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 : service::ReconcileResultRequest,
13 : };
14 : use pageserver_api::controller_api::{
15 : NodeSchedulingPolicy, PlacementPolicy, ShardSchedulingPolicy,
16 : };
17 : use pageserver_api::{
18 : models::{LocationConfig, LocationConfigMode, TenantConfig},
19 : shard::{ShardIdentity, TenantShardId},
20 : };
21 : use serde::{Deserialize, Serialize};
22 : use tokio::task::JoinHandle;
23 : use tokio_util::sync::CancellationToken;
24 : use tracing::{instrument, Instrument};
25 : use utils::{
26 : generation::Generation,
27 : id::NodeId,
28 : seqwait::{SeqWait, SeqWaitError},
29 : sync::gate::GateGuard,
30 : };
31 :
32 : use crate::{
33 : compute_hook::ComputeHook,
34 : node::Node,
35 : persistence::{split_state::SplitState, Persistence},
36 : reconciler::{
37 : attached_location_conf, secondary_location_conf, ReconcileError, Reconciler, TargetState,
38 : },
39 : scheduler::{ScheduleError, Scheduler},
40 : service, Sequence,
41 : };
42 :
43 : /// Serialization helper
44 0 : fn read_last_error<S, T>(v: &std::sync::Mutex<Option<T>>, serializer: S) -> Result<S::Ok, S::Error>
45 0 : where
46 0 : S: serde::ser::Serializer,
47 0 : T: std::fmt::Display,
48 0 : {
49 0 : serializer.collect_str(
50 0 : &v.lock()
51 0 : .unwrap()
52 0 : .as_ref()
53 0 : .map(|e| format!("{e}"))
54 0 : .unwrap_or("".to_string()),
55 0 : )
56 0 : }
57 :
58 : /// In-memory state for a particular tenant shard.
59 : ///
60 : /// This struct implement Serialize for debugging purposes, but is _not_ persisted
61 : /// itself: see [`crate::persistence`] for the subset of tenant shard state that is persisted.
62 0 : #[derive(Serialize)]
63 : pub(crate) struct TenantShard {
64 : pub(crate) tenant_shard_id: TenantShardId,
65 :
66 : pub(crate) shard: ShardIdentity,
67 :
68 : // Runtime only: sequence used to coordinate when updating this object while
69 : // with background reconcilers may be running. A reconciler runs to a particular
70 : // sequence.
71 : pub(crate) sequence: Sequence,
72 :
73 : // Latest generation number: next time we attach, increment this
74 : // and use the incremented number when attaching.
75 : //
76 : // None represents an incompletely onboarded tenant via the [`Service::location_config`]
77 : // API, where this tenant may only run in PlacementPolicy::Secondary.
78 : pub(crate) generation: Option<Generation>,
79 :
80 : // High level description of how the tenant should be set up. Provided
81 : // externally.
82 : pub(crate) policy: PlacementPolicy,
83 :
84 : // Low level description of exactly which pageservers should fulfil
85 : // which role. Generated by `Self::schedule`.
86 : pub(crate) intent: IntentState,
87 :
88 : // Low level description of how the tenant is configured on pageservers:
89 : // if this does not match `Self::intent` then the tenant needs reconciliation
90 : // with `Self::reconcile`.
91 : pub(crate) observed: ObservedState,
92 :
93 : // Tenant configuration, passed through opaquely to the pageserver. Identical
94 : // for all shards in a tenant.
95 : pub(crate) config: TenantConfig,
96 :
97 : /// If a reconcile task is currently in flight, it may be joined here (it is
98 : /// only safe to join if either the result has been received or the reconciler's
99 : /// cancellation token has been fired)
100 : #[serde(skip)]
101 : pub(crate) reconciler: Option<ReconcilerHandle>,
102 :
103 : /// If a tenant is being split, then all shards with that TenantId will have a
104 : /// SplitState set, this acts as a guard against other operations such as background
105 : /// reconciliation, and timeline creation.
106 : pub(crate) splitting: SplitState,
107 :
108 : /// If a tenant was enqueued for later reconcile due to hitting concurrency limit, this flag
109 : /// is set. This flag is cleared when the tenant is popped off the delay queue.
110 : pub(crate) delayed_reconcile: bool,
111 :
112 : /// Optionally wait for reconciliation to complete up to a particular
113 : /// sequence number.
114 : #[serde(skip)]
115 : pub(crate) waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
116 :
117 : /// Indicates sequence number for which we have encountered an error reconciling. If
118 : /// this advances ahead of [`Self::waiter`] then a reconciliation error has occurred,
119 : /// and callers should stop waiting for `waiter` and propagate the error.
120 : #[serde(skip)]
121 : pub(crate) error_waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
122 :
123 : /// The most recent error from a reconcile on this tenant. This is a nested Arc
124 : /// because:
125 : /// - ReconcileWaiters need to Arc-clone the overall object to read it later
126 : /// - ReconcileWaitError needs to use an `Arc<ReconcileError>` because we can construct
127 : /// many waiters for one shard, and the underlying error types are not Clone.
128 : ///
129 : /// TODO: generalize to an array of recent events
130 : /// TOOD: use a ArcSwap instead of mutex for faster reads?
131 : #[serde(serialize_with = "read_last_error")]
132 : pub(crate) last_error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
133 :
134 : /// If we have a pending compute notification that for some reason we weren't able to send,
135 : /// set this to true. If this is set, calls to [`Self::get_reconcile_needed`] will return Yes
136 : /// and trigger a Reconciler run. This is the mechanism by which compute notifications are included in the scope
137 : /// of state that we publish externally in an eventually consistent way.
138 : pub(crate) pending_compute_notification: bool,
139 :
140 : // Support/debug tool: if something is going wrong or flapping with scheduling, this may
141 : // be set to a non-active state to avoid making changes while the issue is fixed.
142 : scheduling_policy: ShardSchedulingPolicy,
143 : }
144 :
145 : #[derive(Default, Clone, Debug, Serialize)]
146 : pub(crate) struct IntentState {
147 : attached: Option<NodeId>,
148 : secondary: Vec<NodeId>,
149 : }
150 :
151 : impl IntentState {
152 4 : pub(crate) fn new() -> Self {
153 4 : Self {
154 4 : attached: None,
155 4 : secondary: vec![],
156 4 : }
157 4 : }
158 0 : pub(crate) fn single(scheduler: &mut Scheduler, node_id: Option<NodeId>) -> Self {
159 0 : if let Some(node_id) = node_id {
160 0 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::Attach);
161 0 : }
162 0 : Self {
163 0 : attached: node_id,
164 0 : secondary: vec![],
165 0 : }
166 0 : }
167 :
168 26 : pub(crate) fn set_attached(&mut self, scheduler: &mut Scheduler, new_attached: Option<NodeId>) {
169 26 : if self.attached != new_attached {
170 26 : if let Some(old_attached) = self.attached.take() {
171 0 : scheduler.update_node_ref_counts(old_attached, RefCountUpdate::Detach);
172 26 : }
173 26 : if let Some(new_attached) = &new_attached {
174 26 : scheduler.update_node_ref_counts(*new_attached, RefCountUpdate::Attach);
175 26 : }
176 26 : self.attached = new_attached;
177 0 : }
178 26 : }
179 :
180 : /// Like set_attached, but the node is from [`Self::secondary`]. This swaps the node from
181 : /// secondary to attached while maintaining the scheduler's reference counts.
182 10 : pub(crate) fn promote_attached(
183 10 : &mut self,
184 10 : scheduler: &mut Scheduler,
185 10 : promote_secondary: NodeId,
186 10 : ) {
187 10 : // If we call this with a node that isn't in secondary, it would cause incorrect
188 10 : // scheduler reference counting, since we assume the node is already referenced as a secondary.
