Line data Source code
1 : use std::{
2 : collections::{HashMap, HashSet},
3 : sync::Arc,
4 : time::Duration,
5 : };
6 :
7 : use crate::{
8 : metrics::{
9 : self, ReconcileCompleteLabelGroup, ReconcileLongRunningLabelGroup, ReconcileOutcome,
10 : },
11 : persistence::TenantShardPersistence,
12 : reconciler::{ReconcileUnits, ReconcilerConfig},
13 : scheduler::{
14 : AffinityScore, AttachedShardTag, NodeSchedulingScore, NodeSecondarySchedulingScore,
15 : RefCountUpdate, ScheduleContext, SecondaryShardTag, ShardTag,
16 : },
17 : service::ReconcileResultRequest,
18 : };
19 : use futures::future::{self, Either};
20 : use itertools::Itertools;
21 : use pageserver_api::controller_api::{AvailabilityZone, PlacementPolicy, ShardSchedulingPolicy};
22 : use pageserver_api::{
23 : models::{LocationConfig, LocationConfigMode, TenantConfig},
24 : shard::{ShardIdentity, TenantShardId},
25 : };
26 : use serde::{Deserialize, Serialize};
27 : use tokio::task::JoinHandle;
28 : use tokio_util::sync::CancellationToken;
29 : use tracing::{instrument, Instrument};
30 : use utils::{
31 : generation::Generation,
32 : id::NodeId,
33 : seqwait::{SeqWait, SeqWaitError},
34 : shard::ShardCount,
35 : sync::gate::GateGuard,
36 : };
37 :
38 : use crate::{
39 : compute_hook::ComputeHook,
40 : node::Node,
41 : persistence::{split_state::SplitState, Persistence},
42 : reconciler::{
43 : attached_location_conf, secondary_location_conf, ReconcileError, Reconciler, TargetState,
44 : },
45 : scheduler::{ScheduleError, Scheduler},
46 : service, Sequence,
47 : };
48 :
49 : /// Serialization helper
50 0 : fn read_last_error<S, T>(v: &std::sync::Mutex<Option<T>>, serializer: S) -> Result<S::Ok, S::Error>
51 0 : where
52 0 : S: serde::ser::Serializer,
53 0 : T: std::fmt::Display,
54 0 : {
55 0 : serializer.collect_str(
56 0 : &v.lock()
57 0 : .unwrap()
58 0 : .as_ref()
59 0 : .map(|e| format!("{e}"))
60 0 : .unwrap_or("".to_string()),
61 0 : )
62 0 : }
63 :
64 : /// In-memory state for a particular tenant shard.
65 : ///
66 : /// This struct implement Serialize for debugging purposes, but is _not_ persisted
67 : /// itself: see [`crate::persistence`] for the subset of tenant shard state that is persisted.
68 0 : #[derive(Serialize)]
69 : pub(crate) struct TenantShard {
70 : pub(crate) tenant_shard_id: TenantShardId,
71 :
72 : pub(crate) shard: ShardIdentity,
73 :
74 : // Runtime only: sequence used to coordinate when updating this object while
75 : // with background reconcilers may be running. A reconciler runs to a particular
76 : // sequence.
77 : pub(crate) sequence: Sequence,
78 :
79 : // Latest generation number: next time we attach, increment this
80 : // and use the incremented number when attaching.
81 : //
82 : // None represents an incompletely onboarded tenant via the [`Service::location_config`]
83 : // API, where this tenant may only run in PlacementPolicy::Secondary.
84 : pub(crate) generation: Option<Generation>,
85 :
86 : // High level description of how the tenant should be set up. Provided
87 : // externally.
88 : pub(crate) policy: PlacementPolicy,
89 :
90 : // Low level description of exactly which pageservers should fulfil
91 : // which role. Generated by `Self::schedule`.
92 : pub(crate) intent: IntentState,
93 :
94 : // Low level description of how the tenant is configured on pageservers:
95 : // if this does not match `Self::intent` then the tenant needs reconciliation
96 : // with `Self::reconcile`.
97 : pub(crate) observed: ObservedState,
98 :
99 : // Tenant configuration, passed through opaquely to the pageserver. Identical
100 : // for all shards in a tenant.
101 : pub(crate) config: TenantConfig,
102 :
103 : /// If a reconcile task is currently in flight, it may be joined here (it is
104 : /// only safe to join if either the result has been received or the reconciler's
105 : /// cancellation token has been fired)
106 : #[serde(skip)]
107 : pub(crate) reconciler: Option<ReconcilerHandle>,
108 :
109 : /// If a tenant is being split, then all shards with that TenantId will have a
110 : /// SplitState set, this acts as a guard against other operations such as background
111 : /// reconciliation, and timeline creation.
112 : pub(crate) splitting: SplitState,
113 :
114 : /// If a tenant was enqueued for later reconcile due to hitting concurrency limit, this flag
115 : /// is set. This flag is cleared when the tenant is popped off the delay queue.
116 : pub(crate) delayed_reconcile: bool,
117 :
118 : /// Optionally wait for reconciliation to complete up to a particular
119 : /// sequence number.
120 : #[serde(skip)]
121 : pub(crate) waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
122 :
123 : /// Indicates sequence number for which we have encountered an error reconciling. If
124 : /// this advances ahead of [`Self::waiter`] then a reconciliation error has occurred,
125 : /// and callers should stop waiting for `waiter` and propagate the error.
126 : #[serde(skip)]
127 : pub(crate) error_waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
128 :
129 : /// The most recent error from a reconcile on this tenant. This is a nested Arc
130 : /// because:
131 : /// - ReconcileWaiters need to Arc-clone the overall object to read it later
132 : /// - ReconcileWaitError needs to use an `Arc<ReconcileError>` because we can construct
133 : /// many waiters for one shard, and the underlying error types are not Clone.
134 : ///
135 : /// TODO: generalize to an array of recent events
136 : /// TOOD: use a ArcSwap instead of mutex for faster reads?
137 : #[serde(serialize_with = "read_last_error")]
138 : pub(crate) last_error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
139 :
140 : /// If we have a pending compute notification that for some reason we weren't able to send,
141 : /// set this to true. If this is set, calls to [`Self::get_reconcile_needed`] will return Yes
142 : /// and trigger a Reconciler run. This is the mechanism by which compute notifications are included in the scope
143 : /// of state that we publish externally in an eventually consistent way.
144 : pub(crate) pending_compute_notification: bool,
145 :
146 : // Support/debug tool: if something is going wrong or flapping with scheduling, this may
147 : // be set to a non-active state to avoid making changes while the issue is fixed.
148 : scheduling_policy: ShardSchedulingPolicy,
149 : }
150 :
151 : #[derive(Clone, Debug, Serialize)]
152 : pub(crate) struct IntentState {
153 : attached: Option<NodeId>,
154 : secondary: Vec<NodeId>,
155 :
156 : // We should attempt to schedule this shard in the provided AZ to
157 : // decrease chances of cross-AZ compute.
158 : preferred_az_id: Option<AvailabilityZone>,
159 : }
160 :
161 : impl IntentState {
162 12858 : pub(crate) fn new(preferred_az_id: Option<AvailabilityZone>) -> Self {
163 12858 : Self {
164 12858 : attached: None,
165 12858 : secondary: vec![],
166 12858 : preferred_az_id,
167 12858 : }
168 12858 : }
169 0 : pub(crate) fn single(
170 0 : scheduler: &mut Scheduler,
171 0 : node_id: Option<NodeId>,
172 0 : preferred_az_id: Option<AvailabilityZone>,
173 0 : ) -> Self {
174 0 : if let Some(node_id) = node_id {
175 0 : scheduler.update_node_ref_counts(
176 0 : node_id,
177 0 : preferred_az_id.as_ref(),
178 0 : RefCountUpdate::Attach,
179 0 : );
180 0 : }
181 0 : Self {
182 0 : attached: node_id,
183 0 : secondary: vec![],
184 0 : preferred_az_id,
185 0 : }
186 0 : }
187 :
188 12857 : pub(crate) fn set_attached(&mut self, scheduler: &mut Scheduler, new_attached: Option<NodeId>) {
189 12857 : if self.attached != new_attached {
190 12857 : if let Some(old_attached) = self.attached.take() {
191 0 : scheduler.update_node_ref_counts(
192 0 : old_attached,
193 0 : self.preferred_az_id.as_ref(),
194 0 : RefCountUpdate::Detach,
195 0 : );
196 12857 : }
197 12857 : if let Some(new_attached) = &new_attached {
198 12857 : scheduler.update_node_ref_counts(
199 12857 : *new_attached,
200 12857 : self.preferred_az_id.as_ref(),
201 12857 : RefCountUpdate::Attach,
202 12857 : );
203 12857 : }
204 12857 : self.attached = new_attached;
205 0 : }
206 :
207 12857 : if let Some(new_attached) = &new_attached {
208 12857 : assert!(!self.secondary.contains(new_attached));
209 0 : }
210 12857 : }
211 :
212 : /// Like set_attached, but the node is from [`Self::secondary`]. This swaps the node from
213 : /// secondary to attached while maintaining the scheduler's reference counts.
214 6 : pub(crate) fn promote_attached(
215 6 : &mut self,
216 6 : scheduler: &mut Scheduler,
217 6 : promote_secondary: NodeId,
218 6 : ) {
219 6 : // If we call this with a node that isn't in secondary, it would cause incorrect
220 6 : // scheduler reference counting, since we assume the node is already referenced as a secondary.
221 6 : debug_assert!(self.secondary.contains(&promote_secondary));
222 :
223 13 : self.secondary.retain(|n| n != &promote_secondary);
224 6 :
225 6 : let demoted = self.attached;
226 6 : self.attached = Some(promote_secondary);
227 6 :
228 6 : scheduler.update_node_ref_counts(
229 6 : promote_secondary,
230 6 : self.preferred_az_id.as_ref(),
231 6 : RefCountUpdate::PromoteSecondary,
232 6 : );
233 6 : if let Some(demoted) = demoted {
234 0 : scheduler.update_node_ref_counts(
235 0 : demoted,
236 0 : self.preferred_az_id.as_ref(),
237 0 : RefCountUpdate::DemoteAttached,
238 0 : );
239 6 : }
240 6 : }
241 :
242 12845 : pub(crate) fn push_secondary(&mut self, scheduler: &mut Scheduler, new_secondary: NodeId) {
243 12845 : assert!(!self.secondary.contains(&new_secondary));
244 12845 : assert!(self.attached != Some(new_secondary));
245 12845 : scheduler.update_node_ref_counts(
246 12845 : new_secondary,
247 12845 : self.preferred_az_id.as_ref(),
248 12845 : RefCountUpdate::AddSecondary,
249 12845 : );
250 12845 : self.secondary.push(new_secondary);
251 12845 : }
252 :
253 : /// It is legal to call this with a node that is not currently a secondary: that is a no-op
254 8 : pub(crate) fn remove_secondary(&mut self, scheduler: &mut Scheduler, node_id: NodeId) {
255 11 : let index = self.secondary.iter().position(|n| *n == node_id);
256 8 : if let Some(index) = index {
257 8 : scheduler.update_node_ref_counts(
258 8 : node_id,
259 8 : self.preferred_az_id.as_ref(),
260 8 : RefCountUpdate::RemoveSecondary,
261 8 : );
262 8 : self.secondary.remove(index);
263 8 : }
264 8 : }
265 :
266 12857 : pub(crate) fn clear_secondary(&mut self, scheduler: &mut Scheduler) {
267 12857 : for secondary in self.secondary.drain(..) {
268 12838 : scheduler.update_node_ref_counts(
269 12838 : secondary,
270 12838 : self.preferred_az_id.as_ref(),
271 12838 : RefCountUpdate::RemoveSecondary,
272 12838 : );
273 12838 : }
274 12857 : }
275 :
276 : /// Remove the last secondary node from the list of secondaries
277 0 : pub(crate) fn pop_secondary(&mut self, scheduler: &mut Scheduler) {
278 0 : if let Some(node_id) = self.secondary.pop() {
279 0 : scheduler.update_node_ref_counts(
280 0 : node_id,
281 0 : self.preferred_az_id.as_ref(),
282 0 : RefCountUpdate::RemoveSecondary,
283 0 : );
284 0 : }
285 0 : }
286 :
287 12857 : pub(crate) fn clear(&mut self, scheduler: &mut Scheduler) {
288 12857 : if let Some(old_attached) = self.attached.take() {
289 12855 : scheduler.update_node_ref_counts(
290 12855 : old_attached,
291 12855 : self.preferred_az_id.as_ref(),
292 12855 : RefCountUpdate::Detach,
293 12855 : );
294 12855 : }
295 :
296 12857 : self.clear_secondary(scheduler);
297 12857 : }
298 :
299 12897 : pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
300 12897 : let mut result = Vec::new();
301 12897 : if let Some(p) = self.attached {
302 12892 : result.push(p)
303 5 : }
304 :
305 12897 : result.extend(self.secondary.iter().copied());
306 12897 :
307 12897 : result
308 12897 : }
309 :
310 12618 : pub(crate) fn get_attached(&self) -> &Option<NodeId> {
311 12618 : &self.attached
312 12618 : }
313 :
314 12696 : pub(crate) fn get_secondary(&self) -> &Vec<NodeId> {
315 12696 : &self.secondary
316 12696 : }
317 :
318 : /// If the node is in use as the attached location, demote it into
319 : /// the list of secondary locations. This is used when a node goes offline,
320 : /// and we want to use a different node for attachment, but not permanently
321 : /// forget the location on the offline node.
322 : ///
323 : /// Returns true if a change was made
324 7 : pub(crate) fn demote_attached(&mut self, scheduler: &mut Scheduler, node_id: NodeId) -> bool {
325 7 : if self.attached == Some(node_id) {
326 7 : self.attached = None;
327 7 : self.secondary.push(node_id);
328 7 : scheduler.update_node_ref_counts(
329 7 : node_id,
330 7 : self.preferred_az_id.as_ref(),
331 7 : RefCountUpdate::DemoteAttached,
332 7 : );
333 7 : true
334 : } else {
335 0 : false
336 : }
337 7 : }
338 : }
339 :
340 : impl Drop for IntentState {
341 12858 : fn drop(&mut self) {
342 12858 : // Must clear before dropping, to avoid leaving stale refcounts in the Scheduler.
343 12858 : // We do not check this while panicking, to avoid polluting unit test failures or
344 12858 : // other assertions with this assertion's output. It's still wrong to leak these,
345 12858 : // but if we already have a panic then we don't need to independently flag this case.
346 12858 : if !(std::thread::panicking()) {
347 12858 : debug_assert!(self.attached.is_none() && self.secondary.is_empty());
348 0 : }
349 12857 : }
350 : }
351 :
352 0 : #[derive(Default, Clone, Serialize, Deserialize, Debug)]
353 : pub(crate) struct ObservedState {
354 : pub(crate) locations: HashMap<NodeId, ObservedStateLocation>,
355 : }
356 :
357 : /// Our latest knowledge of how this tenant is configured in the outside world.
358 : ///
359 : /// Meaning:
360 : /// * No instance of this type exists for a node: we are certain that we have nothing configured on that
361 : /// node for this shard.
362 : /// * Instance exists with conf==None: we *might* have some state on that node, but we don't know
363 : /// what it is (e.g. we failed partway through configuring it)
364 : /// * Instance exists with conf==Some: this tells us what we last successfully configured on this node,
365 : /// and that configuration will still be present unless something external interfered.
366 0 : #[derive(Clone, Serialize, Deserialize, Debug)]
367 : pub(crate) struct ObservedStateLocation {
368 : /// If None, it means we do not know the status of this shard's location on this node, but
369 : /// we know that we might have some state on this node.
