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 : /// Amount of consecutive [`crate::service::Service::reconcile_all`] iterations that have been
135 : /// scheduled a reconciliation for this shard.
136 : ///
137 : /// If this reaches `MAX_CONSECUTIVE_RECONCILES`, the shard is considered "stuck" and will be
138 : /// ignored when deciding whether optimizations can run. This includes both successful and failed
139 : /// reconciliations.
140 : ///
141 : /// Incremented in [`crate::service::Service::process_result`], and reset to 0 when
142 : /// [`crate::service::Service::reconcile_all`] determines no reconciliation is needed for this shard.
143 : pub(crate) consecutive_reconciles_count: usize,
144 :
145 : /// If we have a pending compute notification that for some reason we weren't able to send,
146 : /// set this to true. If this is set, calls to [`Self::get_reconcile_needed`] will return Yes
147 : /// and trigger a Reconciler run. This is the mechanism by which compute notifications are included in the scope
148 : /// of state that we publish externally in an eventually consistent way.
149 : pub(crate) pending_compute_notification: bool,
150 :
151 : /// To do a graceful migration, set this field to the destination pageserver, and optimization
152 : /// functions will consider this node the best location and react appropriately.
153 : preferred_node: Option<NodeId>,
154 :
155 : // Support/debug tool: if something is going wrong or flapping with scheduling, this may
156 : // be set to a non-active state to avoid making changes while the issue is fixed.
157 : scheduling_policy: ShardSchedulingPolicy,
158 : }
159 :
160 : #[derive(Clone, Debug, Serialize)]
161 : pub(crate) struct IntentState {
162 : attached: Option<NodeId>,
163 : secondary: Vec<NodeId>,
164 :
165 : // We should attempt to schedule this shard in the provided AZ to
166 : // decrease chances of cross-AZ compute.
167 : preferred_az_id: Option<AvailabilityZone>,
168 : }
169 :
170 : impl IntentState {
171 12860 : pub(crate) fn new(preferred_az_id: Option<AvailabilityZone>) -> Self {
172 12860 : Self {
173 12860 : attached: None,
174 12860 : secondary: vec![],
175 12860 : preferred_az_id,
176 12860 : }
177 12860 : }
178 0 : pub(crate) fn single(
179 0 : scheduler: &mut Scheduler,
180 0 : node_id: Option<NodeId>,
181 0 : preferred_az_id: Option<AvailabilityZone>,
182 0 : ) -> Self {
183 0 : if let Some(node_id) = node_id {
184 0 : scheduler.update_node_ref_counts(
185 0 : node_id,
186 0 : preferred_az_id.as_ref(),
187 0 : RefCountUpdate::Attach,
188 0 : );
189 0 : }
190 0 : Self {
191 0 : attached: node_id,
192 0 : secondary: vec![],
193 0 : preferred_az_id,
194 0 : }
195 0 : }
196 :
197 12859 : pub(crate) fn set_attached(&mut self, scheduler: &mut Scheduler, new_attached: Option<NodeId>) {
198 12859 : if self.attached != new_attached {
199 12859 : if let Some(old_attached) = self.attached.take() {
200 0 : scheduler.update_node_ref_counts(
201 0 : old_attached,
202 0 : self.preferred_az_id.as_ref(),
203 0 : RefCountUpdate::Detach,
204 0 : );
205 12859 : }
206 12859 : if let Some(new_attached) = &new_attached {
207 12859 : scheduler.update_node_ref_counts(
208 12859 : *new_attached,
209 12859 : self.preferred_az_id.as_ref(),
210 12859 : RefCountUpdate::Attach,
211 12859 : );
212 12859 : }
213 12859 : self.attached = new_attached;
214 0 : }
215 :
216 12859 : if let Some(new_attached) = &new_attached {
217 12859 : assert!(!self.secondary.contains(new_attached));
218 0 : }
219 12859 : }
220 :
221 : /// Like set_attached, but the node is from [`Self::secondary`]. This swaps the node from
222 : /// secondary to attached while maintaining the scheduler's reference counts.
223 7 : pub(crate) fn promote_attached(
224 7 : &mut self,
225 7 : scheduler: &mut Scheduler,
226 7 : promote_secondary: NodeId,
227 7 : ) {
228 : // If we call this with a node that isn't in secondary, it would cause incorrect
229 : // scheduler reference counting, since we assume the node is already referenced as a secondary.
230 7 : debug_assert!(self.secondary.contains(&promote_secondary));
231 :
232 16 : self.secondary.retain(|n| n != &promote_secondary);
233 :
234 7 : let demoted = self.attached;
235 7 : self.attached = Some(promote_secondary);
236 :
237 7 : scheduler.update_node_ref_counts(
238 7 : promote_secondary,
239 7 : self.preferred_az_id.as_ref(),
240 7 : RefCountUpdate::PromoteSecondary,
241 : );
242 7 : if let Some(demoted) = demoted {
243 0 : scheduler.update_node_ref_counts(
244 0 : demoted,
245 0 : self.preferred_az_id.as_ref(),
246 0 : RefCountUpdate::DemoteAttached,
247 0 : );
248 7 : }
249 7 : }
250 :
251 12848 : pub(crate) fn push_secondary(&mut self, scheduler: &mut Scheduler, new_secondary: NodeId) {
252 12848 : assert!(!self.secondary.contains(&new_secondary));
253 12848 : assert!(self.attached != Some(new_secondary));
254 12848 : scheduler.update_node_ref_counts(
255 12848 : new_secondary,
256 12848 : self.preferred_az_id.as_ref(),
257 12848 : RefCountUpdate::AddSecondary,
258 : );
259 12848 : self.secondary.push(new_secondary);
260 12848 : }
261 :
262 : /// It is legal to call this with a node that is not currently a secondary: that is a no-op
263 9 : pub(crate) fn remove_secondary(&mut self, scheduler: &mut Scheduler, node_id: NodeId) {
264 13 : let index = self.secondary.iter().position(|n| *n == node_id);
265 9 : if let Some(index) = index {
266 9 : scheduler.update_node_ref_counts(
267 9 : node_id,
268 9 : self.preferred_az_id.as_ref(),
269 9 : RefCountUpdate::RemoveSecondary,
270 9 : );
271 9 : self.secondary.remove(index);
272 9 : }
273 9 : }
274 :
275 12859 : pub(crate) fn clear_secondary(&mut self, scheduler: &mut Scheduler) {
276 12859 : for secondary in self.secondary.drain(..) {
277 12840 : scheduler.update_node_ref_counts(
278 12840 : secondary,
279 12840 : self.preferred_az_id.as_ref(),
280 12840 : RefCountUpdate::RemoveSecondary,
281 12840 : );
282 12840 : }
283 12859 : }
284 :
285 : /// Remove the last secondary node from the list of secondaries
286 0 : pub(crate) fn pop_secondary(&mut self, scheduler: &mut Scheduler) {
287 0 : if let Some(node_id) = self.secondary.pop() {
288 0 : scheduler.update_node_ref_counts(
289 0 : node_id,
290 0 : self.preferred_az_id.as_ref(),
291 0 : RefCountUpdate::RemoveSecondary,
292 0 : );
293 0 : }
294 0 : }
295 :
296 12859 : pub(crate) fn clear(&mut self, scheduler: &mut Scheduler) {
297 12859 : if let Some(old_attached) = self.attached.take() {
298 12857 : scheduler.update_node_ref_counts(
299 12857 : old_attached,
300 12857 : self.preferred_az_id.as_ref(),
301 12857 : RefCountUpdate::Detach,
302 12857 : );
303 12857 : }
304 :
305 12859 : self.clear_secondary(scheduler);
306 12859 : }
307 :
308 12901 : pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
309 12901 : let mut result = Vec::new();
310 12901 : if let Some(p) = self.attached {
311 12896 : result.push(p)
312 5 : }
313 :
314 12901 : result.extend(self.secondary.iter().copied());
315 :
316 12901 : result
317 12901 : }
318 :
319 12624 : pub(crate) fn get_attached(&self) -> &Option<NodeId> {
320 12624 : &self.attached
321 12624 : }
322 :
323 12706 : pub(crate) fn get_secondary(&self) -> &Vec<NodeId> {
324 12706 : &self.secondary
325 12706 : }
326 :
327 : /// If the node is in use as the attached location, demote it into
328 : /// the list of secondary locations. This is used when a node goes offline,
329 : /// and we want to use a different node for attachment, but not permanently
330 : /// forget the location on the offline node.
331 : ///
332 : /// Returns true if a change was made
333 8 : pub(crate) fn demote_attached(&mut self, scheduler: &mut Scheduler, node_id: NodeId) -> bool {
334 8 : if self.attached == Some(node_id) {
335 8 : self.attached = None;
336 8 : self.secondary.push(node_id);
337 8 : scheduler.update_node_ref_counts(
338 8 : node_id,
339 8 : self.preferred_az_id.as_ref(),
340 8 : RefCountUpdate::DemoteAttached,
341 : );
342 8 : true
343 : } else {
344 0 : false
345 : }
346 8 : }
347 :
348 2 : pub(crate) fn set_preferred_az(
349 2 : &mut self,
350 2 : scheduler: &mut Scheduler,
351 2 : preferred_az: Option<AvailabilityZone>,
352 2 : ) {
353 2 : let new_az = preferred_az.as_ref();
354 2 : let old_az = self.preferred_az_id.as_ref();
355 :
356 2 : if old_az != new_az {
357 2 : if let Some(node_id) = self.attached {
358 2 : scheduler.update_node_ref_counts(
359 2 : node_id,
360 2 : new_az,
361 2 : RefCountUpdate::ChangePreferredAzFrom(old_az),
362 2 : );
363 2 : }
364 4 : for node_id in &self.secondary {
365 2 : scheduler.update_node_ref_counts(
366 2 : *node_id,
367 2 : new_az,
368 2 : RefCountUpdate::ChangePreferredAzFrom(old_az),
369 2 : );
370 2 : }
371 2 : self.preferred_az_id = preferred_az;
372 0 : }
373 2 : }
374 :
375 12508 : pub(crate) fn get_preferred_az(&self) -> Option<&AvailabilityZone> {
376 12508 : self.preferred_az_id.as_ref()
377 12508 : }
378 : }
379 :
380 : impl Drop for IntentState {
381 12860 : fn drop(&mut self) {
382 : // Must clear before dropping, to avoid leaving stale refcounts in the Scheduler.
383 : // We do not check this while panicking, to avoid polluting unit test failures or
384 : // other assertions with this assertion's output. It's still wrong to leak these,
385 : // but if we already have a panic then we don't need to independently flag this case.
386 12860 : if !(std::thread::panicking()) {
387 12860 : debug_assert!(self.attached.is_none() && self.secondary.is_empty());
388 0 : }
389 12859 : }
390 : }
391 :
392 0 : #[derive(Default, Clone, Serialize, Deserialize, Debug)]
393 : pub(crate) struct ObservedState {
394 : pub(crate) locations: HashMap<NodeId, ObservedStateLocation>,
395 : }
396 :
397 : /// Our latest knowledge of how this tenant is configured in the outside world.
398 : ///
399 : /// Meaning:
400 : /// * No instance of this type exists for a node: we are certain that we have nothing configured on that
401 : /// node for this shard.
402 : /// * Instance exists with conf==None: we *might* have some state on that node, but we don't know
403 : /// what it is (e.g. we failed partway through configuring it)
404 : /// * Instance exists with conf==Some: this tells us what we last successfully configured on this node,
405 : /// and that configuration will still be present unless something external interfered.
406 0 : #[derive(Clone, Serialize, Deserialize, Debug)]
407 : pub(crate) struct ObservedStateLocation {
408 : /// If None, it means we do not know the status of this shard's location on this node, but
409 : /// we know that we might have some state on this node.
410 : pub(crate) conf: Option<LocationConfig>,
411 : }
412 :
413 : pub(crate) struct ReconcilerWaiter {
414 : // For observability purposes, remember the ID of the shard we're
415 : // waiting for.
416 : pub(crate) tenant_shard_id: TenantShardId,
417 :
418 : seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
419 : error_seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
420 : error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
421 : seq: Sequence,
422 : }
423 :
424 : pub(crate) enum ReconcilerStatus {
425 : Done,
426 : Failed,
427 : InProgress,
428 : }
429 :
430 : #[derive(thiserror::Error, Debug)]
431 : pub(crate) enum ReconcileWaitError {
432 : #[error("Timeout waiting for shard {0}")]
433 : Timeout(TenantShardId),
434 : #[error("shutting down")]
435 : Shutdown,
436 : #[error("Reconcile error on shard {0}: {1}")]
437 : Failed(TenantShardId, Arc<ReconcileError>),
438 : }
439 :
440 : #[derive(Eq, PartialEq, Debug, Clone)]
441 : pub(crate) struct ReplaceSecondary {
442 : old_node_id: NodeId,
443 : new_node_id: NodeId,
444 : }
445 :
446 : #[derive(Eq, PartialEq, Debug, Clone)]
447 : pub(crate) struct MigrateAttachment {
448 : pub(crate) old_attached_node_id: NodeId,
449 : pub(crate) new_attached_node_id: NodeId,
450 : }
451 :
452 : #[derive(Eq, PartialEq, Debug, Clone)]
453 : pub(crate) enum ScheduleOptimizationAction {
454 : // Replace one of our secondary locations with a different node
455 : ReplaceSecondary(ReplaceSecondary),
456 : // Migrate attachment to an existing secondary location
457 : MigrateAttachment(MigrateAttachment),
458 : // Create a secondary location, with the intent of later migrating to it
459 : CreateSecondary(NodeId),
460 : // Remove a secondary location that we previously created to facilitate a migration
461 : RemoveSecondary(NodeId),
462 : }
463 :
464 : #[derive(Eq, PartialEq, Debug, Clone)]
465 : pub(crate) struct ScheduleOptimization {
466 : // What was the reconcile sequence when we generated this optimization? The optimization
467 : // should only be applied if the shard's sequence is still at this value, in case other changes
468 : // happened between planning the optimization and applying it.
469 : sequence: Sequence,
470 :
471 : pub(crate) action: ScheduleOptimizationAction,
472 : }
473 :
474 : impl ReconcilerWaiter {
475 0 : pub(crate) async fn wait_timeout(&self, timeout: Duration) -> Result<(), ReconcileWaitError> {
476 0 : tokio::select! {
477 0 : result = self.seq_wait.wait_for_timeout(self.seq, timeout)=> {
478 0 : result.map_err(|e| match e {
479 0 : SeqWaitError::Timeout => ReconcileWaitError::Timeout(self.tenant_shard_id),
480 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown
481 0 : })?;
482 : },
483 0 : result = self.error_seq_wait.wait_for(self.seq) => {
484 0 : result.map_err(|e| match e {
485 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown,
486 0 : SeqWaitError::Timeout => unreachable!()
487 0 : })?;
488 :
489 0 : return Err(ReconcileWaitError::Failed(self.tenant_shard_id,
490 0 : self.error.lock().unwrap().clone().expect("If error_seq_wait was advanced error was set").clone()))
491 : }
492 : }
493 :
494 0 : Ok(())
495 0 : }
496 :
497 0 : pub(crate) fn get_status(&self) -> ReconcilerStatus {
498 0 : if self.seq_wait.would_wait_for(self.seq).is_ok() {
499 0 : ReconcilerStatus::Done
500 0 : } else if self.error_seq_wait.would_wait_for(self.seq).is_ok() {
501 0 : ReconcilerStatus::Failed
502 : } else {
503 0 : ReconcilerStatus::InProgress
504 : }
505 0 : }
506 : }
507 :
508 : /// Having spawned a reconciler task, the tenant shard's state will carry enough
509 : /// information to optionally cancel & await it later.
