
45 : ShardsPreferredAzsRequest, ShardsPreferredAzsResponse, TenantCreateRequest,
46 : TenantCreateResponse, TenantCreateResponseShard, TenantDescribeResponse,
47 : TenantDescribeResponseShard, TenantLocateResponse, TenantPolicyRequest,
48 : TenantShardMigrateRequest, TenantShardMigrateResponse,
49 : },
50 : models::{
51 : SecondaryProgress, TenantConfigRequest, TimelineArchivalConfigRequest,
52 : TopTenantShardsRequest,
53 : },
54 : };
55 : use reqwest::StatusCode;
56 : use tracing::{instrument, Instrument};
57 :
58 : use crate::pageserver_client::PageserverClient;
59 : use pageserver_api::{
60 : models::{
61 : self, LocationConfig, LocationConfigListResponse, LocationConfigMode,
62 : PageserverUtilization, ShardParameters, TenantConfig, TenantLocationConfigRequest,
63 : TenantLocationConfigResponse, TenantShardLocation, TenantShardSplitRequest,
64 : TenantShardSplitResponse, TenantTimeTravelRequest, TimelineCreateRequest, TimelineInfo,
65 : },
66 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
67 : upcall_api::{
68 : ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
69 : ValidateResponse, ValidateResponseTenant,
70 : },
71 : };
72 : use pageserver_client::{mgmt_api, BlockUnblock};
73 : use tokio::sync::mpsc::error::TrySendError;
74 : use tokio_util::sync::CancellationToken;
75 : use utils::{
76 : completion::Barrier,
77 : failpoint_support,
78 : generation::Generation,
79 : http::error::ApiError,
80 : id::{NodeId, TenantId, TimelineId},
81 : sync::gate::Gate,
82 : };
83 :
84 : use crate::{
85 : compute_hook::ComputeHook,
86 : heartbeater::{Heartbeater, PageserverState},
87 : node::{AvailabilityTransition, Node},
88 : persistence::{split_state::SplitState, DatabaseError, Persistence, TenantShardPersistence},
89 : reconciler::attached_location_conf,
90 : scheduler::Scheduler,
91 : tenant_shard::{
92 : IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
93 : ReconcilerWaiter, TenantShard,
94 : },
95 : };
96 :
97 : pub mod chaos_injector;
98 :
99 : // For operations that should be quick, like attaching a new tenant
100 : const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
101 :
102 : // For operations that might be slow, like migrating a tenant with
103 : // some data in it.
104 : pub const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
105 :
106 : // If we receive a call using Secondary mode initially, it will omit generation. We will initialize
107 : // tenant shards into this generation, and as long as it remains in this generation, we will accept
108 : // input generation from future requests as authoritative.
109 : const INITIAL_GENERATION: Generation = Generation::new(0);
110 :
111 : /// How long [`Service::startup_reconcile`] is allowed to take before it should give
112 : /// up on unresponsive pageservers and proceed.
113 : pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
114 :
115 : /// How long a node may be unresponsive to heartbeats before we declare it offline.
116 : /// This must be long enough to cover node restarts as well as normal operations: in future
117 : pub const MAX_OFFLINE_INTERVAL_DEFAULT: Duration = Duration::from_secs(30);
118 :
119 : /// How long a node may be unresponsive to heartbeats during start up before we declare it
120 : /// offline.
121 : ///
122 : /// This is much more lenient than [`MAX_OFFLINE_INTERVAL_DEFAULT`] since the pageserver's
123 : /// handling of the re-attach response may take a long time and blocks heartbeats from
124 : /// being handled on the pageserver side.
125 : pub const MAX_WARMING_UP_INTERVAL_DEFAULT: Duration = Duration::from_secs(300);
126 :
127 : /// How often to send heartbeats to registered nodes?
128 : pub const HEARTBEAT_INTERVAL_DEFAULT: Duration = Duration::from_secs(5);
129 :
130 0 : #[derive(Clone, strum_macros::Display)]
131 : enum TenantOperations {
132 : Create,
133 : LocationConfig,
134 : ConfigSet,
135 : TimeTravelRemoteStorage,
136 : Delete,
137 : UpdatePolicy,
138 : ShardSplit,
139 : SecondaryDownload,
140 : TimelineCreate,
141 : TimelineDelete,
142 : AttachHook,
143 : TimelineArchivalConfig,
144 : TimelineDetachAncestor,
145 : TimelineGcBlockUnblock,
146 : }
147 :
148 0 : #[derive(Clone, strum_macros::Display)]
149 : enum NodeOperations {
150 : Register,
151 : Configure,
152 : Delete,
153 : }
154 :
155 : /// The leadership status for the storage controller process.
156 : /// Allowed transitions are:
157 : /// 1. Leader -> SteppedDown
158 : /// 2. Candidate -> Leader
159 : #[derive(
160 : Eq,
161 : PartialEq,
162 : Copy,
163 : Clone,
164 0 : strum_macros::Display,
165 0 : strum_macros::EnumIter,
166 : measured::FixedCardinalityLabel,
167 : )]
168 : #[strum(serialize_all = "snake_case")]
169 : pub(crate) enum LeadershipStatus {
170 : /// This is the steady state where the storage controller can produce
171 : /// side effects in the cluster.
172 : Leader,
173 : /// We've been notified to step down by another candidate. No reconciliations
174 : /// take place in this state.
175 : SteppedDown,
176 : /// Initial state for a new storage controller instance. Will attempt to assume leadership.
177 : #[allow(unused)]
178 : Candidate,
179 : }
180 :
181 : pub const RECONCILER_CONCURRENCY_DEFAULT: usize = 128;
182 :
183 : // Depth of the channel used to enqueue shards for reconciliation when they can't do it immediately.
184 : // This channel is finite-size to avoid using excessive memory if we get into a state where reconciles are finishing more slowly
185 : // than they're being pushed onto the queue.
186 : const MAX_DELAYED_RECONCILES: usize = 10000;
187 :
188 : // Top level state available to all HTTP handlers
189 : struct ServiceState {
190 : leadership_status: LeadershipStatus,
191 :
192 : tenants: BTreeMap<TenantShardId, TenantShard>,
193 :
194 : nodes: Arc<HashMap<NodeId, Node>>,
195 :
196 : scheduler: Scheduler,
197 :
198 : /// Ongoing background operation on the cluster if any is running.
199 : /// Note that only one such operation may run at any given time,
200 : /// hence the type choice.
201 : ongoing_operation: Option<OperationHandler>,
202 :
203 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
204 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
205 : }
206 :
207 : /// Transform an error from a pageserver into an error to return to callers of a storage
208 : /// controller API.
209 0 : fn passthrough_api_error(node: &Node, e: mgmt_api::Error) -> ApiError {
210 0 : match e {
211 0 : mgmt_api::Error::SendRequest(e) => {
212 0 : // Presume errors sending requests are connectivity/availability issues
213 0 : ApiError::ResourceUnavailable(format!("{node} error sending request: {e}").into())
214 : }
215 0 : mgmt_api::Error::ReceiveErrorBody(str) => {
216 0 : // Presume errors receiving body are connectivity/availability issues
217 0 : ApiError::ResourceUnavailable(
218 0 : format!("{node} error receiving error body: {str}").into(),
219 0 : )
220 : }
221 0 : mgmt_api::Error::ReceiveBody(str) => {
222 0 : // Presume errors receiving body are connectivity/availability issues
223 0 : ApiError::ResourceUnavailable(format!("{node} error receiving body: {str}").into())
224 : }
225 0 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, msg) => {
226 0 : ApiError::NotFound(anyhow::anyhow!(format!("{node}: {msg}")).into())
227 : }
228 0 : mgmt_api::Error::ApiError(StatusCode::SERVICE_UNAVAILABLE, msg) => {
229 0 : ApiError::ResourceUnavailable(format!("{node}: {msg}").into())
230 : }
231 0 : mgmt_api::Error::ApiError(status @ StatusCode::UNAUTHORIZED, msg)
232 0 : | mgmt_api::Error::ApiError(status @ StatusCode::FORBIDDEN, msg) => {
233 : // Auth errors talking to a pageserver are not auth errors for the caller: they are
234 : // internal server errors, showing that something is wrong with the pageserver or
235 : // storage controller's auth configuration.
236 0 : ApiError::InternalServerError(anyhow::anyhow!("{node} {status}: {msg}"))
237 : }
238 0 : mgmt_api::Error::ApiError(status, msg) => {
239 0 : // Presume general case of pageserver API errors is that we tried to do something
240 0 : // that can't be done right now.
241 0 : ApiError::Conflict(format!("{node} {status}: {status} {msg}"))
242 : }
243 0 : mgmt_api::Error::Cancelled => ApiError::ShuttingDown,
244 : }
245 0 : }
246 :
247 : impl ServiceState {
248 0 : fn new(
249 0 : nodes: HashMap<NodeId, Node>,
250 0 : tenants: BTreeMap<TenantShardId, TenantShard>,
251 0 : scheduler: Scheduler,
252 0 : delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
253 0 : initial_leadership_status: LeadershipStatus,
254 0 : ) -> Self {
255 0 : metrics::update_leadership_status(initial_leadership_status);
256 0 :
257 0 : Self {
258 0 : leadership_status: initial_leadership_status,
259 0 : tenants,
260 0 : nodes: Arc::new(nodes),
261 0 : scheduler,
262 0 : ongoing_operation: None,
263 0 : delayed_reconcile_rx,
264 0 : }
265 0 : }
266 :
267 0 : fn parts_mut(
268 0 : &mut self,
269 0 : ) -> (
270 0 : &mut Arc<HashMap<NodeId, Node>>,
271 0 : &mut BTreeMap<TenantShardId, TenantShard>,
272 0 : &mut Scheduler,
273 0 : ) {
274 0 : (&mut self.nodes, &mut self.tenants, &mut self.scheduler)
275 0 : }
276 :
277 0 : fn get_leadership_status(&self) -> LeadershipStatus {
278 0 : self.leadership_status
279 0 : }
280 :
281 0 : fn step_down(&mut self) {
282 0 : self.leadership_status = LeadershipStatus::SteppedDown;
283 0 : metrics::update_leadership_status(self.leadership_status);
284 0 : }
285 :
286 0 : fn become_leader(&mut self) {
287 0 : self.leadership_status = LeadershipStatus::Leader;
288 0 : metrics::update_leadership_status(self.leadership_status);
289 0 : }
290 : }
291 :
292 : #[derive(Clone)]
293 : pub struct Config {
294 : // All pageservers managed by one instance of this service must have
295 : // the same public key. This JWT token will be used to authenticate
296 : // this service to the pageservers it manages.
297 : pub jwt_token: Option<String>,
298 :
299 : // This JWT token will be used to authenticate this service to the control plane.
300 : pub control_plane_jwt_token: Option<String>,
301 :
302 : // This JWT token will be used to authenticate with other storage controller instances
303 : pub peer_jwt_token: Option<String>,
304 :
305 : /// Where the compute hook should send notifications of pageserver attachment locations
306 : /// (this URL points to the control plane in prod). If this is None, the compute hook will
307 : /// assume it is running in a test environment and try to update neon_local.
308 : pub compute_hook_url: Option<String>,
309 :
310 : /// Grace period within which a pageserver does not respond to heartbeats, but is still
311 : /// considered active. Once the grace period elapses, the next heartbeat failure will
312 : /// mark the pagseserver offline.
313 : pub max_offline_interval: Duration,
314 :
315 : /// Extended grace period within which pageserver may not respond to heartbeats.
316 : /// This extended grace period kicks in after the node has been drained for restart
317 : /// and/or upon handling the re-attach request from a node.
318 : pub max_warming_up_interval: Duration,
319 :
320 : /// How many Reconcilers may be spawned concurrently
321 : pub reconciler_concurrency: usize,
322 :
323 : /// How large must a shard grow in bytes before we split it?
324 : /// None disables auto-splitting.
325 : pub split_threshold: Option<u64>,
326 :
327 : // TODO: make this cfg(feature = "testing")
328 : pub neon_local_repo_dir: Option<PathBuf>,
329 :
330 : // Maximum acceptable download lag for the secondary location
331 : // while draining a node. If the secondary location is lagging
332 : // by more than the configured amount, then the secondary is not
333 : // upgraded to primary.
334 : pub max_secondary_lag_bytes: Option<u64>,
335 :
336 : pub heartbeat_interval: Duration,
337 :
338 : pub address_for_peers: Option<Uri>,
339 :
340 : pub start_as_candidate: bool,
341 :
342 : pub http_service_port: i32,
343 : }
344 :
345 : impl From<DatabaseError> for ApiError {
346 0 : fn from(err: DatabaseError) -> ApiError {
347 0 : match err {
348 0 : DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
349 : // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
350 : DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
351 0 : ApiError::ShuttingDown
352 : }
353 0 : DatabaseError::Logical(reason) | DatabaseError::Migration(reason) => {
354 0 : ApiError::InternalServerError(anyhow::anyhow!(reason))
355 : }
356 : }
357 0 : }
358 : }
359 :
360 : enum InitialShardScheduleOutcome {
361 : Scheduled(TenantCreateResponseShard),
362 : NotScheduled,
363 : ShardScheduleError(ScheduleError),
364 : }
365 :
366 : pub struct Service {
367 : inner: Arc<std::sync::RwLock<ServiceState>>,
368 : config: Config,
369 : persistence: Arc<Persistence>,
370 : compute_hook: Arc<ComputeHook>,
371 : result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
372 :
373 : heartbeater: Heartbeater,
374 :
375 : // Channel for background cleanup from failed operations that require cleanup, such as shard split
376 : abort_tx: tokio::sync::mpsc::UnboundedSender<TenantShardSplitAbort>,
377 :
378 : // Locking on a tenant granularity (covers all shards in the tenant):
379 : // - Take exclusively for rare operations that mutate the tenant's persistent state (e.g. create/delete/split)
380 : // - Take in shared mode for operations that need the set of shards to stay the same to complete reliably (e.g. timeline CRUD)
381 : tenant_op_locks: IdLockMap<TenantId, TenantOperations>,
382 :
383 : // Locking for node-mutating operations: take exclusively for operations that modify the node's persistent state, or
384 : // that transition it to/from Active.
385 : node_op_locks: IdLockMap<NodeId, NodeOperations>,
386 :
387 : // Limit how many Reconcilers we will spawn concurrently
388 : reconciler_concurrency: Arc<tokio::sync::Semaphore>,
389 :
390 : /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
391 : /// Send into this queue to promptly attempt to reconcile this shard next time units are available.
392 : ///
393 : /// Note that this state logically lives inside ServiceInner, but carrying Sender here makes the code simpler
394 : /// by avoiding needing a &mut ref to something inside the ServiceInner. This could be optimized to
395 : /// use a VecDeque instead of a channel to reduce synchronization overhead, at the cost of some code complexity.
396 : delayed_reconcile_tx: tokio::sync::mpsc::Sender<TenantShardId>,
397 :
398 : // Process shutdown will fire this token
399 : cancel: CancellationToken,
400 :
401 : // Child token of [`Service::cancel`] used by reconcilers
402 : reconcilers_cancel: CancellationToken,
403 :
404 : // Background tasks will hold this gate
405 : gate: Gate,
406 :
407 : // Reconcilers background tasks will hold this gate
408 : reconcilers_gate: Gate,
409 :
410 : /// This waits for initial reconciliation with pageservers to complete. Until this barrier
411 : /// passes, it isn't safe to do any actions that mutate tenants.
412 : pub(crate) startup_complete: Barrier,
413 : }
414 :
415 : impl From<ReconcileWaitError> for ApiError {
416 0 : fn from(value: ReconcileWaitError) -> Self {
417 0 : match value {
418 0 : ReconcileWaitError::Shutdown => ApiError::ShuttingDown,
419 0 : e @ ReconcileWaitError::Timeout(_) => ApiError::Timeout(format!("{e}").into()),
420 0 : e @ ReconcileWaitError::Failed(..) => ApiError::InternalServerError(anyhow::anyhow!(e)),
421 : }
422 0 : }
423 : }
424 :
425 : impl From<OperationError> for ApiError {
426 0 : fn from(value: OperationError) -> Self {
427 0 : match value {
428 0 : OperationError::NodeStateChanged(err) | OperationError::FinalizeError(err) => {
429 0 : ApiError::InternalServerError(anyhow::anyhow!(err))
430 : }
431 0 : OperationError::Cancelled => ApiError::Conflict("Operation was cancelled".into()),
432 : }
433 0 : }
434 : }
435 :
436 : #[allow(clippy::large_enum_variant)]
437 : enum TenantCreateOrUpdate {
438 : Create(TenantCreateRequest),
439 : Update(Vec<ShardUpdate>),
440 : }
441 :
442 : struct ShardSplitParams {
443 : old_shard_count: ShardCount,
444 : new_shard_count: ShardCount,
445 : new_stripe_size: Option<ShardStripeSize>,
446 : targets: Vec<ShardSplitTarget>,
447 : policy: PlacementPolicy,
448 : config: TenantConfig,
449 : shard_ident: ShardIdentity,
450 : }
451 :
452 : // When preparing for a shard split, we may either choose to proceed with the split,
453 : // or find that the work is already done and return NoOp.
454 : enum ShardSplitAction {
455 : Split(Box<ShardSplitParams>),
456 : NoOp(TenantShardSplitResponse),
457 : }
458 :
459 : // A parent shard which will be split
460 : struct ShardSplitTarget {
461 : parent_id: TenantShardId,
462 : node: Node,
463 : child_ids: Vec<TenantShardId>,
464 : }
465 :
466 : /// When we tenant shard split operation fails, we may not be able to clean up immediately, because nodes
467 : /// might not be available. We therefore use a queue of abort operations processed in the background.
468 : struct TenantShardSplitAbort {
469 : tenant_id: TenantId,
470 : /// The target values from the request that failed
471 : new_shard_count: ShardCount,
472 : new_stripe_size: Option<ShardStripeSize>,
473 : /// Until this abort op is complete, no other operations may be done on the tenant
474 : _tenant_lock: TracingExclusiveGuard<TenantOperations>,
475 : }
476 :
477 0 : #[derive(thiserror::Error, Debug)]
478 : enum TenantShardSplitAbortError {
479 : #[error(transparent)]
480 : Database(#[from] DatabaseError),
481 : #[error(transparent)]
482 : Remote(#[from] mgmt_api::Error),
483 : #[error("Unavailable")]
484 : Unavailable,
485 : }
486 :
487 : struct ShardUpdate {
488 : tenant_shard_id: TenantShardId,
489 : placement_policy: PlacementPolicy,
490 : tenant_config: TenantConfig,
491 :
492 : /// If this is None, generation is not updated.
493 : generation: Option<Generation>,
494 : }
495 :
496 : enum StopReconciliationsReason {
497 : ShuttingDown,
498 : SteppingDown,
499 : }
500 :
501 : impl std::fmt::Display for StopReconciliationsReason {
502 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
503 0 : let s = match self {
504 0 : Self::ShuttingDown => "Shutting down",
505 0 : Self::SteppingDown => "Stepping down",
506 : };
507 0 : write!(writer, "{}", s)
508 0 : }
509 : }
510 :
511 : pub(crate) enum ReconcileResultRequest {
512 : ReconcileResult(ReconcileResult),
513 : Stop,
514 : }
515 :
516 : impl Service {
517 0 : pub fn get_config(&self) -> &Config {
518 0 : &self.config
519 0 : }
520 :
521 : /// Called once on startup, this function attempts to contact all pageservers to build an up-to-date
522 : /// view of the world, and determine which pageservers are responsive.
523 0 : #[instrument(skip_all)]
524 : async fn startup_reconcile(
525 : self: &Arc<Service>,
526 : current_leader: Option<ControllerPersistence>,
527 : leader_step_down_state: Option<GlobalObservedState>,
528 : bg_compute_notify_result_tx: tokio::sync::mpsc::Sender<
529 : Result<(), (TenantShardId, NotifyError)>,
530 : >,
531 : ) {
532 : // Startup reconciliation does I/O to other services: whether they
533 : // are responsive or not, we should aim to finish within our deadline, because:
534 : // - If we don't, a k8s readiness hook watching /ready will kill us.
535 : // - While we're waiting for startup reconciliation, we are not fully
536 : // available for end user operations like creating/deleting tenants and timelines.
537 : //
538 : // We set multiple deadlines to break up the time available between the phases of work: this is
539 : // arbitrary, but avoids a situation where the first phase could burn our entire timeout period.
540 : let start_at = Instant::now();
541 : let node_scan_deadline = start_at
542 : .checked_add(STARTUP_RECONCILE_TIMEOUT / 2)
543 : .expect("Reconcile timeout is a modest constant");
544 :
545 : let observed = if let Some(state) = leader_step_down_state {
546 : tracing::info!(
547 : "Using observed state received from leader at {}",
548 : current_leader.as_ref().unwrap().address
549 : );
550 :
551 : state
552 : } else {
553 : self.build_global_observed_state(node_scan_deadline).await
554 : };
555 :
556 : // Accumulate a list of any tenant locations that ought to be detached
557 : let mut cleanup = Vec::new();
558 :
559 : // Send initial heartbeat requests to all nodes loaded from the database
560 : let all_nodes = {
561 : let locked = self.inner.read().unwrap();
562 : locked.nodes.clone()
563 : };
564 : let mut nodes_online = self.initial_heartbeat_round(all_nodes.keys()).await;
565 :
566 : // List of tenants for which we will attempt to notify compute of their location at startup
567 : let mut compute_notifications = Vec::new();
568 :
569 : // Populate intent and observed states for all tenants, based on reported state on pageservers
570 : tracing::info!("Populating tenant shards' states from initial pageserver scan...");
571 : let shard_count = {
572 : let mut locked = self.inner.write().unwrap();
573 : let (nodes, tenants, scheduler) = locked.parts_mut();
574 :
575 : // Mark nodes online if they responded to us: nodes are offline by default after a restart.
576 : let mut new_nodes = (**nodes).clone();
577 : for (node_id, node) in new_nodes.iter_mut() {
578 : if let Some(utilization) = nodes_online.remove(node_id) {
579 : node.set_availability(NodeAvailability::Active(utilization));
580 : scheduler.node_upsert(node);
581 : }
582 : }
583 : *nodes = Arc::new(new_nodes);
584 :
585 : for (tenant_shard_id, observed_state) in observed.0 {
586 : let Some(tenant_shard) = tenants.get_mut(&tenant_shard_id) else {
587 : for node_id in observed_state.locations.keys() {
588 : cleanup.push((tenant_shard_id, *node_id));
589 : }
590 :
591 : continue;
592 : };
593 :
594 : tenant_shard.observed = observed_state;
595 : }
596 :
597 : // Populate each tenant's intent state
598 : let mut schedule_context = ScheduleContext::default();
599 : for (tenant_shard_id, tenant_shard) in tenants.iter_mut() {
600 : if tenant_shard_id.shard_number == ShardNumber(0) {
601 : // Reset scheduling context each time we advance to the next Tenant
602 : schedule_context = ScheduleContext::default();
603 : }
604 :
605 : tenant_shard.intent_from_observed(scheduler);
606 : if let Err(e) = tenant_shard.schedule(scheduler, &mut schedule_context) {
607 : // Non-fatal error: we are unable to properly schedule the tenant, perhaps because
608 : // not enough pageservers are available. The tenant may well still be available
609 : // to clients.
610 : tracing::error!("Failed to schedule tenant {tenant_shard_id} at startup: {e}");
611 : } else {
612 : // If we're both intending and observed to be attached at a particular node, we will
613 : // emit a compute notification for this. In the case where our observed state does not
614 : // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
615 : if let Some(attached_at) = tenant_shard.stably_attached() {
616 : compute_notifications.push((
617 : *tenant_shard_id,
618 : attached_at,
619 : tenant_shard.shard.stripe_size,
620 : ));
621 : }
622 : }
623 : }
624 :
625 : tenants.len()
626 : };
627 :
628 : // Before making any obeservable changes to the cluster, persist self
629 : // as leader in database and memory.
630 : let leadership = Leadership::new(
631 : self.persistence.clone(),
632 : self.config.clone(),
633 : self.cancel.child_token(),
634 : );
635 :
636 : if let Err(e) = leadership.become_leader(current_leader).await {
637 : tracing::error!("Failed to persist self as leader: {e}. Aborting start-up ...");
638 : std::process::exit(1);
639 : }
640 :
641 : self.inner.write().unwrap().become_leader();
642 :
643 : // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
644 : // generation_pageserver in the database.
645 :
646 : // Emit compute hook notifications for all tenants which are already stably attached. Other tenants
647 : // will emit compute hook notifications when they reconcile.
648 : //
649 : // Ordering: our calls to notify_background synchronously establish a relative order for these notifications vs. any later
650 : // calls into the ComputeHook for the same tenant: we can leave these to run to completion in the background and any later
651 : // calls will be correctly ordered wrt these.
652 : //
653 : // Concurrency: we call notify_background for all tenants, which will create O(N) tokio tasks, but almost all of them
654 : // will just wait on the ComputeHook::API_CONCURRENCY semaphore immediately, so very cheap until they get that semaphore
655 : // unit and start doing I/O.
656 : tracing::info!(
657 : "Sending {} compute notifications",
658 : compute_notifications.len()
659 : );
660 : self.compute_hook.notify_background(
661 : compute_notifications,
662 : bg_compute_notify_result_tx.clone(),
663 : &self.cancel,
664 : );
665 :
666 : // Finally, now that the service is up and running, launch reconcile operations for any tenants
667 : // which require it: under normal circumstances this should only include tenants that were in some
668 : // transient state before we restarted, or any tenants whose compute hooks failed above.
669 : tracing::info!("Checking for shards in need of reconciliation...");
670 : let reconcile_tasks = self.reconcile_all();
671 : // We will not wait for these reconciliation tasks to run here: we're now done with startup and
672 : // normal operations may proceed.
673 :
674 : // Clean up any tenants that were found on pageservers but are not known to us. Do this in the
675 : // background because it does not need to complete in order to proceed with other work.
676 : if !cleanup.is_empty() {
677 : tracing::info!("Cleaning up {} locations in the background", cleanup.len());
678 : tokio::task::spawn({
679 : let cleanup_self = self.clone();
680 0 : async move { cleanup_self.cleanup_locations(cleanup).await }
681 : });
682 : }
683 :
684 : tracing::info!("Startup complete, spawned {reconcile_tasks} reconciliation tasks ({shard_count} shards total)");
685 : }
686 :
687 0 : async fn initial_heartbeat_round<'a>(
688 0 : &self,
689 0 : node_ids: impl Iterator<Item = &'a NodeId>,
690 0 : ) -> HashMap<NodeId, PageserverUtilization> {
691 0 : assert!(!self.startup_complete.is_ready());
692 :
693 0 : let all_nodes = {
694 0 : let locked = self.inner.read().unwrap();
695 0 : locked.nodes.clone()
696 0 : };
697 0 :
698 0 : let mut nodes_to_heartbeat = HashMap::new();
699 0 : for node_id in node_ids {
700 0 : match all_nodes.get(node_id) {
701 0 : Some(node) => {
702 0 : nodes_to_heartbeat.insert(*node_id, node.clone());
703 0 : }
704 : None => {
705 0 : tracing::warn!("Node {node_id} was removed during start-up");
706 : }
707 : }
708 : }
709 :
710 0 : tracing::info!("Sending initial heartbeats...");
711 0 : let res = self
712 0 : .heartbeater
713 0 : .heartbeat(Arc::new(nodes_to_heartbeat))
714 0 : .await;
715 :
716 0 : let mut online_nodes = HashMap::new();
717 0 : if let Ok(deltas) = res {
718 0 : for (node_id, status) in deltas.0 {
719 0 : match status {
720 0 : PageserverState::Available { utilization, .. } => {
721 0 : online_nodes.insert(node_id, utilization);
722 0 : }
723 0 : PageserverState::Offline => {}
724 : PageserverState::WarmingUp { .. } => {
725 0 : unreachable!("Nodes are never marked warming-up during startup reconcile")
726 : }
727 : }
728 : }
729 0 : }
730 :
731 0 : online_nodes
732 0 : }
733 :
734 : /// Used during [`Self::startup_reconcile`]: issue GETs to all nodes concurrently, with a deadline.
735 : ///
736 : /// The result includes only nodes which responded within the deadline
737 0 : async fn scan_node_locations(
738 0 : &self,
739 0 : deadline: Instant,
740 0 : ) -> HashMap<NodeId, LocationConfigListResponse> {
741 0 : let nodes = {
742 0 : let locked = self.inner.read().unwrap();
743 0 : locked.nodes.clone()
744 0 : };
745 0 :
746 0 : let mut node_results = HashMap::new();
747 0 :
748 0 : let mut node_list_futs = FuturesUnordered::new();
749 0 :
750 0 : tracing::info!("Scanning shards on {} nodes...", nodes.len());
751 0 : for node in nodes.values() {
752 0 : node_list_futs.push({
753 0 : async move {
754 0 : tracing::info!("Scanning shards on node {node}...");
755 0 : let timeout = Duration::from_secs(1);
756 0 : let response = node
757 0 : .with_client_retries(
758 0 : |client| async move { client.list_location_config().await },
759 0 : &self.config.jwt_token,
760 0 : 1,
761 0 : 5,
762 0 : timeout,
763 0 : &self.cancel,
764 0 : )
765 0 : .await;
766 0 : (node.get_id(), response)
767 0 : }
768 0 : });
769 0 : }
770 :
771 : loop {
772 0 : let (node_id, result) = tokio::select! {
773 0 : next = node_list_futs.next() => {
774 0 : match next {
775 0 : Some(result) => result,
776 : None =>{
777 : // We got results for all our nodes
778 0 : break;
779 : }
780 :
781 : }
782 : },
783 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
784 : // Give up waiting for anyone who hasn't responded: we will yield the results that we have
785 0 : tracing::info!("Reached deadline while waiting for nodes to respond to location listing requests");
786 0 : break;
787 : }
788 : };
789 :
790 0 : let Some(list_response) = result else {
791 0 : tracing::info!("Shutdown during startup_reconcile");
792 0 : break;
793 : };
794 :
795 0 : match list_response {
796 0 : Err(e) => {
797 0 : tracing::warn!("Could not scan node {} ({e})", node_id);
798 : }
799 0 : Ok(listing) => {
800 0 : node_results.insert(node_id, listing);
801 0 : }
802 : }
803 : }
804 :
805 0 : node_results
806 0 : }
807 :
808 0 : async fn build_global_observed_state(&self, deadline: Instant) -> GlobalObservedState {
809 0 : let node_listings = self.scan_node_locations(deadline).await;
810 0 : let mut observed = GlobalObservedState::default();
811 :
812 0 : for (node_id, location_confs) in node_listings {
813 0 : tracing::info!(
814 0 : "Received {} shard statuses from pageserver {}",
815 0 : location_confs.tenant_shards.len(),
816 : node_id
817 : );
818 :
819 0 : for (tid, location_conf) in location_confs.tenant_shards {
820 0 : let entry = observed.0.entry(tid).or_default();
821 0 : entry.locations.insert(
822 0 : node_id,
823 0 : ObservedStateLocation {
824 0 : conf: location_conf,
825 0 : },
826 0 : );
827 0 : }
828 : }
829 :
830 0 : observed
831 0 : }
832 :
833 : /// Used during [`Self::startup_reconcile`]: detach a list of unknown-to-us tenants from pageservers.
834 : ///
835 : /// This is safe to run in the background, because if we don't have this TenantShardId in our map of
836 : /// tenants, then it is probably something incompletely deleted before: we will not fight with any
837 : /// other task trying to attach it.
838 0 : #[instrument(skip_all)]
839 : async fn cleanup_locations(&self, cleanup: Vec<(TenantShardId, NodeId)>) {
840 : let nodes = self.inner.read().unwrap().nodes.clone();
841 :
842 : for (tenant_shard_id, node_id) in cleanup {
843 : // A node reported a tenant_shard_id which is unknown to us: detach it.
844 : let Some(node) = nodes.get(&node_id) else {
845 : // This is legitimate; we run in the background and [`Self::startup_reconcile`] might have identified
846 : // a location to clean up on a node that has since been removed.
847 : tracing::info!(
848 : "Not cleaning up location {node_id}/{tenant_shard_id}: node not found"
849 : );
850 : continue;
851 : };
852 :
853 : if self.cancel.is_cancelled() {
854 : break;
855 : }
856 :
857 : let client = PageserverClient::new(
858 : node.get_id(),
859 : node.base_url(),
860 : self.config.jwt_token.as_deref(),
861 : );
862 : match client
863 : .location_config(
864 : tenant_shard_id,
865 : LocationConfig {
866 : mode: LocationConfigMode::Detached,
867 : generation: None,
868 : secondary_conf: None,
869 : shard_number: tenant_shard_id.shard_number.0,
870 : shard_count: tenant_shard_id.shard_count.literal(),
871 : shard_stripe_size: 0,
872 : tenant_conf: models::TenantConfig::default(),
873 : },
874 : None,
875 : false,
876 : )
877 : .await
878 : {
879 : Ok(()) => {
880 : tracing::info!(
881 : "Detached unknown shard {tenant_shard_id} on pageserver {node_id}"
882 : );
883 : }
884 : Err(e) => {
885 : // Non-fatal error: leaving a tenant shard behind that we are not managing shouldn't
886 : // break anything.
887 : tracing::error!(
888 : "Failed to detach unknkown shard {tenant_shard_id} on pageserver {node_id}: {e}"
889 : );
890 : }
891 : }
892 : }
893 : }
894 :
895 : /// Long running background task that periodically wakes up and looks for shards that need
896 : /// reconciliation. Reconciliation is fallible, so any reconciliation tasks that fail during
897 : /// e.g. a tenant create/attach/migrate must eventually be retried: this task is responsible
898 : /// for those retries.