189 10 : debug_assert!(self.secondary.contains(&promote_secondary));
190 :
191 20 : self.secondary.retain(|n| n != &promote_secondary);
192 10 :
193 10 : let demoted = self.attached;
194 10 : self.attached = Some(promote_secondary);
195 10 :
196 10 : scheduler.update_node_ref_counts(promote_secondary, RefCountUpdate::PromoteSecondary);
197 10 : if let Some(demoted) = demoted {
198 0 : scheduler.update_node_ref_counts(demoted, RefCountUpdate::DemoteAttached);
199 10 : }
200 10 : }
201 :
202 34 : pub(crate) fn push_secondary(&mut self, scheduler: &mut Scheduler, new_secondary: NodeId) {
203 34 : debug_assert!(!self.secondary.contains(&new_secondary));
204 34 : scheduler.update_node_ref_counts(new_secondary, RefCountUpdate::AddSecondary);
205 34 : self.secondary.push(new_secondary);
206 34 : }
207 :
208 : /// It is legal to call this with a node that is not currently a secondary: that is a no-op
209 10 : pub(crate) fn remove_secondary(&mut self, scheduler: &mut Scheduler, node_id: NodeId) {
210 10 : let index = self.secondary.iter().position(|n| *n == node_id);
211 10 : if let Some(index) = index {
212 10 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::RemoveSecondary);
213 10 : self.secondary.remove(index);
214 10 : }
215 10 : }
216 :
217 24 : pub(crate) fn clear_secondary(&mut self, scheduler: &mut Scheduler) {
218 24 : for secondary in self.secondary.drain(..) {
219 24 : scheduler.update_node_ref_counts(secondary, RefCountUpdate::RemoveSecondary);
220 24 : }
221 24 : }
222 :
223 : /// Remove the last secondary node from the list of secondaries
224 0 : pub(crate) fn pop_secondary(&mut self, scheduler: &mut Scheduler) {
225 0 : if let Some(node_id) = self.secondary.pop() {
226 0 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::RemoveSecondary);
227 0 : }
228 0 : }
229 :
230 24 : pub(crate) fn clear(&mut self, scheduler: &mut Scheduler) {
231 24 : if let Some(old_attached) = self.attached.take() {
232 24 : scheduler.update_node_ref_counts(old_attached, RefCountUpdate::Detach);
233 24 : }
234 :
235 24 : self.clear_secondary(scheduler);
236 24 : }
237 :
238 140 : pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
239 140 : let mut result = Vec::new();
240 140 : if let Some(p) = self.attached {
241 136 : result.push(p)
242 4 : }
243 :
244 140 : result.extend(self.secondary.iter().copied());
245 140 :
246 140 : result
247 140 : }
248 :
249 118 : pub(crate) fn get_attached(&self) -> &Option<NodeId> {
250 118 : &self.attached
251 118 : }
252 :
253 32 : pub(crate) fn get_secondary(&self) -> &Vec<NodeId> {
254 32 : &self.secondary
255 32 : }
256 :
257 : /// If the node is in use as the attached location, demote it into
258 : /// the list of secondary locations. This is used when a node goes offline,
259 : /// and we want to use a different node for attachment, but not permanently
260 : /// forget the location on the offline node.
261 : ///
262 : /// Returns true if a change was made
263 10 : pub(crate) fn demote_attached(&mut self, scheduler: &mut Scheduler, node_id: NodeId) -> bool {
264 10 : if self.attached == Some(node_id) {
265 10 : self.attached = None;
266 10 : self.secondary.push(node_id);
267 10 : scheduler.update_node_ref_counts(node_id, RefCountUpdate::DemoteAttached);
268 10 : true
269 : } else {
270 0 : false
271 : }
272 10 : }
273 : }
274 :
275 : impl Drop for IntentState {
276 26 : fn drop(&mut self) {
277 26 : // Must clear before dropping, to avoid leaving stale refcounts in the Scheduler.
278 26 : // We do not check this while panicking, to avoid polluting unit test failures or
279 26 : // other assertions with this assertion's output. It's still wrong to leak these,
280 26 : // but if we already have a panic then we don't need to independently flag this case.
281 26 : if !(std::thread::panicking()) {
282 26 : debug_assert!(self.attached.is_none() && self.secondary.is_empty());
283 0 : }
284 24 : }
285 : }
286 :
287 0 : #[derive(Default, Clone, Serialize, Deserialize, Debug)]
288 : pub(crate) struct ObservedState {
289 : pub(crate) locations: HashMap<NodeId, ObservedStateLocation>,
290 : }
291 :
292 : /// Our latest knowledge of how this tenant is configured in the outside world.
293 : ///
294 : /// Meaning:
295 : /// * No instance of this type exists for a node: we are certain that we have nothing configured on that
296 : /// node for this shard.
297 : /// * Instance exists with conf==None: we *might* have some state on that node, but we don't know
298 : /// what it is (e.g. we failed partway through configuring it)
299 : /// * Instance exists with conf==Some: this tells us what we last successfully configured on this node,
300 : /// and that configuration will still be present unless something external interfered.
301 0 : #[derive(Clone, Serialize, Deserialize, Debug)]
302 : pub(crate) struct ObservedStateLocation {
303 : /// If None, it means we do not know the status of this shard's location on this node, but
304 : /// we know that we might have some state on this node.
305 : pub(crate) conf: Option<LocationConfig>,
306 : }
307 : pub(crate) struct ReconcilerWaiter {
308 : // For observability purposes, remember the ID of the shard we're
309 : // waiting for.
310 : pub(crate) tenant_shard_id: TenantShardId,
311 :
312 : seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
313 : error_seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
314 : error: std::sync::Arc<std::sync::Mutex<Option<Arc<ReconcileError>>>>,
315 : seq: Sequence,
316 : }
317 :
318 : pub(crate) enum ReconcilerStatus {
319 : Done,
320 : Failed,
321 : InProgress,
322 : }
323 :
324 0 : #[derive(thiserror::Error, Debug)]
325 : pub(crate) enum ReconcileWaitError {
326 : #[error("Timeout waiting for shard {0}")]
327 : Timeout(TenantShardId),
328 : #[error("shutting down")]
329 : Shutdown,
330 : #[error("Reconcile error on shard {0}: {1}")]
331 : Failed(TenantShardId, Arc<ReconcileError>),
332 : }
333 :
334 : #[derive(Eq, PartialEq, Debug)]
335 : pub(crate) struct ReplaceSecondary {
336 : old_node_id: NodeId,
337 : new_node_id: NodeId,
338 : }
339 :
340 : #[derive(Eq, PartialEq, Debug)]
341 : pub(crate) struct MigrateAttachment {
342 : pub(crate) old_attached_node_id: NodeId,
343 : pub(crate) new_attached_node_id: NodeId,
344 : }
345 :
346 : #[derive(Eq, PartialEq, Debug)]
347 : pub(crate) enum ScheduleOptimizationAction {
348 : // Replace one of our secondary locations with a different node
349 : ReplaceSecondary(ReplaceSecondary),
350 : // Migrate attachment to an existing secondary location
351 : MigrateAttachment(MigrateAttachment),
352 : }
353 :
354 : #[derive(Eq, PartialEq, Debug)]
355 : pub(crate) struct ScheduleOptimization {
356 : // What was the reconcile sequence when we generated this optimization? The optimization
357 : // should only be applied if the shard's sequence is still at this value, in case other changes
358 : // happened between planning the optimization and applying it.