510 : pub(crate) struct ReconcilerHandle {
511 : sequence: Sequence,
512 : handle: JoinHandle<()>,
513 : cancel: CancellationToken,
514 : }
515 :
516 : pub(crate) enum ReconcileNeeded {
517 : /// shard either doesn't need reconciliation, or is forbidden from spawning a reconciler
518 : /// in its current state (e.g. shard split in progress, or ShardSchedulingPolicy forbids it)
519 : No,
520 : /// shard has a reconciler running, and its intent hasn't changed since that one was
521 : /// spawned: wait for the existing reconciler rather than spawning a new one.
522 : WaitExisting(ReconcilerWaiter),
523 : /// shard needs reconciliation: call into [`TenantShard::spawn_reconciler`]
524 : Yes(ReconcileReason),
525 : }
526 :
527 : #[derive(Debug)]
528 : pub(crate) enum ReconcileReason {
529 : ActiveNodesDirty,
530 : UnknownLocation,
531 : PendingComputeNotification,
532 : }
533 :
534 : /// Pending modification to the observed state of a tenant shard.
535 : /// Produced by [`Reconciler::observed_deltas`] and applied in [`crate::service::Service::process_result`].
536 : pub(crate) enum ObservedStateDelta {
537 : Upsert(Box<(NodeId, ObservedStateLocation)>),
538 : Delete(NodeId),
539 : }
540 :
541 : impl ObservedStateDelta {
542 0 : pub(crate) fn node_id(&self) -> &NodeId {
543 0 : match self {
544 0 : Self::Upsert(up) => &up.0,
545 0 : Self::Delete(nid) => nid,
546 : }
547 0 : }
548 : }
549 :
550 : /// When a reconcile task completes, it sends this result object
551 : /// to be applied to the primary TenantShard.
552 : pub(crate) struct ReconcileResult {
553 : pub(crate) sequence: Sequence,
554 : /// On errors, `observed` should be treated as an incompleted description
555 : /// of state (i.e. any nodes present in the result should override nodes
556 : /// present in the parent tenant state, but any unmentioned nodes should
557 : /// not be removed from parent tenant state)
558 : pub(crate) result: Result<(), ReconcileError>,
559 :
560 : pub(crate) tenant_shard_id: TenantShardId,
561 : pub(crate) generation: Option<Generation>,
562 : pub(crate) observed_deltas: Vec<ObservedStateDelta>,
563 :
564 : /// Set [`TenantShard::pending_compute_notification`] from this flag
565 : pub(crate) pending_compute_notification: bool,
566 : }
567 :
568 : impl ObservedState {
569 0 : pub(crate) fn new() -> Self {
570 0 : Self {
571 0 : locations: HashMap::new(),
572 0 : }
573 0 : }
574 :
575 0 : pub(crate) fn is_empty(&self) -> bool {
576 0 : self.locations.is_empty()
577 0 : }
578 : }
579 :
580 : impl TenantShard {
581 12843 : pub(crate) fn new(
582 12843 : tenant_shard_id: TenantShardId,
583 12843 : shard: ShardIdentity,
584 12843 : policy: PlacementPolicy,
585 12843 : preferred_az_id: Option<AvailabilityZone>,
586 12843 : ) -> Self {
587 12843 : metrics::METRICS_REGISTRY
588 12843 : .metrics_group
589 12843 : .storage_controller_tenant_shards
590 12843 : .inc();
591 :
592 12843 : Self {
593 12843 : tenant_shard_id,
594 12843 : policy,
595 12843 : intent: IntentState::new(preferred_az_id),
596 12843 : generation: Some(Generation::new(0)),
597 12843 : shard,
598 12843 : observed: ObservedState::default(),
599 12843 : config: TenantConfig::default(),
600 12843 : reconciler: None,
601 12843 : splitting: SplitState::Idle,
602 12843 : importing: TimelineImportState::Idle,
603 12843 : sequence: Sequence(1),
604 12843 : delayed_reconcile: false,
605 12843 : waiter: Arc::new(SeqWait::new(Sequence(0))),
606 12843 : error_waiter: Arc::new(SeqWait::new(Sequence(0))),
607 12843 : last_error: Arc::default(),
608 12843 : consecutive_reconciles_count: 0,
609 12843 : pending_compute_notification: false,
610 12843 : scheduling_policy: ShardSchedulingPolicy::default(),
611 12843 : preferred_node: None,
612 12843 : }
613 12843 : }
614 :
615 : /// For use on startup when learning state from pageservers: generate my [`IntentState`] from my
616 : /// [`ObservedState`], even if it violates my [`PlacementPolicy`]. Call [`Self::schedule`] next,
617 : /// to get an intent state that complies with placement policy. The overall goal is to do scheduling
618 : /// in a way that makes use of any configured locations that already exist in the outside world.
619 1 : pub(crate) fn intent_from_observed(&mut self, scheduler: &mut Scheduler) {
620 : // Choose an attached location by filtering observed locations, and then sorting to get the highest
621 : // generation
622 1 : let mut attached_locs = self
623 1 : .observed
624 1 : .locations
625 1 : .iter()
626 2 : .filter_map(|(node_id, l)| {
627 2 : if let Some(conf) = &l.conf {
628 2 : if conf.mode == LocationConfigMode::AttachedMulti
629 1 : || conf.mode == LocationConfigMode::AttachedSingle
630 1 : || conf.mode == LocationConfigMode::AttachedStale
631 : {
632 2 : Some((node_id, conf.generation))
633 : } else {
634 0 : None
635 : }
636 : } else {
637 0 : None
638 : }
639 2 : })
640 1 : .collect::<Vec<_>>();
641 :
642 1 : attached_locs.sort_by_key(|i| i.1);
643 1 : if let Some((node_id, _gen)) = attached_locs.into_iter().next_back() {
644 1 : self.intent.set_attached(scheduler, Some(*node_id));
645 1 : }
646 :
647 : // All remaining observed locations generate secondary intents. This includes None
648 : // observations, as these may well have some local content on disk that is usable (this
649 : // is an edge case that might occur if we restarted during a migration or other change)
650 : //
651 : // We may leave intent.attached empty if we didn't find any attached locations: [`Self::schedule`]
652 : // will take care of promoting one of these secondaries to be attached.
653 2 : self.observed.locations.keys().for_each(|node_id| {
654 2 : if Some(*node_id) != self.intent.attached {
655 1 : self.intent.push_secondary(scheduler, *node_id);
656 1 : }
657 2 : });
658 1 : }
659 :
660 : /// Part of [`Self::schedule`] that is used to choose exactly one node to act as the
661 : /// attached pageserver for a shard.
662 : ///
663 : /// Returns whether we modified it, and the NodeId selected.
664 12832 : fn schedule_attached(
665 12832 : &mut self,
666 12832 : scheduler: &mut Scheduler,
667 12832 : context: &ScheduleContext,
668 12832 : ) -> Result<(bool, NodeId), ScheduleError> {
669 : // No work to do if we already have an attached tenant
670 12832 : if let Some(node_id) = self.intent.attached {
671 0 : return Ok((false, node_id));
672 12832 : }
673 :
674 12832 : if let Some(promote_secondary) = self.preferred_secondary(scheduler) {
675 : // Promote a secondary
676 2 : tracing::debug!("Promoted secondary {} to attached", promote_secondary);
677 2 : self.intent.promote_attached(scheduler, promote_secondary);
678 2 : Ok((true, promote_secondary))
679 : } else {
680 : // Pick a fresh node: either we had no secondaries or none were schedulable
681 12830 : let node_id = scheduler.schedule_shard::<AttachedShardTag>(
682 12830 : &self.intent.secondary,
683 12830 : &self.intent.preferred_az_id,
684 12830 : context,
685 0 : )?;
686 12830 : tracing::debug!("Selected {} as attached", node_id);
687 12830 : self.intent.set_attached(scheduler, Some(node_id));
688 12830 : Ok((true, node_id))
689 : }
690 12832 : }
691 :
692 : #[instrument(skip_all, fields(
693 : tenant_id=%self.tenant_shard_id.tenant_id,
694 : shard_id=%self.tenant_shard_id.shard_slug(),
695 : sequence=%self.sequence
696 : ))]
697 : pub(crate) fn schedule(
698 : &mut self,
699 : scheduler: &mut Scheduler,
700 : context: &mut ScheduleContext,
701 : ) -> Result<(), ScheduleError> {
702 : let r = self.do_schedule(scheduler, context);
703 :
704 : context.avoid(&self.intent.all_pageservers());
705 :
706 : r
707 : }
708 :
709 12836 : pub(crate) fn do_schedule(
710 12836 : &mut self,
711 12836 : scheduler: &mut Scheduler,
712 12836 : context: &ScheduleContext,
713 12836 : ) -> Result<(), ScheduleError> {
714 : // TODO: before scheduling new nodes, check if any existing content in
715 : // self.intent refers to pageservers that are offline, and pick other
716 : // pageservers if so.
717 :
718 : // TODO: respect the splitting bit on tenants: if they are currently splitting then we may not
719 : // change their attach location.
720 :
721 12836 : match self.scheduling_policy {
722 12835 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {}
723 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
724 : // Warn to make it obvious why other things aren't happening/working, if we skip scheduling
725 1 : tracing::warn!(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(),
726 0 : "Scheduling is disabled by policy {:?}", self.scheduling_policy);
727 1 : return Ok(());
728 : }
729 : }
730 :
731 : // Build the set of pageservers already in use by this tenant, to avoid scheduling
732 : // more work on the same pageservers we're already using.
733 12835 : let mut modified = false;
734 :
735 : // Add/remove nodes to fulfil policy
736 : use PlacementPolicy::*;
737 12835 : match self.policy {
738 12832 : Attached(secondary_count) => {
739 : // Should have exactly one attached, and at least N secondaries
740 12832 : let (modified_attached, attached_node_id) =
741 12832 : self.schedule_attached(scheduler, context)?;
742 12832 : modified |= modified_attached;
743 :
744 12832 : let mut used_pageservers = vec![attached_node_id];
745 25663 : while self.intent.secondary.len() < secondary_count {
746 12831 : let node_id = scheduler.schedule_shard::<SecondaryShardTag>(
747 12831 : &used_pageservers,
748 12831 : &self.intent.preferred_az_id,
749 12831 : context,
750 0 : )?;
751 12831 : self.intent.push_secondary(scheduler, node_id);
752 12831 : used_pageservers.push(node_id);
753 12831 : modified = true;
754 : }
755 : }
756 : Secondary => {
757 3 : if let Some(node_id) = self.intent.get_attached() {
758 2 : // Populate secondary by demoting the attached node
759 2 : self.intent.demote_attached(scheduler, *node_id);
760 2 :
761 2 : modified = true;
762 2 : } else if self.intent.secondary.is_empty() {
763 : // Populate secondary by scheduling a fresh node
764 : //
765 : // We use [`AttachedShardTag`] because when a secondary location is the only one
766 : // a shard has, we expect that its next use will be as an attached location: we want
767 : // the tenant to be ready to warm up and run fast in their preferred AZ.
768 1 : let node_id = scheduler.schedule_shard::<AttachedShardTag>(
769 1 : &[],
770 1 : &self.intent.preferred_az_id,
771 1 : context,
772 0 : )?;
773 1 : self.intent.push_secondary(scheduler, node_id);
774 1 : modified = true;
775 0 : }
776 4 : while self.intent.secondary.len() > 1 {
777 : // If we have multiple secondaries (e.g. when transitioning from Attached to Secondary and
778 : // having just demoted our attached location), then we should prefer to keep the location
779 : // in our preferred AZ. Tenants in Secondary mode want to be in the preferred AZ so that
780 : // they have a warm location to become attached when transitioning back into Attached.
781 :
782 1 : let mut candidates = self.intent.get_secondary().clone();
783 : // Sort to get secondaries outside preferred AZ last
784 1 : candidates
785 2 : .sort_by_key(|n| scheduler.get_node_az(n).as_ref() != self.preferred_az());
786 1 : let secondary_to_remove = candidates.pop().unwrap();
787 1 : self.intent.remove_secondary(scheduler, secondary_to_remove);
788 1 : modified = true;
789 : }
790 : }
791 : Detached => {
792 : // Never add locations in this mode
793 0 : if self.intent.get_attached().is_some() || !self.intent.get_secondary().is_empty() {
794 0 : self.intent.clear(scheduler);
795 0 : modified = true;
796 0 : }
797 : }
798 : }
799 :
800 12835 : if modified {
801 12835 : self.sequence.0 += 1;
802 12835 : }
803 :
804 12835 : Ok(())
805 12836 : }
806 :
807 : /// Reschedule this tenant shard to one of its secondary locations. Returns a scheduling error
808 : /// if the swap is not possible and leaves the intent state in its original state.
809 : ///
810 : /// Arguments:
811 : /// `attached_to`: the currently attached location matching the intent state (may be None if the
812 : /// shard is not attached)
813 : /// `promote_to`: an optional secondary location of this tenant shard. If set to None, we ask
814 : /// the scheduler to recommend a node
815 0 : pub(crate) fn reschedule_to_secondary(
816 0 : &mut self,
817 0 : promote_to: Option<NodeId>,
818 0 : scheduler: &mut Scheduler,
819 0 : ) -> Result<(), ScheduleError> {
820 0 : let promote_to = match promote_to {
821 0 : Some(node) => node,
822 0 : None => match self.preferred_secondary(scheduler) {
823 0 : Some(node) => node,
824 : None => {
825 0 : return Err(ScheduleError::ImpossibleConstraint);
826 : }
827 : },
828 : };
829 :
830 0 : assert!(self.intent.get_secondary().contains(&promote_to));
831 :
832 0 : if let Some(node) = self.intent.get_attached() {
833 0 : let demoted = self.intent.demote_attached(scheduler, *node);
834 0 : if !demoted {
835 0 : return Err(ScheduleError::ImpossibleConstraint);
836 0 : }
837 0 : }
838 :
839 0 : self.intent.promote_attached(scheduler, promote_to);
840 :
841 : // Increment the sequence number for the edge case where a
842 : // reconciler is already running to avoid waiting on the
843 : // current reconcile instead of spawning a new one.
844 0 : self.sequence = self.sequence.next();
845 :
846 0 : Ok(())
847 0 : }
848 :
849 : /// Returns None if the current location's score is unavailable, i.e. cannot draw a conclusion
850 68 : fn is_better_location<T: ShardTag>(
851 68 : &self,
852 68 : scheduler: &mut Scheduler,
853 68 : schedule_context: &ScheduleContext,
854 68 : current: NodeId,
855 68 : candidate: NodeId,
856 68 : ) -> Option<bool> {
857 68 : let Some(candidate_score) = scheduler.compute_node_score::<T::Score>(
858 68 : candidate,
859 68 : &self.intent.preferred_az_id,
860 68 : schedule_context,
861 68 : ) else {
862 : // The candidate node is unavailable for scheduling or otherwise couldn't get a score
863 1 : return None;
864 : };
865 :
866 : // If the candidate is our preferred node, then it is better than the current location, as long
867 : // as it is online -- the online check is part of the score calculation we did above, so it's
868 : // important that this check comes after that one.
869 67 : if let Some(preferred) = self.preferred_node.as_ref() {
870 3 : if preferred == &candidate {
871 1 : return Some(true);
872 2 : }
873 64 : }
874 :
875 66 : match scheduler.compute_node_score::<T::Score>(
876 66 : current,
877 66 : &self.intent.preferred_az_id,
878 66 : schedule_context,
879 66 : ) {
880 66 : Some(current_score) => {
881 : // Ignore utilization components when comparing scores: we don't want to migrate
882 : // because of transient load variations, it risks making the system thrash, and
883 : // migrating for utilization requires a separate high level view of the system to
884 : // e.g. prioritize moving larger or smaller tenants, rather than arbitrarily
885 : // moving things around in the order that we hit this function.