899 0 : #[instrument(skip_all)]
900 : async fn background_reconcile(self: &Arc<Self>) {
901 : self.startup_complete.clone().wait().await;
902 :
903 : const BACKGROUND_RECONCILE_PERIOD: Duration = Duration::from_secs(20);
904 :
905 : let mut interval = tokio::time::interval(BACKGROUND_RECONCILE_PERIOD);
906 : while !self.reconcilers_cancel.is_cancelled() {
907 : tokio::select! {
908 : _ = interval.tick() => {
909 : let reconciles_spawned = self.reconcile_all();
910 : if reconciles_spawned == 0 {
911 : // Run optimizer only when we didn't find any other work to do
912 : let optimizations = self.optimize_all().await;
913 : if optimizations == 0 {
914 : // Run new splits only when no optimizations are pending
915 : self.autosplit_tenants().await;
916 : }
917 : }
918 : }
919 : _ = self.reconcilers_cancel.cancelled() => return
920 : }
921 : }
922 : }
923 0 : #[instrument(skip_all)]
924 : async fn spawn_heartbeat_driver(&self) {
925 : self.startup_complete.clone().wait().await;
926 :
927 : let mut interval = tokio::time::interval(self.config.heartbeat_interval);
928 : while !self.cancel.is_cancelled() {
929 : tokio::select! {
930 : _ = interval.tick() => { }
931 : _ = self.cancel.cancelled() => return
932 : };
933 :
934 : let nodes = {
935 : let locked = self.inner.read().unwrap();
936 : locked.nodes.clone()
937 : };
938 :
939 : let res = self.heartbeater.heartbeat(nodes).await;
940 : if let Ok(deltas) = res {
941 : for (node_id, state) in deltas.0 {
942 : let new_availability = match state {
943 : PageserverState::Available { utilization, .. } => {
944 : NodeAvailability::Active(utilization)
945 : }
946 : PageserverState::WarmingUp { started_at } => {
947 : NodeAvailability::WarmingUp(started_at)
948 : }
949 : PageserverState::Offline => {
950 : // The node might have been placed in the WarmingUp state
951 : // while the heartbeat round was on-going. Hence, filter out
952 : // offline transitions for WarmingUp nodes that are still within
953 : // their grace period.
954 : if let Ok(NodeAvailability::WarmingUp(started_at)) = self
955 : .get_node(node_id)
956 : .await
957 : .as_ref()
958 0 : .map(|n| n.get_availability())
959 : {
960 : let now = Instant::now();
961 : if now - *started_at >= self.config.max_warming_up_interval {
962 : NodeAvailability::Offline
963 : } else {
964 : NodeAvailability::WarmingUp(*started_at)
965 : }
966 : } else {
967 : NodeAvailability::Offline
968 : }
969 : }
970 : };
971 :
972 : // This is the code path for geniune availability transitions (i.e node
973 : // goes unavailable and/or comes back online).
974 : let res = self
975 : .node_configure(node_id, Some(new_availability), None)
976 : .await;
977 :
978 : match res {
979 : Ok(()) => {}
980 : Err(ApiError::NotFound(_)) => {
981 : // This should be rare, but legitimate since the heartbeats are done
982 : // on a snapshot of the nodes.
983 : tracing::info!("Node {} was not found after heartbeat round", node_id);
984 : }
985 : Err(err) => {
986 : // Transition to active involves reconciling: if a node responds to a heartbeat then
987 : // becomes unavailable again, we may get an error here.
988 : tracing::error!(
989 : "Failed to update node {} after heartbeat round: {}",
990 : node_id,
991 : err
992 : );
993 : }
994 : }
995 : }
996 : }
997 : }
998 : }
999 :
1000 : /// Apply the contents of a [`ReconcileResult`] to our in-memory state: if the reconciliation
1001 : /// was successful and intent hasn't changed since the Reconciler was spawned, this will update
1002 : /// the observed state of the tenant such that subsequent calls to [`TenantShard::get_reconcile_needed`]
1003 : /// will indicate that reconciliation is not needed.
1004 0 : #[instrument(skip_all, fields(
1005 : tenant_id=%result.tenant_shard_id.tenant_id, shard_id=%result.tenant_shard_id.shard_slug(),
1006 : sequence=%result.sequence
1007 0 : ))]
1008 : fn process_result(&self, mut result: ReconcileResult) {
1009 : let mut locked = self.inner.write().unwrap();
1010 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1011 : let Some(tenant) = tenants.get_mut(&result.tenant_shard_id) else {
1012 : // A reconciliation result might race with removing a tenant: drop results for
1013 : // tenants that aren't in our map.
1014 : return;
1015 : };
1016 :
1017 : // Usually generation should only be updated via this path, so the max() isn't
1018 : // needed, but it is used to handle out-of-band updates via. e.g. test hook.
1019 : tenant.generation = std::cmp::max(tenant.generation, result.generation);
1020 :
1021 : // If the reconciler signals that it failed to notify compute, set this state on
1022 : // the shard so that a future [`TenantShard::maybe_reconcile`] will try again.
1023 : tenant.pending_compute_notification = result.pending_compute_notification;
1024 :
1025 : // Let the TenantShard know it is idle.
1026 : tenant.reconcile_complete(result.sequence);
1027 :
1028 : // In case a node was deleted while this reconcile is in flight, filter it out of the update we will
1029 : // make to the tenant
1030 : result
1031 : .observed
1032 : .locations
1033 0 : .retain(|node_id, _loc| nodes.contains_key(node_id));
1034 :
1035 : match result.result {
1036 : Ok(()) => {
1037 : for (node_id, loc) in &result.observed.locations {
1038 : if let Some(conf) = &loc.conf {
1039 : tracing::info!("Updating observed location {}: {:?}", node_id, conf);
1040 : } else {
1041 : tracing::info!("Setting observed location {} to None", node_id,)
1042 : }
1043 : }
1044 :
1045 : tenant.observed = result.observed;
1046 : tenant.waiter.advance(result.sequence);
1047 : }
1048 : Err(e) => {
1049 : match e {
1050 : ReconcileError::Cancel => {
1051 : tracing::info!("Reconciler was cancelled");
1052 : }
1053 : ReconcileError::Remote(mgmt_api::Error::Cancelled) => {
1054 : // This might be due to the reconciler getting cancelled, or it might
1055 : // be due to the `Node` being marked offline.
1056 : tracing::info!("Reconciler cancelled during pageserver API call");
1057 : }
1058 : _ => {
1059 : tracing::warn!("Reconcile error: {}", e);
1060 : }
1061 : }
1062 :
1063 : // Ordering: populate last_error before advancing error_seq,
1064 : // so that waiters will see the correct error after waiting.
1065 : tenant.set_last_error(result.sequence, e);
1066 :
1067 : for (node_id, o) in result.observed.locations {
1068 : tenant.observed.locations.insert(node_id, o);
1069 : }
1070 : }
1071 : }
1072 :
1073 : // Maybe some other work can proceed now that this job finished.
1074 : if self.reconciler_concurrency.available_permits() > 0 {
1075 : while let Ok(tenant_shard_id) = locked.delayed_reconcile_rx.try_recv() {
1076 : let (nodes, tenants, _scheduler) = locked.parts_mut();
1077 : if let Some(shard) = tenants.get_mut(&tenant_shard_id) {
1078 : shard.delayed_reconcile = false;
1079 : self.maybe_reconcile_shard(shard, nodes);
1080 : }
1081 :
1082 : if self.reconciler_concurrency.available_permits() == 0 {
1083 : break;
1084 : }
1085 : }
1086 : }
1087 : }
1088 :
1089 0 : async fn process_results(
1090 0 : &self,
1091 0 : mut result_rx: tokio::sync::mpsc::UnboundedReceiver<ReconcileResultRequest>,
1092 0 : mut bg_compute_hook_result_rx: tokio::sync::mpsc::Receiver<
1093 0 : Result<(), (TenantShardId, NotifyError)>,
1094 0 : >,
1095 0 : ) {
1096 : loop {
1097 : // Wait for the next result, or for cancellation
1098 0 : tokio::select! {
1099 0 : r = result_rx.recv() => {
1100 0 : match r {
1101 0 : Some(ReconcileResultRequest::ReconcileResult(result)) => {self.process_result(result);},
1102 0 : None | Some(ReconcileResultRequest::Stop) => {break;}
1103 : }
1104 : }
1105 0 : _ = async{
1106 0 : match bg_compute_hook_result_rx.recv().await {
1107 0 : Some(result) => {
1108 0 : if let Err((tenant_shard_id, notify_error)) = result {
1109 0 : tracing::warn!("Marking shard {tenant_shard_id} for notification retry, due to error {notify_error}");
1110 0 : let mut locked = self.inner.write().unwrap();
1111 0 : if let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) {
1112 0 : shard.pending_compute_notification = true;
1113 0 : }
1114 :
1115 0 : }
1116 : },
1117 : None => {
1118 : // This channel is dead, but we don't want to terminate the outer loop{}: just wait for shutdown
1119 0 : self.cancel.cancelled().await;
1120 : }
1121 : }
1122 0 : } => {},
1123 0 : _ = self.cancel.cancelled() => {
1124 0 : break;
1125 : }
1126 : };
1127 : }
1128 0 : }
1129 :
1130 0 : async fn process_aborts(
1131 0 : &self,
1132 0 : mut abort_rx: tokio::sync::mpsc::UnboundedReceiver<TenantShardSplitAbort>,
1133 0 : ) {
1134 : loop {
1135 : // Wait for the next result, or for cancellation
1136 0 : let op = tokio::select! {
1137 0 : r = abort_rx.recv() => {
1138 0 : match r {
1139 0 : Some(op) => {op},
1140 0 : None => {break;}
1141 : }
1142 : }
1143 0 : _ = self.cancel.cancelled() => {
1144 0 : break;
1145 : }
1146 : };
1147 :
1148 : // Retry until shutdown: we must keep this request object alive until it is properly
1149 : // processed, as it holds a lock guard that prevents other operations trying to do things
1150 : // to the tenant while it is in a weird part-split state.
1151 0 : while !self.cancel.is_cancelled() {
1152 0 : match self.abort_tenant_shard_split(&op).await {
1153 0 : Ok(_) => break,
1154 0 : Err(e) => {
1155 0 : tracing::warn!(
1156 0 : "Failed to abort shard split on {}, will retry: {e}",
1157 : op.tenant_id
1158 : );
1159 :
1160 : // If a node is unavailable, we hope that it has been properly marked Offline
1161 : // when we retry, so that the abort op will succeed. If the abort op is failing
1162 : // for some other reason, we will keep retrying forever, or until a human notices
1163 : // and does something about it (either fixing a pageserver or restarting the controller).
1164 0 : tokio::time::timeout(Duration::from_secs(5), self.cancel.cancelled())
1165 0 : .await
1166 0 : .ok();
1167 : }
1168 : }
1169 : }
1170 : }
1171 0 : }
1172 :
1173 0 : pub async fn spawn(config: Config, persistence: Arc<Persistence>) -> anyhow::Result<Arc<Self>> {
1174 0 : let (result_tx, result_rx) = tokio::sync::mpsc::unbounded_channel();
1175 0 : let (abort_tx, abort_rx) = tokio::sync::mpsc::unbounded_channel();
1176 0 :
1177 0 : let leadership_cancel = CancellationToken::new();
1178 0 : let leadership = Leadership::new(persistence.clone(), config.clone(), leadership_cancel);
1179 0 : let (leader, leader_step_down_state) = leadership.step_down_current_leader().await?;
1180 :
1181 : // Apply the migrations **after** the current leader has stepped down
1182 : // (or we've given up waiting for it), but **before** reading from the
1183 : // database. The only exception is reading the current leader before
1184 : // migrating.
1185 0 : persistence.migration_run().await?;
1186 :
1187 0 : tracing::info!("Loading nodes from database...");
1188 0 : let nodes = persistence
1189 0 : .list_nodes()
1190 0 : .await?
1191 0 : .into_iter()
1192 0 : .map(Node::from_persistent)
1193 0 : .collect::<Vec<_>>();
1194 0 : let nodes: HashMap<NodeId, Node> = nodes.into_iter().map(|n| (n.get_id(), n)).collect();
1195 0 : tracing::info!("Loaded {} nodes from database.", nodes.len());
1196 :
1197 0 : tracing::info!("Loading shards from database...");
1198 0 : let mut tenant_shard_persistence = persistence.list_tenant_shards().await?;
1199 0 : tracing::info!(
1200 0 : "Loaded {} shards from database.",
1201 0 : tenant_shard_persistence.len()
1202 : );
1203 :
1204 : // If any shard splits were in progress, reset the database state to abort them
1205 0 : let mut tenant_shard_count_min_max: HashMap<TenantId, (ShardCount, ShardCount)> =
1206 0 : HashMap::new();
1207 0 : for tsp in &mut tenant_shard_persistence {
1208 0 : let shard = tsp.get_shard_identity()?;
1209 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1210 0 : let entry = tenant_shard_count_min_max
1211 0 : .entry(tenant_shard_id.tenant_id)
1212 0 : .or_insert_with(|| (shard.count, shard.count));
1213 0 : entry.0 = std::cmp::min(entry.0, shard.count);
1214 0 : entry.1 = std::cmp::max(entry.1, shard.count);
1215 0 : }
1216 :
1217 0 : for (tenant_id, (count_min, count_max)) in tenant_shard_count_min_max {
1218 0 : if count_min != count_max {
1219 : // Aborting the split in the database and dropping the child shards is sufficient: the reconciliation in
1220 : // [`Self::startup_reconcile`] will implicitly drop the child shards on remote pageservers, or they'll
1221 : // be dropped later in [`Self::node_activate_reconcile`] if it isn't available right now.
1222 0 : tracing::info!("Aborting shard split {tenant_id} {count_min:?} -> {count_max:?}");
1223 0 : let abort_status = persistence.abort_shard_split(tenant_id, count_max).await?;
1224 :
1225 : // We may never see the Complete status here: if the split was complete, we wouldn't have
1226 : // identified this tenant has having mismatching min/max counts.
1227 0 : assert!(matches!(abort_status, AbortShardSplitStatus::Aborted));
1228 :
1229 : // Clear the splitting status in-memory, to reflect that we just aborted in the database
1230 0 : tenant_shard_persistence.iter_mut().for_each(|tsp| {
1231 0 : // Set idle split state on those shards that we will retain.
1232 0 : let tsp_tenant_id = TenantId::from_str(tsp.tenant_id.as_str()).unwrap();
1233 0 : if tsp_tenant_id == tenant_id
1234 0 : && tsp.get_shard_identity().unwrap().count == count_min
1235 0 : {
1236 0 : tsp.splitting = SplitState::Idle;
1237 0 : } else if tsp_tenant_id == tenant_id {
1238 : // Leave the splitting state on the child shards: this will be used next to
1239 : // drop them.
1240 0 : tracing::info!(
1241 0 : "Shard {tsp_tenant_id} will be dropped after shard split abort",
1242 : );
1243 0 : }
1244 0 : });
1245 0 :
1246 0 : // Drop shards for this tenant which we didn't just mark idle (i.e. child shards of the aborted split)
1247 0 : tenant_shard_persistence.retain(|tsp| {
1248 0 : TenantId::from_str(tsp.tenant_id.as_str()).unwrap() != tenant_id
1249 0 : || tsp.splitting == SplitState::Idle
1250 0 : });
1251 0 : }
1252 : }
1253 :
1254 0 : let mut tenants = BTreeMap::new();
1255 0 :
1256 0 : let mut scheduler = Scheduler::new(nodes.values());
1257 0 :
1258 0 : #[cfg(feature = "testing")]
1259 0 : {
1260 0 : // Hack: insert scheduler state for all nodes referenced by shards, as compatibility
1261 0 : // tests only store the shards, not the nodes. The nodes will be loaded shortly
1262 0 : // after when pageservers start up and register.
1263 0 : let mut node_ids = HashSet::new();
1264 0 : for tsp in &tenant_shard_persistence {
1265 0 : if let Some(node_id) = tsp.generation_pageserver {
1266 0 : node_ids.insert(node_id);
1267 0 : }
1268 : }
1269 0 : for node_id in node_ids {
1270 0 : tracing::info!("Creating node {} in scheduler for tests", node_id);
1271 0 : let node = Node::new(
1272 0 : NodeId(node_id as u64),
1273 0 : "".to_string(),
1274 0 : 123,
1275 0 : "".to_string(),
1276 0 : 123,
1277 0 : "test_az".to_string(),
1278 0 : );
1279 0 :
1280 0 : scheduler.node_upsert(&node);
1281 : }
1282 : }
1283 0 : for tsp in tenant_shard_persistence {
1284 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1285 :
1286 : // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
1287 : // it with what we can infer: the node for which a generation was most recently issued.
1288 0 : let mut intent = IntentState::new();
1289 0 : if let Some(generation_pageserver) = tsp.generation_pageserver.map(|n| NodeId(n as u64))
1290 : {
1291 0 : if nodes.contains_key(&generation_pageserver) {
1292 0 : intent.set_attached(&mut scheduler, Some(generation_pageserver));
1293 0 : } else {
1294 : // If a node was removed before being completely drained, it is legal for it to leave behind a `generation_pageserver` referring
1295 : // to a non-existent node, because node deletion doesn't block on completing the reconciliations that will issue new generations
1296 : // on different pageservers.
1297 0 : tracing::warn!("Tenant shard {tenant_shard_id} references non-existent node {generation_pageserver} in database, will be rescheduled");
1298 : }
1299 0 : }
1300 0 : let new_tenant = TenantShard::from_persistent(tsp, intent)?;
1301 :
1302 0 : tenants.insert(tenant_shard_id, new_tenant);
1303 : }
1304 :
1305 0 : let (startup_completion, startup_complete) = utils::completion::channel();
1306 0 :
1307 0 : // This channel is continuously consumed by process_results, so doesn't need to be very large.
1308 0 : let (bg_compute_notify_result_tx, bg_compute_notify_result_rx) =
1309 0 : tokio::sync::mpsc::channel(512);
1310 0 :
1311 0 : let (delayed_reconcile_tx, delayed_reconcile_rx) =
1312 0 : tokio::sync::mpsc::channel(MAX_DELAYED_RECONCILES);
1313 0 :
1314 0 : let cancel = CancellationToken::new();
1315 0 : let reconcilers_cancel = cancel.child_token();
1316 0 :
1317 0 : let heartbeater = Heartbeater::new(
1318 0 : config.jwt_token.clone(),
1319 0 : config.max_offline_interval,
1320 0 : config.max_warming_up_interval,
1321 0 : cancel.clone(),
1322 0 : );
1323 :
1324 0 : let initial_leadership_status = if config.start_as_candidate {
1325 0 : LeadershipStatus::Candidate
1326 : } else {
1327 0 : LeadershipStatus::Leader
1328 : };
1329 :
1330 0 : let this = Arc::new(Self {
1331 0 : inner: Arc::new(std::sync::RwLock::new(ServiceState::new(
1332 0 : nodes,
1333 0 : tenants,
1334 0 : scheduler,
1335 0 : delayed_reconcile_rx,
1336 0 : initial_leadership_status,
1337 0 : ))),
1338 0 : config: config.clone(),
1339 0 : persistence,
1340 0 : compute_hook: Arc::new(ComputeHook::new(config.clone())),
1341 0 : result_tx,
1342 0 : heartbeater,
1343 0 : reconciler_concurrency: Arc::new(tokio::sync::Semaphore::new(
1344 0 : config.reconciler_concurrency,
1345 0 : )),
1346 0 : delayed_reconcile_tx,
1347 0 : abort_tx,
1348 0 : startup_complete: startup_complete.clone(),
1349 0 : cancel,
1350 0 : reconcilers_cancel,
1351 0 : gate: Gate::default(),
1352 0 : reconcilers_gate: Gate::default(),
1353 0 : tenant_op_locks: Default::default(),
1354 0 : node_op_locks: Default::default(),
1355 0 : });
1356 0 :
1357 0 : let result_task_this = this.clone();
1358 0 : tokio::task::spawn(async move {
1359 : // Block shutdown until we're done (we must respect self.cancel)
1360 0 : if let Ok(_gate) = result_task_this.gate.enter() {
1361 0 : result_task_this
1362 0 : .process_results(result_rx, bg_compute_notify_result_rx)
1363 0 : .await
1364 0 : }
1365 0 : });
1366 0 :
1367 0 : tokio::task::spawn({
1368 0 : let this = this.clone();
1369 0 : async move {
1370 : // Block shutdown until we're done (we must respect self.cancel)
1371 0 : if let Ok(_gate) = this.gate.enter() {
1372 0 : this.process_aborts(abort_rx).await
1373 0 : }
1374 0 : }
1375 0 : });
1376 0 :
1377 0 : tokio::task::spawn({
1378 0 : let this = this.clone();
1379 0 : async move {
1380 0 : if let Ok(_gate) = this.gate.enter() {
1381 : loop {
1382 0 : tokio::select! {
1383 0 : _ = this.cancel.cancelled() => {
1384 0 : break;
1385 : },
1386 0 : _ = tokio::time::sleep(Duration::from_secs(60)) => {}
1387 0 : };
1388 0 : this.tenant_op_locks.housekeeping();
1389 : }
1390 0 : }
1391 0 : }
1392 0 : });
1393 0 :
1394 0 : tokio::task::spawn({
1395 0 : let this = this.clone();
1396 0 : // We will block the [`Service::startup_complete`] barrier until [`Self::startup_reconcile`]
1397 0 : // is done.
1398 0 : let startup_completion = startup_completion.clone();
1399 0 : async move {
1400 : // Block shutdown until we're done (we must respect self.cancel)
1401 0 : let Ok(_gate) = this.gate.enter() else {
1402 0 : return;
1403 : };
1404 :
1405 0 : this.startup_reconcile(leader, leader_step_down_state, bg_compute_notify_result_tx)
1406 0 : .await;
1407 :
1408 0 : drop(startup_completion);
1409 0 : }
1410 0 : });
1411 0 :
1412 0 : tokio::task::spawn({
1413 0 : let this = this.clone();
1414 0 : let startup_complete = startup_complete.clone();
1415 0 : async move {
1416 0 : startup_complete.wait().await;
1417 0 : this.background_reconcile().await;
1418 0 : }
1419 0 : });
1420 0 :
1421 0 : tokio::task::spawn({
1422 0 : let this = this.clone();
1423 0 : let startup_complete = startup_complete.clone();
1424 0 : async move {
1425 0 : startup_complete.wait().await;
1426 0 : this.spawn_heartbeat_driver().await;
1427 0 : }
1428 0 : });
1429 0 :
1430 0 : Ok(this)
1431 0 : }
1432 :
1433 0 : pub(crate) async fn attach_hook(
1434 0 : &self,
1435 0 : attach_req: AttachHookRequest,
1436 0 : ) -> anyhow::Result<AttachHookResponse> {
1437 0 : let _tenant_lock = trace_exclusive_lock(
1438 0 : &self.tenant_op_locks,
1439 0 : attach_req.tenant_shard_id.tenant_id,
1440 0 : TenantOperations::AttachHook,
1441 0 : )
1442 0 : .await;
1443 :
1444 : // This is a test hook. To enable using it on tenants that were created directly with
1445 : // the pageserver API (not via this service), we will auto-create any missing tenant
1446 : // shards with default state.
1447 0 : let insert = {
1448 0 : let locked = self.inner.write().unwrap();
1449 0 : !locked.tenants.contains_key(&attach_req.tenant_shard_id)
1450 0 : };
1451 0 :
1452 0 : if insert {
1453 0 : let tsp = TenantShardPersistence {
1454 0 : tenant_id: attach_req.tenant_shard_id.tenant_id.to_string(),
1455 0 : shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
1456 0 : shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
1457 0 : shard_stripe_size: 0,
1458 0 : generation: attach_req.generation_override.or(Some(0)),
1459 0 : generation_pageserver: None,
1460 0 : placement_policy: serde_json::to_string(&PlacementPolicy::Attached(0)).unwrap(),
1461 0 : config: serde_json::to_string(&TenantConfig::default()).unwrap(),
1462 0 : splitting: SplitState::default(),
1463 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
1464 0 : .unwrap(),
1465 0 : preferred_az_id: None,
1466 0 : };
1467 0 :
1468 0 : match self.persistence.insert_tenant_shards(vec![tsp]).await {
1469 0 : Err(e) => match e {
1470 : DatabaseError::Query(diesel::result::Error::DatabaseError(
1471 : DatabaseErrorKind::UniqueViolation,
1472 : _,
1473 : )) => {
1474 0 : tracing::info!(
1475 0 : "Raced with another request to insert tenant {}",
1476 : attach_req.tenant_shard_id
1477 : )
1478 : }
1479 0 : _ => return Err(e.into()),
1480 : },
1481 : Ok(()) => {
1482 0 : tracing::info!("Inserted shard {} in database", attach_req.tenant_shard_id);
1483 :
1484 0 : let mut locked = self.inner.write().unwrap();
1485 0 : locked.tenants.insert(
1486 0 : attach_req.tenant_shard_id,
1487 0 : TenantShard::new(
1488 0 : attach_req.tenant_shard_id,
1489 0 : ShardIdentity::unsharded(),
1490 0 : PlacementPolicy::Attached(0),
1491 0 : ),
1492 0 : );
1493 0 : tracing::info!("Inserted shard {} in memory", attach_req.tenant_shard_id);
1494 : }
1495 : }
1496 0 : }
1497 :
1498 0 : let new_generation = if let Some(req_node_id) = attach_req.node_id {
1499 0 : let maybe_tenant_conf = {
1500 0 : let locked = self.inner.write().unwrap();
1501 0 : locked
1502 0 : .tenants
1503 0 : .get(&attach_req.tenant_shard_id)
1504 0 : .map(|t| t.config.clone())
1505 0 : };
1506 0 :
1507 0 : match maybe_tenant_conf {
1508 0 : Some(conf) => {
1509 0 : let new_generation = self
1510 0 : .persistence
1511 0 : .increment_generation(attach_req.tenant_shard_id, req_node_id)
1512 0 : .await?;
1513 :
1514 : // Persist the placement policy update. This is required
1515 : // when we reattaching a detached tenant.
1516 0 : self.persistence
1517 0 : .update_tenant_shard(
1518 0 : TenantFilter::Shard(attach_req.tenant_shard_id),
1519 0 : Some(PlacementPolicy::Attached(0)),
1520 0 : Some(conf),
1521 0 : None,
1522 0 : None,
1523 0 : )
1524 0 : .await?;
1525 0 : Some(new_generation)
1526 : }
1527 : None => {
1528 0 : anyhow::bail!("Attach hook handling raced with tenant removal")
1529 : }
1530 : }
1531 : } else {
1532 0 : self.persistence.detach(attach_req.tenant_shard_id).await?;
1533 0 : None
1534 : };
1535 :
1536 0 : let mut locked = self.inner.write().unwrap();
1537 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
1538 0 :
1539 0 : let tenant_shard = tenants
1540 0 : .get_mut(&attach_req.tenant_shard_id)
1541 0 : .expect("Checked for existence above");
1542 :
1543 0 : if let Some(new_generation) = new_generation {
1544 0 : tenant_shard.generation = Some(new_generation);
1545 0 : tenant_shard.policy = PlacementPolicy::Attached(0);
1546 0 : } else {
1547 : // This is a detach notification. We must update placement policy to avoid re-attaching
1548 : // during background scheduling/reconciliation, or during storage controller restart.
1549 0 : assert!(attach_req.node_id.is_none());
1550 0 : tenant_shard.policy = PlacementPolicy::Detached;
1551 : }
1552 :
1553 0 : if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
1554 0 : tracing::info!(
1555 : tenant_id = %attach_req.tenant_shard_id,
1556 : ps_id = %attaching_pageserver,
1557 : generation = ?tenant_shard.generation,
1558 0 : "issuing",
1559 : );
1560 0 : } else if let Some(ps_id) = tenant_shard.intent.get_attached() {
1561 0 : tracing::info!(
1562 : tenant_id = %attach_req.tenant_shard_id,
1563 : %ps_id,
1564 : generation = ?tenant_shard.generation,
1565 0 : "dropping",
1566 : );
1567 : } else {
1568 0 : tracing::info!(
1569 : tenant_id = %attach_req.tenant_shard_id,
1570 0 : "no-op: tenant already has no pageserver");
1571 : }
1572 0 : tenant_shard
1573 0 : .intent
1574 0 : .set_attached(scheduler, attach_req.node_id);
1575 0 :
1576 0 : tracing::info!(
1577 0 : "attach_hook: tenant {} set generation {:?}, pageserver {}",
1578 0 : attach_req.tenant_shard_id,
1579 0 : tenant_shard.generation,
1580 0 : // TODO: this is an odd number of 0xf's
1581 0 : attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
1582 : );
1583 :
1584 : // Trick the reconciler into not doing anything for this tenant: this helps
1585 : // tests that manually configure a tenant on the pagesrever, and then call this
1586 : // attach hook: they don't want background reconciliation to modify what they
1587 : // did to the pageserver.
1588 : #[cfg(feature = "testing")]
1589 : {
1590 0 : if let Some(node_id) = attach_req.node_id {
1591 0 : tenant_shard.observed.locations = HashMap::from([(
1592 0 : node_id,
1593 0 : ObservedStateLocation {
1594 0 : conf: Some(attached_location_conf(
1595 0 : tenant_shard.generation.unwrap(),
1596 0 : &tenant_shard.shard,
1597 0 : &tenant_shard.config,
1598 0 : &PlacementPolicy::Attached(0),
1599 0 : )),
1600 0 : },
1601 0 : )]);
1602 0 : } else {
1603 0 : tenant_shard.observed.locations.clear();
1604 0 : }
1605 : }
1606 :
1607 0 : Ok(AttachHookResponse {
1608 0 : gen: attach_req
1609 0 : .node_id
1610 0 : .map(|_| tenant_shard.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
1611 0 : })
1612 0 : }
1613 :
1614 0 : pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
1615 0 : let locked = self.inner.read().unwrap();
1616 0 :
1617 0 : let tenant_shard = locked.tenants.get(&inspect_req.tenant_shard_id);
1618 0 :
1619 0 : InspectResponse {
1620 0 : attachment: tenant_shard.and_then(|s| {
1621 0 : s.intent
1622 0 : .get_attached()
1623 0 : .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
1624 0 : }),
1625 0 : }
1626 0 : }
1627 :
1628 : // When the availability state of a node transitions to active, we must do a full reconciliation
1629 : // of LocationConfigs on that node. This is because while a node was offline:
1630 : // - we might have proceeded through startup_reconcile without checking for extraneous LocationConfigs on this node
1631 : // - aborting a tenant shard split might have left rogue child shards behind on this node.
1632 : //
1633 : // This function must complete _before_ setting a `Node` to Active: once it is set to Active, other
1634 : // Reconcilers might communicate with the node, and these must not overlap with the work we do in
1635 : // this function.
1636 : //
1637 : // The reconciliation logic in here is very similar to what [`Self::startup_reconcile`] does, but
1638 : // for written for a single node rather than as a batch job for all nodes.
1639 0 : #[tracing::instrument(skip_all, fields(node_id=%node.get_id()))]
1640 : async fn node_activate_reconcile(
1641 : &self,
1642 : mut node: Node,
1643 : _lock: &TracingExclusiveGuard<NodeOperations>,
1644 : ) -> Result<(), ApiError> {
1645 : // This Node is a mutable local copy: we will set it active so that we can use its
1646 : // API client to reconcile with the node. The Node in [`Self::nodes`] will get updated
1647 : // later.
1648 : node.set_availability(NodeAvailability::Active(PageserverUtilization::full()));
1649 :
1650 : let configs = match node
1651 : .with_client_retries(
1652 0 : |client| async move { client.list_location_config().await },
1653 : &self.config.jwt_token,
1654 : 1,
1655 : 5,
1656 : SHORT_RECONCILE_TIMEOUT,
1657 : &self.cancel,
1658 : )
1659 : .await
1660 : {
1661 : None => {
1662 : // We're shutting down (the Node's cancellation token can't have fired, because
1663 : // we're the only scope that has a reference to it, and we didn't fire it).
1664 : return Err(ApiError::ShuttingDown);
1665 : }
1666 : Some(Err(e)) => {
1667 : // This node didn't succeed listing its locations: it may not proceed to active state
1668 : // as it is apparently unavailable.
1669 : return Err(ApiError::PreconditionFailed(
1670 : format!("Failed to query node location configs, cannot activate ({e})").into(),
1671 : ));
1672 : }
1673 : Some(Ok(configs)) => configs,
1674 : };
1675 : tracing::info!("Loaded {} LocationConfigs", configs.tenant_shards.len());
1676 :
1677 : let mut cleanup = Vec::new();
1678 : {
1679 : let mut locked = self.inner.write().unwrap();
1680 :
1681 : for (tenant_shard_id, observed_loc) in configs.tenant_shards {
1682 : let Some(tenant_shard) = locked.tenants.get_mut(&tenant_shard_id) else {
1683 : cleanup.push(tenant_shard_id);
1684 : continue;
1685 : };
1686 : tenant_shard
1687 : .observed
1688 : .locations
1689 : .insert(node.get_id(), ObservedStateLocation { conf: observed_loc });
1690 : }
1691 : }
1692 :
1693 : for tenant_shard_id in cleanup {
1694 : tracing::info!("Detaching {tenant_shard_id}");
1695 : match node
1696 : .with_client_retries(
1697 0 : |client| async move {
1698 0 : let config = LocationConfig {
1699 0 : mode: LocationConfigMode::Detached,
1700 0 : generation: None,
1701 0 : secondary_conf: None,
1702 0 : shard_number: tenant_shard_id.shard_number.0,
1703 0 : shard_count: tenant_shard_id.shard_count.literal(),
1704 0 : shard_stripe_size: 0,
1705 0 : tenant_conf: models::TenantConfig::default(),
1706 0 : };
1707 0 : client
1708 0 : .location_config(tenant_shard_id, config, None, false)
1709 0 : .await
1710 0 : },
1711 : &self.config.jwt_token,
1712 : 1,
1713 : 5,
1714 : SHORT_RECONCILE_TIMEOUT,
1715 : &self.cancel,
1716 : )
1717 : .await
1718 : {
1719 : None => {
1720 : // We're shutting down (the Node's cancellation token can't have fired, because
1721 : // we're the only scope that has a reference to it, and we didn't fire it).
1722 : return Err(ApiError::ShuttingDown);
1723 : }
1724 : Some(Err(e)) => {
1725 : // Do not let the node proceed to Active state if it is not responsive to requests
1726 : // to detach. This could happen if e.g. a shutdown bug in the pageserver is preventing
1727 : // detach completing: we should not let this node back into the set of nodes considered
1728 : // okay for scheduling.
1729 : return Err(ApiError::Conflict(format!(
1730 : "Node {node} failed to detach {tenant_shard_id}: {e}"
1731 : )));
1732 : }
1733 : Some(Ok(_)) => {}
1734 : };
1735 : }
1736 :
1737 : Ok(())
1738 : }
1739 :
1740 0 : pub(crate) async fn re_attach(
1741 0 : &self,
1742 0 : reattach_req: ReAttachRequest,
1743 0 : ) -> Result<ReAttachResponse, ApiError> {
1744 0 : if let Some(register_req) = reattach_req.register {
1745 0 : self.node_register(register_req).await?;
1746 0 : }
1747 :
1748 : // Ordering: we must persist generation number updates before making them visible in the in-memory state
1749 0 : let incremented_generations = self.persistence.re_attach(reattach_req.node_id).await?;
1750 :
1751 0 : tracing::info!(
1752 : node_id=%reattach_req.node_id,
1753 0 : "Incremented {} tenant shards' generations",
1754 0 : incremented_generations.len()
1755 : );
1756 :
1757 : // Apply the updated generation to our in-memory state, and
1758 : // gather discover secondary locations.