359 : sequence: Sequence,
360 :
361 : pub(crate) action: ScheduleOptimizationAction,
362 : }
363 :
364 : impl ReconcilerWaiter {
365 0 : pub(crate) async fn wait_timeout(&self, timeout: Duration) -> Result<(), ReconcileWaitError> {
366 : tokio::select! {
367 : result = self.seq_wait.wait_for_timeout(self.seq, timeout)=> {
368 0 : result.map_err(|e| match e {
369 0 : SeqWaitError::Timeout => ReconcileWaitError::Timeout(self.tenant_shard_id),
370 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown
371 0 : })?;
372 : },
373 : result = self.error_seq_wait.wait_for(self.seq) => {
374 0 : result.map_err(|e| match e {
375 0 : SeqWaitError::Shutdown => ReconcileWaitError::Shutdown,
376 0 : SeqWaitError::Timeout => unreachable!()
377 0 : })?;
378 :
379 : return Err(ReconcileWaitError::Failed(self.tenant_shard_id,
380 : self.error.lock().unwrap().clone().expect("If error_seq_wait was advanced error was set").clone()))
381 : }
382 : }
383 :
384 0 : Ok(())
385 0 : }
386 :
387 0 : pub(crate) fn get_status(&self) -> ReconcilerStatus {
388 0 : if self.seq_wait.would_wait_for(self.seq).is_ok() {
389 0 : ReconcilerStatus::Done
390 0 : } else if self.error_seq_wait.would_wait_for(self.seq).is_ok() {
391 0 : ReconcilerStatus::Failed
392 : } else {
393 0 : ReconcilerStatus::InProgress
394 : }
395 0 : }
396 : }
397 :
398 : /// Having spawned a reconciler task, the tenant shard's state will carry enough
399 : /// information to optionally cancel & await it later.
400 : pub(crate) struct ReconcilerHandle {
401 : sequence: Sequence,
402 : handle: JoinHandle<()>,
403 : cancel: CancellationToken,
404 : }
405 :
406 : pub(crate) enum ReconcileNeeded {
407 : /// shard either doesn't need reconciliation, or is forbidden from spawning a reconciler
408 : /// in its current state (e.g. shard split in progress, or ShardSchedulingPolicy forbids it)
409 : No,
410 : /// shard has a reconciler running, and its intent hasn't changed since that one was
411 : /// spawned: wait for the existing reconciler rather than spawning a new one.
412 : WaitExisting(ReconcilerWaiter),
413 : /// shard needs reconciliation: call into [`TenantShard::spawn_reconciler`]
414 : Yes,
415 : }
416 :
417 : /// When a reconcile task completes, it sends this result object
418 : /// to be applied to the primary TenantShard.
419 : pub(crate) struct ReconcileResult {
420 : pub(crate) sequence: Sequence,
421 : /// On errors, `observed` should be treated as an incompleted description
422 : /// of state (i.e. any nodes present in the result should override nodes
423 : /// present in the parent tenant state, but any unmentioned nodes should
424 : /// not be removed from parent tenant state)
425 : pub(crate) result: Result<(), ReconcileError>,
426 :
427 : pub(crate) tenant_shard_id: TenantShardId,
428 : pub(crate) generation: Option<Generation>,
429 : pub(crate) observed: ObservedState,
430 :
431 : /// Set [`TenantShard::pending_compute_notification`] from this flag
432 : pub(crate) pending_compute_notification: bool,
433 : }
434 :
435 : impl ObservedState {
436 0 : pub(crate) fn new() -> Self {
437 0 : Self {
438 0 : locations: HashMap::new(),
439 0 : }
440 0 : }
441 : }
442 :
443 : impl TenantShard {
444 22 : pub(crate) fn new(
445 22 : tenant_shard_id: TenantShardId,
446 22 : shard: ShardIdentity,
447 22 : policy: PlacementPolicy,
448 22 : ) -> Self {
449 22 : Self {
450 22 : tenant_shard_id,
451 22 : policy,
452 22 : intent: IntentState::default(),
453 22 : generation: Some(Generation::new(0)),
454 22 : shard,
455 22 : observed: ObservedState::default(),
456 22 : config: TenantConfig::default(),
457 22 : reconciler: None,
458 22 : splitting: SplitState::Idle,
459 22 : sequence: Sequence(1),
460 22 : delayed_reconcile: false,
461 22 : waiter: Arc::new(SeqWait::new(Sequence(0))),
462 22 : error_waiter: Arc::new(SeqWait::new(Sequence(0))),
463 22 : last_error: Arc::default(),
464 22 : pending_compute_notification: false,
465 22 : scheduling_policy: ShardSchedulingPolicy::default(),
466 22 : }
467 22 : }
468 :
469 : /// For use on startup when learning state from pageservers: generate my [`IntentState`] from my
470 : /// [`ObservedState`], even if it violates my [`PlacementPolicy`]. Call [`Self::schedule`] next,
471 : /// to get an intent state that complies with placement policy. The overall goal is to do scheduling
472 : /// in a way that makes use of any configured locations that already exist in the outside world.
473 2 : pub(crate) fn intent_from_observed(&mut self, scheduler: &mut Scheduler) {
474 2 : // Choose an attached location by filtering observed locations, and then sorting to get the highest
475 2 : // generation
476 2 : let mut attached_locs = self
477 2 : .observed
478 2 : .locations
479 2 : .iter()
480 4 : .filter_map(|(node_id, l)| {
481 4 : if let Some(conf) = &l.conf {
482 4 : if conf.mode == LocationConfigMode::AttachedMulti
483 2 : || conf.mode == LocationConfigMode::AttachedSingle
484 2 : || conf.mode == LocationConfigMode::AttachedStale
485 : {
486 4 : Some((node_id, conf.generation))
487 : } else {
488 0 : None
489 : }
490 : } else {
491 0 : None
492 : }
493 4 : })
494 2 : .collect::<Vec<_>>();
495 2 :
496 4 : attached_locs.sort_by_key(|i| i.1);
497 2 : if let Some((node_id, _gen)) = attached_locs.into_iter().last() {
498 2 : self.intent.set_attached(scheduler, Some(*node_id));
499 2 : }
500 :
501 : // All remaining observed locations generate secondary intents. This includes None
502 : // observations, as these may well have some local content on disk that is usable (this
503 : // is an edge case that might occur if we restarted during a migration or other change)
504 : //
505 : // We may leave intent.attached empty if we didn't find any attached locations: [`Self::schedule`]
506 : // will take care of promoting one of these secondaries to be attached.
507 4 : self.observed.locations.keys().for_each(|node_id| {
508 4 : if Some(*node_id) != self.intent.attached {
509 2 : self.intent.push_secondary(scheduler, *node_id);
510 2 : }
511 4 : });
512 2 : }
513 :
514 : /// Part of [`Self::schedule`] that is used to choose exactly one node to act as the
515 : /// attached pageserver for a shard.
516 : ///
517 : /// Returns whether we modified it, and the NodeId selected.
518 14 : fn schedule_attached(
519 14 : &mut self,
520 14 : scheduler: &mut Scheduler,
521 14 : context: &ScheduleContext,
522 14 : ) -> Result<(bool, NodeId), ScheduleError> {
523 : // No work to do if we already have an attached tenant
524 14 : if let Some(node_id) = self.intent.attached {
525 0 : return Ok((false, node_id));
526 14 : }
527 :
528 14 : if let Some(promote_secondary) = scheduler.node_preferred(&self.intent.secondary) {
529 : // Promote a secondary
530 2 : tracing::debug!("Promoted secondary {} to attached", promote_secondary);
531 2 : self.intent.promote_attached(scheduler, promote_secondary);
532 2 : Ok((true, promote_secondary))
533 : } else {
534 : // Pick a fresh node: either we had no secondaries or none were schedulable
535 12 : let node_id = scheduler.schedule_shard(&self.intent.secondary, context)?;
536 12 : tracing::debug!("Selected {} as attached", node_id);
537 12 : self.intent.set_attached(scheduler, Some(node_id));
538 12 : Ok((true, node_id))
539 : }
540 14 : }
541 :
542 16 : pub(crate) fn schedule(
543 16 : &mut self,
544 16 : scheduler: &mut Scheduler,
545 16 : context: &mut ScheduleContext,
546 16 : ) -> Result<(), ScheduleError> {
547 16 : let r = self.do_schedule(scheduler, context);
548 16 :
549 16 : context.avoid(&self.intent.all_pageservers());
550 16 : if let Some(attached) = self.intent.get_attached() {
551 14 : context.push_attached(*attached);
552 14 : }
553 :
554 16 : r
555 16 : }
556 :
557 16 : pub(crate) fn do_schedule(
558 16 : &mut self,
559 16 : scheduler: &mut Scheduler,
560 16 : context: &ScheduleContext,
561 16 : ) -> Result<(), ScheduleError> {
562 16 : // TODO: before scheduling new nodes, check if any existing content in
563 16 : // self.intent refers to pageservers that are offline, and pick other
564 16 : // pageservers if so.