370 : pub(crate) conf: Option<LocationConfig>,
371 : }
372 :
373 : pub(crate) struct ReconcilerWaiter {
374 : // For observability purposes, remember the ID of the shard we're
375 : // waiting for.
376 : pub(crate) tenant_shard_id: TenantShardId,
377 :
378 : seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
379 : error_seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
380 : error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
381 : seq: Sequence,
382 : }
383 :
384 : pub(crate) enum ReconcilerStatus {
385 : Done,
386 : Failed,
387 : InProgress,
388 : }
389 :
390 : #[derive(thiserror::Error, Debug)]
391 : pub(crate) enum ReconcileWaitError {
392 : #[error("Timeout waiting for shard {0}")]
393 : Timeout(TenantShardId),
394 : #[error("shutting down")]
395 : Shutdown,
396 : #[error("Reconcile error on shard {0}: {1}")]
397 : Failed(TenantShardId, Arc<ReconcileError>),
398 : }
399 :
400 : #[derive(Eq, PartialEq, Debug, Clone)]
401 : pub(crate) struct ReplaceSecondary {
402 : old_node_id: NodeId,
403 : new_node_id: NodeId,
404 : }
405 :
406 : #[derive(Eq, PartialEq, Debug, Clone)]
407 : pub(crate) struct MigrateAttachment {
408 : pub(crate) old_attached_node_id: NodeId,
409 : pub(crate) new_attached_node_id: NodeId,
410 : }
411 :
412 : #[derive(Eq, PartialEq, Debug, Clone)]
413 : pub(crate) enum ScheduleOptimizationAction {
414 : // Replace one of our secondary locations with a different node
415 : ReplaceSecondary(ReplaceSecondary),
416 : // Migrate attachment to an existing secondary location
417 : MigrateAttachment(MigrateAttachment),
418 : // Create a secondary location, with the intent of later migrating to it
419 : CreateSecondary(NodeId),
420 : // Remove a secondary location that we previously created to facilitate a migration
421 : RemoveSecondary(NodeId),
422 : }
423 :
424 : #[derive(Eq, PartialEq, Debug, Clone)]
425 : pub(crate) struct ScheduleOptimization {
426 : // What was the reconcile sequence when we generated this optimization? The optimization
427 : // should only be applied if the shard's sequence is still at this value, in case other changes
428 : // happened between planning the optimization and applying it.
429 : sequence: Sequence,
430 :
431 : pub(crate) action: ScheduleOptimizationAction,
432 : }
433 :
434 : impl ReconcilerWaiter {
435 0 : pub(crate) async fn wait_timeout(&self, timeout: Duration) -> Result<(), ReconcileWaitError> {
436 0 : tokio::select! {
437 0 : result = self.seq_wait.wait_for_timeout(self.seq, timeout)=> {
438 0 : result.map_err(|e| match e {
439 0 : SeqWaitError::Timeout => ReconcileWaitError::Timeout(self.tenant_shard_id),
440 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown
441 0 : })?;
442 : },
443 0 : result = self.error_seq_wait.wait_for(self.seq) => {
444 0 : result.map_err(|e| match e {
445 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown,
446 0 : SeqWaitError::Timeout => unreachable!()
447 0 : })?;
448 :
449 0 : return Err(ReconcileWaitError::Failed(self.tenant_shard_id,
450 0 : self.error.lock().unwrap().clone().expect("If error_seq_wait was advanced error was set").clone()))
451 : }
452 : }
453 :
454 0 : Ok(())
455 0 : }
456 :
457 0 : pub(crate) fn get_status(&self) -> ReconcilerStatus {
458 0 : if self.seq_wait.would_wait_for(self.seq).is_ok() {
459 0 : ReconcilerStatus::Done
460 0 : } else if self.error_seq_wait.would_wait_for(self.seq).is_ok() {
461 0 : ReconcilerStatus::Failed
462 : } else {
463 0 : ReconcilerStatus::InProgress
464 : }
465 0 : }
466 : }
467 :
468 : /// Having spawned a reconciler task, the tenant shard's state will carry enough
469 : /// information to optionally cancel & await it later.
470 : pub(crate) struct ReconcilerHandle {
471 : sequence: Sequence,
472 : handle: JoinHandle<()>,
473 : cancel: CancellationToken,
474 : }
475 :
476 : pub(crate) enum ReconcileNeeded {
477 : /// shard either doesn't need reconciliation, or is forbidden from spawning a reconciler
478 : /// in its current state (e.g. shard split in progress, or ShardSchedulingPolicy forbids it)
479 : No,
480 : /// shard has a reconciler running, and its intent hasn't changed since that one was
481 : /// spawned: wait for the existing reconciler rather than spawning a new one.
482 : WaitExisting(ReconcilerWaiter),
483 : /// shard needs reconciliation: call into [`TenantShard::spawn_reconciler`]
484 : Yes,
485 : }
486 :
487 : /// Pending modification to the observed state of a tenant shard.
488 : /// Produced by [`Reconciler::observed_deltas`] and applied in [`crate::service::Service::process_result`].
489 : pub(crate) enum ObservedStateDelta {
490 : Upsert(Box<(NodeId, ObservedStateLocation)>),
491 : Delete(NodeId),
492 : }
493 :
494 : impl ObservedStateDelta {
495 0 : pub(crate) fn node_id(&self) -> &NodeId {
496 0 : match self {
497 0 : Self::Upsert(up) => &up.0,
498 0 : Self::Delete(nid) => nid,
499 : }
500 0 : }
501 : }
502 :
503 : /// When a reconcile task completes, it sends this result object
504 : /// to be applied to the primary TenantShard.
505 : pub(crate) struct ReconcileResult {
506 : pub(crate) sequence: Sequence,
507 : /// On errors, `observed` should be treated as an incompleted description
508 : /// of state (i.e. any nodes present in the result should override nodes
509 : /// present in the parent tenant state, but any unmentioned nodes should
510 : /// not be removed from parent tenant state)
511 : pub(crate) result: Result<(), ReconcileError>,
512 :
513 : pub(crate) tenant_shard_id: TenantShardId,
514 : pub(crate) generation: Option<Generation>,
515 : pub(crate) observed_deltas: Vec<ObservedStateDelta>,
516 :
517 : /// Set [`TenantShard::pending_compute_notification`] from this flag
518 : pub(crate) pending_compute_notification: bool,
519 : }
520 :
521 : impl ObservedState {
522 0 : pub(crate) fn new() -> Self {
523 0 : Self {
524 0 : locations: HashMap::new(),
525 0 : }
526 0 : }
527 :
528 0 : pub(crate) fn is_empty(&self) -> bool {
529 0 : self.locations.is_empty()
530 0 : }
531 : }
532 :
533 : impl TenantShard {
534 12841 : pub(crate) fn new(
535 12841 : tenant_shard_id: TenantShardId,
536 12841 : shard: ShardIdentity,
537 12841 : policy: PlacementPolicy,
538 12841 : preferred_az_id: Option<AvailabilityZone>,
539 12841 : ) -> Self {
540 12841 : metrics::METRICS_REGISTRY
541 12841 : .metrics_group
542 12841 : .storage_controller_tenant_shards
543 12841 : .inc();
544 12841 :
545 12841 : Self {
546 12841 : tenant_shard_id,
547 12841 : policy,
548 12841 : intent: IntentState::new(preferred_az_id),
549 12841 : generation: Some(Generation::new(0)),
550 12841 : shard,
551 12841 : observed: ObservedState::default(),
552 12841 : config: TenantConfig::default(),
553 12841 : reconciler: None,
554 12841 : splitting: SplitState::Idle,
555 12841 : sequence: Sequence(1),
556 12841 : delayed_reconcile: false,
557 12841 : waiter: Arc::new(SeqWait::new(Sequence(0))),
558 12841 : error_waiter: Arc::new(SeqWait::new(Sequence(0))),
559 12841 : last_error: Arc::default(),
560 12841 : pending_compute_notification: false,
561 12841 : scheduling_policy: ShardSchedulingPolicy::default(),
562 12841 : }
563 12841 : }
564 :
565 : /// For use on startup when learning state from pageservers: generate my [`IntentState`] from my
566 : /// [`ObservedState`], even if it violates my [`PlacementPolicy`]. Call [`Self::schedule`] next,
567 : /// to get an intent state that complies with placement policy. The overall goal is to do scheduling
568 : /// in a way that makes use of any configured locations that already exist in the outside world.
569 1 : pub(crate) fn intent_from_observed(&mut self, scheduler: &mut Scheduler) {
570 1 : // Choose an attached location by filtering observed locations, and then sorting to get the highest
571 1 : // generation
572 1 : let mut attached_locs = self
573 1 : .observed
574 1 : .locations
575 1 : .iter()
576 2 : .filter_map(|(node_id, l)| {
577 2 : if let Some(conf) = &l.conf {
578 2 : if conf.mode == LocationConfigMode::AttachedMulti
579 1 : || conf.mode == LocationConfigMode::AttachedSingle
580 1 : || conf.mode == LocationConfigMode::AttachedStale
581 : {
582 2 : Some((node_id, conf.generation))
583 : } else {
584 0 : None
585 : }
586 : } else {
587 0 : None
588 : }
589 2 : })
590 1 : .collect::<Vec<_>>();
591 1 :
592 2 : attached_locs.sort_by_key(|i| i.1);
593 1 : if let Some((node_id, _gen)) = attached_locs.into_iter().last() {
594 1 : self.intent.set_attached(scheduler, Some(*node_id));
595 1 : }
596 :
597 : // All remaining observed locations generate secondary intents. This includes None
598 : // observations, as these may well have some local content on disk that is usable (this
599 : // is an edge case that might occur if we restarted during a migration or other change)
600 : //
601 : // We may leave intent.attached empty if we didn't find any attached locations: [`Self::schedule`]
602 : // will take care of promoting one of these secondaries to be attached.
603 2 : self.observed.locations.keys().for_each(|node_id| {
604 2 : if Some(*node_id) != self.intent.attached {
605 1 : self.intent.push_secondary(scheduler, *node_id);
606 1 : }
607 2 : });
608 1 : }
609 :
610 : /// Part of [`Self::schedule`] that is used to choose exactly one node to act as the
611 : /// attached pageserver for a shard.
612 : ///
613 : /// Returns whether we modified it, and the NodeId selected.
614 12831 : fn schedule_attached(
615 12831 : &mut self,
616 12831 : scheduler: &mut Scheduler,
617 12831 : context: &ScheduleContext,
618 12831 : ) -> Result<(bool, NodeId), ScheduleError> {
619 : // No work to do if we already have an attached tenant
620 12831 : if let Some(node_id) = self.intent.attached {
621 0 : return Ok((false, node_id));
622 12831 : }
623 :
624 12831 : if let Some(promote_secondary) = self.preferred_secondary(scheduler) {
625 : // Promote a secondary
626 2 : tracing::debug!("Promoted secondary {} to attached", promote_secondary);
627 2 : self.intent.promote_attached(scheduler, promote_secondary);
628 2 : Ok((true, promote_secondary))
629 : } else {
630 : // Pick a fresh node: either we had no secondaries or none were schedulable
631 12829 : let node_id = scheduler.schedule_shard::<AttachedShardTag>(
632 12829 : &self.intent.secondary,
633 12829 : &self.intent.preferred_az_id,
634 12829 : context,
635 12829 : )?;
636 12829 : tracing::debug!("Selected {} as attached", node_id);
637 12829 : self.intent.set_attached(scheduler, Some(node_id));
638 12829 : Ok((true, node_id))
639 : }
640 12831 : }
641 :
642 : #[instrument(skip_all, fields(
643 : tenant_id=%self.tenant_shard_id.tenant_id,
644 : shard_id=%self.tenant_shard_id.shard_slug(),
645 : sequence=%self.sequence
646 : ))]
647 : pub(crate) fn schedule(
648 : &mut self,
649 : scheduler: &mut Scheduler,
650 : context: &mut ScheduleContext,
651 : ) -> Result<(), ScheduleError> {
652 : let r = self.do_schedule(scheduler, context);
653 :
654 : context.avoid(&self.intent.all_pageservers());
655 :
656 : r
657 : }
658 :
659 12835 : pub(crate) fn do_schedule(
660 12835 : &mut self,
661 12835 : scheduler: &mut Scheduler,
662 12835 : context: &ScheduleContext,
663 12835 : ) -> Result<(), ScheduleError> {
664 12835 : // TODO: before scheduling new nodes, check if any existing content in
665 12835 : // self.intent refers to pageservers that are offline, and pick other
666 12835 : // pageservers if so.
667 12835 :
668 12835 : // TODO: respect the splitting bit on tenants: if they are currently splitting then we may not
669 12835 : // change their attach location.
670 12835 :
671 12835 : match self.scheduling_policy {
672 12834 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {}
673 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
674 : // Warn to make it obvious why other things aren't happening/working, if we skip scheduling
675 1 : tracing::warn!(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(),
676 0 : "Scheduling is disabled by policy {:?}", self.scheduling_policy);
677 1 : return Ok(());
678 : }
679 : }
680 :
681 : // Build the set of pageservers already in use by this tenant, to avoid scheduling
682 : // more work on the same pageservers we're already using.
683 12834 : let mut modified = false;
684 :
685 : // Add/remove nodes to fulfil policy
686 : use PlacementPolicy::*;
687 12834 : match self.policy {
688 12831 : Attached(secondary_count) => {
689 : // Should have exactly one attached, and at least N secondaries
690 12831 : let (modified_attached, attached_node_id) =
691 12831 : self.schedule_attached(scheduler, context)?;
692 12831 : modified |= modified_attached;
693 12831 :
694 12831 : let mut used_pageservers = vec![attached_node_id];
695 25661 : while self.intent.secondary.len() < secondary_count {
696 12830 : let node_id = scheduler.schedule_shard::<SecondaryShardTag>(
697 12830 : &used_pageservers,
698 12830 : &self.intent.preferred_az_id,
699 12830 : context,
700 12830 : )?;
701 12830 : self.intent.push_secondary(scheduler, node_id);
702 12830 : used_pageservers.push(node_id);
703 12830 : modified = true;
704 : }
705 : }
706 : Secondary => {
707 3 : if let Some(node_id) = self.intent.get_attached() {
708 2 : // Populate secondary by demoting the attached node
709 2 : self.intent.demote_attached(scheduler, *node_id);
710 2 :
711 2 : modified = true;
712 2 : } else if self.intent.secondary.is_empty() {
713 : // Populate secondary by scheduling a fresh node
714 : //
715 : // We use [`AttachedShardTag`] because when a secondary location is the only one
716 : // a shard has, we expect that its next use will be as an attached location: we want
717 : // the tenant to be ready to warm up and run fast in their preferred AZ.
718 1 : let node_id = scheduler.schedule_shard::<AttachedShardTag>(
719 1 : &[],
720 1 : &self.intent.preferred_az_id,
721 1 : context,
722 1 : )?;
723 1 : self.intent.push_secondary(scheduler, node_id);
724 1 : modified = true;
725 0 : }
726 4 : while self.intent.secondary.len() > 1 {
727 1 : // If we have multiple secondaries (e.g. when transitioning from Attached to Secondary and
728 1 : // having just demoted our attached location), then we should prefer to keep the location
729 1 : // in our preferred AZ. Tenants in Secondary mode want to be in the preferred AZ so that
730 1 : // they have a warm location to become attached when transitioning back into Attached.