886 66 : let candidate_score = candidate_score.for_optimization();
887 66 : let current_score = current_score.for_optimization();
888 :
889 66 : if candidate_score < current_score {
890 8 : tracing::info!(
891 0 : "Found a lower scoring location! {candidate} is better than {current} ({candidate_score:?} is better than {current_score:?})"
892 : );
893 8 : Some(true)
894 : } else {
895 : // The candidate node is no better than our current location, so don't migrate
896 58 : tracing::debug!(
897 0 : "Candidate node {candidate} is no better than our current location {current} (candidate {candidate_score:?} vs current {current_score:?})",
898 : );
899 58 : Some(false)
900 : }
901 : }
902 : None => {
903 : // The current node is unavailable for scheduling, so we can't make any sensible
904 : // decisions about optimisation. This should be a transient state -- if the node
905 : // is offline then it will get evacuated, if is blocked by a scheduling mode
906 : // then we will respect that mode by doing nothing.
907 0 : tracing::debug!("Current node {current} is unavailable for scheduling");
908 0 : None
909 : }
910 : }
911 68 : }
912 :
913 48 : pub(crate) fn find_better_location<T: ShardTag>(
914 48 : &self,
915 48 : scheduler: &mut Scheduler,
916 48 : schedule_context: &ScheduleContext,
917 48 : current: NodeId,
918 48 : hard_exclude: &[NodeId],
919 48 : ) -> Option<NodeId> {
920 : // If we have a migration hint, then that is our better location
921 48 : if let Some(hint) = self.preferred_node.as_ref() {
922 1 : if hint == ¤t {
923 0 : return None;
924 1 : }
925 :
926 1 : return Some(*hint);
927 47 : }
928 :
929 : // Look for a lower-scoring location to attach to
930 47 : let Ok(candidate_node) = scheduler.schedule_shard::<T>(
931 47 : hard_exclude,
932 47 : &self.intent.preferred_az_id,
933 47 : schedule_context,
934 47 : ) else {
935 : // A scheduling error means we have no possible candidate replacements
936 0 : tracing::debug!("No candidate node found");
937 0 : return None;
938 : };
939 :
940 47 : if candidate_node == current {
941 : // We're already at the best possible location, so don't migrate
942 24 : tracing::debug!("Candidate node {candidate_node} is already in use");
943 24 : return None;
944 23 : }
945 :
946 23 : self.is_better_location::<T>(scheduler, schedule_context, current, candidate_node)
947 23 : .and_then(|better| if better { Some(candidate_node) } else { None })
948 48 : }
949 :
950 : /// This function is an optimization, used to avoid doing large numbers of scheduling operations
951 : /// when looking for optimizations. This function uses knowledge of how scores work to do some
952 : /// fast checks for whether it may to be possible to improve a score.
953 : ///
954 : /// If we return true, it only means that optimization _might_ be possible, not that it necessarily is. If we
955 : /// return no, it definitely means that calling [`Self::optimize_attachment`] or [`Self::optimize_secondary`] would do no
956 : /// work.
957 3 : pub(crate) fn maybe_optimizable(
958 3 : &self,
959 3 : scheduler: &mut Scheduler,
960 3 : schedule_context: &ScheduleContext,
961 3 : ) -> bool {
962 : // Tenant with preferred node: check if it is not already at the preferred node
963 3 : if let Some(preferred) = self.preferred_node.as_ref() {
964 0 : if Some(preferred) != self.intent.get_attached().as_ref() {
965 0 : return true;
966 0 : }
967 3 : }
968 :
969 : // Sharded tenant: check if any locations have a nonzero affinity score
970 3 : if self.shard.count >= ShardCount(1) {
971 3 : let schedule_context = schedule_context.project_detach(self);
972 5 : for node in self.intent.all_pageservers() {
973 5 : if let Some(af) = schedule_context.nodes.get(&node) {
974 5 : if *af > AffinityScore(0) {
975 3 : return true;
976 2 : }
977 0 : }
978 : }
979 0 : }
980 :
981 : // Attached tenant: check if the attachment is outside the preferred AZ
982 0 : if let PlacementPolicy::Attached(_) = self.policy {
983 0 : if let Some(attached) = self.intent.get_attached() {
984 0 : if scheduler.get_node_az(attached) != self.intent.preferred_az_id {
985 0 : return true;
986 0 : }
987 0 : }
988 0 : }
989 :
990 : // Tenant with secondary locations: check if any are within the preferred AZ
991 0 : for secondary in self.intent.get_secondary() {
992 0 : if scheduler.get_node_az(secondary) == self.intent.preferred_az_id {
993 0 : return true;
994 0 : }
995 : }
996 :
997 : // Does the tenant have excess secondaries?
998 0 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
999 0 : return true;
1000 0 : }
1001 :
1002 : // Fall through: no optimizations possible
1003 0 : false
1004 3 : }
1005 :
1006 : /// Optimize attachments: if a shard has a secondary location that is preferable to
1007 : /// its primary location based on soft constraints, switch that secondary location
1008 : /// to be attached.
1009 : ///
1010 : /// `schedule_context` should have been populated with all shards in the tenant, including
1011 : /// the one we're trying to optimize (this function will subtract its own contribution before making scoring decisions)
1012 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1013 : pub(crate) fn optimize_attachment(
1014 : &self,
1015 : scheduler: &mut Scheduler,
1016 : schedule_context: &ScheduleContext,
1017 : ) -> Option<ScheduleOptimization> {
1018 : let attached = (*self.intent.get_attached())?;
1019 :
1020 : let schedule_context = schedule_context.project_detach(self);
1021 :
1022 : // If we already have a secondary that is higher-scoring than out current location,
1023 : // then simply migrate to it.
1024 : for secondary in self.intent.get_secondary() {
1025 : if let Some(true) = self.is_better_location::<AttachedShardTag>(
1026 : scheduler,
1027 : &schedule_context,
1028 : attached,
1029 : *secondary,
1030 : ) {
1031 : return Some(ScheduleOptimization {
1032 : sequence: self.sequence,
1033 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1034 : old_attached_node_id: attached,
1035 : new_attached_node_id: *secondary,
1036 : }),
1037 : });
1038 : }
1039 : }
1040 :
1041 : // Given that none of our current secondaries is a better location than our current
1042 : // attached location (checked above), we may trim any secondaries that are not needed
1043 : // for the placement policy.
1044 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
1045 : // This code path cleans up extra secondaries after migrating, and/or
1046 : // trims extra secondaries after a PlacementPolicy::Attached(N) was
1047 : // modified to decrease N.
1048 :
1049 : let secondary_scores = self
1050 : .intent
1051 : .get_secondary()
1052 : .iter()
1053 12 : .map(|node_id| {
1054 12 : (
1055 12 : *node_id,
1056 12 : scheduler.compute_node_score::<NodeSecondarySchedulingScore>(
1057 12 : *node_id,
1058 12 : &self.intent.preferred_az_id,
1059 12 : &schedule_context,
1060 12 : ),
1061 12 : )
1062 12 : })
1063 : .collect::<HashMap<_, _>>();
1064 :
1065 12 : if secondary_scores.iter().any(|score| score.1.is_none()) {
1066 : // Trivial case: if we only have one secondary, drop that one
1067 : if self.intent.get_secondary().len() == 1 {
1068 : return Some(ScheduleOptimization {
1069 : sequence: self.sequence,
1070 : action: ScheduleOptimizationAction::RemoveSecondary(
1071 : *self.intent.get_secondary().first().unwrap(),
1072 : ),
1073 : });
1074 : }
1075 :
1076 : // Try to find a "good" secondary to keep, without relying on scores (one or more nodes is in a state
1077 : // where its score can't be calculated), and drop the others. This enables us to make progress in
1078 : // most cases, even if some nodes are offline or have scheduling=pause set.
1079 :
1080 : debug_assert!(self.intent.attached.is_some()); // We should not make it here unless attached -- this
1081 : // logic presumes we are in a mode where we want secondaries to be in non-home AZ
1082 1 : if let Some(retain_secondary) = self.intent.get_secondary().iter().find(|n| {
1083 1 : let in_home_az = scheduler.get_node_az(n) == self.intent.preferred_az_id;
1084 1 : let is_available = secondary_scores
1085 1 : .get(n)
1086 1 : .expect("Built from same list of nodes")
1087 1 : .is_some();
1088 1 : is_available && !in_home_az
1089 1 : }) {
1090 : // Great, we found one to retain. Pick some other to drop.
1091 : if let Some(victim) = self
1092 : .intent
1093 : .get_secondary()
1094 : .iter()
1095 2 : .find(|n| n != &retain_secondary)
1096 : {
1097 : return Some(ScheduleOptimization {
1098 : sequence: self.sequence,
1099 : action: ScheduleOptimizationAction::RemoveSecondary(*victim),
1100 : });
1101 : }
1102 : }
1103 :
1104 : // Fall through: we didn't identify one to remove. This ought to be rare.
1105 : tracing::warn!(
1106 : "Keeping extra secondaries: can't determine which of {:?} to remove (some nodes offline?)",
1107 : self.intent.get_secondary()
1108 : );
1109 : } else {
1110 : let victim = secondary_scores
1111 : .iter()
1112 10 : .max_by_key(|score| score.1.unwrap())
1113 : .unwrap()
1114 : .0;
1115 : return Some(ScheduleOptimization {
1116 : sequence: self.sequence,
1117 : action: ScheduleOptimizationAction::RemoveSecondary(*victim),
1118 : });
1119 : }
1120 : }
1121 :
1122 : let replacement = self.find_better_location::<AttachedShardTag>(
1123 : scheduler,
1124 : &schedule_context,
1125 : attached,
1126 : &[], // Don't exclude secondaries: our preferred attachment location may be a secondary
1127 : );
1128 :
1129 : // We have found a candidate and confirmed that its score is preferable
1130 : // to our current location. See if we have a secondary location in the preferred location already: if not,
1131 : // then create one.
1132 : if let Some(replacement) = replacement {
1133 : // If we are currently in non-preferred AZ, then the scheduler might suggest a location that is better, but still
1134 : // not in our preferred AZ. Migration has a cost in resources an impact to the workload, so we want to avoid doing
1135 : // multiple hops where we might go to some other AZ before eventually finding a suitable location in our preferred
1136 : // AZ: skip this optimization if it is not in our final, preferred AZ.
1137 : //
1138 : // This should be a transient state, there should always be capacity eventually in our preferred AZ (even if nodes
1139 : // there are too overloaded for scheduler to suggest them, more should be provisioned eventually).
1140 : if self.preferred_node.is_none()
1141 : && self.intent.preferred_az_id.is_some()
1142 : && scheduler.get_node_az(&replacement) != self.intent.preferred_az_id
1143 : {
1144 : tracing::debug!(
1145 : "Candidate node {replacement} is not in preferred AZ {:?}",
1146 : self.intent.preferred_az_id
1147 : );
1148 :
1149 : // This should only happen if our current location is not in the preferred AZ, otherwise
1150 : // [`Self::find_better_location`]` should have rejected any other location outside the preferred Az, because
1151 : // AZ is the highest priority part of NodeAttachmentSchedulingScore.
1152 : debug_assert!(scheduler.get_node_az(&attached) != self.intent.preferred_az_id);
1153 :
1154 : return None;
1155 : }
1156 :
1157 : if !self.intent.get_secondary().contains(&replacement) {
1158 : Some(ScheduleOptimization {
1159 : sequence: self.sequence,
1160 : action: ScheduleOptimizationAction::CreateSecondary(replacement),
1161 : })
1162 : } else {
1163 : // We already have a secondary in the preferred location, let's try migrating to it. Our caller
1164 : // will check the warmth of the destination before deciding whether to really execute this.
1165 : Some(ScheduleOptimization {
1166 : sequence: self.sequence,
1167 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1168 : old_attached_node_id: attached,
1169 : new_attached_node_id: replacement,
1170 : }),
1171 : })
1172 : }
1173 : } else {
1174 : // We didn't find somewhere we'd rather be, and we don't have any excess secondaries
1175 : // to clean up: no action required.
1176 : None
1177 : }
1178 : }
1179 :
1180 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1181 : pub(crate) fn optimize_secondary(
1182 : &self,
1183 : scheduler: &mut Scheduler,
1184 : schedule_context: &ScheduleContext,
1185 : ) -> Option<ScheduleOptimization> {
1186 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
1187 : // We have extra secondaries, perhaps to facilitate a migration of the attached location:
1188 : // do nothing, it is up to [`Self::optimize_attachment`] to clean them up. When that's done,
1189 : // and we are called again, we will proceed.
1190 : tracing::debug!("Too many secondaries: skipping");
1191 : return None;
1192 : }
1193 :
1194 : let schedule_context = schedule_context.project_detach(self);
1195 :
1196 : for secondary in self.intent.get_secondary() {
1197 : // Make sure we don't try to migrate a secondary to our attached location: this case happens
1198 : // easily in environments without multiple AZs.
1199 : let mut exclude = match self.intent.attached {
1200 : Some(attached) => vec![attached],
1201 : None => vec![],
1202 : };
1203 :
1204 : // Exclude all other secondaries from the scheduling process to avoid replacing
1205 : // one existing secondary with another existing secondary.
1206 : for another_secondary in self.intent.secondary.iter() {
1207 : if another_secondary != secondary {
1208 : exclude.push(*another_secondary);
1209 : }
1210 : }
1211 :
1212 : let replacement = match &self.policy {
1213 : PlacementPolicy::Attached(_) => {
1214 : // Secondaries for an attached shard should be scheduled using `SecondaryShardTag`
1215 : // to avoid placing them in the preferred AZ.
1216 : self.find_better_location::<SecondaryShardTag>(
1217 : scheduler,
1218 : &schedule_context,
1219 : *secondary,
1220 : &exclude,
1221 : )
1222 : }
1223 : PlacementPolicy::Secondary => {
1224 : // In secondary-only mode, we want our secondary locations in the preferred AZ,
1225 : // so that they're ready to take over as an attached location when we transition
1226 : // into PlacementPolicy::Attached.
1227 : self.find_better_location::<AttachedShardTag>(
1228 : scheduler,
1229 : &schedule_context,
1230 : *secondary,
1231 : &exclude,
1232 : )
1233 : }
1234 : PlacementPolicy::Detached => None,
1235 : };
1236 :
1237 : assert!(replacement != Some(*secondary));
1238 : if let Some(replacement) = replacement {
1239 : // We have found a candidate and confirmed that its score is preferable
1240 : // to our current location. See if we have a secondary location in the preferred location already: if not,
1241 : // then create one.
1242 : return Some(ScheduleOptimization {
1243 : sequence: self.sequence,
1244 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1245 : old_node_id: *secondary,
1246 : new_node_id: replacement,
1247 : }),
1248 : });
1249 : }
1250 : }
1251 :
1252 : None
1253 : }
1254 :
1255 : /// Start or abort a graceful migration of this shard to another pageserver. This works on top of the
1256 : /// other optimisation functions, to bias them to move to the destination node.
1257 0 : pub(crate) fn set_preferred_node(&mut self, node: Option<NodeId>) {
1258 0 : if let Some(hint) = self.preferred_node.as_ref() {
1259 0 : if Some(hint) != node.as_ref() {
1260 : // This is legal but a bit surprising: we expect that administrators wouldn't usually
1261 : // change their mind about where to migrate something.
1262 0 : tracing::warn!(
1263 0 : "Changing migration destination from {hint} to {node:?} (current intent {:?})",
1264 : self.intent
1265 : );
1266 0 : }
1267 0 : }
1268 :
1269 0 : self.preferred_node = node;
1270 0 : }
1271 :
1272 0 : pub(crate) fn get_preferred_node(&self) -> Option<NodeId> {
1273 0 : self.preferred_node
1274 0 : }
1275 :
1276 : /// Return true if the optimization was really applied: it will not be applied if the optimization's
1277 : /// sequence is behind this tenant shard's or if the intent state proposed by the optimization
1278 : /// is not compatible with the current intent state. The later may happen when the background
1279 : /// reconcile loops runs concurrently with HTTP driven optimisations.
