Line data Source code
1 : use std::{
2 : collections::{HashMap, HashSet},
3 : sync::Arc,
4 : time::Duration,
5 : };
6 :
7 : use crate::{
8 : metrics::{self, ReconcileCompleteLabelGroup, ReconcileOutcome},
9 : persistence::TenantShardPersistence,
10 : reconciler::ReconcileUnits,
11 : scheduler::{AffinityScore, MaySchedule, RefCountUpdate, ScheduleContext},
12 : };
13 : use pageserver_api::controller_api::{
14 : NodeSchedulingPolicy, PlacementPolicy, ShardSchedulingPolicy,
15 : };
16 : use pageserver_api::{
17 : models::{LocationConfig, LocationConfigMode, TenantConfig},
18 : shard::{ShardIdentity, TenantShardId},
19 : };
20 : use serde::Serialize;
21 : use tokio::task::JoinHandle;
22 : use tokio_util::sync::CancellationToken;
23 : use tracing::{instrument, Instrument};
24 : use utils::{
25 : generation::Generation,
26 : id::NodeId,
27 : seqwait::{SeqWait, SeqWaitError},
28 : sync::gate::GateGuard,
29 : };
30 :
31 : use crate::{
32 : compute_hook::ComputeHook,
33 : node::Node,
34 : persistence::{split_state::SplitState, Persistence},
35 : reconciler::{
36 : attached_location_conf, secondary_location_conf, ReconcileError, Reconciler, TargetState,
37 : },
38 : scheduler::{ScheduleError, Scheduler},
39 : service, Sequence,
40 : };
41 :
42 : /// Serialization helper
43 0 : fn read_last_error<S, T>(v: &std::sync::Mutex<Option<T>>, serializer: S) -> Result<S::Ok, S::Error>
44 0 : where
45 0 : S: serde::ser::Serializer,
46 0 : T: std::fmt::Display,
47 0 : {
48 0 : serializer.collect_str(
49 0 : &v.lock()
50 0 : .unwrap()
51 0 : .as_ref()
52 0 : .map(|e| format!("{e}"))
53 0 : .unwrap_or("".to_string()),
54 0 : )
55 0 : }
56 :
57 : /// In-memory state for a particular tenant shard.
58 : ///
59 : /// This struct implement Serialize for debugging purposes, but is _not_ persisted
60 : /// itself: see [`crate::persistence`] for the subset of tenant shard state that is persisted.
61 0 : #[derive(Serialize)]
62 : pub(crate) struct TenantShard {
63 : pub(crate) tenant_shard_id: TenantShardId,
64 :
65 : pub(crate) shard: ShardIdentity,
66 :
67 : // Runtime only: sequence used to coordinate when updating this object while
68 : // with background reconcilers may be running. A reconciler runs to a particular
69 : // sequence.
70 : pub(crate) sequence: Sequence,
71 :
72 : // Latest generation number: next time we attach, increment this
73 : // and use the incremented number when attaching.
74 : //
75 : // None represents an incompletely onboarded tenant via the [`Service::location_config`]
76 : // API, where this tenant may only run in PlacementPolicy::Secondary.
77 : pub(crate) generation: Option<Generation>,
78 :
79 : // High level description of how the tenant should be set up. Provided
80 : // externally.
81 : pub(crate) policy: PlacementPolicy,
82 :
83 : // Low level description of exactly which pageservers should fulfil
84 : // which role. Generated by `Self::schedule`.
85 : pub(crate) intent: IntentState,
86 :
87 : // Low level description of how the tenant is configured on pageservers:
88 : // if this does not match `Self::intent` then the tenant needs reconciliation
89 : // with `Self::reconcile`.
90 : pub(crate) observed: ObservedState,
91 :
92 : // Tenant configuration, passed through opaquely to the pageserver. Identical
93 : // for all shards in a tenant.
94 : pub(crate) config: TenantConfig,
95 :
96 : /// If a reconcile task is currently in flight, it may be joined here (it is
97 : /// only safe to join if either the result has been received or the reconciler's
98 : /// cancellation token has been fired)
99 : #[serde(skip)]
100 : pub(crate) reconciler: Option<ReconcilerHandle>,
101 :
102 : /// If a tenant is being split, then all shards with that TenantId will have a
103 : /// SplitState set, this acts as a guard against other operations such as background
104 : /// reconciliation, and timeline creation.
105 : pub(crate) splitting: SplitState,
106 :
107 : /// If a tenant was enqueued for later reconcile due to hitting concurrency limit, this flag
108 : /// is set. This flag is cleared when the tenant is popped off the delay queue.
109 : pub(crate) delayed_reconcile: bool,
110 :
111 : /// Optionally wait for reconciliation to complete up to a particular
112 : /// sequence number.
113 : #[serde(skip)]
114 : pub(crate) waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
115 :
116 : /// Indicates sequence number for which we have encountered an error reconciling. If
117 : /// this advances ahead of [`Self::waiter`] then a reconciliation error has occurred,
118 : /// and callers should stop waiting for `waiter` and propagate the error.
119 : #[serde(skip)]
120 : pub(crate) error_waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
121 :
122 : /// The most recent error from a reconcile on this tenant. This is a nested Arc
123 : /// because:
124 : /// - ReconcileWaiters need to Arc-clone the overall object to read it later
125 : /// - ReconcileWaitError needs to use an `Arc<ReconcileError>` because we can construct
126 : /// many waiters for one shard, and the underlying error types are not Clone.
127 : /// TODO: generalize to an array of recent events
128 : /// TOOD: use a ArcSwap instead of mutex for faster reads?
129 : #[serde(serialize_with = "read_last_error")]
130 : pub(crate) last_error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
131 :
132 : /// If we have a pending compute notification that for some reason we weren't able to send,
133 : /// set this to true. If this is set, calls to [`Self::get_reconcile_needed`] will return Yes
134 : /// and trigger a Reconciler run. This is the mechanism by which compute notifications are included in the scope
135 : /// of state that we publish externally in an eventually consistent way.
136 : pub(crate) pending_compute_notification: bool,
137 :
138 : // Support/debug tool: if something is going wrong or flapping with scheduling, this may
139 : // be set to a non-active state to avoid making changes while the issue is fixed.
140 : scheduling_policy: ShardSchedulingPolicy,
141 : }
142 :
143 : #[derive(Default, Clone, Debug, Serialize)]
144 : pub(crate) struct IntentState {
145 : attached: Option<NodeId>,
146 : secondary: Vec<NodeId>,
147 : }
148 :
149 : impl IntentState {
150 4 : pub(crate) fn new() -> Self {
151 4 : Self {
152 4 : attached: None,
153 4 : secondary: vec![],
154 4 : }
155 4 : }
156 0 : pub(crate) fn single(scheduler: &mut Scheduler, node_id: Option<NodeId>) -> Self {
157 0 : if let Some(node_id) = node_id {
158 0 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::Attach);
159 0 : }
160 0 : Self {
161 0 : attached: node_id,
162 0 : secondary: vec![],
163 0 : }
164 0 : }
165 :
166 26 : pub(crate) fn set_attached(&mut self, scheduler: &mut Scheduler, new_attached: Option<NodeId>) {
167 26 : if self.attached != new_attached {
168 26 : if let Some(old_attached) = self.attached.take() {
169 0 : scheduler.update_node_ref_counts(old_attached, RefCountUpdate::Detach);
170 26 : }
171 26 : if let Some(new_attached) = &new_attached {
172 26 : scheduler.update_node_ref_counts(*new_attached, RefCountUpdate::Attach);
173 26 : }
174 26 : self.attached = new_attached;
175 0 : }
176 26 : }
177 :
178 : /// Like set_attached, but the node is from [`Self::secondary`]. This swaps the node from
179 : /// secondary to attached while maintaining the scheduler's reference counts.
180 10 : pub(crate) fn promote_attached(
181 10 : &mut self,
182 10 : scheduler: &mut Scheduler,
183 10 : promote_secondary: NodeId,
184 10 : ) {
185 10 : // If we call this with a node that isn't in secondary, it would cause incorrect
186 10 : // scheduler reference counting, since we assume the node is already referenced as a secondary.