731 1 :
732 1 : let mut candidates = self.intent.get_secondary().clone();
733 1 : // Sort to get secondaries outside preferred AZ last
734 1 : candidates
735 2 : .sort_by_key(|n| scheduler.get_node_az(n).as_ref() != self.preferred_az());
736 1 : let secondary_to_remove = candidates.pop().unwrap();
737 1 : self.intent.remove_secondary(scheduler, secondary_to_remove);
738 1 : modified = true;
739 1 : }
740 : }
741 : Detached => {
742 : // Never add locations in this mode
743 0 : if self.intent.get_attached().is_some() || !self.intent.get_secondary().is_empty() {
744 0 : self.intent.clear(scheduler);
745 0 : modified = true;
746 0 : }
747 : }
748 : }
749 :
750 12834 : if modified {
751 12834 : self.sequence.0 += 1;
752 12834 : }
753 :
754 12834 : Ok(())
755 12835 : }
756 :
757 : /// Reschedule this tenant shard to one of its secondary locations. Returns a scheduling error
758 : /// if the swap is not possible and leaves the intent state in its original state.
759 : ///
760 : /// Arguments:
761 : /// `attached_to`: the currently attached location matching the intent state (may be None if the
762 : /// shard is not attached)
763 : /// `promote_to`: an optional secondary location of this tenant shard. If set to None, we ask
764 : /// the scheduler to recommend a node
765 0 : pub(crate) fn reschedule_to_secondary(
766 0 : &mut self,
767 0 : promote_to: Option<NodeId>,
768 0 : scheduler: &mut Scheduler,
769 0 : ) -> Result<(), ScheduleError> {
770 0 : let promote_to = match promote_to {
771 0 : Some(node) => node,
772 0 : None => match self.preferred_secondary(scheduler) {
773 0 : Some(node) => node,
774 : None => {
775 0 : return Err(ScheduleError::ImpossibleConstraint);
776 : }
777 : },
778 : };
779 :
780 0 : assert!(self.intent.get_secondary().contains(&promote_to));
781 :
782 0 : if let Some(node) = self.intent.get_attached() {
783 0 : let demoted = self.intent.demote_attached(scheduler, *node);
784 0 : if !demoted {
785 0 : return Err(ScheduleError::ImpossibleConstraint);
786 0 : }
787 0 : }
788 :
789 0 : self.intent.promote_attached(scheduler, promote_to);
790 0 :
791 0 : // Increment the sequence number for the edge case where a
792 0 : // reconciler is already running to avoid waiting on the
793 0 : // current reconcile instead of spawning a new one.
794 0 : self.sequence = self.sequence.next();
795 0 :
796 0 : Ok(())
797 0 : }
798 :
799 : /// Returns None if the current location's score is unavailable, i.e. cannot draw a conclusion
800 63 : fn is_better_location<T: ShardTag>(
801 63 : &self,
802 63 : scheduler: &mut Scheduler,
803 63 : schedule_context: &ScheduleContext,
804 63 : current: NodeId,
805 63 : candidate: NodeId,
806 63 : ) -> Option<bool> {
807 63 : let Some(candidate_score) = scheduler.compute_node_score::<T::Score>(
808 63 : candidate,
809 63 : &self.intent.preferred_az_id,
810 63 : schedule_context,
811 63 : ) else {
812 : // The candidate node is unavailable for scheduling or otherwise couldn't get a score
813 1 : return None;
814 : };
815 :
816 62 : match scheduler.compute_node_score::<T::Score>(
817 62 : current,
818 62 : &self.intent.preferred_az_id,
819 62 : schedule_context,
820 62 : ) {
821 62 : Some(current_score) => {
822 62 : // Ignore utilization components when comparing scores: we don't want to migrate
823 62 : // because of transient load variations, it risks making the system thrash, and
824 62 : // migrating for utilization requires a separate high level view of the system to
825 62 : // e.g. prioritize moving larger or smaller tenants, rather than arbitrarily
826 62 : // moving things around in the order that we hit this function.
827 62 : let candidate_score = candidate_score.for_optimization();
828 62 : let current_score = current_score.for_optimization();
829 62 :
830 62 : if candidate_score < current_score {
831 8 : tracing::info!("Found a lower scoring location! {candidate} is better than {current} ({candidate_score:?} is better than {current_score:?})");
832 8 : Some(true)
833 : } else {
834 : // The candidate node is no better than our current location, so don't migrate
835 54 : tracing::debug!(
836 0 : "Candidate node {candidate} is no better than our current location {current} (candidate {candidate_score:?} vs current {current_score:?})",
837 : );
838 54 : Some(false)
839 : }
840 : }
841 : None => {
842 : // The current node is unavailable for scheduling, so we can't make any sensible
843 : // decisions about optimisation. This should be a transient state -- if the node
844 : // is offline then it will get evacuated, if is blocked by a scheduling mode
845 : // then we will respect that mode by doing nothing.
846 0 : tracing::debug!("Current node {current} is unavailable for scheduling");
847 0 : None
848 : }
849 : }
850 63 : }
851 :
852 47 : fn find_better_location<T: ShardTag>(
853 47 : &self,
854 47 : scheduler: &mut Scheduler,
855 47 : schedule_context: &ScheduleContext,
856 47 : current: NodeId,
857 47 : hard_exclude: &[NodeId],
858 47 : ) -> Option<NodeId> {
859 : // Look for a lower-scoring location to attach to
860 47 : let Ok(candidate_node) = scheduler.schedule_shard::<T>(
861 47 : hard_exclude,
862 47 : &self.intent.preferred_az_id,
863 47 : schedule_context,
864 47 : ) else {
865 : // A scheduling error means we have no possible candidate replacements
866 0 : tracing::debug!("No candidate node found");
867 0 : return None;
868 : };
869 :
870 47 : if candidate_node == current {
871 : // We're already at the best possible location, so don't migrate
872 24 : tracing::debug!("Candidate node {candidate_node} is already in use");
873 24 : return None;
874 23 : }
875 23 :
876 23 : self.is_better_location::<T>(scheduler, schedule_context, current, candidate_node)
877 23 : .and_then(|better| if better { Some(candidate_node) } else { None })
878 47 : }
879 :
880 : /// This function is an optimization, used to avoid doing large numbers of scheduling operations
881 : /// when looking for optimizations. This function uses knowledge of how scores work to do some
882 : /// fast checks for whether it may to be possible to improve a score.
883 : ///
884 : /// If we return true, it only means that optimization _might_ be possible, not that it necessarily is. If we
885 : /// return no, it definitely means that calling [`Self::optimize_attachment`] or [`Self::optimize_secondary`] would do no
886 : /// work.
887 3 : pub(crate) fn maybe_optimizable(
888 3 : &self,
889 3 : scheduler: &mut Scheduler,
890 3 : schedule_context: &ScheduleContext,
891 3 : ) -> bool {
892 3 : // Sharded tenant: check if any locations have a nonzero affinity score
893 3 : if self.shard.count >= ShardCount(1) {
894 3 : let schedule_context = schedule_context.project_detach(self);
895 5 : for node in self.intent.all_pageservers() {
896 5 : if let Some(af) = schedule_context.nodes.get(&node) {
897 5 : if *af > AffinityScore(0) {
898 3 : return true;
899 2 : }
900 0 : }
901 : }
902 0 : }
903 :
904 : // Attached tenant: check if the attachment is outside the preferred AZ
905 0 : if let PlacementPolicy::Attached(_) = self.policy {
906 0 : if let Some(attached) = self.intent.get_attached() {
907 0 : if scheduler.get_node_az(attached) != self.intent.preferred_az_id {
908 0 : return true;
909 0 : }
910 0 : }
911 0 : }
912 :
913 : // Tenant with secondary locations: check if any are within the preferred AZ
914 0 : for secondary in self.intent.get_secondary() {
915 0 : if scheduler.get_node_az(secondary) == self.intent.preferred_az_id {
916 0 : return true;
917 0 : }
918 : }
919 :
920 : // Does the tenant have excess secondaries?
921 0 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
922 0 : return true;
923 0 : }
924 0 :
925 0 : // Fall through: no optimizations possible
926 0 : false
927 3 : }
928 :
929 : /// Optimize attachments: if a shard has a secondary location that is preferable to
930 : /// its primary location based on soft constraints, switch that secondary location
931 : /// to be attached.
932 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
933 : pub(crate) fn optimize_attachment(
934 : &self,
935 : scheduler: &mut Scheduler,
936 : schedule_context: &ScheduleContext,
937 : ) -> Option<ScheduleOptimization> {
938 : let attached = (*self.intent.get_attached())?;
939 :
940 : let schedule_context = schedule_context.project_detach(self);
941 :
942 : // If we already have a secondary that is higher-scoring than out current location,
943 : // then simply migrate to it.
944 : for secondary in self.intent.get_secondary() {
945 : if let Some(true) = self.is_better_location::<AttachedShardTag>(
946 : scheduler,
947 : &schedule_context,
948 : attached,
949 : *secondary,
950 : ) {
951 : return Some(ScheduleOptimization {
952 : sequence: self.sequence,
953 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
954 : old_attached_node_id: attached,
955 : new_attached_node_id: *secondary,
956 : }),
957 : });
958 : }
959 : }
960 :
961 : // Given that none of our current secondaries is a better location than our current
962 : // attached location (checked above), we may trim any secondaries that are not needed
963 : // for the placement policy.
964 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
965 : // This code path cleans up extra secondaries after migrating, and/or
966 : // trims extra secondaries after a PlacementPolicy::Attached(N) was
967 : // modified to decrease N.
968 :
969 : let secondary_scores = self
970 : .intent
971 : .get_secondary()
972 : .iter()
973 10 : .map(|node_id| {
974 10 : (
975 10 : *node_id,
976 10 : scheduler.compute_node_score::<NodeSecondarySchedulingScore>(
977 10 : *node_id,
978 10 : &self.intent.preferred_az_id,
979 10 : &schedule_context,
980 10 : ),
981 10 : )
982 10 : })
983 : .collect::<HashMap<_, _>>();
984 :
985 10 : if secondary_scores.iter().any(|score| score.1.is_none()) {
986 : // Trivial case: if we only have one secondary, drop that one
987 : if self.intent.get_secondary().len() == 1 {
988 : return Some(ScheduleOptimization {
989 : sequence: self.sequence,
990 : action: ScheduleOptimizationAction::RemoveSecondary(
991 : *self.intent.get_secondary().first().unwrap(),
992 : ),
993 : });
994 : }
995 :
996 : // Try to find a "good" secondary to keep, without relying on scores (one or more nodes is in a state
997 : // where its score can't be calculated), and drop the others. This enables us to make progress in
998 : // most cases, even if some nodes are offline or have scheduling=pause set.
999 :
1000 : debug_assert!(self.intent.attached.is_some()); // We should not make it here unless attached -- this
1001 : // logic presumes we are in a mode where we want secondaries to be in non-home AZ
1002 1 : if let Some(retain_secondary) = self.intent.get_secondary().iter().find(|n| {
1003 1 : let in_home_az = scheduler.get_node_az(n) == self.intent.preferred_az_id;
1004 1 : let is_available = secondary_scores
1005 1 : .get(n)
1006 1 : .expect("Built from same list of nodes")
1007 1 : .is_some();
1008 1 : is_available && !in_home_az
1009 1 : }) {
1010 : // Great, we found one to retain. Pick some other to drop.
1011 : if let Some(victim) = self
1012 : .intent
1013 : .get_secondary()
1014 : .iter()
1015 2 : .find(|n| n != &retain_secondary)
1016 : {
1017 : return Some(ScheduleOptimization {
1018 : sequence: self.sequence,
1019 : action: ScheduleOptimizationAction::RemoveSecondary(*victim),
1020 : });
1021 : }
1022 : }
1023 :
1024 : // Fall through: we didn't identify one to remove. This ought to be rare.
1025 : tracing::warn!("Keeping extra secondaries: can't determine which of {:?} to remove (some nodes offline?)",
1026 : self.intent.get_secondary()
1027 : );
1028 : } else {
1029 : let victim = secondary_scores
1030 : .iter()
1031 8 : .max_by_key(|score| score.1.unwrap())
1032 : .unwrap()
1033 : .0;
1034 : return Some(ScheduleOptimization {
1035 : sequence: self.sequence,
1036 : action: ScheduleOptimizationAction::RemoveSecondary(*victim),
1037 : });
1038 : }
1039 : }
1040 :
1041 : let replacement = self.find_better_location::<AttachedShardTag>(
1042 : scheduler,
1043 : &schedule_context,
1044 : attached,
1045 : &[], // Don't exclude secondaries: our preferred attachment location may be a secondary
1046 : );
1047 :
1048 : // We have found a candidate and confirmed that its score is preferable
1049 : // to our current location. See if we have a secondary location in the preferred location already: if not,
1050 : // then create one.
1051 : if let Some(replacement) = replacement {
1052 : // If we are currently in non-preferred AZ, then the scheduler might suggest a location that is better, but still
1053 : // not in our preferred AZ. Migration has a cost in resources an impact to the workload, so we want to avoid doing
1054 : // multiple hops where we might go to some other AZ before eventually finding a suitable location in our preferred
1055 : // AZ: skip this optimization if it is not in our final, preferred AZ.
1056 : //
1057 : // This should be a transient state, there should always be capacity eventually in our preferred AZ (even if nodes
1058 : // there are too overloaded for scheduler to suggest them, more should be provisioned eventually).
1059 : if self.intent.preferred_az_id.is_some()
1060 : && scheduler.get_node_az(&replacement) != self.intent.preferred_az_id
1061 : {
1062 : tracing::debug!(
1063 : "Candidate node {replacement} is not in preferred AZ {:?}",
1064 : self.intent.preferred_az_id
1065 : );
1066 :
1067 : // This should only happen if our current location is not in the preferred AZ, otherwise
1068 : // [`Self::find_better_location`]` should have rejected any other location outside the preferred Az, because
1069 : // AZ is the highest priority part of NodeAttachmentSchedulingScore.
1070 : debug_assert!(scheduler.get_node_az(&attached) != self.intent.preferred_az_id);
1071 :
1072 : return None;
1073 : }
1074 :
1075 : if !self.intent.get_secondary().contains(&replacement) {
1076 : Some(ScheduleOptimization {
1077 : sequence: self.sequence,
1078 : action: ScheduleOptimizationAction::CreateSecondary(replacement),
1079 : })
1080 : } else {
1081 : // We already have a secondary in the preferred location, let's try migrating to it. Our caller
1082 : // will check the warmth of the destination before deciding whether to really execute this.
1083 : Some(ScheduleOptimization {
1084 : sequence: self.sequence,
1085 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1086 : old_attached_node_id: attached,
1087 : new_attached_node_id: replacement,
1088 : }),
1089 : })
1090 : }
1091 : } else {
1092 : // We didn't find somewhere we'd rather be, and we don't have any excess secondaries
1093 : // to clean up: no action required.
1094 : None
1095 : }
1096 : }
1097 :
1098 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1099 : pub(crate) fn optimize_secondary(
1100 : &self,
1101 : scheduler: &mut Scheduler,
1102 : schedule_context: &ScheduleContext,
1103 : ) -> Option<ScheduleOptimization> {
1104 : if self.intent.get_secondary().len() > self.policy.want_secondaries() {
1105 : // We have extra secondaries, perhaps to facilitate a migration of the attached location:
1106 : // do nothing, it is up to [`Self::optimize_attachment`] to clean them up. When that's done,
1107 : // and we are called again, we will proceed.