1759 0 : let mut locked = self.inner.write().unwrap();
1760 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
1761 0 :
1762 0 : let mut response = ReAttachResponse {
1763 0 : tenants: Vec::new(),
1764 0 : };
1765 :
1766 : // TODO: cancel/restart any running reconciliation for this tenant, it might be trying
1767 : // to call location_conf API with an old generation. Wait for cancellation to complete
1768 : // before responding to this request. Requires well implemented CancellationToken logic
1769 : // all the way to where we call location_conf. Even then, there can still be a location_conf
1770 : // request in flight over the network: TODO handle that by making location_conf API refuse
1771 : // to go backward in generations.
1772 :
1773 : // Scan through all shards, applying updates for ones where we updated generation
1774 : // and identifying shards that intend to have a secondary location on this node.
1775 0 : for (tenant_shard_id, shard) in tenants {
1776 0 : if let Some(new_gen) = incremented_generations.get(tenant_shard_id) {
1777 0 : let new_gen = *new_gen;
1778 0 : response.tenants.push(ReAttachResponseTenant {
1779 0 : id: *tenant_shard_id,
1780 0 : gen: Some(new_gen.into().unwrap()),
1781 0 : // A tenant is only put into multi or stale modes in the middle of a [`Reconciler::live_migrate`]
1782 0 : // execution. If a pageserver is restarted during that process, then the reconcile pass will
1783 0 : // fail, and start from scratch, so it doesn't make sense for us to try and preserve
1784 0 : // the stale/multi states at this point.
1785 0 : mode: LocationConfigMode::AttachedSingle,
1786 0 : });
1787 0 :
1788 0 : shard.generation = std::cmp::max(shard.generation, Some(new_gen));
1789 0 : if let Some(observed) = shard.observed.locations.get_mut(&reattach_req.node_id) {
1790 : // Why can we update `observed` even though we're not sure our response will be received
1791 : // by the pageserver? Because the pageserver will not proceed with startup until
1792 : // it has processed response: if it loses it, we'll see another request and increment
1793 : // generation again, avoiding any uncertainty about dirtiness of tenant's state.
1794 0 : if let Some(conf) = observed.conf.as_mut() {
1795 0 : conf.generation = new_gen.into();
1796 0 : }
1797 0 : } else {
1798 0 : // This node has no observed state for the shard: perhaps it was offline
1799 0 : // when the pageserver restarted. Insert a None, so that the Reconciler
1800 0 : // will be prompted to learn the location's state before it makes changes.
1801 0 : shard
1802 0 : .observed
1803 0 : .locations
1804 0 : .insert(reattach_req.node_id, ObservedStateLocation { conf: None });
1805 0 : }
1806 0 : } else if shard.intent.get_secondary().contains(&reattach_req.node_id) {
1807 0 : // Ordering: pageserver will not accept /location_config requests until it has
1808 0 : // finished processing the response from re-attach. So we can update our in-memory state
1809 0 : // now, and be confident that we are not stamping on the result of some later location config.
1810 0 : // TODO: however, we are not strictly ordered wrt ReconcileResults queue,
1811 0 : // so we might update observed state here, and then get over-written by some racing
1812 0 : // ReconcileResult. The impact is low however, since we have set state on pageserver something
1813 0 : // that matches intent, so worst case if we race then we end up doing a spurious reconcile.
1814 0 :
1815 0 : response.tenants.push(ReAttachResponseTenant {
1816 0 : id: *tenant_shard_id,
1817 0 : gen: None,
1818 0 : mode: LocationConfigMode::Secondary,
1819 0 : });
1820 0 :
1821 0 : // We must not update observed, because we have no guarantee that our
1822 0 : // response will be received by the pageserver. This could leave it
1823 0 : // falsely dirty, but the resulting reconcile should be idempotent.
1824 0 : }
1825 : }
1826 :
1827 : // We consider a node Active once we have composed a re-attach response, but we
1828 : // do not call [`Self::node_activate_reconcile`]: the handling of the re-attach response
1829 : // implicitly synchronizes the LocationConfigs on the node.
1830 : //
1831 : // Setting a node active unblocks any Reconcilers that might write to the location config API,
1832 : // but those requests will not be accepted by the node until it has finished processing
1833 : // the re-attach response.
1834 : //
1835 : // Additionally, reset the nodes scheduling policy to match the conditional update done
1836 : // in [`Persistence::re_attach`].
1837 0 : if let Some(node) = nodes.get(&reattach_req.node_id) {
1838 0 : let reset_scheduling = matches!(
1839 0 : node.get_scheduling(),
1840 : NodeSchedulingPolicy::PauseForRestart
1841 : | NodeSchedulingPolicy::Draining
1842 : | NodeSchedulingPolicy::Filling
1843 : );
1844 :
1845 0 : let mut new_nodes = (**nodes).clone();
1846 0 : if let Some(node) = new_nodes.get_mut(&reattach_req.node_id) {
1847 0 : if reset_scheduling {
1848 0 : node.set_scheduling(NodeSchedulingPolicy::Active);
1849 0 : }
1850 :
1851 0 : tracing::info!("Marking {} warming-up on reattach", reattach_req.node_id);
1852 0 : node.set_availability(NodeAvailability::WarmingUp(std::time::Instant::now()));
1853 0 :
1854 0 : scheduler.node_upsert(node);
1855 0 : let new_nodes = Arc::new(new_nodes);
1856 0 : *nodes = new_nodes;
1857 : } else {
1858 0 : tracing::error!(
1859 0 : "Reattaching node {} was removed while processing the request",
1860 : reattach_req.node_id
1861 : );
1862 : }
1863 0 : }
1864 :
1865 0 : Ok(response)
1866 0 : }
1867 :
1868 0 : pub(crate) async fn validate(
1869 0 : &self,
1870 0 : validate_req: ValidateRequest,
1871 0 : ) -> Result<ValidateResponse, DatabaseError> {
1872 : // Fast in-memory check: we may reject validation on anything that doesn't match our
1873 : // in-memory generation for a shard
1874 0 : let in_memory_result = {
1875 0 : let mut in_memory_result = Vec::new();
1876 0 : let locked = self.inner.read().unwrap();
1877 0 : for req_tenant in validate_req.tenants {
1878 0 : if let Some(tenant_shard) = locked.tenants.get(&req_tenant.id) {
1879 0 : let valid = tenant_shard.generation == Some(Generation::new(req_tenant.gen));
1880 0 : tracing::info!(
1881 0 : "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
1882 : req_tenant.id,
1883 : req_tenant.gen,
1884 : tenant_shard.generation
1885 : );
1886 :
1887 0 : in_memory_result.push((req_tenant.id, Generation::new(req_tenant.gen), valid));
1888 : } else {
1889 : // This is legal: for example during a shard split the pageserver may still
1890 : // have deletions in its queue from the old pre-split shard, or after deletion
1891 : // of a tenant that was busy with compaction/gc while being deleted.
1892 0 : tracing::info!(
1893 0 : "Refusing deletion validation for missing shard {}",
1894 : req_tenant.id
1895 : );
1896 : }
1897 : }
1898 :
1899 0 : in_memory_result
1900 : };
1901 :
1902 : // Database calls to confirm validity for anything that passed the in-memory check. We must do this
1903 : // in case of controller split-brain, where some other controller process might have incremented the generation.
1904 0 : let db_generations = self
1905 0 : .persistence
1906 0 : .shard_generations(in_memory_result.iter().filter_map(|i| {
1907 0 : if i.2 {
1908 0 : Some(&i.0)
1909 : } else {
1910 0 : None
1911 : }
1912 0 : }))
1913 0 : .await?;
1914 0 : let db_generations = db_generations.into_iter().collect::<HashMap<_, _>>();
1915 0 :
1916 0 : let mut response = ValidateResponse {
1917 0 : tenants: Vec::new(),
1918 0 : };
1919 0 : for (tenant_shard_id, validate_generation, valid) in in_memory_result.into_iter() {
1920 0 : let valid = if valid {
1921 0 : let db_generation = db_generations.get(&tenant_shard_id);
1922 0 : db_generation == Some(&Some(validate_generation))
1923 : } else {
1924 : // If in-memory state says it's invalid, trust that. It's always safe to fail a validation, at worst
1925 : // this prevents a pageserver from cleaning up an object in S3.
1926 0 : false
1927 : };
1928 :
1929 0 : response.tenants.push(ValidateResponseTenant {
1930 0 : id: tenant_shard_id,
1931 0 : valid,
1932 0 : })
1933 : }
1934 :
1935 0 : Ok(response)
1936 0 : }
1937 :
1938 0 : pub(crate) async fn tenant_create(
1939 0 : &self,
1940 0 : create_req: TenantCreateRequest,
1941 0 : ) -> Result<TenantCreateResponse, ApiError> {
1942 0 : let tenant_id = create_req.new_tenant_id.tenant_id;
1943 :
1944 : // Exclude any concurrent attempts to create/access the same tenant ID
1945 0 : let _tenant_lock = trace_exclusive_lock(
1946 0 : &self.tenant_op_locks,
1947 0 : create_req.new_tenant_id.tenant_id,
1948 0 : TenantOperations::Create,
1949 0 : )
1950 0 : .await;
1951 0 : let (response, waiters) = self.do_tenant_create(create_req).await?;
1952 :
1953 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
1954 : // Avoid deadlock: reconcile may fail while notifying compute, if the cloud control plane refuses to
1955 : // accept compute notifications while it is in the process of creating. Reconciliation will
1956 : // be retried in the background.
1957 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while creating tenant ({e})");
1958 0 : }
1959 0 : Ok(response)
1960 0 : }
1961 :
1962 0 : pub(crate) async fn do_tenant_create(
1963 0 : &self,
1964 0 : create_req: TenantCreateRequest,
1965 0 : ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
1966 0 : let placement_policy = create_req
1967 0 : .placement_policy
1968 0 : .clone()
1969 0 : // As a default, zero secondaries is convenient for tests that don't choose a policy.
1970 0 : .unwrap_or(PlacementPolicy::Attached(0));
1971 :
1972 : // This service expects to handle sharding itself: it is an error to try and directly create
1973 : // a particular shard here.
1974 0 : let tenant_id = if !create_req.new_tenant_id.is_unsharded() {
1975 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
1976 0 : "Attempted to create a specific shard, this API is for creating the whole tenant"
1977 0 : )));
1978 : } else {
1979 0 : create_req.new_tenant_id.tenant_id
1980 0 : };
1981 0 :
1982 0 : tracing::info!(
1983 0 : "Creating tenant {}, shard_count={:?}",
1984 : create_req.new_tenant_id,
1985 : create_req.shard_parameters.count,
1986 : );
1987 :
1988 0 : let create_ids = (0..create_req.shard_parameters.count.count())
1989 0 : .map(|i| TenantShardId {
1990 0 : tenant_id,
1991 0 : shard_number: ShardNumber(i),
1992 0 : shard_count: create_req.shard_parameters.count,
1993 0 : })
1994 0 : .collect::<Vec<_>>();
1995 :
1996 : // If the caller specifies a None generation, it means "start from default". This is different
1997 : // to [`Self::tenant_location_config`], where a None generation is used to represent
1998 : // an incompletely-onboarded tenant.
1999 0 : let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
2000 0 : tracing::info!(
2001 0 : "tenant_create: secondary mode, generation is_some={}",
2002 0 : create_req.generation.is_some()
2003 : );
2004 0 : create_req.generation.map(Generation::new)
2005 : } else {
2006 0 : tracing::info!(
2007 0 : "tenant_create: not secondary mode, generation is_some={}",
2008 0 : create_req.generation.is_some()
2009 : );
2010 0 : Some(
2011 0 : create_req
2012 0 : .generation
2013 0 : .map(Generation::new)
2014 0 : .unwrap_or(INITIAL_GENERATION),
2015 0 : )
2016 : };
2017 :
2018 : // Ordering: we persist tenant shards before creating them on the pageserver. This enables a caller
2019 : // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
2020 : // during the creation, rather than risking leaving orphan objects in S3.
2021 0 : let persist_tenant_shards = create_ids
2022 0 : .iter()
2023 0 : .map(|tenant_shard_id| TenantShardPersistence {
2024 0 : tenant_id: tenant_shard_id.tenant_id.to_string(),
2025 0 : shard_number: tenant_shard_id.shard_number.0 as i32,
2026 0 : shard_count: tenant_shard_id.shard_count.literal() as i32,
2027 0 : shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
2028 0 : generation: initial_generation.map(|g| g.into().unwrap() as i32),
2029 0 : // The pageserver is not known until scheduling happens: we will set this column when
2030 0 : // incrementing the generation the first time we attach to a pageserver.
2031 0 : generation_pageserver: None,
2032 0 : placement_policy: serde_json::to_string(&placement_policy).unwrap(),
2033 0 : config: serde_json::to_string(&create_req.config).unwrap(),
2034 0 : splitting: SplitState::default(),
2035 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
2036 0 : .unwrap(),
2037 0 : preferred_az_id: None,
2038 0 : })
2039 0 : .collect();
2040 0 :
2041 0 : match self
2042 0 : .persistence
2043 0 : .insert_tenant_shards(persist_tenant_shards)
2044 0 : .await
2045 : {
2046 0 : Ok(_) => {}
2047 : Err(DatabaseError::Query(diesel::result::Error::DatabaseError(
2048 : DatabaseErrorKind::UniqueViolation,
2049 : _,
2050 : ))) => {
2051 : // Unique key violation: this is probably a retry. Because the shard count is part of the unique key,
2052 : // if we see a unique key violation it means that the creation request's shard count matches the previous
2053 : // creation's shard count.
2054 0 : tracing::info!("Tenant shards already present in database, proceeding with idempotent creation...");
2055 : }
2056 : // Any other database error is unexpected and a bug.
2057 0 : Err(e) => return Err(ApiError::InternalServerError(anyhow::anyhow!(e))),
2058 : };
2059 :
2060 0 : let mut schedule_context = ScheduleContext::default();
2061 0 : let mut schedule_error = None;
2062 0 : let mut response_shards = Vec::new();
2063 0 : for tenant_shard_id in create_ids {
2064 0 : tracing::info!("Creating shard {tenant_shard_id}...");
2065 :
2066 0 : let outcome = self
2067 0 : .do_initial_shard_scheduling(
2068 0 : tenant_shard_id,
2069 0 : initial_generation,
2070 0 : &create_req.shard_parameters,
2071 0 : create_req.config.clone(),
2072 0 : placement_policy.clone(),
2073 0 : &mut schedule_context,
2074 0 : )
2075 0 : .await;
2076 :
2077 0 : match outcome {
2078 0 : InitialShardScheduleOutcome::Scheduled(resp) => response_shards.push(resp),
2079 0 : InitialShardScheduleOutcome::NotScheduled => {}
2080 0 : InitialShardScheduleOutcome::ShardScheduleError(err) => {
2081 0 : schedule_error = Some(err);
2082 0 : }
2083 : }
2084 : }
2085 :
2086 0 : let preferred_azs = {
2087 0 : let locked = self.inner.read().unwrap();
2088 0 : response_shards
2089 0 : .iter()
2090 0 : .filter_map(|resp| {
2091 0 : let az_id = locked
2092 0 : .nodes
2093 0 : .get(&resp.node_id)
2094 0 : .map(|n| n.get_availability_zone_id().to_string())?;
2095 :
2096 0 : Some((resp.shard_id, az_id))
2097 0 : })
2098 0 : .collect::<Vec<_>>()
2099 : };
2100 :
2101 : // Note that we persist the preferred AZ for the new shards separately.
2102 : // In theory, we could "peek" the scheduler to determine where the shard will
2103 : // land, but the subsequent "real" call into the scheduler might select a different
2104 : // node. Hence, we do this awkward update to keep things consistent.
2105 0 : let updated = self
2106 0 : .persistence
2107 0 : .set_tenant_shard_preferred_azs(preferred_azs)
2108 0 : .await
2109 0 : .map_err(|err| {
2110 0 : ApiError::InternalServerError(anyhow::anyhow!(
2111 0 : "Failed to persist preferred az ids: {err}"
2112 0 : ))
2113 0 : })?;
2114 :
2115 : {
2116 0 : let mut locked = self.inner.write().unwrap();
2117 0 : for (tid, az_id) in updated {
2118 0 : if let Some(shard) = locked.tenants.get_mut(&tid) {
2119 0 : shard.set_preferred_az(az_id);
2120 0 : }
2121 : }
2122 : }
2123 :
2124 : // If we failed to schedule shards, then they are still created in the controller,
2125 : // but we return an error to the requester to avoid a silent failure when someone
2126 : // tries to e.g. create a tenant whose placement policy requires more nodes than
2127 : // are present in the system. We do this here rather than in the above loop, to
2128 : // avoid situations where we only create a subset of shards in the tenant.
2129 0 : if let Some(e) = schedule_error {
2130 0 : return Err(ApiError::Conflict(format!(
2131 0 : "Failed to schedule shard(s): {e}"
2132 0 : )));
2133 0 : }
2134 0 :
2135 0 : let waiters = {
2136 0 : let mut locked = self.inner.write().unwrap();
2137 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2138 0 : tenants
2139 0 : .range_mut(TenantShardId::tenant_range(tenant_id))
2140 0 : .filter_map(|(_shard_id, shard)| self.maybe_reconcile_shard(shard, nodes))
2141 0 : .collect::<Vec<_>>()
2142 0 : };
2143 0 :
2144 0 : Ok((
2145 0 : TenantCreateResponse {
2146 0 : shards: response_shards,
2147 0 : },
2148 0 : waiters,
2149 0 : ))
2150 0 : }
2151 :
2152 : /// Helper for tenant creation that does the scheduling for an individual shard. Covers both the
2153 : /// case of a new tenant and a pre-existing one.
2154 0 : async fn do_initial_shard_scheduling(
2155 0 : &self,
2156 0 : tenant_shard_id: TenantShardId,
2157 0 : initial_generation: Option<Generation>,
2158 0 : shard_params: &ShardParameters,
2159 0 : config: TenantConfig,
2160 0 : placement_policy: PlacementPolicy,
2161 0 : schedule_context: &mut ScheduleContext,
2162 0 : ) -> InitialShardScheduleOutcome {
2163 0 : let mut locked = self.inner.write().unwrap();
2164 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2165 :
2166 : use std::collections::btree_map::Entry;
2167 0 : match tenants.entry(tenant_shard_id) {
2168 0 : Entry::Occupied(mut entry) => {
2169 0 : tracing::info!("Tenant shard {tenant_shard_id} already exists while creating");
2170 :
2171 : // TODO: schedule() should take an anti-affinity expression that pushes
2172 : // attached and secondary locations (independently) away frorm those
2173 : // pageservers also holding a shard for this tenant.
2174 :
2175 0 : if let Err(err) = entry.get_mut().schedule(scheduler, schedule_context) {
2176 0 : return InitialShardScheduleOutcome::ShardScheduleError(err);
2177 0 : }
2178 :
2179 0 : if let Some(node_id) = entry.get().intent.get_attached() {
2180 0 : let generation = entry
2181 0 : .get()
2182 0 : .generation
2183 0 : .expect("Generation is set when in attached mode");
2184 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2185 0 : shard_id: tenant_shard_id,
2186 0 : node_id: *node_id,
2187 0 : generation: generation.into().unwrap(),
2188 0 : })
2189 : } else {
2190 0 : InitialShardScheduleOutcome::NotScheduled
2191 : }
2192 : }
2193 0 : Entry::Vacant(entry) => {
2194 0 : let state = entry.insert(TenantShard::new(
2195 0 : tenant_shard_id,
2196 0 : ShardIdentity::from_params(tenant_shard_id.shard_number, shard_params),
2197 0 : placement_policy,
2198 0 : ));
2199 0 :
2200 0 : state.generation = initial_generation;
2201 0 : state.config = config;
2202 0 : if let Err(e) = state.schedule(scheduler, schedule_context) {
2203 0 : return InitialShardScheduleOutcome::ShardScheduleError(e);
2204 0 : }
2205 :
2206 : // Only include shards in result if we are attaching: the purpose
2207 : // of the response is to tell the caller where the shards are attached.
2208 0 : if let Some(node_id) = state.intent.get_attached() {
2209 0 : let generation = state
2210 0 : .generation
2211 0 : .expect("Generation is set when in attached mode");
2212 0 : InitialShardScheduleOutcome::Scheduled(TenantCreateResponseShard {
2213 0 : shard_id: tenant_shard_id,
2214 0 : node_id: *node_id,
2215 0 : generation: generation.into().unwrap(),
2216 0 : })
2217 : } else {
2218 0 : InitialShardScheduleOutcome::NotScheduled
2219 : }
2220 : }
2221 : }
2222 0 : }
2223 :
2224 : /// Helper for functions that reconcile a number of shards, and would like to do a timeout-bounded
2225 : /// wait for reconciliation to complete before responding.
2226 0 : async fn await_waiters(
2227 0 : &self,
2228 0 : waiters: Vec<ReconcilerWaiter>,
2229 0 : timeout: Duration,
2230 0 : ) -> Result<(), ReconcileWaitError> {
2231 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2232 0 : for waiter in waiters {
2233 0 : let timeout = deadline.duration_since(Instant::now());
2234 0 : waiter.wait_timeout(timeout).await?;
2235 : }
2236 :
2237 0 : Ok(())
2238 0 : }
2239 :
2240 : /// Same as [`Service::await_waiters`], but returns the waiters which are still
2241 : /// in progress
2242 0 : async fn await_waiters_remainder(
2243 0 : &self,
2244 0 : waiters: Vec<ReconcilerWaiter>,
2245 0 : timeout: Duration,
2246 0 : ) -> Vec<ReconcilerWaiter> {
2247 0 : let deadline = Instant::now().checked_add(timeout).unwrap();
2248 0 : for waiter in waiters.iter() {
2249 0 : let timeout = deadline.duration_since(Instant::now());
2250 0 : let _ = waiter.wait_timeout(timeout).await;
2251 : }
2252 :
2253 0 : waiters
2254 0 : .into_iter()
2255 0 : .filter(|waiter| matches!(waiter.get_status(), ReconcilerStatus::InProgress))
2256 0 : .collect::<Vec<_>>()
2257 0 : }
2258 :
2259 : /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
2260 : /// and transform it into either a tenant creation of a series of shard updates.
2261 : ///
2262 : /// If the incoming request makes no changes, a [`TenantCreateOrUpdate::Update`] result will
2263 : /// still be returned.
2264 0 : fn tenant_location_config_prepare(
2265 0 : &self,
2266 0 : tenant_id: TenantId,
2267 0 : req: TenantLocationConfigRequest,
2268 0 : ) -> TenantCreateOrUpdate {
2269 0 : let mut updates = Vec::new();
2270 0 : let mut locked = self.inner.write().unwrap();
2271 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2272 0 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2273 :
2274 : // Use location config mode as an indicator of policy.
2275 0 : let placement_policy = match req.config.mode {
2276 0 : LocationConfigMode::Detached => PlacementPolicy::Detached,
2277 0 : LocationConfigMode::Secondary => PlacementPolicy::Secondary,
2278 : LocationConfigMode::AttachedMulti
2279 : | LocationConfigMode::AttachedSingle
2280 : | LocationConfigMode::AttachedStale => {
2281 0 : if nodes.len() > 1 {
2282 0 : PlacementPolicy::Attached(1)
2283 : } else {
2284 : // Convenience for dev/test: if we just have one pageserver, import
2285 : // tenants into non-HA mode so that scheduling will succeed.
2286 0 : PlacementPolicy::Attached(0)
2287 : }
2288 : }
2289 : };
2290 :
2291 0 : let mut create = true;
2292 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2293 : // Saw an existing shard: this is not a creation
2294 0 : create = false;
2295 :
2296 : // Shards may have initially been created by a Secondary request, where we
2297 : // would have left generation as None.
2298 : //
2299 : // We only update generation the first time we see an attached-mode request,
2300 : // and if there is no existing generation set. The caller is responsible for
2301 : // ensuring that no non-storage-controller pageserver ever uses a higher
2302 : // generation than they passed in here.
2303 : use LocationConfigMode::*;
2304 0 : let set_generation = match req.config.mode {
2305 0 : AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
2306 0 : req.config.generation.map(Generation::new)
2307 : }
2308 0 : _ => None,
2309 : };
2310 :
2311 0 : updates.push(ShardUpdate {
2312 0 : tenant_shard_id: *shard_id,
2313 0 : placement_policy: placement_policy.clone(),
2314 0 : tenant_config: req.config.tenant_conf.clone(),
2315 0 : generation: set_generation,
2316 0 : });
2317 : }
2318 :
2319 0 : if create {
2320 : use LocationConfigMode::*;
2321 0 : let generation = match req.config.mode {
2322 0 : AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
2323 : // If a caller provided a generation in a non-attached request, ignore it
2324 : // and leave our generation as None: this enables a subsequent update to set
2325 : // the generation when setting an attached mode for the first time.
2326 0 : _ => None,
2327 : };
2328 :
2329 0 : TenantCreateOrUpdate::Create(
2330 0 : // Synthesize a creation request
2331 0 : TenantCreateRequest {
2332 0 : new_tenant_id: tenant_shard_id,
2333 0 : generation,
2334 0 : shard_parameters: ShardParameters {
2335 0 : count: tenant_shard_id.shard_count,
2336 0 : // We only import un-sharded or single-sharded tenants, so stripe
2337 0 : // size can be made up arbitrarily here.
2338 0 : stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
2339 0 : },
2340 0 : placement_policy: Some(placement_policy),
2341 0 : config: req.config.tenant_conf,
2342 0 : },
2343 0 : )
2344 : } else {
2345 0 : assert!(!updates.is_empty());
2346 0 : TenantCreateOrUpdate::Update(updates)
2347 : }
2348 0 : }
2349 :
2350 : /// This API is used by the cloud control plane to migrate unsharded tenants that it created
2351 : /// directly with pageservers into this service.
2352 : ///
2353 : /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
2354 : /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
2355 : /// Think of the first attempt to call this API as a transfer of absolute authority over the
2356 : /// tenant's source of generation numbers.
2357 : ///
2358 : /// The mode in this request coarse-grained control of tenants:
2359 : /// - Call with mode Attached* to upsert the tenant.
2360 : /// - Call with mode Secondary to either onboard a tenant without attaching it, or
2361 : /// to set an existing tenant to PolicyMode::Secondary
2362 : /// - Call with mode Detached to switch to PolicyMode::Detached
2363 0 : pub(crate) async fn tenant_location_config(
2364 0 : &self,
2365 0 : tenant_shard_id: TenantShardId,
2366 0 : req: TenantLocationConfigRequest,
2367 0 : ) -> Result<TenantLocationConfigResponse, ApiError> {
2368 : // We require an exclusive lock, because we are updating both persistent and in-memory state
2369 0 : let _tenant_lock = trace_exclusive_lock(
2370 0 : &self.tenant_op_locks,
2371 0 : tenant_shard_id.tenant_id,
2372 0 : TenantOperations::LocationConfig,
2373 0 : )
2374 0 : .await;
2375 :
2376 0 : if !tenant_shard_id.is_unsharded() {
2377 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
2378 0 : "This API is for importing single-sharded or unsharded tenants"
2379 0 : )));
2380 0 : }
2381 0 :
2382 0 : // First check if this is a creation or an update
2383 0 : let create_or_update = self.tenant_location_config_prepare(tenant_shard_id.tenant_id, req);
2384 0 :
2385 0 : let mut result = TenantLocationConfigResponse {
2386 0 : shards: Vec::new(),
2387 0 : stripe_size: None,
2388 0 : };
2389 0 : let waiters = match create_or_update {
2390 0 : TenantCreateOrUpdate::Create(create_req) => {
2391 0 : let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
2392 0 : result.shards = create_resp
2393 0 : .shards
2394 0 : .into_iter()
2395 0 : .map(|s| TenantShardLocation {
2396 0 : node_id: s.node_id,
2397 0 : shard_id: s.shard_id,
2398 0 : })
2399 0 : .collect();
2400 0 : waiters
2401 : }
2402 0 : TenantCreateOrUpdate::Update(updates) => {
2403 0 : // Persist updates
2404 0 : // Ordering: write to the database before applying changes in-memory, so that
2405 0 : // we will not appear time-travel backwards on a restart.
2406 0 : let mut schedule_context = ScheduleContext::default();
2407 : for ShardUpdate {
2408 0 : tenant_shard_id,
2409 0 : placement_policy,
2410 0 : tenant_config,
2411 0 : generation,
2412 0 : } in &updates
2413 : {
2414 0 : self.persistence
2415 0 : .update_tenant_shard(
2416 0 : TenantFilter::Shard(*tenant_shard_id),
2417 0 : Some(placement_policy.clone()),
2418 0 : Some(tenant_config.clone()),
2419 0 : *generation,
2420 0 : None,
2421 0 : )
2422 0 : .await?;
2423 : }
2424 :
2425 : // Apply updates in-memory
2426 0 : let mut waiters = Vec::new();
2427 0 : {
2428 0 : let mut locked = self.inner.write().unwrap();
2429 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2430 :
2431 : for ShardUpdate {
2432 0 : tenant_shard_id,
2433 0 : placement_policy,
2434 0 : tenant_config,
2435 0 : generation: update_generation,
2436 0 : } in updates
2437 : {
2438 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
2439 0 : tracing::warn!("Shard {tenant_shard_id} removed while updating");
2440 0 : continue;
2441 : };
2442 :
2443 : // Update stripe size
2444 0 : if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
2445 0 : result.stripe_size = Some(shard.shard.stripe_size);
2446 0 : }
2447 :
2448 0 : shard.policy = placement_policy;
2449 0 : shard.config = tenant_config;
2450 0 : if let Some(generation) = update_generation {
2451 0 : shard.generation = Some(generation);
2452 0 : }
2453 :
2454 0 : shard.schedule(scheduler, &mut schedule_context)?;
2455 :
2456 0 : let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
2457 0 : if let Some(waiter) = maybe_waiter {
2458 0 : waiters.push(waiter);
2459 0 : }
2460 :
2461 0 : if let Some(node_id) = shard.intent.get_attached() {
2462 0 : result.shards.push(TenantShardLocation {
2463 0 : shard_id: tenant_shard_id,
2464 0 : node_id: *node_id,
2465 0 : })
2466 0 : }
2467 : }
2468 : }
2469 0 : waiters
2470 : }
2471 : };
2472 :
2473 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2474 : // Do not treat a reconcile error as fatal: we have already applied any requested
2475 : // Intent changes, and the reconcile can fail for external reasons like unavailable
2476 : // compute notification API. In these cases, it is important that we do not
2477 : // cause the cloud control plane to retry forever on this API.
2478 0 : tracing::warn!(
2479 0 : "Failed to reconcile after /location_config: {e}, returning success anyway"
2480 : );
2481 0 : }
2482 :
2483 : // Logging the full result is useful because it lets us cross-check what the cloud control
2484 : // plane's tenant_shards table should contain.
2485 0 : tracing::info!("Complete, returning {result:?}");
2486 :
2487 0 : Ok(result)
2488 0 : }
2489 :
2490 0 : pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
2491 : // We require an exclusive lock, because we are updating persistent and in-memory state
2492 0 : let _tenant_lock = trace_exclusive_lock(
2493 0 : &self.tenant_op_locks,
2494 0 : req.tenant_id,
2495 0 : TenantOperations::ConfigSet,
2496 0 : )
2497 0 : .await;
2498 :
2499 0 : let tenant_id = req.tenant_id;
2500 0 : let config = req.config;
2501 0 :
2502 0 : self.persistence
2503 0 : .update_tenant_shard(
2504 0 : TenantFilter::Tenant(req.tenant_id),
2505 0 : None,
2506 0 : Some(config.clone()),
2507 0 : None,
2508 0 : None,
2509 0 : )
2510 0 : .await?;
2511 :
2512 0 : let waiters = {
2513 0 : let mut waiters = Vec::new();
2514 0 : let mut locked = self.inner.write().unwrap();
2515 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
2516 0 : for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2517 0 : shard.config = config.clone();
2518 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2519 0 : waiters.push(waiter);
2520 0 : }
2521 : }
2522 0 : waiters
2523 : };
2524 :
2525 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
2526 : // Treat this as success because we have stored the configuration. If e.g.
2527 : // a node was unavailable at this time, it should not stop us accepting a
2528 : // configuration change.
2529 0 : tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
2530 0 : }
2531 :
2532 0 : Ok(())
2533 0 : }
2534 :
2535 0 : pub(crate) fn tenant_config_get(
2536 0 : &self,
2537 0 : tenant_id: TenantId,
2538 0 : ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
2539 0 : let config = {
2540 0 : let locked = self.inner.read().unwrap();
2541 0 :
2542 0 : match locked
2543 0 : .tenants
2544 0 : .range(TenantShardId::tenant_range(tenant_id))
2545 0 : .next()
2546 : {
2547 0 : Some((_tenant_shard_id, shard)) => shard.config.clone(),
2548 : None => {
2549 0 : return Err(ApiError::NotFound(
2550 0 : anyhow::anyhow!("Tenant not found").into(),
2551 0 : ))
2552 : }
2553 : }
2554 : };
2555 :
2556 : // Unlike the pageserver, we do not have a set of global defaults: the config is
2557 : // entirely per-tenant. Therefore the distinction between `tenant_specific_overrides`
2558 : // and `effective_config` in the response is meaningless, but we retain that syntax
2559 : // in order to remain compatible with the pageserver API.
2560 :
2561 0 : let response = HashMap::from([
2562 : (
2563 : "tenant_specific_overrides",
2564 0 : serde_json::to_value(&config)
2565 0 : .context("serializing tenant specific overrides")
2566 0 : .map_err(ApiError::InternalServerError)?,
2567 : ),
2568 : (
2569 0 : "effective_config",
2570 0 : serde_json::to_value(&config)
2571 0 : .context("serializing effective config")
2572 0 : .map_err(ApiError::InternalServerError)?,
2573 : ),
2574 : ]);
2575 :
2576 0 : Ok(response)
2577 0 : }
2578 :
2579 0 : pub(crate) async fn tenant_time_travel_remote_storage(
2580 0 : &self,
2581 0 : time_travel_req: &TenantTimeTravelRequest,
2582 0 : tenant_id: TenantId,
2583 0 : timestamp: Cow<'_, str>,
2584 0 : done_if_after: Cow<'_, str>,
2585 0 : ) -> Result<(), ApiError> {
2586 0 : let _tenant_lock = trace_exclusive_lock(
2587 0 : &self.tenant_op_locks,
2588 0 : tenant_id,
2589 0 : TenantOperations::TimeTravelRemoteStorage,
2590 0 : )
2591 0 : .await;
2592 :
2593 0 : let node = {
2594 0 : let mut locked = self.inner.write().unwrap();
2595 : // Just a sanity check to prevent misuse: the API expects that the tenant is fully
2596 : // detached everywhere, and nothing writes to S3 storage. Here, we verify that,
2597 : // but only at the start of the process, so it's really just to prevent operator
2598 : // mistakes.