565 16 :
566 16 : // TODO: respect the splitting bit on tenants: if they are currently splitting then we may not
567 16 : // change their attach location.
568 16 :
569 16 : match self.scheduling_policy {
570 14 : ShardSchedulingPolicy::Active | ShardSchedulingPolicy::Essential => {}
571 : ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
572 : // Warn to make it obvious why other things aren't happening/working, if we skip scheduling
573 2 : tracing::warn!(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(),
574 0 : "Scheduling is disabled by policy {:?}", self.scheduling_policy);
575 2 : return Ok(());
576 : }
577 : }
578 :
579 : // Build the set of pageservers already in use by this tenant, to avoid scheduling
580 : // more work on the same pageservers we're already using.
581 14 : let mut modified = false;
582 14 :
583 14 : // Add/remove nodes to fulfil policy
584 14 : use PlacementPolicy::*;
585 14 : match self.policy {
586 14 : Attached(secondary_count) => {
587 14 : let retain_secondaries = if self.intent.attached.is_none()
588 14 : && scheduler.node_preferred(&self.intent.secondary).is_some()
589 : {
590 : // If we have no attached, and one of the secondaries is elegible to be promoted, retain
591 : // one more secondary than we usually would, as one of them will become attached futher down this function.
592 2 : secondary_count + 1
593 : } else {
594 12 : secondary_count
595 : };
596 :
597 14 : while self.intent.secondary.len() > retain_secondaries {
598 0 : // We have no particular preference for one secondary location over another: just
599 0 : // arbitrarily drop from the end
600 0 : self.intent.pop_secondary(scheduler);
601 0 : modified = true;
602 0 : }
603 :
604 : // Should have exactly one attached, and N secondaries
605 14 : let (modified_attached, attached_node_id) =
606 14 : self.schedule_attached(scheduler, context)?;
607 14 : modified |= modified_attached;
608 14 :
609 14 : let mut used_pageservers = vec![attached_node_id];
610 26 : while self.intent.secondary.len() < secondary_count {
611 12 : let node_id = scheduler.schedule_shard(&used_pageservers, context)?;
612 12 : self.intent.push_secondary(scheduler, node_id);
613 12 : used_pageservers.push(node_id);
614 12 : modified = true;
615 : }
616 : }
617 : Secondary => {
618 0 : if let Some(node_id) = self.intent.get_attached() {
619 0 : // Populate secondary by demoting the attached node
620 0 : self.intent.demote_attached(scheduler, *node_id);
621 0 : modified = true;
622 0 : } else if self.intent.secondary.is_empty() {
623 0 : // Populate secondary by scheduling a fresh node
624 0 : let node_id = scheduler.schedule_shard(&[], context)?;
625 0 : self.intent.push_secondary(scheduler, node_id);
626 0 : modified = true;
627 0 : }
628 0 : while self.intent.secondary.len() > 1 {
629 0 : // We have no particular preference for one secondary location over another: just
630 0 : // arbitrarily drop from the end
631 0 : self.intent.pop_secondary(scheduler);
632 0 : modified = true;
633 0 : }
634 : }
635 : Detached => {
636 : // Never add locations in this mode
637 0 : if self.intent.get_attached().is_some() || !self.intent.get_secondary().is_empty() {
638 0 : self.intent.clear(scheduler);
639 0 : modified = true;
640 0 : }
641 : }
642 : }
643 :
644 14 : if modified {
645 14 : self.sequence.0 += 1;
646 14 : }
647 :
648 14 : Ok(())
649 16 : }
650 :
651 : /// Reschedule this tenant shard to one of its secondary locations. Returns a scheduling error
652 : /// if the swap is not possible and leaves the intent state in its original state.
653 : ///
654 : /// Arguments:
655 : /// `attached_to`: the currently attached location matching the intent state (may be None if the
656 : /// shard is not attached)
657 : /// `promote_to`: an optional secondary location of this tenant shard. If set to None, we ask
658 : /// the scheduler to recommend a node
659 0 : pub(crate) fn reschedule_to_secondary(
660 0 : &mut self,
661 0 : promote_to: Option<NodeId>,
662 0 : scheduler: &mut Scheduler,
663 0 : ) -> Result<(), ScheduleError> {
664 0 : let promote_to = match promote_to {
665 0 : Some(node) => node,
666 0 : None => match scheduler.node_preferred(self.intent.get_secondary()) {
667 0 : Some(node) => node,
668 : None => {
669 0 : return Err(ScheduleError::ImpossibleConstraint);
670 : }
671 : },
672 : };
673 :
674 0 : assert!(self.intent.get_secondary().contains(&promote_to));
675 :
676 0 : if let Some(node) = self.intent.get_attached() {
677 0 : let demoted = self.intent.demote_attached(scheduler, *node);
678 0 : if !demoted {
679 0 : return Err(ScheduleError::ImpossibleConstraint);
680 0 : }
681 0 : }
682 :
683 0 : self.intent.promote_attached(scheduler, promote_to);
684 0 :
685 0 : // Increment the sequence number for the edge case where a
686 0 : // reconciler is already running to avoid waiting on the
687 0 : // current reconcile instead of spawning a new one.
688 0 : self.sequence = self.sequence.next();
689 0 :
690 0 : Ok(())
691 0 : }
692 :
693 : /// Optimize attachments: if a shard has a secondary location that is preferable to
694 : /// its primary location based on soft constraints, switch that secondary location
695 : /// to be attached.
696 30 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
697 : pub(crate) fn optimize_attachment(
698 : &self,
699 : nodes: &HashMap<NodeId, Node>,
700 : schedule_context: &ScheduleContext,
701 : ) -> Option<ScheduleOptimization> {
702 : let attached = (*self.intent.get_attached())?;
703 : if self.intent.secondary.is_empty() {
704 : // We can only do useful work if we have both attached and secondary locations: this
705 : // function doesn't schedule new locations, only swaps between attached and secondaries.
706 : return None;
707 : }
708 :
709 : let current_affinity_score = schedule_context.get_node_affinity(attached);
710 : let current_attachment_count = schedule_context.get_node_attachments(attached);
711 :
712 : // Generate score for each node, dropping any un-schedulable nodes.
713 : let all_pageservers = self.intent.all_pageservers();
714 : let mut scores = all_pageservers
715 : .iter()
716 60 : .flat_map(|node_id| {
717 60 : let node = nodes.get(node_id);
718 60 : if node.is_none() {
719 0 : None
720 60 : } else if matches!(
721 60 : node.unwrap().get_scheduling(),
722 : NodeSchedulingPolicy::Filling
723 : ) {
724 : // If the node is currently filling, don't count it as a candidate to avoid,
725 : // racing with the background fill.