1280 17 : pub(crate) fn apply_optimization(
1281 17 : &mut self,
1282 17 : scheduler: &mut Scheduler,
1283 17 : optimization: ScheduleOptimization,
1284 17 : ) -> bool {
1285 17 : if optimization.sequence != self.sequence {
1286 0 : return false;
1287 17 : }
1288 :
1289 17 : if !self.validate_optimization(&optimization) {
1290 0 : tracing::info!(
1291 0 : "Skipping optimization for {} because it does not match current intent: {:?}",
1292 : self.tenant_shard_id,
1293 : optimization,
1294 : );
1295 0 : return false;
1296 17 : }
1297 :
1298 17 : metrics::METRICS_REGISTRY
1299 17 : .metrics_group
1300 17 : .storage_controller_schedule_optimization
1301 17 : .inc();
1302 :
1303 17 : match optimization.action {
1304 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1305 5 : old_attached_node_id,
1306 5 : new_attached_node_id,
1307 : }) => {
1308 5 : self.intent.demote_attached(scheduler, old_attached_node_id);
1309 5 : self.intent
1310 5 : .promote_attached(scheduler, new_attached_node_id);
1311 :
1312 5 : if let Some(hint) = self.preferred_node.as_ref() {
1313 1 : if hint == &new_attached_node_id {
1314 : // The migration target is not a long term pin: once we are done with the migration, clear it.
1315 1 : tracing::info!("Graceful migration to {hint} complete");
1316 1 : self.preferred_node = None;
1317 0 : }
1318 4 : }
1319 : }
1320 : ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1321 1 : old_node_id,
1322 1 : new_node_id,
1323 1 : }) => {
1324 1 : self.intent.remove_secondary(scheduler, old_node_id);
1325 1 : self.intent.push_secondary(scheduler, new_node_id);
1326 1 : }
1327 4 : ScheduleOptimizationAction::CreateSecondary(new_node_id) => {
1328 4 : self.intent.push_secondary(scheduler, new_node_id);
1329 4 : }
1330 7 : ScheduleOptimizationAction::RemoveSecondary(old_secondary) => {
1331 7 : self.intent.remove_secondary(scheduler, old_secondary);
1332 7 : }
1333 : }
1334 :
1335 17 : true
1336 17 : }
1337 :
1338 : /// Check that the desired modifications to the intent state are compatible with
1339 : /// the current intent state
1340 17 : fn validate_optimization(&self, optimization: &ScheduleOptimization) -> bool {
1341 17 : match optimization.action {
1342 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1343 5 : old_attached_node_id,
1344 5 : new_attached_node_id,
1345 : }) => {
1346 5 : self.intent.attached == Some(old_attached_node_id)
1347 5 : && self.intent.secondary.contains(&new_attached_node_id)
1348 : }
1349 : ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1350 : old_node_id: _,
1351 1 : new_node_id,
1352 : }) => {
1353 : // It's legal to remove a secondary that is not present in the intent state
1354 1 : !self.intent.secondary.contains(&new_node_id)
1355 : }
1356 4 : ScheduleOptimizationAction::CreateSecondary(new_node_id) => {
1357 4 : !self.intent.secondary.contains(&new_node_id)
1358 : }
1359 : ScheduleOptimizationAction::RemoveSecondary(_) => {
1360 : // It's legal to remove a secondary that is not present in the intent state
1361 7 : true
1362 : }
1363 : }
1364 17 : }
1365 :
1366 : /// When a shard has several secondary locations, we need to pick one in situations where
1367 : /// we promote one of them to an attached location:
1368 : /// - When draining a node for restart
1369 : /// - When responding to a node failure
1370 : ///
1371 : /// In this context, 'preferred' does not mean the node with the best scheduling score: instead
1372 : /// we want to pick the node which is best for use _temporarily_ while the previous attached location
1373 : /// is unavailable (e.g. because it's down or deploying). That means we prefer to use secondary
1374 : /// locations in a non-preferred AZ, as they're more likely to have awarm cache than a temporary
1375 : /// secondary in the preferred AZ (which are usually only created for migrations, and if they exist
1376 : /// they're probably not warmed up yet). The latter behavior is based oni
1377 : ///
1378 : /// If the input is empty, or all the nodes are not elegible for scheduling, return None: the
1379 : /// caller needs to a pick a node some other way.
1380 12832 : pub(crate) fn preferred_secondary(&self, scheduler: &Scheduler) -> Option<NodeId> {
1381 12832 : let candidates = scheduler.filter_usable_nodes(&self.intent.secondary);
1382 :
1383 : // We will sort candidates to prefer nodes which are _not_ in our preferred AZ, i.e. we prefer
1384 : // to migrate to a long-lived secondary location (which would have been scheduled in a non-preferred AZ),
1385 : // rather than a short-lived secondary location being used for optimization/migration (which would have
1386 : // been scheduled in our preferred AZ).
1387 12832 : let mut candidates = candidates
1388 12832 : .iter()
1389 12832 : .map(|(node_id, node_az)| {
1390 2 : if node_az == &self.intent.preferred_az_id {
1391 1 : (1, *node_id)
1392 : } else {
1393 1 : (0, *node_id)
1394 : }
1395 2 : })
1396 12832 : .collect::<Vec<_>>();
1397 :
1398 12832 : candidates.sort();
1399 :
1400 12832 : candidates.first().map(|i| i.1)
1401 12832 : }
1402 :
1403 : /// Query whether the tenant's observed state for attached node matches its intent state, and if so,
1404 : /// yield the node ID. This is appropriate for emitting compute hook notifications: we are checking that
1405 : /// the node in question is not only where we intend to attach, but that the tenant is indeed already attached there.
1406 : ///
1407 : /// Reconciliation may still be needed for other aspects of state such as secondaries (see [`Self::dirty`]): this
1408 : /// funciton should not be used to decide whether to reconcile.
1409 0 : pub(crate) fn stably_attached(&self) -> Option<NodeId> {
1410 : // We have an intent to attach for this node
1411 0 : let attach_intent = self.intent.attached?;
1412 : // We have an observed state for this node
1413 0 : let location = self.observed.locations.get(&attach_intent)?;
1414 : // Our observed state is not None, i.e. not in flux
1415 0 : let location_config = location.conf.as_ref()?;
1416 :
1417 : // Check if our intent and observed state agree that this node is in an attached state.
1418 0 : match location_config.mode {
1419 : LocationConfigMode::AttachedMulti
1420 : | LocationConfigMode::AttachedSingle
1421 0 : | LocationConfigMode::AttachedStale => Some(attach_intent),
1422 0 : _ => None,
1423 : }
1424 0 : }
1425 :
1426 0 : fn dirty(&self, nodes: &Arc<HashMap<NodeId, Node>>) -> bool {
1427 0 : let mut dirty_nodes = HashSet::new();
1428 :
1429 0 : if let Some(node_id) = self.intent.attached {
1430 : // Maybe panic: it is a severe bug if we try to attach while generation is null.
1431 0 : let generation = self
1432 0 : .generation
1433 0 : .expect("Attempted to enter attached state without a generation");
1434 :
1435 0 : let wanted_conf = attached_location_conf(
1436 0 : generation,
1437 0 : &self.shard,
1438 0 : &self.config,
1439 0 : &self.policy,
1440 0 : self.intent.get_secondary().len(),
1441 : );
1442 0 : match self.observed.locations.get(&node_id) {
1443 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
1444 0 : Some(_) | None => {
1445 0 : dirty_nodes.insert(node_id);
1446 0 : }
1447 : }
1448 0 : }
1449 :
1450 0 : for node_id in &self.intent.secondary {
1451 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
1452 0 : match self.observed.locations.get(node_id) {
1453 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
1454 0 : Some(_) | None => {
1455 0 : dirty_nodes.insert(*node_id);
1456 0 : }
1457 : }
1458 : }
1459 :
1460 0 : for node_id in self.observed.locations.keys() {
1461 0 : if self.intent.attached != Some(*node_id) && !self.intent.secondary.contains(node_id) {
1462 0 : // We have observed state that isn't part of our intent: need to clean it up.
1463 0 : dirty_nodes.insert(*node_id);
1464 0 : }
1465 : }
1466 :
1467 0 : dirty_nodes.retain(|node_id| {
1468 0 : nodes
1469 0 : .get(node_id)
1470 0 : .map(|n| n.is_available())
1471 0 : .unwrap_or(false)
1472 0 : });
1473 :
1474 0 : !dirty_nodes.is_empty()
1475 0 : }
1476 :
1477 : #[allow(clippy::too_many_arguments)]
1478 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1479 : pub(crate) fn get_reconcile_needed(
1480 : &mut self,
1481 : pageservers: &Arc<HashMap<NodeId, Node>>,
1482 : ) -> ReconcileNeeded {
1483 : // If there are any ambiguous observed states, and the nodes they refer to are available,
1484 : // we should reconcile to clean them up.
1485 : let mut dirty_observed = false;
1486 : for (node_id, observed_loc) in &self.observed.locations {
1487 : let node = pageservers
1488 : .get(node_id)
1489 : .expect("Nodes may not be removed while referenced");
1490 : if observed_loc.conf.is_none() && node.is_available() {
1491 : dirty_observed = true;
1492 : break;
1493 : }
1494 : }
1495 :
1496 : let active_nodes_dirty = self.dirty(pageservers);
1497 :
1498 : let reconcile_needed = match (
1499 : active_nodes_dirty,
1500 : dirty_observed,
1501 : self.pending_compute_notification,
1502 : ) {
1503 : (true, _, _) => ReconcileNeeded::Yes(ReconcileReason::ActiveNodesDirty),
1504 : (_, true, _) => ReconcileNeeded::Yes(ReconcileReason::UnknownLocation),
1505 : (_, _, true) => ReconcileNeeded::Yes(ReconcileReason::PendingComputeNotification),
1506 : _ => ReconcileNeeded::No,
1507 : };
1508 :
1509 : if matches!(reconcile_needed, ReconcileNeeded::No) {
1510 : tracing::debug!("Not dirty, no reconciliation needed.");
1511 : return ReconcileNeeded::No;
1512 : }
1513 :
1514 : // If we are currently splitting, then never start a reconciler task: the splitting logic
1515 : // requires that shards are not interfered with while it runs. Do this check here rather than
1516 : // up top, so that we only log this message if we would otherwise have done a reconciliation.
1517 : if !matches!(self.splitting, SplitState::Idle) {
1518 : tracing::info!("Refusing to reconcile, splitting in progress");
1519 : return ReconcileNeeded::No;
1520 : }
1521 :
1522 : // Reconcile already in flight for the current sequence?
1523 : if let Some(handle) = &self.reconciler {
1524 : if handle.sequence == self.sequence {
1525 : tracing::info!(
1526 : "Reconciliation already in progress for sequence {:?}",
1527 : self.sequence,
1528 : );
1529 : return ReconcileNeeded::WaitExisting(ReconcilerWaiter {
1530 : tenant_shard_id: self.tenant_shard_id,
1531 : seq_wait: self.waiter.clone(),
1532 : error_seq_wait: self.error_waiter.clone(),
1533 : error: self.last_error.clone(),
1534 : seq: self.sequence,
1535 : });
1536 : }
1537 : }
1538 :
1539 : // Pre-checks done: finally check whether we may actually do the work
1540 : match self.scheduling_policy {
1541 : ShardSchedulingPolicy::Active
1542 : | ShardSchedulingPolicy::Essential
1543 : | ShardSchedulingPolicy::Pause => {}
1544 : ShardSchedulingPolicy::Stop => {
1545 : // We only reach this point if there is work to do and we're going to skip
1546 : // doing it: warn it obvious why this tenant isn't doing what it ought to.
1547 : tracing::warn!("Skipping reconcile for policy {:?}", self.scheduling_policy);
1548 : return ReconcileNeeded::No;
1549 : }
1550 : }
1551 :
1552 : reconcile_needed
1553 : }
1554 :
1555 : /// Ensure the sequence number is set to a value where waiting for this value will make us wait
1556 : /// for the next reconcile: i.e. it is ahead of all completed or running reconcilers.
1557 : ///
1558 : /// Constructing a ReconcilerWaiter with the resulting sequence number gives the property
1559 : /// that the waiter will not complete until some future Reconciler is constructed and run.
1560 0 : fn ensure_sequence_ahead(&mut self) {
1561 : // Find the highest sequence for which a Reconciler has previously run or is currently
1562 : // running
1563 0 : let max_seen = std::cmp::max(
1564 0 : self.reconciler
1565 0 : .as_ref()
1566 0 : .map(|r| r.sequence)
1567 0 : .unwrap_or(Sequence(0)),
1568 0 : std::cmp::max(self.waiter.load(), self.error_waiter.load()),
1569 : );
1570 :
1571 0 : if self.sequence <= max_seen {
1572 0 : self.sequence = max_seen.next();
1573 0 : }
1574 0 : }
1575 :
1576 : /// Create a waiter that will wait for some future Reconciler that hasn't been spawned yet.
1577 : ///
1578 : /// This is appropriate when you can't spawn a reconciler (e.g. due to resource limits), but
1579 : /// you would like to wait on the next reconciler that gets spawned in the background.
1580 0 : pub(crate) fn future_reconcile_waiter(&mut self) -> ReconcilerWaiter {
1581 0 : self.ensure_sequence_ahead();
1582 :
1583 0 : ReconcilerWaiter {
1584 0 : tenant_shard_id: self.tenant_shard_id,
1585 0 : seq_wait: self.waiter.clone(),
1586 0 : error_seq_wait: self.error_waiter.clone(),
1587 0 : error: self.last_error.clone(),
1588 0 : seq: self.sequence,
1589 0 : }
1590 0 : }
1591 :
1592 0 : async fn reconcile(
1593 0 : sequence: Sequence,
1594 0 : mut reconciler: Reconciler,
1595 0 : must_notify: bool,
1596 0 : ) -> ReconcileResult {
1597 : // Attempt to make observed state match intent state
1598 0 : let result = reconciler.reconcile().await;
1599 :
1600 : // If we know we had a pending compute notification from some previous action, send a notification irrespective
1601 : // of whether the above reconcile() did any work. It has to be Ok() though, because otherwise we might be
1602 : // sending a notification of a location that isn't really attached.
1603 0 : if result.is_ok() && must_notify {
1604 : // If this fails we will send the need to retry in [`ReconcileResult::pending_compute_notification`]
1605 0 : reconciler.compute_notify().await.ok();
1606 0 : } else if must_notify {
1607 0 : // Carry this flag so that the reconciler's result will indicate that it still needs to retry
1608 0 : // the compute hook notification eventually.
1609 0 : reconciler.compute_notify_failure = true;
1610 0 : }
1611 :
1612 : // Update result counter
1613 0 : let outcome_label = match &result {
1614 : Ok(_) => {
1615 0 : if reconciler.compute_notify_failure {
1616 0 : ReconcileOutcome::SuccessNoNotify
1617 : } else {
1618 0 : ReconcileOutcome::Success
1619 : }
1620 : }
1621 0 : Err(ReconcileError::Cancel) => ReconcileOutcome::Cancel,
1622 0 : Err(_) => ReconcileOutcome::Error,
1623 : };
1624 :
1625 0 : metrics::METRICS_REGISTRY
1626 0 : .metrics_group
1627 0 : .storage_controller_reconcile_complete
1628 0 : .inc(ReconcileCompleteLabelGroup {
1629 0 : status: outcome_label,
1630 0 : });
1631 :
1632 : // Constructing result implicitly drops Reconciler, freeing any ReconcileUnits before the Service might
1633 : // try and schedule more work in response to our result.
