Line data Source code
1 : use std::collections::{HashMap, HashSet};
2 : use std::sync::Arc;
3 : use std::time::Duration;
4 :
5 : use futures::future::{self, Either};
6 : use itertools::Itertools;
7 : use pageserver_api::controller_api::{AvailabilityZone, PlacementPolicy, ShardSchedulingPolicy};
8 : use pageserver_api::models::{LocationConfig, LocationConfigMode, TenantConfig};
9 : use pageserver_api::shard::{ShardIdentity, TenantShardId};
10 : use serde::{Deserialize, Serialize};
11 : use tokio::task::JoinHandle;
12 : use tokio_util::sync::CancellationToken;
13 : use tracing::{Instrument, instrument};
14 : use utils::generation::Generation;
15 : use utils::id::NodeId;
16 : use utils::seqwait::{SeqWait, SeqWaitError};
17 : use utils::shard::ShardCount;
18 : use utils::sync::gate::GateGuard;
19 :
20 : use crate::compute_hook::ComputeHook;
21 : use crate::metrics::{
22 : self, ReconcileCompleteLabelGroup, ReconcileLongRunningLabelGroup, ReconcileOutcome,
23 : };
24 : use crate::node::Node;
25 : use crate::persistence::split_state::SplitState;
26 : use crate::persistence::{Persistence, TenantShardPersistence};
27 : use crate::reconciler::{
28 : ReconcileError, ReconcileUnits, Reconciler, ReconcilerConfig, TargetState,
29 : attached_location_conf, secondary_location_conf,
30 : };
31 : use crate::scheduler::{
32 : AffinityScore, AttachedShardTag, NodeSchedulingScore, NodeSecondarySchedulingScore,
33 : RefCountUpdate, ScheduleContext, ScheduleError, Scheduler, SecondaryShardTag, ShardTag,
34 : };
35 : use crate::service::ReconcileResultRequest;
36 : use crate::timeline_import::TimelineImportState;
37 : use crate::{Sequence, service};
38 :
39 : /// Serialization helper
40 0 : fn read_last_error<S, T>(v: &std::sync::Mutex<Option<T>>, serializer: S) -> Result<S::Ok, S::Error>
41 0 : where
42 0 : S: serde::ser::Serializer,
43 0 : T: std::fmt::Display,
44 : {
45 0 : serializer.collect_str(
46 0 : &v.lock()
47 0 : .unwrap()
48 0 : .as_ref()
49 0 : .map(|e| format!("{e}"))
50 0 : .unwrap_or("".to_string()),
51 : )
52 0 : }
53 :
54 : /// In-memory state for a particular tenant shard.
55 : ///
56 : /// This struct implement Serialize for debugging purposes, but is _not_ persisted
57 : /// itself: see [`crate::persistence`] for the subset of tenant shard state that is persisted.
58 : #[derive(Serialize)]
59 : pub(crate) struct TenantShard {
60 : pub(crate) tenant_shard_id: TenantShardId,
61 :
62 : pub(crate) shard: ShardIdentity,
63 :
64 : // Runtime only: sequence used to coordinate when updating this object while
65 : // with background reconcilers may be running. A reconciler runs to a particular
66 : // sequence.
67 : pub(crate) sequence: Sequence,
68 :
69 : // Latest generation number: next time we attach, increment this
70 : // and use the incremented number when attaching.
71 : //
72 : // None represents an incompletely onboarded tenant via the [`Service::location_config`]
73 : // API, where this tenant may only run in PlacementPolicy::Secondary.
74 : pub(crate) generation: Option<Generation>,
75 :
76 : // High level description of how the tenant should be set up. Provided
77 : // externally.
78 : pub(crate) policy: PlacementPolicy,
79 :
80 : // Low level description of exactly which pageservers should fulfil
81 : // which role. Generated by `Self::schedule`.
82 : pub(crate) intent: IntentState,
83 :
84 : // Low level description of how the tenant is configured on pageservers:
85 : // if this does not match `Self::intent` then the tenant needs reconciliation
86 : // with `Self::reconcile`.
87 : pub(crate) observed: ObservedState,
88 :
89 : // Tenant configuration, passed through opaquely to the pageserver. Identical
90 : // for all shards in a tenant.
91 : pub(crate) config: TenantConfig,
92 :
93 : /// If a reconcile task is currently in flight, it may be joined here (it is
94 : /// only safe to join if either the result has been received or the reconciler's
95 : /// cancellation token has been fired)
96 : #[serde(skip)]
97 : pub(crate) reconciler: Option<ReconcilerHandle>,
98 :
99 : /// If a tenant is being split, then all shards with that TenantId will have a
100 : /// SplitState set, this acts as a guard against other operations such as background
101 : /// reconciliation, and timeline creation.
102 : pub(crate) splitting: SplitState,
103 :
104 : /// Flag indicating whether the tenant has an in-progress timeline import.
105 : /// Used to disallow shard splits while an import is in progress.
106 : pub(crate) importing: TimelineImportState,
107 :
108 : /// If a tenant was enqueued for later reconcile due to hitting concurrency limit, this flag
109 : /// is set. This flag is cleared when the tenant is popped off the delay queue.
110 : pub(crate) delayed_reconcile: bool,
111 :
112 : /// Optionally wait for reconciliation to complete up to a particular
113 : /// sequence number.
114 : #[serde(skip)]
115 : pub(crate) waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
116 :
117 : /// Indicates sequence number for which we have encountered an error reconciling. If
118 : /// this advances ahead of [`Self::waiter`] then a reconciliation error has occurred,
119 : /// and callers should stop waiting for `waiter` and propagate the error.
120 : #[serde(skip)]
121 : pub(crate) error_waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
122 :
123 : /// The most recent error from a reconcile on this tenant. This is a nested Arc
124 : /// because:
125 : /// - ReconcileWaiters need to Arc-clone the overall object to read it later
126 : /// - ReconcileWaitError needs to use an `Arc<ReconcileError>` because we can construct
127 : /// many waiters for one shard, and the underlying error types are not Clone.
128 : ///
129 : /// TODO: generalize to an array of recent events
130 : /// TOOD: use a ArcSwap instead of mutex for faster reads?
131 : #[serde(serialize_with = "read_last_error")]
132 : pub(crate) last_error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
133 :
134 : /// Number of consecutive reconciliation errors that have occurred for this shard.
135 : ///
136 : /// When this count reaches MAX_CONSECUTIVE_RECONCILIATION_ERRORS, the tenant shard
137 : /// will be countered as keep-failing in `reconcile_all` calculations. This will lead to
138 : /// allowing optimizations to run even with some failing shards.
139 : ///
140 : /// The counter is reset to 0 after a successful reconciliation.
141 : pub(crate) consecutive_errors_count: usize,
142 :
143 : /// If we have a pending compute notification that for some reason we weren't able to send,
144 : /// set this to true. If this is set, calls to [`Self::get_reconcile_needed`] will return Yes
145 : /// and trigger a Reconciler run. This is the mechanism by which compute notifications are included in the scope
146 : /// of state that we publish externally in an eventually consistent way.
147 : pub(crate) pending_compute_notification: bool,
148 :
149 : /// To do a graceful migration, set this field to the destination pageserver, and optimization
150 : /// functions will consider this node the best location and react appropriately.
151 : preferred_node: Option<NodeId>,
152 :
153 : // Support/debug tool: if something is going wrong or flapping with scheduling, this may
154 : // be set to a non-active state to avoid making changes while the issue is fixed.
155 : scheduling_policy: ShardSchedulingPolicy,
156 : }
157 :
158 : #[derive(Clone, Debug, Serialize)]
159 : pub(crate) struct IntentState {
160 : attached: Option<NodeId>,
161 : secondary: Vec<NodeId>,
162 :
163 : // We should attempt to schedule this shard in the provided AZ to
164 : // decrease chances of cross-AZ compute.
165 : preferred_az_id: Option<AvailabilityZone>,
166 : }
167 :
168 : impl IntentState {
169 12860 : pub(crate) fn new(preferred_az_id: Option<AvailabilityZone>) -> Self {
170 12860 : Self {
171 12860 : attached: None,
172 12860 : secondary: vec![],
173 12860 : preferred_az_id,
174 12860 : }
175 12860 : }
176 0 : pub(crate) fn single(
177 0 : scheduler: &mut Scheduler,
178 0 : node_id: Option<NodeId>,
179 0 : preferred_az_id: Option<AvailabilityZone>,
180 0 : ) -> Self {
181 0 : if let Some(node_id) = node_id {
182 0 : scheduler.update_node_ref_counts(
183 0 : node_id,
184 0 : preferred_az_id.as_ref(),
185 0 : RefCountUpdate::Attach,
186 0 : );
187 0 : }
188 0 : Self {
189 0 : attached: node_id,
190 0 : secondary: vec![],
191 0 : preferred_az_id,
192 0 : }
193 0 : }
194 :
195 12859 : pub(crate) fn set_attached(&mut self, scheduler: &mut Scheduler, new_attached: Option<NodeId>) {
196 12859 : if self.attached != new_attached {
197 12859 : if let Some(old_attached) = self.attached.take() {
198 0 : scheduler.update_node_ref_counts(
199 0 : old_attached,
200 0 : self.preferred_az_id.as_ref(),
201 0 : RefCountUpdate::Detach,
202 0 : );
203 12859 : }
204 12859 : if let Some(new_attached) = &new_attached {
205 12859 : scheduler.update_node_ref_counts(
206 12859 : *new_attached,
207 12859 : self.preferred_az_id.as_ref(),
208 12859 : RefCountUpdate::Attach,
209 12859 : );
210 12859 : }
211 12859 : self.attached = new_attached;
212 0 : }
213 :
214 12859 : if let Some(new_attached) = &new_attached {
215 12859 : assert!(!self.secondary.contains(new_attached));
216 0 : }
217 12859 : }
218 :
219 : /// Like set_attached, but the node is from [`Self::secondary`]. This swaps the node from
220 : /// secondary to attached while maintaining the scheduler's reference counts.
221 7 : pub(crate) fn promote_attached(
222 7 : &mut self,
223 7 : scheduler: &mut Scheduler,
224 7 : promote_secondary: NodeId,
225 7 : ) {
226 : // If we call this with a node that isn't in secondary, it would cause incorrect
227 : // scheduler reference counting, since we assume the node is already referenced as a secondary.
228 7 : debug_assert!(self.secondary.contains(&promote_secondary));
229 :
230 16 : self.secondary.retain(|n| n != &promote_secondary);
231 :
232 7 : let demoted = self.attached;
233 7 : self.attached = Some(promote_secondary);
234 :
235 7 : scheduler.update_node_ref_counts(
236 7 : promote_secondary,
237 7 : self.preferred_az_id.as_ref(),
238 7 : RefCountUpdate::PromoteSecondary,
239 : );
240 7 : if let Some(demoted) = demoted {
241 0 : scheduler.update_node_ref_counts(
242 0 : demoted,
243 0 : self.preferred_az_id.as_ref(),
244 0 : RefCountUpdate::DemoteAttached,
245 0 : );
246 7 : }
247 7 : }
248 :
249 12848 : pub(crate) fn push_secondary(&mut self, scheduler: &mut Scheduler, new_secondary: NodeId) {
250 12848 : assert!(!self.secondary.contains(&new_secondary));
251 12848 : assert!(self.attached != Some(new_secondary));
252 12848 : scheduler.update_node_ref_counts(
253 12848 : new_secondary,
254 12848 : self.preferred_az_id.as_ref(),
255 12848 : RefCountUpdate::AddSecondary,
256 : );
257 12848 : self.secondary.push(new_secondary);
258 12848 : }
259 :
260 : /// It is legal to call this with a node that is not currently a secondary: that is a no-op
261 9 : pub(crate) fn remove_secondary(&mut self, scheduler: &mut Scheduler, node_id: NodeId) {
262 13 : let index = self.secondary.iter().position(|n| *n == node_id);
263 9 : if let Some(index) = index {
264 9 : scheduler.update_node_ref_counts(
265 9 : node_id,
266 9 : self.preferred_az_id.as_ref(),
267 9 : RefCountUpdate::RemoveSecondary,
268 9 : );
269 9 : self.secondary.remove(index);
270 9 : }
271 9 : }
272 :
273 12859 : pub(crate) fn clear_secondary(&mut self, scheduler: &mut Scheduler) {
274 12859 : for secondary in self.secondary.drain(..) {
275 12840 : scheduler.update_node_ref_counts(
276 12840 : secondary,
277 12840 : self.preferred_az_id.as_ref(),
278 12840 : RefCountUpdate::RemoveSecondary,
279 12840 : );
280 12840 : }
281 12859 : }
282 :
283 : /// Remove the last secondary node from the list of secondaries
284 0 : pub(crate) fn pop_secondary(&mut self, scheduler: &mut Scheduler) {
285 0 : if let Some(node_id) = self.secondary.pop() {
286 0 : scheduler.update_node_ref_counts(
287 0 : node_id,
288 0 : self.preferred_az_id.as_ref(),
289 0 : RefCountUpdate::RemoveSecondary,
290 0 : );
291 0 : }
292 0 : }
293 :
294 12859 : pub(crate) fn clear(&mut self, scheduler: &mut Scheduler) {
295 12859 : if let Some(old_attached) = self.attached.take() {
296 12857 : scheduler.update_node_ref_counts(
297 12857 : old_attached,
298 12857 : self.preferred_az_id.as_ref(),
299 12857 : RefCountUpdate::Detach,
300 12857 : );
301 12857 : }
302 :
303 12859 : self.clear_secondary(scheduler);
304 12859 : }
305 :
306 12901 : pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
307 12901 : let mut result = Vec::new();
308 12901 : if let Some(p) = self.attached {
309 12896 : result.push(p)
310 5 : }
311 :
312 12901 : result.extend(self.secondary.iter().copied());
313 :
314 12901 : result
315 12901 : }
316 :
317 12624 : pub(crate) fn get_attached(&self) -> &Option<NodeId> {
318 12624 : &self.attached
319 12624 : }
320 :
321 12706 : pub(crate) fn get_secondary(&self) -> &Vec<NodeId> {
322 12706 : &self.secondary
323 12706 : }
324 :
325 : /// If the node is in use as the attached location, demote it into
326 : /// the list of secondary locations. This is used when a node goes offline,
327 : /// and we want to use a different node for attachment, but not permanently
328 : /// forget the location on the offline node.
329 : ///
330 : /// Returns true if a change was made
331 8 : pub(crate) fn demote_attached(&mut self, scheduler: &mut Scheduler, node_id: NodeId) -> bool {
332 8 : if self.attached == Some(node_id) {
333 8 : self.attached = None;
334 8 : self.secondary.push(node_id);
335 8 : scheduler.update_node_ref_counts(
336 8 : node_id,
337 8 : self.preferred_az_id.as_ref(),
338 8 : RefCountUpdate::DemoteAttached,
339 : );
340 8 : true
341 : } else {
342 0 : false
343 : }
344 8 : }
345 :
346 2 : pub(crate) fn set_preferred_az(
347 2 : &mut self,
348 2 : scheduler: &mut Scheduler,
349 2 : preferred_az: Option<AvailabilityZone>,
350 2 : ) {
351 2 : let new_az = preferred_az.as_ref();
352 2 : let old_az = self.preferred_az_id.as_ref();
353 :
354 2 : if old_az != new_az {
355 2 : if let Some(node_id) = self.attached {
356 2 : scheduler.update_node_ref_counts(
357 2 : node_id,
358 2 : new_az,
359 2 : RefCountUpdate::ChangePreferredAzFrom(old_az),
360 2 : );
361 2 : }
362 4 : for node_id in &self.secondary {
363 2 : scheduler.update_node_ref_counts(
364 2 : *node_id,
365 2 : new_az,
366 2 : RefCountUpdate::ChangePreferredAzFrom(old_az),
367 2 : );
368 2 : }
369 2 : self.preferred_az_id = preferred_az;
370 0 : }
371 2 : }
372 :
373 12508 : pub(crate) fn get_preferred_az(&self) -> Option<&AvailabilityZone> {
374 12508 : self.preferred_az_id.as_ref()
375 12508 : }
376 : }
377 :
378 : impl Drop for IntentState {
379 12860 : fn drop(&mut self) {
380 : // Must clear before dropping, to avoid leaving stale refcounts in the Scheduler.