726 0 : None
727 60 : } else if matches!(node.unwrap().may_schedule(), MaySchedule::No) {
728 0 : None
729 : } else {
730 60 : let affinity_score = schedule_context.get_node_affinity(*node_id);
731 60 : let attachment_count = schedule_context.get_node_attachments(*node_id);
732 60 : Some((*node_id, affinity_score, attachment_count))
733 : }
734 60 : })
735 : .collect::<Vec<_>>();
736 :
737 : // Sort precedence:
738 : // 1st - prefer nodes with the lowest total affinity score
739 : // 2nd - prefer nodes with the lowest number of attachments in this context
740 : // 3rd - if all else is equal, sort by node ID for determinism in tests.
741 60 : scores.sort_by_key(|i| (i.1, i.2, i.0));
742 :
743 : if let Some((preferred_node, preferred_affinity_score, preferred_attachment_count)) =
744 : scores.first()
745 : {
746 : if attached != *preferred_node {
747 : // The best alternative must be more than 1 better than us, otherwise we could end
748 : // up flapping back next time we're called (e.g. there's no point migrating from
749 : // a location with score 1 to a score zero, because on next location the situation
750 : // would be the same, but in reverse).
751 : if current_affinity_score > *preferred_affinity_score + AffinityScore(1)
752 : || current_attachment_count > *preferred_attachment_count + 1
753 : {
754 : tracing::info!(
755 : "Identified optimization: migrate attachment {attached}->{preferred_node} (secondaries {:?})",
756 : self.intent.get_secondary()
757 : );
758 : return Some(ScheduleOptimization {
759 : sequence: self.sequence,
760 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
761 : old_attached_node_id: attached,
762 : new_attached_node_id: *preferred_node,
763 : }),
764 : });
765 : }
766 : } else {
767 : tracing::debug!(
768 : "Node {} is already preferred (score {:?})",
769 : preferred_node,
770 : preferred_affinity_score
771 : );
772 : }
773 : }
774 :
775 : // Fall-through: we didn't find an optimization
776 : None
777 : }
778 :
779 24 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
780 : pub(crate) fn optimize_secondary(
781 : &self,
782 : scheduler: &Scheduler,
783 : schedule_context: &ScheduleContext,
784 : ) -> Option<ScheduleOptimization> {
785 : if self.intent.secondary.is_empty() {
786 : // We can only do useful work if we have both attached and secondary locations: this
787 : // function doesn't schedule new locations, only swaps between attached and secondaries.
788 : return None;
789 : }
790 :
791 : for secondary in self.intent.get_secondary() {
792 : let Some(affinity_score) = schedule_context.nodes.get(secondary) else {
793 : // We're already on a node unaffected any affinity constraints,
794 : // so we won't change it.
795 : continue;
796 : };
797 :
798 : // Let the scheduler suggest a node, where it would put us if we were scheduling afresh
799 : // This implicitly limits the choice to nodes that are available, and prefers nodes
800 : // with lower utilization.
801 : let Ok(candidate_node) =
802 : scheduler.schedule_shard(&self.intent.all_pageservers(), schedule_context)
803 : else {
804 : // A scheduling error means we have no possible candidate replacements
805 : continue;
806 : };
807 :
808 : let candidate_affinity_score = schedule_context
809 : .nodes
810 : .get(&candidate_node)
811 : .unwrap_or(&AffinityScore::FREE);
812 :
813 : // The best alternative must be more than 1 better than us, otherwise we could end
814 : // up flapping back next time we're called.
815 : if *candidate_affinity_score + AffinityScore(1) < *affinity_score {
816 : // If some other node is available and has a lower score than this node, then
817 : // that other node is a good place to migrate to.
818 : tracing::info!(
819 : "Identified optimization: replace secondary {secondary}->{candidate_node} (current secondaries {:?})",
820 : self.intent.get_secondary()
821 : );
822 : return Some(ScheduleOptimization {
823 : sequence: self.sequence,
824 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
825 : old_node_id: *secondary,
826 : new_node_id: candidate_node,
827 : }),
828 : });
829 : }
830 : }
831 :
832 : None
833 : }
834 :
835 : /// Return true if the optimization was really applied: it will not be applied if the optimization's
836 : /// sequence is behind this tenant shard's
837 18 : pub(crate) fn apply_optimization(
838 18 : &mut self,
839 18 : scheduler: &mut Scheduler,
840 18 : optimization: ScheduleOptimization,
841 18 : ) -> bool {
842 18 : if optimization.sequence != self.sequence {
843 0 : return false;
844 18 : }
845 18 :
846 18 : metrics::METRICS_REGISTRY
847 18 : .metrics_group
848 18 : .storage_controller_schedule_optimization
849 18 : .inc();
850 18 :
851 18 : match optimization.action {
852 : ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
853 8 : old_attached_node_id,
854 8 : new_attached_node_id,
855 8 : }) => {
856 8 : self.intent.demote_attached(scheduler, old_attached_node_id);
857 8 : self.intent
858 8 : .promote_attached(scheduler, new_attached_node_id);
859 8 : }
860 : ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
861 10 : old_node_id,
862 10 : new_node_id,
863 10 : }) => {
864 10 : self.intent.remove_secondary(scheduler, old_node_id);
865 10 : self.intent.push_secondary(scheduler, new_node_id);
866 10 : }
867 : }
868 :
869 18 : true
870 18 : }
871 :
872 : /// Query whether the tenant's observed state for attached node matches its intent state, and if so,
873 : /// yield the node ID. This is appropriate for emitting compute hook notifications: we are checking that
874 : /// the node in question is not only where we intend to attach, but that the tenant is indeed already attached there.
875 : ///
876 : /// Reconciliation may still be needed for other aspects of state such as secondaries (see [`Self::dirty`]): this
877 : /// funciton should not be used to decide whether to reconcile.
878 0 : pub(crate) fn stably_attached(&self) -> Option<NodeId> {
879 0 : if let Some(attach_intent) = self.intent.attached {
880 0 : match self.observed.locations.get(&attach_intent) {
881 0 : Some(loc) => match &loc.conf {
882 0 : Some(conf) => match conf.mode {
883 : LocationConfigMode::AttachedMulti
884 : | LocationConfigMode::AttachedSingle
885 : | LocationConfigMode::AttachedStale => {
886 : // Our intent and observed state agree that this node is in an attached state.
887 0 : Some(attach_intent)
888 : }
889 : // Our observed config is not an attached state
890 0 : _ => None,
891 : },
892 : // Our observed state is None, i.e. in flux
893 0 : None => None,
894 : },
895 : // We have no observed state for this node
896 0 : None => None,
897 : }
898 : } else {
899 : // Our intent is not to attach
900 0 : None
901 : }
902 0 : }
903 :
904 0 : fn dirty(&self, nodes: &Arc<HashMap<NodeId, Node>>) -> bool {
905 0 : let mut dirty_nodes = HashSet::new();
906 :
907 0 : if let Some(node_id) = self.intent.attached {
908 : // Maybe panic: it is a severe bug if we try to attach while generation is null.
909 0 : let generation = self
910 0 : .generation
911 0 : .expect("Attempted to enter attached state without a generation");
912 0 :
913 0 : let wanted_conf =
914 0 : attached_location_conf(generation, &self.shard, &self.config, &self.policy);
915 0 : match self.observed.locations.get(&node_id) {
916 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
917 0 : Some(_) | None => {
918 0 : dirty_nodes.insert(node_id);
919 0 : }
920 : }
921 0 : }
922 :
923 0 : for node_id in &self.intent.secondary {
924 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
925 0 : match self.observed.locations.get(node_id) {
926 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
927 0 : Some(_) | None => {
928 0 : dirty_nodes.insert(*node_id);
929 0 : }
930 : }
931 : }
932 :
933 0 : for node_id in self.observed.locations.keys() {
934 0 : if self.intent.attached != Some(*node_id) && !self.intent.secondary.contains(node_id) {
935 0 : // We have observed state that isn't part of our intent: need to clean it up.