1108 : tracing::debug!("Too many secondaries: skipping");
1109 : return None;
1110 : }
1111 :
1112 : let schedule_context = schedule_context.project_detach(self);
1113 :
1114 : for secondary in self.intent.get_secondary() {
1115 : // Make sure we don't try to migrate a secondary to our attached location: this case happens
1116 : // easily in environments without multiple AZs.
1117 : let exclude = match self.intent.attached {
1118 : Some(attached) => vec![attached],
1119 : None => vec![],
1120 : };
1121 :
1122 : let replacement = match &self.policy {
1123 : PlacementPolicy::Attached(_) => {
1124 : // Secondaries for an attached shard should be scheduled using `SecondaryShardTag`
1125 : // to avoid placing them in the preferred AZ.
1126 : self.find_better_location::<SecondaryShardTag>(
1127 : scheduler,
1128 : &schedule_context,
1129 : *secondary,
1130 : &exclude,
1131 : )
1132 : }
1133 : PlacementPolicy::Secondary => {
1134 : // In secondary-only mode, we want our secondary locations in the preferred AZ,
1135 : // so that they're ready to take over as an attached location when we transition
1136 : // into PlacementPolicy::Attached.
1137 : self.find_better_location::<AttachedShardTag>(
1138 : scheduler,
1139 : &schedule_context,
1140 : *secondary,
1141 : &exclude,
1142 : )
1143 : }
1144 : PlacementPolicy::Detached => None,
1145 : };
1146 :
1147 : assert!(replacement != Some(*secondary));
1148 : if let Some(replacement) = replacement {
1149 : // We have found a candidate and confirmed that its score is preferable
1150 : // to our current location. See if we have a secondary location in the preferred location already: if not,
1151 : // then create one.
1152 : return Some(ScheduleOptimization {
1153 : sequence: self.sequence,
1154 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1155 : old_node_id: *secondary,
1156 : new_node_id: replacement,
1157 : }),
1158 : });
1159 : }
1160 : }
1161 :
1162 : None
1163 : }
1164 :
1165 : /// Return true if the optimization was really applied: it will not be applied if the optimization's
1166 : /// sequence is behind this tenant shard's
1167 14 : pub(crate) fn apply_optimization(
1168 14 : &mut self,
1169 14 : scheduler: &mut Scheduler,
1170 14 : optimization: ScheduleOptimization,
1171 14 : ) -> bool {
1172 14 : if optimization.sequence != self.sequence {
1173 0 : return false;
1174 14 : }
1175 14 :
1176 14 : metrics::METRICS_REGISTRY
1177 14 : .metrics_group
1178 14 : .storage_controller_schedule_optimization
1179 14 : .inc();
1180 14 :
1181 14 : match optimization.action {
1182 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1183 4 : old_attached_node_id,
1184 4 : new_attached_node_id,
1185 4 : }) => {
1186 4 : self.intent.demote_attached(scheduler, old_attached_node_id);
1187 4 : self.intent
1188 4 : .promote_attached(scheduler, new_attached_node_id);
1189 4 : }
1190 : ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1191 1 : old_node_id,
1192 1 : new_node_id,
1193 1 : }) => {
1194 1 : self.intent.remove_secondary(scheduler, old_node_id);
1195 1 : self.intent.push_secondary(scheduler, new_node_id);
1196 1 : }
1197 3 : ScheduleOptimizationAction::CreateSecondary(new_node_id) => {
1198 3 : self.intent.push_secondary(scheduler, new_node_id);
1199 3 : }
1200 6 : ScheduleOptimizationAction::RemoveSecondary(old_secondary) => {
1201 6 : self.intent.remove_secondary(scheduler, old_secondary);
1202 6 : }
1203 : }
1204 :
1205 14 : true
1206 14 : }
1207 :
1208 : /// When a shard has several secondary locations, we need to pick one in situations where
1209 : /// we promote one of them to an attached location:
1210 : /// - When draining a node for restart
1211 : /// - When responding to a node failure
1212 : ///
1213 : /// In this context, 'preferred' does not mean the node with the best scheduling score: instead
1214 : /// we want to pick the node which is best for use _temporarily_ while the previous attached location
1215 : /// is unavailable (e.g. because it's down or deploying). That means we prefer to use secondary
1216 : /// locations in a non-preferred AZ, as they're more likely to have awarm cache than a temporary
1217 : /// secondary in the preferred AZ (which are usually only created for migrations, and if they exist
1218 : /// they're probably not warmed up yet). The latter behavior is based oni
1219 : ///
1220 : /// If the input is empty, or all the nodes are not elegible for scheduling, return None: the
1221 : /// caller needs to a pick a node some other way.
1222 12831 : pub(crate) fn preferred_secondary(&self, scheduler: &Scheduler) -> Option<NodeId> {
1223 12831 : let candidates = scheduler.filter_usable_nodes(&self.intent.secondary);
1224 12831 :
1225 12831 : // We will sort candidates to prefer nodes which are _not_ in our preferred AZ, i.e. we prefer
1226 12831 : // to migrate to a long-lived secondary location (which would have been scheduled in a non-preferred AZ),
1227 12831 : // rather than a short-lived secondary location being used for optimization/migration (which would have
1228 12831 : // been scheduled in our preferred AZ).
1229 12831 : let mut candidates = candidates
1230 12831 : .iter()
1231 12831 : .map(|(node_id, node_az)| {
1232 2 : if node_az == &self.intent.preferred_az_id {
1233 1 : (1, *node_id)
1234 : } else {
1235 1 : (0, *node_id)
1236 : }
1237 12831 : })
1238 12831 : .collect::<Vec<_>>();
1239 12831 :
1240 12831 : candidates.sort();
1241 12831 :
1242 12831 : candidates.first().map(|i| i.1)
1243 12831 : }
1244 :
1245 : /// Query whether the tenant's observed state for attached node matches its intent state, and if so,
1246 : /// yield the node ID. This is appropriate for emitting compute hook notifications: we are checking that
1247 : /// the node in question is not only where we intend to attach, but that the tenant is indeed already attached there.
1248 : ///
1249 : /// Reconciliation may still be needed for other aspects of state such as secondaries (see [`Self::dirty`]): this
1250 : /// funciton should not be used to decide whether to reconcile.
1251 0 : pub(crate) fn stably_attached(&self) -> Option<NodeId> {
1252 0 : if let Some(attach_intent) = self.intent.attached {
1253 0 : match self.observed.locations.get(&attach_intent) {
1254 0 : Some(loc) => match &loc.conf {
1255 0 : Some(conf) => match conf.mode {
1256 : LocationConfigMode::AttachedMulti
1257 : | LocationConfigMode::AttachedSingle
1258 : | LocationConfigMode::AttachedStale => {
1259 : // Our intent and observed state agree that this node is in an attached state.
1260 0 : Some(attach_intent)
1261 : }
1262 : // Our observed config is not an attached state
1263 0 : _ => None,
1264 : },
1265 : // Our observed state is None, i.e. in flux
1266 0 : None => None,
1267 : },
1268 : // We have no observed state for this node
1269 0 : None => None,
1270 : }
1271 : } else {
1272 : // Our intent is not to attach
1273 0 : None
1274 : }
1275 0 : }
1276 :
1277 0 : fn dirty(&self, nodes: &Arc<HashMap<NodeId, Node>>) -> bool {
1278 0 : let mut dirty_nodes = HashSet::new();
1279 :
1280 0 : if let Some(node_id) = self.intent.attached {
1281 : // Maybe panic: it is a severe bug if we try to attach while generation is null.
1282 0 : let generation = self
1283 0 : .generation
1284 0 : .expect("Attempted to enter attached state without a generation");
1285 0 :
1286 0 : let wanted_conf =
1287 0 : attached_location_conf(generation, &self.shard, &self.config, &self.policy);
1288 0 : match self.observed.locations.get(&node_id) {
1289 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
1290 0 : Some(_) | None => {
1291 0 : dirty_nodes.insert(node_id);
1292 0 : }
1293 : }
1294 0 : }
1295 :
1296 0 : for node_id in &self.intent.secondary {
1297 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
1298 0 : match self.observed.locations.get(node_id) {
1299 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
1300 0 : Some(_) | None => {
1301 0 : dirty_nodes.insert(*node_id);
1302 0 : }
1303 : }
1304 : }
1305 :
1306 0 : for node_id in self.observed.locations.keys() {
1307 0 : if self.intent.attached != Some(*node_id) && !self.intent.secondary.contains(node_id) {
1308 0 : // We have observed state that isn't part of our intent: need to clean it up.
1309 0 : dirty_nodes.insert(*node_id);
1310 0 : }
1311 : }
1312 :
1313 0 : dirty_nodes.retain(|node_id| {
1314 0 : nodes
1315 0 : .get(node_id)
1316 0 : .map(|n| n.is_available())
1317 0 : .unwrap_or(false)
1318 0 : });
1319 0 :
1320 0 : !dirty_nodes.is_empty()
1321 0 : }
1322 :
1323 : #[allow(clippy::too_many_arguments)]
1324 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1325 : pub(crate) fn get_reconcile_needed(
1326 : &mut self,
1327 : pageservers: &Arc<HashMap<NodeId, Node>>,
1328 : ) -> ReconcileNeeded {
1329 : // If there are any ambiguous observed states, and the nodes they refer to are available,
1330 : // we should reconcile to clean them up.
1331 : let mut dirty_observed = false;
1332 : for (node_id, observed_loc) in &self.observed.locations {
1333 : let node = pageservers
1334 : .get(node_id)
1335 : .expect("Nodes may not be removed while referenced");
1336 : if observed_loc.conf.is_none() && node.is_available() {
1337 : dirty_observed = true;
1338 : break;
1339 : }
1340 : }
1341 :
1342 : let active_nodes_dirty = self.dirty(pageservers);
1343 :
1344 : // Even if there is no pageserver work to be done, if we have a pending notification to computes,
1345 : // wake up a reconciler to send it.
1346 : let do_reconcile =
1347 : active_nodes_dirty || dirty_observed || self.pending_compute_notification;
1348 :
1349 : if !do_reconcile {
1350 : tracing::debug!("Not dirty, no reconciliation needed.");
1351 : return ReconcileNeeded::No;
1352 : }
1353 :
1354 : // If we are currently splitting, then never start a reconciler task: the splitting logic
1355 : // requires that shards are not interfered with while it runs. Do this check here rather than
1356 : // up top, so that we only log this message if we would otherwise have done a reconciliation.
1357 : if !matches!(self.splitting, SplitState::Idle) {
1358 : tracing::info!("Refusing to reconcile, splitting in progress");
1359 : return ReconcileNeeded::No;
1360 : }
1361 :
1362 : // Reconcile already in flight for the current sequence?
1363 : if let Some(handle) = &self.reconciler {
1364 : if handle.sequence == self.sequence {
1365 : tracing::info!(
1366 : "Reconciliation already in progress for sequence {:?}",
1367 : self.sequence,
1368 : );
1369 : return ReconcileNeeded::WaitExisting(ReconcilerWaiter {
1370 : tenant_shard_id: self.tenant_shard_id,
1371 : seq_wait: self.waiter.clone(),
1372 : error_seq_wait: self.error_waiter.clone(),
1373 : error: self.last_error.clone(),
1374 : seq: self.sequence,
1375 : });
1376 : }
1377 : }
1378 :
1379 : // Pre-checks done: finally check whether we may actually do the work
1380 : match self.scheduling_policy {
1381 : ShardSchedulingPolicy::Active
1382 : | ShardSchedulingPolicy::Essential
1383 : | ShardSchedulingPolicy::Pause => {}
1384 : ShardSchedulingPolicy::Stop => {
1385 : // We only reach this point if there is work to do and we're going to skip
1386 : // doing it: warn it obvious why this tenant isn't doing what it ought to.
1387 : tracing::warn!("Skipping reconcile for policy {:?}", self.scheduling_policy);
1388 : return ReconcileNeeded::No;
1389 : }
1390 : }
1391 :
1392 : ReconcileNeeded::Yes
1393 : }
1394 :
1395 : /// Ensure the sequence number is set to a value where waiting for this value will make us wait
1396 : /// for the next reconcile: i.e. it is ahead of all completed or running reconcilers.
1397 : ///
1398 : /// Constructing a ReconcilerWaiter with the resulting sequence number gives the property
1399 : /// that the waiter will not complete until some future Reconciler is constructed and run.
1400 0 : fn ensure_sequence_ahead(&mut self) {
1401 0 : // Find the highest sequence for which a Reconciler has previously run or is currently
1402 0 : // running
1403 0 : let max_seen = std::cmp::max(
1404 0 : self.reconciler
1405 0 : .as_ref()
1406 0 : .map(|r| r.sequence)
1407 0 : .unwrap_or(Sequence(0)),
1408 0 : std::cmp::max(self.waiter.load(), self.error_waiter.load()),
1409 0 : );
1410 0 :
1411 0 : if self.sequence <= max_seen {
1412 0 : self.sequence = max_seen.next();
1413 0 : }
1414 0 : }
1415 :
1416 : /// Create a waiter that will wait for some future Reconciler that hasn't been spawned yet.
1417 : ///
1418 : /// This is appropriate when you can't spawn a reconciler (e.g. due to resource limits), but
1419 : /// you would like to wait on the next reconciler that gets spawned in the background.
1420 0 : pub(crate) fn future_reconcile_waiter(&mut self) -> ReconcilerWaiter {
1421 0 : self.ensure_sequence_ahead();
1422 0 :
1423 0 : ReconcilerWaiter {
1424 0 : tenant_shard_id: self.tenant_shard_id,
1425 0 : seq_wait: self.waiter.clone(),
1426 0 : error_seq_wait: self.error_waiter.clone(),
1427 0 : error: self.last_error.clone(),
1428 0 : seq: self.sequence,
1429 0 : }
1430 0 : }
1431 :
1432 0 : async fn reconcile(
1433 0 : sequence: Sequence,
1434 0 : mut reconciler: Reconciler,
1435 0 : must_notify: bool,
1436 0 : ) -> ReconcileResult {
1437 : // Attempt to make observed state match intent state
1438 0 : let result = reconciler.reconcile().await;
1439 :
1440 : // If we know we had a pending compute notification from some previous action, send a notification irrespective
1441 : // of whether the above reconcile() did any work. It has to be Ok() though, because otherwise we might be
1442 : // sending a notification of a location that isn't really attached.
1443 0 : if result.is_ok() && must_notify {
1444 : // If this fails we will send the need to retry in [`ReconcileResult::pending_compute_notification`]
1445 0 : reconciler.compute_notify().await.ok();
1446 0 : } else if must_notify {
1447 0 : // Carry this flag so that the reconciler's result will indicate that it still needs to retry
1448 0 : // the compute hook notification eventually.
1449 0 : reconciler.compute_notify_failure = true;
1450 0 : }
1451 :
1452 : // Update result counter
1453 0 : let outcome_label = match &result {
1454 0 : Ok(_) => ReconcileOutcome::Success,
1455 0 : Err(ReconcileError::Cancel) => ReconcileOutcome::Cancel,
1456 0 : Err(_) => ReconcileOutcome::Error,
1457 : };
1458 :
1459 0 : metrics::METRICS_REGISTRY
1460 0 : .metrics_group
1461 0 : .storage_controller_reconcile_complete
1462 0 : .inc(ReconcileCompleteLabelGroup {
1463 0 : status: outcome_label,
1464 0 : });
1465 0 :
1466 0 : // Constructing result implicitly drops Reconciler, freeing any ReconcileUnits before the Service might
1467 0 : // try and schedule more work in response to our result.