381 : // We do not check this while panicking, to avoid polluting unit test failures or
382 : // other assertions with this assertion's output. It's still wrong to leak these,
383 : // but if we already have a panic then we don't need to independently flag this case.
384 12860 : if !(std::thread::panicking()) {
385 12860 : debug_assert!(self.attached.is_none() && self.secondary.is_empty());
386 0 : }
387 12859 : }
388 : }
389 :
390 0 : #[derive(Default, Clone, Serialize, Deserialize, Debug)]
391 : pub(crate) struct ObservedState {
392 : pub(crate) locations: HashMap<NodeId, ObservedStateLocation>,
393 : }
394 :
395 : /// Our latest knowledge of how this tenant is configured in the outside world.
396 : ///
397 : /// Meaning:
398 : /// * No instance of this type exists for a node: we are certain that we have nothing configured on that
399 : /// node for this shard.
400 : /// * Instance exists with conf==None: we *might* have some state on that node, but we don't know
401 : /// what it is (e.g. we failed partway through configuring it)
402 : /// * Instance exists with conf==Some: this tells us what we last successfully configured on this node,
403 : /// and that configuration will still be present unless something external interfered.
404 0 : #[derive(Clone, Serialize, Deserialize, Debug)]
405 : pub(crate) struct ObservedStateLocation {
406 : /// If None, it means we do not know the status of this shard's location on this node, but
407 : /// we know that we might have some state on this node.
408 : pub(crate) conf: Option<LocationConfig>,
409 : }
410 :
411 : pub(crate) struct ReconcilerWaiter {
412 : // For observability purposes, remember the ID of the shard we're
413 : // waiting for.
414 : pub(crate) tenant_shard_id: TenantShardId,
415 :
416 : seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
417 : error_seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
418 : error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
419 : seq: Sequence,
420 : }
421 :
422 : pub(crate) enum ReconcilerStatus {
423 : Done,
424 : Failed,
425 : InProgress,
426 : }
427 :
428 : #[derive(thiserror::Error, Debug)]
429 : pub(crate) enum ReconcileWaitError {
430 : #[error("Timeout waiting for shard {0}")]
431 : Timeout(TenantShardId),
432 : #[error("shutting down")]
433 : Shutdown,
434 : #[error("Reconcile error on shard {0}: {1}")]
435 : Failed(TenantShardId, Arc<ReconcileError>),
436 : }
437 :
438 : #[derive(Eq, PartialEq, Debug, Clone)]
439 : pub(crate) struct ReplaceSecondary {
440 : old_node_id: NodeId,
441 : new_node_id: NodeId,
442 : }
443 :
444 : #[derive(Eq, PartialEq, Debug, Clone)]
445 : pub(crate) struct MigrateAttachment {
446 : pub(crate) old_attached_node_id: NodeId,
447 : pub(crate) new_attached_node_id: NodeId,
448 : }
449 :
450 : #[derive(Eq, PartialEq, Debug, Clone)]
451 : pub(crate) enum ScheduleOptimizationAction {
452 : // Replace one of our secondary locations with a different node
453 : ReplaceSecondary(ReplaceSecondary),
454 : // Migrate attachment to an existing secondary location
455 : MigrateAttachment(MigrateAttachment),
456 : // Create a secondary location, with the intent of later migrating to it
457 : CreateSecondary(NodeId),
458 : // Remove a secondary location that we previously created to facilitate a migration
459 : RemoveSecondary(NodeId),
460 : }
461 :
462 : #[derive(Eq, PartialEq, Debug, Clone)]
463 : pub(crate) struct ScheduleOptimization {
464 : // What was the reconcile sequence when we generated this optimization? The optimization
465 : // should only be applied if the shard's sequence is still at this value, in case other changes
466 : // happened between planning the optimization and applying it.
467 : sequence: Sequence,
468 :
469 : pub(crate) action: ScheduleOptimizationAction,
470 : }
471 :
472 : impl ReconcilerWaiter {
473 0 : pub(crate) async fn wait_timeout(&self, timeout: Duration) -> Result<(), ReconcileWaitError> {
474 0 : tokio::select! {
475 0 : result = self.seq_wait.wait_for_timeout(self.seq, timeout)=> {
476 0 : result.map_err(|e| match e {
477 0 : SeqWaitError::Timeout => ReconcileWaitError::Timeout(self.tenant_shard_id),
478 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown
479 0 : })?;
480 : },
481 0 : result = self.error_seq_wait.wait_for(self.seq) => {
482 0 : result.map_err(|e| match e {
483 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown,
484 0 : SeqWaitError::Timeout => unreachable!()
485 0 : })?;
486 :
487 0 : return Err(ReconcileWaitError::Failed(self.tenant_shard_id,
488 0 : self.error.lock().unwrap().clone().expect("If error_seq_wait was advanced error was set").clone()))
489 : }
490 : }
491 :
492 0 : Ok(())
493 0 : }
494 :
495 0 : pub(crate) fn get_status(&self) -> ReconcilerStatus {
496 0 : if self.seq_wait.would_wait_for(self.seq).is_ok() {
497 0 : ReconcilerStatus::Done
498 0 : } else if self.error_seq_wait.would_wait_for(self.seq).is_ok() {
499 0 : ReconcilerStatus::Failed
500 : } else {
501 0 : ReconcilerStatus::InProgress
502 : }
503 0 : }
504 : }
505 :
506 : /// Having spawned a reconciler task, the tenant shard's state will carry enough
507 : /// information to optionally cancel & await it later.
508 : pub(crate) struct ReconcilerHandle {
509 : sequence: Sequence,
510 : handle: JoinHandle<()>,
511 : cancel: CancellationToken,
512 : }
513 :
514 : pub(crate) enum ReconcileNeeded {
515 : /// shard either doesn't need reconciliation, or is forbidden from spawning a reconciler
516 : /// in its current state (e.g. shard split in progress, or ShardSchedulingPolicy forbids it)
517 : No,
518 : /// shard has a reconciler running, and its intent hasn't changed since that one was
519 : /// spawned: wait for the existing reconciler rather than spawning a new one.
520 : WaitExisting(ReconcilerWaiter),
521 : /// shard needs reconciliation: call into [`TenantShard::spawn_reconciler`]
522 : Yes(ReconcileReason),
523 : }
524 :
525 : #[derive(Debug)]
526 : pub(crate) enum ReconcileReason {
527 : ActiveNodesDirty,
528 : UnknownLocation,
529 : PendingComputeNotification,
530 : }
531 :
532 : /// Pending modification to the observed state of a tenant shard.
533 : /// Produced by [`Reconciler::observed_deltas`] and applied in [`crate::service::Service::process_result`].
534 : pub(crate) enum ObservedStateDelta {
535 : Upsert(Box<(NodeId, ObservedStateLocation)>),
536 : Delete(NodeId),
537 : }
538 :
539 : impl ObservedStateDelta {
540 0 : pub(crate) fn node_id(&self) -> &NodeId {
541 0 : match self {
542 0 : Self::Upsert(up) => &up.0,
543 0 : Self::Delete(nid) => nid,
544 : }
545 0 : }
546 : }
547 :
548 : /// When a reconcile task completes, it sends this result object
549 : /// to be applied to the primary TenantShard.
550 : pub(crate) struct ReconcileResult {
551 : pub(crate) sequence: Sequence,
552 : /// On errors, `observed` should be treated as an incompleted description
553 : /// of state (i.e. any nodes present in the result should override nodes
554 : /// present in the parent tenant state, but any unmentioned nodes should
555 : /// not be removed from parent tenant state)
556 : pub(crate) result: Result<(), ReconcileError>,
557 :
558 : pub(crate) tenant_shard_id: TenantShardId,
559 : pub(crate) generation: Option<Generation>,
560 : pub(crate) observed_deltas: Vec<ObservedStateDelta>,
561 :
562 : /// Set [`TenantShard::pending_compute_notification`] from this flag
563 : pub(crate) pending_compute_notification: bool,
564 : }
565 :
566 : impl ObservedState {
567 0 : pub(crate) fn new() -> Self {
568 0 : Self {
569 0 : locations: HashMap::new(),
570 0 : }
571 0 : }
572 :
573 0 : pub(crate) fn is_empty(&self) -> bool {
574 0 : self.locations.is_empty()
575 0 : }
576 : }
577 :
578 : impl TenantShard {
579 12843 : pub(crate) fn new(
580 12843 : tenant_shard_id: TenantShardId,
581 12843 : shard: ShardIdentity,
582 12843 : policy: PlacementPolicy,
583 12843 : preferred_az_id: Option<AvailabilityZone>,
584 12843 : ) -> Self {
585 12843 : metrics::METRICS_REGISTRY
586 12843 : .metrics_group
587 12843 : .storage_controller_tenant_shards
588 12843 : .inc();
589 :
590 12843 : Self {
591 12843 : tenant_shard_id,
592 12843 : policy,
593 12843 : intent: IntentState::new(preferred_az_id),
594 12843 : generation: Some(Generation::new(0)),
595 12843 : shard,
596 12843 : observed: ObservedState::default(),
597 12843 : config: TenantConfig::default(),
598 12843 : reconciler: None,
599 12843 : splitting: SplitState::Idle,
600 12843 : importing: TimelineImportState::Idle,
601 12843 : sequence: Sequence(1),
602 12843 : delayed_reconcile: false,
603 12843 : waiter: Arc::new(SeqWait::new(Sequence(0))),
604 12843 : error_waiter: Arc::new(SeqWait::new(Sequence(0))),
605 12843 : last_error: Arc::default(),
606 12843 : consecutive_errors_count: 0,
607 12843 : pending_compute_notification: false,
608 12843 : scheduling_policy: ShardSchedulingPolicy::default(),
609 12843 : preferred_node: None,
610 12843 : }
611 12843 : }
612 :
613 : /// For use on startup when learning state from pageservers: generate my [`IntentState`] from my
614 : /// [`ObservedState`], even if it violates my [`PlacementPolicy`]. Call [`Self::schedule`] next,
615 : /// to get an intent state that complies with placement policy. The overall goal is to do scheduling
616 : /// in a way that makes use of any configured locations that already exist in the outside world.
617 1 : pub(crate) fn intent_from_observed(&mut self, scheduler: &mut Scheduler) {
618 : // Choose an attached location by filtering observed locations, and then sorting to get the highest
619 : // generation
620 1 : let mut attached_locs = self
621 1 : .observed
622 1 : .locations
623 1 : .iter()
624 2 : .filter_map(|(node_id, l)| {
625 2 : if let Some(conf) = &l.conf {
626 2 : if conf.mode == LocationConfigMode::AttachedMulti
627 1 : || conf.mode == LocationConfigMode::AttachedSingle
628 1 : || conf.mode == LocationConfigMode::AttachedStale
629 : {
630 2 : Some((node_id, conf.generation))
631 : } else {
632 0 : None
633 : }
634 : } else {
635 0 : None
636 : }
637 2 : })
638 1 : .collect::<Vec<_>>();
639 :
640 1 : attached_locs.sort_by_key(|i| i.1);
641 1 : if let Some((node_id, _gen)) = attached_locs.into_iter().next_back() {
642 1 : self.intent.set_attached(scheduler, Some(*node_id));
643 1 : }
644 :
645 : // All remaining observed locations generate secondary intents. This includes None
646 : // observations, as these may well have some local content on disk that is usable (this
647 : // is an edge case that might occur if we restarted during a migration or other change)
648 : //
649 : // We may leave intent.attached empty if we didn't find any attached locations: [`Self::schedule`]
650 : // will take care of promoting one of these secondaries to be attached.
651 2 : self.observed.locations.keys().for_each(|node_id| {
652 2 : if Some(*node_id) != self.intent.attached {
653 1 : self.intent.push_secondary(scheduler, *node_id);
654 1 : }
655 2 : });
656 1 : }
657 :
658 : /// Part of [`Self::schedule`] that is used to choose exactly one node to act as the
659 : /// attached pageserver for a shard.
660 : ///
661 : /// Returns whether we modified it, and the NodeId selected.
662 12832 : fn schedule_attached(
663 12832 : &mut self,
664 12832 : scheduler: &mut Scheduler,
665 12832 : context: &ScheduleContext,
666 12832 : ) -> Result<(bool, NodeId), ScheduleError> {
667 : // No work to do if we already have an attached tenant
668 12832 : if let Some(node_id) = self.intent.attached {
669 0 : return Ok((false, node_id));
670 12832 : }
671 :
672 12832 : if let Some(promote_secondary) = self.preferred_secondary(scheduler) {
673 : // Promote a secondary
674 2 : tracing::debug!("Promoted secondary {} to attached", promote_secondary);
675 2 : self.intent.promote_attached(scheduler, promote_secondary);
676 2 : Ok((true, promote_secondary))
677 : } else {
678 : // Pick a fresh node: either we had no secondaries or none were schedulable
679 12830 : let node_id = scheduler.schedule_shard::<AttachedShardTag>(
680 12830 : &self.intent.secondary,
681 12830 : &self.intent.preferred_az_id,
682 12830 : context,
683 0 : )?;
684 12830 : tracing::debug!("Selected {} as attached", node_id);
685 12830 : self.intent.set_attached(scheduler, Some(node_id));
686 12830 : Ok((true, node_id))
687 : }
688 12832 : }
689 :
690 : #[instrument(skip_all, fields(
691 : tenant_id=%self.tenant_shard_id.tenant_id,
692 : shard_id=%self.tenant_shard_id.shard_slug(),
693 : sequence=%self.sequence
694 : ))]
695 : pub(crate) fn schedule(
696 : &mut self,
697 : scheduler: &mut Scheduler,
698 : context: &mut ScheduleContext,
699 : ) -> Result<(), ScheduleError> {
700 : let r = self.do_schedule(scheduler, context);
701 :
702 : context.avoid(&self.intent.all_pageservers());
703 :
704 : r
705 : }
706 :
707 12836 : pub(crate) fn do_schedule(
708 12836 : &mut self,
709 12836 : scheduler: &mut Scheduler,
710 12836 : context: &ScheduleContext,
711 12836 : ) -> Result<(), ScheduleError> {
712 : // TODO: before scheduling new nodes, check if any existing content in
713 : // self.intent refers to pageservers that are offline, and pick other
714 : // pageservers if so.
715 :
716 : // TODO: respect the splitting bit on tenants: if they are currently splitting then we may not
717 : // change their attach location.
718 :
719 12836 : match self.scheduling_policy {
720 12835 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {}
721 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
722 : // Warn to make it obvious why other things aren't happening/working, if we skip scheduling
723 1 : tracing::warn!(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(),
724 0 : "Scheduling is disabled by policy {:?}", self.scheduling_policy);
725 1 : return Ok(());
726 : }
727 : }
728 :
729 : // Build the set of pageservers already in use by this tenant, to avoid scheduling
730 : // more work on the same pageservers we're already using.
731 12835 : let mut modified = false;
732 :
733 : // Add/remove nodes to fulfil policy
734 : use PlacementPolicy::*;
735 12835 : match self.policy {
736 12832 : Attached(secondary_count) => {
737 : // Should have exactly one attached, and at least N secondaries
738 12832 : let (modified_attached, attached_node_id) =
739 12832 : self.schedule_attached(scheduler, context)?;
740 12832 : modified |= modified_attached;
741 :
742 12832 : let mut used_pageservers = vec![attached_node_id];
743 25663 : while self.intent.secondary.len() < secondary_count {
744 12831 : let node_id = scheduler.schedule_shard::<SecondaryShardTag>(
745 12831 : &used_pageservers,
746 12831 : &self.intent.preferred_az_id,
747 12831 : context,
748 0 : )?;
749 12831 : self.intent.push_secondary(scheduler, node_id);
750 12831 : used_pageservers.push(node_id);
751 12831 : modified = true;
752 : }
753 : }
754 : Secondary => {
755 3 : if let Some(node_id) = self.intent.get_attached() {
756 2 : // Populate secondary by demoting the attached node
757 2 : self.intent.demote_attached(scheduler, *node_id);
758 2 :
759 2 : modified = true;
760 2 : } else if self.intent.secondary.is_empty() {
761 : // Populate secondary by scheduling a fresh node
762 : //
763 : // We use [`AttachedShardTag`] because when a secondary location is the only one
764 : // a shard has, we expect that its next use will be as an attached location: we want
765 : // the tenant to be ready to warm up and run fast in their preferred AZ.