187 10 : debug_assert!(self.secondary.contains(&promote_secondary));
188 :
189 20 : self.secondary.retain(|n| n != &promote_secondary);
190 10 :
191 10 : let demoted = self.attached;
192 10 : self.attached = Some(promote_secondary);
193 10 :
194 10 : scheduler.update_node_ref_counts(promote_secondary, RefCountUpdate::PromoteSecondary);
195 10 : if let Some(demoted) = demoted {
196 0 : scheduler.update_node_ref_counts(demoted, RefCountUpdate::DemoteAttached);
197 10 : }
198 10 : }
199 :
200 34 : pub(crate) fn push_secondary(&mut self, scheduler: &mut Scheduler, new_secondary: NodeId) {
201 34 : debug_assert!(!self.secondary.contains(&new_secondary));
202 34 : scheduler.update_node_ref_counts(new_secondary, RefCountUpdate::AddSecondary);
203 34 : self.secondary.push(new_secondary);
204 34 : }
205 :
206 : /// It is legal to call this with a node that is not currently a secondary: that is a no-op
207 10 : pub(crate) fn remove_secondary(&mut self, scheduler: &mut Scheduler, node_id: NodeId) {
208 10 : let index = self.secondary.iter().position(|n| *n == node_id);
209 10 : if let Some(index) = index {
210 10 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::RemoveSecondary);
211 10 : self.secondary.remove(index);
212 10 : }
213 10 : }
214 :
215 24 : pub(crate) fn clear_secondary(&mut self, scheduler: &mut Scheduler) {
216 24 : for secondary in self.secondary.drain(..) {
217 24 : scheduler.update_node_ref_counts(secondary, RefCountUpdate::RemoveSecondary);
218 24 : }
219 24 : }
220 :
221 : /// Remove the last secondary node from the list of secondaries
222 0 : pub(crate) fn pop_secondary(&mut self, scheduler: &mut Scheduler) {
223 0 : if let Some(node_id) = self.secondary.pop() {
224 0 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::RemoveSecondary);
225 0 : }
226 0 : }
227 :
228 24 : pub(crate) fn clear(&mut self, scheduler: &mut Scheduler) {
229 24 : if let Some(old_attached) = self.attached.take() {
230 24 : scheduler.update_node_ref_counts(old_attached, RefCountUpdate::Detach);
231 24 : }
232 :
233 24 : self.clear_secondary(scheduler);
234 24 : }
235 :
236 140 : pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
237 140 : let mut result = Vec::new();
238 140 : if let Some(p) = self.attached {
239 136 : result.push(p)
240 4 : }
241 :
242 140 : result.extend(self.secondary.iter().copied());
243 140 :
244 140 : result
245 140 : }
246 :
247 118 : pub(crate) fn get_attached(&self) -> &Option<NodeId> {
248 118 : &self.attached
249 118 : }
250 :
251 32 : pub(crate) fn get_secondary(&self) -> &Vec<NodeId> {
252 32 : &self.secondary
253 32 : }
254 :
255 : /// If the node is in use as the attached location, demote it into
256 : /// the list of secondary locations. This is used when a node goes offline,
257 : /// and we want to use a different node for attachment, but not permanently
258 : /// forget the location on the offline node.
259 : ///
260 : /// Returns true if a change was made
261 10 : pub(crate) fn demote_attached(&mut self, scheduler: &mut Scheduler, node_id: NodeId) -> bool {
262 10 : if self.attached == Some(node_id) {
263 10 : self.attached = None;
264 10 : self.secondary.push(node_id);
265 10 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::DemoteAttached);
266 10 : true
267 : } else {
268 0 : false
269 : }
270 10 : }
271 : }
272 :
273 : impl Drop for IntentState {
274 26 : fn drop(&mut self) {
275 26 : // Must clear before dropping, to avoid leaving stale refcounts in the Scheduler.
276 26 : // We do not check this while panicking, to avoid polluting unit test failures or
277 26 : // other assertions with this assertion's output. It's still wrong to leak these,
278 26 : // but if we already have a panic then we don't need to independently flag this case.
279 26 : if !(std::thread::panicking()) {
280 26 : debug_assert!(self.attached.is_none() && self.secondary.is_empty());
281 0 : }
282 24 : }
283 : }
284 :
285 : #[derive(Default, Clone, Serialize)]
286 : pub(crate) struct ObservedState {
287 : pub(crate) locations: HashMap<NodeId, ObservedStateLocation>,
288 : }
289 :
290 : /// Our latest knowledge of how this tenant is configured in the outside world.
291 : ///
292 : /// Meaning:
293 : /// * No instance of this type exists for a node: we are certain that we have nothing configured on that
294 : /// node for this shard.
295 : /// * Instance exists with conf==None: we *might* have some state on that node, but we don't know
296 : /// what it is (e.g. we failed partway through configuring it)
297 : /// * Instance exists with conf==Some: this tells us what we last successfully configured on this node,
298 : /// and that configuration will still be present unless something external interfered.
299 : #[derive(Clone, Serialize)]
300 : pub(crate) struct ObservedStateLocation {
301 : /// If None, it means we do not know the status of this shard's location on this node, but
302 : /// we know that we might have some state on this node.
303 : pub(crate) conf: Option<LocationConfig>,
304 : }
305 : pub(crate) struct ReconcilerWaiter {
306 : // For observability purposes, remember the ID of the shard we're
307 : // waiting for.
308 : pub(crate) tenant_shard_id: TenantShardId,
309 :
310 : seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
311 : error_seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
312 : error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
313 : seq: Sequence,
314 : }
315 :
316 : pub(crate) enum ReconcilerStatus {
317 : Done,
318 : Failed,
319 : InProgress,
320 : }
321 :
322 0 : #[derive(thiserror::Error, Debug)]
323 : pub(crate) enum ReconcileWaitError {
324 : #[error("Timeout waiting for shard {0}")]
325 : Timeout(TenantShardId),
326 : #[error("shutting down")]
327 : Shutdown,
328 : #[error("Reconcile error on shard {0}: {1}")]
329 : Failed(TenantShardId, Arc<ReconcileError>),
330 : }
331 :
332 : #[derive(Eq, PartialEq, Debug)]
333 : pub(crate) struct ReplaceSecondary {
334 : old_node_id: NodeId,
335 : new_node_id: NodeId,
336 : }
337 :
338 : #[derive(Eq, PartialEq, Debug)]
339 : pub(crate) struct MigrateAttachment {
340 : pub(crate) old_attached_node_id: NodeId,
341 : pub(crate) new_attached_node_id: NodeId,
342 : }
343 :
344 : #[derive(Eq, PartialEq, Debug)]
345 : pub(crate) enum ScheduleOptimizationAction {
346 : // Replace one of our secondary locations with a different node
347 : ReplaceSecondary(ReplaceSecondary),
348 : // Migrate attachment to an existing secondary location
349 : MigrateAttachment(MigrateAttachment),
350 : }
351 :
352 : #[derive(Eq, PartialEq, Debug)]
353 : pub(crate) struct ScheduleOptimization {
354 : // What was the reconcile sequence when we generated this optimization? The optimization
355 : // should only be applied if the shard's sequence is still at this value, in case other changes
356 : // happened between planning the optimization and applying it.
357 : sequence: Sequence,
358 :
359 : pub(crate) action: ScheduleOptimizationAction,
360 : }
361 :
362 : impl ReconcilerWaiter {
363 0 : pub(crate) async fn wait_timeout(&self, timeout: Duration) -> Result<(), ReconcileWaitError> {
364 : tokio::select! {
365 : result = self.seq_wait.wait_for_timeout(self.seq, timeout)=> {
366 0 : result.map_err(|e| match e {
367 0 : SeqWaitError::Timeout => ReconcileWaitError::Timeout(self.tenant_shard_id),
368 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown
369 0 : })?;
370 : },
371 : result = self.error_seq_wait.wait_for(self.seq) => {
372 0 : result.map_err(|e| match e {
373 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown,
374 0 : SeqWaitError::Timeout => unreachable!()
375 0 : })?;
376 :
377 : return Err(ReconcileWaitError::Failed(self.tenant_shard_id,
378 : self.error.lock().unwrap().clone().expect("If error_seq_wait was advanced error was set").clone()))
379 : }
380 : }
381 :
382 0 : Ok(())
383 0 : }
384 :
385 0 : pub(crate) fn get_status(&self) -> ReconcilerStatus {
386 0 : if self.seq_wait.would_wait_for(self.seq).is_err() {
387 0 : ReconcilerStatus::Done
388 0 : } else if self.error_seq_wait.would_wait_for(self.seq).is_err() {
389 0 : ReconcilerStatus::Failed
390 : } else {
391 0 : ReconcilerStatus::InProgress
392 : }
393 0 : }
394 : }
395 :
396 : /// Having spawned a reconciler task, the tenant shard's state will carry enough
397 : /// information to optionally cancel & await it later.
398 : pub(crate) struct ReconcilerHandle {
399 : sequence: Sequence,
400 : handle: JoinHandle<()>,
401 : cancel: CancellationToken,
402 : }
403 :
404 : pub(crate) enum ReconcileNeeded {
405 : /// shard either doesn't need reconciliation, or is forbidden from spawning a reconciler
406 : /// in its current state (e.g. shard split in progress, or ShardSchedulingPolicy forbids it)
407 : No,
408 : /// shard has a reconciler running, and its intent hasn't changed since that one was
409 : /// spawned: wait for the existing reconciler rather than spawning a new one.
410 : WaitExisting(ReconcilerWaiter),
411 : /// shard needs reconciliation: call into [`TenantShard::spawn_reconciler`]
412 : Yes,
413 : }
414 :
415 : /// When a reconcile task completes, it sends this result object
416 : /// to be applied to the primary TenantShard.