2599 0 : for (shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
2600 0 : if shard.intent.get_attached().is_some() || !shard.intent.get_secondary().is_empty()
2601 : {
2602 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2603 0 : "We want tenant to be attached in shard with tenant_shard_id={shard_id}"
2604 0 : )));
2605 0 : }
2606 0 : let maybe_attached = shard
2607 0 : .observed
2608 0 : .locations
2609 0 : .iter()
2610 0 : .filter_map(|(node_id, observed_location)| {
2611 0 : observed_location
2612 0 : .conf
2613 0 : .as_ref()
2614 0 : .map(|loc| (node_id, observed_location, loc.mode))
2615 0 : })
2616 0 : .find(|(_, _, mode)| *mode != LocationConfigMode::Detached);
2617 0 : if let Some((node_id, _observed_location, mode)) = maybe_attached {
2618 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("We observed attached={mode:?} tenant in node_id={node_id} shard with tenant_shard_id={shard_id}")));
2619 0 : }
2620 : }
2621 0 : let scheduler = &mut locked.scheduler;
2622 : // Right now we only perform the operation on a single node without parallelization
2623 : // TODO fan out the operation to multiple nodes for better performance
2624 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2625 0 : let node = locked
2626 0 : .nodes
2627 0 : .get(&node_id)
2628 0 : .expect("Pageservers may not be deleted while lock is active");
2629 0 : node.clone()
2630 0 : };
2631 0 :
2632 0 : // The shard count is encoded in the remote storage's URL, so we need to handle all historically used shard counts
2633 0 : let mut counts = time_travel_req
2634 0 : .shard_counts
2635 0 : .iter()
2636 0 : .copied()
2637 0 : .collect::<HashSet<_>>()
2638 0 : .into_iter()
2639 0 : .collect::<Vec<_>>();
2640 0 : counts.sort_unstable();
2641 :
2642 0 : for count in counts {
2643 0 : let shard_ids = (0..count.count())
2644 0 : .map(|i| TenantShardId {
2645 0 : tenant_id,
2646 0 : shard_number: ShardNumber(i),
2647 0 : shard_count: count,
2648 0 : })
2649 0 : .collect::<Vec<_>>();
2650 0 : for tenant_shard_id in shard_ids {
2651 0 : let client = PageserverClient::new(
2652 0 : node.get_id(),
2653 0 : node.base_url(),
2654 0 : self.config.jwt_token.as_deref(),
2655 0 : );
2656 0 :
2657 0 : tracing::info!("Doing time travel recovery for shard {tenant_shard_id}",);
2658 :
2659 0 : client
2660 0 : .tenant_time_travel_remote_storage(
2661 0 : tenant_shard_id,
2662 0 : ×tamp,
2663 0 : &done_if_after,
2664 0 : )
2665 0 : .await
2666 0 : .map_err(|e| {
2667 0 : ApiError::InternalServerError(anyhow::anyhow!(
2668 0 : "Error doing time travel recovery for shard {tenant_shard_id} on node {}: {e}",
2669 0 : node
2670 0 : ))
2671 0 : })?;
2672 : }
2673 : }
2674 0 : Ok(())
2675 0 : }
2676 :
2677 0 : pub(crate) async fn tenant_secondary_download(
2678 0 : &self,
2679 0 : tenant_id: TenantId,
2680 0 : wait: Option<Duration>,
2681 0 : ) -> Result<(StatusCode, SecondaryProgress), ApiError> {
2682 0 : let _tenant_lock = trace_shared_lock(
2683 0 : &self.tenant_op_locks,
2684 0 : tenant_id,
2685 0 : TenantOperations::SecondaryDownload,
2686 0 : )
2687 0 : .await;
2688 :
2689 : // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
2690 0 : let targets = {
2691 0 : let locked = self.inner.read().unwrap();
2692 0 : let mut targets = Vec::new();
2693 :
2694 0 : for (tenant_shard_id, shard) in
2695 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
2696 : {
2697 0 : for node_id in shard.intent.get_secondary() {
2698 0 : let node = locked
2699 0 : .nodes
2700 0 : .get(node_id)
2701 0 : .expect("Pageservers may not be deleted while referenced");
2702 0 :
2703 0 : targets.push((*tenant_shard_id, node.clone()));
2704 0 : }
2705 : }
2706 0 : targets
2707 0 : };
2708 0 :
2709 0 : // Issue concurrent requests to all shards' locations
2710 0 : let mut futs = FuturesUnordered::new();
2711 0 : for (tenant_shard_id, node) in targets {
2712 0 : let client = PageserverClient::new(
2713 0 : node.get_id(),
2714 0 : node.base_url(),
2715 0 : self.config.jwt_token.as_deref(),
2716 0 : );
2717 0 : futs.push(async move {
2718 0 : let result = client
2719 0 : .tenant_secondary_download(tenant_shard_id, wait)
2720 0 : .await;
2721 0 : (result, node, tenant_shard_id)
2722 0 : })
2723 : }
2724 :
2725 : // Handle any errors returned by pageservers. This includes cases like this request racing with
2726 : // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
2727 : // well as more general cases like 503s, 500s, or timeouts.
2728 0 : let mut aggregate_progress = SecondaryProgress::default();
2729 0 : let mut aggregate_status: Option<StatusCode> = None;
2730 0 : let mut error: Option<mgmt_api::Error> = None;
2731 0 : while let Some((result, node, tenant_shard_id)) = futs.next().await {
2732 0 : match result {
2733 0 : Err(e) => {
2734 0 : // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
2735 0 : // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
2736 0 : // than they had hoped for.
2737 0 : tracing::warn!("Secondary download error from pageserver {node}: {e}",);
2738 0 : error = Some(e)
2739 : }
2740 0 : Ok((status_code, progress)) => {
2741 0 : tracing::info!(%tenant_shard_id, "Shard status={status_code} progress: {progress:?}");
2742 0 : aggregate_progress.layers_downloaded += progress.layers_downloaded;
2743 0 : aggregate_progress.layers_total += progress.layers_total;
2744 0 : aggregate_progress.bytes_downloaded += progress.bytes_downloaded;
2745 0 : aggregate_progress.bytes_total += progress.bytes_total;
2746 0 : aggregate_progress.heatmap_mtime =
2747 0 : std::cmp::max(aggregate_progress.heatmap_mtime, progress.heatmap_mtime);
2748 0 : aggregate_status = match aggregate_status {
2749 0 : None => Some(status_code),
2750 0 : Some(StatusCode::OK) => Some(status_code),
2751 0 : Some(cur) => {
2752 0 : // Other status codes (e.g. 202) -- do not overwrite.
2753 0 : Some(cur)
2754 : }
2755 : };
2756 : }
2757 : }
2758 : }
2759 :
2760 : // If any of the shards return 202, indicate our result as 202.
2761 0 : match aggregate_status {
2762 : None => {
2763 0 : match error {
2764 0 : Some(e) => {
2765 0 : // No successes, and an error: surface it
2766 0 : Err(ApiError::Conflict(format!("Error from pageserver: {e}")))
2767 : }
2768 : None => {
2769 : // No shards found
2770 0 : Err(ApiError::NotFound(
2771 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
2772 0 : ))
2773 : }
2774 : }
2775 : }
2776 0 : Some(aggregate_status) => Ok((aggregate_status, aggregate_progress)),
2777 : }
2778 0 : }
2779 :
2780 0 : pub(crate) async fn tenant_delete(&self, tenant_id: TenantId) -> Result<StatusCode, ApiError> {
2781 0 : let _tenant_lock =
2782 0 : trace_exclusive_lock(&self.tenant_op_locks, tenant_id, TenantOperations::Delete).await;
2783 :
2784 : // Detach all shards
2785 0 : let (detach_waiters, shard_ids, node) = {
2786 0 : let mut shard_ids = Vec::new();
2787 0 : let mut detach_waiters = Vec::new();
2788 0 : let mut locked = self.inner.write().unwrap();
2789 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2790 0 : for (tenant_shard_id, shard) in
2791 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2792 : {
2793 0 : shard_ids.push(*tenant_shard_id);
2794 0 :
2795 0 : // Update the tenant's intent to remove all attachments
2796 0 : shard.policy = PlacementPolicy::Detached;
2797 0 : shard
2798 0 : .schedule(scheduler, &mut ScheduleContext::default())
2799 0 : .expect("De-scheduling is infallible");
2800 0 : debug_assert!(shard.intent.get_attached().is_none());
2801 0 : debug_assert!(shard.intent.get_secondary().is_empty());
2802 :
2803 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
2804 0 : detach_waiters.push(waiter);
2805 0 : }
2806 : }
2807 :
2808 : // Pick an arbitrary node to use for remote deletions (does not have to be where the tenant
2809 : // was attached, just has to be able to see the S3 content)
2810 0 : let node_id = scheduler.schedule_shard(&[], &ScheduleContext::default())?;
2811 0 : let node = nodes
2812 0 : .get(&node_id)
2813 0 : .expect("Pageservers may not be deleted while lock is active");
2814 0 : (detach_waiters, shard_ids, node.clone())
2815 0 : };
2816 0 :
2817 0 : // This reconcile wait can fail in a few ways:
2818 0 : // A there is a very long queue for the reconciler semaphore
2819 0 : // B some pageserver is failing to handle a detach promptly
2820 0 : // C some pageserver goes offline right at the moment we send it a request.
2821 0 : //
2822 0 : // A and C are transient: the semaphore will eventually become available, and once a node is marked offline
2823 0 : // the next attempt to reconcile will silently skip detaches for an offline node and succeed. If B happens,
2824 0 : // it's a bug, and needs resolving at the pageserver level (we shouldn't just leave attachments behind while
2825 0 : // deleting the underlying data).
2826 0 : self.await_waiters(detach_waiters, RECONCILE_TIMEOUT)
2827 0 : .await?;
2828 :
2829 0 : let locations = shard_ids
2830 0 : .into_iter()
2831 0 : .map(|s| (s, node.clone()))
2832 0 : .collect::<Vec<_>>();
2833 0 : let results = self.tenant_for_shards_api(
2834 0 : locations,
2835 0 : |tenant_shard_id, client| async move { client.tenant_delete(tenant_shard_id).await },
2836 0 : 1,
2837 0 : 3,
2838 0 : RECONCILE_TIMEOUT,
2839 0 : &self.cancel,
2840 0 : )
2841 0 : .await;
2842 0 : for result in results {
2843 0 : match result {
2844 : Ok(StatusCode::ACCEPTED) => {
2845 : // This should never happen: we waited for detaches to finish above
2846 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
2847 0 : "Unexpectedly still attached on {}",
2848 0 : node
2849 0 : )));
2850 : }
2851 0 : Ok(_) => {}
2852 : Err(mgmt_api::Error::Cancelled) => {
2853 0 : return Err(ApiError::ShuttingDown);
2854 : }
2855 0 : Err(e) => {
2856 0 : // This is unexpected: remote deletion should be infallible, unless the object store
2857 0 : // at large is unavailable.
2858 0 : tracing::error!("Error deleting via node {}: {e}", node);
2859 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(e)));
2860 : }
2861 : }
2862 : }
2863 :
2864 : // Fall through: deletion of the tenant on pageservers is complete, we may proceed to drop
2865 : // our in-memory state and database state.
2866 :
2867 : // Ordering: we delete persistent state first: if we then
2868 : // crash, we will drop the in-memory state.
2869 :
2870 : // Drop persistent state.
2871 0 : self.persistence.delete_tenant(tenant_id).await?;
2872 :
2873 : // Drop in-memory state
2874 : {
2875 0 : let mut locked = self.inner.write().unwrap();
2876 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
2877 :
2878 : // Dereference Scheduler from shards before dropping them
2879 0 : for (_tenant_shard_id, shard) in
2880 0 : tenants.range_mut(TenantShardId::tenant_range(tenant_id))
2881 0 : {
2882 0 : shard.intent.clear(scheduler);
2883 0 : }
2884 :
2885 0 : tenants.retain(|tenant_shard_id, _shard| tenant_shard_id.tenant_id != tenant_id);
2886 0 : tracing::info!(
2887 0 : "Deleted tenant {tenant_id}, now have {} tenants",
2888 0 : locked.tenants.len()
2889 : );
2890 : };
2891 :
2892 : // Success is represented as 404, to imitate the existing pageserver deletion API
2893 0 : Ok(StatusCode::NOT_FOUND)
2894 0 : }
2895 :
2896 : /// Naming: this configures the storage controller's policies for a tenant, whereas [`Self::tenant_config_set`] is "set the TenantConfig"
2897 : /// for a tenant. The TenantConfig is passed through to pageservers, whereas this function modifies
2898 : /// the tenant's policies (configuration) within the storage controller
2899 0 : pub(crate) async fn tenant_update_policy(
2900 0 : &self,
2901 0 : tenant_id: TenantId,
2902 0 : req: TenantPolicyRequest,
2903 0 : ) -> Result<(), ApiError> {
2904 : // We require an exclusive lock, because we are updating persistent and in-memory state
2905 0 : let _tenant_lock = trace_exclusive_lock(
2906 0 : &self.tenant_op_locks,
2907 0 : tenant_id,
2908 0 : TenantOperations::UpdatePolicy,
2909 0 : )
2910 0 : .await;
2911 :
2912 0 : failpoint_support::sleep_millis_async!("tenant-update-policy-exclusive-lock");
2913 :
2914 : let TenantPolicyRequest {
2915 0 : placement,
2916 0 : scheduling,
2917 0 : } = req;
2918 0 :
2919 0 : self.persistence
2920 0 : .update_tenant_shard(
2921 0 : TenantFilter::Tenant(tenant_id),
2922 0 : placement.clone(),
2923 0 : None,
2924 0 : None,
2925 0 : scheduling,
2926 0 : )
2927 0 : .await?;
2928 :
2929 0 : let mut schedule_context = ScheduleContext::default();
2930 0 : let mut locked = self.inner.write().unwrap();
2931 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
2932 0 : for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
2933 0 : if let Some(placement) = &placement {
2934 0 : shard.policy = placement.clone();
2935 0 :
2936 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2937 0 : "Updated placement policy to {placement:?}");
2938 0 : }
2939 :
2940 0 : if let Some(scheduling) = &scheduling {
2941 0 : shard.set_scheduling_policy(*scheduling);
2942 0 :
2943 0 : tracing::info!(tenant_id=%shard_id.tenant_id, shard_id=%shard_id.shard_slug(),
2944 0 : "Updated scheduling policy to {scheduling:?}");
2945 0 : }
2946 :
2947 : // In case scheduling is being switched back on, try it now.
2948 0 : shard.schedule(scheduler, &mut schedule_context).ok();
2949 0 : self.maybe_reconcile_shard(shard, nodes);
2950 : }
2951 :
2952 0 : Ok(())
2953 0 : }
2954 :
2955 0 : pub(crate) async fn tenant_timeline_create(
2956 0 : &self,
2957 0 : tenant_id: TenantId,
2958 0 : mut create_req: TimelineCreateRequest,
2959 0 : ) -> Result<TimelineInfo, ApiError> {
2960 0 : tracing::info!(
2961 0 : "Creating timeline {}/{}",
2962 : tenant_id,
2963 : create_req.new_timeline_id,
2964 : );
2965 :
2966 0 : let _tenant_lock = trace_shared_lock(
2967 0 : &self.tenant_op_locks,
2968 0 : tenant_id,
2969 0 : TenantOperations::TimelineCreate,
2970 0 : )
2971 0 : .await;
2972 0 : failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
2973 :
2974 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
2975 0 : if targets.is_empty() {
2976 0 : return Err(ApiError::NotFound(
2977 0 : anyhow::anyhow!("Tenant not found").into(),
2978 0 : ));
2979 0 : };
2980 0 : let shard_zero = targets.remove(0);
2981 :
2982 0 : async fn create_one(
2983 0 : tenant_shard_id: TenantShardId,
2984 0 : node: Node,
2985 0 : jwt: Option<String>,
2986 0 : create_req: TimelineCreateRequest,
2987 0 : ) -> Result<TimelineInfo, ApiError> {
2988 0 : tracing::info!(
2989 0 : "Creating timeline on shard {}/{}, attached to node {node}",
2990 : tenant_shard_id,
2991 : create_req.new_timeline_id,
2992 : );
2993 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
2994 0 :
2995 0 : client
2996 0 : .timeline_create(tenant_shard_id, &create_req)
2997 0 : .await
2998 0 : .map_err(|e| passthrough_api_error(&node, e))
2999 0 : }
3000 :
3001 : // Because the caller might not provide an explicit LSN, we must do the creation first on a single shard, and then
3002 : // use whatever LSN that shard picked when creating on subsequent shards. We arbitrarily use shard zero as the shard
3003 : // that will get the first creation request, and propagate the LSN to all the >0 shards.
3004 0 : let timeline_info = create_one(
3005 0 : shard_zero.0,
3006 0 : shard_zero.1,
3007 0 : self.config.jwt_token.clone(),
3008 0 : create_req.clone(),
3009 0 : )
3010 0 : .await?;
3011 :
3012 : // Propagate the LSN that shard zero picked, if caller didn't provide one
3013 0 : if create_req.ancestor_timeline_id.is_some() && create_req.ancestor_start_lsn.is_none()
3014 0 : {
3015 0 : create_req.ancestor_start_lsn = timeline_info.ancestor_lsn;
3016 0 : }
3017 :
3018 : // Create timeline on remaining shards with number >0
3019 0 : if !targets.is_empty() {
3020 : // If we had multiple shards, issue requests for the remainder now.
3021 0 : let jwt = &self.config.jwt_token;
3022 0 : self.tenant_for_shards(
3023 0 : targets.iter().map(|t| (t.0, t.1.clone())).collect(),
3024 0 : |tenant_shard_id: TenantShardId, node: Node| {
3025 0 : let create_req = create_req.clone();
3026 0 : Box::pin(create_one(tenant_shard_id, node, jwt.clone(), create_req))
3027 0 : },
3028 0 : )
3029 0 : .await?;
3030 0 : }
3031 :
3032 0 : Ok(timeline_info)
3033 0 : })
3034 0 : .await?
3035 0 : }
3036 :
3037 0 : pub(crate) async fn tenant_timeline_archival_config(
3038 0 : &self,
3039 0 : tenant_id: TenantId,
3040 0 : timeline_id: TimelineId,
3041 0 : req: TimelineArchivalConfigRequest,
3042 0 : ) -> Result<(), ApiError> {
3043 0 : tracing::info!(
3044 0 : "Setting archival config of timeline {tenant_id}/{timeline_id} to '{:?}'",
3045 : req.state
3046 : );
3047 :
3048 0 : let _tenant_lock = trace_shared_lock(
3049 0 : &self.tenant_op_locks,
3050 0 : tenant_id,
3051 0 : TenantOperations::TimelineArchivalConfig,
3052 0 : )
3053 0 : .await;
3054 :
3055 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3056 0 : if targets.is_empty() {
3057 0 : return Err(ApiError::NotFound(
3058 0 : anyhow::anyhow!("Tenant not found").into(),
3059 0 : ));
3060 0 : }
3061 0 : async fn config_one(
3062 0 : tenant_shard_id: TenantShardId,
3063 0 : timeline_id: TimelineId,
3064 0 : node: Node,
3065 0 : jwt: Option<String>,
3066 0 : req: TimelineArchivalConfigRequest,
3067 0 : ) -> Result<(), ApiError> {
3068 0 : tracing::info!(
3069 0 : "Setting archival config of timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3070 : );
3071 :
3072 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3073 0 :
3074 0 : client
3075 0 : .timeline_archival_config(tenant_shard_id, timeline_id, &req)
3076 0 : .await
3077 0 : .map_err(|e| match e {
3078 0 : mgmt_api::Error::ApiError(StatusCode::PRECONDITION_FAILED, msg) => {
3079 0 : ApiError::PreconditionFailed(msg.into_boxed_str())
3080 : }
3081 0 : _ => passthrough_api_error(&node, e),
3082 0 : })
3083 0 : }
3084 :
3085 : // no shard needs to go first/last; the operation should be idempotent
3086 : // TODO: it would be great to ensure that all shards return the same error
3087 0 : let results = self
3088 0 : .tenant_for_shards(targets, |tenant_shard_id, node| {
3089 0 : futures::FutureExt::boxed(config_one(
3090 0 : tenant_shard_id,
3091 0 : timeline_id,
3092 0 : node,
3093 0 : self.config.jwt_token.clone(),
3094 0 : req.clone(),
3095 0 : ))
3096 0 : })
3097 0 : .await?;
3098 0 : assert!(!results.is_empty(), "must have at least one result");
3099 :
3100 0 : Ok(())
3101 0 : }).await?
3102 0 : }
3103 :
3104 0 : pub(crate) async fn tenant_timeline_detach_ancestor(
3105 0 : &self,
3106 0 : tenant_id: TenantId,
3107 0 : timeline_id: TimelineId,
3108 0 : ) -> Result<models::detach_ancestor::AncestorDetached, ApiError> {
3109 0 : tracing::info!("Detaching timeline {tenant_id}/{timeline_id}",);
3110 :
3111 0 : let _tenant_lock = trace_shared_lock(
3112 0 : &self.tenant_op_locks,
3113 0 : tenant_id,
3114 0 : TenantOperations::TimelineDetachAncestor,
3115 0 : )
3116 0 : .await;
3117 :
3118 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3119 0 : if targets.is_empty() {
3120 0 : return Err(ApiError::NotFound(
3121 0 : anyhow::anyhow!("Tenant not found").into(),
3122 0 : ));
3123 0 : }
3124 :
3125 0 : async fn detach_one(
3126 0 : tenant_shard_id: TenantShardId,
3127 0 : timeline_id: TimelineId,
3128 0 : node: Node,
3129 0 : jwt: Option<String>,
3130 0 : ) -> Result<(ShardNumber, models::detach_ancestor::AncestorDetached), ApiError> {
3131 0 : tracing::info!(
3132 0 : "Detaching timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3133 : );
3134 :
3135 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3136 0 :
3137 0 : client
3138 0 : .timeline_detach_ancestor(tenant_shard_id, timeline_id)
3139 0 : .await
3140 0 : .map_err(|e| {
3141 : use mgmt_api::Error;
3142 :
3143 0 : match e {
3144 : // no ancestor (ever)
3145 0 : Error::ApiError(StatusCode::CONFLICT, msg) => ApiError::Conflict(format!(
3146 0 : "{node}: {}",
3147 0 : msg.strip_prefix("Conflict: ").unwrap_or(&msg)
3148 0 : )),
3149 : // too many ancestors
3150 0 : Error::ApiError(StatusCode::BAD_REQUEST, msg) => {
3151 0 : ApiError::BadRequest(anyhow::anyhow!("{node}: {msg}"))
3152 : }
3153 0 : Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, msg) => {
3154 0 : // avoid turning these into conflicts to remain compatible with
3155 0 : // pageservers, 500 errors are sadly retryable with timeline ancestor
3156 0 : // detach
3157 0 : ApiError::InternalServerError(anyhow::anyhow!("{node}: {msg}"))
3158 : }
3159 : // rest can be mapped as usual
3160 0 : other => passthrough_api_error(&node, other),
3161 : }
3162 0 : })
3163 0 : .map(|res| (tenant_shard_id.shard_number, res))
3164 0 : }
3165 :
3166 : // no shard needs to go first/last; the operation should be idempotent
3167 0 : let mut results = self
3168 0 : .tenant_for_shards(targets, |tenant_shard_id, node| {
3169 0 : futures::FutureExt::boxed(detach_one(
3170 0 : tenant_shard_id,
3171 0 : timeline_id,
3172 0 : node,
3173 0 : self.config.jwt_token.clone(),
3174 0 : ))
3175 0 : })
3176 0 : .await?;
3177 :
3178 0 : let any = results.pop().expect("we must have at least one response");
3179 0 :
3180 0 : let mismatching = results
3181 0 : .iter()
3182 0 : .filter(|(_, res)| res != &any.1)
3183 0 : .collect::<Vec<_>>();
3184 0 : if !mismatching.is_empty() {
3185 : // this can be hit by races which should not happen because operation lock on cplane
3186 0 : let matching = results.len() - mismatching.len();
3187 0 : tracing::error!(
3188 : matching,
3189 : compared_against=?any,
3190 : ?mismatching,
3191 0 : "shards returned different results"
3192 : );
3193 :
3194 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!("pageservers returned mixed results for ancestor detach; manual intervention is required.")));
3195 0 : }
3196 0 :
3197 0 : Ok(any.1)
3198 0 : }).await?
3199 0 : }
3200 :
3201 0 : pub(crate) async fn tenant_timeline_block_unblock_gc(
3202 0 : &self,
3203 0 : tenant_id: TenantId,
3204 0 : timeline_id: TimelineId,
3205 0 : dir: BlockUnblock,
3206 0 : ) -> Result<(), ApiError> {
3207 0 : let _tenant_lock = trace_shared_lock(
3208 0 : &self.tenant_op_locks,
3209 0 : tenant_id,
3210 0 : TenantOperations::TimelineGcBlockUnblock,
3211 0 : )
3212 0 : .await;
3213 :
3214 0 : self.tenant_remote_mutation(tenant_id, move |targets| async move {
3215 0 : if targets.is_empty() {
3216 0 : return Err(ApiError::NotFound(
3217 0 : anyhow::anyhow!("Tenant not found").into(),
3218 0 : ));
3219 0 : }
3220 :
3221 0 : async fn do_one(
3222 0 : tenant_shard_id: TenantShardId,
3223 0 : timeline_id: TimelineId,
3224 0 : node: Node,
3225 0 : jwt: Option<String>,
3226 0 : dir: BlockUnblock,
3227 0 : ) -> Result<(), ApiError> {
3228 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3229 0 :
3230 0 : client
3231 0 : .timeline_block_unblock_gc(tenant_shard_id, timeline_id, dir)
3232 0 : .await
3233 0 : .map_err(|e| passthrough_api_error(&node, e))
3234 0 : }
3235 :
3236 : // no shard needs to go first/last; the operation should be idempotent
3237 0 : self.tenant_for_shards(targets, |tenant_shard_id, node| {
3238 0 : futures::FutureExt::boxed(do_one(
3239 0 : tenant_shard_id,
3240 0 : timeline_id,
3241 0 : node,
3242 0 : self.config.jwt_token.clone(),
3243 0 : dir,
3244 0 : ))
3245 0 : })
3246 0 : .await
3247 0 : })
3248 0 : .await??;
3249 0 : Ok(())
3250 0 : }
3251 :
3252 : /// Helper for concurrently calling a pageserver API on a number of shards, such as timeline creation.
3253 : ///
3254 : /// On success, the returned vector contains exactly the same number of elements as the input `locations`.
3255 0 : async fn tenant_for_shards<F, R>(
3256 0 : &self,
3257 0 : locations: Vec<(TenantShardId, Node)>,
3258 0 : mut req_fn: F,
3259 0 : ) -> Result<Vec<R>, ApiError>
3260 0 : where
3261 0 : F: FnMut(
3262 0 : TenantShardId,
3263 0 : Node,
3264 0 : )
3265 0 : -> std::pin::Pin<Box<dyn futures::Future<Output = Result<R, ApiError>> + Send>>,
3266 0 : {
3267 0 : let mut futs = FuturesUnordered::new();
3268 0 : let mut results = Vec::with_capacity(locations.len());
3269 :
3270 0 : for (tenant_shard_id, node) in locations {
3271 0 : futs.push(req_fn(tenant_shard_id, node));
3272 0 : }
3273 :
3274 0 : while let Some(r) = futs.next().await {
3275 0 : results.push(r?);
3276 : }
3277 :
3278 0 : Ok(results)
3279 0 : }
3280 :
3281 : /// Concurrently invoke a pageserver API call on many shards at once
3282 0 : pub(crate) async fn tenant_for_shards_api<T, O, F>(
3283 0 : &self,
3284 0 : locations: Vec<(TenantShardId, Node)>,
3285 0 : op: O,
3286 0 : warn_threshold: u32,
3287 0 : max_retries: u32,
3288 0 : timeout: Duration,
3289 0 : cancel: &CancellationToken,
3290 0 : ) -> Vec<mgmt_api::Result<T>>
3291 0 : where
3292 0 : O: Fn(TenantShardId, PageserverClient) -> F + Copy,
3293 0 : F: std::future::Future<Output = mgmt_api::Result<T>>,
3294 0 : {
3295 0 : let mut futs = FuturesUnordered::new();
3296 0 : let mut results = Vec::with_capacity(locations.len());
3297 :
3298 0 : for (tenant_shard_id, node) in locations {
3299 0 : futs.push(async move {
3300 0 : node.with_client_retries(
3301 0 : |client| op(tenant_shard_id, client),
3302 0 : &self.config.jwt_token,
3303 0 : warn_threshold,
3304 0 : max_retries,
3305 0 : timeout,
3306 0 : cancel,
3307 0 : )
3308 0 : .await
3309 0 : });
3310 0 : }
3311 :
3312 0 : while let Some(r) = futs.next().await {
3313 0 : let r = r.unwrap_or(Err(mgmt_api::Error::Cancelled));
3314 0 : results.push(r);
3315 0 : }
3316 :
3317 0 : results
3318 0 : }
3319 :
3320 : /// Helper for safely working with the shards in a tenant remotely on pageservers, for example
3321 : /// when creating and deleting timelines:
3322 : /// - Makes sure shards are attached somewhere if they weren't already
3323 : /// - Looks up the shards and the nodes where they were most recently attached
3324 : /// - Guarantees that after the inner function returns, the shards' generations haven't moved on: this
3325 : /// ensures that the remote operation acted on the most recent generation, and is therefore durable.
3326 0 : async fn tenant_remote_mutation<R, O, F>(
3327 0 : &self,
3328 0 : tenant_id: TenantId,
3329 0 : op: O,
3330 0 : ) -> Result<R, ApiError>
3331 0 : where
3332 0 : O: FnOnce(Vec<(TenantShardId, Node)>) -> F,
3333 0 : F: std::future::Future<Output = R>,
3334 0 : {
3335 0 : let target_gens = {
3336 0 : let mut targets = Vec::new();
3337 :
3338 : // Load the currently attached pageservers for the latest generation of each shard. This can
3339 : // run concurrently with reconciliations, and it is not guaranteed that the node we find here
3340 : // will still be the latest when we're done: we will check generations again at the end of
3341 : // this function to handle that.
3342 0 : let generations = self.persistence.tenant_generations(tenant_id).await?;
3343 :
3344 0 : if generations
3345 0 : .iter()
3346 0 : .any(|i| i.generation.is_none() || i.generation_pageserver.is_none())
3347 : {
3348 : // One or more shards has not been attached to a pageserver. Check if this is because it's configured
3349 : // to be detached (409: caller should give up), or because it's meant to be attached but isn't yet (503: caller should retry)
3350 0 : let locked = self.inner.read().unwrap();
3351 0 : for (shard_id, shard) in
3352 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3353 : {
3354 0 : match shard.policy {
3355 0 : PlacementPolicy::Attached(_) => {
3356 0 : // This shard is meant to be attached: the caller is not wrong to try and
3357 0 : // use this function, but we can't service the request right now.
3358 0 : }
3359 : PlacementPolicy::Secondary | PlacementPolicy::Detached => {
3360 0 : return Err(ApiError::Conflict(format!(
3361 0 : "Shard {shard_id} tenant has policy {:?}",
3362 0 : shard.policy
3363 0 : )));
3364 : }
3365 : }
3366 : }
3367 :
3368 0 : return Err(ApiError::ResourceUnavailable(
3369 0 : "One or more shards in tenant is not yet attached".into(),
3370 0 : ));
3371 0 : }
3372 0 :
3373 0 : let locked = self.inner.read().unwrap();
3374 : for ShardGenerationState {
3375 0 : tenant_shard_id,
3376 0 : generation,
3377 0 : generation_pageserver,
3378 0 : } in generations
3379 : {
3380 0 : let node_id = generation_pageserver.expect("We checked for None above");
3381 0 : let node = locked
3382 0 : .nodes
3383 0 : .get(&node_id)
3384 0 : .ok_or(ApiError::Conflict(format!(
3385 0 : "Raced with removal of node {node_id}"
3386 0 : )))?;
3387 0 : targets.push((tenant_shard_id, node.clone(), generation));
3388 : }
3389 :
3390 0 : targets
3391 0 : };
3392 0 :
3393 0 : let targets = target_gens.iter().map(|t| (t.0, t.1.clone())).collect();
3394 0 : let result = op(targets).await;
3395 :
3396 : // Post-check: are all the generations of all the shards the same as they were initially? This proves that
3397 : // our remote operation executed on the latest generation and is therefore persistent.
3398 : {
3399 0 : let latest_generations = self.persistence.tenant_generations(tenant_id).await?;
3400 0 : if latest_generations
3401 0 : .into_iter()
3402 0 : .map(
3403 0 : |ShardGenerationState {
3404 : tenant_shard_id,
3405 : generation,
3406 : generation_pageserver: _,
3407 0 : }| (tenant_shard_id, generation),
3408 0 : )
3409 0 : .collect::<Vec<_>>()
3410 0 : != target_gens
3411 0 : .into_iter()
3412 0 : .map(|i| (i.0, i.2))
3413 0 : .collect::<Vec<_>>()
3414 : {
3415 : // We raced with something that incremented the generation, and therefore cannot be
3416 : // confident that our actions are persistent (they might have hit an old generation).
3417 : //
3418 : // This is safe but requires a retry: ask the client to do that by giving them a 503 response.