1634 0 : ReconcileResult {
1635 0 : sequence,
1636 0 : result,
1637 0 : tenant_shard_id: reconciler.tenant_shard_id,
1638 0 : generation: reconciler.generation,
1639 0 : observed_deltas: reconciler.observed_deltas(),
1640 0 : pending_compute_notification: reconciler.compute_notify_failure,
1641 0 : }
1642 0 : }
1643 :
1644 : #[allow(clippy::too_many_arguments)]
1645 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1646 : pub(crate) fn spawn_reconciler(
1647 : &mut self,
1648 : reason: ReconcileReason,
1649 : result_tx: &tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
1650 : pageservers: &Arc<HashMap<NodeId, Node>>,
1651 : compute_hook: &Arc<ComputeHook>,
1652 : reconciler_config: ReconcilerConfig,
1653 : service_config: &service::Config,
1654 : persistence: &Arc<Persistence>,
1655 : units: ReconcileUnits,
1656 : gate_guard: GateGuard,
1657 : cancel: &CancellationToken,
1658 : http_client: reqwest::Client,
1659 : ) -> Option<ReconcilerWaiter> {
1660 : // Reconcile in flight for a stale sequence? Our sequence's task will wait for it before
1661 : // doing our sequence's work.
1662 : let old_handle = self.reconciler.take();
1663 :
1664 : // Build list of nodes from which the reconciler should detach
1665 : let mut detach = Vec::new();
1666 : for node_id in self.observed.locations.keys() {
1667 : if self.intent.get_attached() != &Some(*node_id)
1668 : && !self.intent.secondary.contains(node_id)
1669 : {
1670 : detach.push(
1671 : pageservers
1672 : .get(node_id)
1673 : .expect("Intent references non-existent pageserver")
1674 : .clone(),
1675 : )
1676 : }
1677 : }
1678 :
1679 : // Advance the sequence before spawning a reconciler, so that sequence waiters
1680 : // can distinguish between before+after the reconcile completes.
1681 : self.ensure_sequence_ahead();
1682 :
1683 : let reconciler_cancel = cancel.child_token();
1684 : let reconciler_intent = TargetState::from_intent(pageservers, &self.intent);
1685 : let reconciler = Reconciler {
1686 : tenant_shard_id: self.tenant_shard_id,
1687 : shard: self.shard,
1688 : placement_policy: self.policy.clone(),
1689 : generation: self.generation,
1690 : intent: reconciler_intent,
1691 : detach,
1692 : reconciler_config,
1693 : config: self.config.clone(),
1694 : preferred_az: self.intent.preferred_az_id.clone(),
1695 : observed: self.observed.clone(),
1696 : original_observed: self.observed.clone(),
1697 : compute_hook: compute_hook.clone(),
1698 : service_config: service_config.clone(),
1699 : _gate_guard: gate_guard,
1700 : _resource_units: units,
1701 : cancel: reconciler_cancel.clone(),
1702 : persistence: persistence.clone(),
1703 : compute_notify_failure: false,
1704 : http_client,
1705 : };
1706 :
1707 : let reconcile_seq = self.sequence;
1708 : let long_reconcile_threshold = service_config.long_reconcile_threshold;
1709 :
1710 : tracing::info!(seq=%reconcile_seq, "Spawning Reconciler ({reason:?})");
1711 : let must_notify = self.pending_compute_notification;
1712 : let reconciler_span = tracing::info_span!(parent: None, "reconciler", seq=%reconcile_seq,
1713 : tenant_id=%reconciler.tenant_shard_id.tenant_id,
1714 : shard_id=%reconciler.tenant_shard_id.shard_slug());
1715 : metrics::METRICS_REGISTRY
1716 : .metrics_group
1717 : .storage_controller_reconcile_spawn
1718 : .inc();
1719 : let result_tx = result_tx.clone();
1720 : let join_handle = tokio::task::spawn(
1721 0 : async move {
1722 : // Wait for any previous reconcile task to complete before we start
1723 0 : if let Some(old_handle) = old_handle {
1724 0 : old_handle.cancel.cancel();
1725 0 : if let Err(e) = old_handle.handle.await {
1726 : // We can't do much with this other than log it: the task is done, so
1727 : // we may proceed with our work.
1728 0 : tracing::error!("Unexpected join error waiting for reconcile task: {e}");
1729 0 : }
1730 0 : }
1731 :
1732 : // Early check for cancellation before doing any work
1733 : // TODO: wrap all remote API operations in cancellation check
1734 : // as well.
1735 0 : if reconciler.cancel.is_cancelled() {
1736 0 : metrics::METRICS_REGISTRY
1737 0 : .metrics_group
1738 0 : .storage_controller_reconcile_complete
1739 0 : .inc(ReconcileCompleteLabelGroup {
1740 0 : status: ReconcileOutcome::Cancel,
1741 0 : });
1742 0 : return;
1743 0 : }
1744 :
1745 0 : let (tenant_id_label, shard_number_label, sequence_label) = {
1746 0 : (
1747 0 : reconciler.tenant_shard_id.tenant_id.to_string(),
1748 0 : reconciler.tenant_shard_id.shard_number.0.to_string(),
1749 0 : reconcile_seq.to_string(),
1750 0 : )
1751 0 : };
1752 :
1753 0 : let label_group = ReconcileLongRunningLabelGroup {
1754 0 : tenant_id: &tenant_id_label,
1755 0 : shard_number: &shard_number_label,
1756 0 : sequence: &sequence_label,
1757 0 : };
1758 :
1759 0 : let reconcile_fut = Self::reconcile(reconcile_seq, reconciler, must_notify);
1760 0 : let long_reconcile_fut = {
1761 0 : let label_group = label_group.clone();
1762 0 : async move {
1763 0 : tokio::time::sleep(long_reconcile_threshold).await;
1764 :
1765 0 : tracing::warn!("Reconcile passed the long running threshold of {long_reconcile_threshold:?}");
1766 :
1767 0 : metrics::METRICS_REGISTRY
1768 0 : .metrics_group
1769 0 : .storage_controller_reconcile_long_running
1770 0 : .inc(label_group);
1771 0 : }
1772 : };
1773 :
1774 0 : let reconcile_fut = std::pin::pin!(reconcile_fut);
1775 0 : let long_reconcile_fut = std::pin::pin!(long_reconcile_fut);
1776 :
1777 0 : let (was_long, result) =
1778 0 : match future::select(reconcile_fut, long_reconcile_fut).await {
1779 0 : Either::Left((reconcile_result, _)) => (false, reconcile_result),
1780 0 : Either::Right((_, reconcile_fut)) => (true, reconcile_fut.await),
1781 : };
1782 :
1783 0 : if was_long {
1784 0 : let id = metrics::METRICS_REGISTRY
1785 0 : .metrics_group
1786 0 : .storage_controller_reconcile_long_running
1787 0 : .with_labels(label_group);
1788 0 : metrics::METRICS_REGISTRY
1789 0 : .metrics_group
1790 0 : .storage_controller_reconcile_long_running
1791 0 : .remove_metric(id);
1792 0 : }
1793 :
1794 0 : result_tx
1795 0 : .send(ReconcileResultRequest::ReconcileResult(result))
1796 0 : .ok();
1797 0 : }
1798 : .instrument(reconciler_span),
1799 : );
1800 :
1801 : self.reconciler = Some(ReconcilerHandle {
1802 : sequence: self.sequence,
1803 : handle: join_handle,
1804 : cancel: reconciler_cancel,
1805 : });
1806 :
1807 : Some(ReconcilerWaiter {
1808 : tenant_shard_id: self.tenant_shard_id,
1809 : seq_wait: self.waiter.clone(),
1810 : error_seq_wait: self.error_waiter.clone(),
1811 : error: self.last_error.clone(),
1812 : seq: self.sequence,
1813 : })
1814 : }
1815 :
1816 0 : pub(crate) fn cancel_reconciler(&self) {
1817 0 : if let Some(handle) = self.reconciler.as_ref() {
1818 0 : handle.cancel.cancel()
1819 0 : }
1820 0 : }
1821 :
1822 : /// Get a waiter for any reconciliation in flight, but do not start reconciliation
1823 : /// if it is not already running
1824 0 : pub(crate) fn get_waiter(&self) -> Option<ReconcilerWaiter> {
1825 0 : if self.reconciler.is_some() {
1826 0 : Some(ReconcilerWaiter {
1827 0 : tenant_shard_id: self.tenant_shard_id,
1828 0 : seq_wait: self.waiter.clone(),
1829 0 : error_seq_wait: self.error_waiter.clone(),
1830 0 : error: self.last_error.clone(),
1831 0 : seq: self.sequence,
1832 0 : })
1833 : } else {
1834 0 : None
1835 : }
1836 0 : }
1837 :
1838 : /// Called when a ReconcileResult has been emitted and the service is updating
1839 : /// our state: if the result is from a sequence >= my ReconcileHandle, then drop
1840 : /// the handle to indicate there is no longer a reconciliation in progress.
1841 0 : pub(crate) fn reconcile_complete(&mut self, sequence: Sequence) {
1842 0 : if let Some(reconcile_handle) = &self.reconciler {
1843 0 : if reconcile_handle.sequence <= sequence {
1844 0 : self.reconciler = None;
1845 0 : }
1846 0 : }
1847 0 : }
1848 :
1849 : /// If we had any state at all referring to this node ID, drop it. Does not
1850 : /// attempt to reschedule.
1851 : ///
1852 : /// Returns true if we modified the node's intent state.
1853 0 : pub(crate) fn deref_node(&mut self, node_id: NodeId) -> bool {
1854 0 : let mut intent_modified = false;
1855 :
1856 : // Drop if this node was our attached intent
1857 0 : if self.intent.attached == Some(node_id) {
1858 0 : self.intent.attached = None;
1859 0 : intent_modified = true;
1860 0 : }
1861 :
1862 : // Drop from the list of secondaries, and check if we modified it
1863 0 : let had_secondaries = self.intent.secondary.len();
1864 0 : self.intent.secondary.retain(|n| n != &node_id);
1865 0 : intent_modified |= self.intent.secondary.len() != had_secondaries;
1866 :
1867 0 : debug_assert!(!self.intent.all_pageservers().contains(&node_id));
1868 :
1869 0 : if self.preferred_node == Some(node_id) {
1870 0 : self.preferred_node = None;
1871 0 : }
1872 :
1873 0 : intent_modified
1874 0 : }
1875 :
1876 0 : pub(crate) fn set_scheduling_policy(&mut self, p: ShardSchedulingPolicy) {
1877 0 : self.scheduling_policy = p;
1878 0 : }
1879 :
1880 0 : pub(crate) fn get_scheduling_policy(&self) -> ShardSchedulingPolicy {
1881 0 : self.scheduling_policy
1882 0 : }
1883 :
1884 0 : pub(crate) fn set_last_error(&mut self, sequence: Sequence, error: ReconcileError) {
1885 : // Ordering: always set last_error before advancing sequence, so that sequence
1886 : // waiters are guaranteed to see a Some value when they see an error.
1887 0 : *(self.last_error.lock().unwrap()) = Some(Arc::new(error));
1888 0 : self.error_waiter.advance(sequence);
1889 0 : }
1890 :
1891 0 : pub(crate) fn from_persistent(
1892 0 : tsp: TenantShardPersistence,
1893 0 : intent: IntentState,
1894 0 : ) -> anyhow::Result<Self> {
1895 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1896 0 : let shard_identity = tsp.get_shard_identity()?;
1897 :
1898 0 : metrics::METRICS_REGISTRY
1899 0 : .metrics_group
1900 0 : .storage_controller_tenant_shards
1901 0 : .inc();
1902 :
1903 : Ok(Self {
1904 0 : tenant_shard_id,
1905 0 : shard: shard_identity,
1906 0 : sequence: Sequence::initial(),
1907 0 : generation: tsp.generation.map(|g| Generation::new(g as u32)),
1908 0 : policy: serde_json::from_str(&tsp.placement_policy).unwrap(),
1909 0 : intent,
1910 0 : observed: ObservedState::new(),
1911 0 : config: serde_json::from_str(&tsp.config).unwrap(),
1912 0 : reconciler: None,
1913 0 : splitting: tsp.splitting,
1914 : // Filled in during [`Service::startup_reconcile`]
1915 0 : importing: TimelineImportState::Idle,
1916 0 : waiter: Arc::new(SeqWait::new(Sequence::initial())),
1917 0 : error_waiter: Arc::new(SeqWait::new(Sequence::initial())),
1918 0 : last_error: Arc::default(),
1919 : consecutive_reconciles_count: 0,
1920 : pending_compute_notification: false,
1921 : delayed_reconcile: false,
1922 0 : scheduling_policy: serde_json::from_str(&tsp.scheduling_policy).unwrap(),
1923 0 : preferred_node: None,
1924 : })
1925 0 : }
1926 :
1927 0 : pub(crate) fn to_persistent(&self) -> TenantShardPersistence {
1928 : TenantShardPersistence {
1929 0 : tenant_id: self.tenant_shard_id.tenant_id.to_string(),
1930 0 : shard_number: self.tenant_shard_id.shard_number.0 as i32,
1931 0 : shard_count: self.tenant_shard_id.shard_count.literal() as i32,
1932 0 : shard_stripe_size: self.shard.stripe_size.0 as i32,
1933 0 : generation: self.generation.map(|g| g.into().unwrap_or(0) as i32),
1934 0 : generation_pageserver: self.intent.get_attached().map(|n| n.0 as i64),
1935 0 : placement_policy: serde_json::to_string(&self.policy).unwrap(),
1936 0 : config: serde_json::to_string(&self.config).unwrap(),
1937 0 : splitting: SplitState::default(),
1938 0 : scheduling_policy: serde_json::to_string(&self.scheduling_policy).unwrap(),
1939 0 : preferred_az_id: self.intent.preferred_az_id.as_ref().map(|az| az.0.clone()),
1940 : }
1941 0 : }
1942 :
1943 12508 : pub(crate) fn preferred_az(&self) -> Option<&AvailabilityZone> {
1944 12508 : self.intent.get_preferred_az()
1945 12508 : }
1946 :
1947 2 : pub(crate) fn set_preferred_az(
1948 2 : &mut self,
1949 2 : scheduler: &mut Scheduler,
1950 2 : preferred_az_id: Option<AvailabilityZone>,
1951 2 : ) {
1952 2 : self.intent.set_preferred_az(scheduler, preferred_az_id);
1953 2 : }
1954 :
1955 : /// Returns all the nodes to which this tenant shard is attached according to the
1956 : /// observed state and the generations. Return vector is sorted from latest generation
1957 : /// to earliest.
1958 0 : pub(crate) fn attached_locations(&self) -> Vec<(NodeId, Generation)> {
1959 0 : self.observed
1960 0 : .locations
1961 0 : .iter()
1962 0 : .filter_map(|(node_id, observed)| {
1963 : use LocationConfigMode::{AttachedMulti, AttachedSingle, AttachedStale};
1964 :
1965 0 : let conf = observed.conf.as_ref()?;
1966 :
1967 0 : match (conf.generation, conf.mode) {
1968 0 : (Some(gen_), AttachedMulti | AttachedSingle | AttachedStale) => {
1969 0 : Some((*node_id, gen_))
1970 : }
1971 0 : _ => None,
1972 : }
1973 0 : })
1974 0 : .sorted_by(|(_lhs_node_id, lhs_gen), (_rhs_node_id, rhs_gen)| {
1975 0 : lhs_gen.cmp(rhs_gen).reverse()
1976 0 : })
1977 0 : .map(|(node_id, gen_)| (node_id, Generation::new(gen_)))
1978 0 : .collect()
1979 0 : }
1980 :
1981 : /// Update the observed state of the tenant by applying incremental deltas
1982 : ///
1983 : /// Deltas are generated by reconcilers via [`Reconciler::observed_deltas`].