1468 0 : ReconcileResult {
1469 0 : sequence,
1470 0 : result,
1471 0 : tenant_shard_id: reconciler.tenant_shard_id,
1472 0 : generation: reconciler.generation,
1473 0 : observed_deltas: reconciler.observed_deltas(),
1474 0 : pending_compute_notification: reconciler.compute_notify_failure,
1475 0 : }
1476 0 : }
1477 :
1478 : #[allow(clippy::too_many_arguments)]
1479 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1480 : pub(crate) fn spawn_reconciler(
1481 : &mut self,
1482 : result_tx: &tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
1483 : pageservers: &Arc<HashMap<NodeId, Node>>,
1484 : compute_hook: &Arc<ComputeHook>,
1485 : reconciler_config: ReconcilerConfig,
1486 : service_config: &service::Config,
1487 : persistence: &Arc<Persistence>,
1488 : units: ReconcileUnits,
1489 : gate_guard: GateGuard,
1490 : cancel: &CancellationToken,
1491 : ) -> Option<ReconcilerWaiter> {
1492 : // Reconcile in flight for a stale sequence? Our sequence's task will wait for it before
1493 : // doing our sequence's work.
1494 : let old_handle = self.reconciler.take();
1495 :
1496 : // Build list of nodes from which the reconciler should detach
1497 : let mut detach = Vec::new();
1498 : for node_id in self.observed.locations.keys() {
1499 : if self.intent.get_attached() != &Some(*node_id)
1500 : && !self.intent.secondary.contains(node_id)
1501 : {
1502 : detach.push(
1503 : pageservers
1504 : .get(node_id)
1505 : .expect("Intent references non-existent pageserver")
1506 : .clone(),
1507 : )
1508 : }
1509 : }
1510 :
1511 : // Advance the sequence before spawning a reconciler, so that sequence waiters
1512 : // can distinguish between before+after the reconcile completes.
1513 : self.ensure_sequence_ahead();
1514 :
1515 : let reconciler_cancel = cancel.child_token();
1516 : let reconciler_intent = TargetState::from_intent(pageservers, &self.intent);
1517 : let reconciler = Reconciler {
1518 : tenant_shard_id: self.tenant_shard_id,
1519 : shard: self.shard,
1520 : placement_policy: self.policy.clone(),
1521 : generation: self.generation,
1522 : intent: reconciler_intent,
1523 : detach,
1524 : reconciler_config,
1525 : config: self.config.clone(),
1526 : preferred_az: self.intent.preferred_az_id.clone(),
1527 : observed: self.observed.clone(),
1528 : original_observed: self.observed.clone(),
1529 : compute_hook: compute_hook.clone(),
1530 : service_config: service_config.clone(),
1531 : _gate_guard: gate_guard,
1532 : _resource_units: units,
1533 : cancel: reconciler_cancel.clone(),
1534 : persistence: persistence.clone(),
1535 : compute_notify_failure: false,
1536 : };
1537 :
1538 : let reconcile_seq = self.sequence;
1539 : let long_reconcile_threshold = service_config.long_reconcile_threshold;
1540 :
1541 : tracing::info!(seq=%reconcile_seq, "Spawning Reconciler for sequence {}", self.sequence);
1542 : let must_notify = self.pending_compute_notification;
1543 : let reconciler_span = tracing::info_span!(parent: None, "reconciler", seq=%reconcile_seq,
1544 : tenant_id=%reconciler.tenant_shard_id.tenant_id,
1545 : shard_id=%reconciler.tenant_shard_id.shard_slug());
1546 : metrics::METRICS_REGISTRY
1547 : .metrics_group
1548 : .storage_controller_reconcile_spawn
1549 : .inc();
1550 : let result_tx = result_tx.clone();
1551 : let join_handle = tokio::task::spawn(
1552 0 : async move {
1553 : // Wait for any previous reconcile task to complete before we start
1554 0 : if let Some(old_handle) = old_handle {
1555 0 : old_handle.cancel.cancel();
1556 0 : if let Err(e) = old_handle.handle.await {
1557 : // We can't do much with this other than log it: the task is done, so
1558 : // we may proceed with our work.
1559 0 : tracing::error!("Unexpected join error waiting for reconcile task: {e}");
1560 0 : }
1561 0 : }
1562 :
1563 : // Early check for cancellation before doing any work
1564 : // TODO: wrap all remote API operations in cancellation check
1565 : // as well.
1566 0 : if reconciler.cancel.is_cancelled() {
1567 0 : metrics::METRICS_REGISTRY
1568 0 : .metrics_group
1569 0 : .storage_controller_reconcile_complete
1570 0 : .inc(ReconcileCompleteLabelGroup {
1571 0 : status: ReconcileOutcome::Cancel,
1572 0 : });
1573 0 : return;
1574 0 : }
1575 0 :
1576 0 : let (tenant_id_label, shard_number_label, sequence_label) = {
1577 0 : (
1578 0 : reconciler.tenant_shard_id.tenant_id.to_string(),
1579 0 : reconciler.tenant_shard_id.shard_number.0.to_string(),
1580 0 : reconcile_seq.to_string(),
1581 0 : )
1582 0 : };
1583 0 :
1584 0 : let label_group = ReconcileLongRunningLabelGroup {
1585 0 : tenant_id: &tenant_id_label,
1586 0 : shard_number: &shard_number_label,
1587 0 : sequence: &sequence_label,
1588 0 : };
1589 0 :
1590 0 : let reconcile_fut = Self::reconcile(reconcile_seq, reconciler, must_notify);
1591 0 : let long_reconcile_fut = {
1592 0 : let label_group = label_group.clone();
1593 0 : async move {
1594 0 : tokio::time::sleep(long_reconcile_threshold).await;
1595 :
1596 0 : tracing::warn!("Reconcile passed the long running threshold of {long_reconcile_threshold:?}");
1597 :
1598 0 : metrics::METRICS_REGISTRY
1599 0 : .metrics_group
1600 0 : .storage_controller_reconcile_long_running
1601 0 : .inc(label_group);
1602 0 : }
1603 : };
1604 :
1605 0 : let reconcile_fut = std::pin::pin!(reconcile_fut);
1606 0 : let long_reconcile_fut = std::pin::pin!(long_reconcile_fut);
1607 :
1608 0 : let (was_long, result) =
1609 0 : match future::select(reconcile_fut, long_reconcile_fut).await {
1610 0 : Either::Left((reconcile_result, _)) => (false, reconcile_result),
1611 0 : Either::Right((_, reconcile_fut)) => (true, reconcile_fut.await),
1612 : };
1613 :
1614 0 : if was_long {
1615 0 : let id = metrics::METRICS_REGISTRY
1616 0 : .metrics_group
1617 0 : .storage_controller_reconcile_long_running
1618 0 : .with_labels(label_group);
1619 0 : metrics::METRICS_REGISTRY
1620 0 : .metrics_group
1621 0 : .storage_controller_reconcile_long_running
1622 0 : .remove_metric(id);
1623 0 : }
1624 :
1625 0 : result_tx
1626 0 : .send(ReconcileResultRequest::ReconcileResult(result))
1627 0 : .ok();
1628 0 : }
1629 : .instrument(reconciler_span),
1630 : );
1631 :
1632 : self.reconciler = Some(ReconcilerHandle {
1633 : sequence: self.sequence,
1634 : handle: join_handle,
1635 : cancel: reconciler_cancel,
1636 : });
1637 :
1638 : Some(ReconcilerWaiter {
1639 : tenant_shard_id: self.tenant_shard_id,
1640 : seq_wait: self.waiter.clone(),
1641 : error_seq_wait: self.error_waiter.clone(),
1642 : error: self.last_error.clone(),
1643 : seq: self.sequence,
1644 : })
1645 : }
1646 :
1647 0 : pub(crate) fn cancel_reconciler(&self) {
1648 0 : if let Some(handle) = self.reconciler.as_ref() {
1649 0 : handle.cancel.cancel()
1650 0 : }
1651 0 : }
1652 :
1653 : /// Get a waiter for any reconciliation in flight, but do not start reconciliation
1654 : /// if it is not already running
1655 0 : pub(crate) fn get_waiter(&self) -> Option<ReconcilerWaiter> {
1656 0 : if self.reconciler.is_some() {
1657 0 : Some(ReconcilerWaiter {
1658 0 : tenant_shard_id: self.tenant_shard_id,
1659 0 : seq_wait: self.waiter.clone(),
1660 0 : error_seq_wait: self.error_waiter.clone(),
1661 0 : error: self.last_error.clone(),
1662 0 : seq: self.sequence,
1663 0 : })
1664 : } else {
1665 0 : None
1666 : }
1667 0 : }
1668 :
1669 : /// Called when a ReconcileResult has been emitted and the service is updating
1670 : /// our state: if the result is from a sequence >= my ReconcileHandle, then drop
1671 : /// the handle to indicate there is no longer a reconciliation in progress.
1672 0 : pub(crate) fn reconcile_complete(&mut self, sequence: Sequence) {
1673 0 : if let Some(reconcile_handle) = &self.reconciler {
1674 0 : if reconcile_handle.sequence <= sequence {
1675 0 : self.reconciler = None;
1676 0 : }
1677 0 : }
1678 0 : }
1679 :
1680 : /// If we had any state at all referring to this node ID, drop it. Does not
1681 : /// attempt to reschedule.
1682 : ///
1683 : /// Returns true if we modified the node's intent state.
1684 0 : pub(crate) fn deref_node(&mut self, node_id: NodeId) -> bool {
1685 0 : let mut intent_modified = false;
1686 0 :
1687 0 : // Drop if this node was our attached intent
1688 0 : if self.intent.attached == Some(node_id) {
1689 0 : self.intent.attached = None;
1690 0 : intent_modified = true;
1691 0 : }
1692 :
1693 : // Drop from the list of secondaries, and check if we modified it
1694 0 : let had_secondaries = self.intent.secondary.len();
1695 0 : self.intent.secondary.retain(|n| n != &node_id);
1696 0 : intent_modified |= self.intent.secondary.len() != had_secondaries;
1697 0 :
1698 0 : debug_assert!(!self.intent.all_pageservers().contains(&node_id));
1699 :
1700 0 : intent_modified
1701 0 : }
1702 :
1703 0 : pub(crate) fn set_scheduling_policy(&mut self, p: ShardSchedulingPolicy) {
1704 0 : self.scheduling_policy = p;
1705 0 : }
1706 :
1707 0 : pub(crate) fn get_scheduling_policy(&self) -> &ShardSchedulingPolicy {
1708 0 : &self.scheduling_policy
1709 0 : }
1710 :
1711 0 : pub(crate) fn set_last_error(&mut self, sequence: Sequence, error: ReconcileError) {
1712 0 : // Ordering: always set last_error before advancing sequence, so that sequence
1713 0 : // waiters are guaranteed to see a Some value when they see an error.
1714 0 : *(self.last_error.lock().unwrap()) = Some(Arc::new(error));
1715 0 : self.error_waiter.advance(sequence);
1716 0 : }
1717 :
1718 0 : pub(crate) fn from_persistent(
1719 0 : tsp: TenantShardPersistence,
1720 0 : intent: IntentState,
1721 0 : ) -> anyhow::Result<Self> {
1722 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1723 0 : let shard_identity = tsp.get_shard_identity()?;
1724 :
1725 0 : metrics::METRICS_REGISTRY
1726 0 : .metrics_group
1727 0 : .storage_controller_tenant_shards
1728 0 : .inc();
1729 0 :
1730 0 : Ok(Self {
1731 0 : tenant_shard_id,
1732 0 : shard: shard_identity,
1733 0 : sequence: Sequence::initial(),
1734 0 : generation: tsp.generation.map(|g| Generation::new(g as u32)),
1735 0 : policy: serde_json::from_str(&tsp.placement_policy).unwrap(),
1736 0 : intent,
1737 0 : observed: ObservedState::new(),
1738 0 : config: serde_json::from_str(&tsp.config).unwrap(),
1739 0 : reconciler: None,
1740 0 : splitting: tsp.splitting,
1741 0 : waiter: Arc::new(SeqWait::new(Sequence::initial())),
1742 0 : error_waiter: Arc::new(SeqWait::new(Sequence::initial())),
1743 0 : last_error: Arc::default(),
1744 0 : pending_compute_notification: false,
1745 0 : delayed_reconcile: false,
1746 0 : scheduling_policy: serde_json::from_str(&tsp.scheduling_policy).unwrap(),
1747 0 : })
1748 0 : }
1749 :
1750 0 : pub(crate) fn to_persistent(&self) -> TenantShardPersistence {
1751 0 : TenantShardPersistence {
1752 0 : tenant_id: self.tenant_shard_id.tenant_id.to_string(),
1753 0 : shard_number: self.tenant_shard_id.shard_number.0 as i32,
1754 0 : shard_count: self.tenant_shard_id.shard_count.literal() as i32,
1755 0 : shard_stripe_size: self.shard.stripe_size.0 as i32,
1756 0 : generation: self.generation.map(|g| g.into().unwrap_or(0) as i32),
1757 0 : generation_pageserver: self.intent.get_attached().map(|n| n.0 as i64),
1758 0 : placement_policy: serde_json::to_string(&self.policy).unwrap(),
1759 0 : config: serde_json::to_string(&self.config).unwrap(),
1760 0 : splitting: SplitState::default(),
1761 0 : scheduling_policy: serde_json::to_string(&self.scheduling_policy).unwrap(),
1762 0 : preferred_az_id: self.intent.preferred_az_id.as_ref().map(|az| az.0.clone()),
1763 0 : }
1764 0 : }
1765 :
1766 12508 : pub(crate) fn preferred_az(&self) -> Option<&AvailabilityZone> {
1767 12508 : self.intent.preferred_az_id.as_ref()
1768 12508 : }
1769 :
1770 0 : pub(crate) fn set_preferred_az(&mut self, preferred_az_id: Option<AvailabilityZone>) {
1771 0 : self.intent.preferred_az_id = preferred_az_id;
1772 0 : }
1773 :
1774 : /// Returns all the nodes to which this tenant shard is attached according to the
1775 : /// observed state and the generations. Return vector is sorted from latest generation
1776 : /// to earliest.
1777 0 : pub(crate) fn attached_locations(&self) -> Vec<(NodeId, Generation)> {
1778 0 : self.observed
1779 0 : .locations
1780 0 : .iter()
1781 0 : .filter_map(|(node_id, observed)| {
1782 : use LocationConfigMode::{AttachedMulti, AttachedSingle, AttachedStale};
1783 :
1784 0 : let conf = observed.conf.as_ref()?;
1785 :
1786 0 : match (conf.generation, conf.mode) {
1787 0 : (Some(gen), AttachedMulti | AttachedSingle | AttachedStale) => {
1788 0 : Some((*node_id, gen))
1789 : }
1790 0 : _ => None,
1791 : }
1792 0 : })
1793 0 : .sorted_by(|(_lhs_node_id, lhs_gen), (_rhs_node_id, rhs_gen)| {
1794 0 : lhs_gen.cmp(rhs_gen).reverse()
1795 0 : })
1796 0 : .map(|(node_id, gen)| (node_id, Generation::new(gen)))
1797 0 : .collect()
1798 0 : }
1799 :
1800 : /// Update the observed state of the tenant by applying incremental deltas
1801 : ///
1802 : /// Deltas are generated by reconcilers via [`Reconciler::observed_deltas`].
1803 : /// They are then filtered in [`crate::service::Service::process_result`].
1804 0 : pub(crate) fn apply_observed_deltas(
1805 0 : &mut self,
1806 0 : deltas: impl Iterator<Item = ObservedStateDelta>,
1807 0 : ) {
1808 0 : for delta in deltas {
1809 0 : match delta {
1810 0 : ObservedStateDelta::Upsert(ups) => {
1811 0 : let (node_id, loc) = *ups;
1812 0 :
1813 0 : // If the generation of the observed location in the delta is lagging
1814 0 : // behind the current one, then we have a race condition and cannot
1815 0 : // be certain about the true observed state. Set the observed state
1816 0 : // to None in order to reflect this.