766 1 : let node_id = scheduler.schedule_shard::<AttachedShardTag>(
767 1 : &[],
768 1 : &self.intent.preferred_az_id,
769 1 : context,
770 0 : )?;
771 1 : self.intent.push_secondary(scheduler, node_id);
772 1 : modified = true;
773 0 : }
774 4 : while self.intent.secondary.len() > 1 {
775 : // If we have multiple secondaries (e.g. when transitioning from Attached to Secondary and
776 : // having just demoted our attached location), then we should prefer to keep the location
777 : // in our preferred AZ. Tenants in Secondary mode want to be in the preferred AZ so that
778 : // they have a warm location to become attached when transitioning back into Attached.
779 :
780 1 : let mut candidates = self.intent.get_secondary().clone();
781 : // Sort to get secondaries outside preferred AZ last
782 1 : candidates
783 2 : .sort_by_key(|n| scheduler.get_node_az(n).as_ref() != self.preferred_az());
784 1 : let secondary_to_remove = candidates.pop().unwrap();
785 1 : self.intent.remove_secondary(scheduler, secondary_to_remove);
786 1 : modified = true;
787 : }
788 : }
789 : Detached => {
790 : // Never add locations in this mode
791 0 : if self.intent.get_attached().is_some() || !self.intent.get_secondary().is_empty() {
792 0 : self.intent.clear(scheduler);
793 0 : modified = true;
794 0 : }
795 : }
796 : }
797 :
798 12835 : if modified {
799 12835 : self.sequence.0 += 1;
800 12835 : }
801 :
802 12835 : Ok(())
803 12836 : }
804 :
805 : /// Reschedule this tenant shard to one of its secondary locations. Returns a scheduling error
806 : /// if the swap is not possible and leaves the intent state in its original state.
807 : ///
808 : /// Arguments:
809 : /// `attached_to`: the currently attached location matching the intent state (may be None if the
810 : /// shard is not attached)
811 : /// `promote_to`: an optional secondary location of this tenant shard. If set to None, we ask
812 : /// the scheduler to recommend a node
813 0 : pub(crate) fn reschedule_to_secondary(
814 0 : &mut self,
815 0 : promote_to: Option<NodeId>,
816 0 : scheduler: &mut Scheduler,
817 0 : ) -> Result<(), ScheduleError> {
818 0 : let promote_to = match promote_to {
819 0 : Some(node) => node,
820 0 : None => match self.preferred_secondary(scheduler) {
821 0 : Some(node) => node,
822 : None => {
823 0 : return Err(ScheduleError::ImpossibleConstraint);
824 : }
825 : },
826 : };
827 :
828 0 : assert!(self.intent.get_secondary().contains(&promote_to));
829 :
830 0 : if let Some(node) = self.intent.get_attached() {
831 0 : let demoted = self.intent.demote_attached(scheduler, *node);
832 0 : if !demoted {
833 0 : return Err(ScheduleError::ImpossibleConstraint);
834 0 : }
835 0 : }
836 :
837 0 : self.intent.promote_attached(scheduler, promote_to);
838 :
839 : // Increment the sequence number for the edge case where a
840 : // reconciler is already running to avoid waiting on the
841 : // current reconcile instead of spawning a new one.
842 0 : self.sequence = self.sequence.next();
843 :
844 0 : Ok(())
845 0 : }
846 :
847 : /// Returns None if the current location's score is unavailable, i.e. cannot draw a conclusion
848 68 : fn is_better_location<T: ShardTag>(
849 68 : &self,
850 68 : scheduler: &mut Scheduler,
851 68 : schedule_context: &ScheduleContext,
852 68 : current: NodeId,
853 68 : candidate: NodeId,
854 68 : ) -> Option<bool> {
855 68 : let Some(candidate_score) = scheduler.compute_node_score::<T::Score>(
856 68 : candidate,
857 68 : &self.intent.preferred_az_id,
858 68 : schedule_context,
859 68 : ) else {
860 : // The candidate node is unavailable for scheduling or otherwise couldn't get a score
861 1 : return None;
862 : };
863 :
864 : // If the candidate is our preferred node, then it is better than the current location, as long
865 : // as it is online -- the online check is part of the score calculation we did above, so it's
866 : // important that this check comes after that one.
867 67 : if let Some(preferred) = self.preferred_node.as_ref() {
868 3 : if preferred == &candidate {
869 1 : return Some(true);
870 2 : }
871 64 : }
872 :
873 66 : match scheduler.compute_node_score::<T::Score>(
874 66 : current,
875 66 : &self.intent.preferred_az_id,
876 66 : schedule_context,
877 66 : ) {
878 66 : Some(current_score) => {
879 : // Ignore utilization components when comparing scores: we don't want to migrate
880 : // because of transient load variations, it risks making the system thrash, and
881 : // migrating for utilization requires a separate high level view of the system to
882 : // e.g. prioritize moving larger or smaller tenants, rather than arbitrarily
883 : // moving things around in the order that we hit this function.
884 66 : let candidate_score = candidate_score.for_optimization();
885 66 : let current_score = current_score.for_optimization();
886 :
887 66 : if candidate_score < current_score {
888 8 : tracing::info!(
889 0 : "Found a lower scoring location! {candidate} is better than {current} ({candidate_score:?} is better than {current_score:?})"
890 : );
891 8 : Some(true)
892 : } else {
893 : // The candidate node is no better than our current location, so don't migrate
894 58 : tracing::debug!(
895 0 : "Candidate node {candidate} is no better than our current location {current} (candidate {candidate_score:?} vs current {current_score:?})",
896 : );
897 58 : Some(false)
898 : }
899 : }
900 : None => {
901 : // The current node is unavailable for scheduling, so we can't make any sensible
902 : // decisions about optimisation. This should be a transient state -- if the node
903 : // is offline then it will get evacuated, if is blocked by a scheduling mode
904 : // then we will respect that mode by doing nothing.
905 0 : tracing::debug!("Current node {current} is unavailable for scheduling");
906 0 : None
907 : }
908 : }
909 68 : }
910 :
911 48 : pub(crate) fn find_better_location<T: ShardTag>(
912 48 : &self,
913 48 : scheduler: &mut Scheduler,
914 48 : schedule_context: &ScheduleContext,
915 48 : current: NodeId,
916 48 : hard_exclude: &[NodeId],
917 48 : ) -> Option<NodeId> {
918 : // If we have a migration hint, then that is our better location
919 48 : if let Some(hint) = self.preferred_node.as_ref() {
920 1 : if hint == ¤t {
921 0 : return None;
922 1 : }
923 :
924 1 : return Some(*hint);
925 47 : }
926 :
927 : // Look for a lower-scoring location to attach to
928 47 : let Ok(candidate_node) = scheduler.schedule_shard::<T>(
929 47 : hard_exclude,
930 47 : &self.intent.preferred_az_id,
931 47 : schedule_context,
932 47 : ) else {
933 : // A scheduling error means we have no possible candidate replacements
934 0 : tracing::debug!("No candidate node found");
935 0 : return None;
936 : };
937 :
938 47 : if candidate_node == current {
939 : // We're already at the best possible location, so don't migrate
940 24 : tracing::debug!("Candidate node {candidate_node} is already in use");
941 24 : return None;
942 23 : }
943 :
944 23 : self.is_better_location::<T>(scheduler, schedule_context, current, candidate_node)
945 23 : .and_then(|better| if better { Some(candidate_node) } else { None })
946 48 : }
947 :
948 : /// This function is an optimization, used to avoid doing large numbers of scheduling operations
949 : /// when looking for optimizations. This function uses knowledge of how scores work to do some
950 : /// fast checks for whether it may to be possible to improve a score.
951 : ///
952 : /// If we return true, it only means that optimization _might_ be possible, not that it necessarily is. If we
953 : /// return no, it definitely means that calling [`Self::optimize_attachment`] or [`Self::optimize_secondary`] would do no
954 : /// work.
955 3 : pub(crate) fn maybe_optimizable(
956 3 : &self,
957 3 : scheduler: &mut Scheduler,
958 3 : schedule_context: &ScheduleContext,
959 3 : ) -> bool {
960 : // Tenant with preferred node: check if it is not already at the preferred node
961 3 : if let Some(preferred) = self.preferred_node.as_ref() {
962 0 : if Some(preferred) != self.intent.get_attached().as_ref() {
963 0 : return true;
964 0 : }
965 3 : }
966 :
967 : // Sharded tenant: check if any locations have a nonzero affinity score
968 3 : if self.shard.count >= ShardCount(1) {
969 3 : let schedule_context = schedule_context.project_detach(self);
970 5 : for node in self.intent.all_pageservers() {
971 5 : if let Some(af) = schedule_context.nodes.get(&node) {
972 5 : if *af > AffinityScore(0) {
973 3 : return true;
974 2 : }
975 0 : }
976 : }
977 0 : }
978 :
979 : // Attached tenant: check if the attachment is outside the preferred AZ
980 0 : if let PlacementPolicy::Attached(_) = self.policy {
981 0 : if let Some(attached) = self.intent.get_attached() {
982 0 : if scheduler.get_node_az(attached) != self.intent.preferred_az_id {
983 0 : return true;
984 0 : }
985 0 : }
986 0 : }
987 :
988 : // Tenant with secondary locations: check if any are within the preferred AZ
989 0 : for secondary in self.intent.get_secondary() {
990 0 : if scheduler.get_node_az(secondary) == self.intent.preferred_az_id {
991 0 : return true;
992 0 : }
993 : }
994 :
995 : // Does the tenant have excess secondaries?
996 0 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
997 0 : return true;
998 0 : }
999 :
1000 : // Fall through: no optimizations possible
1001 0 : false
1002 3 : }
1003 :
1004 : /// Optimize attachments: if a shard has a secondary location that is preferable to
1005 : /// its primary location based on soft constraints, switch that secondary location
1006 : /// to be attached.
1007 : ///
1008 : /// `schedule_context` should have been populated with all shards in the tenant, including
1009 : /// the one we're trying to optimize (this function will subtract its own contribution before making scoring decisions)
1010 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1011 : pub(crate) fn optimize_attachment(
1012 : &self,
1013 : scheduler: &mut Scheduler,
1014 : schedule_context: &ScheduleContext,
1015 : ) -> Option<ScheduleOptimization> {
1016 : let attached = (*self.intent.get_attached())?;
1017 :
1018 : let schedule_context = schedule_context.project_detach(self);
1019 :
1020 : // If we already have a secondary that is higher-scoring than out current location,
1021 : // then simply migrate to it.
1022 : for secondary in self.intent.get_secondary() {
1023 : if let Some(true) = self.is_better_location::<AttachedShardTag>(
1024 : scheduler,
1025 : &schedule_context,
1026 : attached,
1027 : *secondary,
1028 : ) {
1029 : return Some(ScheduleOptimization {
1030 : sequence: self.sequence,
1031 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1032 : old_attached_node_id: attached,
1033 : new_attached_node_id: *secondary,
1034 : }),
1035 : });
1036 : }
1037 : }
1038 :
1039 : // Given that none of our current secondaries is a better location than our current
1040 : // attached location (checked above), we may trim any secondaries that are not needed
1041 : // for the placement policy.
1042 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
1043 : // This code path cleans up extra secondaries after migrating, and/or
1044 : // trims extra secondaries after a PlacementPolicy::Attached(N) was
1045 : // modified to decrease N.
1046 :
1047 : let secondary_scores = self
1048 : .intent
1049 : .get_secondary()
1050 : .iter()
1051 12 : .map(|node_id| {
1052 12 : (
1053 12 : *node_id,
1054 12 : scheduler.compute_node_score::<NodeSecondarySchedulingScore>(
1055 12 : *node_id,
1056 12 : &self.intent.preferred_az_id,
1057 12 : &schedule_context,
1058 12 : ),
1059 12 : )
1060 12 : })
1061 : .collect::<HashMap<_, _>>();
1062 :
1063 11 : if secondary_scores.iter().any(|score| score.1.is_none()) {
1064 : // Trivial case: if we only have one secondary, drop that one
1065 : if self.intent.get_secondary().len() == 1 {
1066 : return Some(ScheduleOptimization {
1067 : sequence: self.sequence,
1068 : action: ScheduleOptimizationAction::RemoveSecondary(
1069 : *self.intent.get_secondary().first().unwrap(),
1070 : ),
1071 : });
1072 : }
1073 :
1074 : // Try to find a "good" secondary to keep, without relying on scores (one or more nodes is in a state
1075 : // where its score can't be calculated), and drop the others. This enables us to make progress in
1076 : // most cases, even if some nodes are offline or have scheduling=pause set.
1077 :
1078 : debug_assert!(self.intent.attached.is_some()); // We should not make it here unless attached -- this
1079 : // logic presumes we are in a mode where we want secondaries to be in non-home AZ
1080 1 : if let Some(retain_secondary) = self.intent.get_secondary().iter().find(|n| {
1081 1 : let in_home_az = scheduler.get_node_az(n) == self.intent.preferred_az_id;
1082 1 : let is_available = secondary_scores
1083 1 : .get(n)
1084 1 : .expect("Built from same list of nodes")
1085 1 : .is_some();
1086 1 : is_available && !in_home_az
1087 1 : }) {
1088 : // Great, we found one to retain. Pick some other to drop.
1089 : if let Some(victim) = self
1090 : .intent
1091 : .get_secondary()
1092 : .iter()
1093 2 : .find(|n| n != &retain_secondary)
1094 : {
1095 : return Some(ScheduleOptimization {
1096 : sequence: self.sequence,
1097 : action: ScheduleOptimizationAction::RemoveSecondary(*victim),
1098 : });
1099 : }
1100 : }
1101 :
1102 : // Fall through: we didn't identify one to remove. This ought to be rare.
1103 : tracing::warn!(
1104 : "Keeping extra secondaries: can't determine which of {:?} to remove (some nodes offline?)",
1105 : self.intent.get_secondary()
1106 : );
1107 : } else {
1108 : let victim = secondary_scores
1109 : .iter()
1110 10 : .max_by_key(|score| score.1.unwrap())
1111 : .unwrap()
1112 : .0;
1113 : return Some(ScheduleOptimization {
1114 : sequence: self.sequence,
1115 : action: ScheduleOptimizationAction::RemoveSecondary(*victim),
1116 : });
1117 : }
1118 : }
1119 :
1120 : let replacement = self.find_better_location::<AttachedShardTag>(
1121 : scheduler,
1122 : &schedule_context,
1123 : attached,
1124 : &[], // Don't exclude secondaries: our preferred attachment location may be a secondary
1125 : );
1126 :
1127 : // We have found a candidate and confirmed that its score is preferable
1128 : // to our current location. See if we have a secondary location in the preferred location already: if not,
1129 : // then create one.
1130 : if let Some(replacement) = replacement {
1131 : // If we are currently in non-preferred AZ, then the scheduler might suggest a location that is better, but still
1132 : // not in our preferred AZ. Migration has a cost in resources an impact to the workload, so we want to avoid doing
1133 : // multiple hops where we might go to some other AZ before eventually finding a suitable location in our preferred
1134 : // AZ: skip this optimization if it is not in our final, preferred AZ.
1135 : //
1136 : // This should be a transient state, there should always be capacity eventually in our preferred AZ (even if nodes
1137 : // there are too overloaded for scheduler to suggest them, more should be provisioned eventually).
1138 : if self.preferred_node.is_none()
1139 : && self.intent.preferred_az_id.is_some()
1140 : && scheduler.get_node_az(&replacement) != self.intent.preferred_az_id
1141 : {
1142 : tracing::debug!(
1143 : "Candidate node {replacement} is not in preferred AZ {:?}",
1144 : self.intent.preferred_az_id
1145 : );
1146 :
1147 : // This should only happen if our current location is not in the preferred AZ, otherwise
1148 : // [`Self::find_better_location`]` should have rejected any other location outside the preferred Az, because
1149 : // AZ is the highest priority part of NodeAttachmentSchedulingScore.
1150 : debug_assert!(scheduler.get_node_az(&attached) != self.intent.preferred_az_id);
1151 :
1152 : return None;
1153 : }
1154 :
1155 : if !self.intent.get_secondary().contains(&replacement) {
1156 : Some(ScheduleOptimization {
1157 : sequence: self.sequence,
1158 : action: ScheduleOptimizationAction::CreateSecondary(replacement),
1159 : })
1160 : } else {
1161 : // We already have a secondary in the preferred location, let's try migrating to it. Our caller
1162 : // will check the warmth of the destination before deciding whether to really execute this.