3419 0 : return Err(ApiError::ResourceUnavailable(
3420 0 : "Tenant attachment changed, please retry".into(),
3421 0 : ));
3422 0 : }
3423 0 : }
3424 0 :
3425 0 : Ok(result)
3426 0 : }
3427 :
3428 0 : pub(crate) async fn tenant_timeline_delete(
3429 0 : &self,
3430 0 : tenant_id: TenantId,
3431 0 : timeline_id: TimelineId,
3432 0 : ) -> Result<StatusCode, ApiError> {
3433 0 : tracing::info!("Deleting timeline {}/{}", tenant_id, timeline_id,);
3434 0 : let _tenant_lock = trace_shared_lock(
3435 0 : &self.tenant_op_locks,
3436 0 : tenant_id,
3437 0 : TenantOperations::TimelineDelete,
3438 0 : )
3439 0 : .await;
3440 :
3441 0 : self.tenant_remote_mutation(tenant_id, move |mut targets| async move {
3442 0 : if targets.is_empty() {
3443 0 : return Err(ApiError::NotFound(
3444 0 : anyhow::anyhow!("Tenant not found").into(),
3445 0 : ));
3446 0 : }
3447 0 : let shard_zero = targets.remove(0);
3448 :
3449 0 : async fn delete_one(
3450 0 : tenant_shard_id: TenantShardId,
3451 0 : timeline_id: TimelineId,
3452 0 : node: Node,
3453 0 : jwt: Option<String>,
3454 0 : ) -> Result<StatusCode, ApiError> {
3455 0 : tracing::info!(
3456 0 : "Deleting timeline on shard {tenant_shard_id}/{timeline_id}, attached to node {node}",
3457 : );
3458 :
3459 0 : let client = PageserverClient::new(node.get_id(), node.base_url(), jwt.as_deref());
3460 0 : client
3461 0 : .timeline_delete(tenant_shard_id, timeline_id)
3462 0 : .await
3463 0 : .map_err(|e| {
3464 0 : ApiError::InternalServerError(anyhow::anyhow!(
3465 0 : "Error deleting timeline {timeline_id} on {tenant_shard_id} on node {node}: {e}",
3466 0 : ))
3467 0 : })
3468 0 : }
3469 :
3470 0 : let statuses = self
3471 0 : .tenant_for_shards(targets, |tenant_shard_id: TenantShardId, node: Node| {
3472 0 : Box::pin(delete_one(
3473 0 : tenant_shard_id,
3474 0 : timeline_id,
3475 0 : node,
3476 0 : self.config.jwt_token.clone(),
3477 0 : ))
3478 0 : })
3479 0 : .await?;
3480 :
3481 : // If any shards >0 haven't finished deletion yet, don't start deletion on shard zero
3482 0 : if statuses.iter().any(|s| s != &StatusCode::NOT_FOUND) {
3483 0 : return Ok(StatusCode::ACCEPTED);
3484 0 : }
3485 :
3486 : // Delete shard zero last: this is not strictly necessary, but since a caller's GET on a timeline will be routed
3487 : // to shard zero, it gives a more obvious behavior that a GET returns 404 once the deletion is done.
3488 0 : let shard_zero_status = delete_one(
3489 0 : shard_zero.0,
3490 0 : timeline_id,
3491 0 : shard_zero.1,
3492 0 : self.config.jwt_token.clone(),
3493 0 : )
3494 0 : .await?;
3495 0 : Ok(shard_zero_status)
3496 0 : }).await?
3497 0 : }
3498 :
3499 : /// When you need to send an HTTP request to the pageserver that holds shard0 of a tenant, this
3500 : /// function looks up and returns node. If the tenant isn't found, returns Err(ApiError::NotFound)
3501 0 : pub(crate) async fn tenant_shard0_node(
3502 0 : &self,
3503 0 : tenant_id: TenantId,
3504 0 : ) -> Result<(Node, TenantShardId), ApiError> {
3505 0 : // Look up in-memory state and maybe use the node from there.
3506 0 : {
3507 0 : let locked = self.inner.read().unwrap();
3508 0 : let Some((tenant_shard_id, shard)) = locked
3509 0 : .tenants
3510 0 : .range(TenantShardId::tenant_range(tenant_id))
3511 0 : .next()
3512 : else {
3513 0 : return Err(ApiError::NotFound(
3514 0 : anyhow::anyhow!("Tenant {tenant_id} not found").into(),
3515 0 : ));
3516 : };
3517 :
3518 0 : let Some(intent_node_id) = shard.intent.get_attached() else {
3519 0 : tracing::warn!(
3520 0 : tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
3521 0 : "Shard not scheduled (policy {:?}), cannot generate pass-through URL",
3522 : shard.policy
3523 : );
3524 0 : return Err(ApiError::Conflict(
3525 0 : "Cannot call timeline API on non-attached tenant".to_string(),
3526 0 : ));
3527 : };
3528 :
3529 0 : if shard.reconciler.is_none() {
3530 : // Optimization: while no reconcile is in flight, we may trust our in-memory state
3531 : // to tell us which pageserver to use. Otherwise we will fall through and hit the database
3532 0 : let Some(node) = locked.nodes.get(intent_node_id) else {
3533 : // This should never happen
3534 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3535 0 : "Shard refers to nonexistent node"
3536 0 : )));
3537 : };
3538 0 : return Ok((node.clone(), *tenant_shard_id));
3539 0 : }
3540 : };
3541 :
3542 : // Look up the latest attached pageserver location from the database
3543 : // generation state: this will reflect the progress of any ongoing migration.
3544 : // Note that it is not guaranteed to _stay_ here, our caller must still handle
3545 : // the case where they call through to the pageserver and get a 404.
3546 0 : let db_result = self.persistence.tenant_generations(tenant_id).await?;
3547 : let Some(ShardGenerationState {
3548 0 : tenant_shard_id,
3549 0 : generation: _,
3550 0 : generation_pageserver: Some(node_id),
3551 0 : }) = db_result.first()
3552 : else {
3553 : // This can happen if we raced with a tenant deletion or a shard split. On a retry
3554 : // the caller will either succeed (shard split case), get a proper 404 (deletion case),
3555 : // or a conflict response (case where tenant was detached in background)
3556 0 : return Err(ApiError::ResourceUnavailable(
3557 0 : "Shard {} not found in database, or is not attached".into(),
3558 0 : ));
3559 : };
3560 0 : let locked = self.inner.read().unwrap();
3561 0 : let Some(node) = locked.nodes.get(node_id) else {
3562 : // This should never happen
3563 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3564 0 : "Shard refers to nonexistent node"
3565 0 : )));
3566 : };
3567 :
3568 0 : Ok((node.clone(), *tenant_shard_id))
3569 0 : }
3570 :
3571 0 : pub(crate) fn tenant_locate(
3572 0 : &self,
3573 0 : tenant_id: TenantId,
3574 0 : ) -> Result<TenantLocateResponse, ApiError> {
3575 0 : let locked = self.inner.read().unwrap();
3576 0 : tracing::info!("Locating shards for tenant {tenant_id}");
3577 :
3578 0 : let mut result = Vec::new();
3579 0 : let mut shard_params: Option<ShardParameters> = None;
3580 :
3581 0 : for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
3582 : {
3583 0 : let node_id =
3584 0 : shard
3585 0 : .intent
3586 0 : .get_attached()
3587 0 : .ok_or(ApiError::BadRequest(anyhow::anyhow!(
3588 0 : "Cannot locate a tenant that is not attached"
3589 0 : )))?;
3590 :
3591 0 : let node = locked
3592 0 : .nodes
3593 0 : .get(&node_id)
3594 0 : .expect("Pageservers may not be deleted while referenced");
3595 0 :
3596 0 : result.push(node.shard_location(*tenant_shard_id));
3597 0 :
3598 0 : match &shard_params {
3599 0 : None => {
3600 0 : shard_params = Some(ShardParameters {
3601 0 : stripe_size: shard.shard.stripe_size,
3602 0 : count: shard.shard.count,
3603 0 : });
3604 0 : }
3605 0 : Some(params) => {
3606 0 : if params.stripe_size != shard.shard.stripe_size {
3607 : // This should never happen. We enforce at runtime because it's simpler than
3608 : // adding an extra per-tenant data structure to store the things that should be the same
3609 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
3610 0 : "Inconsistent shard stripe size parameters!"
3611 0 : )));
3612 0 : }
3613 : }
3614 : }
3615 : }
3616 :
3617 0 : if result.is_empty() {
3618 0 : return Err(ApiError::NotFound(
3619 0 : anyhow::anyhow!("No shards for this tenant ID found").into(),
3620 0 : ));
3621 0 : }
3622 0 : let shard_params = shard_params.expect("result is non-empty, therefore this is set");
3623 0 : tracing::info!(
3624 0 : "Located tenant {} with params {:?} on shards {}",
3625 0 : tenant_id,
3626 0 : shard_params,
3627 0 : result
3628 0 : .iter()
3629 0 : .map(|s| format!("{:?}", s))
3630 0 : .collect::<Vec<_>>()
3631 0 : .join(",")
3632 : );
3633 :
3634 0 : Ok(TenantLocateResponse {
3635 0 : shards: result,
3636 0 : shard_params,
3637 0 : })
3638 0 : }
3639 :
3640 : /// Returns None if the input iterator of shards does not include a shard with number=0
3641 0 : fn tenant_describe_impl<'a>(
3642 0 : &self,
3643 0 : shards: impl Iterator<Item = &'a TenantShard>,
3644 0 : ) -> Option<TenantDescribeResponse> {
3645 0 : let mut shard_zero = None;
3646 0 : let mut describe_shards = Vec::new();
3647 :
3648 0 : for shard in shards {
3649 0 : if shard.tenant_shard_id.is_shard_zero() {
3650 0 : shard_zero = Some(shard);
3651 0 : }
3652 :
3653 0 : describe_shards.push(TenantDescribeResponseShard {
3654 0 : tenant_shard_id: shard.tenant_shard_id,
3655 0 : node_attached: *shard.intent.get_attached(),
3656 0 : node_secondary: shard.intent.get_secondary().to_vec(),
3657 0 : last_error: shard
3658 0 : .last_error
3659 0 : .lock()
3660 0 : .unwrap()
3661 0 : .as_ref()
3662 0 : .map(|e| format!("{e}"))
3663 0 : .unwrap_or("".to_string())
3664 0 : .clone(),
3665 0 : is_reconciling: shard.reconciler.is_some(),
3666 0 : is_pending_compute_notification: shard.pending_compute_notification,
3667 0 : is_splitting: matches!(shard.splitting, SplitState::Splitting),
3668 0 : scheduling_policy: *shard.get_scheduling_policy(),
3669 0 : preferred_az_id: shard.preferred_az().map(ToString::to_string),
3670 : })
3671 : }
3672 :
3673 0 : let shard_zero = shard_zero?;
3674 :
3675 0 : Some(TenantDescribeResponse {
3676 0 : tenant_id: shard_zero.tenant_shard_id.tenant_id,
3677 0 : shards: describe_shards,
3678 0 : stripe_size: shard_zero.shard.stripe_size,
3679 0 : policy: shard_zero.policy.clone(),
3680 0 : config: shard_zero.config.clone(),
3681 0 : })
3682 0 : }
3683 :
3684 0 : pub(crate) fn tenant_describe(
3685 0 : &self,
3686 0 : tenant_id: TenantId,
3687 0 : ) -> Result<TenantDescribeResponse, ApiError> {
3688 0 : let locked = self.inner.read().unwrap();
3689 0 :
3690 0 : self.tenant_describe_impl(
3691 0 : locked
3692 0 : .tenants
3693 0 : .range(TenantShardId::tenant_range(tenant_id))
3694 0 : .map(|(_k, v)| v),
3695 0 : )
3696 0 : .ok_or_else(|| ApiError::NotFound(anyhow::anyhow!("Tenant {tenant_id} not found").into()))
3697 0 : }
3698 :
3699 0 : pub(crate) fn tenant_list(&self) -> Vec<TenantDescribeResponse> {
3700 0 : let locked = self.inner.read().unwrap();
3701 0 :
3702 0 : let mut result = Vec::new();
3703 0 : for (_tenant_id, tenant_shards) in
3704 0 : &locked.tenants.iter().group_by(|(id, _shard)| id.tenant_id)
3705 0 : {
3706 0 : result.push(
3707 0 : self.tenant_describe_impl(tenant_shards.map(|(_k, v)| v))
3708 0 : .expect("Groups are always non-empty"),
3709 0 : );
3710 0 : }
3711 :
3712 0 : result
3713 0 : }
3714 :
3715 0 : #[instrument(skip_all, fields(tenant_id=%op.tenant_id))]
3716 : async fn abort_tenant_shard_split(
3717 : &self,
3718 : op: &TenantShardSplitAbort,
3719 : ) -> Result<(), TenantShardSplitAbortError> {
3720 : // Cleaning up a split:
3721 : // - Parent shards are not destroyed during a split, just detached.
3722 : // - Failed pageserver split API calls can leave the remote node with just the parent attached,
3723 : // just the children attached, or both.
3724 : //
3725 : // Therefore our work to do is to:
3726 : // 1. Clean up storage controller's internal state to just refer to parents, no children
3727 : // 2. Call out to pageservers to ensure that children are detached
3728 : // 3. Call out to pageservers to ensure that parents are attached.
3729 : //
3730 : // Crash safety:
3731 : // - If the storage controller stops running during this cleanup *after* clearing the splitting state
3732 : // from our database, then [`Self::startup_reconcile`] will regard child attachments as garbage
3733 : // and detach them.
3734 : // - TODO: If the storage controller stops running during this cleanup *before* clearing the splitting state
3735 : // from our database, then we will re-enter this cleanup routine on startup.
3736 :
3737 : let TenantShardSplitAbort {
3738 : tenant_id,
3739 : new_shard_count,
3740 : new_stripe_size,
3741 : ..
3742 : } = op;
3743 :
3744 : // First abort persistent state, if any exists.
3745 : match self
3746 : .persistence
3747 : .abort_shard_split(*tenant_id, *new_shard_count)
3748 : .await?
3749 : {
3750 : AbortShardSplitStatus::Aborted => {
3751 : // Proceed to roll back any child shards created on pageservers
3752 : }
3753 : AbortShardSplitStatus::Complete => {
3754 : // The split completed (we might hit that path if e.g. our database transaction
3755 : // to write the completion landed in the database, but we dropped connection
3756 : // before seeing the result).
3757 : //
3758 : // We must update in-memory state to reflect the successful split.
3759 : self.tenant_shard_split_commit_inmem(
3760 : *tenant_id,
3761 : *new_shard_count,
3762 : *new_stripe_size,
3763 : );
3764 : return Ok(());
3765 : }
3766 : }
3767 :
3768 : // Clean up in-memory state, and accumulate the list of child locations that need detaching
3769 : let detach_locations: Vec<(Node, TenantShardId)> = {
3770 : let mut detach_locations = Vec::new();
3771 : let mut locked = self.inner.write().unwrap();
3772 : let (nodes, tenants, scheduler) = locked.parts_mut();
3773 :
3774 : for (tenant_shard_id, shard) in
3775 : tenants.range_mut(TenantShardId::tenant_range(op.tenant_id))
3776 : {
3777 : if shard.shard.count == op.new_shard_count {
3778 : // Surprising: the phase of [`Self::do_tenant_shard_split`] which inserts child shards in-memory
3779 : // is infallible, so if we got an error we shouldn't have got that far.
3780 : tracing::warn!(
3781 : "During split abort, child shard {tenant_shard_id} found in-memory"
3782 : );
3783 : continue;
3784 : }
3785 :
3786 : // Add the children of this shard to this list of things to detach
3787 : if let Some(node_id) = shard.intent.get_attached() {
3788 : for child_id in tenant_shard_id.split(*new_shard_count) {
3789 : detach_locations.push((
3790 : nodes
3791 : .get(node_id)
3792 : .expect("Intent references nonexistent node")
3793 : .clone(),
3794 : child_id,
3795 : ));
3796 : }
3797 : } else {
3798 : tracing::warn!(
3799 : "During split abort, shard {tenant_shard_id} has no attached location"
3800 : );
3801 : }
3802 :
3803 : tracing::info!("Restoring parent shard {tenant_shard_id}");
3804 : shard.splitting = SplitState::Idle;
3805 : if let Err(e) = shard.schedule(scheduler, &mut ScheduleContext::default()) {
3806 : // If this shard can't be scheduled now (perhaps due to offline nodes or
3807 : // capacity issues), that must not prevent us rolling back a split. In this
3808 : // case it should be eventually scheduled in the background.
3809 : tracing::warn!("Failed to schedule {tenant_shard_id} during shard abort: {e}")
3810 : }
3811 :
3812 : self.maybe_reconcile_shard(shard, nodes);
3813 : }
3814 :
3815 : // We don't expect any new_shard_count shards to exist here, but drop them just in case
3816 0 : tenants.retain(|_id, s| s.shard.count != *new_shard_count);
3817 :
3818 : detach_locations
3819 : };
3820 :
3821 : for (node, child_id) in detach_locations {
3822 : if !node.is_available() {
3823 : // An unavailable node cannot be cleaned up now: to avoid blocking forever, we will permit this, and
3824 : // rely on the reconciliation that happens when a node transitions to Active to clean up. Since we have
3825 : // removed child shards from our in-memory state and database, the reconciliation will implicitly remove
3826 : // them from the node.
3827 : tracing::warn!("Node {node} unavailable, can't clean up during split abort. It will be cleaned up when it is reactivated.");
3828 : continue;
3829 : }
3830 :
3831 : // Detach the remote child. If the pageserver split API call is still in progress, this call will get
3832 : // a 503 and retry, up to our limit.
3833 : tracing::info!("Detaching {child_id} on {node}...");
3834 : match node
3835 : .with_client_retries(
3836 0 : |client| async move {
3837 0 : let config = LocationConfig {
3838 0 : mode: LocationConfigMode::Detached,
3839 0 : generation: None,
3840 0 : secondary_conf: None,
3841 0 : shard_number: child_id.shard_number.0,
3842 0 : shard_count: child_id.shard_count.literal(),
3843 0 : // Stripe size and tenant config don't matter when detaching
3844 0 : shard_stripe_size: 0,
3845 0 : tenant_conf: TenantConfig::default(),
3846 0 : };
3847 0 :
3848 0 : client.location_config(child_id, config, None, false).await
3849 0 : },
3850 : &self.config.jwt_token,
3851 : 1,
3852 : 10,
3853 : Duration::from_secs(5),
3854 : &self.cancel,
3855 : )
3856 : .await
3857 : {
3858 : Some(Ok(_)) => {}
3859 : Some(Err(e)) => {
3860 : // We failed to communicate with the remote node. This is problematic: we may be
3861 : // leaving it with a rogue child shard.
3862 : tracing::warn!(
3863 : "Failed to detach child {child_id} from node {node} during abort"
3864 : );
3865 : return Err(e.into());
3866 : }
3867 : None => {
3868 : // Cancellation: we were shutdown or the node went offline. Shutdown is fine, we'll
3869 : // clean up on restart. The node going offline requires a retry.
3870 : return Err(TenantShardSplitAbortError::Unavailable);
3871 : }
3872 : };
3873 : }
3874 :
3875 : tracing::info!("Successfully aborted split");
3876 : Ok(())
3877 : }
3878 :
3879 : /// Infallible final stage of [`Self::tenant_shard_split`]: update the contents
3880 : /// of the tenant map to reflect the child shards that exist after the split.
3881 0 : fn tenant_shard_split_commit_inmem(
3882 0 : &self,
3883 0 : tenant_id: TenantId,
3884 0 : new_shard_count: ShardCount,
3885 0 : new_stripe_size: Option<ShardStripeSize>,
3886 0 : ) -> (
3887 0 : TenantShardSplitResponse,
3888 0 : Vec<(TenantShardId, NodeId, ShardStripeSize)>,
3889 0 : Vec<ReconcilerWaiter>,
3890 0 : ) {
3891 0 : let mut response = TenantShardSplitResponse {
3892 0 : new_shards: Vec::new(),
3893 0 : };
3894 0 : let mut child_locations = Vec::new();
3895 0 : let mut waiters = Vec::new();
3896 0 :
3897 0 : {
3898 0 : let mut locked = self.inner.write().unwrap();
3899 0 :
3900 0 : let parent_ids = locked
3901 0 : .tenants
3902 0 : .range(TenantShardId::tenant_range(tenant_id))
3903 0 : .map(|(shard_id, _)| *shard_id)
3904 0 : .collect::<Vec<_>>();
3905 0 :
3906 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
3907 0 : for parent_id in parent_ids {
3908 0 : let child_ids = parent_id.split(new_shard_count);
3909 :
3910 0 : let (pageserver, generation, policy, parent_ident, config) = {
3911 0 : let mut old_state = tenants
3912 0 : .remove(&parent_id)
3913 0 : .expect("It was present, we just split it");
3914 0 :
3915 0 : // A non-splitting state is impossible, because [`Self::tenant_shard_split`] holds
3916 0 : // a TenantId lock and passes it through to [`TenantShardSplitAbort`] in case of cleanup:
3917 0 : // nothing else can clear this.
3918 0 : assert!(matches!(old_state.splitting, SplitState::Splitting));
3919 :
3920 0 : let old_attached = old_state.intent.get_attached().unwrap();
3921 0 : old_state.intent.clear(scheduler);
3922 0 : let generation = old_state.generation.expect("Shard must have been attached");
3923 0 : (
3924 0 : old_attached,
3925 0 : generation,
3926 0 : old_state.policy,
3927 0 : old_state.shard,
3928 0 : old_state.config,
3929 0 : )
3930 0 : };
3931 0 :
3932 0 : let mut schedule_context = ScheduleContext::default();
3933 0 : for child in child_ids {
3934 0 : let mut child_shard = parent_ident;
3935 0 : child_shard.number = child.shard_number;
3936 0 : child_shard.count = child.shard_count;
3937 0 : if let Some(stripe_size) = new_stripe_size {
3938 0 : child_shard.stripe_size = stripe_size;
3939 0 : }
3940 :
3941 0 : let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
3942 0 : child_observed.insert(
3943 0 : pageserver,
3944 0 : ObservedStateLocation {
3945 0 : conf: Some(attached_location_conf(
3946 0 : generation,
3947 0 : &child_shard,
3948 0 : &config,
3949 0 : &policy,
3950 0 : )),
3951 0 : },
3952 0 : );
3953 0 :
3954 0 : let mut child_state = TenantShard::new(child, child_shard, policy.clone());
3955 0 : child_state.intent = IntentState::single(scheduler, Some(pageserver));
3956 0 : child_state.observed = ObservedState {
3957 0 : locations: child_observed,
3958 0 : };
3959 0 : child_state.generation = Some(generation);
3960 0 : child_state.config = config.clone();
3961 0 :
3962 0 : // The child's TenantShard::splitting is intentionally left at the default value of Idle,
3963 0 : // as at this point in the split process we have succeeded and this part is infallible:
3964 0 : // we will never need to do any special recovery from this state.
3965 0 :
3966 0 : child_locations.push((child, pageserver, child_shard.stripe_size));
3967 :
3968 0 : if let Err(e) = child_state.schedule(scheduler, &mut schedule_context) {
3969 : // This is not fatal, because we've implicitly already got an attached
3970 : // location for the child shard. Failure here just means we couldn't
3971 : // find a secondary (e.g. because cluster is overloaded).
3972 0 : tracing::warn!("Failed to schedule child shard {child}: {e}");
3973 0 : }
3974 : // In the background, attach secondary locations for the new shards
3975 0 : if let Some(waiter) = self.maybe_reconcile_shard(&mut child_state, nodes) {
3976 0 : waiters.push(waiter);
3977 0 : }
3978 :
3979 0 : tenants.insert(child, child_state);
3980 0 : response.new_shards.push(child);
3981 : }
3982 : }
3983 0 : (response, child_locations, waiters)
3984 0 : }
3985 0 : }
3986 :
3987 0 : async fn tenant_shard_split_start_secondaries(
3988 0 : &self,
3989 0 : tenant_id: TenantId,
3990 0 : waiters: Vec<ReconcilerWaiter>,
3991 0 : ) {
3992 : // Wait for initial reconcile of child shards, this creates the secondary locations
3993 0 : if let Err(e) = self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
3994 : // This is not a failure to split: it's some issue reconciling the new child shards, perhaps
3995 : // their secondaries couldn't be attached.
3996 0 : tracing::warn!("Failed to reconcile after split: {e}");
3997 0 : return;
3998 0 : }
3999 :
4000 : // Take the state lock to discover the attached & secondary intents for all shards
4001 0 : let (attached, secondary) = {
4002 0 : let locked = self.inner.read().unwrap();
4003 0 : let mut attached = Vec::new();
4004 0 : let mut secondary = Vec::new();
4005 :
4006 0 : for (tenant_shard_id, shard) in
4007 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4008 : {
4009 0 : let Some(node_id) = shard.intent.get_attached() else {
4010 : // Unexpected. Race with a PlacementPolicy change?
4011 0 : tracing::warn!(
4012 0 : "No attached node on {tenant_shard_id} immediately after shard split!"
4013 : );
4014 0 : continue;
4015 : };
4016 :
4017 0 : let Some(secondary_node_id) = shard.intent.get_secondary().first() else {
4018 : // No secondary location. Nothing for us to do.
4019 0 : continue;
4020 : };
4021 :
4022 0 : let attached_node = locked
4023 0 : .nodes
4024 0 : .get(node_id)
4025 0 : .expect("Pageservers may not be deleted while referenced");
4026 0 :
4027 0 : let secondary_node = locked
4028 0 : .nodes
4029 0 : .get(secondary_node_id)
4030 0 : .expect("Pageservers may not be deleted while referenced");
4031 0 :
4032 0 : attached.push((*tenant_shard_id, attached_node.clone()));
4033 0 : secondary.push((*tenant_shard_id, secondary_node.clone()));
4034 : }
4035 0 : (attached, secondary)
4036 0 : };
4037 0 :
4038 0 : if secondary.is_empty() {
4039 : // No secondary locations; nothing for us to do
4040 0 : return;
4041 0 : }
4042 :
4043 0 : for result in self
4044 0 : .tenant_for_shards_api(
4045 0 : attached,
4046 0 : |tenant_shard_id, client| async move {
4047 0 : client.tenant_heatmap_upload(tenant_shard_id).await
4048 0 : },
4049 0 : 1,
4050 0 : 1,
4051 0 : SHORT_RECONCILE_TIMEOUT,
4052 0 : &self.cancel,
4053 0 : )
4054 0 : .await
4055 : {
4056 0 : if let Err(e) = result {
4057 0 : tracing::warn!("Error calling heatmap upload after shard split: {e}");
4058 0 : return;
4059 0 : }
4060 : }
4061 :
4062 0 : for result in self
4063 0 : .tenant_for_shards_api(
4064 0 : secondary,
4065 0 : |tenant_shard_id, client| async move {
4066 0 : client
4067 0 : .tenant_secondary_download(tenant_shard_id, Some(Duration::ZERO))
4068 0 : .await
4069 0 : },
4070 0 : 1,
4071 0 : 1,
4072 0 : SHORT_RECONCILE_TIMEOUT,
4073 0 : &self.cancel,
4074 0 : )
4075 0 : .await
4076 : {
4077 0 : if let Err(e) = result {
4078 0 : tracing::warn!("Error calling secondary download after shard split: {e}");
4079 0 : return;
4080 0 : }
4081 : }
4082 0 : }
4083 :
4084 0 : pub(crate) async fn tenant_shard_split(
4085 0 : &self,
4086 0 : tenant_id: TenantId,
4087 0 : split_req: TenantShardSplitRequest,
4088 0 : ) -> Result<TenantShardSplitResponse, ApiError> {
4089 : // TODO: return 503 if we get stuck waiting for this lock
4090 : // (issue https://github.com/neondatabase/neon/issues/7108)
4091 0 : let _tenant_lock = trace_exclusive_lock(
4092 0 : &self.tenant_op_locks,
4093 0 : tenant_id,
4094 0 : TenantOperations::ShardSplit,
4095 0 : )
4096 0 : .await;
4097 :
4098 0 : let new_shard_count = ShardCount::new(split_req.new_shard_count);
4099 0 : let new_stripe_size = split_req.new_stripe_size;
4100 :
4101 : // Validate the request and construct parameters. This phase is fallible, but does not require
4102 : // rollback on errors, as it does no I/O and mutates no state.
4103 0 : let shard_split_params = match self.prepare_tenant_shard_split(tenant_id, split_req)? {
4104 0 : ShardSplitAction::NoOp(resp) => return Ok(resp),
4105 0 : ShardSplitAction::Split(params) => params,
4106 : };
4107 :
4108 : // Execute this split: this phase mutates state and does remote I/O on pageservers. If it fails,
4109 : // we must roll back.
4110 0 : let r = self
4111 0 : .do_tenant_shard_split(tenant_id, shard_split_params)
4112 0 : .await;
4113 :
4114 0 : let (response, waiters) = match r {
4115 0 : Ok(r) => r,
4116 0 : Err(e) => {
4117 0 : // Split might be part-done, we must do work to abort it.
4118 0 : tracing::warn!("Enqueuing background abort of split on {tenant_id}");
4119 0 : self.abort_tx
4120 0 : .send(TenantShardSplitAbort {
4121 0 : tenant_id,
4122 0 : new_shard_count,
4123 0 : new_stripe_size,
4124 0 : _tenant_lock,
4125 0 : })
4126 0 : // Ignore error sending: that just means we're shutting down: aborts are ephemeral so it's fine to drop it.
4127 0 : .ok();
4128 0 : return Err(e);
4129 : }
4130 : };
4131 :
4132 : // The split is now complete. As an optimization, we will trigger all the child shards to upload
4133 : // a heatmap immediately, and all their secondary locations to start downloading: this avoids waiting
4134 : // for the background heatmap/download interval before secondaries get warm enough to migrate shards
4135 : // in [`Self::optimize_all`]
4136 0 : self.tenant_shard_split_start_secondaries(tenant_id, waiters)
4137 0 : .await;
4138 0 : Ok(response)
4139 0 : }
4140 :
4141 0 : fn prepare_tenant_shard_split(
4142 0 : &self,
4143 0 : tenant_id: TenantId,
4144 0 : split_req: TenantShardSplitRequest,
4145 0 : ) -> Result<ShardSplitAction, ApiError> {
4146 0 : fail::fail_point!("shard-split-validation", |_| Err(ApiError::BadRequest(
4147 0 : anyhow::anyhow!("failpoint")
4148 0 : )));
4149 :
4150 0 : let mut policy = None;
4151 0 : let mut config = None;
4152 0 : let mut shard_ident = None;
4153 : // Validate input, and calculate which shards we will create
4154 0 : let (old_shard_count, targets) =
4155 : {
4156 0 : let locked = self.inner.read().unwrap();
4157 0 :
4158 0 : let pageservers = locked.nodes.clone();
4159 0 :
4160 0 : let mut targets = Vec::new();
4161 0 :
4162 0 : // In case this is a retry, count how many already-split shards we found
4163 0 : let mut children_found = Vec::new();
4164 0 : let mut old_shard_count = None;
4165 :
4166 0 : for (tenant_shard_id, shard) in
4167 0 : locked.tenants.range(TenantShardId::tenant_range(tenant_id))
4168 : {
4169 0 : match shard.shard.count.count().cmp(&split_req.new_shard_count) {
4170 : Ordering::Equal => {
4171 : // Already split this
4172 0 : children_found.push(*tenant_shard_id);
4173 0 : continue;
4174 : }
4175 : Ordering::Greater => {
4176 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4177 0 : "Requested count {} but already have shards at count {}",
4178 0 : split_req.new_shard_count,
4179 0 : shard.shard.count.count()
4180 0 : )));
4181 : }
4182 0 : Ordering::Less => {
4183 0 : // Fall through: this shard has lower count than requested,
4184 0 : // is a candidate for splitting.
4185 0 : }
4186 0 : }
4187 0 :
4188 0 : match old_shard_count {
4189 0 : None => old_shard_count = Some(shard.shard.count),
4190 0 : Some(old_shard_count) => {
4191 0 : if old_shard_count != shard.shard.count {
4192 : // We may hit this case if a caller asked for two splits to
4193 : // different sizes, before the first one is complete.
4194 : // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
4195 : // of shard_count=1 and shard_count=2 shards in the map.
4196 0 : return Err(ApiError::Conflict(
4197 0 : "Cannot split, currently mid-split".to_string(),
4198 0 : ));
4199 0 : }
4200 : }
4201 : }
4202 0 : if policy.is_none() {
4203 0 : policy = Some(shard.policy.clone());
4204 0 : }
4205 0 : if shard_ident.is_none() {
4206 0 : shard_ident = Some(shard.shard);
4207 0 : }
4208 0 : if config.is_none() {
4209 0 : config = Some(shard.config.clone());
4210 0 : }
4211 :
4212 0 : if tenant_shard_id.shard_count.count() == split_req.new_shard_count {
4213 0 : tracing::info!(
4214 0 : "Tenant shard {} already has shard count {}",
4215 : tenant_shard_id,
4216 : split_req.new_shard_count
4217 : );
4218 0 : continue;
4219 0 : }
4220 :
4221 0 : let node_id = shard.intent.get_attached().ok_or(ApiError::BadRequest(
4222 0 : anyhow::anyhow!("Cannot split a tenant that is not attached"),
4223 0 : ))?;
4224 :
4225 0 : let node = pageservers
4226 0 : .get(&node_id)
4227 0 : .expect("Pageservers may not be deleted while referenced");
4228 0 :
4229 0 : targets.push(ShardSplitTarget {
4230 0 : parent_id: *tenant_shard_id,
4231 0 : node: node.clone(),
4232 0 : child_ids: tenant_shard_id
4233 0 : .split(ShardCount::new(split_req.new_shard_count)),
4234 0 : });
4235 : }
4236 :
4237 0 : if targets.is_empty() {
4238 0 : if children_found.len() == split_req.new_shard_count as usize {
4239 0 : return Ok(ShardSplitAction::NoOp(TenantShardSplitResponse {
4240 0 : new_shards: children_found,
4241 0 : }));
4242 : } else {
4243 : // No shards found to split, and no existing children found: the
4244 : // tenant doesn't exist at all.
4245 0 : return Err(ApiError::NotFound(
4246 0 : anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
4247 0 : ));
4248 : }
4249 0 : }
4250 0 :
4251 0 : (old_shard_count, targets)
4252 0 : };
4253 0 :
4254 0 : // unwrap safety: we would have returned above if we didn't find at least one shard to split
4255 0 : let old_shard_count = old_shard_count.unwrap();
4256 0 : let shard_ident = if let Some(new_stripe_size) = split_req.new_stripe_size {
4257 : // This ShardIdentity will be used as the template for all children, so this implicitly
4258 : // applies the new stripe size to the children.