1817 0 : let crnt_gen = self
1818 0 : .observed
1819 0 : .locations
1820 0 : .get(&node_id)
1821 0 : .and_then(|loc| loc.conf.as_ref())
1822 0 : .and_then(|conf| conf.generation);
1823 0 : let new_gen = loc.conf.as_ref().and_then(|conf| conf.generation);
1824 0 : match (crnt_gen, new_gen) {
1825 0 : (Some(crnt), Some(new)) if crnt_gen > new_gen => {
1826 0 : tracing::warn!(
1827 0 : "Skipping observed state update {}: {:?} and using None due to stale generation ({} > {})",
1828 : node_id, loc, crnt, new
1829 : );
1830 :
1831 0 : self.observed
1832 0 : .locations
1833 0 : .insert(node_id, ObservedStateLocation { conf: None });
1834 0 :
1835 0 : continue;
1836 : }
1837 0 : _ => {}
1838 : }
1839 :
1840 0 : if let Some(conf) = &loc.conf {
1841 0 : tracing::info!("Updating observed location {}: {:?}", node_id, conf);
1842 : } else {
1843 0 : tracing::info!("Setting observed location {} to None", node_id,)
1844 : }
1845 :
1846 0 : self.observed.locations.insert(node_id, loc);
1847 : }
1848 0 : ObservedStateDelta::Delete(node_id) => {
1849 0 : tracing::info!("Deleting observed location {}", node_id);
1850 0 : self.observed.locations.remove(&node_id);
1851 : }
1852 : }
1853 : }
1854 0 : }
1855 :
1856 : /// Returns true if the tenant shard is attached to a node that is outside the preferred AZ.
1857 : ///
1858 : /// If the shard does not have a preferred AZ, returns false.
1859 0 : pub(crate) fn is_attached_outside_preferred_az(&self, nodes: &HashMap<NodeId, Node>) -> bool {
1860 0 : self.intent
1861 0 : .get_attached()
1862 0 : .map(|node_id| {
1863 0 : Some(
1864 0 : nodes
1865 0 : .get(&node_id)
1866 0 : .expect("referenced node exists")
1867 0 : .get_availability_zone_id(),
1868 0 : ) != self.intent.preferred_az_id.as_ref()
1869 0 : })
1870 0 : .unwrap_or(false)
1871 0 : }
1872 : }
1873 :
1874 : impl Drop for TenantShard {
1875 12841 : fn drop(&mut self) {
1876 12841 : metrics::METRICS_REGISTRY
1877 12841 : .metrics_group
1878 12841 : .storage_controller_tenant_shards
1879 12841 : .dec();
1880 12841 : }
1881 : }
1882 :
1883 : #[cfg(test)]
1884 : pub(crate) mod tests {
1885 : use std::{cell::RefCell, rc::Rc};
1886 :
1887 : use pageserver_api::{
1888 : controller_api::NodeAvailability,
1889 : shard::{ShardCount, ShardNumber},
1890 : };
1891 : use rand::{rngs::StdRng, SeedableRng};
1892 : use utils::id::TenantId;
1893 :
1894 : use crate::scheduler::test_utils::make_test_nodes;
1895 :
1896 : use super::*;
1897 :
1898 11 : fn make_test_tenant_shard(policy: PlacementPolicy) -> TenantShard {
1899 11 : let tenant_id = TenantId::generate();
1900 11 : let shard_number = ShardNumber(0);
1901 11 : let shard_count = ShardCount::new(1);
1902 11 :
1903 11 : let tenant_shard_id = TenantShardId {
1904 11 : tenant_id,
1905 11 : shard_number,
1906 11 : shard_count,
1907 11 : };
1908 11 : TenantShard::new(
1909 11 : tenant_shard_id,
1910 11 : ShardIdentity::new(
1911 11 : shard_number,
1912 11 : shard_count,
1913 11 : pageserver_api::shard::ShardStripeSize(32768),
1914 11 : )
1915 11 : .unwrap(),
1916 11 : policy,
1917 11 : None,
1918 11 : )
1919 11 : }
1920 :
1921 5004 : pub(crate) fn make_test_tenant(
1922 5004 : policy: PlacementPolicy,
1923 5004 : shard_count: ShardCount,
1924 5004 : preferred_az: Option<AvailabilityZone>,
1925 5004 : ) -> Vec<TenantShard> {
1926 5004 : make_test_tenant_with_id(TenantId::generate(), policy, shard_count, preferred_az)
1927 5004 : }
1928 :
1929 5007 : pub(crate) fn make_test_tenant_with_id(
1930 5007 : tenant_id: TenantId,
1931 5007 : policy: PlacementPolicy,
1932 5007 : shard_count: ShardCount,
1933 5007 : preferred_az: Option<AvailabilityZone>,
1934 5007 : ) -> Vec<TenantShard> {
1935 5007 : (0..shard_count.count())
1936 12522 : .map(|i| {
1937 12522 : let shard_number = ShardNumber(i);
1938 12522 :
1939 12522 : let tenant_shard_id = TenantShardId {
1940 12522 : tenant_id,
1941 12522 : shard_number,
1942 12522 : shard_count,
1943 12522 : };
1944 12522 : TenantShard::new(
1945 12522 : tenant_shard_id,
1946 12522 : ShardIdentity::new(
1947 12522 : shard_number,
1948 12522 : shard_count,
1949 12522 : pageserver_api::shard::ShardStripeSize(32768),
1950 12522 : )
1951 12522 : .unwrap(),
1952 12522 : policy.clone(),
1953 12522 : preferred_az.clone(),
1954 12522 : )
1955 12522 : })
1956 5007 : .collect()
1957 5007 : }
1958 :
1959 : /// Test the scheduling behaviors used when a tenant configured for HA is subject
1960 : /// to nodes being marked offline.
1961 : #[test]
1962 1 : fn tenant_ha_scheduling() -> anyhow::Result<()> {
1963 1 : // Start with three nodes. Our tenant will only use two. The third one is
1964 1 : // expected to remain unused.
1965 1 : let mut nodes = make_test_nodes(3, &[]);
1966 1 :
1967 1 : let mut scheduler = Scheduler::new(nodes.values());
1968 1 : let mut context = ScheduleContext::default();
1969 1 :
1970 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
1971 1 : tenant_shard
1972 1 : .schedule(&mut scheduler, &mut context)
1973 1 : .expect("we have enough nodes, scheduling should work");
1974 1 :
1975 1 : // Expect to initially be schedule on to different nodes
1976 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
1977 1 : assert!(tenant_shard.intent.attached.is_some());
1978 :
1979 1 : let attached_node_id = tenant_shard.intent.attached.unwrap();
1980 1 : let secondary_node_id = *tenant_shard.intent.secondary.iter().last().unwrap();
1981 1 : assert_ne!(attached_node_id, secondary_node_id);
1982 :
1983 : // Notifying the attached node is offline should demote it to a secondary
1984 1 : let changed = tenant_shard
1985 1 : .intent
1986 1 : .demote_attached(&mut scheduler, attached_node_id);
1987 1 : assert!(changed);
1988 1 : assert!(tenant_shard.intent.attached.is_none());
1989 1 : assert_eq!(tenant_shard.intent.secondary.len(), 2);
1990 :
1991 : // Update the scheduler state to indicate the node is offline
1992 1 : nodes
1993 1 : .get_mut(&attached_node_id)
1994 1 : .unwrap()
1995 1 : .set_availability(NodeAvailability::Offline);
1996 1 : scheduler.node_upsert(nodes.get(&attached_node_id).unwrap());
1997 1 :
1998 1 : // Scheduling the node should promote the still-available secondary node to attached
1999 1 : tenant_shard
2000 1 : .schedule(&mut scheduler, &mut context)
2001 1 : .expect("active nodes are available");
2002 1 : assert_eq!(tenant_shard.intent.attached.unwrap(), secondary_node_id);
2003 :
2004 : // The original attached node should have been retained as a secondary
2005 1 : assert_eq!(
2006 1 : *tenant_shard.intent.secondary.iter().last().unwrap(),
2007 1 : attached_node_id
2008 1 : );
2009 :
2010 1 : tenant_shard.intent.clear(&mut scheduler);
2011 1 :
2012 1 : Ok(())
2013 1 : }
2014 :
2015 : #[test]
2016 1 : fn intent_from_observed() -> anyhow::Result<()> {
2017 1 : let nodes = make_test_nodes(3, &[]);
2018 1 : let mut scheduler = Scheduler::new(nodes.values());
2019 1 :
2020 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2021 1 :
2022 1 : tenant_shard.observed.locations.insert(
2023 1 : NodeId(3),
2024 1 : ObservedStateLocation {
2025 1 : conf: Some(LocationConfig {
2026 1 : mode: LocationConfigMode::AttachedMulti,
2027 1 : generation: Some(2),
2028 1 : secondary_conf: None,
2029 1 : shard_number: tenant_shard.shard.number.0,
2030 1 : shard_count: tenant_shard.shard.count.literal(),
2031 1 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
2032 1 : tenant_conf: TenantConfig::default(),
2033 1 : }),
2034 1 : },
2035 1 : );
2036 1 :
2037 1 : tenant_shard.observed.locations.insert(
2038 1 : NodeId(2),
2039 1 : ObservedStateLocation {
2040 1 : conf: Some(LocationConfig {
2041 1 : mode: LocationConfigMode::AttachedStale,
2042 1 : generation: Some(1),
2043 1 : secondary_conf: None,
2044 1 : shard_number: tenant_shard.shard.number.0,
2045 1 : shard_count: tenant_shard.shard.count.literal(),
2046 1 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
2047 1 : tenant_conf: TenantConfig::default(),
2048 1 : }),
2049 1 : },
2050 1 : );
2051 1 :
2052 1 : tenant_shard.intent_from_observed(&mut scheduler);
2053 1 :
2054 1 : // The highest generationed attached location gets used as attached
2055 1 : assert_eq!(tenant_shard.intent.attached, Some(NodeId(3)));
2056 : // Other locations get used as secondary
2057 1 : assert_eq!(tenant_shard.intent.secondary, vec![NodeId(2)]);
2058 :
2059 1 : scheduler.consistency_check(nodes.values(), [&tenant_shard].into_iter())?;
2060 :
2061 1 : tenant_shard.intent.clear(&mut scheduler);
2062 1 : Ok(())
2063 1 : }
2064 :
2065 : #[test]
2066 1 : fn scheduling_mode() -> anyhow::Result<()> {
2067 1 : let nodes = make_test_nodes(3, &[]);
2068 1 : let mut scheduler = Scheduler::new(nodes.values());
2069 1 :
2070 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
2071 1 :
2072 1 : // In pause mode, schedule() shouldn't do anything
2073 1 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Pause;
2074 1 : assert!(tenant_shard
2075 1 : .schedule(&mut scheduler, &mut ScheduleContext::default())
2076 1 : .is_ok());
2077 1 : assert!(tenant_shard.intent.all_pageservers().is_empty());
2078 :
2079 : // In active mode, schedule() works
2080 1 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Active;
2081 1 : assert!(tenant_shard
2082 1 : .schedule(&mut scheduler, &mut ScheduleContext::default())
2083 1 : .is_ok());
2084 1 : assert!(!tenant_shard.intent.all_pageservers().is_empty());
2085 :
2086 1 : tenant_shard.intent.clear(&mut scheduler);
2087 1 : Ok(())
2088 1 : }
2089 :
2090 : #[test]
2091 : /// Simple case: moving attachment to somewhere better where we already have a secondary
2092 1 : fn optimize_attachment_simple() -> anyhow::Result<()> {
2093 1 : let nodes = make_test_nodes(
2094 1 : 3,
2095 1 : &[
2096 1 : AvailabilityZone("az-a".to_string()),
2097 1 : AvailabilityZone("az-b".to_string()),
2098 1 : AvailabilityZone("az-c".to_string()),
2099 1 : ],
2100 1 : );
2101 1 : let mut scheduler = Scheduler::new(nodes.values());
2102 1 :
2103 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2104 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2105 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2106 1 : shard_b.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2107 1 :
2108 1 : // Initially: both nodes attached on shard 1, and both have secondary locations
2109 1 : // on different nodes.
2110 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2111 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(1));
2112 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2113 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(2));
2114 :
2115 1 : fn make_schedule_context(shard_a: &TenantShard, shard_b: &TenantShard) -> ScheduleContext {
2116 1 : let mut schedule_context = ScheduleContext::default();
2117 1 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2118 1 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2119 1 : schedule_context
2120 1 : }
2121 :
2122 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2123 1 : let optimization_a = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2124 1 : assert_eq!(
2125 1 : optimization_a,
2126 1 : Some(ScheduleOptimization {
2127 1 : sequence: shard_a.sequence,
2128 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2129 1 : old_attached_node_id: NodeId(2),
2130 1 : new_attached_node_id: NodeId(1)
2131 1 : })
2132 1 : })
2133 1 : );
2134 1 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
2135 1 :
2136 1 : // // Either shard should recognize that it has the option to switch to a secondary location where there
2137 1 : // // would be no other shards from the same tenant, and request to do so.
2138 1 : // assert_eq!(
2139 1 : // optimization_a_prepare,
2140 1 : // Some(ScheduleOptimization {
2141 1 : // sequence: shard_a.sequence,
2142 1 : // action: ScheduleOptimizationAction::CreateSecondary(NodeId(2))
2143 1 : // })
2144 1 : // );
2145 1 : // shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2146 1 :
2147 1 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2148 1 : // let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2149 1 : // assert_eq!(
2150 1 : // optimization_a_migrate,
2151 1 : // Some(ScheduleOptimization {
2152 1 : // sequence: shard_a.sequence,
2153 1 : // action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2154 1 : // old_attached_node_id: NodeId(1),
2155 1 : // new_attached_node_id: NodeId(2)
2156 1 : // })
2157 1 : // })
2158 1 : // );
2159 1 : // shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2160 1 :
2161 1 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2162 1 : // let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2163 1 : // assert_eq!(
2164 1 : // optimization_a_cleanup,
2165 1 : // Some(ScheduleOptimization {
2166 1 : // sequence: shard_a.sequence,
2167 1 : // action: ScheduleOptimizationAction::RemoveSecondary(NodeId(1))
2168 1 : // })
2169 1 : // );
2170 1 : // shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2171 1 :
2172 1 : // // Shard B should not be moved anywhere, since the pressure on node 1 was relieved by moving shard A
2173 1 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2174 1 : // assert_eq!(shard_b.optimize_attachment(&mut scheduler, &schedule_context), None);
2175 1 :
2176 1 : shard_a.intent.clear(&mut scheduler);
2177 1 : shard_b.intent.clear(&mut scheduler);
2178 1 :
2179 1 : Ok(())
2180 1 : }
2181 :
2182 : #[test]
2183 : /// Complicated case: moving attachment to somewhere better where we do not have a secondary
2184 : /// already, creating one as needed.