936 0 : dirty_nodes.insert(*node_id);
937 0 : }
938 : }
939 :
940 0 : dirty_nodes.retain(|node_id| {
941 0 : nodes
942 0 : .get(node_id)
943 0 : .map(|n| n.is_available())
944 0 : .unwrap_or(false)
945 0 : });
946 0 :
947 0 : !dirty_nodes.is_empty()
948 0 : }
949 :
950 : #[allow(clippy::too_many_arguments)]
951 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
952 : pub(crate) fn get_reconcile_needed(
953 : &mut self,
954 : pageservers: &Arc<HashMap<NodeId, Node>>,
955 : ) -> ReconcileNeeded {
956 : // If there are any ambiguous observed states, and the nodes they refer to are available,
957 : // we should reconcile to clean them up.
958 : let mut dirty_observed = false;
959 : for (node_id, observed_loc) in &self.observed.locations {
960 : let node = pageservers
961 : .get(node_id)
962 : .expect("Nodes may not be removed while referenced");
963 : if observed_loc.conf.is_none() && node.is_available() {
964 : dirty_observed = true;
965 : break;
966 : }
967 : }
968 :
969 : let active_nodes_dirty = self.dirty(pageservers);
970 :
971 : // Even if there is no pageserver work to be done, if we have a pending notification to computes,
972 : // wake up a reconciler to send it.
973 : let do_reconcile =
974 : active_nodes_dirty || dirty_observed || self.pending_compute_notification;
975 :
976 : if !do_reconcile {
977 : tracing::debug!("Not dirty, no reconciliation needed.");
978 : return ReconcileNeeded::No;
979 : }
980 :
981 : // If we are currently splitting, then never start a reconciler task: the splitting logic
982 : // requires that shards are not interfered with while it runs. Do this check here rather than
983 : // up top, so that we only log this message if we would otherwise have done a reconciliation.
984 : if !matches!(self.splitting, SplitState::Idle) {
985 : tracing::info!("Refusing to reconcile, splitting in progress");
986 : return ReconcileNeeded::No;
987 : }
988 :
989 : // Reconcile already in flight for the current sequence?
990 : if let Some(handle) = &self.reconciler {
991 : if handle.sequence == self.sequence {
992 : tracing::info!(
993 : "Reconciliation already in progress for sequence {:?}",
994 : self.sequence,
995 : );
996 : return ReconcileNeeded::WaitExisting(ReconcilerWaiter {
997 : tenant_shard_id: self.tenant_shard_id,
998 : seq_wait: self.waiter.clone(),
999 : error_seq_wait: self.error_waiter.clone(),
1000 : error: self.last_error.clone(),
1001 : seq: self.sequence,
1002 : });
1003 : }
1004 : }
1005 :
1006 : // Pre-checks done: finally check whether we may actually do the work
1007 : match self.scheduling_policy {
1008 : ShardSchedulingPolicy::Active
1009 : | ShardSchedulingPolicy::Essential
1010 : | ShardSchedulingPolicy::Pause => {}
1011 : ShardSchedulingPolicy::Stop => {
1012 : // We only reach this point if there is work to do and we're going to skip
1013 : // doing it: warn it obvious why this tenant isn't doing what it ought to.
1014 : tracing::warn!("Skipping reconcile for policy {:?}", self.scheduling_policy);
1015 : return ReconcileNeeded::No;
1016 : }
1017 : }
1018 :
1019 : ReconcileNeeded::Yes
1020 : }
1021 :
1022 : /// Ensure the sequence number is set to a value where waiting for this value will make us wait
1023 : /// for the next reconcile: i.e. it is ahead of all completed or running reconcilers.
1024 : ///
1025 : /// Constructing a ReconcilerWaiter with the resulting sequence number gives the property
1026 : /// that the waiter will not complete until some future Reconciler is constructed and run.
1027 0 : fn ensure_sequence_ahead(&mut self) {
1028 0 : // Find the highest sequence for which a Reconciler has previously run or is currently
1029 0 : // running
1030 0 : let max_seen = std::cmp::max(
1031 0 : self.reconciler
1032 0 : .as_ref()
1033 0 : .map(|r| r.sequence)
1034 0 : .unwrap_or(Sequence(0)),
1035 0 : std::cmp::max(self.waiter.load(), self.error_waiter.load()),
1036 0 : );
1037 0 :
1038 0 : if self.sequence <= max_seen {
1039 0 : self.sequence = max_seen.next();
1040 0 : }
1041 0 : }
1042 :
1043 : /// Create a waiter that will wait for some future Reconciler that hasn't been spawned yet.
1044 : ///
1045 : /// This is appropriate when you can't spawn a reconciler (e.g. due to resource limits), but
1046 : /// you would like to wait on the next reconciler that gets spawned in the background.
1047 0 : pub(crate) fn future_reconcile_waiter(&mut self) -> ReconcilerWaiter {
1048 0 : self.ensure_sequence_ahead();
1049 0 :
1050 0 : ReconcilerWaiter {
1051 0 : tenant_shard_id: self.tenant_shard_id,
1052 0 : seq_wait: self.waiter.clone(),
1053 0 : error_seq_wait: self.error_waiter.clone(),
1054 0 : error: self.last_error.clone(),
1055 0 : seq: self.sequence,
1056 0 : }
1057 0 : }
1058 :
1059 : #[allow(clippy::too_many_arguments)]
1060 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
1061 : pub(crate) fn spawn_reconciler(
1062 : &mut self,
1063 : result_tx: &tokio::sync::mpsc::UnboundedSender<ReconcileResultRequest>,
1064 : pageservers: &Arc<HashMap<NodeId, Node>>,
1065 : compute_hook: &Arc<ComputeHook>,
1066 : service_config: &service::Config,
1067 : persistence: &Arc<Persistence>,
1068 : units: ReconcileUnits,
1069 : gate_guard: GateGuard,
1070 : cancel: &CancellationToken,
1071 : ) -> Option<ReconcilerWaiter> {
1072 : // Reconcile in flight for a stale sequence? Our sequence's task will wait for it before
1073 : // doing our sequence's work.
1074 : let old_handle = self.reconciler.take();
1075 :
1076 : // Build list of nodes from which the reconciler should detach
1077 : let mut detach = Vec::new();
1078 : for node_id in self.observed.locations.keys() {
1079 : if self.intent.get_attached() != &Some(*node_id)
1080 : && !self.intent.secondary.contains(node_id)
1081 : {
1082 : detach.push(
1083 : pageservers
1084 : .get(node_id)
1085 : .expect("Intent references non-existent pageserver")
1086 : .clone(),
1087 : )
1088 : }
1089 : }
1090 :
1091 : // Advance the sequence before spawning a reconciler, so that sequence waiters
1092 : // can distinguish between before+after the reconcile completes.
1093 : self.ensure_sequence_ahead();
1094 :
1095 : let reconciler_cancel = cancel.child_token();
1096 : let reconciler_intent = TargetState::from_intent(pageservers, &self.intent);
1097 : let mut reconciler = Reconciler {
1098 : tenant_shard_id: self.tenant_shard_id,
1099 : shard: self.shard,
1100 : placement_policy: self.policy.clone(),
1101 : generation: self.generation,
1102 : intent: reconciler_intent,
1103 : detach,
1104 : config: self.config.clone(),
1105 : observed: self.observed.clone(),
1106 : compute_hook: compute_hook.clone(),
1107 : service_config: service_config.clone(),
1108 : _gate_guard: gate_guard,
1109 : _resource_units: units,
1110 : cancel: reconciler_cancel.clone(),
1111 : persistence: persistence.clone(),
1112 : compute_notify_failure: false,
1113 : };
1114 :
1115 : let reconcile_seq = self.sequence;
1116 :
1117 : tracing::info!(seq=%reconcile_seq, "Spawning Reconciler for sequence {}", self.sequence);
1118 : let must_notify = self.pending_compute_notification;
1119 : let reconciler_span = tracing::info_span!(parent: None, "reconciler", seq=%reconcile_seq,
1120 : tenant_id=%reconciler.tenant_shard_id.tenant_id,
1121 : shard_id=%reconciler.tenant_shard_id.shard_slug());
1122 : metrics::METRICS_REGISTRY
1123 : .metrics_group
1124 : .storage_controller_reconcile_spawn
1125 : .inc();
1126 : let result_tx = result_tx.clone();
1127 : let join_handle = tokio::task::spawn(
1128 0 : async move {
1129 : // Wait for any previous reconcile task to complete before we start
1130 0 : if let Some(old_handle) = old_handle {
1131 0 : old_handle.cancel.cancel();
1132 0 : if let Err(e) = old_handle.handle.await {
1133 : // We can't do much with this other than log it: the task is done, so
1134 : // we may proceed with our work.