417 : pub(crate) struct ReconcileResult {
418 : pub(crate) sequence: Sequence,
419 : /// On errors, `observed` should be treated as an incompleted description
420 : /// of state (i.e. any nodes present in the result should override nodes
421 : /// present in the parent tenant state, but any unmentioned nodes should
422 : /// not be removed from parent tenant state)
423 : pub(crate) result: Result<(), ReconcileError>,
424 :
425 : pub(crate) tenant_shard_id: TenantShardId,
426 : pub(crate) generation: Option<Generation>,
427 : pub(crate) observed: ObservedState,
428 :
429 : /// Set [`TenantShard::pending_compute_notification`] from this flag
430 : pub(crate) pending_compute_notification: bool,
431 : }
432 :
433 : impl ObservedState {
434 0 : pub(crate) fn new() -> Self {
435 0 : Self {
436 0 : locations: HashMap::new(),
437 0 : }
438 0 : }
439 : }
440 :
441 : impl TenantShard {
442 22 : pub(crate) fn new(
443 22 : tenant_shard_id: TenantShardId,
444 22 : shard: ShardIdentity,
445 22 : policy: PlacementPolicy,
446 22 : ) -> Self {
447 22 : Self {
448 22 : tenant_shard_id,
449 22 : policy,
450 22 : intent: IntentState::default(),
451 22 : generation: Some(Generation::new(0)),
452 22 : shard,
453 22 : observed: ObservedState::default(),
454 22 : config: TenantConfig::default(),
455 22 : reconciler: None,
456 22 : splitting: SplitState::Idle,
457 22 : sequence: Sequence(1),
458 22 : delayed_reconcile: false,
459 22 : waiter: Arc::new(SeqWait::new(Sequence(0))),
460 22 : error_waiter: Arc::new(SeqWait::new(Sequence(0))),
461 22 : last_error: Arc::default(),
462 22 : pending_compute_notification: false,
463 22 : scheduling_policy: ShardSchedulingPolicy::default(),
464 22 : }
465 22 : }
466 :
467 : /// For use on startup when learning state from pageservers: generate my [`IntentState`] from my
468 : /// [`ObservedState`], even if it violates my [`PlacementPolicy`]. Call [`Self::schedule`] next,
469 : /// to get an intent state that complies with placement policy. The overall goal is to do scheduling
470 : /// in a way that makes use of any configured locations that already exist in the outside world.
471 2 : pub(crate) fn intent_from_observed(&mut self, scheduler: &mut Scheduler) {
472 2 : // Choose an attached location by filtering observed locations, and then sorting to get the highest
473 2 : // generation
474 2 : let mut attached_locs = self
475 2 : .observed
476 2 : .locations
477 2 : .iter()
478 4 : .filter_map(|(node_id, l)| {
479 4 : if let Some(conf) = &l.conf {
480 4 : if conf.mode == LocationConfigMode::AttachedMulti
481 2 : || conf.mode == LocationConfigMode::AttachedSingle
482 2 : || conf.mode == LocationConfigMode::AttachedStale
483 : {
484 4 : Some((node_id, conf.generation))
485 : } else {
486 0 : None
487 : }
488 : } else {
489 0 : None
490 : }
491 4 : })
492 2 : .collect::<Vec<_>>();
493 2 :
494 4 : attached_locs.sort_by_key(|i| i.1);
495 2 : if let Some((node_id, _gen)) = attached_locs.into_iter().last() {
496 2 : self.intent.set_attached(scheduler, Some(*node_id));
497 2 : }
498 :
499 : // All remaining observed locations generate secondary intents. This includes None
500 : // observations, as these may well have some local content on disk that is usable (this
501 : // is an edge case that might occur if we restarted during a migration or other change)
502 : //
503 : // We may leave intent.attached empty if we didn't find any attached locations: [`Self::schedule`]
504 : // will take care of promoting one of these secondaries to be attached.
505 4 : self.observed.locations.keys().for_each(|node_id| {
506 4 : if Some(*node_id) != self.intent.attached {
507 2 : self.intent.push_secondary(scheduler, *node_id);
508 2 : }
509 4 : });
510 2 : }
511 :
512 : /// Part of [`Self::schedule`] that is used to choose exactly one node to act as the
513 : /// attached pageserver for a shard.
514 : ///
515 : /// Returns whether we modified it, and the NodeId selected.
516 14 : fn schedule_attached(
517 14 : &mut self,
518 14 : scheduler: &mut Scheduler,
519 14 : context: &ScheduleContext,
520 14 : ) -> Result<(bool, NodeId), ScheduleError> {
521 : // No work to do if we already have an attached tenant
522 14 : if let Some(node_id) = self.intent.attached {
523 0 : return Ok((false, node_id));
524 14 : }
525 :
526 14 : if let Some(promote_secondary) = scheduler.node_preferred(&self.intent.secondary) {
527 : // Promote a secondary
528 2 : tracing::debug!("Promoted secondary {} to attached", promote_secondary);
529 2 : self.intent.promote_attached(scheduler, promote_secondary);
530 2 : Ok((true, promote_secondary))
531 : } else {
532 : // Pick a fresh node: either we had no secondaries or none were schedulable
533 12 : let node_id = scheduler.schedule_shard(&self.intent.secondary, context)?;
534 12 : tracing::debug!("Selected {} as attached", node_id);
535 12 : self.intent.set_attached(scheduler, Some(node_id));
536 12 : Ok((true, node_id))
537 : }
538 14 : }
539 :
540 16 : pub(crate) fn schedule(
541 16 : &mut self,
542 16 : scheduler: &mut Scheduler,
543 16 : context: &mut ScheduleContext,
544 16 : ) -> Result<(), ScheduleError> {
545 16 : let r = self.do_schedule(scheduler, context);
546 16 :
547 16 : context.avoid(&self.intent.all_pageservers());
548 16 : if let Some(attached) = self.intent.get_attached() {
549 14 : context.push_attached(*attached);
550 14 : }
551 :
552 16 : r
553 16 : }
554 :
555 16 : pub(crate) fn do_schedule(
556 16 : &mut self,
557 16 : scheduler: &mut Scheduler,
558 16 : context: &ScheduleContext,
559 16 : ) -> Result<(), ScheduleError> {
560 16 : // TODO: before scheduling new nodes, check if any existing content in
561 16 : // self.intent refers to pageservers that are offline, and pick other
562 16 : // pageservers if so.
563 16 :
564 16 : // TODO: respect the splitting bit on tenants: if they are currently splitting then we may not
565 16 : // change their attach location.
566 16 :
567 16 : match self.scheduling_policy {
568 14 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {}
569 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
570 : // Warn to make it obvious why other things aren't happening/working, if we skip scheduling
571 2 : tracing::warn!(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(),
572 0 : "Scheduling is disabled by policy {:?}", self.scheduling_policy);
573 2 : return Ok(());
574 : }
575 : }
576 :
577 : // Build the set of pageservers already in use by this tenant, to avoid scheduling
578 : // more work on the same pageservers we're already using.
579 14 : let mut modified = false;
580 14 :
581 14 : // Add/remove nodes to fulfil policy
582 14 : use PlacementPolicy::*;
583 14 : match self.policy {
584 14 : Attached(secondary_count) => {
585 14 : let retain_secondaries = if self.intent.attached.is_none()
586 14 : && scheduler.node_preferred(&self.intent.secondary).is_some()
587 : {
588 : // If we have no attached, and one of the secondaries is elegible to be promoted, retain
589 : // one more secondary than we usually would, as one of them will become attached futher down this function.
590 2 : secondary_count + 1
591 : } else {
592 12 : secondary_count
593 : };
594 :
595 14 : while self.intent.secondary.len() > retain_secondaries {
596 0 : // We have no particular preference for one secondary location over another: just
597 0 : // arbitrarily drop from the end
598 0 : self.intent.pop_secondary(scheduler);
599 0 : modified = true;
600 0 : }
601 :
602 : // Should have exactly one attached, and N secondaries
603 14 : let (modified_attached, attached_node_id) =
604 14 : self.schedule_attached(scheduler, context)?;
605 14 : modified |= modified_attached;
606 14 :
607 14 : let mut used_pageservers = vec![attached_node_id];
608 26 : while self.intent.secondary.len() < secondary_count {
609 12 : let node_id = scheduler.schedule_shard(&used_pageservers, context)?;
610 12 : self.intent.push_secondary(scheduler, node_id);
611 12 : used_pageservers.push(node_id);
612 12 : modified = true;
613 : }
614 : }
615 : Secondary => {
616 0 : if let Some(node_id) = self.intent.get_attached() {
617 0 : // Populate secondary by demoting the attached node
618 0 : self.intent.demote_attached(scheduler, *node_id);
619 0 : modified = true;
620 0 : } else if self.intent.secondary.is_empty() {
621 0 : // Populate secondary by scheduling a fresh node
622 0 : let node_id = scheduler.schedule_shard(&[], context)?;
623 0 : self.intent.push_secondary(scheduler, node_id);
624 0 : modified = true;
625 0 : }
626 0 : while self.intent.secondary.len() > 1 {
627 0 : // We have no particular preference for one secondary location over another: just
628 0 : // arbitrarily drop from the end
629 0 : self.intent.pop_secondary(scheduler);
630 0 : modified = true;
631 0 : }
632 : }
633 : Detached => {
634 : // Never add locations in this mode
635 0 : if self.intent.get_attached().is_some() || !self.intent.get_secondary().is_empty() {
636 0 : self.intent.clear(scheduler);
637 0 : modified = true;
638 0 : }
639 : }
640 : }
641 :
642 14 : if modified {
643 14 : self.sequence.0 += 1;
644 14 : }
645 :
646 14 : Ok(())
647 16 : }
648 :
649 : /// Reschedule this tenant shard to one of its secondary locations. Returns a scheduling error
650 : /// if the swap is not possible and leaves the intent state in its original state.