1984 : /// They are then filtered in [`crate::service::Service::process_result`].
1985 0 : pub(crate) fn apply_observed_deltas(
1986 0 : &mut self,
1987 0 : deltas: impl Iterator<Item = ObservedStateDelta>,
1988 0 : ) {
1989 0 : for delta in deltas {
1990 0 : match delta {
1991 0 : ObservedStateDelta::Upsert(ups) => {
1992 0 : let (node_id, loc) = *ups;
1993 :
1994 : // If the generation of the observed location in the delta is lagging
1995 : // behind the current one, then we have a race condition and cannot
1996 : // be certain about the true observed state. Set the observed state
1997 : // to None in order to reflect this.
1998 0 : let crnt_gen = self
1999 0 : .observed
2000 0 : .locations
2001 0 : .get(&node_id)
2002 0 : .and_then(|loc| loc.conf.as_ref())
2003 0 : .and_then(|conf| conf.generation);
2004 0 : let new_gen = loc.conf.as_ref().and_then(|conf| conf.generation);
2005 0 : match (crnt_gen, new_gen) {
2006 0 : (Some(crnt), Some(new)) if crnt_gen > new_gen => {
2007 0 : tracing::warn!(
2008 0 : "Skipping observed state update {}: {:?} and using None due to stale generation ({} > {})",
2009 : node_id,
2010 : loc,
2011 : crnt,
2012 : new
2013 : );
2014 :
2015 0 : self.observed
2016 0 : .locations
2017 0 : .insert(node_id, ObservedStateLocation { conf: None });
2018 :
2019 0 : continue;
2020 : }
2021 0 : _ => {}
2022 : }
2023 :
2024 0 : if let Some(conf) = &loc.conf {
2025 0 : tracing::info!("Updating observed location {}: {:?}", node_id, conf);
2026 : } else {
2027 0 : tracing::info!("Setting observed location {} to None", node_id,)
2028 : }
2029 :
2030 0 : self.observed.locations.insert(node_id, loc);
2031 : }
2032 0 : ObservedStateDelta::Delete(node_id) => {
2033 0 : tracing::info!("Deleting observed location {}", node_id);
2034 0 : self.observed.locations.remove(&node_id);
2035 : }
2036 : }
2037 : }
2038 0 : }
2039 :
2040 : /// Returns true if the tenant shard is attached to a node that is outside the preferred AZ.
2041 : ///
2042 : /// If the shard does not have a preferred AZ, returns false.
2043 0 : pub(crate) fn is_attached_outside_preferred_az(&self, nodes: &HashMap<NodeId, Node>) -> bool {
2044 0 : self.intent
2045 0 : .get_attached()
2046 0 : .map(|node_id| {
2047 0 : Some(
2048 0 : nodes
2049 0 : .get(&node_id)
2050 0 : .expect("referenced node exists")
2051 0 : .get_availability_zone_id(),
2052 0 : ) != self.intent.preferred_az_id.as_ref()
2053 0 : })
2054 0 : .unwrap_or(false)
2055 0 : }
2056 : }
2057 :
2058 : impl Drop for TenantShard {
2059 12843 : fn drop(&mut self) {
2060 12843 : metrics::METRICS_REGISTRY
2061 12843 : .metrics_group
2062 12843 : .storage_controller_tenant_shards
2063 12843 : .dec();
2064 12843 : }
2065 : }
2066 :
2067 : #[cfg(test)]
2068 : pub(crate) mod tests {
2069 : use std::cell::RefCell;
2070 : use std::rc::Rc;
2071 :
2072 : use pageserver_api::controller_api::NodeAvailability;
2073 : use pageserver_api::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardNumber};
2074 : use rand::SeedableRng;
2075 : use rand::rngs::StdRng;
2076 : use utils::id::TenantId;
2077 :
2078 : use super::*;
2079 : use crate::scheduler::test_utils::make_test_nodes;
2080 :
2081 12 : fn make_test_tenant_shard(policy: PlacementPolicy) -> TenantShard {
2082 12 : let tenant_id = TenantId::generate();
2083 12 : let shard_number = ShardNumber(0);
2084 12 : let shard_count = ShardCount::new(1);
2085 12 : let stripe_size = DEFAULT_STRIPE_SIZE;
2086 :
2087 12 : let tenant_shard_id = TenantShardId {
2088 12 : tenant_id,
2089 12 : shard_number,
2090 12 : shard_count,
2091 12 : };
2092 12 : TenantShard::new(
2093 12 : tenant_shard_id,
2094 12 : ShardIdentity::new(shard_number, shard_count, stripe_size).unwrap(),
2095 12 : policy,
2096 12 : None,
2097 : )
2098 12 : }
2099 :
2100 5004 : pub(crate) fn make_test_tenant(
2101 5004 : policy: PlacementPolicy,
2102 5004 : shard_count: ShardCount,
2103 5004 : preferred_az: Option<AvailabilityZone>,
2104 5004 : ) -> Vec<TenantShard> {
2105 5004 : make_test_tenant_with_id(TenantId::generate(), policy, shard_count, preferred_az)
2106 5004 : }
2107 :
2108 5007 : pub(crate) fn make_test_tenant_with_id(
2109 5007 : tenant_id: TenantId,
2110 5007 : policy: PlacementPolicy,
2111 5007 : shard_count: ShardCount,
2112 5007 : preferred_az: Option<AvailabilityZone>,
2113 5007 : ) -> Vec<TenantShard> {
2114 5007 : let stripe_size = DEFAULT_STRIPE_SIZE;
2115 5007 : (0..shard_count.count())
2116 12522 : .map(|i| {
2117 12522 : let shard_number = ShardNumber(i);
2118 :
2119 12522 : let tenant_shard_id = TenantShardId {
2120 12522 : tenant_id,
2121 12522 : shard_number,
2122 12522 : shard_count,
2123 12522 : };
2124 12522 : TenantShard::new(
2125 12522 : tenant_shard_id,
2126 12522 : ShardIdentity::new(shard_number, shard_count, stripe_size).unwrap(),
2127 12522 : policy.clone(),
2128 12522 : preferred_az.clone(),
2129 : )
2130 12522 : })
2131 5007 : .collect()
2132 5007 : }
2133 :
2134 : /// Test the scheduling behaviors used when a tenant configured for HA is subject
2135 : /// to nodes being marked offline.
2136 : #[test]
2137 1 : fn tenant_ha_scheduling() -> anyhow::Result<()> {
2138 : // Start with three nodes. Our tenant will only use two. The third one is
2139 : // expected to remain unused.
2140 1 : let mut nodes = make_test_nodes(3, &[]);
2141 :
2142 1 : let mut scheduler = Scheduler::new(nodes.values());
2143 1 : let mut context = ScheduleContext::default();
2144 :
2145 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2146 1 : tenant_shard
2147 1 : .schedule(&mut scheduler, &mut context)
2148 1 : .expect("we have enough nodes, scheduling should work");
2149 :
2150 : // Expect to initially be schedule on to different nodes
2151 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
2152 1 : assert!(tenant_shard.intent.attached.is_some());
2153 :
2154 1 : let attached_node_id = tenant_shard.intent.attached.unwrap();
2155 1 : let secondary_node_id = *tenant_shard.intent.secondary.iter().last().unwrap();
2156 1 : assert_ne!(attached_node_id, secondary_node_id);
2157 :
2158 : // Notifying the attached node is offline should demote it to a secondary
2159 1 : let changed = tenant_shard
2160 1 : .intent
2161 1 : .demote_attached(&mut scheduler, attached_node_id);
2162 1 : assert!(changed);
2163 1 : assert!(tenant_shard.intent.attached.is_none());
2164 1 : assert_eq!(tenant_shard.intent.secondary.len(), 2);
2165 :
2166 : // Update the scheduler state to indicate the node is offline
2167 1 : nodes
2168 1 : .get_mut(&attached_node_id)
2169 1 : .unwrap()
2170 1 : .set_availability(NodeAvailability::Offline);
2171 1 : scheduler.node_upsert(nodes.get(&attached_node_id).unwrap());
2172 :
2173 : // Scheduling the node should promote the still-available secondary node to attached
2174 1 : tenant_shard
2175 1 : .schedule(&mut scheduler, &mut context)
2176 1 : .expect("active nodes are available");
2177 1 : assert_eq!(tenant_shard.intent.attached.unwrap(), secondary_node_id);
2178 :
2179 : // The original attached node should have been retained as a secondary
2180 1 : assert_eq!(
2181 1 : *tenant_shard.intent.secondary.iter().last().unwrap(),
2182 : attached_node_id
2183 : );
2184 :
2185 1 : tenant_shard.intent.clear(&mut scheduler);
2186 :
2187 1 : Ok(())
2188 1 : }
2189 :
2190 : #[test]
2191 1 : fn intent_from_observed() -> anyhow::Result<()> {
2192 1 : let nodes = make_test_nodes(3, &[]);
2193 1 : let mut scheduler = Scheduler::new(nodes.values());
2194 :
2195 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2196 :
2197 1 : tenant_shard.observed.locations.insert(
2198 1 : NodeId(3),
2199 1 : ObservedStateLocation {
2200 1 : conf: Some(LocationConfig {
2201 1 : mode: LocationConfigMode::AttachedMulti,
2202 1 : generation: Some(2),
2203 1 : secondary_conf: None,
2204 1 : shard_number: tenant_shard.shard.number.0,
2205 1 : shard_count: tenant_shard.shard.count.literal(),
2206 1 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
2207 1 : tenant_conf: TenantConfig::default(),
2208 1 : }),
2209 1 : },
2210 : );
2211 :
2212 1 : tenant_shard.observed.locations.insert(
2213 1 : NodeId(2),
2214 1 : ObservedStateLocation {
2215 1 : conf: Some(LocationConfig {
2216 1 : mode: LocationConfigMode::AttachedStale,
2217 1 : generation: Some(1),
2218 1 : secondary_conf: None,
2219 1 : shard_number: tenant_shard.shard.number.0,
2220 1 : shard_count: tenant_shard.shard.count.literal(),
2221 1 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
2222 1 : tenant_conf: TenantConfig::default(),
2223 1 : }),
2224 1 : },
2225 : );
2226 :
2227 1 : tenant_shard.intent_from_observed(&mut scheduler);
2228 :
2229 : // The highest generationed attached location gets used as attached
2230 1 : assert_eq!(tenant_shard.intent.attached, Some(NodeId(3)));
2231 : // Other locations get used as secondary
2232 1 : assert_eq!(tenant_shard.intent.secondary, vec![NodeId(2)]);
2233 :
2234 1 : scheduler.consistency_check(nodes.values(), [&tenant_shard].into_iter())?;
2235 :
2236 1 : tenant_shard.intent.clear(&mut scheduler);
2237 1 : Ok(())
2238 1 : }
2239 :
2240 : #[test]
2241 1 : fn scheduling_mode() -> anyhow::Result<()> {
2242 1 : let nodes = make_test_nodes(3, &[]);
2243 1 : let mut scheduler = Scheduler::new(nodes.values());
2244 :
2245 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2246 :
2247 : // In pause mode, schedule() shouldn't do anything
2248 1 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Pause;
2249 1 : assert!(
2250 1 : tenant_shard
2251 1 : .schedule(&mut scheduler, &mut ScheduleContext::default())
2252 1 : .is_ok()
2253 : );
2254 1 : assert!(tenant_shard.intent.all_pageservers().is_empty());
2255 :
2256 : // In active mode, schedule() works
2257 1 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Active;
2258 1 : assert!(
2259 1 : tenant_shard
2260 1 : .schedule(&mut scheduler, &mut ScheduleContext::default())
2261 1 : .is_ok()
2262 : );
2263 1 : assert!(!tenant_shard.intent.all_pageservers().is_empty());
2264 :
2265 1 : tenant_shard.intent.clear(&mut scheduler);
2266 1 : Ok(())
2267 1 : }
2268 :
2269 : #[test]
2270 : /// Simple case: moving attachment to somewhere better where we already have a secondary
2271 1 : fn optimize_attachment_simple() -> anyhow::Result<()> {
2272 1 : let nodes = make_test_nodes(
2273 : 3,
2274 1 : &[
2275 1 : AvailabilityZone("az-a".to_string()),
2276 1 : AvailabilityZone("az-b".to_string()),
2277 1 : AvailabilityZone("az-c".to_string()),
2278 1 : ],
2279 : );
2280 1 : let mut scheduler = Scheduler::new(nodes.values());
2281 :
2282 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2283 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2284 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2285 1 : shard_b.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2286 :
2287 : // Initially: both nodes attached on shard 1, and both have secondary locations
2288 : // on different nodes.
2289 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2290 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(1));
2291 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2292 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(2));
2293 :
2294 1 : fn make_schedule_context(shard_a: &TenantShard, shard_b: &TenantShard) -> ScheduleContext {
2295 1 : let mut schedule_context = ScheduleContext::default();
2296 1 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2297 1 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2298 1 : schedule_context
2299 1 : }
2300 :
2301 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2302 1 : let optimization_a = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2303 1 : assert_eq!(
2304 : optimization_a,
2305 1 : Some(ScheduleOptimization {
2306 1 : sequence: shard_a.sequence,
2307 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2308 1 : old_attached_node_id: NodeId(2),
2309 1 : new_attached_node_id: NodeId(1)
2310 1 : })
2311 1 : })
2312 : );
2313 1 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
2314 :
2315 : // // Either shard should recognize that it has the option to switch to a secondary location where there
2316 : // // would be no other shards from the same tenant, and request to do so.
2317 : // assert_eq!(
2318 : // optimization_a_prepare,
2319 : // Some(ScheduleOptimization {
2320 : // sequence: shard_a.sequence,
2321 : // action: ScheduleOptimizationAction::CreateSecondary(NodeId(2))
2322 : // })
2323 : // );
2324 : // shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2325 :
2326 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2327 : // let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2328 : // assert_eq!(
2329 : // optimization_a_migrate,
2330 : // Some(ScheduleOptimization {
2331 : // sequence: shard_a.sequence,
2332 : // action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2333 : // old_attached_node_id: NodeId(1),
2334 : // new_attached_node_id: NodeId(2)
2335 : // })
2336 : // })
2337 : // );
2338 : // shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2339 :
2340 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2341 : // let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2342 : // assert_eq!(
2343 : // optimization_a_cleanup,
2344 : // Some(ScheduleOptimization {
2345 : // sequence: shard_a.sequence,
2346 : // action: ScheduleOptimizationAction::RemoveSecondary(NodeId(1))
2347 : // })
2348 : // );
2349 : // shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2350 :
2351 : // // Shard B should not be moved anywhere, since the pressure on node 1 was relieved by moving shard A
2352 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2353 : // assert_eq!(shard_b.optimize_attachment(&mut scheduler, &schedule_context), None);
2354 :
2355 1 : shard_a.intent.clear(&mut scheduler);
2356 1 : shard_b.intent.clear(&mut scheduler);
2357 :
2358 1 : Ok(())
2359 1 : }
2360 :
2361 : #[test]
2362 : /// Complicated case: moving attachment to somewhere better where we do not have a secondary
2363 : /// already, creating one as needed.