1135 0 : tracing::error!("Unexpected join error waiting for reconcile task: {e}");
1136 0 : }
1137 0 : }
1138 :
1139 : // Early check for cancellation before doing any work
1140 : // TODO: wrap all remote API operations in cancellation check
1141 : // as well.
1142 0 : if reconciler.cancel.is_cancelled() {
1143 0 : metrics::METRICS_REGISTRY
1144 0 : .metrics_group
1145 0 : .storage_controller_reconcile_complete
1146 0 : .inc(ReconcileCompleteLabelGroup {
1147 0 : status: ReconcileOutcome::Cancel,
1148 0 : });
1149 0 : return;
1150 0 : }
1151 :
1152 : // Attempt to make observed state match intent state
1153 0 : let result = reconciler.reconcile().await;
1154 :
1155 : // If we know we had a pending compute notification from some previous action, send a notification irrespective
1156 : // of whether the above reconcile() did any work
1157 0 : if result.is_ok() && must_notify {
1158 : // If this fails we will send the need to retry in [`ReconcileResult::pending_compute_notification`]
1159 0 : reconciler.compute_notify().await.ok();
1160 0 : }
1161 :
1162 : // Update result counter
1163 0 : let outcome_label = match &result {
1164 0 : Ok(_) => ReconcileOutcome::Success,
1165 0 : Err(ReconcileError::Cancel) => ReconcileOutcome::Cancel,
1166 0 : Err(_) => ReconcileOutcome::Error,
1167 : };
1168 :
1169 0 : metrics::METRICS_REGISTRY
1170 0 : .metrics_group
1171 0 : .storage_controller_reconcile_complete
1172 0 : .inc(ReconcileCompleteLabelGroup {
1173 0 : status: outcome_label,
1174 0 : });
1175 0 :
1176 0 : // Constructing result implicitly drops Reconciler, freeing any ReconcileUnits before the Service might
1177 0 : // try and schedule more work in response to our result.
1178 0 : let result = ReconcileResult {
1179 0 : sequence: reconcile_seq,
1180 0 : result,
1181 0 : tenant_shard_id: reconciler.tenant_shard_id,
1182 0 : generation: reconciler.generation,
1183 0 : observed: reconciler.observed,
1184 0 : pending_compute_notification: reconciler.compute_notify_failure,
1185 0 : };
1186 0 :
1187 0 : result_tx
1188 0 : .send(ReconcileResultRequest::ReconcileResult(result))
1189 0 : .ok();
1190 0 : }
1191 : .instrument(reconciler_span),
1192 : );
1193 :
1194 : self.reconciler = Some(ReconcilerHandle {
1195 : sequence: self.sequence,
1196 : handle: join_handle,
1197 : cancel: reconciler_cancel,
1198 : });
1199 :
1200 : Some(ReconcilerWaiter {
1201 : tenant_shard_id: self.tenant_shard_id,
1202 : seq_wait: self.waiter.clone(),
1203 : error_seq_wait: self.error_waiter.clone(),
1204 : error: self.last_error.clone(),
1205 : seq: self.sequence,
1206 : })
1207 : }
1208 :
1209 : /// Get a waiter for any reconciliation in flight, but do not start reconciliation
1210 : /// if it is not already running
1211 0 : pub(crate) fn get_waiter(&self) -> Option<ReconcilerWaiter> {
1212 0 : if self.reconciler.is_some() {
1213 0 : Some(ReconcilerWaiter {
1214 0 : tenant_shard_id: self.tenant_shard_id,
1215 0 : seq_wait: self.waiter.clone(),
1216 0 : error_seq_wait: self.error_waiter.clone(),
1217 0 : error: self.last_error.clone(),
1218 0 : seq: self.sequence,
1219 0 : })
1220 : } else {
1221 0 : None
1222 : }
1223 0 : }
1224 :
1225 : /// Called when a ReconcileResult has been emitted and the service is updating
1226 : /// our state: if the result is from a sequence >= my ReconcileHandle, then drop
1227 : /// the handle to indicate there is no longer a reconciliation in progress.
1228 0 : pub(crate) fn reconcile_complete(&mut self, sequence: Sequence) {
1229 0 : if let Some(reconcile_handle) = &self.reconciler {
1230 0 : if reconcile_handle.sequence <= sequence {
1231 0 : self.reconciler = None;
1232 0 : }
1233 0 : }
1234 0 : }
1235 :
1236 : /// If we had any state at all referring to this node ID, drop it. Does not
1237 : /// attempt to reschedule.
1238 : ///
1239 : /// Returns true if we modified the node's intent state.
1240 0 : pub(crate) fn deref_node(&mut self, node_id: NodeId) -> bool {
1241 0 : let mut intent_modified = false;
1242 0 :
1243 0 : // Drop if this node was our attached intent
1244 0 : if self.intent.attached == Some(node_id) {
1245 0 : self.intent.attached = None;
1246 0 : intent_modified = true;
1247 0 : }
1248 :
1249 : // Drop from the list of secondaries, and check if we modified it
1250 0 : let had_secondaries = self.intent.secondary.len();
1251 0 : self.intent.secondary.retain(|n| n != &node_id);
1252 0 : intent_modified |= self.intent.secondary.len() != had_secondaries;
1253 0 :
1254 0 : debug_assert!(!self.intent.all_pageservers().contains(&node_id));
1255 :
1256 0 : intent_modified
1257 0 : }
1258 :
1259 0 : pub(crate) fn set_scheduling_policy(&mut self, p: ShardSchedulingPolicy) {
1260 0 : self.scheduling_policy = p;
1261 0 : }
1262 :
1263 0 : pub(crate) fn get_scheduling_policy(&self) -> &ShardSchedulingPolicy {
1264 0 : &self.scheduling_policy
1265 0 : }
1266 :
1267 0 : pub(crate) fn set_last_error(&mut self, sequence: Sequence, error: ReconcileError) {
1268 0 : // Ordering: always set last_error before advancing sequence, so that sequence
1269 0 : // waiters are guaranteed to see a Some value when they see an error.