2185 1 : fn optimize_attachment_multistep() -> anyhow::Result<()> {
2186 1 : let nodes = make_test_nodes(
2187 1 : 3,
2188 1 : &[
2189 1 : AvailabilityZone("az-a".to_string()),
2190 1 : AvailabilityZone("az-b".to_string()),
2191 1 : AvailabilityZone("az-c".to_string()),
2192 1 : ],
2193 1 : );
2194 1 : let mut scheduler = Scheduler::new(nodes.values());
2195 1 :
2196 1 : // Two shards of a tenant that wants to be in AZ A
2197 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2198 1 : shard_a.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2199 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2200 1 : shard_b.intent.preferred_az_id = Some(AvailabilityZone("az-a".to_string()));
2201 1 :
2202 1 : // Both shards are initially attached in non-home AZ _and_ have secondaries in non-home AZs
2203 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2204 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2205 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(3)));
2206 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(2));
2207 :
2208 3 : fn make_schedule_context(shard_a: &TenantShard, shard_b: &TenantShard) -> ScheduleContext {
2209 3 : let mut schedule_context = ScheduleContext::default();
2210 3 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2211 3 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2212 3 : schedule_context
2213 3 : }
2214 :
2215 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2216 1 : let optimization_a_prepare = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2217 1 : assert_eq!(
2218 1 : optimization_a_prepare,
2219 1 : Some(ScheduleOptimization {
2220 1 : sequence: shard_a.sequence,
2221 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(1))
2222 1 : })
2223 1 : );
2224 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2225 1 :
2226 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2227 1 : let optimization_a_migrate = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2228 1 : assert_eq!(
2229 1 : optimization_a_migrate,
2230 1 : Some(ScheduleOptimization {
2231 1 : sequence: shard_a.sequence,
2232 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2233 1 : old_attached_node_id: NodeId(2),
2234 1 : new_attached_node_id: NodeId(1)
2235 1 : })
2236 1 : })
2237 1 : );
2238 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2239 1 :
2240 1 : let schedule_context = make_schedule_context(&shard_a, &shard_b);
2241 1 : let optimization_a_cleanup = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2242 1 : assert_eq!(
2243 1 : optimization_a_cleanup,
2244 1 : Some(ScheduleOptimization {
2245 1 : sequence: shard_a.sequence,
2246 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(3))
2247 1 : })
2248 1 : );
2249 1 : shard_a.apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2250 1 :
2251 1 : // // Shard B should not be moved anywhere, since the pressure on node 1 was relieved by moving shard A
2252 1 : // let schedule_context = make_schedule_context(&shard_a, &shard_b);
2253 1 : // assert_eq!(shard_b.optimize_attachment(&mut scheduler, &schedule_context), None);
2254 1 :
2255 1 : shard_a.intent.clear(&mut scheduler);
2256 1 : shard_b.intent.clear(&mut scheduler);
2257 1 :
2258 1 : Ok(())
2259 1 : }
2260 :
2261 : #[test]
2262 : /// Check that multi-step migration works when moving to somewhere that is only better by
2263 : /// 1 AffinityScore -- this ensures that we don't have a bug like the intermediate secondary
2264 : /// counting toward the affinity score such that it prevents the rest of the migration from happening.
2265 1 : fn optimize_attachment_marginal() -> anyhow::Result<()> {
2266 1 : let nodes = make_test_nodes(2, &[]);
2267 1 : let mut scheduler = Scheduler::new(nodes.values());
2268 1 :
2269 1 : // Multi-sharded tenant, we will craft a situation where affinity
2270 1 : // scores differ only slightly
2271 1 : let mut shards = make_test_tenant(PlacementPolicy::Attached(0), ShardCount::new(4), None);
2272 1 :
2273 1 : // 1 attached on node 1
2274 1 : shards[0]
2275 1 : .intent
2276 1 : .set_attached(&mut scheduler, Some(NodeId(1)));
2277 1 : // 3 attached on node 2
2278 1 : shards[1]
2279 1 : .intent
2280 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2281 1 : shards[2]
2282 1 : .intent
2283 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2284 1 : shards[3]
2285 1 : .intent
2286 1 : .set_attached(&mut scheduler, Some(NodeId(2)));
2287 :
2288 : // The scheduler should figure out that we need to:
2289 : // - Create a secondary for shard 3 on node 1
2290 : // - Migrate shard 3 to node 1
2291 : // - Remove shard 3's location on node 2
2292 :
2293 4 : fn make_schedule_context(shards: &Vec<TenantShard>) -> ScheduleContext {
2294 4 : let mut schedule_context = ScheduleContext::default();
2295 20 : for shard in shards {
2296 16 : schedule_context.avoid(&shard.intent.all_pageservers());
2297 16 : }
2298 4 : schedule_context
2299 4 : }
2300 :
2301 1 : let schedule_context = make_schedule_context(&shards);
2302 1 : let optimization_a_prepare =
2303 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2304 1 : assert_eq!(
2305 1 : optimization_a_prepare,
2306 1 : Some(ScheduleOptimization {
2307 1 : sequence: shards[1].sequence,
2308 1 : action: ScheduleOptimizationAction::CreateSecondary(NodeId(1))
2309 1 : })
2310 1 : );
2311 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_prepare.unwrap());
2312 1 :
2313 1 : let schedule_context = make_schedule_context(&shards);
2314 1 : let optimization_a_migrate =
2315 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2316 1 : assert_eq!(
2317 1 : optimization_a_migrate,
2318 1 : Some(ScheduleOptimization {
2319 1 : sequence: shards[1].sequence,
2320 1 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
2321 1 : old_attached_node_id: NodeId(2),
2322 1 : new_attached_node_id: NodeId(1)
2323 1 : })
2324 1 : })
2325 1 : );
2326 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_migrate.unwrap());
2327 1 :
2328 1 : let schedule_context = make_schedule_context(&shards);
2329 1 : let optimization_a_cleanup =
2330 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context);
2331 1 : assert_eq!(
2332 1 : optimization_a_cleanup,
2333 1 : Some(ScheduleOptimization {
2334 1 : sequence: shards[1].sequence,
2335 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(2))
2336 1 : })
2337 1 : );
2338 1 : shards[1].apply_optimization(&mut scheduler, optimization_a_cleanup.unwrap());
2339 1 :
2340 1 : // Everything should be stable now
2341 1 : let schedule_context = make_schedule_context(&shards);
2342 1 : assert_eq!(
2343 1 : shards[0].optimize_attachment(&mut scheduler, &schedule_context),
2344 1 : None
2345 1 : );
2346 1 : assert_eq!(
2347 1 : shards[1].optimize_attachment(&mut scheduler, &schedule_context),
2348 1 : None
2349 1 : );
2350 1 : assert_eq!(
2351 1 : shards[2].optimize_attachment(&mut scheduler, &schedule_context),
2352 1 : None
2353 1 : );
2354 1 : assert_eq!(
2355 1 : shards[3].optimize_attachment(&mut scheduler, &schedule_context),
2356 1 : None
2357 1 : );
2358 :
2359 5 : for mut shard in shards {
2360 4 : shard.intent.clear(&mut scheduler);
2361 4 : }
2362 :
2363 1 : Ok(())
2364 1 : }
2365 :
2366 : #[test]
2367 1 : fn optimize_secondary() -> anyhow::Result<()> {
2368 1 : let nodes = make_test_nodes(4, &[]);
2369 1 : let mut scheduler = Scheduler::new(nodes.values());
2370 1 :
2371 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2372 1 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
2373 1 :
2374 1 : // Initially: both nodes attached on shard 1, and both have secondary locations
2375 1 : // on different nodes.
2376 1 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
2377 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2378 1 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(2)));
2379 1 : shard_b.intent.push_secondary(&mut scheduler, NodeId(3));
2380 1 :
2381 1 : let mut schedule_context = ScheduleContext::default();
2382 1 : schedule_context.avoid(&shard_a.intent.all_pageservers());
2383 1 : schedule_context.avoid(&shard_b.intent.all_pageservers());
2384 1 :
2385 1 : let optimization_a = shard_a.optimize_secondary(&mut scheduler, &schedule_context);
2386 1 :
2387 1 : // Since there is a node with no locations available, the node with two locations for the
2388 1 : // same tenant should generate an optimization to move one away
2389 1 : assert_eq!(
2390 1 : optimization_a,
2391 1 : Some(ScheduleOptimization {
2392 1 : sequence: shard_a.sequence,
2393 1 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
2394 1 : old_node_id: NodeId(3),
2395 1 : new_node_id: NodeId(4)
2396 1 : })
2397 1 : })
2398 1 : );
2399 :
2400 1 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
2401 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2402 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(4)]);
2403 :
2404 1 : shard_a.intent.clear(&mut scheduler);
2405 1 : shard_b.intent.clear(&mut scheduler);
2406 1 :
2407 1 : Ok(())
2408 1 : }
2409 :
2410 : /// Test how the optimisation code behaves with an extra secondary
2411 : #[test]
2412 1 : fn optimize_removes_secondary() -> anyhow::Result<()> {
2413 1 : let az_a_tag = AvailabilityZone("az-a".to_string());
2414 1 : let az_b_tag = AvailabilityZone("az-b".to_string());
2415 1 : let mut nodes = make_test_nodes(
2416 1 : 4,
2417 1 : &[
2418 1 : az_a_tag.clone(),
2419 1 : az_b_tag.clone(),
2420 1 : az_a_tag.clone(),
2421 1 : az_b_tag.clone(),
2422 1 : ],
2423 1 : );
2424 1 : let mut scheduler = Scheduler::new(nodes.values());
2425 1 :
2426 1 : let mut schedule_context = ScheduleContext::default();
2427 1 :
2428 1 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
2429 1 : shard_a.intent.preferred_az_id = Some(az_a_tag.clone());
2430 1 : shard_a
2431 1 : .schedule(&mut scheduler, &mut schedule_context)
2432 1 : .unwrap();
2433 1 :
2434 1 : // Attached on node 1, secondary on node 2
2435 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2436 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(2)]);
2437 :
2438 : // Initially optimiser is idle
2439 1 : assert_eq!(
2440 1 : shard_a.optimize_attachment(&mut scheduler, &schedule_context),
2441 1 : None
2442 1 : );
2443 1 : assert_eq!(
2444 1 : shard_a.optimize_secondary(&mut scheduler, &schedule_context),
2445 1 : None
2446 1 : );
2447 :
2448 : // 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
2449 : // to our new location
2450 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
2451 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2452 1 : assert_eq!(
2453 1 : optimization,
2454 1 : Some(ScheduleOptimization {
2455 1 : sequence: shard_a.sequence,
2456 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(3))
2457 1 : })
2458 1 : );
2459 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2460 1 :
2461 1 : // A spare secondary in the non-home AZ, and one of them is offline
2462 1 : shard_a.intent.push_secondary(&mut scheduler, NodeId(4));
2463 1 : nodes
2464 1 : .get_mut(&NodeId(4))
2465 1 : .unwrap()
2466 1 : .set_availability(NodeAvailability::Offline);
2467 1 : scheduler.node_upsert(nodes.get(&NodeId(4)).unwrap());
2468 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2469 1 : assert_eq!(
2470 1 : optimization,
2471 1 : Some(ScheduleOptimization {
2472 1 : sequence: shard_a.sequence,
2473 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(4))
2474 1 : })
2475 1 : );
2476 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2477 1 :
2478 1 : // A spare secondary when should have none
2479 1 : shard_a.policy = PlacementPolicy::Attached(0);
2480 1 : let optimization = shard_a.optimize_attachment(&mut scheduler, &schedule_context);
2481 1 : assert_eq!(
2482 1 : optimization,
2483 1 : Some(ScheduleOptimization {
2484 1 : sequence: shard_a.sequence,
2485 1 : action: ScheduleOptimizationAction::RemoveSecondary(NodeId(2))
2486 1 : })
2487 1 : );
2488 1 : shard_a.apply_optimization(&mut scheduler, optimization.unwrap());
2489 1 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
2490 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![]);
2491 :
2492 : // Check that in secondary mode, we preserve the secondary in the preferred AZ
2493 1 : let mut schedule_context = ScheduleContext::default(); // Fresh context, we're about to call schedule()
2494 1 : shard_a.policy = PlacementPolicy::Secondary;
2495 1 : shard_a
2496 1 : .schedule(&mut scheduler, &mut schedule_context)
2497 1 : .unwrap();
2498 1 : assert_eq!(shard_a.intent.get_attached(), &None);
2499 1 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(1)]);
2500 1 : assert_eq!(
2501 1 : shard_a.optimize_attachment(&mut scheduler, &schedule_context),
2502 1 : None
2503 1 : );
2504 1 : assert_eq!(
2505 1 : shard_a.optimize_secondary(&mut scheduler, &schedule_context),
2506 1 : None
2507 1 : );
2508 :
2509 1 : shard_a.intent.clear(&mut scheduler);
2510 1 :
2511 1 : Ok(())
2512 1 : }
2513 :
2514 : // Optimize til quiescent: this emulates what Service::optimize_all does, when
2515 : // called repeatedly in the background.
2516 : // Returns the applied optimizations
2517 3 : fn optimize_til_idle(
2518 3 : scheduler: &mut Scheduler,
2519 3 : shards: &mut [TenantShard],
2520 3 : ) -> Vec<ScheduleOptimization> {
2521 3 : let mut loop_n = 0;
2522 3 : let mut optimizations = Vec::default();
2523 : loop {
2524 6 : let mut schedule_context = ScheduleContext::default();
2525 6 : let mut any_changed = false;
2526 :
2527 24 : for shard in shards.iter() {
2528 24 : schedule_context.avoid(&shard.intent.all_pageservers());
2529 24 : }
2530 :
2531 15 : for shard in shards.iter_mut() {
2532 15 : let optimization = shard.optimize_attachment(scheduler, &schedule_context);
2533 15 : tracing::info!(
2534 0 : "optimize_attachment({})={:?}",
2535 : shard.tenant_shard_id,
2536 : optimization
2537 : );
2538 15 : if let Some(optimization) = optimization {
2539 : // Check that maybe_optimizable wouldn't have wrongly claimed this optimization didn't exist
2540 3 : assert!(shard.maybe_optimizable(scheduler, &schedule_context));
2541 3 : optimizations.push(optimization.clone());
2542 3 : shard.apply_optimization(scheduler, optimization);
2543 3 : any_changed = true;
2544 3 : break;
2545 12 : }
2546 12 :
2547 12 : let optimization = shard.optimize_secondary(scheduler, &schedule_context);
2548 12 : tracing::info!(
2549 0 : "optimize_secondary({})={:?}",
2550 : shard.tenant_shard_id,
2551 : optimization
2552 : );
2553 12 : if let Some(optimization) = optimization {
2554 : // Check that maybe_optimizable wouldn't have wrongly claimed this optimization didn't exist
2555 0 : assert!(shard.maybe_optimizable(scheduler, &schedule_context));
2556 :
2557 0 : optimizations.push(optimization.clone());
2558 0 : shard.apply_optimization(scheduler, optimization);
2559 0 : any_changed = true;
2560 0 : break;
2561 12 : }
2562 : }
2563 :
2564 6 : if !any_changed {
2565 3 : break;
2566 3 : }
2567 3 :
2568 3 : // Assert no infinite loop
2569 3 : loop_n += 1;
2570 3 : assert!(loop_n < 1000);
2571 : }
2572 :
2573 3 : optimizations
2574 3 : }
2575 :
2576 : /// Test the balancing behavior of shard scheduling: that it achieves a balance, and
2577 : /// that it converges.