651 : ///
652 : /// Arguments:
653 : /// `attached_to`: the currently attached location matching the intent state (may be None if the
654 : /// shard is not attached)
655 : /// `promote_to`: an optional secondary location of this tenant shard. If set to None, we ask
656 : /// the scheduler to recommend a node
657 0 : pub(crate) fn reschedule_to_secondary(
658 0 : &mut self,
659 0 : promote_to: Option<NodeId>,
660 0 : scheduler: &mut Scheduler,
661 0 : ) -> Result<(), ScheduleError> {
662 0 : let promote_to = match promote_to {
663 0 : Some(node) => node,
664 0 : None => match scheduler.node_preferred(self.intent.get_secondary()) {
665 0 : Some(node) => node,
666 : None => {
667 0 : return Err(ScheduleError::ImpossibleConstraint);
668 : }
669 : },
670 : };
671 :
672 0 : assert!(self.intent.get_secondary().contains(&promote_to));
673 :
674 0 : if let Some(node) = self.intent.get_attached() {
675 0 : let demoted = self.intent.demote_attached(scheduler, *node);
676 0 : if !demoted {
677 0 : return Err(ScheduleError::ImpossibleConstraint);
678 0 : }
679 0 : }
680 :
681 0 : self.intent.promote_attached(scheduler, promote_to);
682 0 :
683 0 : // Increment the sequence number for the edge case where a
684 0 : // reconciler is already running to avoid waiting on the
685 0 : // current reconcile instead of spawning a new one.
686 0 : self.sequence = self.sequence.next();
687 0 :
688 0 : Ok(())
689 0 : }
690 :
691 : /// Optimize attachments: if a shard has a secondary location that is preferable to
692 : /// its primary location based on soft constraints, switch that secondary location
693 : /// to be attached.
694 30 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
695 : pub(crate) fn optimize_attachment(
696 : &self,
697 : nodes: &HashMap<NodeId, Node>,
698 : schedule_context: &ScheduleContext,
699 : ) -> Option<ScheduleOptimization> {
700 : let attached = (*self.intent.get_attached())?;
701 : if self.intent.secondary.is_empty() {
702 : // We can only do useful work if we have both attached and secondary locations: this
703 : // function doesn't schedule new locations, only swaps between attached and secondaries.
704 : return None;
705 : }
706 :
707 : let current_affinity_score = schedule_context.get_node_affinity(attached);
708 : let current_attachment_count = schedule_context.get_node_attachments(attached);
709 :
710 : // Generate score for each node, dropping any un-schedulable nodes.
711 : let all_pageservers = self.intent.all_pageservers();
712 : let mut scores = all_pageservers
713 : .iter()
714 60 : .flat_map(|node_id| {
715 60 : let node = nodes.get(node_id);
716 60 : if node.is_none() {
717 0 : None
718 60 : } else if matches!(
719 60 : node.unwrap().get_scheduling(),
720 : NodeSchedulingPolicy::Filling
721 : ) {
722 : // If the node is currently filling, don't count it as a candidate to avoid,
723 : // racing with the background fill.
724 0 : None
725 60 : } else if matches!(node.unwrap().may_schedule(), MaySchedule::No) {
726 0 : None
727 : } else {
728 60 : let affinity_score = schedule_context.get_node_affinity(*node_id);
729 60 : let attachment_count = schedule_context.get_node_attachments(*node_id);
730 60 : Some((*node_id, affinity_score, attachment_count))
731 : }
732 60 : })
733 : .collect::<Vec<_>>();
734 :
735 : // Sort precedence:
736 : // 1st - prefer nodes with the lowest total affinity score
737 : // 2nd - prefer nodes with the lowest number of attachments in this context
738 : // 3rd - if all else is equal, sort by node ID for determinism in tests.
739 60 : scores.sort_by_key(|i| (i.1, i.2, i.0));
740 :
741 : if let Some((preferred_node, preferred_affinity_score, preferred_attachment_count)) =
742 : scores.first()
743 : {
744 : if attached != *preferred_node {
745 : // The best alternative must be more than 1 better than us, otherwise we could end
746 : // up flapping back next time we're called (e.g. there's no point migrating from
747 : // a location with score 1 to a score zero, because on next location the situation
748 : // would be the same, but in reverse).
749 : if current_affinity_score > *preferred_affinity_score + AffinityScore(1)
750 : || current_attachment_count > *preferred_attachment_count + 1
751 : {
752 : tracing::info!(
753 : "Identified optimization: migrate attachment {attached}->{preferred_node} (secondaries {:?})",
754 : self.intent.get_secondary()
755 : );
756 : return Some(ScheduleOptimization {
757 : sequence: self.sequence,
758 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
759 : old_attached_node_id: attached,
760 : new_attached_node_id: *preferred_node,
761 : }),
762 : });
763 : }
764 : } else {
765 : tracing::debug!(
766 : "Node {} is already preferred (score {:?})",
767 : preferred_node,
768 : preferred_affinity_score
769 : );
770 : }
771 : }
772 :
773 : // Fall-through: we didn't find an optimization
774 : None
775 : }
776 :
777 24 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
778 : pub(crate) fn optimize_secondary(
779 : &self,
780 : scheduler: &Scheduler,
781 : schedule_context: &ScheduleContext,
782 : ) -> Option<ScheduleOptimization> {
783 : if self.intent.secondary.is_empty() {
784 : // We can only do useful work if we have both attached and secondary locations: this
785 : // function doesn't schedule new locations, only swaps between attached and secondaries.
786 : return None;
787 : }
788 :
789 : for secondary in self.intent.get_secondary() {
790 : let Some(affinity_score) = schedule_context.nodes.get(secondary) else {
791 : // We're already on a node unaffected any affinity constraints,
792 : // so we won't change it.
793 : continue;
794 : };
795 :
796 : // Let the scheduler suggest a node, where it would put us if we were scheduling afresh
797 : // This implicitly limits the choice to nodes that are available, and prefers nodes
798 : // with lower utilization.
799 : let Ok(candidate_node) =
800 : scheduler.schedule_shard(&self.intent.all_pageservers(), schedule_context)
801 : else {
802 : // A scheduling error means we have no possible candidate replacements
803 : continue;
804 : };
805 :
806 : let candidate_affinity_score = schedule_context
807 : .nodes
808 : .get(&candidate_node)
809 : .unwrap_or(&AffinityScore::FREE);
810 :
811 : // The best alternative must be more than 1 better than us, otherwise we could end
812 : // up flapping back next time we're called.
813 : if *candidate_affinity_score + AffinityScore(1) < *affinity_score {
814 : // If some other node is available and has a lower score than this node, then
815 : // that other node is a good place to migrate to.
816 : tracing::info!(
817 : "Identified optimization: replace secondary {secondary}->{candidate_node} (current secondaries {:?})",
818 : self.intent.get_secondary()
819 : );
820 : return Some(ScheduleOptimization {
821 : sequence: self.sequence,
822 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
823 : old_node_id: *secondary,
824 : new_node_id: candidate_node,
825 : }),
826 : });
827 : }
828 : }
829 :
830 : None
831 : }
832 :
833 : /// Return true if the optimization was really applied: it will not be applied if the optimization's
834 : /// sequence is behind this tenant shard's
835 18 : pub(crate) fn apply_optimization(
836 18 : &mut self,
837 18 : scheduler: &mut Scheduler,
838 18 : optimization: ScheduleOptimization,
839 18 : ) -> bool {
840 18 : if optimization.sequence != self.sequence {
841 0 : return false;
842 18 : }
843 18 :
844 18 : metrics::METRICS_REGISTRY
845 18 : .metrics_group
846 18 : .storage_controller_schedule_optimization
847 18 : .inc();
848 18 :
849 18 : match optimization.action {
850 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
851 8 : old_attached_node_id,
852 8 : new_attached_node_id,
853 8 : }) => {
854 8 : self.intent.demote_attached(scheduler, old_attached_node_id);
855 8 : self.intent
856 8 : .promote_attached(scheduler, new_attached_node_id);
857 8 : }
858 : ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
859 10 : old_node_id,
860 10 : new_node_id,
861 10 : }) => {
862 10 : self.intent.remove_secondary(scheduler, old_node_id);
863 10 : self.intent.push_secondary(scheduler, new_node_id);
864 10 : }
865 : }
866 :
867 18 : true
868 18 : }
869 :
870 : /// Query whether the tenant's observed state for attached node matches its intent state, and if so,
871 : /// yield the node ID. This is appropriate for emitting compute hook notifications: we are checking that
872 : /// the node in question is not only where we intend to attach, but that the tenant is indeed already attached there.
