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