1163 : Some(ScheduleOptimization {
1164 : sequence: self.sequence,
1165 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1166 : old_attached_node_id: attached,
1167 : new_attached_node_id: replacement,
1168 : }),
1169 : })
1170 : }
1171 : } else {
1172 : // We didn't find somewhere we'd rather be, and we don't have any excess secondaries
1173 : // to clean up: no action required.
1174 : None
1175 : }
1176 : }
1177 :
1178 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1179 : pub(crate) fn optimize_secondary(
1180 : &self,
1181 : scheduler: &mut Scheduler,
1182 : schedule_context: &ScheduleContext,
1183 : ) -> Option<ScheduleOptimization> {
1184 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
1185 : // We have extra secondaries, perhaps to facilitate a migration of the attached location:
1186 : // do nothing, it is up to [`Self::optimize_attachment`] to clean them up. When that's done,
1187 : // and we are called again, we will proceed.
1188 : tracing::debug!("Too many secondaries: skipping");
1189 : return None;
1190 : }
1191 :
1192 : let schedule_context = schedule_context.project_detach(self);
1193 :
1194 : for secondary in self.intent.get_secondary() {
1195 : // Make sure we don't try to migrate a secondary to our attached location: this case happens
1196 : // easily in environments without multiple AZs.
1197 : let mut exclude = match self.intent.attached {
1198 : Some(attached) => vec![attached],
1199 : None => vec![],
1200 : };
1201 :
1202 : // Exclude all other secondaries from the scheduling process to avoid replacing
1203 : // one existing secondary with another existing secondary.
1204 : for another_secondary in self.intent.secondary.iter() {
1205 : if another_secondary != secondary {
1206 : exclude.push(*another_secondary);
1207 : }
1208 : }
1209 :
1210 : let replacement = match &self.policy {
1211 : PlacementPolicy::Attached(_) => {
1212 : // Secondaries for an attached shard should be scheduled using `SecondaryShardTag`
1213 : // to avoid placing them in the preferred AZ.
1214 : self.find_better_location::<SecondaryShardTag>(
1215 : scheduler,
1216 : &schedule_context,
1217 : *secondary,
1218 : &exclude,
1219 : )
1220 : }
1221 : PlacementPolicy::Secondary => {
1222 : // In secondary-only mode, we want our secondary locations in the preferred AZ,
1223 : // so that they're ready to take over as an attached location when we transition
1224 : // into PlacementPolicy::Attached.
1225 : self.find_better_location::<AttachedShardTag>(
1226 : scheduler,
1227 : &schedule_context,
1228 : *secondary,
1229 : &exclude,
1230 : )
1231 : }
1232 : PlacementPolicy::Detached => None,
1233 : };
1234 :
1235 : assert!(replacement != Some(*secondary));
1236 : if let Some(replacement) = replacement {
1237 : // We have found a candidate and confirmed that its score is preferable
1238 : // to our current location. See if we have a secondary location in the preferred location already: if not,
1239 : // then create one.
1240 : return Some(ScheduleOptimization {
1241 : sequence: self.sequence,
1242 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1243 : old_node_id: *secondary,
1244 : new_node_id: replacement,
1245 : }),
1246 : });
1247 : }
1248 : }
1249 :
1250 : None
1251 : }
1252 :
1253 : /// Start or abort a graceful migration of this shard to another pageserver. This works on top of the
1254 : /// other optimisation functions, to bias them to move to the destination node.
1255 0 : pub(crate) fn set_preferred_node(&mut self, node: Option<NodeId>) {
1256 0 : if let Some(hint) = self.preferred_node.as_ref() {
1257 0 : if Some(hint) != node.as_ref() {
1258 : // This is legal but a bit surprising: we expect that administrators wouldn't usually
1259 : // change their mind about where to migrate something.
1260 0 : tracing::warn!(
1261 0 : "Changing migration destination from {hint} to {node:?} (current intent {:?})",
1262 : self.intent
1263 : );
1264 0 : }
1265 0 : }
1266 :
1267 0 : self.preferred_node = node;
1268 0 : }
1269 :
1270 0 : pub(crate) fn get_preferred_node(&self) -> Option<NodeId> {
1271 0 : self.preferred_node
1272 0 : }
1273 :
1274 : /// Return true if the optimization was really applied: it will not be applied if the optimization's
1275 : /// sequence is behind this tenant shard's
1276 17 : pub(crate) fn apply_optimization(
1277 17 : &mut self,
1278 17 : scheduler: &mut Scheduler,
1279 17 : optimization: ScheduleOptimization,
1280 17 : ) -> bool {
1281 17 : if optimization.sequence != self.sequence {
1282 0 : return false;
1283 17 : }
1284 :
1285 17 : metrics::METRICS_REGISTRY
1286 17 : .metrics_group
1287 17 : .storage_controller_schedule_optimization
1288 17 : .inc();
1289 :
1290 17 : match optimization.action {
1291 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1292 5 : old_attached_node_id,
1293 5 : new_attached_node_id,
1294 : }) => {
1295 5 : self.intent.demote_attached(scheduler, old_attached_node_id);
1296 5 : self.intent
1297 5 : .promote_attached(scheduler, new_attached_node_id);
1298 :
1299 5 : if let Some(hint) = self.preferred_node.as_ref() {
1300 1 : if hint == &new_attached_node_id {
1301 : // The migration target is not a long term pin: once we are done with the migration, clear it.
1302 1 : tracing::info!("Graceful migration to {hint} complete");
1303 1 : self.preferred_node = None;
1304 0 : }
1305 4 : }
1306 : }
1307 : ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1308 1 : old_node_id,
1309 1 : new_node_id,
1310 1 : }) => {
1311 1 : self.intent.remove_secondary(scheduler, old_node_id);
1312 1 : self.intent.push_secondary(scheduler, new_node_id);
1313 1 : }
1314 4 : ScheduleOptimizationAction::CreateSecondary(new_node_id) => {
1315 4 : self.intent.push_secondary(scheduler, new_node_id);
1316 4 : }
1317 7 : ScheduleOptimizationAction::RemoveSecondary(old_secondary) => {
1318 7 : self.intent.remove_secondary(scheduler, old_secondary);
1319 7 : }
1320 : }
1321 :
1322 17 : true
1323 17 : }
1324 :
1325 : /// When a shard has several secondary locations, we need to pick one in situations where
1326 : /// we promote one of them to an attached location:
1327 : /// - When draining a node for restart
1328 : /// - When responding to a node failure
1329 : ///
1330 : /// In this context, 'preferred' does not mean the node with the best scheduling score: instead
1331 : /// we want to pick the node which is best for use _temporarily_ while the previous attached location
1332 : /// is unavailable (e.g. because it's down or deploying). That means we prefer to use secondary
1333 : /// locations in a non-preferred AZ, as they're more likely to have awarm cache than a temporary
1334 : /// secondary in the preferred AZ (which are usually only created for migrations, and if they exist
1335 : /// they're probably not warmed up yet). The latter behavior is based oni
1336 : ///
1337 : /// If the input is empty, or all the nodes are not elegible for scheduling, return None: the
1338 : /// caller needs to a pick a node some other way.
1339 12832 : pub(crate) fn preferred_secondary(&self, scheduler: &Scheduler) -> Option<NodeId> {
1340 12832 : let candidates = scheduler.filter_usable_nodes(&self.intent.secondary);
1341 :
1342 : // We will sort candidates to prefer nodes which are _not_ in our preferred AZ, i.e. we prefer
1343 : // to migrate to a long-lived secondary location (which would have been scheduled in a non-preferred AZ),
1344 : // rather than a short-lived secondary location being used for optimization/migration (which would have
1345 : // been scheduled in our preferred AZ).
1346 12832 : let mut candidates = candidates
1347 12832 : .iter()
1348 12832 : .map(|(node_id, node_az)| {
1349 2 : if node_az == &self.intent.preferred_az_id {
1350 1 : (1, *node_id)
1351 : } else {
1352 1 : (0, *node_id)
1353 : }
1354 2 : })
1355 12832 : .collect::<Vec<_>>();
1356 :
1357 12832 : candidates.sort();
1358 :
1359 12832 : candidates.first().map(|i| i.1)
1360 12832 : }
1361 :
1362 : /// Query whether the tenant's observed state for attached node matches its intent state, and if so,
1363 : /// yield the node ID. This is appropriate for emitting compute hook notifications: we are checking that
1364 : /// the node in question is not only where we intend to attach, but that the tenant is indeed already attached there.
1365 : ///
1366 : /// Reconciliation may still be needed for other aspects of state such as secondaries (see [`Self::dirty`]): this
1367 : /// funciton should not be used to decide whether to reconcile.
1368 0 : pub(crate) fn stably_attached(&self) -> Option<NodeId> {
1369 : // We have an intent to attach for this node
1370 0 : let attach_intent = self.intent.attached?;
1371 : // We have an observed state for this node
1372 0 : let location = self.observed.locations.get(&attach_intent)?;
1373 : // Our observed state is not None, i.e. not in flux
1374 0 : let location_config = location.conf.as_ref()?;
1375 :
1376 : // Check if our intent and observed state agree that this node is in an attached state.
1377 0 : match location_config.mode {
1378 : LocationConfigMode::AttachedMulti
1379 : | LocationConfigMode::AttachedSingle
1380 0 : | LocationConfigMode::AttachedStale => Some(attach_intent),
1381 0 : _ => None,
1382 : }
1383 0 : }
1384 :
1385 0 : fn dirty(&self, nodes: &Arc<HashMap<NodeId, Node>>) -> bool {
1386 0 : let mut dirty_nodes = HashSet::new();
1387 :
1388 0 : if let Some(node_id) = self.intent.attached {
1389 : // Maybe panic: it is a severe bug if we try to attach while generation is null.
1390 0 : let generation = self
1391 0 : .generation
1392 0 : .expect("Attempted to enter attached state without a generation");
1393 :
1394 0 : let wanted_conf = attached_location_conf(
1395 0 : generation,
1396 0 : &self.shard,
1397 0 : &self.config,
1398 0 : &self.policy,
1399 0 : self.intent.get_secondary().len(),
1400 : );
1401 0 : match self.observed.locations.get(&node_id) {
1402 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
1403 0 : Some(_) | None => {
1404 0 : dirty_nodes.insert(node_id);
1405 0 : }
1406 : }
1407 0 : }
1408 :
1409 0 : for node_id in &self.intent.secondary {
1410 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
1411 0 : match self.observed.locations.get(node_id) {
1412 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
1413 0 : Some(_) | None => {
1414 0 : dirty_nodes.insert(*node_id);
1415 0 : }
1416 : }
1417 : }
1418 :
1419 0 : for node_id in self.observed.locations.keys() {
1420 0 : if self.intent.attached != Some(*node_id) && !self.intent.secondary.contains(node_id) {
1421 0 : // We have observed state that isn't part of our intent: need to clean it up.
1422 0 : dirty_nodes.insert(*node_id);
1423 0 : }
1424 : }
1425 :
1426 0 : dirty_nodes.retain(|node_id| {
1427 0 : nodes
1428 0 : .get(node_id)
1429 0 : .map(|n| n.is_available())
1430 0 : .unwrap_or(false)
1431 0 : });
1432 :
1433 0 : !dirty_nodes.is_empty()
1434 0 : }
1435 :
1436 : #[allow(clippy::too_many_arguments)]
1437 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1438 : pub(crate) fn get_reconcile_needed(
1439 : &mut self,
1440 : pageservers: &Arc<HashMap<NodeId, Node>>,
1441 : ) -> ReconcileNeeded {
1442 : // If there are any ambiguous observed states, and the nodes they refer to are available,
1443 : // we should reconcile to clean them up.
1444 : let mut dirty_observed = false;
1445 : for (node_id, observed_loc) in &self.observed.locations {
1446 : let node = pageservers
1447 : .get(node_id)
1448 : .expect("Nodes may not be removed while referenced");
1449 : if observed_loc.conf.is_none() && node.is_available() {
1450 : dirty_observed = true;
1451 : break;
1452 : }
1453 : }
1454 :
1455 : let active_nodes_dirty = self.dirty(pageservers);
1456 :
1457 : let reconcile_needed = match (
1458 : active_nodes_dirty,
1459 : dirty_observed,
1460 : self.pending_compute_notification,
1461 : ) {
1462 : (true, _, _) => ReconcileNeeded::Yes(ReconcileReason::ActiveNodesDirty),
1463 : (_, true, _) => ReconcileNeeded::Yes(ReconcileReason::UnknownLocation),
1464 : (_, _, true) => ReconcileNeeded::Yes(ReconcileReason::PendingComputeNotification),
1465 : _ => ReconcileNeeded::No,
1466 : };
1467 :
1468 : if matches!(reconcile_needed, ReconcileNeeded::No) {
1469 : tracing::debug!("Not dirty, no reconciliation needed.");
1470 : return ReconcileNeeded::No;
1471 : }
1472 :
1473 : // If we are currently splitting, then never start a reconciler task: the splitting logic
1474 : // requires that shards are not interfered with while it runs. Do this check here rather than
1475 : // up top, so that we only log this message if we would otherwise have done a reconciliation.
1476 : if !matches!(self.splitting, SplitState::Idle) {
1477 : tracing::info!("Refusing to reconcile, splitting in progress");
1478 : return ReconcileNeeded::No;
1479 : }
1480 :
1481 : // Reconcile already in flight for the current sequence?
1482 : if let Some(handle) = &self.reconciler {
1483 : if handle.sequence == self.sequence {
1484 : tracing::info!(
1485 : "Reconciliation already in progress for sequence {:?}",
1486 : self.sequence,
1487 : );
1488 : return ReconcileNeeded::WaitExisting(ReconcilerWaiter {
1489 : tenant_shard_id: self.tenant_shard_id,
1490 : seq_wait: self.waiter.clone(),
1491 : error_seq_wait: self.error_waiter.clone(),
1492 : error: self.last_error.clone(),
1493 : seq: self.sequence,
1494 : });
1495 : }
1496 : }
1497 :
1498 : // Pre-checks done: finally check whether we may actually do the work
1499 : match self.scheduling_policy {
1500 : ShardSchedulingPolicy::Active
1501 : | ShardSchedulingPolicy::Essential
1502 : | ShardSchedulingPolicy::Pause => {}
1503 : ShardSchedulingPolicy::Stop => {
1504 : // We only reach this point if there is work to do and we're going to skip
1505 : // doing it: warn it obvious why this tenant isn't doing what it ought to.
1506 : tracing::warn!("Skipping reconcile for policy {:?}", self.scheduling_policy);
1507 : return ReconcileNeeded::No;
1508 : }
1509 : }
1510 :
1511 : reconcile_needed
1512 : }
1513 :
1514 : /// Ensure the sequence number is set to a value where waiting for this value will make us wait
1515 : /// for the next reconcile: i.e. it is ahead of all completed or running reconcilers.
1516 : ///
1517 : /// Constructing a ReconcilerWaiter with the resulting sequence number gives the property
1518 : /// that the waiter will not complete until some future Reconciler is constructed and run.
1519 0 : fn ensure_sequence_ahead(&mut self) {
1520 : // Find the highest sequence for which a Reconciler has previously run or is currently
1521 : // running
1522 0 : let max_seen = std::cmp::max(
1523 0 : self.reconciler
1524 0 : .as_ref()
1525 0 : .map(|r| r.sequence)
1526 0 : .unwrap_or(Sequence(0)),
1527 0 : std::cmp::max(self.waiter.load(), self.error_waiter.load()),
1528 : );
1529 :
1530 0 : if self.sequence <= max_seen {
1531 0 : self.sequence = max_seen.next();
1532 0 : }
1533 0 : }
1534 :
1535 : /// Create a waiter that will wait for some future Reconciler that hasn't been spawned yet.
1536 : ///
1537 : /// This is appropriate when you can't spawn a reconciler (e.g. due to resource limits), but
1538 : /// you would like to wait on the next reconciler that gets spawned in the background.