2364 1 : fn optimize_attachment_multistep() -> anyhow::Result<()> {
2365 1 : let nodes = make_test_nodes(
2366 : 3,
2367 1 : &[
2368 1 : AvailabilityZone("az-a".to_string()),
2369 1 : AvailabilityZone("az-b".to_string()),
2370 1 : AvailabilityZone("az-c".to_string()),
2371 1 : ],
2372 : );
2373 1 : let mut scheduler = Scheduler::new(nodes.values());
2374 :
2375 : // Two shards of a tenant that wants to be in AZ A
2376 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2377 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2378 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2379 1 : shard_b.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2380 :
2381 : // Both shards are initially attached in non-home AZ _and_ have secondaries in non-home AZs
2382 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2383 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2384 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(3)));
2385 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(2));
2386 :
2387 3 : fn make_schedule_context(shard_a: &TenantShard, shard_b: &TenantShard) -> ScheduleContext {
2388 3 : let mut schedule_context = ScheduleContext::default();
2389 3 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2390 3 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2391 3 : schedule_context
2392 3 : }
2393 :
2394 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2395 1 : let optimization_a_prepare = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2396 1 : assert_eq!(
2397 : optimization_a_prepare,
2398 1 : Some(ScheduleOptimization {
2399 1 : sequence: shard_a.sequence,
2400 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(1))
2401 1 : })
2402 : );
2403 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2404 :
2405 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2406 1 : let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2407 1 : assert_eq!(
2408 : optimization_a_migrate,
2409 1 : Some(ScheduleOptimization {
2410 1 : sequence: shard_a.sequence,
2411 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2412 1 : old_attached_node_id: NodeId(2),
2413 1 : new_attached_node_id: NodeId(1)
2414 1 : })
2415 1 : })
2416 : );
2417 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2418 :
2419 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2420 1 : let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2421 1 : assert_eq!(
2422 : optimization_a_cleanup,
2423 1 : Some(ScheduleOptimization {
2424 1 : sequence: shard_a.sequence,
2425 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(3))
2426 1 : })
2427 : );
2428 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2429 :
2430 : // // Shard B should not be moved anywhere, since the pressure on node 1 was relieved by moving shard A
2431 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2432 : // assert_eq!(shard_b.optimize_attachment(&mut scheduler, &schedule_context), None);
2433 :
2434 1 : shard_a.intent.clear(&mut scheduler);
2435 1 : shard_b.intent.clear(&mut scheduler);
2436 :
2437 1 : Ok(())
2438 1 : }
2439 :
2440 : #[test]
2441 : /// How the optimisation code handles a shard with a preferred node set; this is an example
2442 : /// of the multi-step migration, but driven by a different input.
2443 1 : fn optimize_attachment_multi_preferred_node() -> anyhow::Result<()> {
2444 1 : let nodes = make_test_nodes(
2445 : 4,
2446 1 : &[
2447 1 : AvailabilityZone("az-a".to_string()),
2448 1 : AvailabilityZone("az-a".to_string()),
2449 1 : AvailabilityZone("az-b".to_string()),
2450 1 : AvailabilityZone("az-b".to_string()),
2451 1 : ],
2452 : );
2453 1 : let mut scheduler = Scheduler::new(nodes.values());
2454 :
2455 : // Two shards of a tenant that wants to be in AZ A
2456 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2457 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2458 :
2459 : // Initially attached in a stable location
2460 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2461 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2462 :
2463 : // Set the preferred node to node 2, an equally high scoring node to its current location
2464 1 : shard_a.preferred_node = Some(NodeId(2));
2465 :
2466 3 : fn make_schedule_context(shard_a: &TenantShard) -> ScheduleContext {
2467 3 : let mut schedule_context = ScheduleContext::default();
2468 3 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2469 3 : schedule_context
2470 3 : }
2471 :
2472 1 : let schedule_context = make_schedule_context(&shard_a);
2473 1 : let optimization_a_prepare = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2474 1 : assert_eq!(
2475 : optimization_a_prepare,
2476 1 : Some(ScheduleOptimization {
2477 1 : sequence: shard_a.sequence,
2478 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(2))
2479 1 : })
2480 : );
2481 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2482 :
2483 : // The first step of the optimisation should not have cleared the preferred node
2484 1 : assert_eq!(shard_a.preferred_node, Some(NodeId(2)));
2485 :
2486 1 : let schedule_context = make_schedule_context(&shard_a);
2487 1 : let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2488 1 : assert_eq!(
2489 : optimization_a_migrate,
2490 1 : Some(ScheduleOptimization {
2491 1 : sequence: shard_a.sequence,
2492 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2493 1 : old_attached_node_id: NodeId(1),
2494 1 : new_attached_node_id: NodeId(2)
2495 1 : })
2496 1 : })
2497 : );
2498 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2499 :
2500 : // The cutover step of the optimisation should have cleared the preferred node
2501 1 : assert_eq!(shard_a.preferred_node, None);
2502 :
2503 1 : let schedule_context = make_schedule_context(&shard_a);
2504 1 : let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2505 1 : assert_eq!(
2506 : optimization_a_cleanup,
2507 1 : Some(ScheduleOptimization {
2508 1 : sequence: shard_a.sequence,
2509 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(1))
2510 1 : })
2511 : );
2512 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2513 :
2514 1 : shard_a.intent.clear(&mut scheduler);
2515 :
2516 1 : Ok(())
2517 1 : }
2518 :
2519 : #[test]
2520 : /// Check that multi-step migration works when moving to somewhere that is only better by
2521 : /// 1 AffinityScore -- this ensures that we don't have a bug like the intermediate secondary
2522 : /// counting toward the affinity score such that it prevents the rest of the migration from happening.
2523 1 : fn optimize_attachment_marginal() -> anyhow::Result<()> {
2524 1 : let nodes = make_test_nodes(2, &[]);
2525 1 : let mut scheduler = Scheduler::new(nodes.values());
2526 :
2527 : // Multi-sharded tenant, we will craft a situation where affinity
2528 : // scores differ only slightly
2529 1 : let mut shards = make_test_tenant(PlacementPolicy::Attached(0), ShardCount::new(4), None);
2530 :
2531 : // 1 attached on node 1
2532 1 : shards[0]
2533 1 : .intent
2534 1 : .set_attached(&mut scheduler, Some(NodeId(1)));
2535 : // 3 attached on node 2
2536 1 : shards[1]
2537 1 : .intent
2538 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2539 1 : shards[2]
2540 1 : .intent
2541 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2542 1 : shards[3]
2543 1 : .intent
2544 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2545 :
2546 : // The scheduler should figure out that we need to:
2547 : // - Create a secondary for shard 3 on node 1
2548 : // - Migrate shard 3 to node 1
2549 : // - Remove shard 3's location on node 2
2550 :
2551 4 : fn make_schedule_context(shards: &Vec<TenantShard>) -> ScheduleContext {
2552 4 : let mut schedule_context = ScheduleContext::default();
2553 20 : for shard in shards {
2554 16 : schedule_context.avoid(&shard.intent.all_pageservers());
2555 16 : }
2556 4 : schedule_context
2557 4 : }
2558 :
2559 1 : let schedule_context = make_schedule_context(&shards);
2560 1 : let optimization_a_prepare =
2561 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2562 1 : assert_eq!(
2563 : optimization_a_prepare,
2564 1 : Some(ScheduleOptimization {
2565 1 : sequence: shards[1].sequence,
2566 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(1))
2567 1 : })
2568 : );
2569 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2570 :
2571 1 : let schedule_context = make_schedule_context(&shards);
2572 1 : let optimization_a_migrate =
2573 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2574 1 : assert_eq!(
2575 : optimization_a_migrate,
2576 1 : Some(ScheduleOptimization {
2577 1 : sequence: shards[1].sequence,
2578 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2579 1 : old_attached_node_id: NodeId(2),
2580 1 : new_attached_node_id: NodeId(1)
2581 1 : })
2582 1 : })
2583 : );
2584 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2585 :
2586 1 : let schedule_context = make_schedule_context(&shards);
2587 1 : let optimization_a_cleanup =
2588 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2589 1 : assert_eq!(
2590 : optimization_a_cleanup,
2591 1 : Some(ScheduleOptimization {
2592 1 : sequence: shards[1].sequence,
2593 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(2))
2594 1 : })
2595 : );
2596 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2597 :
2598 : // Everything should be stable now
2599 1 : let schedule_context = make_schedule_context(&shards);
2600 1 : assert_eq!(
2601 1 : shards[0].optimize_attachment(&mut scheduler, &schedule_context),
2602 : None
2603 : );
2604 1 : assert_eq!(
2605 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context),
2606 : None
2607 : );
2608 1 : assert_eq!(
2609 1 : shards[2].optimize_attachment(&mut scheduler, &schedule_context),
2610 : None
2611 : );
2612 1 : assert_eq!(
2613 1 : shards[3].optimize_attachment(&mut scheduler, &schedule_context),
2614 : None
2615 : );
2616 :
2617 5 : for mut shard in shards {
2618 4 : shard.intent.clear(&mut scheduler);
2619 4 : }
2620 :
2621 1 : Ok(())
2622 1 : }
2623 :
2624 : #[test]
2625 1 : fn optimize_secondary() -> anyhow::Result<()> {
2626 1 : let nodes = make_test_nodes(4, &[]);
2627 1 : let mut scheduler = Scheduler::new(nodes.values());
2628 :
2629 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2630 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2631 :
2632 : // Initially: both nodes attached on shard 1, and both have secondary locations
2633 : // on different nodes.
2634 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2635 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2636 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2637 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(3));
2638 :
2639 1 : let mut schedule_context = ScheduleContext::default();
2640 1 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2641 1 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2642 :
2643 1 : let optimization_a = shard_a.optimize_secondary(&mut scheduler, &schedule_context);
2644 :
2645 : // Since there is a node with no locations available, the node with two locations for the
2646 : // same tenant should generate an optimization to move one away
2647 1 : assert_eq!(
2648 : optimization_a,
2649 1 : Some(ScheduleOptimization {
2650 1 : sequence: shard_a.sequence,
2651 1 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
2652 1 : old_node_id: NodeId(3),
2653 1 : new_node_id: NodeId(4)
2654 1 : })
2655 1 : })
2656 : );
2657 :
2658 1 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
2659 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2660 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(4)]);
2661 :
2662 1 : shard_a.intent.clear(&mut scheduler);
2663 1 : shard_b.intent.clear(&mut scheduler);
2664 :
2665 1 : Ok(())
2666 1 : }
2667 :
2668 : /// Test how the optimisation code behaves with an extra secondary
2669 : #[test]
2670 1 : fn optimize_removes_secondary() -> anyhow::Result<()> {
2671 1 : let az_a_tag = AvailabilityZone("az-a".to_string());
2672 1 : let az_b_tag = AvailabilityZone("az-b".to_string());
2673 1 : let mut nodes = make_test_nodes(
2674 : 4,
2675 1 : &[
2676 1 : az_a_tag.clone(),
2677 1 : az_b_tag.clone(),
2678 1 : az_a_tag.clone(),
2679 1 : az_b_tag.clone(),
2680 1 : ],
2681 : );
2682 1 : let mut scheduler = Scheduler::new(nodes.values());
2683 :
2684 1 : let mut schedule_context = ScheduleContext::default();
2685 :
2686 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2687 1 : shard_a.intent.preferred_az_id = Some(az_a_tag.clone());
2688 1 : shard_a
2689 1 : .schedule(&mut scheduler, &mut schedule_context)
2690 1 : .unwrap();
2691 :
2692 : // Attached on node 1, secondary on node 2
2693 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2694 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(2)]);
2695 :
2696 : // Initially optimiser is idle
2697 1 : assert_eq!(
2698 1 : shard_a.optimize_attachment(&mut scheduler, &schedule_context),
2699 : None
2700 : );
2701 1 : assert_eq!(
2702 1 : shard_a.optimize_secondary(&mut scheduler, &schedule_context),
2703 : None
2704 : );
2705 :
2706 : // 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
2707 : // to our new location
2708 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2709 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2710 1 : assert_eq!(
2711 : optimization,
2712 1 : Some(ScheduleOptimization {
2713 1 : sequence: shard_a.sequence,
2714 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(3))
2715 1 : })
2716 : );
2717 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2718 :
2719 : // A spare secondary in the non-home AZ, and one of them is offline
2720 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(4));
2721 1 : nodes
2722 1 : .get_mut(&NodeId(4))
2723 1 : .unwrap()
2724 1 : .set_availability(NodeAvailability::Offline);
2725 1 : scheduler.node_upsert(nodes.get(&NodeId(4)).unwrap());
2726 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2727 1 : assert_eq!(
2728 : optimization,
2729 1 : Some(ScheduleOptimization {
2730 1 : sequence: shard_a.sequence,
2731 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(4))
2732 1 : })
2733 : );
2734 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2735 :
2736 : // A spare secondary when should have none
2737 1 : shard_a.policy = PlacementPolicy::Attached(0);
2738 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2739 1 : assert_eq!(
2740 : optimization,
2741 1 : Some(ScheduleOptimization {
2742 1 : sequence: shard_a.sequence,
2743 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(2))
2744 1 : })
2745 : );
2746 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2747 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2748 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![]);
2749 :
2750 : // Check that in secondary mode, we preserve the secondary in the preferred AZ
2751 1 : let mut schedule_context = ScheduleContext::default(); // Fresh context, we're about to call schedule()
2752 1 : shard_a.policy = PlacementPolicy::Secondary;
2753 1 : shard_a
2754 1 : .schedule(&mut scheduler, &mut schedule_context)
2755 1 : .unwrap();
2756 1 : assert_eq!(shard_a.intent.get_attached(), &None);
2757 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(1)]);
2758 1 : assert_eq!(
2759 1 : shard_a.optimize_attachment(&mut scheduler, &schedule_context),
2760 : None
2761 : );
2762 1 : assert_eq!(
2763 1 : shard_a.optimize_secondary(&mut scheduler, &schedule_context),
2764 : None
2765 : );
2766 :
2767 1 : shard_a.intent.clear(&mut scheduler);
2768 :
2769 1 : Ok(())
2770 1 : }
2771 :
2772 : // Optimize til quiescent: this emulates what Service::optimize_all does, when
2773 : // called repeatedly in the background.
2774 : // Returns the applied optimizations
2775 3 : fn optimize_til_idle(
2776 3 : scheduler: &mut Scheduler,
2777 3 : shards: &mut [TenantShard],
2778 3 : ) -> Vec<ScheduleOptimization> {
2779 3 : let mut loop_n = 0;
2780 3 : let mut optimizations = Vec::default();
2781 : loop {
2782 6 : let mut schedule_context = ScheduleContext::default();
2783 6 : let mut any_changed = false;
2784 :
2785 24 : for shard in shards.iter() {
2786 24 : schedule_context.avoid(&shard.intent.all_pageservers());
2787 24 : }
2788 :
2789 15 : for shard in shards.iter_mut() {
2790 15 : let optimization = shard.optimize_attachment(scheduler, &schedule_context);
2791 15 : tracing::info!(
2792 0 : "optimize_attachment({})={:?}",
2793 : shard.tenant_shard_id,
2794 : optimization
2795 : );
2796 15 : if let Some(optimization) = optimization {
2797 : // Check that maybe_optimizable wouldn't have wrongly claimed this optimization didn't exist
2798 3 : assert!(shard.maybe_optimizable(scheduler, &schedule_context));
2799 3 : optimizations.push(optimization.clone());
2800 3 : shard.apply_optimization(scheduler, optimization);
2801 3 : any_changed = true;
2802 3 : break;
2803 12 : }
2804 :
2805 12 : let optimization = shard.optimize_secondary(scheduler, &schedule_context);
2806 12 : tracing::info!(
2807 0 : "optimize_secondary({})={:?}",
2808 : shard.tenant_shard_id,
2809 : optimization
2810 : );
2811 12 : if let Some(optimization) = optimization {
2812 : // Check that maybe_optimizable wouldn't have wrongly claimed this optimization didn't exist
2813 0 : assert!(shard.maybe_optimizable(scheduler, &schedule_context));
2814 :
2815 0 : optimizations.push(optimization.clone());
2816 0 : shard.apply_optimization(scheduler, optimization);
2817 0 : any_changed = true;
2818 0 : break;
2819 12 : }
2820 : }
2821 :
2822 6 : if !any_changed {
2823 3 : break;
2824 3 : }
2825 :
2826 : // Assert no infinite loop
2827 3 : loop_n += 1;
2828 3 : assert!(loop_n < 1000);
2829 : }
2830 :
2831 3 : optimizations
2832 3 : }
2833 :
2834 : /// Test the balancing behavior of shard scheduling: that it achieves a balance, and
2835 : /// that it converges.