4259 0 : let mut shard_ident = shard_ident.unwrap();
4260 0 : if shard_ident.count.count() > 1 && shard_ident.stripe_size != new_stripe_size {
4261 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("Attempted to change stripe size ({:?}->{new_stripe_size:?}) on a tenant with multiple shards", shard_ident.stripe_size)));
4262 0 : }
4263 0 :
4264 0 : shard_ident.stripe_size = new_stripe_size;
4265 0 : tracing::info!("applied stripe size {}", shard_ident.stripe_size.0);
4266 0 : shard_ident
4267 : } else {
4268 0 : shard_ident.unwrap()
4269 : };
4270 0 : let policy = policy.unwrap();
4271 0 : let config = config.unwrap();
4272 0 :
4273 0 : Ok(ShardSplitAction::Split(Box::new(ShardSplitParams {
4274 0 : old_shard_count,
4275 0 : new_shard_count: ShardCount::new(split_req.new_shard_count),
4276 0 : new_stripe_size: split_req.new_stripe_size,
4277 0 : targets,
4278 0 : policy,
4279 0 : config,
4280 0 : shard_ident,
4281 0 : })))
4282 0 : }
4283 :
4284 0 : async fn do_tenant_shard_split(
4285 0 : &self,
4286 0 : tenant_id: TenantId,
4287 0 : params: Box<ShardSplitParams>,
4288 0 : ) -> Result<(TenantShardSplitResponse, Vec<ReconcilerWaiter>), ApiError> {
4289 0 : // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
4290 0 : // request could occur here, deleting or mutating the tenant. begin_shard_split checks that the
4291 0 : // parent shards exist as expected, but it would be neater to do the above pre-checks within the
4292 0 : // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
4293 0 : // (https://github.com/neondatabase/neon/issues/6676)
4294 0 :
4295 0 : let ShardSplitParams {
4296 0 : old_shard_count,
4297 0 : new_shard_count,
4298 0 : new_stripe_size,
4299 0 : mut targets,
4300 0 : policy,
4301 0 : config,
4302 0 : shard_ident,
4303 0 : } = *params;
4304 :
4305 : // Drop any secondary locations: pageservers do not support splitting these, and in any case the
4306 : // end-state for a split tenant will usually be to have secondary locations on different nodes.
4307 : // The reconciliation calls in this block also implicitly cancel+barrier wrt any ongoing reconciliation
4308 : // at the time of split.
4309 0 : let waiters = {
4310 0 : let mut locked = self.inner.write().unwrap();
4311 0 : let mut waiters = Vec::new();
4312 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4313 0 : for target in &mut targets {
4314 0 : let Some(shard) = tenants.get_mut(&target.parent_id) else {
4315 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4316 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4317 0 : "Shard {} not found",
4318 0 : target.parent_id
4319 0 : )));
4320 : };
4321 :
4322 0 : if shard.intent.get_attached() != &Some(target.node.get_id()) {
4323 : // Paranoia check: this shouldn't happen: we have the oplock for this tenant ID.
4324 0 : return Err(ApiError::Conflict(format!(
4325 0 : "Shard {} unexpectedly rescheduled during split",
4326 0 : target.parent_id
4327 0 : )));
4328 0 : }
4329 0 :
4330 0 : // Irrespective of PlacementPolicy, clear secondary locations from intent
4331 0 : shard.intent.clear_secondary(scheduler);
4332 :
4333 : // Run Reconciler to execute detach fo secondary locations.
4334 0 : if let Some(waiter) = self.maybe_reconcile_shard(shard, nodes) {
4335 0 : waiters.push(waiter);
4336 0 : }
4337 : }
4338 0 : waiters
4339 0 : };
4340 0 : self.await_waiters(waiters, RECONCILE_TIMEOUT).await?;
4341 :
4342 : // Before creating any new child shards in memory or on the pageservers, persist them: this
4343 : // enables us to ensure that we will always be able to clean up if something goes wrong. This also
4344 : // acts as the protection against two concurrent attempts to split: one of them will get a database
4345 : // error trying to insert the child shards.
4346 0 : let mut child_tsps = Vec::new();
4347 0 : for target in &targets {
4348 0 : let mut this_child_tsps = Vec::new();
4349 0 : for child in &target.child_ids {
4350 0 : let mut child_shard = shard_ident;
4351 0 : child_shard.number = child.shard_number;
4352 0 : child_shard.count = child.shard_count;
4353 0 :
4354 0 : tracing::info!(
4355 0 : "Create child shard persistence with stripe size {}",
4356 : shard_ident.stripe_size.0
4357 : );
4358 :
4359 0 : this_child_tsps.push(TenantShardPersistence {
4360 0 : tenant_id: child.tenant_id.to_string(),
4361 0 : shard_number: child.shard_number.0 as i32,
4362 0 : shard_count: child.shard_count.literal() as i32,
4363 0 : shard_stripe_size: shard_ident.stripe_size.0 as i32,
4364 0 : // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
4365 0 : // populate the correct generation as part of its transaction, to protect us
4366 0 : // against racing with changes in the state of the parent.
4367 0 : generation: None,
4368 0 : generation_pageserver: Some(target.node.get_id().0 as i64),
4369 0 : placement_policy: serde_json::to_string(&policy).unwrap(),
4370 0 : config: serde_json::to_string(&config).unwrap(),
4371 0 : splitting: SplitState::Splitting,
4372 0 :
4373 0 : // Scheduling policies and preferred AZ do not carry through to children
4374 0 : scheduling_policy: serde_json::to_string(&ShardSchedulingPolicy::default())
4375 0 : .unwrap(),
4376 0 : preferred_az_id: None,
4377 0 : });
4378 : }
4379 :
4380 0 : child_tsps.push((target.parent_id, this_child_tsps));
4381 : }
4382 :
4383 0 : if let Err(e) = self
4384 0 : .persistence
4385 0 : .begin_shard_split(old_shard_count, tenant_id, child_tsps)
4386 0 : .await
4387 : {
4388 0 : match e {
4389 : DatabaseError::Query(diesel::result::Error::DatabaseError(
4390 : DatabaseErrorKind::UniqueViolation,
4391 : _,
4392 : )) => {
4393 : // Inserting a child shard violated a unique constraint: we raced with another call to
4394 : // this function
4395 0 : tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
4396 0 : return Err(ApiError::Conflict("Tenant is already splitting".into()));
4397 : }
4398 0 : _ => return Err(ApiError::InternalServerError(e.into())),
4399 : }
4400 0 : }
4401 0 : fail::fail_point!("shard-split-post-begin", |_| Err(
4402 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4403 0 : ));
4404 :
4405 : // Now that I have persisted the splitting state, apply it in-memory. This is infallible, so
4406 : // callers may assume that if splitting is set in memory, then it was persisted, and if splitting
4407 : // is not set in memory, then it was not persisted.
4408 : {
4409 0 : let mut locked = self.inner.write().unwrap();
4410 0 : for target in &targets {
4411 0 : if let Some(parent_shard) = locked.tenants.get_mut(&target.parent_id) {
4412 0 : parent_shard.splitting = SplitState::Splitting;
4413 0 : // Put the observed state to None, to reflect that it is indeterminate once we start the
4414 0 : // split operation.
4415 0 : parent_shard
4416 0 : .observed
4417 0 : .locations
4418 0 : .insert(target.node.get_id(), ObservedStateLocation { conf: None });
4419 0 : }
4420 : }
4421 : }
4422 :
4423 : // TODO: issue split calls concurrently (this only matters once we're splitting
4424 : // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
4425 :
4426 0 : for target in &targets {
4427 : let ShardSplitTarget {
4428 0 : parent_id,
4429 0 : node,
4430 0 : child_ids,
4431 0 : } = target;
4432 0 : let client = PageserverClient::new(
4433 0 : node.get_id(),
4434 0 : node.base_url(),
4435 0 : self.config.jwt_token.as_deref(),
4436 0 : );
4437 0 : let response = client
4438 0 : .tenant_shard_split(
4439 0 : *parent_id,
4440 0 : TenantShardSplitRequest {
4441 0 : new_shard_count: new_shard_count.literal(),
4442 0 : new_stripe_size,
4443 0 : },
4444 0 : )
4445 0 : .await
4446 0 : .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
4447 :
4448 0 : fail::fail_point!("shard-split-post-remote", |_| Err(ApiError::Conflict(
4449 0 : "failpoint".to_string()
4450 0 : )));
4451 :
4452 0 : failpoint_support::sleep_millis_async!("shard-split-post-remote-sleep", &self.cancel);
4453 :
4454 0 : tracing::info!(
4455 0 : "Split {} into {}",
4456 0 : parent_id,
4457 0 : response
4458 0 : .new_shards
4459 0 : .iter()
4460 0 : .map(|s| format!("{:?}", s))
4461 0 : .collect::<Vec<_>>()
4462 0 : .join(",")
4463 : );
4464 :
4465 0 : if &response.new_shards != child_ids {
4466 : // This should never happen: the pageserver should agree with us on how shard splits work.
4467 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4468 0 : "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
4469 0 : parent_id,
4470 0 : response.new_shards,
4471 0 : child_ids
4472 0 : )));
4473 0 : }
4474 : }
4475 :
4476 : // TODO: if the pageserver restarted concurrently with our split API call,
4477 : // the actual generation of the child shard might differ from the generation
4478 : // we expect it to have. In order for our in-database generation to end up
4479 : // correct, we should carry the child generation back in the response and apply it here
4480 : // in complete_shard_split (and apply the correct generation in memory)
4481 : // (or, we can carry generation in the request and reject the request if
4482 : // it doesn't match, but that requires more retry logic on this side)
4483 :
4484 0 : self.persistence
4485 0 : .complete_shard_split(tenant_id, old_shard_count)
4486 0 : .await?;
4487 :
4488 0 : fail::fail_point!("shard-split-post-complete", |_| Err(
4489 0 : ApiError::InternalServerError(anyhow::anyhow!("failpoint"))
4490 0 : ));
4491 :
4492 : // Replace all the shards we just split with their children: this phase is infallible.
4493 0 : let (response, child_locations, waiters) =
4494 0 : self.tenant_shard_split_commit_inmem(tenant_id, new_shard_count, new_stripe_size);
4495 0 :
4496 0 : // Now that we have scheduled the child shards, attempt to set their preferred AZ
4497 0 : // to that of the pageserver they've been attached on.
4498 0 : let preferred_azs = {
4499 0 : let locked = self.inner.read().unwrap();
4500 0 : child_locations
4501 0 : .iter()
4502 0 : .filter_map(|(tid, node_id, _stripe_size)| {
4503 0 : let az_id = locked
4504 0 : .nodes
4505 0 : .get(node_id)
4506 0 : .map(|n| n.get_availability_zone_id().to_string())?;
4507 :
4508 0 : Some((*tid, az_id))
4509 0 : })
4510 0 : .collect::<Vec<_>>()
4511 : };
4512 :
4513 0 : let updated = self
4514 0 : .persistence
4515 0 : .set_tenant_shard_preferred_azs(preferred_azs)
4516 0 : .await
4517 0 : .map_err(|err| {
4518 0 : ApiError::InternalServerError(anyhow::anyhow!(
4519 0 : "Failed to persist preferred az ids: {err}"
4520 0 : ))
4521 0 : });
4522 0 :
4523 0 : match updated {
4524 0 : Ok(updated) => {
4525 0 : let mut locked = self.inner.write().unwrap();
4526 0 : for (tid, az_id) in updated {
4527 0 : if let Some(shard) = locked.tenants.get_mut(&tid) {
4528 0 : shard.set_preferred_az(az_id);
4529 0 : }
4530 : }
4531 : }
4532 0 : Err(err) => {
4533 0 : tracing::warn!("Failed to persist preferred AZs after split: {err}");
4534 : }
4535 : }
4536 :
4537 : // Send compute notifications for all the new shards
4538 0 : let mut failed_notifications = Vec::new();
4539 0 : for (child_id, child_ps, stripe_size) in child_locations {
4540 0 : if let Err(e) = self
4541 0 : .compute_hook
4542 0 : .notify(child_id, child_ps, stripe_size, &self.cancel)
4543 0 : .await
4544 : {
4545 0 : tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
4546 : child_id, child_ps);
4547 0 : failed_notifications.push(child_id);
4548 0 : }
4549 : }
4550 :
4551 : // If we failed any compute notifications, make a note to retry later.
4552 0 : if !failed_notifications.is_empty() {
4553 0 : let mut locked = self.inner.write().unwrap();
4554 0 : for failed in failed_notifications {
4555 0 : if let Some(shard) = locked.tenants.get_mut(&failed) {
4556 0 : shard.pending_compute_notification = true;
4557 0 : }
4558 : }
4559 0 : }
4560 :
4561 0 : Ok((response, waiters))
4562 0 : }
4563 :
4564 0 : pub(crate) async fn tenant_shard_migrate(
4565 0 : &self,
4566 0 : tenant_shard_id: TenantShardId,
4567 0 : migrate_req: TenantShardMigrateRequest,
4568 0 : ) -> Result<TenantShardMigrateResponse, ApiError> {
4569 0 : let waiter = {
4570 0 : let mut locked = self.inner.write().unwrap();
4571 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4572 :
4573 0 : let Some(node) = nodes.get(&migrate_req.node_id) else {
4574 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4575 0 : "Node {} not found",
4576 0 : migrate_req.node_id
4577 0 : )));
4578 : };
4579 :
4580 0 : if !node.is_available() {
4581 : // Warn but proceed: the caller may intend to manually adjust the placement of
4582 : // a shard even if the node is down, e.g. if intervening during an incident.
4583 0 : tracing::warn!("Migrating to unavailable node {node}");
4584 0 : }
4585 :
4586 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
4587 0 : return Err(ApiError::NotFound(
4588 0 : anyhow::anyhow!("Tenant shard not found").into(),
4589 0 : ));
4590 : };
4591 :
4592 0 : if shard.intent.get_attached() == &Some(migrate_req.node_id) {
4593 : // No-op case: we will still proceed to wait for reconciliation in case it is
4594 : // incomplete from an earlier update to the intent.
4595 0 : tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
4596 : } else {
4597 0 : let old_attached = *shard.intent.get_attached();
4598 0 :
4599 0 : match shard.policy {
4600 0 : PlacementPolicy::Attached(n) => {
4601 0 : // If our new attached node was a secondary, it no longer should be.
4602 0 : shard.intent.remove_secondary(scheduler, migrate_req.node_id);
4603 :
4604 : // If we were already attached to something, demote that to a secondary
4605 0 : if let Some(old_attached) = old_attached {
4606 0 : if n > 0 {
4607 : // Remove other secondaries to make room for the location we'll demote
4608 0 : while shard.intent.get_secondary().len() >= n {
4609 0 : shard.intent.pop_secondary(scheduler);
4610 0 : }
4611 :
4612 0 : shard.intent.push_secondary(scheduler, old_attached);
4613 0 : }
4614 0 : }
4615 :
4616 0 : shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
4617 : }
4618 0 : PlacementPolicy::Secondary => {
4619 0 : shard.intent.clear(scheduler);
4620 0 : shard.intent.push_secondary(scheduler, migrate_req.node_id);
4621 0 : }
4622 : PlacementPolicy::Detached => {
4623 0 : return Err(ApiError::BadRequest(anyhow::anyhow!(
4624 0 : "Cannot migrate a tenant that is PlacementPolicy::Detached: configure it to an attached policy first"
4625 0 : )))
4626 : }
4627 : }
4628 :
4629 0 : tracing::info!("Migrating: new intent {:?}", shard.intent);
4630 0 : shard.sequence = shard.sequence.next();
4631 : }
4632 :
4633 0 : self.maybe_reconcile_shard(shard, nodes)
4634 : };
4635 :
4636 0 : if let Some(waiter) = waiter {
4637 0 : waiter.wait_timeout(RECONCILE_TIMEOUT).await?;
4638 : } else {
4639 0 : tracing::info!("Migration is a no-op");
4640 : }
4641 :
4642 0 : Ok(TenantShardMigrateResponse {})
4643 0 : }
4644 :
4645 : /// This is for debug/support only: we simply drop all state for a tenant, without
4646 : /// detaching or deleting it on pageservers.
4647 0 : pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
4648 0 : self.persistence.delete_tenant(tenant_id).await?;
4649 :
4650 0 : let mut locked = self.inner.write().unwrap();
4651 0 : let (_nodes, tenants, scheduler) = locked.parts_mut();
4652 0 : let mut shards = Vec::new();
4653 0 : for (tenant_shard_id, _) in tenants.range(TenantShardId::tenant_range(tenant_id)) {
4654 0 : shards.push(*tenant_shard_id);
4655 0 : }
4656 :
4657 0 : for shard_id in shards {
4658 0 : if let Some(mut shard) = tenants.remove(&shard_id) {
4659 0 : shard.intent.clear(scheduler);
4660 0 : }
4661 : }
4662 :
4663 0 : Ok(())
4664 0 : }
4665 :
4666 : /// This is for debug/support only: assuming tenant data is already present in S3, we "create" a
4667 : /// tenant with a very high generation number so that it will see the existing data.
4668 0 : pub(crate) async fn tenant_import(
4669 0 : &self,
4670 0 : tenant_id: TenantId,
4671 0 : ) -> Result<TenantCreateResponse, ApiError> {
4672 0 : // Pick an arbitrary available pageserver to use for scanning the tenant in remote storage
4673 0 : let maybe_node = {
4674 0 : self.inner
4675 0 : .read()
4676 0 : .unwrap()
4677 0 : .nodes
4678 0 : .values()
4679 0 : .find(|n| n.is_available())
4680 0 : .cloned()
4681 : };
4682 0 : let Some(node) = maybe_node else {
4683 0 : return Err(ApiError::BadRequest(anyhow::anyhow!("No nodes available")));
4684 : };
4685 :
4686 0 : let client = PageserverClient::new(
4687 0 : node.get_id(),
4688 0 : node.base_url(),
4689 0 : self.config.jwt_token.as_deref(),
4690 0 : );
4691 :
4692 0 : let scan_result = client
4693 0 : .tenant_scan_remote_storage(tenant_id)
4694 0 : .await
4695 0 : .map_err(|e| passthrough_api_error(&node, e))?;
4696 :
4697 : // A post-split tenant may contain a mixture of shard counts in remote storage: pick the highest count.
4698 0 : let Some(shard_count) = scan_result
4699 0 : .shards
4700 0 : .iter()
4701 0 : .map(|s| s.tenant_shard_id.shard_count)
4702 0 : .max()
4703 : else {
4704 0 : return Err(ApiError::NotFound(
4705 0 : anyhow::anyhow!("No shards found").into(),
4706 0 : ));
4707 : };
4708 :
4709 : // Ideally we would set each newly imported shard's generation independently, but for correctness it is sufficient
4710 : // to
4711 0 : let generation = scan_result
4712 0 : .shards
4713 0 : .iter()
4714 0 : .map(|s| s.generation)
4715 0 : .max()
4716 0 : .expect("We already validated >0 shards");
4717 0 :
4718 0 : // FIXME: we have no way to recover the shard stripe size from contents of remote storage: this will
4719 0 : // only work if they were using the default stripe size.
4720 0 : let stripe_size = ShardParameters::DEFAULT_STRIPE_SIZE;
4721 :
4722 0 : let (response, waiters) = self
4723 0 : .do_tenant_create(TenantCreateRequest {
4724 0 : new_tenant_id: TenantShardId::unsharded(tenant_id),
4725 0 : generation,
4726 0 :
4727 0 : shard_parameters: ShardParameters {
4728 0 : count: shard_count,
4729 0 : stripe_size,
4730 0 : },
4731 0 : placement_policy: Some(PlacementPolicy::Attached(0)), // No secondaries, for convenient debug/hacking
4732 0 :
4733 0 : // There is no way to know what the tenant's config was: revert to defaults
4734 0 : //
4735 0 : // TODO: remove `switch_aux_file_policy` once we finish auxv2 migration
4736 0 : //
4737 0 : // we write to both v1+v2 storage, so that the test case can use either storage format for testing
4738 0 : config: TenantConfig {
4739 0 : switch_aux_file_policy: Some(models::AuxFilePolicy::CrossValidation),
4740 0 : ..TenantConfig::default()
4741 0 : },
4742 0 : })
4743 0 : .await?;
4744 :
4745 0 : if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
4746 : // Since this is a debug/support operation, all kinds of weird issues are possible (e.g. this
4747 : // tenant doesn't exist in the control plane), so don't fail the request if it can't fully
4748 : // reconcile, as reconciliation includes notifying compute.
4749 0 : tracing::warn!(%tenant_id, "Reconcile not done yet while importing tenant ({e})");
4750 0 : }
4751 :
4752 0 : Ok(response)
4753 0 : }
4754 :
4755 : /// For debug/support: a full JSON dump of TenantShards. Returns a response so that
4756 : /// we don't have to make TenantShard clonable in the return path.
4757 0 : pub(crate) fn tenants_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4758 0 : let serialized = {
4759 0 : let locked = self.inner.read().unwrap();
4760 0 : let result = locked.tenants.values().collect::<Vec<_>>();
4761 0 : serde_json::to_string(&result).map_err(|e| ApiError::InternalServerError(e.into()))?
4762 : };
4763 :
4764 0 : hyper::Response::builder()
4765 0 : .status(hyper::StatusCode::OK)
4766 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4767 0 : .body(hyper::Body::from(serialized))
4768 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4769 0 : }
4770 :
4771 : /// Check the consistency of in-memory state vs. persistent state, and check that the
4772 : /// scheduler's statistics are up to date.
4773 : ///
4774 : /// These consistency checks expect an **idle** system. If changes are going on while
4775 : /// we run, then we can falsely indicate a consistency issue. This is sufficient for end-of-test
4776 : /// checks, but not suitable for running continuously in the background in the field.
4777 0 : pub(crate) async fn consistency_check(&self) -> Result<(), ApiError> {
4778 0 : let (mut expect_nodes, mut expect_shards) = {
4779 0 : let locked = self.inner.read().unwrap();
4780 0 :
4781 0 : locked
4782 0 : .scheduler
4783 0 : .consistency_check(locked.nodes.values(), locked.tenants.values())
4784 0 : .context("Scheduler checks")
4785 0 : .map_err(ApiError::InternalServerError)?;
4786 :
4787 0 : let expect_nodes = locked
4788 0 : .nodes
4789 0 : .values()
4790 0 : .map(|n| n.to_persistent())
4791 0 : .collect::<Vec<_>>();
4792 0 :
4793 0 : let expect_shards = locked
4794 0 : .tenants
4795 0 : .values()
4796 0 : .map(|t| t.to_persistent())
4797 0 : .collect::<Vec<_>>();
4798 :
4799 : // This method can only validate the state of an idle system: if a reconcile is in
4800 : // progress, fail out early to avoid giving false errors on state that won't match
4801 : // between database and memory under a ReconcileResult is processed.
4802 0 : for t in locked.tenants.values() {
4803 0 : if t.reconciler.is_some() {
4804 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4805 0 : "Shard {} reconciliation in progress",
4806 0 : t.tenant_shard_id
4807 0 : )));
4808 0 : }
4809 : }
4810 :
4811 0 : (expect_nodes, expect_shards)
4812 : };
4813 :
4814 0 : let mut nodes = self.persistence.list_nodes().await?;
4815 0 : expect_nodes.sort_by_key(|n| n.node_id);
4816 0 : nodes.sort_by_key(|n| n.node_id);
4817 0 :
4818 0 : if nodes != expect_nodes {
4819 0 : tracing::error!("Consistency check failed on nodes.");
4820 0 : tracing::error!(
4821 0 : "Nodes in memory: {}",
4822 0 : serde_json::to_string(&expect_nodes)
4823 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4824 : );
4825 0 : tracing::error!(
4826 0 : "Nodes in database: {}",
4827 0 : serde_json::to_string(&nodes)
4828 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4829 : );
4830 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4831 0 : "Node consistency failure"
4832 0 : )));
4833 0 : }
4834 :
4835 0 : let mut shards = self.persistence.list_tenant_shards().await?;
4836 0 : shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4837 0 : expect_shards.sort_by_key(|tsp| (tsp.tenant_id.clone(), tsp.shard_number, tsp.shard_count));
4838 0 :
4839 0 : if shards != expect_shards {
4840 0 : tracing::error!("Consistency check failed on shards.");
4841 0 : tracing::error!(
4842 0 : "Shards in memory: {}",
4843 0 : serde_json::to_string(&expect_shards)
4844 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4845 : );
4846 0 : tracing::error!(
4847 0 : "Shards in database: {}",
4848 0 : serde_json::to_string(&shards)
4849 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4850 : );
4851 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
4852 0 : "Shard consistency failure"
4853 0 : )));
4854 0 : }
4855 0 :
4856 0 : Ok(())
4857 0 : }
4858 :
4859 : /// For debug/support: a JSON dump of the [`Scheduler`]. Returns a response so that
4860 : /// we don't have to make TenantShard clonable in the return path.
4861 0 : pub(crate) fn scheduler_dump(&self) -> Result<hyper::Response<hyper::Body>, ApiError> {
4862 0 : let serialized = {
4863 0 : let locked = self.inner.read().unwrap();
4864 0 : serde_json::to_string(&locked.scheduler)
4865 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?
4866 : };
4867 :
4868 0 : hyper::Response::builder()
4869 0 : .status(hyper::StatusCode::OK)
4870 0 : .header(hyper::header::CONTENT_TYPE, "application/json")
4871 0 : .body(hyper::Body::from(serialized))
4872 0 : .map_err(|e| ApiError::InternalServerError(e.into()))
4873 0 : }
4874 :
4875 : /// This is for debug/support only: we simply drop all state for a tenant, without
4876 : /// detaching or deleting it on pageservers. We do not try and re-schedule any
4877 : /// tenants that were on this node.
4878 0 : pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
4879 0 : self.persistence.delete_node(node_id).await?;
4880 :
4881 0 : let mut locked = self.inner.write().unwrap();
4882 :
4883 0 : for shard in locked.tenants.values_mut() {
4884 0 : shard.deref_node(node_id);
4885 0 : shard.observed.locations.remove(&node_id);
4886 0 : }
4887 :
4888 0 : let mut nodes = (*locked.nodes).clone();
4889 0 : nodes.remove(&node_id);
4890 0 : locked.nodes = Arc::new(nodes);
4891 0 :
4892 0 : locked.scheduler.node_remove(node_id);
4893 0 :
4894 0 : Ok(())
4895 0 : }
4896 :
4897 : /// If a node has any work on it, it will be rescheduled: this is "clean" in the sense
4898 : /// that we don't leave any bad state behind in the storage controller, but unclean
4899 : /// in the sense that we are not carefully draining the node.
4900 0 : pub(crate) async fn node_delete(&self, node_id: NodeId) -> Result<(), ApiError> {
4901 0 : let _node_lock =
4902 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Delete).await;
4903 :
4904 : // 1. Atomically update in-memory state:
4905 : // - set the scheduling state to Pause to make subsequent scheduling ops skip it
4906 : // - update shards' intents to exclude the node, and reschedule any shards whose intents we modified.
4907 : // - drop the node from the main nodes map, so that when running reconciles complete they do not
4908 : // re-insert references to this node into the ObservedState of shards
4909 : // - drop the node from the scheduler
4910 : {
4911 0 : let mut locked = self.inner.write().unwrap();
4912 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
4913 0 :
4914 0 : {
4915 0 : let mut nodes_mut = (*nodes).deref().clone();
4916 0 : match nodes_mut.get_mut(&node_id) {
4917 0 : Some(node) => {
4918 0 : // We do not bother setting this in the database, because we're about to delete the row anyway, and
4919 0 : // if we crash it would not be desirable to leave the node paused after a restart.
4920 0 : node.set_scheduling(NodeSchedulingPolicy::Pause);
4921 0 : }
4922 : None => {
4923 0 : tracing::info!(
4924 0 : "Node not found: presuming this is a retry and returning success"
4925 : );
4926 0 : return Ok(());
4927 : }
4928 : }
4929 :
4930 0 : *nodes = Arc::new(nodes_mut);
4931 : }
4932 :
4933 0 : for (tenant_shard_id, shard) in tenants {
4934 0 : if shard.deref_node(node_id) {
4935 : // FIXME: we need to build a ScheduleContext that reflects this shard's peers, otherwise
4936 : // it won't properly do anti-affinity.
4937 0 : let mut schedule_context = ScheduleContext::default();
4938 :
4939 0 : if let Err(e) = shard.schedule(scheduler, &mut schedule_context) {
4940 : // TODO: implement force flag to remove a node even if we can't reschedule
4941 : // a tenant
4942 0 : tracing::error!("Refusing to delete node, shard {tenant_shard_id} can't be rescheduled: {e}");
4943 0 : return Err(e.into());
4944 : } else {
4945 0 : tracing::info!(
4946 0 : "Rescheduled shard {tenant_shard_id} away from node during deletion"
4947 : )
4948 : }
4949 :
4950 0 : self.maybe_reconcile_shard(shard, nodes);
4951 0 : }
4952 :
4953 : // Here we remove an existing observed location for the node we're removing, and it will
4954 : // not be re-added by a reconciler's completion because we filter out removed nodes in
4955 : // process_result.
4956 : //
4957 : // Note that we update the shard's observed state _after_ calling maybe_reconcile_shard: that
4958 : // means any reconciles we spawned will know about the node we're deleting, enabling them
4959 : // to do live migrations if it's still online.
4960 0 : shard.observed.locations.remove(&node_id);
4961 : }
4962 :
4963 0 : scheduler.node_remove(node_id);
4964 0 :
4965 0 : {
4966 0 : let mut nodes_mut = (**nodes).clone();
4967 0 : nodes_mut.remove(&node_id);
4968 0 : *nodes = Arc::new(nodes_mut);
4969 0 : }
4970 0 : }
4971 0 :
4972 0 : // Note: some `generation_pageserver` columns on tenant shards in the database may still refer to
4973 0 : // the removed node, as this column means "The pageserver to which this generation was issued", and
4974 0 : // their generations won't get updated until the reconcilers moving them away from this node complete.
4975 0 : // That is safe because in Service::spawn we only use generation_pageserver if it refers to a node
4976 0 : // that exists.
4977 0 :
4978 0 : // 2. Actually delete the node from the database and from in-memory state
4979 0 : tracing::info!("Deleting node from database");
4980 0 : self.persistence.delete_node(node_id).await?;
4981 :
4982 0 : Ok(())
4983 0 : }
4984 :
4985 0 : pub(crate) async fn node_list(&self) -> Result<Vec<Node>, ApiError> {
4986 0 : let nodes = {
4987 0 : self.inner
4988 0 : .read()
4989 0 : .unwrap()
4990 0 : .nodes
4991 0 : .values()
4992 0 : .cloned()
4993 0 : .collect::<Vec<_>>()
4994 0 : };
4995 0 :
4996 0 : Ok(nodes)
4997 0 : }
4998 :
4999 0 : pub(crate) async fn get_node(&self, node_id: NodeId) -> Result<Node, ApiError> {
5000 0 : self.inner
5001 0 : .read()
5002 0 : .unwrap()
5003 0 : .nodes
5004 0 : .get(&node_id)
5005 0 : .cloned()
5006 0 : .ok_or(ApiError::NotFound(
5007 0 : format!("Node {node_id} not registered").into(),
5008 0 : ))
5009 0 : }
5010 :
5011 0 : pub(crate) async fn get_node_shards(
5012 0 : &self,
5013 0 : node_id: NodeId,
5014 0 : ) -> Result<NodeShardResponse, ApiError> {
5015 0 : let locked = self.inner.read().unwrap();
5016 0 : let mut shards = Vec::new();
5017 0 : for (tid, tenant) in locked.tenants.iter() {
5018 0 : let is_intended_secondary = match (
5019 0 : tenant.intent.get_attached() == &Some(node_id),
5020 0 : tenant.intent.get_secondary().contains(&node_id),
5021 0 : ) {
5022 : (true, true) => {
5023 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
5024 0 : "{} attached as primary+secondary on the same node",
5025 0 : tid
5026 0 : )))
5027 : }
5028 0 : (true, false) => Some(false),
5029 0 : (false, true) => Some(true),
5030 0 : (false, false) => None,
5031 : };
5032 0 : let is_observed_secondary = if let Some(ObservedStateLocation { conf: Some(conf) }) =
5033 0 : tenant.observed.locations.get(&node_id)
5034 : {
5035 0 : Some(conf.secondary_conf.is_some())
5036 : } else {
5037 0 : None
5038 : };
5039 0 : if is_intended_secondary.is_some() || is_observed_secondary.is_some() {
5040 0 : shards.push(NodeShard {
5041 0 : tenant_shard_id: *tid,
5042 0 : is_intended_secondary,
5043 0 : is_observed_secondary,
5044 0 : });
5045 0 : }
5046 : }
5047 0 : Ok(NodeShardResponse { node_id, shards })
5048 0 : }
5049 :
5050 0 : pub(crate) async fn get_leader(&self) -> DatabaseResult<Option<ControllerPersistence>> {
5051 0 : self.persistence.get_leader().await
5052 0 : }
5053 :
5054 0 : pub(crate) async fn node_register(
5055 0 : &self,
5056 0 : register_req: NodeRegisterRequest,
5057 0 : ) -> Result<(), ApiError> {
5058 0 : let _node_lock = trace_exclusive_lock(
5059 0 : &self.node_op_locks,
5060 0 : register_req.node_id,
5061 0 : NodeOperations::Register,
5062 0 : )
5063 0 : .await;
5064 :
5065 : enum RegistrationStatus {
5066 : Matched,
5067 : Mismatched,
5068 : New,
5069 : }
5070 :
5071 0 : let registration_status = {
5072 0 : let locked = self.inner.read().unwrap();
5073 0 : if let Some(node) = locked.nodes.get(®ister_req.node_id) {
5074 0 : if node.registration_match(®ister_req) {
5075 0 : RegistrationStatus::Matched
5076 : } else {
5077 0 : RegistrationStatus::Mismatched
5078 : }
5079 : } else {
5080 0 : RegistrationStatus::New
5081 : }
5082 : };
5083 :
5084 0 : match registration_status {
5085 : RegistrationStatus::Matched => {
5086 0 : tracing::info!(
5087 0 : "Node {} re-registered with matching address",
5088 : register_req.node_id
5089 : );
5090 :
5091 0 : return Ok(());
5092 : }
5093 : RegistrationStatus::Mismatched => {
5094 : // TODO: decide if we want to allow modifying node addresses without removing and re-adding
5095 : // the node. Safest/simplest thing is to refuse it, and usually we deploy with
5096 : // a fixed address through the lifetime of a node.