1539 0 : pub(crate) fn future_reconcile_waiter(&mut self) -> ReconcilerWaiter {
1540 0 : self.ensure_sequence_ahead();
1541 :
1542 0 : ReconcilerWaiter {
1543 0 : tenant_shard_id: self.tenant_shard_id,
1544 0 : seq_wait: self.waiter.clone(),
1545 0 : error_seq_wait: self.error_waiter.clone(),
1546 0 : error: self.last_error.clone(),
1547 0 : seq: self.sequence,
1548 0 : }
1549 0 : }
1550 :
1551 0 : async fn reconcile(
1552 0 : sequence: Sequence,
1553 0 : mut reconciler: Reconciler,
1554 0 : must_notify: bool,
1555 0 : ) -> ReconcileResult {
1556 : // Attempt to make observed state match intent state
1557 0 : let result = reconciler.reconcile().await;
1558 :
1559 : // If we know we had a pending compute notification from some previous action, send a notification irrespective
1560 : // of whether the above reconcile() did any work. It has to be Ok() though, because otherwise we might be
1561 : // sending a notification of a location that isn't really attached.
1562 0 : if result.is_ok() && must_notify {
1563 : // If this fails we will send the need to retry in [`ReconcileResult::pending_compute_notification`]
1564 0 : reconciler.compute_notify().await.ok();
1565 0 : } else if must_notify {
1566 0 : // Carry this flag so that the reconciler's result will indicate that it still needs to retry
1567 0 : // the compute hook notification eventually.
1568 0 : reconciler.compute_notify_failure = true;
1569 0 : }
1570 :
1571 : // Update result counter
1572 0 : let outcome_label = match &result {
1573 0 : Ok(_) => ReconcileOutcome::Success,
1574 0 : Err(ReconcileError::Cancel) => ReconcileOutcome::Cancel,
1575 0 : Err(_) => ReconcileOutcome::Error,
1576 : };
1577 :
1578 0 : metrics::METRICS_REGISTRY
1579 0 : .metrics_group
1580 0 : .storage_controller_reconcile_complete
1581 0 : .inc(ReconcileCompleteLabelGroup {
1582 0 : status: outcome_label,
1583 0 : });
1584 :
1585 : // Constructing result implicitly drops Reconciler, freeing any ReconcileUnits before the Service might
1586 : // try and schedule more work in response to our result.
1587 0 : ReconcileResult {
1588 0 : sequence,
1589 0 : result,
1590 0 : tenant_shard_id: reconciler.tenant_shard_id,
1591 0 : generation: reconciler.generation,
1592 0 : observed_deltas: reconciler.observed_deltas(),
1593 0 : pending_compute_notification: reconciler.compute_notify_failure,
1594 0 : }
1595 0 : }
1596 :
1597 : #[allow(clippy::too_many_arguments)]
1598 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1599 : pub(crate) fn spawn_reconciler(
1600 : &mut self,
1601 : reason: ReconcileReason,
1602 : result_tx: &tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
1603 : pageservers: &Arc<HashMap<NodeId, Node>>,
1604 : compute_hook: &Arc<ComputeHook>,
1605 : reconciler_config: ReconcilerConfig,
1606 : service_config: &service::Config,
1607 : persistence: &Arc<Persistence>,
1608 : units: ReconcileUnits,
1609 : gate_guard: GateGuard,
1610 : cancel: &CancellationToken,
1611 : http_client: reqwest::Client,
1612 : ) -> Option<ReconcilerWaiter> {
1613 : // Reconcile in flight for a stale sequence? Our sequence's task will wait for it before
1614 : // doing our sequence's work.
1615 : let old_handle = self.reconciler.take();
1616 :
1617 : // Build list of nodes from which the reconciler should detach
1618 : let mut detach = Vec::new();
1619 : for node_id in self.observed.locations.keys() {
1620 : if self.intent.get_attached() != &Some(*node_id)
1621 : && !self.intent.secondary.contains(node_id)
1622 : {
1623 : detach.push(
1624 : pageservers
1625 : .get(node_id)
1626 : .expect("Intent references non-existent pageserver")
1627 : .clone(),
1628 : )
1629 : }
1630 : }
1631 :
1632 : // Advance the sequence before spawning a reconciler, so that sequence waiters
1633 : // can distinguish between before+after the reconcile completes.
1634 : self.ensure_sequence_ahead();
1635 :
1636 : let reconciler_cancel = cancel.child_token();
1637 : let reconciler_intent = TargetState::from_intent(pageservers, &self.intent);
1638 : let reconciler = Reconciler {
1639 : tenant_shard_id: self.tenant_shard_id,
1640 : shard: self.shard,
1641 : placement_policy: self.policy.clone(),
1642 : generation: self.generation,
1643 : intent: reconciler_intent,
1644 : detach,
1645 : reconciler_config,
1646 : config: self.config.clone(),
1647 : preferred_az: self.intent.preferred_az_id.clone(),
1648 : observed: self.observed.clone(),
1649 : original_observed: self.observed.clone(),
1650 : compute_hook: compute_hook.clone(),
1651 : service_config: service_config.clone(),
1652 : _gate_guard: gate_guard,
1653 : _resource_units: units,
1654 : cancel: reconciler_cancel.clone(),
1655 : persistence: persistence.clone(),
1656 : compute_notify_failure: false,
1657 : http_client,
1658 : };
1659 :
1660 : let reconcile_seq = self.sequence;
1661 : let long_reconcile_threshold = service_config.long_reconcile_threshold;
1662 :
1663 : tracing::info!(seq=%reconcile_seq, "Spawning Reconciler ({reason:?})");
1664 : let must_notify = self.pending_compute_notification;
1665 : let reconciler_span = tracing::info_span!(parent: None, "reconciler", seq=%reconcile_seq,
1666 : tenant_id=%reconciler.tenant_shard_id.tenant_id,
1667 : shard_id=%reconciler.tenant_shard_id.shard_slug());
1668 : metrics::METRICS_REGISTRY
1669 : .metrics_group
1670 : .storage_controller_reconcile_spawn
1671 : .inc();
1672 : let result_tx = result_tx.clone();
1673 : let join_handle = tokio::task::spawn(
1674 0 : async move {
1675 : // Wait for any previous reconcile task to complete before we start
1676 0 : if let Some(old_handle) = old_handle {
1677 0 : old_handle.cancel.cancel();
1678 0 : if let Err(e) = old_handle.handle.await {
1679 : // We can't do much with this other than log it: the task is done, so
1680 : // we may proceed with our work.
1681 0 : tracing::error!("Unexpected join error waiting for reconcile task: {e}");
1682 0 : }
1683 0 : }
1684 :
1685 : // Early check for cancellation before doing any work
1686 : // TODO: wrap all remote API operations in cancellation check
1687 : // as well.
1688 0 : if reconciler.cancel.is_cancelled() {
1689 0 : metrics::METRICS_REGISTRY
1690 0 : .metrics_group
1691 0 : .storage_controller_reconcile_complete
1692 0 : .inc(ReconcileCompleteLabelGroup {
1693 0 : status: ReconcileOutcome::Cancel,
1694 0 : });
1695 0 : return;
1696 0 : }
1697 :
1698 0 : let (tenant_id_label, shard_number_label, sequence_label) = {
1699 0 : (
1700 0 : reconciler.tenant_shard_id.tenant_id.to_string(),
1701 0 : reconciler.tenant_shard_id.shard_number.0.to_string(),
1702 0 : reconcile_seq.to_string(),
1703 0 : )
1704 0 : };
1705 :
1706 0 : let label_group = ReconcileLongRunningLabelGroup {
1707 0 : tenant_id: &tenant_id_label,
1708 0 : shard_number: &shard_number_label,
1709 0 : sequence: &sequence_label,
1710 0 : };
1711 :
1712 0 : let reconcile_fut = Self::reconcile(reconcile_seq, reconciler, must_notify);
1713 0 : let long_reconcile_fut = {
1714 0 : let label_group = label_group.clone();
1715 0 : async move {
1716 0 : tokio::time::sleep(long_reconcile_threshold).await;
1717 :
1718 0 : tracing::warn!("Reconcile passed the long running threshold of {long_reconcile_threshold:?}");
1719 :
1720 0 : metrics::METRICS_REGISTRY
1721 0 : .metrics_group
1722 0 : .storage_controller_reconcile_long_running
1723 0 : .inc(label_group);
1724 0 : }
1725 : };
1726 :
1727 0 : let reconcile_fut = std::pin::pin!(reconcile_fut);
1728 0 : let long_reconcile_fut = std::pin::pin!(long_reconcile_fut);
1729 :
1730 0 : let (was_long, result) =
1731 0 : match future::select(reconcile_fut, long_reconcile_fut).await {
1732 0 : Either::Left((reconcile_result, _)) => (false, reconcile_result),
1733 0 : Either::Right((_, reconcile_fut)) => (true, reconcile_fut.await),
1734 : };
1735 :
1736 0 : if was_long {
1737 0 : let id = metrics::METRICS_REGISTRY
1738 0 : .metrics_group
1739 0 : .storage_controller_reconcile_long_running
1740 0 : .with_labels(label_group);
1741 0 : metrics::METRICS_REGISTRY
1742 0 : .metrics_group
1743 0 : .storage_controller_reconcile_long_running
1744 0 : .remove_metric(id);
1745 0 : }
1746 :
1747 0 : result_tx
1748 0 : .send(ReconcileResultRequest::ReconcileResult(result))
1749 0 : .ok();
1750 0 : }
1751 : .instrument(reconciler_span),
1752 : );
1753 :
1754 : self.reconciler = Some(ReconcilerHandle {
1755 : sequence: self.sequence,
1756 : handle: join_handle,
1757 : cancel: reconciler_cancel,
1758 : });
1759 :
1760 : Some(ReconcilerWaiter {
1761 : tenant_shard_id: self.tenant_shard_id,
1762 : seq_wait: self.waiter.clone(),
1763 : error_seq_wait: self.error_waiter.clone(),
1764 : error: self.last_error.clone(),
1765 : seq: self.sequence,
1766 : })
1767 : }
1768 :
1769 0 : pub(crate) fn cancel_reconciler(&self) {
1770 0 : if let Some(handle) = self.reconciler.as_ref() {
1771 0 : handle.cancel.cancel()
1772 0 : }
1773 0 : }
1774 :
1775 : /// Get a waiter for any reconciliation in flight, but do not start reconciliation
1776 : /// if it is not already running
1777 0 : pub(crate) fn get_waiter(&self) -> Option<ReconcilerWaiter> {
1778 0 : if self.reconciler.is_some() {
1779 0 : Some(ReconcilerWaiter {
1780 0 : tenant_shard_id: self.tenant_shard_id,
1781 0 : seq_wait: self.waiter.clone(),
1782 0 : error_seq_wait: self.error_waiter.clone(),
1783 0 : error: self.last_error.clone(),
1784 0 : seq: self.sequence,
1785 0 : })
1786 : } else {
1787 0 : None
1788 : }
1789 0 : }
1790 :
1791 : /// Called when a ReconcileResult has been emitted and the service is updating
1792 : /// our state: if the result is from a sequence >= my ReconcileHandle, then drop
1793 : /// the handle to indicate there is no longer a reconciliation in progress.
1794 0 : pub(crate) fn reconcile_complete(&mut self, sequence: Sequence) {
1795 0 : if let Some(reconcile_handle) = &self.reconciler {
1796 0 : if reconcile_handle.sequence <= sequence {
1797 0 : self.reconciler = None;
1798 0 : }
1799 0 : }
1800 0 : }
1801 :
1802 : /// If we had any state at all referring to this node ID, drop it. Does not
1803 : /// attempt to reschedule.
1804 : ///
1805 : /// Returns true if we modified the node's intent state.
1806 0 : pub(crate) fn deref_node(&mut self, node_id: NodeId) -> bool {
1807 0 : let mut intent_modified = false;
1808 :
1809 : // Drop if this node was our attached intent
1810 0 : if self.intent.attached == Some(node_id) {
1811 0 : self.intent.attached = None;
1812 0 : intent_modified = true;
1813 0 : }
1814 :
1815 : // Drop from the list of secondaries, and check if we modified it
1816 0 : let had_secondaries = self.intent.secondary.len();
1817 0 : self.intent.secondary.retain(|n| n != &node_id);
1818 0 : intent_modified |= self.intent.secondary.len() != had_secondaries;
1819 :
1820 0 : debug_assert!(!self.intent.all_pageservers().contains(&node_id));
1821 :
1822 0 : if self.preferred_node == Some(node_id) {
1823 0 : self.preferred_node = None;
1824 0 : }
1825 :
1826 0 : intent_modified
1827 0 : }
1828 :
1829 0 : pub(crate) fn set_scheduling_policy(&mut self, p: ShardSchedulingPolicy) {
1830 0 : self.scheduling_policy = p;
1831 0 : }
1832 :
1833 0 : pub(crate) fn get_scheduling_policy(&self) -> ShardSchedulingPolicy {
1834 0 : self.scheduling_policy
1835 0 : }
1836 :
1837 0 : pub(crate) fn set_last_error(&mut self, sequence: Sequence, error: ReconcileError) {
1838 : // Ordering: always set last_error before advancing sequence, so that sequence
1839 : // waiters are guaranteed to see a Some value when they see an error.
1840 0 : *(self.last_error.lock().unwrap()) = Some(Arc::new(error));
1841 0 : self.error_waiter.advance(sequence);
1842 0 : }
1843 :
1844 0 : pub(crate) fn from_persistent(
1845 0 : tsp: TenantShardPersistence,
1846 0 : intent: IntentState,
1847 0 : ) -> anyhow::Result<Self> {
1848 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1849 0 : let shard_identity = tsp.get_shard_identity()?;
1850 :
1851 0 : metrics::METRICS_REGISTRY
1852 0 : .metrics_group
1853 0 : .storage_controller_tenant_shards
1854 0 : .inc();
1855 :
1856 : Ok(Self {
1857 0 : tenant_shard_id,
1858 0 : shard: shard_identity,
1859 0 : sequence: Sequence::initial(),
1860 0 : generation: tsp.generation.map(|g| Generation::new(g as u32)),
1861 0 : policy: serde_json::from_str(&tsp.placement_policy).unwrap(),
1862 0 : intent,
1863 0 : observed: ObservedState::new(),
1864 0 : config: serde_json::from_str(&tsp.config).unwrap(),
1865 0 : reconciler: None,
1866 0 : splitting: tsp.splitting,
1867 : // Filled in during [`Service::startup_reconcile`]
1868 0 : importing: TimelineImportState::Idle,
1869 0 : waiter: Arc::new(SeqWait::new(Sequence::initial())),
1870 0 : error_waiter: Arc::new(SeqWait::new(Sequence::initial())),
1871 0 : last_error: Arc::default(),
1872 : consecutive_errors_count: 0,
1873 : pending_compute_notification: false,
1874 : delayed_reconcile: false,
1875 0 : scheduling_policy: serde_json::from_str(&tsp.scheduling_policy).unwrap(),
1876 0 : preferred_node: None,
1877 : })
1878 0 : }
1879 :
1880 0 : pub(crate) fn to_persistent(&self) -> TenantShardPersistence {
1881 : TenantShardPersistence {
1882 0 : tenant_id: self.tenant_shard_id.tenant_id.to_string(),
1883 0 : shard_number: self.tenant_shard_id.shard_number.0 as i32,
1884 0 : shard_count: self.tenant_shard_id.shard_count.literal() as i32,
1885 0 : shard_stripe_size: self.shard.stripe_size.0 as i32,
1886 0 : generation: self.generation.map(|g| g.into().unwrap_or(0) as i32),
1887 0 : generation_pageserver: self.intent.get_attached().map(|n| n.0 as i64),
1888 0 : placement_policy: serde_json::to_string(&self.policy).unwrap(),
1889 0 : config: serde_json::to_string(&self.config).unwrap(),
1890 0 : splitting: SplitState::default(),
1891 0 : scheduling_policy: serde_json::to_string(&self.scheduling_policy).unwrap(),
1892 0 : preferred_az_id: self.intent.preferred_az_id.as_ref().map(|az| az.0.clone()),
1893 : }
1894 0 : }
1895 :
1896 12508 : pub(crate) fn preferred_az(&self) -> Option<&AvailabilityZone> {
1897 12508 : self.intent.get_preferred_az()
1898 12508 : }
1899 :
1900 2 : pub(crate) fn set_preferred_az(
1901 2 : &mut self,
1902 2 : scheduler: &mut Scheduler,
1903 2 : preferred_az_id: Option<AvailabilityZone>,
1904 2 : ) {
1905 2 : self.intent.set_preferred_az(scheduler, preferred_az_id);
1906 2 : }
1907 :
1908 : /// Returns all the nodes to which this tenant shard is attached according to the
1909 : /// observed state and the generations. Return vector is sorted from latest generation
1910 : /// to earliest.