873 : ///
874 : /// Reconciliation may still be needed for other aspects of state such as secondaries (see [`Self::dirty`]): this
875 : /// funciton should not be used to decide whether to reconcile.
876 0 : pub(crate) fn stably_attached(&self) -> Option<NodeId> {
877 0 : if let Some(attach_intent) = self.intent.attached {
878 0 : match self.observed.locations.get(&attach_intent) {
879 0 : Some(loc) => match &loc.conf {
880 0 : Some(conf) => match conf.mode {
881 : LocationConfigMode::AttachedMulti
882 : | LocationConfigMode::AttachedSingle
883 : | LocationConfigMode::AttachedStale => {
884 : // Our intent and observed state agree that this node is in an attached state.
885 0 : Some(attach_intent)
886 : }
887 : // Our observed config is not an attached state
888 0 : _ => None,
889 : },
890 : // Our observed state is None, i.e. in flux
891 0 : None => None,
892 : },
893 : // We have no observed state for this node
894 0 : None => None,
895 : }
896 : } else {
897 : // Our intent is not to attach
898 0 : None
899 : }
900 0 : }
901 :
902 0 : fn dirty(&self, nodes: &Arc<HashMap<NodeId, Node>>) -> bool {
903 0 : let mut dirty_nodes = HashSet::new();
904 :
905 0 : if let Some(node_id) = self.intent.attached {
906 : // Maybe panic: it is a severe bug if we try to attach while generation is null.
907 0 : let generation = self
908 0 : .generation
909 0 : .expect("Attempted to enter attached state without a generation");
910 0 :
911 0 : let wanted_conf = attached_location_conf(
912 0 : generation,
913 0 : &self.shard,
914 0 : &self.config,
915 0 : !self.intent.secondary.is_empty(),
916 0 : );
917 0 : match self.observed.locations.get(&node_id) {
918 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
919 0 : Some(_) | None => {
920 0 : dirty_nodes.insert(node_id);
921 0 : }
922 : }
923 0 : }
924 :
925 0 : for node_id in &self.intent.secondary {
926 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
927 0 : match self.observed.locations.get(node_id) {
928 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
929 0 : Some(_) | None => {
930 0 : dirty_nodes.insert(*node_id);
931 0 : }
932 : }
933 : }
934 :
935 0 : for node_id in self.observed.locations.keys() {
936 0 : if self.intent.attached != Some(*node_id) && !self.intent.secondary.contains(node_id) {
937 0 : // We have observed state that isn't part of our intent: need to clean it up.
938 0 : dirty_nodes.insert(*node_id);
939 0 : }
940 : }
941 :
942 0 : dirty_nodes.retain(|node_id| {
943 0 : nodes
944 0 : .get(node_id)
945 0 : .map(|n| n.is_available())
946 0 : .unwrap_or(false)
947 0 : });
948 0 :
949 0 : !dirty_nodes.is_empty()
950 0 : }
951 :
952 : #[allow(clippy::too_many_arguments)]
953 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
954 : pub(crate) fn get_reconcile_needed(
955 : &mut self,
956 : pageservers: &Arc<HashMap<NodeId, Node>>,
957 : ) -> ReconcileNeeded {
958 : // If there are any ambiguous observed states, and the nodes they refer to are available,
959 : // we should reconcile to clean them up.
960 : let mut dirty_observed = false;
961 : for (node_id, observed_loc) in &self.observed.locations {
962 : let node = pageservers
963 : .get(node_id)
964 : .expect("Nodes may not be removed while referenced");
965 : if observed_loc.conf.is_none() && node.is_available() {
966 : dirty_observed = true;
967 : break;
968 : }
969 : }
970 :
971 : let active_nodes_dirty = self.dirty(pageservers);
972 :
973 : // Even if there is no pageserver work to be done, if we have a pending notification to computes,
974 : // wake up a reconciler to send it.
975 : let do_reconcile =
976 : active_nodes_dirty || dirty_observed || self.pending_compute_notification;
977 :
978 : if !do_reconcile {
979 : tracing::debug!("Not dirty, no reconciliation needed.");
980 : return ReconcileNeeded::No;
981 : }
982 :
983 : // If we are currently splitting, then never start a reconciler task: the splitting logic
984 : // requires that shards are not interfered with while it runs. Do this check here rather than
985 : // up top, so that we only log this message if we would otherwise have done a reconciliation.
986 : if !matches!(self.splitting, SplitState::Idle) {
987 : tracing::info!("Refusing to reconcile, splitting in progress");
988 : return ReconcileNeeded::No;
989 : }
990 :
991 : // Reconcile already in flight for the current sequence?
992 : if let Some(handle) = &self.reconciler {
993 : if handle.sequence == self.sequence {
994 : tracing::info!(
995 : "Reconciliation already in progress for sequence {:?}",
996 : self.sequence,
997 : );
998 : return ReconcileNeeded::WaitExisting(ReconcilerWaiter {
999 : tenant_shard_id: self.tenant_shard_id,
1000 : seq_wait: self.waiter.clone(),
1001 : error_seq_wait: self.error_waiter.clone(),
1002 : error: self.last_error.clone(),
1003 : seq: self.sequence,
1004 : });
1005 : }
1006 : }
1007 :
1008 : // Pre-checks done: finally check whether we may actually do the work
1009 : match self.scheduling_policy {
1010 : ShardSchedulingPolicy::Active
1011 : | ShardSchedulingPolicy::Essential
1012 : | ShardSchedulingPolicy::Pause => {}
1013 : ShardSchedulingPolicy::Stop => {
1014 : // We only reach this point if there is work to do and we're going to skip
1015 : // doing it: warn it obvious why this tenant isn't doing what it ought to.
1016 : tracing::warn!("Skipping reconcile for policy {:?}", self.scheduling_policy);
1017 : return ReconcileNeeded::No;
1018 : }
1019 : }
1020 :
1021 : ReconcileNeeded::Yes
1022 : }
1023 :
1024 : /// Ensure the sequence number is set to a value where waiting for this value will make us wait
1025 : /// for the next reconcile: i.e. it is ahead of all completed or running reconcilers.
1026 : ///
1027 : /// Constructing a ReconcilerWaiter with the resulting sequence number gives the property
1028 : /// that the waiter will not complete until some future Reconciler is constructed and run.
1029 0 : fn ensure_sequence_ahead(&mut self) {
1030 0 : // Find the highest sequence for which a Reconciler has previously run or is currently
1031 0 : // running
1032 0 : let max_seen = std::cmp::max(
1033 0 : self.reconciler
1034 0 : .as_ref()
1035 0 : .map(|r| r.sequence)
1036 0 : .unwrap_or(Sequence(0)),
1037 0 : std::cmp::max(self.waiter.load(), self.error_waiter.load()),
1038 0 : );
1039 0 :
1040 0 : if self.sequence <= max_seen {
1041 0 : self.sequence = max_seen.next();
1042 0 : }
1043 0 : }
1044 :
1045 : /// Create a waiter that will wait for some future Reconciler that hasn't been spawned yet.
1046 : ///
1047 : /// This is appropriate when you can't spawn a reconciler (e.g. due to resource limits), but
1048 : /// you would like to wait on the next reconciler that gets spawned in the background.
1049 0 : pub(crate) fn future_reconcile_waiter(&mut self) -> ReconcilerWaiter {
1050 0 : self.ensure_sequence_ahead();
1051 0 :
1052 0 : ReconcilerWaiter {
1053 0 : tenant_shard_id: self.tenant_shard_id,
1054 0 : seq_wait: self.waiter.clone(),
1055 0 : error_seq_wait: self.error_waiter.clone(),
1056 0 : error: self.last_error.clone(),
1057 0 : seq: self.sequence,
1058 0 : }
1059 0 : }
1060 :
1061 : #[allow(clippy::too_many_arguments)]
1062 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1063 : pub(crate) fn spawn_reconciler(
1064 : &mut self,
1065 : result_tx: &tokio::sync::mpsc::UnboundedSender<ReconcileResult>,
1066 : pageservers: &Arc<HashMap<NodeId, Node>>,
1067 : compute_hook: &Arc<ComputeHook>,
1068 : service_config: &service::Config,
1069 : persistence: &Arc<Persistence>,
1070 : units: ReconcileUnits,
1071 : gate_guard: GateGuard,
1072 : cancel: &CancellationToken,
1073 : ) -> Option<ReconcilerWaiter> {
1074 : // Reconcile in flight for a stale sequence? Our sequence's task will wait for it before
1075 : // doing our sequence's work.