5097 0 : tracing::warn!(
5098 0 : "Node {} tried to register with different address",
5099 : register_req.node_id
5100 : );
5101 0 : return Err(ApiError::Conflict(
5102 0 : "Node is already registered with different address".to_string(),
5103 0 : ));
5104 : }
5105 0 : RegistrationStatus::New => {
5106 0 : // fallthrough
5107 0 : }
5108 0 : }
5109 0 :
5110 0 : // We do not require that a node is actually online when registered (it will start life
5111 0 : // with it's availability set to Offline), but we _do_ require that its DNS record exists. We're
5112 0 : // therefore not immune to asymmetric L3 connectivity issues, but we are protected against nodes
5113 0 : // that register themselves with a broken DNS config. We check only the HTTP hostname, because
5114 0 : // the postgres hostname might only be resolvable to clients (e.g. if we're on a different VPC than clients).
5115 0 : if tokio::net::lookup_host(format!(
5116 0 : "{}:{}",
5117 0 : register_req.listen_http_addr, register_req.listen_http_port
5118 0 : ))
5119 0 : .await
5120 0 : .is_err()
5121 : {
5122 : // If we have a transient DNS issue, it's up to the caller to retry their registration. Because
5123 : // we can't robustly distinguish between an intermittent issue and a totally bogus DNS situation,
5124 : // we return a soft 503 error, to encourage callers to retry past transient issues.
5125 0 : return Err(ApiError::ResourceUnavailable(
5126 0 : format!(
5127 0 : "Node {} tried to register with unknown DNS name '{}'",
5128 0 : register_req.node_id, register_req.listen_http_addr
5129 0 : )
5130 0 : .into(),
5131 0 : ));
5132 0 : }
5133 0 :
5134 0 : // Ordering: we must persist the new node _before_ adding it to in-memory state.
5135 0 : // This ensures that before we use it for anything or expose it via any external
5136 0 : // API, it is guaranteed to be available after a restart.
5137 0 : let new_node = Node::new(
5138 0 : register_req.node_id,
5139 0 : register_req.listen_http_addr,
5140 0 : register_req.listen_http_port,
5141 0 : register_req.listen_pg_addr,
5142 0 : register_req.listen_pg_port,
5143 0 : register_req.availability_zone_id,
5144 0 : );
5145 0 :
5146 0 : // TODO: idempotency if the node already exists in the database
5147 0 : self.persistence.insert_node(&new_node).await?;
5148 :
5149 0 : let mut locked = self.inner.write().unwrap();
5150 0 : let mut new_nodes = (*locked.nodes).clone();
5151 0 :
5152 0 : locked.scheduler.node_upsert(&new_node);
5153 0 : new_nodes.insert(register_req.node_id, new_node);
5154 0 :
5155 0 : locked.nodes = Arc::new(new_nodes);
5156 0 :
5157 0 : tracing::info!(
5158 0 : "Registered pageserver {}, now have {} pageservers",
5159 0 : register_req.node_id,
5160 0 : locked.nodes.len()
5161 : );
5162 0 : Ok(())
5163 0 : }
5164 :
5165 0 : pub(crate) async fn node_configure(
5166 0 : &self,
5167 0 : node_id: NodeId,
5168 0 : availability: Option<NodeAvailability>,
5169 0 : scheduling: Option<NodeSchedulingPolicy>,
5170 0 : ) -> Result<(), ApiError> {
5171 0 : let _node_lock =
5172 0 : trace_exclusive_lock(&self.node_op_locks, node_id, NodeOperations::Configure).await;
5173 :
5174 0 : if let Some(scheduling) = scheduling {
5175 : // Scheduling is a persistent part of Node: we must write updates to the database before
5176 : // applying them in memory
5177 0 : self.persistence.update_node(node_id, scheduling).await?;
5178 0 : }
5179 :
5180 : // If we're activating a node, then before setting it active we must reconcile any shard locations
5181 : // on that node, in case it is out of sync, e.g. due to being unavailable during controller startup,
5182 : // by calling [`Self::node_activate_reconcile`]
5183 : //
5184 : // The transition we calculate here remains valid later in the function because we hold the op lock on the node:
5185 : // nothing else can mutate its availability while we run.
5186 0 : let availability_transition = if let Some(input_availability) = availability.as_ref() {
5187 0 : let (activate_node, availability_transition) = {
5188 0 : let locked = self.inner.read().unwrap();
5189 0 : let Some(node) = locked.nodes.get(&node_id) else {
5190 0 : return Err(ApiError::NotFound(
5191 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5192 0 : ));
5193 : };
5194 :
5195 0 : (
5196 0 : node.clone(),
5197 0 : node.get_availability_transition(input_availability),
5198 0 : )
5199 : };
5200 :
5201 0 : if matches!(availability_transition, AvailabilityTransition::ToActive) {
5202 0 : self.node_activate_reconcile(activate_node, &_node_lock)
5203 0 : .await?;
5204 0 : }
5205 0 : availability_transition
5206 : } else {
5207 0 : AvailabilityTransition::Unchanged
5208 : };
5209 :
5210 : // Apply changes from the request to our in-memory state for the Node
5211 0 : let mut locked = self.inner.write().unwrap();
5212 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5213 0 :
5214 0 : let mut new_nodes = (**nodes).clone();
5215 :
5216 0 : let Some(node) = new_nodes.get_mut(&node_id) else {
5217 0 : return Err(ApiError::NotFound(
5218 0 : anyhow::anyhow!("Node not registered").into(),
5219 0 : ));
5220 : };
5221 :
5222 0 : if let Some(availability) = availability.as_ref() {
5223 0 : node.set_availability(availability.clone());
5224 0 : }
5225 :
5226 0 : if let Some(scheduling) = scheduling {
5227 0 : node.set_scheduling(scheduling);
5228 0 : }
5229 :
5230 : // Update the scheduler, in case the elegibility of the node for new shards has changed
5231 0 : scheduler.node_upsert(node);
5232 0 :
5233 0 : let new_nodes = Arc::new(new_nodes);
5234 0 :
5235 0 : // Modify scheduling state for any Tenants that are affected by a change in the node's availability state.
5236 0 : match availability_transition {
5237 : AvailabilityTransition::ToOffline => {
5238 0 : tracing::info!("Node {} transition to offline", node_id);
5239 0 : let mut tenants_affected: usize = 0;
5240 :
5241 0 : for (tenant_shard_id, tenant_shard) in tenants {
5242 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
5243 0 : // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
5244 0 : // not assume our knowledge of the node's configuration is accurate until it comes back online
5245 0 : observed_loc.conf = None;
5246 0 : }
5247 :
5248 0 : if new_nodes.len() == 1 {
5249 : // Special case for single-node cluster: there is no point trying to reschedule
5250 : // any tenant shards: avoid doing so, in order to avoid spewing warnings about
5251 : // failures to schedule them.
5252 0 : continue;
5253 0 : }
5254 0 :
5255 0 : if !new_nodes
5256 0 : .values()
5257 0 : .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
5258 : {
5259 : // Special case for when all nodes are unavailable and/or unschedulable: there is no point
5260 : // trying to reschedule since there's nowhere else to go. Without this
5261 : // branch we incorrectly detach tenants in response to node unavailability.
5262 0 : continue;
5263 0 : }
5264 0 :
5265 0 : if tenant_shard.intent.demote_attached(scheduler, node_id) {
5266 0 : tenant_shard.sequence = tenant_shard.sequence.next();
5267 0 :
5268 0 : // TODO: populate a ScheduleContext including all shards in the same tenant_id (only matters
5269 0 : // for tenants without secondary locations: if they have a secondary location, then this
5270 0 : // schedule() call is just promoting an existing secondary)
5271 0 : let mut schedule_context = ScheduleContext::default();
5272 0 :
5273 0 : match tenant_shard.schedule(scheduler, &mut schedule_context) {
5274 0 : Err(e) => {
5275 0 : // It is possible that some tenants will become unschedulable when too many pageservers
5276 0 : // go offline: in this case there isn't much we can do other than make the issue observable.
5277 0 : // TODO: give TenantShard a scheduling error attribute to be queried later.
5278 0 : tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", node_id);
5279 : }
5280 : Ok(()) => {
5281 0 : if self
5282 0 : .maybe_reconcile_shard(tenant_shard, &new_nodes)
5283 0 : .is_some()
5284 0 : {
5285 0 : tenants_affected += 1;
5286 0 : };
5287 : }
5288 : }
5289 0 : }
5290 : }
5291 0 : tracing::info!(
5292 0 : "Launched {} reconciler tasks for tenants affected by node {} going offline",
5293 : tenants_affected,
5294 : node_id
5295 : )
5296 : }
5297 : AvailabilityTransition::ToActive => {
5298 0 : tracing::info!("Node {} transition to active", node_id);
5299 : // When a node comes back online, we must reconcile any tenant that has a None observed
5300 : // location on the node.
5301 0 : for tenant_shard in locked.tenants.values_mut() {
5302 : // If a reconciliation is already in progress, rely on the previous scheduling
5303 : // decision and skip triggering a new reconciliation.
5304 0 : if tenant_shard.reconciler.is_some() {
5305 0 : continue;
5306 0 : }
5307 :
5308 0 : if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
5309 0 : if observed_loc.conf.is_none() {
5310 0 : self.maybe_reconcile_shard(tenant_shard, &new_nodes);
5311 0 : }
5312 0 : }
5313 : }
5314 :
5315 : // TODO: in the background, we should balance work back onto this pageserver
5316 : }
5317 : // No action required for the intermediate unavailable state.
5318 : // When we transition into active or offline from the unavailable state,
5319 : // the correct handling above will kick in.
5320 : AvailabilityTransition::ToWarmingUpFromActive => {
5321 0 : tracing::info!("Node {} transition to unavailable from active", node_id);
5322 : }
5323 : AvailabilityTransition::ToWarmingUpFromOffline => {
5324 0 : tracing::info!("Node {} transition to unavailable from offline", node_id);
5325 : }
5326 : AvailabilityTransition::Unchanged => {
5327 0 : tracing::debug!("Node {} no availability change during config", node_id);
5328 : }
5329 : }
5330 :
5331 0 : locked.nodes = new_nodes;
5332 0 :
5333 0 : Ok(())
5334 0 : }
5335 :
5336 : /// Wrapper around [`Self::node_configure`] which only allows changes while there is no ongoing
5337 : /// operation for HTTP api.
5338 0 : pub(crate) async fn external_node_configure(
5339 0 : &self,
5340 0 : node_id: NodeId,
5341 0 : availability: Option<NodeAvailability>,
5342 0 : scheduling: Option<NodeSchedulingPolicy>,
5343 0 : ) -> Result<(), ApiError> {
5344 0 : {
5345 0 : let locked = self.inner.read().unwrap();
5346 0 : if let Some(op) = locked.ongoing_operation.as_ref().map(|op| op.operation) {
5347 0 : return Err(ApiError::PreconditionFailed(
5348 0 : format!("Ongoing background operation forbids configuring: {op}").into(),
5349 0 : ));
5350 0 : }
5351 0 : }
5352 0 :
5353 0 : self.node_configure(node_id, availability, scheduling).await
5354 0 : }
5355 :
5356 0 : pub(crate) async fn start_node_drain(
5357 0 : self: &Arc<Self>,
5358 0 : node_id: NodeId,
5359 0 : ) -> Result<(), ApiError> {
5360 0 : let (ongoing_op, node_available, node_policy, schedulable_nodes_count) = {
5361 0 : let locked = self.inner.read().unwrap();
5362 0 : let nodes = &locked.nodes;
5363 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5364 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5365 0 : ))?;
5366 0 : let schedulable_nodes_count = nodes
5367 0 : .iter()
5368 0 : .filter(|(_, n)| matches!(n.may_schedule(), MaySchedule::Yes(_)))
5369 0 : .count();
5370 0 :
5371 0 : (
5372 0 : locked
5373 0 : .ongoing_operation
5374 0 : .as_ref()
5375 0 : .map(|ongoing| ongoing.operation),
5376 0 : node.is_available(),
5377 0 : node.get_scheduling(),
5378 0 : schedulable_nodes_count,
5379 0 : )
5380 0 : };
5381 :
5382 0 : if let Some(ongoing) = ongoing_op {
5383 0 : return Err(ApiError::PreconditionFailed(
5384 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
5385 0 : ));
5386 0 : }
5387 0 :
5388 0 : if !node_available {
5389 0 : return Err(ApiError::ResourceUnavailable(
5390 0 : format!("Node {node_id} is currently unavailable").into(),
5391 0 : ));
5392 0 : }
5393 0 :
5394 0 : if schedulable_nodes_count == 0 {
5395 0 : return Err(ApiError::PreconditionFailed(
5396 0 : "No other schedulable nodes to drain to".into(),
5397 0 : ));
5398 0 : }
5399 0 :
5400 0 : match node_policy {
5401 : NodeSchedulingPolicy::Active | NodeSchedulingPolicy::Pause => {
5402 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Draining))
5403 0 : .await?;
5404 :
5405 0 : let cancel = self.cancel.child_token();
5406 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
5407 :
5408 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
5409 0 : operation: Operation::Drain(Drain { node_id }),
5410 0 : cancel: cancel.clone(),
5411 0 : });
5412 :
5413 0 : let span = tracing::info_span!(parent: None, "drain_node", %node_id);
5414 :
5415 0 : tokio::task::spawn({
5416 0 : let service = self.clone();
5417 0 : let cancel = cancel.clone();
5418 0 : async move {
5419 0 : let _gate_guard = gate_guard;
5420 0 :
5421 0 : scopeguard::defer! {
5422 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
5423 0 :
5424 0 : if let Some(Operation::Drain(removed_drain)) = prev.map(|h| h.operation) {
5425 0 : assert_eq!(removed_drain.node_id, node_id, "We always take the same operation");
5426 0 : } else {
5427 0 : panic!("We always remove the same operation")
5428 0 : }
5429 0 : }
5430 0 :
5431 0 : tracing::info!("Drain background operation starting");
5432 0 : let res = service.drain_node(node_id, cancel).await;
5433 0 : match res {
5434 : Ok(()) => {
5435 0 : tracing::info!("Drain background operation completed successfully");
5436 : }
5437 : Err(OperationError::Cancelled) => {
5438 0 : tracing::info!("Drain background operation was cancelled");
5439 : }
5440 0 : Err(err) => {
5441 0 : tracing::error!("Drain background operation encountered: {err}")
5442 : }
5443 : }
5444 0 : }
5445 0 : }.instrument(span));
5446 0 : }
5447 : NodeSchedulingPolicy::Draining => {
5448 0 : return Err(ApiError::Conflict(format!(
5449 0 : "Node {node_id} has drain in progress"
5450 0 : )));
5451 : }
5452 0 : policy => {
5453 0 : return Err(ApiError::PreconditionFailed(
5454 0 : format!("Node {node_id} cannot be drained due to {policy:?} policy").into(),
5455 0 : ));
5456 : }
5457 : }
5458 :
5459 0 : Ok(())
5460 0 : }
5461 :
5462 0 : pub(crate) async fn cancel_node_drain(&self, node_id: NodeId) -> Result<(), ApiError> {
5463 0 : let node_available = {
5464 0 : let locked = self.inner.read().unwrap();
5465 0 : let nodes = &locked.nodes;
5466 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5467 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5468 0 : ))?;
5469 :
5470 0 : node.is_available()
5471 0 : };
5472 0 :
5473 0 : if !node_available {
5474 0 : return Err(ApiError::ResourceUnavailable(
5475 0 : format!("Node {node_id} is currently unavailable").into(),
5476 0 : ));
5477 0 : }
5478 :
5479 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
5480 0 : if let Operation::Drain(drain) = op_handler.operation {
5481 0 : if drain.node_id == node_id {
5482 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
5483 0 : op_handler.cancel.cancel();
5484 0 : return Ok(());
5485 0 : }
5486 0 : }
5487 0 : }
5488 :
5489 0 : Err(ApiError::PreconditionFailed(
5490 0 : format!("Node {node_id} has no drain in progress").into(),
5491 0 : ))
5492 0 : }
5493 :
5494 0 : pub(crate) async fn start_node_fill(self: &Arc<Self>, node_id: NodeId) -> Result<(), ApiError> {
5495 0 : let (ongoing_op, node_available, node_policy, total_nodes_count) = {
5496 0 : let locked = self.inner.read().unwrap();
5497 0 : let nodes = &locked.nodes;
5498 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5499 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5500 0 : ))?;
5501 :
5502 0 : (
5503 0 : locked
5504 0 : .ongoing_operation
5505 0 : .as_ref()
5506 0 : .map(|ongoing| ongoing.operation),
5507 0 : node.is_available(),
5508 0 : node.get_scheduling(),
5509 0 : nodes.len(),
5510 0 : )
5511 0 : };
5512 :
5513 0 : if let Some(ongoing) = ongoing_op {
5514 0 : return Err(ApiError::PreconditionFailed(
5515 0 : format!("Background operation already ongoing for node: {}", ongoing).into(),
5516 0 : ));
5517 0 : }
5518 0 :
5519 0 : if !node_available {
5520 0 : return Err(ApiError::ResourceUnavailable(
5521 0 : format!("Node {node_id} is currently unavailable").into(),
5522 0 : ));
5523 0 : }
5524 0 :
5525 0 : if total_nodes_count <= 1 {
5526 0 : return Err(ApiError::PreconditionFailed(
5527 0 : "No other nodes to fill from".into(),
5528 0 : ));
5529 0 : }
5530 0 :
5531 0 : match node_policy {
5532 : NodeSchedulingPolicy::Active => {
5533 0 : self.node_configure(node_id, None, Some(NodeSchedulingPolicy::Filling))
5534 0 : .await?;
5535 :
5536 0 : let cancel = self.cancel.child_token();
5537 0 : let gate_guard = self.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
5538 :
5539 0 : self.inner.write().unwrap().ongoing_operation = Some(OperationHandler {
5540 0 : operation: Operation::Fill(Fill { node_id }),
5541 0 : cancel: cancel.clone(),
5542 0 : });
5543 :
5544 0 : let span = tracing::info_span!(parent: None, "fill_node", %node_id);
5545 :
5546 0 : tokio::task::spawn({
5547 0 : let service = self.clone();
5548 0 : let cancel = cancel.clone();
5549 0 : async move {
5550 0 : let _gate_guard = gate_guard;
5551 0 :
5552 0 : scopeguard::defer! {
5553 0 : let prev = service.inner.write().unwrap().ongoing_operation.take();
5554 0 :
5555 0 : if let Some(Operation::Fill(removed_fill)) = prev.map(|h| h.operation) {
5556 0 : assert_eq!(removed_fill.node_id, node_id, "We always take the same operation");
5557 0 : } else {
5558 0 : panic!("We always remove the same operation")
5559 0 : }
5560 0 : }
5561 0 :
5562 0 : tracing::info!("Fill background operation starting");
5563 0 : let res = service.fill_node(node_id, cancel).await;
5564 0 : match res {
5565 : Ok(()) => {
5566 0 : tracing::info!("Fill background operation completed successfully");
5567 : }
5568 : Err(OperationError::Cancelled) => {
5569 0 : tracing::info!("Fill background operation was cancelled");
5570 : }
5571 0 : Err(err) => {
5572 0 : tracing::error!("Fill background operation encountered: {err}")
5573 : }
5574 : }
5575 0 : }
5576 0 : }.instrument(span));
5577 0 : }
5578 : NodeSchedulingPolicy::Filling => {
5579 0 : return Err(ApiError::Conflict(format!(
5580 0 : "Node {node_id} has fill in progress"
5581 0 : )));
5582 : }
5583 0 : policy => {
5584 0 : return Err(ApiError::PreconditionFailed(
5585 0 : format!("Node {node_id} cannot be filled due to {policy:?} policy").into(),
5586 0 : ));
5587 : }
5588 : }
5589 :
5590 0 : Ok(())
5591 0 : }
5592 :
5593 0 : pub(crate) async fn cancel_node_fill(&self, node_id: NodeId) -> Result<(), ApiError> {
5594 0 : let node_available = {
5595 0 : let locked = self.inner.read().unwrap();
5596 0 : let nodes = &locked.nodes;
5597 0 : let node = nodes.get(&node_id).ok_or(ApiError::NotFound(
5598 0 : anyhow::anyhow!("Node {} not registered", node_id).into(),
5599 0 : ))?;
5600 :
5601 0 : node.is_available()
5602 0 : };
5603 0 :
5604 0 : if !node_available {
5605 0 : return Err(ApiError::ResourceUnavailable(
5606 0 : format!("Node {node_id} is currently unavailable").into(),
5607 0 : ));
5608 0 : }
5609 :
5610 0 : if let Some(op_handler) = self.inner.read().unwrap().ongoing_operation.as_ref() {
5611 0 : if let Operation::Fill(fill) = op_handler.operation {
5612 0 : if fill.node_id == node_id {
5613 0 : tracing::info!("Cancelling background drain operation for node {node_id}");
5614 0 : op_handler.cancel.cancel();
5615 0 : return Ok(());
5616 0 : }
5617 0 : }
5618 0 : }
5619 :
5620 0 : Err(ApiError::PreconditionFailed(
5621 0 : format!("Node {node_id} has no fill in progress").into(),
5622 0 : ))
5623 0 : }
5624 :
5625 : /// Like [`Self::maybe_configured_reconcile_shard`], but uses the default reconciler
5626 : /// configuration
5627 0 : fn maybe_reconcile_shard(
5628 0 : &self,
5629 0 : shard: &mut TenantShard,
5630 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5631 0 : ) -> Option<ReconcilerWaiter> {
5632 0 : self.maybe_configured_reconcile_shard(shard, nodes, ReconcilerConfig::default())
5633 0 : }
5634 :
5635 : /// Wrap [`TenantShard`] reconciliation methods with acquisition of [`Gate`] and [`ReconcileUnits`],
5636 0 : fn maybe_configured_reconcile_shard(
5637 0 : &self,
5638 0 : shard: &mut TenantShard,
5639 0 : nodes: &Arc<HashMap<NodeId, Node>>,
5640 0 : reconciler_config: ReconcilerConfig,
5641 0 : ) -> Option<ReconcilerWaiter> {
5642 0 : let reconcile_needed = shard.get_reconcile_needed(nodes);
5643 0 :
5644 0 : match reconcile_needed {
5645 0 : ReconcileNeeded::No => return None,
5646 0 : ReconcileNeeded::WaitExisting(waiter) => return Some(waiter),
5647 0 : ReconcileNeeded::Yes => {
5648 0 : // Fall through to try and acquire units for spawning reconciler
5649 0 : }
5650 : };
5651 :
5652 0 : let units = match self.reconciler_concurrency.clone().try_acquire_owned() {
5653 0 : Ok(u) => ReconcileUnits::new(u),
5654 : Err(_) => {
5655 0 : tracing::info!(tenant_id=%shard.tenant_shard_id.tenant_id, shard_id=%shard.tenant_shard_id.shard_slug(),
5656 0 : "Concurrency limited: enqueued for reconcile later");
5657 0 : if !shard.delayed_reconcile {
5658 0 : match self.delayed_reconcile_tx.try_send(shard.tenant_shard_id) {
5659 0 : Err(TrySendError::Closed(_)) => {
5660 0 : // Weird mid-shutdown case?
5661 0 : }
5662 : Err(TrySendError::Full(_)) => {
5663 : // It is safe to skip sending our ID in the channel: we will eventually get retried by the background reconcile task.
5664 0 : tracing::warn!(
5665 0 : "Many shards are waiting to reconcile: delayed_reconcile queue is full"
5666 : );
5667 : }
5668 0 : Ok(()) => {
5669 0 : shard.delayed_reconcile = true;
5670 0 : }
5671 : }
5672 0 : }
5673 :
5674 : // We won't spawn a reconciler, but we will construct a waiter that waits for the shard's sequence
5675 : // number to advance. When this function is eventually called again and succeeds in getting units,
5676 : // it will spawn a reconciler that makes this waiter complete.
5677 0 : return Some(shard.future_reconcile_waiter());
5678 : }
5679 : };
5680 :
5681 0 : let Ok(gate_guard) = self.reconcilers_gate.enter() else {
5682 : // Gate closed: we're shutting down, drop out.
5683 0 : return None;
5684 : };
5685 :
5686 0 : shard.spawn_reconciler(
5687 0 : &self.result_tx,
5688 0 : nodes,
5689 0 : &self.compute_hook,
5690 0 : reconciler_config,
5691 0 : &self.config,
5692 0 : &self.persistence,
5693 0 : units,
5694 0 : gate_guard,
5695 0 : &self.reconcilers_cancel,
5696 0 : )
5697 0 : }
5698 :
5699 : /// Check all tenants for pending reconciliation work, and reconcile those in need.
5700 : /// Additionally, reschedule tenants that require it.
5701 : ///
5702 : /// Returns how many reconciliation tasks were started, or `1` if no reconciles were
5703 : /// spawned but some _would_ have been spawned if `reconciler_concurrency` units where
5704 : /// available. A return value of 0 indicates that everything is fully reconciled already.
5705 0 : fn reconcile_all(&self) -> usize {
5706 0 : let mut locked = self.inner.write().unwrap();
5707 0 : let (nodes, tenants, _scheduler) = locked.parts_mut();
5708 0 : let pageservers = nodes.clone();
5709 0 :
5710 0 : let mut schedule_context = ScheduleContext::default();
5711 0 :
5712 0 : let mut reconciles_spawned = 0;
5713 0 : for (tenant_shard_id, shard) in tenants.iter_mut() {
5714 0 : if tenant_shard_id.is_shard_zero() {
5715 0 : schedule_context = ScheduleContext::default();
5716 0 : }
5717 :
5718 : // Skip checking if this shard is already enqueued for reconciliation
5719 0 : if shard.delayed_reconcile && self.reconciler_concurrency.available_permits() == 0 {
5720 : // If there is something delayed, then return a nonzero count so that
5721 : // callers like reconcile_all_now do not incorrectly get the impression
5722 : // that the system is in a quiescent state.
5723 0 : reconciles_spawned = std::cmp::max(1, reconciles_spawned);
5724 0 : continue;
5725 0 : }
5726 0 :
5727 0 : // Eventual consistency: if an earlier reconcile job failed, and the shard is still
5728 0 : // dirty, spawn another rone
5729 0 : if self.maybe_reconcile_shard(shard, &pageservers).is_some() {
5730 0 : reconciles_spawned += 1;
5731 0 : }
5732 :
5733 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5734 : }
5735 :
5736 0 : reconciles_spawned
5737 0 : }
5738 :
5739 : /// `optimize` in this context means identifying shards which have valid scheduled locations, but
5740 : /// could be scheduled somewhere better:
5741 : /// - Cutting over to a secondary if the node with the secondary is more lightly loaded
5742 : /// * e.g. after a node fails then recovers, to move some work back to it
5743 : /// - Cutting over to a secondary if it improves the spread of shard attachments within a tenant
5744 : /// * e.g. after a shard split, the initial attached locations will all be on the node where
5745 : /// we did the split, but are probably better placed elsewhere.
5746 : /// - Creating new secondary locations if it improves the spreading of a sharded tenant
5747 : /// * e.g. after a shard split, some locations will be on the same node (where the split
5748 : /// happened), and will probably be better placed elsewhere.
5749 : ///
5750 : /// To put it more briefly: whereas the scheduler respects soft constraints in a ScheduleContext at
5751 : /// the time of scheduling, this function looks for cases where a better-scoring location is available
5752 : /// according to those same soft constraints.
5753 0 : async fn optimize_all(&self) -> usize {
5754 : // Limit on how many shards' optmizations each call to this function will execute. Combined
5755 : // with the frequency of background calls, this acts as an implicit rate limit that runs a small
5756 : // trickle of optimizations in the background, rather than executing a large number in parallel
5757 : // when a change occurs.
5758 : const MAX_OPTIMIZATIONS_EXEC_PER_PASS: usize = 2;
5759 :
5760 : // Synchronous prepare: scan shards for possible scheduling optimizations
5761 0 : let candidate_work = self.optimize_all_plan();
5762 0 : let candidate_work_len = candidate_work.len();
5763 :
5764 : // Asynchronous validate: I/O to pageservers to make sure shards are in a good state to apply validation
5765 0 : let validated_work = self.optimize_all_validate(candidate_work).await;
5766 :
5767 0 : let was_work_filtered = validated_work.len() != candidate_work_len;
5768 0 :
5769 0 : // Synchronous apply: update the shards' intent states according to validated optimisations
5770 0 : let mut reconciles_spawned = 0;
5771 0 : let mut optimizations_applied = 0;
5772 0 : let mut locked = self.inner.write().unwrap();
5773 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5774 0 : for (tenant_shard_id, optimization) in validated_work {
5775 0 : let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
5776 : // Shard was dropped between planning and execution;
5777 0 : continue;
5778 : };
5779 0 : if shard.apply_optimization(scheduler, optimization) {
5780 0 : optimizations_applied += 1;
5781 0 : if self.maybe_reconcile_shard(shard, nodes).is_some() {
5782 0 : reconciles_spawned += 1;
5783 0 : }
5784 0 : }
5785 :
5786 0 : if optimizations_applied >= MAX_OPTIMIZATIONS_EXEC_PER_PASS {
5787 0 : break;
5788 0 : }
5789 : }
5790 :
5791 0 : if was_work_filtered {
5792 0 : // If we filtered any work out during validation, ensure we return a nonzero value to indicate
5793 0 : // to callers that the system is not in a truly quiet state, it's going to do some work as soon
5794 0 : // as these validations start passing.
5795 0 : reconciles_spawned = std::cmp::max(reconciles_spawned, 1);
5796 0 : }
5797 :
5798 0 : reconciles_spawned
5799 0 : }
5800 :
5801 0 : fn optimize_all_plan(&self) -> Vec<(TenantShardId, ScheduleOptimization)> {
5802 0 : let mut schedule_context = ScheduleContext::default();
5803 0 :
5804 0 : let mut tenant_shards: Vec<&TenantShard> = Vec::new();
5805 :
5806 : // How many candidate optimizations we will generate, before evaluating them for readniess: setting
5807 : // this higher than the execution limit gives us a chance to execute some work even if the first
5808 : // few optimizations we find are not ready.
5809 : const MAX_OPTIMIZATIONS_PLAN_PER_PASS: usize = 8;
5810 :
5811 0 : let mut work = Vec::new();
5812 0 :
5813 0 : let mut locked = self.inner.write().unwrap();
5814 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
5815 0 : for (tenant_shard_id, shard) in tenants.iter() {
5816 0 : if tenant_shard_id.is_shard_zero() {
5817 0 : // Reset accumulators on the first shard in a tenant
5818 0 : schedule_context = ScheduleContext::default();
5819 0 : schedule_context.mode = ScheduleMode::Speculative;
5820 0 : tenant_shards.clear();
5821 0 : }
5822 :
5823 0 : if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
5824 0 : break;
5825 0 : }
5826 0 :
5827 0 : match shard.get_scheduling_policy() {
5828 0 : ShardSchedulingPolicy::Active => {
5829 0 : // Ok to do optimization
5830 0 : }
5831 : ShardSchedulingPolicy::Essential
5832 : | ShardSchedulingPolicy::Pause
5833 : | ShardSchedulingPolicy::Stop => {
5834 : // Policy prevents optimizing this shard.
5835 0 : continue;
5836 : }
5837 : }
5838 :
5839 : // Accumulate the schedule context for all the shards in a tenant: we must have
5840 : // the total view of all shards before we can try to optimize any of them.
5841 0 : schedule_context.avoid(&shard.intent.all_pageservers());
5842 0 : if let Some(attached) = shard.intent.get_attached() {
5843 0 : schedule_context.push_attached(*attached);
5844 0 : }
5845 0 : tenant_shards.push(shard);
5846 0 :
5847 0 : // Once we have seen the last shard in the tenant, proceed to search across all shards
5848 0 : // in the tenant for optimizations
5849 0 : if shard.shard.number.0 == shard.shard.count.count() - 1 {
5850 0 : if tenant_shards.iter().any(|s| s.reconciler.is_some()) {
5851 : // Do not start any optimizations while another change to the tenant is ongoing: this
5852 : // is not necessary for correctness, but simplifies operations and implicitly throttles
5853 : // optimization changes to happen in a "trickle" over time.
5854 0 : continue;
5855 0 : }
5856 0 :
5857 0 : if tenant_shards.iter().any(|s| {
5858 0 : !matches!(s.splitting, SplitState::Idle)
5859 0 : || matches!(s.policy, PlacementPolicy::Detached)
5860 0 : }) {
5861 : // Never attempt to optimize a tenant that is currently being split, or
5862 : // a tenant that is meant to be detached
5863 0 : continue;
5864 0 : }
5865 :
5866 : // TODO: optimization calculations are relatively expensive: create some fast-path for
5867 : // the common idle case (avoiding the search on tenants that we have recently checked)
5868 :
5869 0 : for shard in &tenant_shards {
5870 0 : if let Some(optimization) =
5871 : // If idle, maybe ptimize attachments: if a shard has a secondary location that is preferable to
5872 : // its primary location based on soft constraints, cut it over.
5873 0 : shard.optimize_attachment(nodes, &schedule_context)
5874 : {
5875 0 : work.push((shard.tenant_shard_id, optimization));
5876 0 : break;
5877 0 : } else if let Some(optimization) =
5878 : // If idle, maybe optimize secondary locations: if a shard has a secondary location that would be
5879 : // better placed on another node, based on ScheduleContext, then adjust it. This
5880 : // covers cases like after a shard split, where we might have too many shards
5881 : // in the same tenant with secondary locations on the node where they originally split.
5882 0 : shard.optimize_secondary(scheduler, &schedule_context)
5883 : {
5884 0 : work.push((shard.tenant_shard_id, optimization));
5885 0 : break;
5886 0 : }
5887 :
5888 : // TODO: extend this mechanism to prefer attaching on nodes with fewer attached
5889 : // tenants (i.e. extend schedule state to distinguish attached from secondary counts),
5890 : // for the total number of attachments on a node (not just within a tenant.)
5891 : }
5892 0 : }
5893 : }
5894 :
5895 0 : work
5896 0 : }
5897 :
5898 0 : async fn optimize_all_validate(
5899 0 : &self,
5900 0 : candidate_work: Vec<(TenantShardId, ScheduleOptimization)>,
5901 0 : ) -> Vec<(TenantShardId, ScheduleOptimization)> {
5902 0 : // Take a clone of the node map to use outside the lock in async validation phase
5903 0 : let validation_nodes = { self.inner.read().unwrap().nodes.clone() };
5904 0 :
5905 0 : let mut want_secondary_status = Vec::new();
5906 0 :
5907 0 : // Validate our plans: this is an async phase where we may do I/O to pageservers to
5908 0 : // check that the state of locations is acceptable to run the optimization, such as
5909 0 : // checking that a secondary location is sufficiently warmed-up to cleanly cut over
5910 0 : // in a live migration.