1911 0 : pub(crate) fn attached_locations(&self) -> Vec<(NodeId, Generation)> {
1912 0 : self.observed
1913 0 : .locations
1914 0 : .iter()
1915 0 : .filter_map(|(node_id, observed)| {
1916 : use LocationConfigMode::{AttachedMulti, AttachedSingle, AttachedStale};
1917 :
1918 0 : let conf = observed.conf.as_ref()?;
1919 :
1920 0 : match (conf.generation, conf.mode) {
1921 0 : (Some(gen_), AttachedMulti | AttachedSingle | AttachedStale) => {
1922 0 : Some((*node_id, gen_))
1923 : }
1924 0 : _ => None,
1925 : }
1926 0 : })
1927 0 : .sorted_by(|(_lhs_node_id, lhs_gen), (_rhs_node_id, rhs_gen)| {
1928 0 : lhs_gen.cmp(rhs_gen).reverse()
1929 0 : })
1930 0 : .map(|(node_id, gen_)| (node_id, Generation::new(gen_)))
1931 0 : .collect()
1932 0 : }
1933 :
1934 : /// Update the observed state of the tenant by applying incremental deltas
1935 : ///
1936 : /// Deltas are generated by reconcilers via [`Reconciler::observed_deltas`].
1937 : /// They are then filtered in [`crate::service::Service::process_result`].
1938 0 : pub(crate) fn apply_observed_deltas(
1939 0 : &mut self,
1940 0 : deltas: impl Iterator<Item = ObservedStateDelta>,
1941 0 : ) {
1942 0 : for delta in deltas {
1943 0 : match delta {
1944 0 : ObservedStateDelta::Upsert(ups) => {
1945 0 : let (node_id, loc) = *ups;
1946 :
1947 : // If the generation of the observed location in the delta is lagging
1948 : // behind the current one, then we have a race condition and cannot
1949 : // be certain about the true observed state. Set the observed state
1950 : // to None in order to reflect this.
1951 0 : let crnt_gen = self
1952 0 : .observed
1953 0 : .locations
1954 0 : .get(&node_id)
1955 0 : .and_then(|loc| loc.conf.as_ref())
1956 0 : .and_then(|conf| conf.generation);
1957 0 : let new_gen = loc.conf.as_ref().and_then(|conf| conf.generation);
1958 0 : match (crnt_gen, new_gen) {
1959 0 : (Some(crnt), Some(new)) if crnt_gen > new_gen => {
1960 0 : tracing::warn!(
1961 0 : "Skipping observed state update {}: {:?} and using None due to stale generation ({} > {})",
1962 : node_id,
1963 : loc,
1964 : crnt,
1965 : new
1966 : );
1967 :
1968 0 : self.observed
1969 0 : .locations
1970 0 : .insert(node_id, ObservedStateLocation { conf: None });
1971 :
1972 0 : continue;
1973 : }
1974 0 : _ => {}
1975 : }
1976 :
1977 0 : if let Some(conf) = &loc.conf {
1978 0 : tracing::info!("Updating observed location {}: {:?}", node_id, conf);
1979 : } else {
1980 0 : tracing::info!("Setting observed location {} to None", node_id,)
1981 : }
1982 :
1983 0 : self.observed.locations.insert(node_id, loc);
1984 : }
1985 0 : ObservedStateDelta::Delete(node_id) => {
1986 0 : tracing::info!("Deleting observed location {}", node_id);
1987 0 : self.observed.locations.remove(&node_id);
1988 : }
1989 : }
1990 : }
1991 0 : }
1992 :
1993 : /// Returns true if the tenant shard is attached to a node that is outside the preferred AZ.
1994 : ///
1995 : /// If the shard does not have a preferred AZ, returns false.
1996 0 : pub(crate) fn is_attached_outside_preferred_az(&self, nodes: &HashMap<NodeId, Node>) -> bool {
1997 0 : self.intent
1998 0 : .get_attached()
1999 0 : .map(|node_id| {
2000 0 : Some(
2001 0 : nodes
2002 0 : .get(&node_id)
2003 0 : .expect("referenced node exists")
2004 0 : .get_availability_zone_id(),
2005 0 : ) != self.intent.preferred_az_id.as_ref()
2006 0 : })
2007 0 : .unwrap_or(false)
2008 0 : }
2009 : }
2010 :
2011 : impl Drop for TenantShard {
2012 12843 : fn drop(&mut self) {
2013 12843 : metrics::METRICS_REGISTRY
2014 12843 : .metrics_group
2015 12843 : .storage_controller_tenant_shards
2016 12843 : .dec();
2017 12843 : }
2018 : }
2019 :
2020 : #[cfg(test)]
2021 : pub(crate) mod tests {
2022 : use std::cell::RefCell;
2023 : use std::rc::Rc;
2024 :
2025 : use pageserver_api::controller_api::NodeAvailability;
2026 : use pageserver_api::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardNumber};
2027 : use rand::SeedableRng;
2028 : use rand::rngs::StdRng;
2029 : use utils::id::TenantId;
2030 :
2031 : use super::*;
2032 : use crate::scheduler::test_utils::make_test_nodes;
2033 :
2034 12 : fn make_test_tenant_shard(policy: PlacementPolicy) -> TenantShard {
2035 12 : let tenant_id = TenantId::generate();
2036 12 : let shard_number = ShardNumber(0);
2037 12 : let shard_count = ShardCount::new(1);
2038 12 : let stripe_size = DEFAULT_STRIPE_SIZE;
2039 :
2040 12 : let tenant_shard_id = TenantShardId {
2041 12 : tenant_id,
2042 12 : shard_number,
2043 12 : shard_count,
2044 12 : };
2045 12 : TenantShard::new(
2046 12 : tenant_shard_id,
2047 12 : ShardIdentity::new(shard_number, shard_count, stripe_size).unwrap(),
2048 12 : policy,
2049 12 : None,
2050 : )
2051 12 : }
2052 :
2053 5004 : pub(crate) fn make_test_tenant(
2054 5004 : policy: PlacementPolicy,
2055 5004 : shard_count: ShardCount,
2056 5004 : preferred_az: Option<AvailabilityZone>,
2057 5004 : ) -> Vec<TenantShard> {
2058 5004 : make_test_tenant_with_id(TenantId::generate(), policy, shard_count, preferred_az)
2059 5004 : }
2060 :
2061 5007 : pub(crate) fn make_test_tenant_with_id(
2062 5007 : tenant_id: TenantId,
2063 5007 : policy: PlacementPolicy,
2064 5007 : shard_count: ShardCount,
2065 5007 : preferred_az: Option<AvailabilityZone>,
2066 5007 : ) -> Vec<TenantShard> {
2067 5007 : let stripe_size = DEFAULT_STRIPE_SIZE;
2068 5007 : (0..shard_count.count())
2069 12522 : .map(|i| {
2070 12522 : let shard_number = ShardNumber(i);
2071 :
2072 12522 : let tenant_shard_id = TenantShardId {
2073 12522 : tenant_id,
2074 12522 : shard_number,
2075 12522 : shard_count,
2076 12522 : };
2077 12522 : TenantShard::new(
2078 12522 : tenant_shard_id,
2079 12522 : ShardIdentity::new(shard_number, shard_count, stripe_size).unwrap(),
2080 12522 : policy.clone(),
2081 12522 : preferred_az.clone(),
2082 : )
2083 12522 : })
2084 5007 : .collect()
2085 5007 : }
2086 :
2087 : /// Test the scheduling behaviors used when a tenant configured for HA is subject
2088 : /// to nodes being marked offline.
2089 : #[test]
2090 1 : fn tenant_ha_scheduling() -> anyhow::Result<()> {
2091 : // Start with three nodes. Our tenant will only use two. The third one is
2092 : // expected to remain unused.
2093 1 : let mut nodes = make_test_nodes(3, &[]);
2094 :
2095 1 : let mut scheduler = Scheduler::new(nodes.values());
2096 1 : let mut context = ScheduleContext::default();
2097 :
2098 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2099 1 : tenant_shard
2100 1 : .schedule(&mut scheduler, &mut context)
2101 1 : .expect("we have enough nodes, scheduling should work");
2102 :
2103 : // Expect to initially be schedule on to different nodes
2104 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
2105 1 : assert!(tenant_shard.intent.attached.is_some());
2106 :
2107 1 : let attached_node_id = tenant_shard.intent.attached.unwrap();
2108 1 : let secondary_node_id = *tenant_shard.intent.secondary.iter().last().unwrap();
2109 1 : assert_ne!(attached_node_id, secondary_node_id);
2110 :
2111 : // Notifying the attached node is offline should demote it to a secondary
2112 1 : let changed = tenant_shard
2113 1 : .intent
2114 1 : .demote_attached(&mut scheduler, attached_node_id);
2115 1 : assert!(changed);
2116 1 : assert!(tenant_shard.intent.attached.is_none());
2117 1 : assert_eq!(tenant_shard.intent.secondary.len(), 2);
2118 :
2119 : // Update the scheduler state to indicate the node is offline
2120 1 : nodes
2121 1 : .get_mut(&attached_node_id)
2122 1 : .unwrap()
2123 1 : .set_availability(NodeAvailability::Offline);
2124 1 : scheduler.node_upsert(nodes.get(&attached_node_id).unwrap());
2125 :
2126 : // Scheduling the node should promote the still-available secondary node to attached
2127 1 : tenant_shard
2128 1 : .schedule(&mut scheduler, &mut context)
2129 1 : .expect("active nodes are available");
2130 1 : assert_eq!(tenant_shard.intent.attached.unwrap(), secondary_node_id);
2131 :
2132 : // The original attached node should have been retained as a secondary
2133 1 : assert_eq!(
2134 1 : *tenant_shard.intent.secondary.iter().last().unwrap(),
2135 : attached_node_id
2136 : );
2137 :
2138 1 : tenant_shard.intent.clear(&mut scheduler);
2139 :
2140 1 : Ok(())
2141 1 : }
2142 :
2143 : #[test]
2144 1 : fn intent_from_observed() -> anyhow::Result<()> {
2145 1 : let nodes = make_test_nodes(3, &[]);
2146 1 : let mut scheduler = Scheduler::new(nodes.values());
2147 :
2148 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2149 :
2150 1 : tenant_shard.observed.locations.insert(
2151 1 : NodeId(3),
2152 1 : ObservedStateLocation {
2153 1 : conf: Some(LocationConfig {
2154 1 : mode: LocationConfigMode::AttachedMulti,
2155 1 : generation: Some(2),
2156 1 : secondary_conf: None,
2157 1 : shard_number: tenant_shard.shard.number.0,
2158 1 : shard_count: tenant_shard.shard.count.literal(),
2159 1 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
2160 1 : tenant_conf: TenantConfig::default(),
2161 1 : }),
2162 1 : },
2163 : );
2164 :
2165 1 : tenant_shard.observed.locations.insert(
2166 1 : NodeId(2),
2167 1 : ObservedStateLocation {
2168 1 : conf: Some(LocationConfig {
2169 1 : mode: LocationConfigMode::AttachedStale,
2170 1 : generation: Some(1),
2171 1 : secondary_conf: None,
2172 1 : shard_number: tenant_shard.shard.number.0,
2173 1 : shard_count: tenant_shard.shard.count.literal(),
2174 1 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
2175 1 : tenant_conf: TenantConfig::default(),
2176 1 : }),
2177 1 : },
2178 : );
2179 :
2180 1 : tenant_shard.intent_from_observed(&mut scheduler);
2181 :
2182 : // The highest generationed attached location gets used as attached
2183 1 : assert_eq!(tenant_shard.intent.attached, Some(NodeId(3)));
2184 : // Other locations get used as secondary
2185 1 : assert_eq!(tenant_shard.intent.secondary, vec![NodeId(2)]);
2186 :
2187 1 : scheduler.consistency_check(nodes.values(), [&tenant_shard].into_iter())?;
2188 :
2189 1 : tenant_shard.intent.clear(&mut scheduler);
2190 1 : Ok(())
2191 1 : }
2192 :
2193 : #[test]
2194 1 : fn scheduling_mode() -> anyhow::Result<()> {
2195 1 : let nodes = make_test_nodes(3, &[]);
2196 1 : let mut scheduler = Scheduler::new(nodes.values());
2197 :
2198 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2199 :
2200 : // In pause mode, schedule() shouldn't do anything
2201 1 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Pause;
2202 1 : assert!(
2203 1 : tenant_shard
2204 1 : .schedule(&mut scheduler, &mut ScheduleContext::default())
2205 1 : .is_ok()
2206 : );
2207 1 : assert!(tenant_shard.intent.all_pageservers().is_empty());
2208 :
2209 : // In active mode, schedule() works
2210 1 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Active;
2211 1 : assert!(
2212 1 : tenant_shard
2213 1 : .schedule(&mut scheduler, &mut ScheduleContext::default())
2214 1 : .is_ok()
2215 : );
2216 1 : assert!(!tenant_shard.intent.all_pageservers().is_empty());
2217 :
2218 1 : tenant_shard.intent.clear(&mut scheduler);
2219 1 : Ok(())
2220 1 : }
2221 :
2222 : #[test]
2223 : /// Simple case: moving attachment to somewhere better where we already have a secondary
2224 1 : fn optimize_attachment_simple() -> anyhow::Result<()> {
2225 1 : let nodes = make_test_nodes(
2226 : 3,
2227 1 : &[
2228 1 : AvailabilityZone("az-a".to_string()),
2229 1 : AvailabilityZone("az-b".to_string()),
2230 1 : AvailabilityZone("az-c".to_string()),
2231 1 : ],
2232 : );
2233 1 : let mut scheduler = Scheduler::new(nodes.values());
2234 :
2235 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2236 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2237 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2238 1 : shard_b.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2239 :
2240 : // Initially: both nodes attached on shard 1, and both have secondary locations
2241 : // on different nodes.
2242 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2243 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(1));
2244 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2245 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(2));
2246 :
2247 1 : fn make_schedule_context(shard_a: &TenantShard, shard_b: &TenantShard) -> ScheduleContext {
2248 1 : let mut schedule_context = ScheduleContext::default();
2249 1 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2250 1 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2251 1 : schedule_context
2252 1 : }
2253 :
2254 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2255 1 : let optimization_a = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2256 1 : assert_eq!(
2257 : optimization_a,
2258 1 : Some(ScheduleOptimization {
2259 1 : sequence: shard_a.sequence,
2260 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2261 1 : old_attached_node_id: NodeId(2),
2262 1 : new_attached_node_id: NodeId(1)
2263 1 : })
2264 1 : })
2265 : );
2266 1 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
2267 :
2268 : // // Either shard should recognize that it has the option to switch to a secondary location where there
2269 : // // would be no other shards from the same tenant, and request to do so.
2270 : // assert_eq!(
2271 : // optimization_a_prepare,
2272 : // Some(ScheduleOptimization {
2273 : // sequence: shard_a.sequence,
2274 : // action: ScheduleOptimizationAction::CreateSecondary(NodeId(2))
2275 : // })
2276 : // );
2277 : // shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2278 :
2279 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2280 : // let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2281 : // assert_eq!(
2282 : // optimization_a_migrate,
2283 : // Some(ScheduleOptimization {
2284 : // sequence: shard_a.sequence,
2285 : // action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2286 : // old_attached_node_id: NodeId(1),
2287 : // new_attached_node_id: NodeId(2)
2288 : // })
2289 : // })
2290 : // );
2291 : // shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2292 :
2293 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2294 : // let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2295 : // assert_eq!(
2296 : // optimization_a_cleanup,
2297 : // Some(ScheduleOptimization {
2298 : // sequence: shard_a.sequence,
2299 : // action: ScheduleOptimizationAction::RemoveSecondary(NodeId(1))
2300 : // })
2301 : // );
2302 : // shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2303 :
2304 : // // Shard B should not be moved anywhere, since the pressure on node 1 was relieved by moving shard A
2305 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2306 : // assert_eq!(shard_b.optimize_attachment(&mut scheduler, &schedule_context), None);
2307 :
2308 1 : shard_a.intent.clear(&mut scheduler);
2309 1 : shard_b.intent.clear(&mut scheduler);
2310 :
2311 1 : Ok(())
2312 1 : }
2313 :
2314 : #[test]
2315 : /// Complicated case: moving attachment to somewhere better where we do not have a secondary
2316 : /// already, creating one as needed.