2836 : #[test]
2837 1 : fn optimize_add_nodes() -> anyhow::Result<()> {
2838 1 : let nodes = make_test_nodes(
2839 : 9,
2840 1 : &[
2841 1 : // Initial 6 nodes
2842 1 : AvailabilityZone("az-a".to_string()),
2843 1 : AvailabilityZone("az-a".to_string()),
2844 1 : AvailabilityZone("az-b".to_string()),
2845 1 : AvailabilityZone("az-b".to_string()),
2846 1 : AvailabilityZone("az-c".to_string()),
2847 1 : AvailabilityZone("az-c".to_string()),
2848 1 : // Three we will add later
2849 1 : AvailabilityZone("az-a".to_string()),
2850 1 : AvailabilityZone("az-b".to_string()),
2851 1 : AvailabilityZone("az-c".to_string()),
2852 1 : ],
2853 : );
2854 :
2855 : // Only show the scheduler two nodes in each AZ to start with
2856 1 : let mut scheduler = Scheduler::new([].iter());
2857 7 : for i in 1..=6 {
2858 6 : scheduler.node_upsert(nodes.get(&NodeId(i)).unwrap());
2859 6 : }
2860 :
2861 1 : let mut shards = make_test_tenant(
2862 1 : PlacementPolicy::Attached(1),
2863 1 : ShardCount::new(4),
2864 1 : Some(AvailabilityZone("az-a".to_string())),
2865 : );
2866 1 : let mut schedule_context = ScheduleContext::default();
2867 5 : for shard in &mut shards {
2868 4 : assert!(
2869 4 : shard
2870 4 : .schedule(&mut scheduler, &mut schedule_context)
2871 4 : .is_ok()
2872 : );
2873 : }
2874 :
2875 : // Initial: attached locations land in the tenant's home AZ.
2876 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 2);
2877 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 2);
2878 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
2879 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
2880 :
2881 : // Initial: secondary locations in a remote AZ
2882 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 1);
2883 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 0);
2884 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 1);
2885 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 0);
2886 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(5)), 1);
2887 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(5)), 0);
2888 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(6)), 1);
2889 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(6)), 0);
2890 :
2891 : // Add another three nodes: we should see the shards spread out when their optimize
2892 : // methods are called
2893 1 : scheduler.node_upsert(nodes.get(&NodeId(7)).unwrap());
2894 1 : scheduler.node_upsert(nodes.get(&NodeId(8)).unwrap());
2895 1 : scheduler.node_upsert(nodes.get(&NodeId(9)).unwrap());
2896 1 : optimize_til_idle(&mut scheduler, &mut shards);
2897 :
2898 : // We expect one attached location was moved to the new node in the tenant's home AZ
2899 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(7)), 1);
2900 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(7)), 1);
2901 : // The original node has one less attached shard
2902 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 1);
2903 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 1);
2904 :
2905 : // One of the original nodes still has two attachments, since there are an odd number of nodes
2906 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
2907 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
2908 :
2909 : // None of our secondaries moved, since we already had enough nodes for those to be
2910 : // scheduled perfectly
2911 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 1);
2912 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 0);
2913 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 1);
2914 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 0);
2915 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(5)), 1);
2916 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(5)), 0);
2917 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(6)), 1);
2918 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(6)), 0);
2919 :
2920 4 : for shard in shards.iter_mut() {
2921 4 : shard.intent.clear(&mut scheduler);
2922 4 : }
2923 :
2924 1 : Ok(())
2925 1 : }
2926 :
2927 : /// Test that initial shard scheduling is optimal. By optimal we mean
2928 : /// that the optimizer cannot find a way to improve it.
2929 : ///
2930 : /// This test is an example of the scheduling issue described in
2931 : /// https://github.com/neondatabase/neon/issues/8969
2932 : #[test]
2933 1 : fn initial_scheduling_is_optimal() -> anyhow::Result<()> {
2934 : use itertools::Itertools;
2935 :
2936 1 : let nodes = make_test_nodes(2, &[]);
2937 :
2938 1 : let mut scheduler = Scheduler::new([].iter());
2939 1 : scheduler.node_upsert(nodes.get(&NodeId(1)).unwrap());
2940 1 : scheduler.node_upsert(nodes.get(&NodeId(2)).unwrap());
2941 :
2942 1 : let mut a = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4), None);
2943 1 : let a_context = Rc::new(RefCell::new(ScheduleContext::default()));
2944 :
2945 1 : let mut b = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4), None);
2946 1 : let b_context = Rc::new(RefCell::new(ScheduleContext::default()));
2947 :
2948 4 : let a_shards_with_context = a.iter_mut().map(|shard| (shard, a_context.clone()));
2949 4 : let b_shards_with_context = b.iter_mut().map(|shard| (shard, b_context.clone()));
2950 :
2951 1 : let schedule_order = a_shards_with_context.interleave(b_shards_with_context);
2952 :
2953 9 : for (shard, context) in schedule_order {
2954 8 : let context = &mut *context.borrow_mut();
2955 8 : shard.schedule(&mut scheduler, context).unwrap();
2956 8 : }
2957 :
2958 1 : let applied_to_a = optimize_til_idle(&mut scheduler, &mut a);
2959 1 : assert_eq!(applied_to_a, vec![]);
2960 :
2961 1 : let applied_to_b = optimize_til_idle(&mut scheduler, &mut b);
2962 1 : assert_eq!(applied_to_b, vec![]);
2963 :
2964 8 : for shard in a.iter_mut().chain(b.iter_mut()) {
2965 8 : shard.intent.clear(&mut scheduler);
2966 8 : }
2967 :
2968 1 : Ok(())
2969 1 : }
2970 :
2971 : #[test]
2972 1 : fn random_az_shard_scheduling() -> anyhow::Result<()> {
2973 : use rand::seq::SliceRandom;
2974 :
2975 51 : for seed in 0..50 {
2976 50 : eprintln!("Running test with seed {seed}");
2977 50 : let mut rng = StdRng::seed_from_u64(seed);
2978 :
2979 50 : let az_a_tag = AvailabilityZone("az-a".to_string());
2980 50 : let az_b_tag = AvailabilityZone("az-b".to_string());
2981 50 : let azs = [az_a_tag, az_b_tag];
2982 50 : let nodes = make_test_nodes(4, &azs);
2983 50 : let mut shards_per_az: HashMap<AvailabilityZone, u32> = HashMap::new();
2984 :
2985 50 : let mut scheduler = Scheduler::new([].iter());
2986 200 : for node in nodes.values() {
2987 200 : scheduler.node_upsert(node);
2988 200 : }
2989 :
2990 50 : let mut shards = Vec::default();
2991 50 : let mut contexts = Vec::default();
2992 50 : let mut az_picker = azs.iter().cycle().cloned();
2993 5050 : for i in 0..100 {
2994 5000 : let az = az_picker.next().unwrap();
2995 5000 : let shard_count = i % 4 + 1;
2996 5000 : *shards_per_az.entry(az.clone()).or_default() += shard_count;
2997 :
2998 5000 : let tenant_shards = make_test_tenant(
2999 5000 : PlacementPolicy::Attached(1),
3000 5000 : ShardCount::new(shard_count.try_into().unwrap()),
3001 5000 : Some(az),
3002 : );
3003 5000 : let context = Rc::new(RefCell::new(ScheduleContext::default()));
3004 :
3005 5000 : contexts.push(context.clone());
3006 5000 : let with_ctx = tenant_shards
3007 5000 : .into_iter()
3008 12500 : .map(|shard| (shard, context.clone()));
3009 17500 : for shard_with_ctx in with_ctx {
3010 12500 : shards.push(shard_with_ctx);
3011 12500 : }
3012 : }
3013 :
3014 50 : shards.shuffle(&mut rng);
3015 :
3016 : #[derive(Default, Debug)]
3017 : struct NodeStats {
3018 : attachments: u32,
3019 : secondaries: u32,
3020 : }
3021 :
3022 50 : let mut node_stats: HashMap<NodeId, NodeStats> = HashMap::default();
3023 50 : let mut attachments_in_wrong_az = 0;
3024 50 : let mut secondaries_in_wrong_az = 0;
3025 :
3026 12550 : for (shard, context) in &mut shards {
3027 12500 : let context = &mut *context.borrow_mut();
3028 12500 : shard.schedule(&mut scheduler, context).unwrap();
3029 :
3030 12500 : let attached_node = shard.intent.get_attached().unwrap();
3031 12500 : let stats = node_stats.entry(attached_node).or_default();
3032 12500 : stats.attachments += 1;
3033 :
3034 12500 : let secondary_node = *shard.intent.get_secondary().first().unwrap();
3035 12500 : let stats = node_stats.entry(secondary_node).or_default();
3036 12500 : stats.secondaries += 1;
3037 :
3038 12500 : let attached_node_az = nodes
3039 12500 : .get(&attached_node)
3040 12500 : .unwrap()
3041 12500 : .get_availability_zone_id();
3042 12500 : let secondary_node_az = nodes
3043 12500 : .get(&secondary_node)
3044 12500 : .unwrap()
3045 12500 : .get_availability_zone_id();
3046 12500 : let preferred_az = shard.preferred_az().unwrap();
3047 :
3048 12500 : if attached_node_az != preferred_az {
3049 0 : eprintln!(
3050 0 : "{} attachment was scheduled in AZ {} but preferred AZ {}",
3051 0 : shard.tenant_shard_id, attached_node_az, preferred_az
3052 0 : );
3053 0 : attachments_in_wrong_az += 1;
3054 12500 : }
3055 :
3056 12500 : if secondary_node_az == preferred_az {
3057 0 : eprintln!(
3058 0 : "{} secondary was scheduled in AZ {} which matches preference",
3059 0 : shard.tenant_shard_id, attached_node_az
3060 0 : );
3061 0 : secondaries_in_wrong_az += 1;
3062 12500 : }
3063 : }
3064 :
3065 50 : let mut violations = Vec::default();
3066 :
3067 50 : if attachments_in_wrong_az > 0 {
3068 0 : violations.push(format!(
3069 0 : "{attachments_in_wrong_az} attachments scheduled to the incorrect AZ"
3070 0 : ));
3071 50 : }
3072 :
3073 50 : if secondaries_in_wrong_az > 0 {
3074 0 : violations.push(format!(
3075 0 : "{secondaries_in_wrong_az} secondaries scheduled to the incorrect AZ"
3076 0 : ));
3077 50 : }
3078 :
3079 50 : eprintln!(
3080 50 : "attachments_in_wrong_az={attachments_in_wrong_az} secondaries_in_wrong_az={secondaries_in_wrong_az}"
3081 : );
3082 :
3083 250 : for (node_id, stats) in &node_stats {
3084 200 : let node_az = nodes.get(node_id).unwrap().get_availability_zone_id();
3085 200 : let ideal_attachment_load = shards_per_az.get(node_az).unwrap() / 2;
3086 200 : let allowed_attachment_load =
3087 200 : (ideal_attachment_load - 1)..(ideal_attachment_load + 2);
3088 :
3089 200 : if !allowed_attachment_load.contains(&stats.attachments) {
3090 0 : violations.push(format!(
3091 0 : "Found {} attachments on node {}, but expected {}",
3092 0 : stats.attachments, node_id, ideal_attachment_load
3093 0 : ));
3094 200 : }
3095 :
3096 200 : eprintln!(
3097 200 : "{}: attachments={} secondaries={} ideal_attachment_load={}",
3098 : node_id, stats.attachments, stats.secondaries, ideal_attachment_load
3099 : );
3100 : }
3101 :
3102 50 : assert!(violations.is_empty(), "{violations:?}");
3103 :
3104 12550 : for (mut shard, _ctx) in shards {
3105 12500 : shard.intent.clear(&mut scheduler);
3106 12500 : }
3107 : }
3108 1 : Ok(())
3109 1 : }
3110 :
3111 : /// Check how the shard's scheduling behaves when in PlacementPolicy::Secondary mode.
3112 : #[test]
3113 1 : fn tenant_secondary_scheduling() -> anyhow::Result<()> {
3114 1 : let az_a = AvailabilityZone("az-a".to_string());
3115 1 : let nodes = make_test_nodes(
3116 : 3,
3117 1 : &[
3118 1 : az_a.clone(),
3119 1 : AvailabilityZone("az-b".to_string()),
3120 1 : AvailabilityZone("az-c".to_string()),
3121 1 : ],
3122 : );
3123 :
3124 1 : let mut scheduler = Scheduler::new(nodes.values());
3125 1 : let mut context = ScheduleContext::default();
3126 :
3127 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Secondary);
3128 1 : tenant_shard.intent.preferred_az_id = Some(az_a.clone());
3129 1 : tenant_shard
3130 1 : .schedule(&mut scheduler, &mut context)
3131 1 : .expect("we have enough nodes, scheduling should work");
3132 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
3133 1 : assert!(tenant_shard.intent.attached.is_none());
3134 :
3135 : // Should have scheduled into the preferred AZ
3136 1 : assert_eq!(
3137 1 : scheduler
3138 1 : .get_node_az(&tenant_shard.intent.secondary[0])
3139 1 : .as_ref(),
3140 1 : tenant_shard.preferred_az()
3141 : );
3142 :
3143 : // Optimizer should agree
3144 1 : assert_eq!(
3145 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
3146 : None
3147 : );
3148 1 : assert_eq!(
3149 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
3150 : None
3151 : );
3152 :
3153 : // Switch to PlacementPolicy::Attached
3154 1 : tenant_shard.policy = PlacementPolicy::Attached(1);
3155 1 : tenant_shard
3156 1 : .schedule(&mut scheduler, &mut context)
3157 1 : .expect("we have enough nodes, scheduling should work");
3158 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
3159 1 : assert!(tenant_shard.intent.attached.is_some());
3160 : // Secondary should now be in non-preferred AZ
3161 1 : assert_ne!(
3162 1 : scheduler
3163 1 : .get_node_az(&tenant_shard.intent.secondary[0])
3164 1 : .as_ref(),
3165 1 : tenant_shard.preferred_az()
3166 : );
3167 : // Attached should be in preferred AZ
3168 1 : assert_eq!(
3169 1 : scheduler
3170 1 : .get_node_az(&tenant_shard.intent.attached.unwrap())
3171 1 : .as_ref(),
3172 1 : tenant_shard.preferred_az()
3173 : );
3174 :
3175 : // Optimizer should agree
3176 1 : assert_eq!(
3177 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
3178 : None
3179 : );
3180 1 : assert_eq!(
3181 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
3182 : None
3183 : );
3184 :
3185 : // Switch back to PlacementPolicy::Secondary
3186 1 : tenant_shard.policy = PlacementPolicy::Secondary;
3187 1 : tenant_shard
3188 1 : .schedule(&mut scheduler, &mut context)
3189 1 : .expect("we have enough nodes, scheduling should work");
3190 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
3191 1 : assert!(tenant_shard.intent.attached.is_none());
3192 : // When we picked a location to keep, we should have kept the one in the preferred AZ
3193 1 : assert_eq!(
3194 1 : scheduler
3195 1 : .get_node_az(&tenant_shard.intent.secondary[0])
3196 1 : .as_ref(),
3197 1 : tenant_shard.preferred_az()
3198 : );
3199 :
3200 : // Optimizer should agree
3201 1 : assert_eq!(
3202 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
3203 : None
3204 : );
3205 1 : assert_eq!(
3206 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
3207 : None
3208 : );
3209 :
3210 1 : tenant_shard.intent.clear(&mut scheduler);
3211 :
3212 1 : Ok(())
3213 1 : }
3214 : }
|