5911 0 : let mut validated_work = Vec::new();
5912 0 : for (tenant_shard_id, optimization) in candidate_work {
5913 0 : match optimization.action {
5914 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
5915 : old_attached_node_id: _,
5916 0 : new_attached_node_id,
5917 0 : }) => {
5918 0 : match validation_nodes.get(&new_attached_node_id) {
5919 0 : None => {
5920 0 : // Node was dropped between planning and validation
5921 0 : }
5922 0 : Some(node) => {
5923 0 : if !node.is_available() {
5924 0 : tracing::info!("Skipping optimization migration of {tenant_shard_id} to {new_attached_node_id} because node unavailable");
5925 0 : } else {
5926 0 : // Accumulate optimizations that require fetching secondary status, so that we can execute these
5927 0 : // remote API requests concurrently.
5928 0 : want_secondary_status.push((
5929 0 : tenant_shard_id,
5930 0 : node.clone(),
5931 0 : optimization,
5932 0 : ));
5933 0 : }
5934 : }
5935 : }
5936 : }
5937 : ScheduleOptimizationAction::ReplaceSecondary(_) => {
5938 : // No extra checks needed to replace a secondary: this does not interrupt client access
5939 0 : validated_work.push((tenant_shard_id, optimization))
5940 : }
5941 : };
5942 : }
5943 :
5944 : // Call into pageserver API to find out if the destination secondary location is warm enough for a reasonably smooth migration: we
5945 : // do this so that we avoid spawning a Reconciler that would have to wait minutes/hours for a destination to warm up: that reconciler
5946 : // would hold a precious reconcile semaphore unit the whole time it was waiting for the destination to warm up.
5947 0 : let results = self
5948 0 : .tenant_for_shards_api(
5949 0 : want_secondary_status
5950 0 : .iter()
5951 0 : .map(|i| (i.0, i.1.clone()))
5952 0 : .collect(),
5953 0 : |tenant_shard_id, client| async move {
5954 0 : client.tenant_secondary_status(tenant_shard_id).await
5955 0 : },
5956 0 : 1,
5957 0 : 1,
5958 0 : SHORT_RECONCILE_TIMEOUT,
5959 0 : &self.cancel,
5960 0 : )
5961 0 : .await;
5962 :
5963 0 : for ((tenant_shard_id, node, optimization), secondary_status) in
5964 0 : want_secondary_status.into_iter().zip(results.into_iter())
5965 : {
5966 0 : match secondary_status {
5967 0 : Err(e) => {
5968 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node}, error querying secondary: {e}");
5969 : }
5970 0 : Ok(progress) => {
5971 : // We require secondary locations to have less than 10GiB of downloads pending before we will use
5972 : // them in an optimization
5973 : const DOWNLOAD_FRESHNESS_THRESHOLD: u64 = 10 * 1024 * 1024 * 1024;
5974 :
5975 0 : if progress.heatmap_mtime.is_none()
5976 0 : || progress.bytes_total < DOWNLOAD_FRESHNESS_THRESHOLD
5977 0 : && progress.bytes_downloaded != progress.bytes_total
5978 0 : || progress.bytes_total - progress.bytes_downloaded
5979 0 : > DOWNLOAD_FRESHNESS_THRESHOLD
5980 : {
5981 0 : tracing::info!("Skipping migration of {tenant_shard_id} to {node} because secondary isn't ready: {progress:?}");
5982 : } else {
5983 : // Location looks ready: proceed
5984 0 : tracing::info!(
5985 0 : "{tenant_shard_id} secondary on {node} is warm enough for migration: {progress:?}"
5986 : );
5987 0 : validated_work.push((tenant_shard_id, optimization))
5988 : }
5989 : }
5990 : }
5991 : }
5992 :
5993 0 : validated_work
5994 0 : }
5995 :
5996 : /// Look for shards which are oversized and in need of splitting
5997 0 : async fn autosplit_tenants(self: &Arc<Self>) {
5998 0 : let Some(split_threshold) = self.config.split_threshold else {
5999 : // Auto-splitting is disabled
6000 0 : return;
6001 : };
6002 :
6003 0 : let nodes = self.inner.read().unwrap().nodes.clone();
6004 :
6005 : const SPLIT_TO_MAX: ShardCount = ShardCount::new(8);
6006 :
6007 0 : let mut top_n = Vec::new();
6008 0 :
6009 0 : // Call into each node to look for big tenants
6010 0 : let top_n_request = TopTenantShardsRequest {
6011 0 : // We currently split based on logical size, for simplicity: logical size is a signal of
6012 0 : // the user's intent to run a large database, whereas physical/resident size can be symptoms
6013 0 : // of compaction issues. Eventually we should switch to using resident size to bound the
6014 0 : // disk space impact of one shard.
6015 0 : order_by: models::TenantSorting::MaxLogicalSize,
6016 0 : limit: 10,
6017 0 : where_shards_lt: Some(SPLIT_TO_MAX),
6018 0 : where_gt: Some(split_threshold),
6019 0 : };
6020 0 : for node in nodes.values() {
6021 0 : let request_ref = &top_n_request;
6022 0 : match node
6023 0 : .with_client_retries(
6024 0 : |client| async move {
6025 0 : let request = request_ref.clone();
6026 0 : client.top_tenant_shards(request.clone()).await
6027 0 : },
6028 0 : &self.config.jwt_token,
6029 0 : 3,
6030 0 : 3,
6031 0 : Duration::from_secs(5),
6032 0 : &self.cancel,
6033 0 : )
6034 0 : .await
6035 : {
6036 0 : Some(Ok(node_top_n)) => {
6037 0 : top_n.extend(node_top_n.shards.into_iter());
6038 0 : }
6039 : Some(Err(mgmt_api::Error::Cancelled)) => {
6040 0 : continue;
6041 : }
6042 0 : Some(Err(e)) => {
6043 0 : tracing::warn!("Failed to fetch top N tenants from {node}: {e}");
6044 0 : continue;
6045 : }
6046 : None => {
6047 : // Node is shutting down
6048 0 : continue;
6049 : }
6050 : };
6051 : }
6052 :
6053 : // Pick the biggest tenant to split first
6054 0 : top_n.sort_by_key(|i| i.resident_size);
6055 0 : let Some(split_candidate) = top_n.into_iter().next() else {
6056 0 : tracing::debug!("No split-elegible shards found");
6057 0 : return;
6058 : };
6059 :
6060 : // We spawn a task to run this, so it's exactly like some external API client requesting it. We don't
6061 : // want to block the background reconcile loop on this.
6062 0 : tracing::info!("Auto-splitting tenant for size threshold {split_threshold}: current size {split_candidate:?}");
6063 :
6064 0 : let this = self.clone();
6065 0 : tokio::spawn(
6066 0 : async move {
6067 0 : match this
6068 0 : .tenant_shard_split(
6069 0 : split_candidate.id.tenant_id,
6070 0 : TenantShardSplitRequest {
6071 0 : // Always split to the max number of shards: this avoids stepping through
6072 0 : // intervening shard counts and encountering the overrhead of a split+cleanup
6073 0 : // each time as a tenant grows, and is not too expensive because our max shard
6074 0 : // count is relatively low anyway.
6075 0 : // This policy will be adjusted in future once we support higher shard count.
6076 0 : new_shard_count: SPLIT_TO_MAX.literal(),
6077 0 : new_stripe_size: Some(ShardParameters::DEFAULT_STRIPE_SIZE),
6078 0 : },
6079 0 : )
6080 0 : .await
6081 : {
6082 : Ok(_) => {
6083 0 : tracing::info!("Successful auto-split");
6084 : }
6085 0 : Err(e) => {
6086 0 : tracing::error!("Auto-split failed: {e}");
6087 : }
6088 : }
6089 0 : }
6090 0 : .instrument(tracing::info_span!("auto_split", tenant_id=%split_candidate.id.tenant_id)),
6091 : );
6092 0 : }
6093 :
6094 : /// Useful for tests: run whatever work a background [`Self::reconcile_all`] would have done, but
6095 : /// also wait for any generated Reconcilers to complete. Calling this until it returns zero should
6096 : /// put the system into a quiescent state where future background reconciliations won't do anything.
6097 0 : pub(crate) async fn reconcile_all_now(&self) -> Result<usize, ReconcileWaitError> {
6098 0 : let reconciles_spawned = self.reconcile_all();
6099 0 : let reconciles_spawned = if reconciles_spawned == 0 {
6100 : // Only optimize when we are otherwise idle
6101 0 : self.optimize_all().await
6102 : } else {
6103 0 : reconciles_spawned
6104 : };
6105 :
6106 0 : let waiters = {
6107 0 : let mut waiters = Vec::new();
6108 0 : let locked = self.inner.read().unwrap();
6109 0 : for (_tenant_shard_id, shard) in locked.tenants.iter() {
6110 0 : if let Some(waiter) = shard.get_waiter() {
6111 0 : waiters.push(waiter);
6112 0 : }
6113 : }
6114 0 : waiters
6115 0 : };
6116 0 :
6117 0 : let waiter_count = waiters.len();
6118 0 : match self.await_waiters(waiters, RECONCILE_TIMEOUT).await {
6119 0 : Ok(()) => {}
6120 0 : Err(ReconcileWaitError::Failed(_, reconcile_error))
6121 0 : if matches!(*reconcile_error, ReconcileError::Cancel) =>
6122 0 : {
6123 0 : // Ignore reconciler cancel errors: this reconciler might have shut down
6124 0 : // because some other change superceded it. We will return a nonzero number,
6125 0 : // so the caller knows they might have to call again to quiesce the system.
6126 0 : }
6127 0 : Err(e) => {
6128 0 : return Err(e);
6129 : }
6130 : };
6131 :
6132 0 : tracing::info!(
6133 0 : "{} reconciles in reconcile_all, {} waiters",
6134 : reconciles_spawned,
6135 : waiter_count
6136 : );
6137 :
6138 0 : Ok(std::cmp::max(waiter_count, reconciles_spawned))
6139 0 : }
6140 :
6141 0 : async fn stop_reconciliations(&self, reason: StopReconciliationsReason) {
6142 0 : // Cancel all on-going reconciles and wait for them to exit the gate.
6143 0 : tracing::info!("{reason}: cancelling and waiting for in-flight reconciles");
6144 0 : self.reconcilers_cancel.cancel();
6145 0 : self.reconcilers_gate.close().await;
6146 :
6147 : // Signal the background loop in [`Service::process_results`] to exit once
6148 : // it has proccessed the results from all the reconciles we cancelled earlier.
6149 0 : tracing::info!("{reason}: processing results from previously in-flight reconciles");
6150 0 : self.result_tx.send(ReconcileResultRequest::Stop).ok();
6151 0 : self.result_tx.closed().await;
6152 0 : }
6153 :
6154 0 : pub async fn shutdown(&self) {
6155 0 : self.stop_reconciliations(StopReconciliationsReason::ShuttingDown)
6156 0 : .await;
6157 :
6158 : // Background tasks hold gate guards: this notifies them of the cancellation and
6159 : // waits for them all to complete.
6160 0 : tracing::info!("Shutting down: cancelling and waiting for background tasks to exit");
6161 0 : self.cancel.cancel();
6162 0 : self.gate.close().await;
6163 0 : }
6164 :
6165 : /// Spot check the download lag for a secondary location of a shard.
6166 : /// Should be used as a heuristic, since it's not always precise: the
6167 : /// secondary might have not downloaded the new heat map yet and, hence,
6168 : /// is not aware of the lag.
6169 : ///
6170 : /// Returns:
6171 : /// * Ok(None) if the lag could not be determined from the status,
6172 : /// * Ok(Some(_)) if the lag could be determind
6173 : /// * Err on failures to query the pageserver.
6174 0 : async fn secondary_lag(
6175 0 : &self,
6176 0 : secondary: &NodeId,
6177 0 : tenant_shard_id: TenantShardId,
6178 0 : ) -> Result<Option<u64>, mgmt_api::Error> {
6179 0 : let nodes = self.inner.read().unwrap().nodes.clone();
6180 0 : let node = nodes.get(secondary).ok_or(mgmt_api::Error::ApiError(
6181 0 : StatusCode::NOT_FOUND,
6182 0 : format!("Node with id {} not found", secondary),
6183 0 : ))?;
6184 :
6185 0 : match node
6186 0 : .with_client_retries(
6187 0 : |client| async move { client.tenant_secondary_status(tenant_shard_id).await },
6188 0 : &self.config.jwt_token,
6189 0 : 1,
6190 0 : 3,
6191 0 : Duration::from_millis(250),
6192 0 : &self.cancel,
6193 0 : )
6194 0 : .await
6195 : {
6196 0 : Some(Ok(status)) => match status.heatmap_mtime {
6197 0 : Some(_) => Ok(Some(status.bytes_total - status.bytes_downloaded)),
6198 0 : None => Ok(None),
6199 : },
6200 0 : Some(Err(e)) => Err(e),
6201 0 : None => Err(mgmt_api::Error::Cancelled),
6202 : }
6203 0 : }
6204 :
6205 : /// Drain a node by moving the shards attached to it as primaries.
6206 : /// This is a long running operation and it should run as a separate Tokio task.
6207 0 : pub(crate) async fn drain_node(
6208 0 : self: &Arc<Self>,
6209 0 : node_id: NodeId,
6210 0 : cancel: CancellationToken,
6211 0 : ) -> Result<(), OperationError> {
6212 : const MAX_SECONDARY_LAG_BYTES_DEFAULT: u64 = 256 * 1024 * 1024;
6213 0 : let max_secondary_lag_bytes = self
6214 0 : .config
6215 0 : .max_secondary_lag_bytes
6216 0 : .unwrap_or(MAX_SECONDARY_LAG_BYTES_DEFAULT);
6217 :
6218 : // By default, live migrations are generous about the wait time for getting
6219 : // the secondary location up to speed. When draining, give up earlier in order
6220 : // to not stall the operation when a cold secondary is encountered.
6221 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
6222 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
6223 0 : let reconciler_config = ReconcilerConfigBuilder::new()
6224 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
6225 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
6226 0 : .build();
6227 0 :
6228 0 : let mut waiters = Vec::new();
6229 0 :
6230 0 : let mut tid_iter = TenantShardIterator::new({
6231 0 : let service = self.clone();
6232 0 : move |last_inspected_shard: Option<TenantShardId>| {
6233 0 : let locked = &service.inner.read().unwrap();
6234 0 : let tenants = &locked.tenants;
6235 0 : let entry = match last_inspected_shard {
6236 0 : Some(skip_past) => {
6237 0 : // Skip to the last seen tenant shard id
6238 0 : let mut cursor = tenants.iter().skip_while(|(tid, _)| **tid != skip_past);
6239 0 :
6240 0 : // Skip past the last seen
6241 0 : cursor.nth(1)
6242 : }
6243 0 : None => tenants.first_key_value(),
6244 : };
6245 :
6246 0 : entry.map(|(tid, _)| tid).copied()
6247 0 : }
6248 0 : });
6249 :
6250 0 : while !tid_iter.finished() {
6251 0 : if cancel.is_cancelled() {
6252 0 : match self
6253 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6254 0 : .await
6255 : {
6256 0 : Ok(()) => return Err(OperationError::Cancelled),
6257 0 : Err(err) => {
6258 0 : return Err(OperationError::FinalizeError(
6259 0 : format!(
6260 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6261 0 : node_id, err
6262 0 : )
6263 0 : .into(),
6264 0 : ));
6265 : }
6266 : }
6267 0 : }
6268 0 :
6269 0 : drain_utils::validate_node_state(&node_id, self.inner.read().unwrap().nodes.clone())?;
6270 :
6271 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
6272 0 : let tid = match tid_iter.next() {
6273 0 : Some(tid) => tid,
6274 : None => {
6275 0 : break;
6276 : }
6277 : };
6278 :
6279 0 : let tid_drain = TenantShardDrain {
6280 0 : drained_node: node_id,
6281 0 : tenant_shard_id: tid,
6282 0 : };
6283 :
6284 0 : let dest_node_id = {
6285 0 : let locked = self.inner.read().unwrap();
6286 0 :
6287 0 : match tid_drain
6288 0 : .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
6289 : {
6290 0 : Some(node_id) => node_id,
6291 : None => {
6292 0 : continue;
6293 : }
6294 : }
6295 : };
6296 :
6297 0 : match self.secondary_lag(&dest_node_id, tid).await {
6298 0 : Ok(Some(lag)) if lag <= max_secondary_lag_bytes => {
6299 0 : // The secondary is reasonably up to date.
6300 0 : // Migrate to it
6301 0 : }
6302 0 : Ok(Some(lag)) => {
6303 0 : tracing::info!(
6304 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6305 0 : "Secondary on node {dest_node_id} is lagging by {lag}. Skipping reconcile."
6306 : );
6307 0 : continue;
6308 : }
6309 : Ok(None) => {
6310 0 : tracing::info!(
6311 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6312 0 : "Could not determine lag for secondary on node {dest_node_id}. Skipping reconcile."
6313 : );
6314 0 : continue;
6315 : }
6316 0 : Err(err) => {
6317 0 : tracing::warn!(
6318 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6319 0 : "Failed to get secondary lag from node {dest_node_id}. Skipping reconcile: {err}"
6320 : );
6321 0 : continue;
6322 : }
6323 : }
6324 :
6325 : {
6326 0 : let mut locked = self.inner.write().unwrap();
6327 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6328 0 : let rescheduled = tid_drain.reschedule_to_secondary(
6329 0 : dest_node_id,
6330 0 : tenants,
6331 0 : scheduler,
6332 0 : nodes,
6333 0 : )?;
6334 :
6335 0 : if let Some(tenant_shard) = rescheduled {
6336 0 : let waiter = self.maybe_configured_reconcile_shard(
6337 0 : tenant_shard,
6338 0 : nodes,
6339 0 : reconciler_config,
6340 0 : );
6341 0 : if let Some(some) = waiter {
6342 0 : waiters.push(some);
6343 0 : }
6344 0 : }
6345 : }
6346 : }
6347 :
6348 0 : waiters = self
6349 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6350 0 : .await;
6351 :
6352 0 : failpoint_support::sleep_millis_async!("sleepy-drain-loop", &cancel);
6353 : }
6354 :
6355 0 : while !waiters.is_empty() {
6356 0 : if cancel.is_cancelled() {
6357 0 : match self
6358 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6359 0 : .await
6360 : {
6361 0 : Ok(()) => return Err(OperationError::Cancelled),
6362 0 : Err(err) => {
6363 0 : return Err(OperationError::FinalizeError(
6364 0 : format!(
6365 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6366 0 : node_id, err
6367 0 : )
6368 0 : .into(),
6369 0 : ));
6370 : }
6371 : }
6372 0 : }
6373 0 :
6374 0 : tracing::info!("Awaiting {} pending drain reconciliations", waiters.len());
6375 :
6376 0 : waiters = self
6377 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6378 0 : .await;
6379 : }
6380 :
6381 : // At this point we have done the best we could to drain shards from this node.
6382 : // Set the node scheduling policy to `[NodeSchedulingPolicy::PauseForRestart]`
6383 : // to complete the drain.
6384 0 : if let Err(err) = self
6385 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::PauseForRestart))
6386 0 : .await
6387 : {
6388 : // This is not fatal. Anything that is polling the node scheduling policy to detect
6389 : // the end of the drain operations will hang, but all such places should enforce an
6390 : // overall timeout. The scheduling policy will be updated upon node re-attach and/or
6391 : // by the counterpart fill operation.
6392 0 : return Err(OperationError::FinalizeError(
6393 0 : format!(
6394 0 : "Failed to finalise drain of {node_id} by setting scheduling policy to PauseForRestart: {err}"
6395 0 : )
6396 0 : .into(),
6397 0 : ));
6398 0 : }
6399 0 :
6400 0 : Ok(())
6401 0 : }
6402 :
6403 : /// Create a node fill plan (pick secondaries to promote) that meets the following requirements:
6404 : /// 1. The node should be filled until it reaches the expected cluster average of
6405 : /// attached shards. If there are not enough secondaries on the node, the plan stops early.
6406 : /// 2. Select tenant shards to promote such that the number of attached shards is balanced
6407 : /// throughout the cluster. We achieve this by picking tenant shards from each node,
6408 : /// starting from the ones with the largest number of attached shards, until the node
6409 : /// reaches the expected cluster average.
6410 : /// 3. Avoid promoting more shards of the same tenant than required. The upper bound
6411 : /// for the number of tenants from the same shard promoted to the node being filled is:
6412 : /// shard count for the tenant divided by the number of nodes in the cluster.
6413 0 : fn fill_node_plan(&self, node_id: NodeId) -> Vec<TenantShardId> {
6414 0 : let mut locked = self.inner.write().unwrap();
6415 0 : let fill_requirement = locked.scheduler.compute_fill_requirement(node_id);
6416 0 :
6417 0 : let mut tids_by_node = locked
6418 0 : .tenants
6419 0 : .iter_mut()
6420 0 : .filter_map(|(tid, tenant_shard)| {
6421 0 : if tenant_shard.intent.get_secondary().contains(&node_id) {
6422 0 : if let Some(primary) = tenant_shard.intent.get_attached() {
6423 0 : return Some((*primary, *tid));
6424 0 : }
6425 0 : }
6426 :
6427 0 : None
6428 0 : })
6429 0 : .into_group_map();
6430 0 :
6431 0 : let expected_attached = locked.scheduler.expected_attached_shard_count();
6432 0 : let nodes_by_load = locked.scheduler.nodes_by_attached_shard_count();
6433 0 :
6434 0 : let mut promoted_per_tenant: HashMap<TenantId, usize> = HashMap::new();
6435 0 : let mut plan = Vec::new();
6436 :
6437 0 : for (node_id, attached) in nodes_by_load {
6438 0 : let available = locked
6439 0 : .nodes
6440 0 : .get(&node_id)
6441 0 : .map_or(false, |n| n.is_available());
6442 0 : if !available {
6443 0 : continue;
6444 0 : }
6445 0 :
6446 0 : if plan.len() >= fill_requirement
6447 0 : || tids_by_node.is_empty()
6448 0 : || attached <= expected_attached
6449 : {
6450 0 : break;
6451 0 : }
6452 0 :
6453 0 : let can_take = attached - expected_attached;
6454 0 : let needed = fill_requirement - plan.len();
6455 0 : let mut take = std::cmp::min(can_take, needed);
6456 0 :
6457 0 : let mut remove_node = false;
6458 0 : while take > 0 {
6459 0 : match tids_by_node.get_mut(&node_id) {
6460 0 : Some(tids) => match tids.pop() {
6461 0 : Some(tid) => {
6462 0 : let max_promote_for_tenant = std::cmp::max(
6463 0 : tid.shard_count.count() as usize / locked.nodes.len(),
6464 0 : 1,
6465 0 : );
6466 0 : let promoted = promoted_per_tenant.entry(tid.tenant_id).or_default();
6467 0 : if *promoted < max_promote_for_tenant {
6468 0 : plan.push(tid);
6469 0 : *promoted += 1;
6470 0 : take -= 1;
6471 0 : }
6472 : }
6473 : None => {
6474 0 : remove_node = true;
6475 0 : break;
6476 : }
6477 : },
6478 : None => {
6479 0 : break;
6480 : }
6481 : }
6482 : }
6483 :
6484 0 : if remove_node {
6485 0 : tids_by_node.remove(&node_id);
6486 0 : }
6487 : }
6488 :
6489 0 : plan
6490 0 : }
6491 :
6492 : /// Fill a node by promoting its secondaries until the cluster is balanced
6493 : /// with regards to attached shard counts. Note that this operation only
6494 : /// makes sense as a counterpart to the drain implemented in [`Service::drain_node`].
6495 : /// This is a long running operation and it should run as a separate Tokio task.
6496 0 : pub(crate) async fn fill_node(
6497 0 : &self,
6498 0 : node_id: NodeId,
6499 0 : cancel: CancellationToken,
6500 0 : ) -> Result<(), OperationError> {
6501 : const SECONDARY_WARMUP_TIMEOUT: Duration = Duration::from_secs(20);
6502 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT: Duration = Duration::from_secs(5);
6503 0 : let reconciler_config = ReconcilerConfigBuilder::new()
6504 0 : .secondary_warmup_timeout(SECONDARY_WARMUP_TIMEOUT)
6505 0 : .secondary_download_request_timeout(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT)
6506 0 : .build();
6507 0 :
6508 0 : let mut tids_to_promote = self.fill_node_plan(node_id);
6509 0 : let mut waiters = Vec::new();
6510 :
6511 : // Execute the plan we've composed above. Before aplying each move from the plan,
6512 : // we validate to ensure that it has not gone stale in the meantime.
6513 0 : while !tids_to_promote.is_empty() {
6514 0 : if cancel.is_cancelled() {
6515 0 : match self
6516 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6517 0 : .await
6518 : {
6519 0 : Ok(()) => return Err(OperationError::Cancelled),
6520 0 : Err(err) => {
6521 0 : return Err(OperationError::FinalizeError(
6522 0 : format!(
6523 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6524 0 : node_id, err
6525 0 : )
6526 0 : .into(),
6527 0 : ));
6528 : }
6529 : }
6530 0 : }
6531 0 :
6532 0 : {
6533 0 : let mut locked = self.inner.write().unwrap();
6534 0 : let (nodes, tenants, scheduler) = locked.parts_mut();
6535 :
6536 0 : let node = nodes.get(&node_id).ok_or(OperationError::NodeStateChanged(
6537 0 : format!("node {node_id} was removed").into(),
6538 0 : ))?;
6539 :
6540 0 : let current_policy = node.get_scheduling();
6541 0 : if !matches!(current_policy, NodeSchedulingPolicy::Filling) {
6542 : // TODO(vlad): maybe cancel pending reconciles before erroring out. need to think
6543 : // about it
6544 0 : return Err(OperationError::NodeStateChanged(
6545 0 : format!("node {node_id} changed state to {current_policy:?}").into(),
6546 0 : ));
6547 0 : }
6548 :
6549 0 : while waiters.len() < MAX_RECONCILES_PER_OPERATION {
6550 0 : if let Some(tid) = tids_to_promote.pop() {
6551 0 : if let Some(tenant_shard) = tenants.get_mut(&tid) {
6552 : // If the node being filled is not a secondary anymore,
6553 : // skip the promotion.
6554 0 : if !tenant_shard.intent.get_secondary().contains(&node_id) {
6555 0 : continue;
6556 0 : }
6557 0 :
6558 0 : let previously_attached_to = *tenant_shard.intent.get_attached();
6559 0 : match tenant_shard.reschedule_to_secondary(Some(node_id), scheduler) {
6560 0 : Err(e) => {
6561 0 : tracing::warn!(
6562 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6563 0 : "Scheduling error when filling pageserver {} : {e}", node_id
6564 : );
6565 : }
6566 : Ok(()) => {
6567 0 : tracing::info!(
6568 0 : tenant_id=%tid.tenant_id, shard_id=%tid.shard_slug(),
6569 0 : "Rescheduled shard while filling node {}: {:?} -> {}",
6570 : node_id,
6571 : previously_attached_to,
6572 : node_id
6573 : );
6574 :
6575 0 : if let Some(waiter) = self.maybe_configured_reconcile_shard(
6576 0 : tenant_shard,
6577 0 : nodes,
6578 0 : reconciler_config,
6579 0 : ) {
6580 0 : waiters.push(waiter);
6581 0 : }
6582 : }
6583 : }
6584 0 : }
6585 : } else {
6586 0 : break;
6587 : }
6588 : }
6589 : }
6590 :
6591 0 : waiters = self
6592 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6593 0 : .await;
6594 : }
6595 :
6596 0 : while !waiters.is_empty() {
6597 0 : if cancel.is_cancelled() {
6598 0 : match self
6599 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6600 0 : .await
6601 : {
6602 0 : Ok(()) => return Err(OperationError::Cancelled),
6603 0 : Err(err) => {
6604 0 : return Err(OperationError::FinalizeError(
6605 0 : format!(
6606 0 : "Failed to finalise drain cancel of {} by setting scheduling policy to Active: {}",
6607 0 : node_id, err
6608 0 : )
6609 0 : .into(),
6610 0 : ));
6611 : }
6612 : }
6613 0 : }
6614 0 :
6615 0 : tracing::info!("Awaiting {} pending fill reconciliations", waiters.len());
6616 :
6617 0 : waiters = self
6618 0 : .await_waiters_remainder(waiters, SHORT_RECONCILE_TIMEOUT)
6619 0 : .await;
6620 : }
6621 :
6622 0 : if let Err(err) = self
6623 0 : .node_configure(node_id, None, Some(NodeSchedulingPolicy::Active))
6624 0 : .await
6625 : {
6626 : // This isn't a huge issue since the filling process starts upon request. However, it
6627 : // will prevent the next drain from starting. The only case in which this can fail
6628 : // is database unavailability. Such a case will require manual intervention.
6629 0 : return Err(OperationError::FinalizeError(
6630 0 : format!("Failed to finalise fill of {node_id} by setting scheduling policy to Active: {err}")
6631 0 : .into(),
6632 0 : ));
6633 0 : }
6634 0 :
6635 0 : Ok(())
6636 0 : }
6637 :
6638 : /// Updates scrubber metadata health check results.
6639 0 : pub(crate) async fn metadata_health_update(
6640 0 : &self,
6641 0 : update_req: MetadataHealthUpdateRequest,
6642 0 : ) -> Result<(), ApiError> {
6643 0 : let now = chrono::offset::Utc::now();
6644 0 : let (healthy_records, unhealthy_records) = {
6645 0 : let locked = self.inner.read().unwrap();
6646 0 : let healthy_records = update_req
6647 0 : .healthy_tenant_shards
6648 0 : .into_iter()
6649 0 : // Retain only health records associated with tenant shards managed by storage controller.
6650 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6651 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, true, now))
6652 0 : .collect();
6653 0 : let unhealthy_records = update_req
6654 0 : .unhealthy_tenant_shards
6655 0 : .into_iter()
6656 0 : .filter(|tenant_shard_id| locked.tenants.contains_key(tenant_shard_id))
6657 0 : .map(|tenant_shard_id| MetadataHealthPersistence::new(tenant_shard_id, false, now))
6658 0 : .collect();
6659 0 :
6660 0 : (healthy_records, unhealthy_records)
6661 0 : };
6662 0 :
6663 0 : self.persistence
6664 0 : .update_metadata_health_records(healthy_records, unhealthy_records, now)
6665 0 : .await?;
6666 0 : Ok(())
6667 0 : }
6668 :
6669 : /// Lists the tenant shards that has unhealthy metadata status.
6670 0 : pub(crate) async fn metadata_health_list_unhealthy(
6671 0 : &self,
6672 0 : ) -> Result<Vec<TenantShardId>, ApiError> {
6673 0 : let result = self
6674 0 : .persistence
6675 0 : .list_unhealthy_metadata_health_records()
6676 0 : .await?
6677 0 : .iter()
6678 0 : .map(|p| p.get_tenant_shard_id().unwrap())
6679 0 : .collect();
6680 0 :
6681 0 : Ok(result)
6682 0 : }
6683 :
6684 : /// Lists the tenant shards that have not been scrubbed for some duration.
6685 0 : pub(crate) async fn metadata_health_list_outdated(
6686 0 : &self,
6687 0 : not_scrubbed_for: Duration,
6688 0 : ) -> Result<Vec<MetadataHealthRecord>, ApiError> {
6689 0 : let earlier = chrono::offset::Utc::now() - not_scrubbed_for;
6690 0 : let result = self
6691 0 : .persistence
6692 0 : .list_outdated_metadata_health_records(earlier)
6693 0 : .await?
6694 0 : .into_iter()
6695 0 : .map(|record| record.into())
6696 0 : .collect();
6697 0 : Ok(result)
6698 0 : }
6699 :
6700 0 : pub(crate) fn get_leadership_status(&self) -> LeadershipStatus {
6701 0 : self.inner.read().unwrap().get_leadership_status()
6702 0 : }
6703 :
6704 0 : pub(crate) async fn step_down(&self) -> GlobalObservedState {
6705 0 : tracing::info!("Received step down request from peer");
6706 0 : failpoint_support::sleep_millis_async!("sleep-on-step-down-handling");
6707 :
6708 0 : self.inner.write().unwrap().step_down();
6709 0 : // TODO: would it make sense to have a time-out for this?
6710 0 : self.stop_reconciliations(StopReconciliationsReason::SteppingDown)
6711 0 : .await;
6712 :
6713 0 : let mut global_observed = GlobalObservedState::default();
6714 0 : let locked = self.inner.read().unwrap();
6715 0 : for (tid, tenant_shard) in locked.tenants.iter() {
6716 0 : global_observed
6717 0 : .0
6718 0 : .insert(*tid, tenant_shard.observed.clone());
6719 0 : }
6720 :
6721 0 : global_observed
6722 0 : }
6723 :
6724 0 : pub(crate) async fn get_safekeeper(
6725 0 : &self,
6726 0 : id: i64,
6727 0 : ) -> Result<crate::persistence::SafekeeperPersistence, DatabaseError> {
6728 0 : self.persistence.safekeeper_get(id).await
6729 0 : }
6730 :
6731 0 : pub(crate) async fn upsert_safekeeper(
6732 0 : &self,
6733 0 : record: crate::persistence::SafekeeperPersistence,
6734 0 : ) -> Result<(), DatabaseError> {
6735 0 : self.persistence.safekeeper_upsert(record).await
6736 0 : }
6737 :
6738 0 : pub(crate) async fn update_shards_preferred_azs(
6739 0 : &self,
6740 0 : req: ShardsPreferredAzsRequest,
6741 0 : ) -> Result<ShardsPreferredAzsResponse, ApiError> {
6742 0 : let preferred_azs = req.preferred_az_ids.into_iter().collect::<Vec<_>>();
6743 0 : let updated = self
6744 0 : .persistence
6745 0 : .set_tenant_shard_preferred_azs(preferred_azs)
6746 0 : .await
6747 0 : .map_err(|err| {
6748 0 : ApiError::InternalServerError(anyhow::anyhow!(
6749 0 : "Failed to persist preferred AZs: {err}"
6750 0 : ))
6751 0 : })?;
6752 :
6753 0 : let mut updated_in_mem_and_db = Vec::default();
6754 0 :
6755 0 : let mut locked = self.inner.write().unwrap();
6756 0 : for (tid, az_id) in updated {
6757 0 : let shard = locked.tenants.get_mut(&tid);
6758 0 : if let Some(shard) = shard {
6759 0 : shard.set_preferred_az(az_id);
6760 0 : updated_in_mem_and_db.push(tid);
6761 0 : }
6762 : }
6763 :
6764 0 : Ok(ShardsPreferredAzsResponse {
6765 0 : updated: updated_in_mem_and_db,
6766 0 : })
6767 0 : }
6768 : }
|