1076 : let old_handle = self.reconciler.take();
1077 :
1078 : // Build list of nodes from which the reconciler should detach
1079 : let mut detach = Vec::new();
1080 : for node_id in self.observed.locations.keys() {
1081 : if self.intent.get_attached() != &Some(*node_id)
1082 : && !self.intent.secondary.contains(node_id)
1083 : {
1084 : detach.push(
1085 : pageservers
1086 : .get(node_id)
1087 : .expect("Intent references non-existent pageserver")
1088 : .clone(),
1089 : )
1090 : }
1091 : }
1092 :
1093 : // Advance the sequence before spawning a reconciler, so that sequence waiters
1094 : // can distinguish between before+after the reconcile completes.
1095 : self.ensure_sequence_ahead();
1096 :
1097 : let reconciler_cancel = cancel.child_token();
1098 : let reconciler_intent = TargetState::from_intent(pageservers, &self.intent);
1099 : let mut reconciler = Reconciler {
1100 : tenant_shard_id: self.tenant_shard_id,
1101 : shard: self.shard,
1102 : generation: self.generation,
1103 : intent: reconciler_intent,
1104 : detach,
1105 : config: self.config.clone(),
1106 : observed: self.observed.clone(),
1107 : compute_hook: compute_hook.clone(),
1108 : service_config: service_config.clone(),
1109 : _gate_guard: gate_guard,
1110 : _resource_units: units,
1111 : cancel: reconciler_cancel.clone(),
1112 : persistence: persistence.clone(),
1113 : compute_notify_failure: false,
1114 : };
1115 :
1116 : let reconcile_seq = self.sequence;
1117 :
1118 : tracing::info!(seq=%reconcile_seq, "Spawning Reconciler for sequence {}", self.sequence);
1119 : let must_notify = self.pending_compute_notification;
1120 : let reconciler_span = tracing::info_span!(parent: None, "reconciler", seq=%reconcile_seq,
1121 : tenant_id=%reconciler.tenant_shard_id.tenant_id,
1122 : shard_id=%reconciler.tenant_shard_id.shard_slug());
1123 : metrics::METRICS_REGISTRY
1124 : .metrics_group
1125 : .storage_controller_reconcile_spawn
1126 : .inc();
1127 : let result_tx = result_tx.clone();
1128 : let join_handle = tokio::task::spawn(
1129 0 : async move {
1130 : // Wait for any previous reconcile task to complete before we start
1131 0 : if let Some(old_handle) = old_handle {
1132 0 : old_handle.cancel.cancel();
1133 0 : if let Err(e) = old_handle.handle.await {
1134 : // We can't do much with this other than log it: the task is done, so
1135 : // we may proceed with our work.
1136 0 : tracing::error!("Unexpected join error waiting for reconcile task: {e}");
1137 0 : }
1138 0 : }
1139 :
1140 : // Early check for cancellation before doing any work
1141 : // TODO: wrap all remote API operations in cancellation check
1142 : // as well.
1143 0 : if reconciler.cancel.is_cancelled() {
1144 0 : metrics::METRICS_REGISTRY
1145 0 : .metrics_group
1146 0 : .storage_controller_reconcile_complete
1147 0 : .inc(ReconcileCompleteLabelGroup {
1148 0 : status: ReconcileOutcome::Cancel,
1149 0 : });
1150 0 : return;
1151 0 : }
1152 :
1153 : // Attempt to make observed state match intent state
1154 0 : let result = reconciler.reconcile().await;
1155 :
1156 : // If we know we had a pending compute notification from some previous action, send a notification irrespective
1157 : // of whether the above reconcile() did any work
1158 0 : if result.is_ok() && must_notify {
1159 : // If this fails we will send the need to retry in [`ReconcileResult::pending_compute_notification`]
1160 0 : reconciler.compute_notify().await.ok();
1161 0 : }
1162 :
1163 : // Update result counter
1164 0 : let outcome_label = match &result {
1165 0 : Ok(_) => ReconcileOutcome::Success,
1166 0 : Err(ReconcileError::Cancel) => ReconcileOutcome::Cancel,
1167 0 : Err(_) => ReconcileOutcome::Error,
1168 : };
1169 :
1170 0 : metrics::METRICS_REGISTRY
1171 0 : .metrics_group
1172 0 : .storage_controller_reconcile_complete
1173 0 : .inc(ReconcileCompleteLabelGroup {
1174 0 : status: outcome_label,
1175 0 : });
1176 0 :
1177 0 : // Constructing result implicitly drops Reconciler, freeing any ReconcileUnits before the Service might
1178 0 : // try and schedule more work in response to our result.
1179 0 : let result = ReconcileResult {
1180 0 : sequence: reconcile_seq,
1181 0 : result,
1182 0 : tenant_shard_id: reconciler.tenant_shard_id,
1183 0 : generation: reconciler.generation,
1184 0 : observed: reconciler.observed,
1185 0 : pending_compute_notification: reconciler.compute_notify_failure,
1186 0 : };
1187 0 :
1188 0 : result_tx.send(result).ok();
1189 0 : }
1190 : .instrument(reconciler_span),
1191 : );
1192 :
1193 : self.reconciler = Some(ReconcilerHandle {
1194 : sequence: self.sequence,
1195 : handle: join_handle,
1196 : cancel: reconciler_cancel,
1197 : });
1198 :
1199 : Some(ReconcilerWaiter {
1200 : tenant_shard_id: self.tenant_shard_id,
1201 : seq_wait: self.waiter.clone(),
1202 : error_seq_wait: self.error_waiter.clone(),
1203 : error: self.last_error.clone(),
1204 : seq: self.sequence,
1205 : })
1206 : }
1207 :
1208 : /// Get a waiter for any reconciliation in flight, but do not start reconciliation
1209 : /// if it is not already running
1210 0 : pub(crate) fn get_waiter(&self) -> Option<ReconcilerWaiter> {
1211 0 : if self.reconciler.is_some() {
1212 0 : Some(ReconcilerWaiter {
1213 0 : tenant_shard_id: self.tenant_shard_id,
1214 0 : seq_wait: self.waiter.clone(),
1215 0 : error_seq_wait: self.error_waiter.clone(),
1216 0 : error: self.last_error.clone(),
1217 0 : seq: self.sequence,
1218 0 : })
1219 : } else {
1220 0 : None
1221 : }
1222 0 : }
1223 :
1224 : /// Called when a ReconcileResult has been emitted and the service is updating
1225 : /// our state: if the result is from a sequence >= my ReconcileHandle, then drop
1226 : /// the handle to indicate there is no longer a reconciliation in progress.
1227 0 : pub(crate) fn reconcile_complete(&mut self, sequence: Sequence) {
1228 0 : if let Some(reconcile_handle) = &self.reconciler {
1229 0 : if reconcile_handle.sequence <= sequence {
1230 0 : self.reconciler = None;
1231 0 : }
1232 0 : }
1233 0 : }
1234 :
1235 : // If we had any state at all referring to this node ID, drop it. Does not
1236 : // attempt to reschedule.
1237 0 : pub(crate) fn deref_node(&mut self, node_id: NodeId) {
1238 0 : if self.intent.attached == Some(node_id) {
1239 0 : self.intent.attached = None;
1240 0 : }
1241 :
1242 0 : self.intent.secondary.retain(|n| n != &node_id);
1243 0 :
1244 0 : self.observed.locations.remove(&node_id);
1245 0 :
1246 0 : debug_assert!(!self.intent.all_pageservers().contains(&node_id));
1247 0 : }
1248 :
1249 0 : pub(crate) fn set_scheduling_policy(&mut self, p: ShardSchedulingPolicy) {
1250 0 : self.scheduling_policy = p;
1251 0 : }
1252 :
1253 0 : pub(crate) fn get_scheduling_policy(&self) -> &ShardSchedulingPolicy {
1254 0 : &self.scheduling_policy
1255 0 : }
1256 :
1257 0 : pub(crate) fn set_last_error(&mut self, sequence: Sequence, error: ReconcileError) {
1258 0 : // Ordering: always set last_error before advancing sequence, so that sequence
1259 0 : // waiters are guaranteed to see a Some value when they see an error.