2317 1 : fn optimize_attachment_multistep() -> anyhow::Result<()> {
2318 1 : let nodes = make_test_nodes(
2319 : 3,
2320 1 : &[
2321 1 : AvailabilityZone("az-a".to_string()),
2322 1 : AvailabilityZone("az-b".to_string()),
2323 1 : AvailabilityZone("az-c".to_string()),
2324 1 : ],
2325 : );
2326 1 : let mut scheduler = Scheduler::new(nodes.values());
2327 :
2328 : // Two shards of a tenant that wants to be in AZ A
2329 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2330 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2331 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2332 1 : shard_b.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2333 :
2334 : // Both shards are initially attached in non-home AZ _and_ have secondaries in non-home AZs
2335 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2336 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2337 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(3)));
2338 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(2));
2339 :
2340 3 : fn make_schedule_context(shard_a: &TenantShard, shard_b: &TenantShard) -> ScheduleContext {
2341 3 : let mut schedule_context = ScheduleContext::default();
2342 3 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2343 3 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2344 3 : schedule_context
2345 3 : }
2346 :
2347 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2348 1 : let optimization_a_prepare = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2349 1 : assert_eq!(
2350 : optimization_a_prepare,
2351 1 : Some(ScheduleOptimization {
2352 1 : sequence: shard_a.sequence,
2353 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(1))
2354 1 : })
2355 : );
2356 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2357 :
2358 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2359 1 : let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2360 1 : assert_eq!(
2361 : optimization_a_migrate,
2362 1 : Some(ScheduleOptimization {
2363 1 : sequence: shard_a.sequence,
2364 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2365 1 : old_attached_node_id: NodeId(2),
2366 1 : new_attached_node_id: NodeId(1)
2367 1 : })
2368 1 : })
2369 : );
2370 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2371 :
2372 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2373 1 : let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2374 1 : assert_eq!(
2375 : optimization_a_cleanup,
2376 1 : Some(ScheduleOptimization {
2377 1 : sequence: shard_a.sequence,
2378 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(3))
2379 1 : })
2380 : );
2381 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2382 :
2383 : // // Shard B should not be moved anywhere, since the pressure on node 1 was relieved by moving shard A
2384 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2385 : // assert_eq!(shard_b.optimize_attachment(&mut scheduler, &schedule_context), None);
2386 :
2387 1 : shard_a.intent.clear(&mut scheduler);
2388 1 : shard_b.intent.clear(&mut scheduler);
2389 :
2390 1 : Ok(())
2391 1 : }
2392 :
2393 : #[test]
2394 : /// How the optimisation code handles a shard with a preferred node set; this is an example
2395 : /// of the multi-step migration, but driven by a different input.
2396 1 : fn optimize_attachment_multi_preferred_node() -> anyhow::Result<()> {
2397 1 : let nodes = make_test_nodes(
2398 : 4,
2399 1 : &[
2400 1 : AvailabilityZone("az-a".to_string()),
2401 1 : AvailabilityZone("az-a".to_string()),
2402 1 : AvailabilityZone("az-b".to_string()),
2403 1 : AvailabilityZone("az-b".to_string()),
2404 1 : ],
2405 : );
2406 1 : let mut scheduler = Scheduler::new(nodes.values());
2407 :
2408 : // Two shards of a tenant that wants to be in AZ A
2409 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2410 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2411 :
2412 : // Initially attached in a stable location
2413 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2414 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2415 :
2416 : // Set the preferred node to node 2, an equally high scoring node to its current location
2417 1 : shard_a.preferred_node = Some(NodeId(2));
2418 :
2419 3 : fn make_schedule_context(shard_a: &TenantShard) -> ScheduleContext {
2420 3 : let mut schedule_context = ScheduleContext::default();
2421 3 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2422 3 : schedule_context
2423 3 : }
2424 :
2425 1 : let schedule_context = make_schedule_context(&shard_a);
2426 1 : let optimization_a_prepare = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2427 1 : assert_eq!(
2428 : optimization_a_prepare,
2429 1 : Some(ScheduleOptimization {
2430 1 : sequence: shard_a.sequence,
2431 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(2))
2432 1 : })
2433 : );
2434 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2435 :
2436 : // The first step of the optimisation should not have cleared the preferred node
2437 1 : assert_eq!(shard_a.preferred_node, Some(NodeId(2)));
2438 :
2439 1 : let schedule_context = make_schedule_context(&shard_a);
2440 1 : let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2441 1 : assert_eq!(
2442 : optimization_a_migrate,
2443 1 : Some(ScheduleOptimization {
2444 1 : sequence: shard_a.sequence,
2445 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2446 1 : old_attached_node_id: NodeId(1),
2447 1 : new_attached_node_id: NodeId(2)
2448 1 : })
2449 1 : })
2450 : );
2451 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2452 :
2453 : // The cutover step of the optimisation should have cleared the preferred node
2454 1 : assert_eq!(shard_a.preferred_node, None);
2455 :
2456 1 : let schedule_context = make_schedule_context(&shard_a);
2457 1 : let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2458 1 : assert_eq!(
2459 : optimization_a_cleanup,
2460 1 : Some(ScheduleOptimization {
2461 1 : sequence: shard_a.sequence,
2462 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(1))
2463 1 : })
2464 : );
2465 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2466 :
2467 1 : shard_a.intent.clear(&mut scheduler);
2468 :
2469 1 : Ok(())
2470 1 : }
2471 :
2472 : #[test]
2473 : /// Check that multi-step migration works when moving to somewhere that is only better by
2474 : /// 1 AffinityScore -- this ensures that we don't have a bug like the intermediate secondary
2475 : /// counting toward the affinity score such that it prevents the rest of the migration from happening.
2476 1 : fn optimize_attachment_marginal() -> anyhow::Result<()> {
2477 1 : let nodes = make_test_nodes(2, &[]);
2478 1 : let mut scheduler = Scheduler::new(nodes.values());
2479 :
2480 : // Multi-sharded tenant, we will craft a situation where affinity
2481 : // scores differ only slightly
2482 1 : let mut shards = make_test_tenant(PlacementPolicy::Attached(0), ShardCount::new(4), None);
2483 :
2484 : // 1 attached on node 1
2485 1 : shards[0]
2486 1 : .intent
2487 1 : .set_attached(&mut scheduler, Some(NodeId(1)));
2488 : // 3 attached on node 2
2489 1 : shards[1]
2490 1 : .intent
2491 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2492 1 : shards[2]
2493 1 : .intent
2494 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2495 1 : shards[3]
2496 1 : .intent
2497 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2498 :
2499 : // The scheduler should figure out that we need to:
2500 : // - Create a secondary for shard 3 on node 1
2501 : // - Migrate shard 3 to node 1
2502 : // - Remove shard 3's location on node 2
2503 :
2504 4 : fn make_schedule_context(shards: &Vec<TenantShard>) -> ScheduleContext {
2505 4 : let mut schedule_context = ScheduleContext::default();
2506 20 : for shard in shards {
2507 16 : schedule_context.avoid(&shard.intent.all_pageservers());
2508 16 : }
2509 4 : schedule_context
2510 4 : }
2511 :
2512 1 : let schedule_context = make_schedule_context(&shards);
2513 1 : let optimization_a_prepare =
2514 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2515 1 : assert_eq!(
2516 : optimization_a_prepare,
2517 1 : Some(ScheduleOptimization {
2518 1 : sequence: shards[1].sequence,
2519 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(1))
2520 1 : })
2521 : );
2522 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2523 :
2524 1 : let schedule_context = make_schedule_context(&shards);
2525 1 : let optimization_a_migrate =
2526 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2527 1 : assert_eq!(
2528 : optimization_a_migrate,
2529 1 : Some(ScheduleOptimization {
2530 1 : sequence: shards[1].sequence,
2531 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2532 1 : old_attached_node_id: NodeId(2),
2533 1 : new_attached_node_id: NodeId(1)
2534 1 : })
2535 1 : })
2536 : );
2537 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2538 :
2539 1 : let schedule_context = make_schedule_context(&shards);
2540 1 : let optimization_a_cleanup =
2541 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2542 1 : assert_eq!(
2543 : optimization_a_cleanup,
2544 1 : Some(ScheduleOptimization {
2545 1 : sequence: shards[1].sequence,
2546 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(2))
2547 1 : })
2548 : );
2549 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2550 :
2551 : // Everything should be stable now
2552 1 : let schedule_context = make_schedule_context(&shards);
2553 1 : assert_eq!(
2554 1 : shards[0].optimize_attachment(&mut scheduler, &schedule_context),
2555 : None
2556 : );
2557 1 : assert_eq!(
2558 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context),
2559 : None
2560 : );
2561 1 : assert_eq!(
2562 1 : shards[2].optimize_attachment(&mut scheduler, &schedule_context),
2563 : None
2564 : );
2565 1 : assert_eq!(
2566 1 : shards[3].optimize_attachment(&mut scheduler, &schedule_context),
2567 : None
2568 : );
2569 :
2570 5 : for mut shard in shards {
2571 4 : shard.intent.clear(&mut scheduler);
2572 4 : }
2573 :
2574 1 : Ok(())
2575 1 : }
2576 :
2577 : #[test]
2578 1 : fn optimize_secondary() -> anyhow::Result<()> {
2579 1 : let nodes = make_test_nodes(4, &[]);
2580 1 : let mut scheduler = Scheduler::new(nodes.values());
2581 :
2582 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2583 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2584 :
2585 : // Initially: both nodes attached on shard 1, and both have secondary locations
2586 : // on different nodes.
2587 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2588 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2589 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2590 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(3));
2591 :
2592 1 : let mut schedule_context = ScheduleContext::default();
2593 1 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2594 1 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2595 :
2596 1 : let optimization_a = shard_a.optimize_secondary(&mut scheduler, &schedule_context);
2597 :
2598 : // Since there is a node with no locations available, the node with two locations for the
2599 : // same tenant should generate an optimization to move one away
2600 1 : assert_eq!(
2601 : optimization_a,
2602 1 : Some(ScheduleOptimization {
2603 1 : sequence: shard_a.sequence,
2604 1 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
2605 1 : old_node_id: NodeId(3),
2606 1 : new_node_id: NodeId(4)
2607 1 : })
2608 1 : })
2609 : );
2610 :
2611 1 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
2612 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2613 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(4)]);
2614 :
2615 1 : shard_a.intent.clear(&mut scheduler);
2616 1 : shard_b.intent.clear(&mut scheduler);
2617 :
2618 1 : Ok(())
2619 1 : }
2620 :
2621 : /// Test how the optimisation code behaves with an extra secondary
2622 : #[test]
2623 1 : fn optimize_removes_secondary() -> anyhow::Result<()> {
2624 1 : let az_a_tag = AvailabilityZone("az-a".to_string());
2625 1 : let az_b_tag = AvailabilityZone("az-b".to_string());
2626 1 : let mut nodes = make_test_nodes(
2627 : 4,
2628 1 : &[
2629 1 : az_a_tag.clone(),
2630 1 : az_b_tag.clone(),
2631 1 : az_a_tag.clone(),
2632 1 : az_b_tag.clone(),
2633 1 : ],
2634 : );
2635 1 : let mut scheduler = Scheduler::new(nodes.values());
2636 :
2637 1 : let mut schedule_context = ScheduleContext::default();
2638 :
2639 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2640 1 : shard_a.intent.preferred_az_id = Some(az_a_tag.clone());
2641 1 : shard_a
2642 1 : .schedule(&mut scheduler, &mut schedule_context)
2643 1 : .unwrap();
2644 :
2645 : // Attached on node 1, secondary on node 2
2646 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2647 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(2)]);
2648 :
2649 : // Initially optimiser is idle
2650 1 : assert_eq!(
2651 1 : shard_a.optimize_attachment(&mut scheduler, &schedule_context),
2652 : None
2653 : );
2654 1 : assert_eq!(
2655 1 : shard_a.optimize_secondary(&mut scheduler, &schedule_context),
2656 : None
2657 : );
2658 :
2659 : // A spare secondary in the home AZ: it should be removed -- this is the situation when we're midway through a graceful migration, after cutting over
2660 : // to our new location
2661 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2662 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2663 1 : assert_eq!(
2664 : optimization,
2665 1 : Some(ScheduleOptimization {
2666 1 : sequence: shard_a.sequence,
2667 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(3))
2668 1 : })
2669 : );
2670 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2671 :
2672 : // A spare secondary in the non-home AZ, and one of them is offline
2673 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(4));
2674 1 : nodes
2675 1 : .get_mut(&NodeId(4))
2676 1 : .unwrap()
2677 1 : .set_availability(NodeAvailability::Offline);
2678 1 : scheduler.node_upsert(nodes.get(&NodeId(4)).unwrap());
2679 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2680 1 : assert_eq!(
2681 : optimization,
2682 1 : Some(ScheduleOptimization {
2683 1 : sequence: shard_a.sequence,
2684 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(4))
2685 1 : })
2686 : );
2687 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2688 :
2689 : // A spare secondary when should have none
2690 1 : shard_a.policy = PlacementPolicy::Attached(0);
2691 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2692 1 : assert_eq!(
2693 : optimization,
2694 1 : Some(ScheduleOptimization {
2695 1 : sequence: shard_a.sequence,
2696 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(2))
2697 1 : })
2698 : );
2699 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2700 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2701 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![]);
2702 :
2703 : // Check that in secondary mode, we preserve the secondary in the preferred AZ
2704 1 : let mut schedule_context = ScheduleContext::default(); // Fresh context, we're about to call schedule()
2705 1 : shard_a.policy = PlacementPolicy::Secondary;
2706 1 : shard_a
2707 1 : .schedule(&mut scheduler, &mut schedule_context)
2708 1 : .unwrap();
2709 1 : assert_eq!(shard_a.intent.get_attached(), &None);
2710 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(1)]);
2711 1 : assert_eq!(
2712 1 : shard_a.optimize_attachment(&mut scheduler, &schedule_context),
2713 : None
2714 : );
2715 1 : assert_eq!(
2716 1 : shard_a.optimize_secondary(&mut scheduler, &schedule_context),
2717 : None
2718 : );
2719 :
2720 1 : shard_a.intent.clear(&mut scheduler);
2721 :
2722 1 : Ok(())
2723 1 : }
2724 :
2725 : // Optimize til quiescent: this emulates what Service::optimize_all does, when
2726 : // called repeatedly in the background.
2727 : // Returns the applied optimizations
2728 3 : fn optimize_til_idle(
2729 3 : scheduler: &mut Scheduler,
2730 3 : shards: &mut [TenantShard],
2731 3 : ) -> Vec<ScheduleOptimization> {
2732 3 : let mut loop_n = 0;
2733 3 : let mut optimizations = Vec::default();
2734 : loop {
2735 6 : let mut schedule_context = ScheduleContext::default();
2736 6 : let mut any_changed = false;
2737 :
2738 24 : for shard in shards.iter() {
2739 24 : schedule_context.avoid(&shard.intent.all_pageservers());
2740 24 : }
2741 :
2742 15 : for shard in shards.iter_mut() {
2743 15 : let optimization = shard.optimize_attachment(scheduler, &schedule_context);
2744 15 : tracing::info!(
2745 0 : "optimize_attachment({})={:?}",
2746 : shard.tenant_shard_id,
2747 : optimization
2748 : );
2749 15 : if let Some(optimization) = optimization {
2750 : // Check that maybe_optimizable wouldn't have wrongly claimed this optimization didn't exist
2751 3 : assert!(shard.maybe_optimizable(scheduler, &schedule_context));
2752 3 : optimizations.push(optimization.clone());
2753 3 : shard.apply_optimization(scheduler, optimization);
2754 3 : any_changed = true;
2755 3 : break;
2756 12 : }
2757 :
2758 12 : let optimization = shard.optimize_secondary(scheduler, &schedule_context);
2759 12 : tracing::info!(
2760 0 : "optimize_secondary({})={:?}",
2761 : shard.tenant_shard_id,
2762 : optimization
2763 : );
2764 12 : if let Some(optimization) = optimization {
2765 : // Check that maybe_optimizable wouldn't have wrongly claimed this optimization didn't exist
2766 0 : assert!(shard.maybe_optimizable(scheduler, &schedule_context));
2767 :
2768 0 : optimizations.push(optimization.clone());
2769 0 : shard.apply_optimization(scheduler, optimization);
2770 0 : any_changed = true;
2771 0 : break;
2772 12 : }
2773 : }
2774 :
2775 6 : if !any_changed {
2776 3 : break;
2777 3 : }
2778 :
2779 : // Assert no infinite loop
2780 3 : loop_n += 1;
2781 3 : assert!(loop_n < 1000);
2782 : }
2783 :
2784 3 : optimizations
2785 3 : }
2786 :
2787 : /// Test the balancing behavior of shard scheduling: that it achieves a balance, and
2788 : /// that it converges.