2578 : #[test]
2579 1 : fn optimize_add_nodes() -> anyhow::Result<()> {
2580 1 : let nodes = make_test_nodes(
2581 1 : 9,
2582 1 : &[
2583 1 : // Initial 6 nodes
2584 1 : AvailabilityZone("az-a".to_string()),
2585 1 : AvailabilityZone("az-a".to_string()),
2586 1 : AvailabilityZone("az-b".to_string()),
2587 1 : AvailabilityZone("az-b".to_string()),
2588 1 : AvailabilityZone("az-c".to_string()),
2589 1 : AvailabilityZone("az-c".to_string()),
2590 1 : // Three we will add later
2591 1 : AvailabilityZone("az-a".to_string()),
2592 1 : AvailabilityZone("az-b".to_string()),
2593 1 : AvailabilityZone("az-c".to_string()),
2594 1 : ],
2595 1 : );
2596 1 :
2597 1 : // Only show the scheduler two nodes in each AZ to start with
2598 1 : let mut scheduler = Scheduler::new([].iter());
2599 7 : for i in 1..=6 {
2600 6 : scheduler.node_upsert(nodes.get(&NodeId(i)).unwrap());
2601 6 : }
2602 :
2603 1 : let mut shards = make_test_tenant(
2604 1 : PlacementPolicy::Attached(1),
2605 1 : ShardCount::new(4),
2606 1 : Some(AvailabilityZone("az-a".to_string())),
2607 1 : );
2608 1 : let mut schedule_context = ScheduleContext::default();
2609 5 : for shard in &mut shards {
2610 4 : assert!(shard
2611 4 : .schedule(&mut scheduler, &mut schedule_context)
2612 4 : .is_ok());
2613 : }
2614 :
2615 : // Initial: attached locations land in the tenant's home AZ.
2616 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 2);
2617 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 2);
2618 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
2619 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
2620 :
2621 : // Initial: secondary locations in a remote AZ
2622 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 1);
2623 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 0);
2624 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 1);
2625 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 0);
2626 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(5)), 1);
2627 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(5)), 0);
2628 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(6)), 1);
2629 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(6)), 0);
2630 :
2631 : // Add another three nodes: we should see the shards spread out when their optimize
2632 : // methods are called
2633 1 : scheduler.node_upsert(nodes.get(&NodeId(7)).unwrap());
2634 1 : scheduler.node_upsert(nodes.get(&NodeId(8)).unwrap());
2635 1 : scheduler.node_upsert(nodes.get(&NodeId(9)).unwrap());
2636 1 : optimize_til_idle(&mut scheduler, &mut shards);
2637 1 :
2638 1 : // We expect one attached location was moved to the new node in the tenant's home AZ
2639 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(7)), 1);
2640 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(7)), 1);
2641 : // The original node has one less attached shard
2642 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 1);
2643 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 1);
2644 :
2645 : // One of the original nodes still has two attachments, since there are an odd number of nodes
2646 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
2647 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
2648 :
2649 : // None of our secondaries moved, since we already had enough nodes for those to be
2650 : // scheduled perfectly
2651 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 1);
2652 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 0);
2653 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 1);
2654 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 0);
2655 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(5)), 1);
2656 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(5)), 0);
2657 1 : assert_eq!(scheduler.get_node_shard_count(NodeId(6)), 1);
2658 1 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(6)), 0);
2659 :
2660 4 : for shard in shards.iter_mut() {
2661 4 : shard.intent.clear(&mut scheduler);
2662 4 : }
2663 :
2664 1 : Ok(())
2665 1 : }
2666 :
2667 : /// Test that initial shard scheduling is optimal. By optimal we mean
2668 : /// that the optimizer cannot find a way to improve it.
2669 : ///
2670 : /// This test is an example of the scheduling issue described in
2671 : /// https://github.com/neondatabase/neon/issues/8969
2672 : #[test]
2673 1 : fn initial_scheduling_is_optimal() -> anyhow::Result<()> {
2674 : use itertools::Itertools;
2675 :
2676 1 : let nodes = make_test_nodes(2, &[]);
2677 1 :
2678 1 : let mut scheduler = Scheduler::new([].iter());
2679 1 : scheduler.node_upsert(nodes.get(&NodeId(1)).unwrap());
2680 1 : scheduler.node_upsert(nodes.get(&NodeId(2)).unwrap());
2681 1 :
2682 1 : let mut a = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4), None);
2683 1 : let a_context = Rc::new(RefCell::new(ScheduleContext::default()));
2684 1 :
2685 1 : let mut b = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4), None);
2686 1 : let b_context = Rc::new(RefCell::new(ScheduleContext::default()));
2687 1 :
2688 4 : let a_shards_with_context = a.iter_mut().map(|shard| (shard, a_context.clone()));
2689 4 : let b_shards_with_context = b.iter_mut().map(|shard| (shard, b_context.clone()));
2690 1 :
2691 1 : let schedule_order = a_shards_with_context.interleave(b_shards_with_context);
2692 :
2693 9 : for (shard, context) in schedule_order {
2694 8 : let context = &mut *context.borrow_mut();
2695 8 : shard.schedule(&mut scheduler, context).unwrap();
2696 8 : }
2697 :
2698 1 : let applied_to_a = optimize_til_idle(&mut scheduler, &mut a);
2699 1 : assert_eq!(applied_to_a, vec![]);
2700 :
2701 1 : let applied_to_b = optimize_til_idle(&mut scheduler, &mut b);
2702 1 : assert_eq!(applied_to_b, vec![]);
2703 :
2704 8 : for shard in a.iter_mut().chain(b.iter_mut()) {
2705 8 : shard.intent.clear(&mut scheduler);
2706 8 : }
2707 :
2708 1 : Ok(())
2709 1 : }
2710 :
2711 : #[test]
2712 1 : fn random_az_shard_scheduling() -> anyhow::Result<()> {
2713 : use rand::seq::SliceRandom;
2714 :
2715 51 : for seed in 0..50 {
2716 50 : eprintln!("Running test with seed {seed}");
2717 50 : let mut rng = StdRng::seed_from_u64(seed);
2718 50 :
2719 50 : let az_a_tag = AvailabilityZone("az-a".to_string());
2720 50 : let az_b_tag = AvailabilityZone("az-b".to_string());
2721 50 : let azs = [az_a_tag, az_b_tag];
2722 50 : let nodes = make_test_nodes(4, &azs);
2723 50 : let mut shards_per_az: HashMap<AvailabilityZone, u32> = HashMap::new();
2724 50 :
2725 50 : let mut scheduler = Scheduler::new([].iter());
2726 200 : for node in nodes.values() {
2727 200 : scheduler.node_upsert(node);
2728 200 : }
2729 :
2730 50 : let mut shards = Vec::default();
2731 50 : let mut contexts = Vec::default();
2732 50 : let mut az_picker = azs.iter().cycle().cloned();
2733 5050 : for i in 0..100 {
2734 5000 : let az = az_picker.next().unwrap();
2735 5000 : let shard_count = i % 4 + 1;
2736 5000 : *shards_per_az.entry(az.clone()).or_default() += shard_count;
2737 5000 :
2738 5000 : let tenant_shards = make_test_tenant(
2739 5000 : PlacementPolicy::Attached(1),
2740 5000 : ShardCount::new(shard_count.try_into().unwrap()),
2741 5000 : Some(az),
2742 5000 : );
2743 5000 : let context = Rc::new(RefCell::new(ScheduleContext::default()));
2744 5000 :
2745 5000 : contexts.push(context.clone());
2746 5000 : let with_ctx = tenant_shards
2747 5000 : .into_iter()
2748 12500 : .map(|shard| (shard, context.clone()));
2749 17500 : for shard_with_ctx in with_ctx {
2750 12500 : shards.push(shard_with_ctx);
2751 12500 : }
2752 : }
2753 :
2754 50 : shards.shuffle(&mut rng);
2755 :
2756 : #[derive(Default, Debug)]
2757 : struct NodeStats {
2758 : attachments: u32,
2759 : secondaries: u32,
2760 : }
2761 :
2762 50 : let mut node_stats: HashMap<NodeId, NodeStats> = HashMap::default();
2763 50 : let mut attachments_in_wrong_az = 0;
2764 50 : let mut secondaries_in_wrong_az = 0;
2765 :
2766 12550 : for (shard, context) in &mut shards {
2767 12500 : let context = &mut *context.borrow_mut();
2768 12500 : shard.schedule(&mut scheduler, context).unwrap();
2769 12500 :
2770 12500 : let attached_node = shard.intent.get_attached().unwrap();
2771 12500 : let stats = node_stats.entry(attached_node).or_default();
2772 12500 : stats.attachments += 1;
2773 12500 :
2774 12500 : let secondary_node = *shard.intent.get_secondary().first().unwrap();
2775 12500 : let stats = node_stats.entry(secondary_node).or_default();
2776 12500 : stats.secondaries += 1;
2777 12500 :
2778 12500 : let attached_node_az = nodes
2779 12500 : .get(&attached_node)
2780 12500 : .unwrap()
2781 12500 : .get_availability_zone_id();
2782 12500 : let secondary_node_az = nodes
2783 12500 : .get(&secondary_node)
2784 12500 : .unwrap()
2785 12500 : .get_availability_zone_id();
2786 12500 : let preferred_az = shard.preferred_az().unwrap();
2787 12500 :
2788 12500 : if attached_node_az != preferred_az {
2789 0 : eprintln!(
2790 0 : "{} attachment was scheduled in AZ {} but preferred AZ {}",
2791 0 : shard.tenant_shard_id, attached_node_az, preferred_az
2792 0 : );
2793 0 : attachments_in_wrong_az += 1;
2794 12500 : }
2795 :
2796 12500 : if secondary_node_az == preferred_az {
2797 0 : eprintln!(
2798 0 : "{} secondary was scheduled in AZ {} which matches preference",
2799 0 : shard.tenant_shard_id, attached_node_az
2800 0 : );
2801 0 : secondaries_in_wrong_az += 1;
2802 12500 : }
2803 : }
2804 :
2805 50 : let mut violations = Vec::default();
2806 50 :
2807 50 : if attachments_in_wrong_az > 0 {
2808 0 : violations.push(format!(
2809 0 : "{} attachments scheduled to the incorrect AZ",
2810 0 : attachments_in_wrong_az
2811 0 : ));
2812 50 : }
2813 :
2814 50 : if secondaries_in_wrong_az > 0 {
2815 0 : violations.push(format!(
2816 0 : "{} secondaries scheduled to the incorrect AZ",
2817 0 : secondaries_in_wrong_az
2818 0 : ));
2819 50 : }
2820 :
2821 50 : eprintln!(
2822 50 : "attachments_in_wrong_az={} secondaries_in_wrong_az={}",
2823 50 : attachments_in_wrong_az, secondaries_in_wrong_az
2824 50 : );
2825 :
2826 250 : for (node_id, stats) in &node_stats {
2827 200 : let node_az = nodes.get(node_id).unwrap().get_availability_zone_id();
2828 200 : let ideal_attachment_load = shards_per_az.get(node_az).unwrap() / 2;
2829 200 : let allowed_attachment_load =
2830 200 : (ideal_attachment_load - 1)..(ideal_attachment_load + 2);
2831 200 :
2832 200 : if !allowed_attachment_load.contains(&stats.attachments) {
2833 0 : violations.push(format!(
2834 0 : "Found {} attachments on node {}, but expected {}",
2835 0 : stats.attachments, node_id, ideal_attachment_load
2836 0 : ));
2837 200 : }
2838 :
2839 200 : eprintln!(
2840 200 : "{}: attachments={} secondaries={} ideal_attachment_load={}",
2841 200 : node_id, stats.attachments, stats.secondaries, ideal_attachment_load
2842 200 : );
2843 : }
2844 :
2845 50 : assert!(violations.is_empty(), "{violations:?}");
2846 :
2847 12550 : for (mut shard, _ctx) in shards {
2848 12500 : shard.intent.clear(&mut scheduler);
2849 12500 : }
2850 : }
2851 1 : Ok(())
2852 1 : }
2853 :
2854 : /// Check how the shard's scheduling behaves when in PlacementPolicy::Secondary mode.
2855 : #[test]
2856 1 : fn tenant_secondary_scheduling() -> anyhow::Result<()> {
2857 1 : let az_a = AvailabilityZone("az-a".to_string());
2858 1 : let nodes = make_test_nodes(
2859 1 : 3,
2860 1 : &[
2861 1 : az_a.clone(),
2862 1 : AvailabilityZone("az-b".to_string()),
2863 1 : AvailabilityZone("az-c".to_string()),
2864 1 : ],
2865 1 : );
2866 1 :
2867 1 : let mut scheduler = Scheduler::new(nodes.values());
2868 1 : let mut context = ScheduleContext::default();
2869 1 :
2870 1 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Secondary);
2871 1 : tenant_shard.intent.preferred_az_id = Some(az_a.clone());
2872 1 : tenant_shard
2873 1 : .schedule(&mut scheduler, &mut context)
2874 1 : .expect("we have enough nodes, scheduling should work");
2875 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
2876 1 : assert!(tenant_shard.intent.attached.is_none());
2877 :
2878 : // Should have scheduled into the preferred AZ
2879 1 : assert_eq!(
2880 1 : scheduler
2881 1 : .get_node_az(&tenant_shard.intent.secondary[0])
2882 1 : .as_ref(),
2883 1 : tenant_shard.preferred_az()
2884 1 : );
2885 :
2886 : // Optimizer should agree
2887 1 : assert_eq!(
2888 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
2889 1 : None
2890 1 : );
2891 1 : assert_eq!(
2892 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
2893 1 : None
2894 1 : );
2895 :
2896 : // Switch to PlacementPolicy::Attached
2897 1 : tenant_shard.policy = PlacementPolicy::Attached(1);
2898 1 : tenant_shard
2899 1 : .schedule(&mut scheduler, &mut context)
2900 1 : .expect("we have enough nodes, scheduling should work");
2901 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
2902 1 : assert!(tenant_shard.intent.attached.is_some());
2903 : // Secondary should now be in non-preferred AZ
2904 1 : assert_ne!(
2905 1 : scheduler
2906 1 : .get_node_az(&tenant_shard.intent.secondary[0])
2907 1 : .as_ref(),
2908 1 : tenant_shard.preferred_az()
2909 1 : );
2910 : // Attached should be in preferred AZ
2911 1 : assert_eq!(
2912 1 : scheduler
2913 1 : .get_node_az(&tenant_shard.intent.attached.unwrap())
2914 1 : .as_ref(),
2915 1 : tenant_shard.preferred_az()
2916 1 : );
2917 :
2918 : // Optimizer should agree
2919 1 : assert_eq!(
2920 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
2921 1 : None
2922 1 : );
2923 1 : assert_eq!(
2924 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
2925 1 : None
2926 1 : );
2927 :
2928 : // Switch back to PlacementPolicy::Secondary
2929 1 : tenant_shard.policy = PlacementPolicy::Secondary;
2930 1 : tenant_shard
2931 1 : .schedule(&mut scheduler, &mut context)
2932 1 : .expect("we have enough nodes, scheduling should work");
2933 1 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
2934 1 : assert!(tenant_shard.intent.attached.is_none());
2935 : // When we picked a location to keep, we should have kept the one in the preferred AZ
2936 1 : assert_eq!(
2937 1 : scheduler
2938 1 : .get_node_az(&tenant_shard.intent.secondary[0])
2939 1 : .as_ref(),
2940 1 : tenant_shard.preferred_az()
2941 1 : );
2942 :
2943 : // Optimizer should agree
2944 1 : assert_eq!(
2945 1 : tenant_shard.optimize_attachment(&mut scheduler, &context),
2946 1 : None
2947 1 : );
2948 1 : assert_eq!(
2949 1 : tenant_shard.optimize_secondary(&mut scheduler, &context),
2950 1 : None
2951 1 : );
2952 :
2953 1 : tenant_shard.intent.clear(&mut scheduler);
2954 1 :
2955 1 : Ok(())
2956 1 : }
2957 : }
|