1260 0 : *(self.last_error.lock().unwrap()) = Some(Arc::new(error));
1261 0 : self.error_waiter.advance(sequence);
1262 0 : }
1263 :
1264 0 : pub(crate) fn from_persistent(
1265 0 : tsp: TenantShardPersistence,
1266 0 : intent: IntentState,
1267 0 : ) -> anyhow::Result<Self> {
1268 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1269 0 : let shard_identity = tsp.get_shard_identity()?;
1270 :
1271 0 : Ok(Self {
1272 0 : tenant_shard_id,
1273 0 : shard: shard_identity,
1274 0 : sequence: Sequence::initial(),
1275 0 : generation: tsp.generation.map(|g| Generation::new(g as u32)),
1276 0 : policy: serde_json::from_str(&tsp.placement_policy).unwrap(),
1277 0 : intent,
1278 0 : observed: ObservedState::new(),
1279 0 : config: serde_json::from_str(&tsp.config).unwrap(),
1280 0 : reconciler: None,
1281 0 : splitting: tsp.splitting,
1282 0 : waiter: Arc::new(SeqWait::new(Sequence::initial())),
1283 0 : error_waiter: Arc::new(SeqWait::new(Sequence::initial())),
1284 0 : last_error: Arc::default(),
1285 0 : pending_compute_notification: false,
1286 0 : delayed_reconcile: false,
1287 0 : scheduling_policy: serde_json::from_str(&tsp.scheduling_policy).unwrap(),
1288 0 : })
1289 0 : }
1290 :
1291 0 : pub(crate) fn to_persistent(&self) -> TenantShardPersistence {
1292 0 : TenantShardPersistence {
1293 0 : tenant_id: self.tenant_shard_id.tenant_id.to_string(),
1294 0 : shard_number: self.tenant_shard_id.shard_number.0 as i32,
1295 0 : shard_count: self.tenant_shard_id.shard_count.literal() as i32,
1296 0 : shard_stripe_size: self.shard.stripe_size.0 as i32,
1297 0 : generation: self.generation.map(|g| g.into().unwrap_or(0) as i32),
1298 0 : generation_pageserver: self.intent.get_attached().map(|n| n.0 as i64),
1299 0 : placement_policy: serde_json::to_string(&self.policy).unwrap(),
1300 0 : config: serde_json::to_string(&self.config).unwrap(),
1301 0 : splitting: SplitState::default(),
1302 0 : scheduling_policy: serde_json::to_string(&self.scheduling_policy).unwrap(),
1303 0 : }
1304 0 : }
1305 : }
1306 :
1307 : #[cfg(test)]
1308 : pub(crate) mod tests {
1309 : use pageserver_api::{
1310 : controller_api::NodeAvailability,
1311 : shard::{ShardCount, ShardNumber},
1312 : };
1313 : use utils::id::TenantId;
1314 :
1315 : use crate::scheduler::test_utils::make_test_nodes;
1316 :
1317 : use super::*;
1318 :
1319 14 : fn make_test_tenant_shard(policy: PlacementPolicy) -> TenantShard {
1320 14 : let tenant_id = TenantId::generate();
1321 14 : let shard_number = ShardNumber(0);
1322 14 : let shard_count = ShardCount::new(1);
1323 14 :
1324 14 : let tenant_shard_id = TenantShardId {
1325 14 : tenant_id,
1326 14 : shard_number,
1327 14 : shard_count,
1328 14 : };
1329 14 : TenantShard::new(
1330 14 : tenant_shard_id,
1331 14 : ShardIdentity::new(
1332 14 : shard_number,
1333 14 : shard_count,
1334 14 : pageserver_api::shard::ShardStripeSize(32768),
1335 14 : )
1336 14 : .unwrap(),
1337 14 : policy,
1338 14 : )
1339 14 : }
1340 :
1341 2 : fn make_test_tenant(policy: PlacementPolicy, shard_count: ShardCount) -> Vec<TenantShard> {
1342 2 : let tenant_id = TenantId::generate();
1343 2 :
1344 2 : (0..shard_count.count())
1345 8 : .map(|i| {
1346 8 : let shard_number = ShardNumber(i);
1347 8 :
1348 8 : let tenant_shard_id = TenantShardId {
1349 8 : tenant_id,
1350 8 : shard_number,
1351 8 : shard_count,
1352 8 : };
1353 8 : TenantShard::new(
1354 8 : tenant_shard_id,
1355 8 : ShardIdentity::new(
1356 8 : shard_number,
1357 8 : shard_count,
1358 8 : pageserver_api::shard::ShardStripeSize(32768),
1359 8 : )
1360 8 : .unwrap(),
1361 8 : policy.clone(),
1362 8 : )
1363 8 : })
1364 2 : .collect()
1365 2 : }
1366 :
1367 : /// Test the scheduling behaviors used when a tenant configured for HA is subject
1368 : /// to nodes being marked offline.
1369 : #[test]
1370 2 : fn tenant_ha_scheduling() -> anyhow::Result<()> {
1371 2 : // Start with three nodes. Our tenant will only use two. The third one is
1372 2 : // expected to remain unused.
1373 2 : let mut nodes = make_test_nodes(3);
1374 2 :
1375 2 : let mut scheduler = Scheduler::new(nodes.values());
1376 2 : let mut context = ScheduleContext::default();
1377 2 :
1378 2 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
1379 2 : tenant_shard
1380 2 : .schedule(&mut scheduler, &mut context)
1381 2 : .expect("we have enough nodes, scheduling should work");
1382 2 :
1383 2 : // Expect to initially be schedule on to different nodes
1384 2 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
1385 2 : assert!(tenant_shard.intent.attached.is_some());
1386 :
1387 2 : let attached_node_id = tenant_shard.intent.attached.unwrap();
1388 2 : let secondary_node_id = *tenant_shard.intent.secondary.iter().last().unwrap();
1389 2 : assert_ne!(attached_node_id, secondary_node_id);
1390 :
1391 : // Notifying the attached node is offline should demote it to a secondary
1392 2 : let changed = tenant_shard
1393 2 : .intent
1394 2 : .demote_attached(&mut scheduler, attached_node_id);
1395 2 : assert!(changed);
1396 2 : assert!(tenant_shard.intent.attached.is_none());
1397 2 : assert_eq!(tenant_shard.intent.secondary.len(), 2);
1398 :
1399 : // Update the scheduler state to indicate the node is offline
1400 2 : nodes
1401 2 : .get_mut(&attached_node_id)
1402 2 : .unwrap()
1403 2 : .set_availability(NodeAvailability::Offline);
1404 2 : scheduler.node_upsert(nodes.get(&attached_node_id).unwrap());
1405 2 :
1406 2 : // Scheduling the node should promote the still-available secondary node to attached
1407 2 : tenant_shard
1408 2 : .schedule(&mut scheduler, &mut context)
1409 2 : .expect("active nodes are available");
1410 2 : assert_eq!(tenant_shard.intent.attached.unwrap(), secondary_node_id);
1411 :
1412 : // The original attached node should have been retained as a secondary
1413 2 : assert_eq!(
1414 2 : *tenant_shard.intent.secondary.iter().last().unwrap(),
1415 2 : attached_node_id
1416 2 : );
1417 :
1418 2 : tenant_shard.intent.clear(&mut scheduler);
1419 2 :
1420 2 : Ok(())
1421 2 : }
1422 :
1423 : #[test]
1424 2 : fn intent_from_observed() -> anyhow::Result<()> {
1425 2 : let nodes = make_test_nodes(3);
1426 2 : let mut scheduler = Scheduler::new(nodes.values());
1427 2 :
1428 2 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
1429 2 :
1430 2 : tenant_shard.observed.locations.insert(
1431 2 : NodeId(3),
1432 2 : ObservedStateLocation {
1433 2 : conf: Some(LocationConfig {
1434 2 : mode: LocationConfigMode::AttachedMulti,
1435 2 : generation: Some(2),
1436 2 : secondary_conf: None,
1437 2 : shard_number: tenant_shard.shard.number.0,
1438 2 : shard_count: tenant_shard.shard.count.literal(),
1439 2 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
1440 2 : tenant_conf: TenantConfig::default(),
1441 2 : }),
1442 2 : },
1443 2 : );
1444 2 :
1445 2 : tenant_shard.observed.locations.insert(
1446 2 : NodeId(2),
1447 2 : ObservedStateLocation {
1448 2 : conf: Some(LocationConfig {
1449 2 : mode: LocationConfigMode::AttachedStale,
1450 2 : generation: Some(1),
1451 2 : secondary_conf: None,
1452 2 : shard_number: tenant_shard.shard.number.0,
1453 2 : shard_count: tenant_shard.shard.count.literal(),
1454 2 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
1455 2 : tenant_conf: TenantConfig::default(),
1456 2 : }),
1457 2 : },
1458 2 : );
1459 2 :
1460 2 : tenant_shard.intent_from_observed(&mut scheduler);
1461 2 :
1462 2 : // The highest generationed attached location gets used as attached
1463 2 : assert_eq!(tenant_shard.intent.attached, Some(NodeId(3)));
1464 : // Other locations get used as secondary
1465 2 : assert_eq!(tenant_shard.intent.secondary, vec![NodeId(2)]);
1466 :
1467 2 : scheduler.consistency_check(nodes.values(), [&tenant_shard].into_iter())?;
1468 :
1469 2 : tenant_shard.intent.clear(&mut scheduler);
1470 2 : Ok(())
1471 2 : }
1472 :
1473 : #[test]
1474 2 : fn scheduling_mode() -> anyhow::Result<()> {
1475 2 : let nodes = make_test_nodes(3);
1476 2 : let mut scheduler = Scheduler::new(nodes.values());
1477 2 :
1478 2 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
1479 2 :
1480 2 : // In pause mode, schedule() shouldn't do anything
1481 2 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Pause;
1482 2 : assert!(tenant_shard
1483 2 : .schedule(&mut scheduler, &mut ScheduleContext::default())
1484 2 : .is_ok());
1485 2 : assert!(tenant_shard.intent.all_pageservers().is_empty());
1486 :
1487 : // In active mode, schedule() works
1488 2 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Active;
1489 2 : assert!(tenant_shard
1490 2 : .schedule(&mut scheduler, &mut ScheduleContext::default())
1491 2 : .is_ok());
1492 2 : assert!(!tenant_shard.intent.all_pageservers().is_empty());
1493 :
1494 2 : tenant_shard.intent.clear(&mut scheduler);
1495 2 : Ok(())
1496 2 : }
1497 :
1498 : #[test]
1499 2 : fn optimize_attachment() -> anyhow::Result<()> {
1500 2 : let nodes = make_test_nodes(3);
1501 2 : let mut scheduler = Scheduler::new(nodes.values());
1502 2 :
1503 2 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
1504 2 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
1505 2 :
1506 2 : // Initially: both nodes attached on shard 1, and both have secondary locations
1507 2 : // on different nodes.