2789 : #[test]
2790 1 : fn optimize_add_nodes() -> anyhow::Result<()> {
2791 1 : let nodes = make_test_nodes(
2792 : 9,
2793 1 : &[
2794 1 : // Initial 6 nodes
2795 1 : AvailabilityZone("az-a".to_string()),
2796 1 : AvailabilityZone("az-a".to_string()),
2797 1 : AvailabilityZone("az-b".to_string()),
2798 1 : AvailabilityZone("az-b".to_string()),
2799 1 : AvailabilityZone("az-c".to_string()),
2800 1 : AvailabilityZone("az-c".to_string()),
2801 1 : // Three we will add later
2802 1 : AvailabilityZone("az-a".to_string()),
2803 1 : AvailabilityZone("az-b".to_string()),
2804 1 : AvailabilityZone("az-c".to_string()),
2805 1 : ],
2806 : );
2807 :
2808 : // Only show the scheduler two nodes in each AZ to start with
2809 1 : let mut scheduler = Scheduler::new([].iter());
2810 7 : for i in 1..=6 {
2811 6 : scheduler.node_upsert(nodes.get(&NodeId(i)).unwrap());
2812 6 : }
2813 :
2814 1 : let mut shards = make_test_tenant(
2815 1 : PlacementPolicy::Attached(1),
2816 1 : ShardCount::new(4),
2817 1 : Some(AvailabilityZone("az-a".to_string())),
2818 : );
2819 1 : let mut schedule_context = ScheduleContext::default();
2820 5 : for shard in &mut shards {
2821 4 : assert!(
2822 4 : shard
2823 4 : .schedule(&mut scheduler, &mut schedule_context)
2824 4 : .is_ok()
2825 : );
2826 : }
2827 :
2828 : // Initial: attached locations land in the tenant's home AZ.
2829 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 2);
2830 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 2);
2831 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
2832 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
2833 :
2834 : // Initial: secondary locations in a remote AZ
2835 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 1);
2836 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 0);
2837 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 1);
2838 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 0);
2839 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(5)), 1);
2840 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(5)), 0);
2841 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(6)), 1);
2842 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(6)), 0);
2843 :
2844 : // Add another three nodes: we should see the shards spread out when their optimize
2845 : // methods are called
2846 1 : scheduler.node_upsert(nodes.get(&NodeId(7)).unwrap());
2847 1 : scheduler.node_upsert(nodes.get(&NodeId(8)).unwrap());
2848 1 : scheduler.node_upsert(nodes.get(&NodeId(9)).unwrap());
2849 1 : optimize_til_idle(&mut scheduler, &mut shards);
2850 :
2851 : // We expect one attached location was moved to the new node in the tenant's home AZ
2852 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(7)), 1);
2853 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(7)), 1);
2854 : // The original node has one less attached shard
2855 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 1);
2856 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 1);
2857 :
2858 : // One of the original nodes still has two attachments, since there are an odd number of nodes
2859 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
2860 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
2861 :
2862 : // None of our secondaries moved, since we already had enough nodes for those to be
2863 : // scheduled perfectly
2864 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 1);
2865 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 0);
2866 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 1);
2867 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 0);
2868 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(5)), 1);
2869 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(5)), 0);
2870 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(6)), 1);
2871 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(6)), 0);
2872 :
2873 4 : for shard in shards.iter_mut() {
2874 4 : shard.intent.clear(&mut scheduler);
2875 4 : }
2876 :
2877 1 : Ok(())
2878 1 : }
2879 :
2880 : /// Test that initial shard scheduling is optimal. By optimal we mean
2881 : /// that the optimizer cannot find a way to improve it.
2882 : ///
2883 : /// This test is an example of the scheduling issue described in
2884 : /// https://github.com/neondatabase/neon/issues/8969
2885 : #[test]
2886 1 : fn initial_scheduling_is_optimal() -> anyhow::Result<()> {
2887 : use itertools::Itertools;
2888 :
2889 1 : let nodes = make_test_nodes(2, &[]);
2890 :
2891 1 : let mut scheduler = Scheduler::new([].iter());
2892 1 : scheduler.node_upsert(nodes.get(&NodeId(1)).unwrap());
2893 1 : scheduler.node_upsert(nodes.get(&NodeId(2)).unwrap());
2894 :
2895 1 : let mut a = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4), None);
2896 1 : let a_context = Rc::new(RefCell::new(ScheduleContext::default()));
2897 :
2898 1 : let mut b = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4), None);
2899 1 : let b_context = Rc::new(RefCell::new(ScheduleContext::default()));
2900 :
2901 4 : let a_shards_with_context = a.iter_mut().map(|shard| (shard, a_context.clone()));
2902 4 : let b_shards_with_context = b.iter_mut().map(|shard| (shard, b_context.clone()));
2903 :
2904 1 : let schedule_order = a_shards_with_context.interleave(b_shards_with_context);
2905 :
2906 9 : for (shard, context) in schedule_order {
2907 8 : let context = &mut *context.borrow_mut();
2908 8 : shard.schedule(&mut scheduler, context).unwrap();
2909 8 : }
2910 :
2911 1 : let applied_to_a = optimize_til_idle(&mut scheduler, &mut a);
2912 1 : assert_eq!(applied_to_a, vec![]);
2913 :
2914 1 : let applied_to_b = optimize_til_idle(&mut scheduler, &mut b);
2915 1 : assert_eq!(applied_to_b, vec![]);
2916 :
2917 8 : for shard in a.iter_mut().chain(b.iter_mut()) {
2918 8 : shard.intent.clear(&mut scheduler);
2919 8 : }
2920 :
2921 1 : Ok(())
2922 1 : }
2923 :
2924 : #[test]
2925 1 : fn random_az_shard_scheduling() -> anyhow::Result<()> {
2926 : use rand::seq::SliceRandom;
2927 :
2928 51 : for seed in 0..50 {
2929 50 : eprintln!("Running test with seed {seed}");
2930 50 : let mut rng = StdRng::seed_from_u64(seed);
2931 :
2932 50 : let az_a_tag = AvailabilityZone("az-a".to_string());
2933 50 : let az_b_tag = AvailabilityZone("az-b".to_string());
2934 50 : let azs = [az_a_tag, az_b_tag];
2935 50 : let nodes = make_test_nodes(4, &azs);
2936 50 : let mut shards_per_az: HashMap<AvailabilityZone, u32> = HashMap::new();
2937 :
2938 50 : let mut scheduler = Scheduler::new([].iter());
2939 200 : for node in nodes.values() {
2940 200 : scheduler.node_upsert(node);
2941 200 : }
2942 :
2943 50 : let mut shards = Vec::default();
2944 50 : let mut contexts = Vec::default();
2945 50 : let mut az_picker = azs.iter().cycle().cloned();
2946 5050 : for i in 0..100 {
2947 5000 : let az = az_picker.next().unwrap();
2948 5000 : let shard_count = i % 4 + 1;
2949 5000 : *shards_per_az.entry(az.clone()).or_default() += shard_count;
2950 :
2951 5000 : let tenant_shards = make_test_tenant(
2952 5000 : PlacementPolicy::Attached(1),
2953 5000 : ShardCount::new(shard_count.try_into().unwrap()),
2954 5000 : Some(az),
2955 : );
2956 5000 : let context = Rc::new(RefCell::new(ScheduleContext::default()));
2957 :
2958 5000 : contexts.push(context.clone());
2959 5000 : let with_ctx = tenant_shards
2960 5000 : .into_iter()
2961 12500 : .map(|shard| (shard, context.clone()));
2962 17500 : for shard_with_ctx in with_ctx {
2963 12500 : shards.push(shard_with_ctx);
2964 12500 : }
2965 : }
2966 :
2967 50 : shards.shuffle(&mut rng);
2968 :
2969 : #[derive(Default, Debug)]
2970 : struct NodeStats {
2971 : attachments: u32,
2972 : secondaries: u32,
2973 : }
2974 :
2975 50 : let mut node_stats: HashMap<NodeId, NodeStats> = HashMap::default();
2976 50 : let mut attachments_in_wrong_az = 0;
2977 50 : let mut secondaries_in_wrong_az = 0;
2978 :
2979 12550 : for (shard, context) in &mut shards {
2980 12500 : let context = &mut *context.borrow_mut();
2981 12500 : shard.schedule(&mut scheduler, context).unwrap();
2982 :
2983 12500 : let attached_node = shard.intent.get_attached().unwrap();
2984 12500 : let stats = node_stats.entry(attached_node).or_default();
2985 12500 : stats.attachments += 1;
2986 :
2987 12500 : let secondary_node = *shard.intent.get_secondary().first().unwrap();
2988 12500 : let stats = node_stats.entry(secondary_node).or_default();
2989 12500 : stats.secondaries += 1;
2990 :
2991 12500 : let attached_node_az = nodes
2992 12500 : .get(&attached_node)
2993 12500 : .unwrap()
2994 12500 : .get_availability_zone_id();
2995 12500 : let secondary_node_az = nodes
2996 12500 : .get(&secondary_node)
2997 12500 : .unwrap()
2998 12500 : .get_availability_zone_id();
2999 12500 : let preferred_az = shard.preferred_az().unwrap();
3000 :
3001 12500 : if attached_node_az != preferred_az {
3002 0 : eprintln!(
3003 0 : "{} attachment was scheduled in AZ {} but preferred AZ {}",
3004 0 : shard.tenant_shard_id, attached_node_az, preferred_az
3005 0 : );
3006 0 : attachments_in_wrong_az += 1;
3007 12500 : }
3008 :
3009 12500 : if secondary_node_az == preferred_az {
3010 0 : eprintln!(
3011 0 : "{} secondary was scheduled in AZ {} which matches preference",
3012 0 : shard.tenant_shard_id, attached_node_az
3013 0 : );
3014 0 : secondaries_in_wrong_az += 1;
3015 12500 : }
3016 : }
3017 :
3018 50 : let mut violations = Vec::default();
3019 :
3020 50 : if attachments_in_wrong_az > 0 {
3021 0 : violations.push(format!(
3022 0 : "{attachments_in_wrong_az} attachments scheduled to the incorrect AZ"
3023 0 : ));
3024 50 : }
3025 :
3026 50 : if secondaries_in_wrong_az > 0 {
3027 0 : violations.push(format!(
3028 0 : "{secondaries_in_wrong_az} secondaries scheduled to the incorrect AZ"
3029 0 : ));
3030 50 : }
3031 :
3032 50 : eprintln!(
3033 50 : "attachments_in_wrong_az={attachments_in_wrong_az} secondaries_in_wrong_az={secondaries_in_wrong_az}"
3034 : );
3035 :
3036 250 : for (node_id, stats) in &node_stats {
3037 200 : let node_az = nodes.get(node_id).unwrap().get_availability_zone_id();
3038 200 : let ideal_attachment_load = shards_per_az.get(node_az).unwrap() / 2;
3039 200 : let allowed_attachment_load =
3040 200 : (ideal_attachment_load - 1)..(ideal_attachment_load + 2);
3041 :
3042 200 : if !allowed_attachment_load.contains(&stats.attachments) {
3043 0 : violations.push(format!(
3044 0 : "Found {} attachments on node {}, but expected {}",
3045 0 : stats.attachments, node_id, ideal_attachment_load
3046 0 : ));
3047 200 : }
3048 :
3049 200 : eprintln!(
3050 200 : "{}: attachments={} secondaries={} ideal_attachment_load={}",
3051 : node_id, stats.attachments, stats.secondaries, ideal_attachment_load
3052 : );
3053 : }
3054 :
3055 50 : assert!(violations.is_empty(), "{violations:?}");
3056 :
3057 12550 : for (mut shard, _ctx) in shards {
3058 12500 : shard.intent.clear(&mut scheduler);
3059 12500 : }
3060 : }
3061 1 : Ok(())
3062 1 : }
3063 :
3064 : /// Check how the shard's scheduling behaves when in PlacementPolicy::Secondary mode.
3065 : #[test]
3066 1 : fn tenant_secondary_scheduling() -> anyhow::Result<()> {
3067 1 : let az_a = AvailabilityZone("az-a".to_string());
3068 1 : let nodes = make_test_nodes(
3069 : 3,
3070 1 : &[
3071 1 : az_a.clone(),
3072 1 : AvailabilityZone("az-b".to_string()),
3073 1 : AvailabilityZone("az-c".to_string()),
3074 1 : ],
3075 : );
3076 :
3077 1 : let mut scheduler = Scheduler::new(nodes.values());
3078 1 : let mut context = ScheduleContext::default();
3079 :
3080 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Secondary);
3081 1 : tenant_shard.intent.preferred_az_id = Some(az_a.clone());
3082 1 : tenant_shard
3083 1 : .schedule(&mut scheduler, &mut context)
3084 1 : .expect("we have enough nodes, scheduling should work");
3085 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
3086 1 : assert!(tenant_shard.intent.attached.is_none());
3087 :
3088 : // Should have scheduled into the preferred AZ
3089 1 : assert_eq!(
3090 1 : scheduler
3091 1 : .get_node_az(&tenant_shard.intent.secondary[0])
3092 1 : .as_ref(),
3093 1 : tenant_shard.preferred_az()
3094 : );
3095 :
3096 : // Optimizer should agree
3097 1 : assert_eq!(
3098 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
3099 : None
3100 : );
3101 1 : assert_eq!(
3102 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
3103 : None
3104 : );
3105 :
3106 : // Switch to PlacementPolicy::Attached
3107 1 : tenant_shard.policy = PlacementPolicy::Attached(1);
3108 1 : tenant_shard
3109 1 : .schedule(&mut scheduler, &mut context)
3110 1 : .expect("we have enough nodes, scheduling should work");
3111 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
3112 1 : assert!(tenant_shard.intent.attached.is_some());
3113 : // Secondary should now be in non-preferred AZ
3114 1 : assert_ne!(
3115 1 : scheduler
3116 1 : .get_node_az(&tenant_shard.intent.secondary[0])
3117 1 : .as_ref(),
3118 1 : tenant_shard.preferred_az()
3119 : );
3120 : // Attached should be in preferred AZ
3121 1 : assert_eq!(
3122 1 : scheduler
3123 1 : .get_node_az(&tenant_shard.intent.attached.unwrap())
3124 1 : .as_ref(),
3125 1 : tenant_shard.preferred_az()
3126 : );
3127 :
3128 : // Optimizer should agree
3129 1 : assert_eq!(
3130 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
3131 : None
3132 : );
3133 1 : assert_eq!(
3134 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
3135 : None
3136 : );
3137 :
3138 : // Switch back to PlacementPolicy::Secondary
3139 1 : tenant_shard.policy = PlacementPolicy::Secondary;
3140 1 : tenant_shard
3141 1 : .schedule(&mut scheduler, &mut context)
3142 1 : .expect("we have enough nodes, scheduling should work");
3143 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
3144 1 : assert!(tenant_shard.intent.attached.is_none());
3145 : // When we picked a location to keep, we should have kept the one in the preferred AZ
3146 1 : assert_eq!(
3147 1 : scheduler
3148 1 : .get_node_az(&tenant_shard.intent.secondary[0])
3149 1 : .as_ref(),
3150 1 : tenant_shard.preferred_az()
3151 : );
3152 :
3153 : // Optimizer should agree
3154 1 : assert_eq!(
3155 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
3156 : None
3157 : );
3158 1 : assert_eq!(
3159 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
3160 : None
3161 : );
3162 :
3163 1 : tenant_shard.intent.clear(&mut scheduler);
3164 :
3165 1 : Ok(())
3166 1 : }
3167 : }
|