1508 2 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
1509 2 : shard_a.intent.push_secondary(&mut scheduler, NodeId(2));
1510 2 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(1)));
1511 2 : shard_b.intent.push_secondary(&mut scheduler, NodeId(3));
1512 2 :
1513 2 : let mut schedule_context = ScheduleContext::default();
1514 2 : schedule_context.avoid(&shard_a.intent.all_pageservers());
1515 2 : schedule_context.push_attached(shard_a.intent.get_attached().unwrap());
1516 2 : schedule_context.avoid(&shard_b.intent.all_pageservers());
1517 2 : schedule_context.push_attached(shard_b.intent.get_attached().unwrap());
1518 2 :
1519 2 : let optimization_a = shard_a.optimize_attachment(&nodes, &schedule_context);
1520 2 :
1521 2 : // Either shard should recognize that it has the option to switch to a secondary location where there
1522 2 : // would be no other shards from the same tenant, and request to do so.
1523 2 : assert_eq!(
1524 2 : optimization_a,
1525 2 : Some(ScheduleOptimization {
1526 2 : sequence: shard_a.sequence,
1527 2 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1528 2 : old_attached_node_id: NodeId(1),
1529 2 : new_attached_node_id: NodeId(2)
1530 2 : })
1531 2 : })
1532 2 : );
1533 :
1534 : // Note that these optimizing two shards in the same tenant with the same ScheduleContext is
1535 : // mutually exclusive (the optimization of one invalidates the stats) -- it is the responsibility
1536 : // of [`Service::optimize_all`] to avoid trying
1537 : // to do optimizations for multiple shards in the same tenant at the same time. Generating
1538 : // both optimizations is just done for test purposes
1539 2 : let optimization_b = shard_b.optimize_attachment(&nodes, &schedule_context);
1540 2 : assert_eq!(
1541 2 : optimization_b,
1542 2 : Some(ScheduleOptimization {
1543 2 : sequence: shard_b.sequence,
1544 2 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1545 2 : old_attached_node_id: NodeId(1),
1546 2 : new_attached_node_id: NodeId(3)
1547 2 : })
1548 2 : })
1549 2 : );
1550 :
1551 : // Applying these optimizations should result in the end state proposed
1552 2 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
1553 2 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(2)));
1554 2 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(1)]);
1555 2 : shard_b.apply_optimization(&mut scheduler, optimization_b.unwrap());
1556 2 : assert_eq!(shard_b.intent.get_attached(), &Some(NodeId(3)));
1557 2 : assert_eq!(shard_b.intent.get_secondary(), &vec![NodeId(1)]);
1558 :
1559 2 : shard_a.intent.clear(&mut scheduler);
1560 2 : shard_b.intent.clear(&mut scheduler);
1561 2 :
1562 2 : Ok(())
1563 2 : }
1564 :
1565 : #[test]
1566 2 : fn optimize_secondary() -> anyhow::Result<()> {
1567 2 : let nodes = make_test_nodes(4);
1568 2 : let mut scheduler = Scheduler::new(nodes.values());
1569 2 :
1570 2 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
1571 2 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
1572 2 :
1573 2 : // Initially: both nodes attached on shard 1, and both have secondary locations
1574 2 : // on different nodes.
1575 2 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
1576 2 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
1577 2 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(2)));
1578 2 : shard_b.intent.push_secondary(&mut scheduler, NodeId(3));
1579 2 :
1580 2 : let mut schedule_context = ScheduleContext::default();
1581 2 : schedule_context.avoid(&shard_a.intent.all_pageservers());
1582 2 : schedule_context.push_attached(shard_a.intent.get_attached().unwrap());
1583 2 : schedule_context.avoid(&shard_b.intent.all_pageservers());
1584 2 : schedule_context.push_attached(shard_b.intent.get_attached().unwrap());
1585 2 :
1586 2 : let optimization_a = shard_a.optimize_secondary(&scheduler, &schedule_context);
1587 2 :
1588 2 : // Since there is a node with no locations available, the node with two locations for the
1589 2 : // same tenant should generate an optimization to move one away
1590 2 : assert_eq!(
1591 2 : optimization_a,
1592 2 : Some(ScheduleOptimization {
1593 2 : sequence: shard_a.sequence,
1594 2 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1595 2 : old_node_id: NodeId(3),
1596 2 : new_node_id: NodeId(4)
1597 2 : })
1598 2 : })
1599 2 : );
1600 :
1601 2 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
1602 2 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
1603 2 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(4)]);
1604 :
1605 2 : shard_a.intent.clear(&mut scheduler);
1606 2 : shard_b.intent.clear(&mut scheduler);
1607 2 :
1608 2 : Ok(())
1609 2 : }
1610 :
1611 : // Optimize til quiescent: this emulates what Service::optimize_all does, when
1612 : // called repeatedly in the background.
1613 2 : fn optimize_til_idle(
1614 2 : nodes: &HashMap<NodeId, Node>,
1615 2 : scheduler: &mut Scheduler,
1616 2 : shards: &mut [TenantShard],
1617 2 : ) {
1618 2 : let mut loop_n = 0;
1619 : loop {
1620 14 : let mut schedule_context = ScheduleContext::default();
1621 14 : let mut any_changed = false;
1622 :
1623 56 : for shard in shards.iter() {
1624 56 : schedule_context.avoid(&shard.intent.all_pageservers());
1625 56 : if let Some(attached) = shard.intent.get_attached() {
1626 56 : schedule_context.push_attached(*attached);
1627 56 : }
1628 : }
1629 :
1630 26 : for shard in shards.iter_mut() {
1631 26 : let optimization = shard.optimize_attachment(nodes, &schedule_context);
1632 26 : if let Some(optimization) = optimization {
1633 4 : shard.apply_optimization(scheduler, optimization);
1634 4 : any_changed = true;
1635 4 : break;
1636 22 : }
1637 22 :
1638 22 : let optimization = shard.optimize_secondary(scheduler, &schedule_context);
1639 22 : if let Some(optimization) = optimization {
1640 8 : shard.apply_optimization(scheduler, optimization);
1641 8 : any_changed = true;
1642 8 : break;
1643 14 : }
1644 : }
1645 :
1646 14 : if !any_changed {
1647 2 : break;
1648 12 : }
1649 12 :
1650 12 : // Assert no infinite loop
1651 12 : loop_n += 1;
1652 12 : assert!(loop_n < 1000);
1653 : }
1654 2 : }
1655 :
1656 : /// Test the balancing behavior of shard scheduling: that it achieves a balance, and
1657 : /// that it converges.
1658 : #[test]
1659 2 : fn optimize_add_nodes() -> anyhow::Result<()> {
1660 2 : let nodes = make_test_nodes(4);
1661 2 :
1662 2 : // Only show the scheduler a couple of nodes
1663 2 : let mut scheduler = Scheduler::new([].iter());
1664 2 : scheduler.node_upsert(nodes.get(&NodeId(1)).unwrap());
1665 2 : scheduler.node_upsert(nodes.get(&NodeId(2)).unwrap());
1666 2 :
1667 2 : let mut shards = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4));
1668 2 : let mut schedule_context = ScheduleContext::default();
1669 10 : for shard in &mut shards {
1670 8 : assert!(shard
1671 8 : .schedule(&mut scheduler, &mut schedule_context)
1672 8 : .is_ok());
1673 : }
1674 :
1675 : // We should see equal number of locations on the two nodes.
1676 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 4);
1677 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 2);
1678 :
1679 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 4);
1680 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
1681 :
1682 : // Add another two nodes: we should see the shards spread out when their optimize
1683 : // methods are called
1684 2 : scheduler.node_upsert(nodes.get(&NodeId(3)).unwrap());
1685 2 : scheduler.node_upsert(nodes.get(&NodeId(4)).unwrap());
1686 2 : optimize_til_idle(&nodes, &mut scheduler, &mut shards);
1687 2 :
1688 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 2);
1689 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 1);
1690 :
1691 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
1692 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 1);
1693 :
1694 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 2);
1695 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 1);
1696 :
1697 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 2);
1698 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 1);
1699 :
1700 8 : for shard in shards.iter_mut() {
1701 8 : shard.intent.clear(&mut scheduler);
1702 8 : }
1703 :
1704 2 : Ok(())
1705 2 : }
1706 : }
|