1270 0 : *(self.last_error.lock().unwrap()) = Some(Arc::new(error));
1271 0 : self.error_waiter.advance(sequence);
1272 0 : }
1273 :
1274 0 : pub(crate) fn from_persistent(
1275 0 : tsp: TenantShardPersistence,
1276 0 : intent: IntentState,
1277 0 : ) -> anyhow::Result<Self> {
1278 0 : let tenant_shard_id = tsp.get_tenant_shard_id()?;
1279 0 : let shard_identity = tsp.get_shard_identity()?;
1280 :
1281 0 : Ok(Self {
1282 0 : tenant_shard_id,
1283 0 : shard: shard_identity,
1284 0 : sequence: Sequence::initial(),
1285 0 : generation: tsp.generation.map(|g| Generation::new(g as u32)),
1286 0 : policy: serde_json::from_str(&tsp.placement_policy).unwrap(),
1287 0 : intent,
1288 0 : observed: ObservedState::new(),
1289 0 : config: serde_json::from_str(&tsp.config).unwrap(),
1290 0 : reconciler: None,
1291 0 : splitting: tsp.splitting,
1292 0 : waiter: Arc::new(SeqWait::new(Sequence::initial())),
1293 0 : error_waiter: Arc::new(SeqWait::new(Sequence::initial())),
1294 0 : last_error: Arc::default(),
1295 0 : pending_compute_notification: false,
1296 0 : delayed_reconcile: false,
1297 0 : scheduling_policy: serde_json::from_str(&tsp.scheduling_policy).unwrap(),
1298 0 : })
1299 0 : }
1300 :
1301 0 : pub(crate) fn to_persistent(&self) -> TenantShardPersistence {
1302 0 : TenantShardPersistence {
1303 0 : tenant_id: self.tenant_shard_id.tenant_id.to_string(),
1304 0 : shard_number: self.tenant_shard_id.shard_number.0 as i32,
1305 0 : shard_count: self.tenant_shard_id.shard_count.literal() as i32,
1306 0 : shard_stripe_size: self.shard.stripe_size.0 as i32,
1307 0 : generation: self.generation.map(|g| g.into().unwrap_or(0) as i32),
1308 0 : generation_pageserver: self.intent.get_attached().map(|n| n.0 as i64),
1309 0 : placement_policy: serde_json::to_string(&self.policy).unwrap(),
1310 0 : config: serde_json::to_string(&self.config).unwrap(),
1311 0 : splitting: SplitState::default(),
1312 0 : scheduling_policy: serde_json::to_string(&self.scheduling_policy).unwrap(),
1313 0 : }
1314 0 : }
1315 : }
1316 :
1317 : #[cfg(test)]
1318 : pub(crate) mod tests {
1319 : use pageserver_api::{
1320 : controller_api::NodeAvailability,
1321 : shard::{ShardCount, ShardNumber},
1322 : };
1323 : use utils::id::TenantId;
1324 :
1325 : use crate::scheduler::test_utils::make_test_nodes;
1326 :
1327 : use super::*;
1328 :
1329 14 : fn make_test_tenant_shard(policy: PlacementPolicy) -> TenantShard {
1330 14 : let tenant_id = TenantId::generate();
1331 14 : let shard_number = ShardNumber(0);
1332 14 : let shard_count = ShardCount::new(1);
1333 14 :
1334 14 : let tenant_shard_id = TenantShardId {
1335 14 : tenant_id,
1336 14 : shard_number,
1337 14 : shard_count,
1338 14 : };
1339 14 : TenantShard::new(
1340 14 : tenant_shard_id,
1341 14 : ShardIdentity::new(
1342 14 : shard_number,
1343 14 : shard_count,
1344 14 : pageserver_api::shard::ShardStripeSize(32768),
1345 14 : )
1346 14 : .unwrap(),
1347 14 : policy,
1348 14 : )
1349 14 : }
1350 :
1351 2 : fn make_test_tenant(policy: PlacementPolicy, shard_count: ShardCount) -> Vec<TenantShard> {
1352 2 : let tenant_id = TenantId::generate();
1353 2 :
1354 2 : (0..shard_count.count())
1355 8 : .map(|i| {
1356 8 : let shard_number = ShardNumber(i);
1357 8 :
1358 8 : let tenant_shard_id = TenantShardId {
1359 8 : tenant_id,
1360 8 : shard_number,
1361 8 : shard_count,
1362 8 : };
1363 8 : TenantShard::new(
1364 8 : tenant_shard_id,
1365 8 : ShardIdentity::new(
1366 8 : shard_number,
1367 8 : shard_count,
1368 8 : pageserver_api::shard::ShardStripeSize(32768),
1369 8 : )
1370 8 : .unwrap(),
1371 8 : policy.clone(),
1372 8 : )
1373 8 : })
1374 2 : .collect()
1375 2 : }
1376 :
1377 : /// Test the scheduling behaviors used when a tenant configured for HA is subject
1378 : /// to nodes being marked offline.
1379 : #[test]
1380 2 : fn tenant_ha_scheduling() -> anyhow::Result<()> {
1381 2 : // Start with three nodes. Our tenant will only use two. The third one is
1382 2 : // expected to remain unused.
1383 2 : let mut nodes = make_test_nodes(3);
1384 2 :
1385 2 : let mut scheduler = Scheduler::new(nodes.values());
1386 2 : let mut context = ScheduleContext::default();
1387 2 :
1388 2 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
1389 2 : tenant_shard
1390 2 : .schedule(&mut scheduler, &mut context)
1391 2 : .expect("we have enough nodes, scheduling should work");
1392 2 :
1393 2 : // Expect to initially be schedule on to different nodes
1394 2 : assert_eq!(tenant_shard.intent.secondary.len(), 1);
1395 2 : assert!(tenant_shard.intent.attached.is_some());
1396 :
1397 2 : let attached_node_id = tenant_shard.intent.attached.unwrap();
1398 2 : let secondary_node_id = *tenant_shard.intent.secondary.iter().last().unwrap();
1399 2 : assert_ne!(attached_node_id, secondary_node_id);
1400 :
1401 : // Notifying the attached node is offline should demote it to a secondary
1402 2 : let changed = tenant_shard
1403 2 : .intent
1404 2 : .demote_attached(&mut scheduler, attached_node_id);
1405 2 : assert!(changed);
1406 2 : assert!(tenant_shard.intent.attached.is_none());
1407 2 : assert_eq!(tenant_shard.intent.secondary.len(), 2);
1408 :
1409 : // Update the scheduler state to indicate the node is offline
1410 2 : nodes
1411 2 : .get_mut(&attached_node_id)
1412 2 : .unwrap()
1413 2 : .set_availability(NodeAvailability::Offline);
1414 2 : scheduler.node_upsert(nodes.get(&attached_node_id).unwrap());
1415 2 :
1416 2 : // Scheduling the node should promote the still-available secondary node to attached
1417 2 : tenant_shard
1418 2 : .schedule(&mut scheduler, &mut context)
1419 2 : .expect("active nodes are available");
1420 2 : assert_eq!(tenant_shard.intent.attached.unwrap(), secondary_node_id);
1421 :
1422 : // The original attached node should have been retained as a secondary
1423 2 : assert_eq!(
1424 2 : *tenant_shard.intent.secondary.iter().last().unwrap(),
1425 2 : attached_node_id
1426 2 : );
1427 :
1428 2 : tenant_shard.intent.clear(&mut scheduler);
1429 2 :
1430 2 : Ok(())
1431 2 : }
1432 :
1433 : #[test]
1434 2 : fn intent_from_observed() -> anyhow::Result<()> {
1435 2 : let nodes = make_test_nodes(3);
1436 2 : let mut scheduler = Scheduler::new(nodes.values());
1437 2 :
1438 2 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
1439 2 :
1440 2 : tenant_shard.observed.locations.insert(
1441 2 : NodeId(3),
1442 2 : ObservedStateLocation {
1443 2 : conf: Some(LocationConfig {
1444 2 : mode: LocationConfigMode::AttachedMulti,
1445 2 : generation: Some(2),
1446 2 : secondary_conf: None,
1447 2 : shard_number: tenant_shard.shard.number.0,
1448 2 : shard_count: tenant_shard.shard.count.literal(),
1449 2 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
1450 2 : tenant_conf: TenantConfig::default(),
1451 2 : }),
1452 2 : },
1453 2 : );
1454 2 :
1455 2 : tenant_shard.observed.locations.insert(
1456 2 : NodeId(2),
1457 2 : ObservedStateLocation {
1458 2 : conf: Some(LocationConfig {
1459 2 : mode: LocationConfigMode::AttachedStale,
1460 2 : generation: Some(1),
1461 2 : secondary_conf: None,
1462 2 : shard_number: tenant_shard.shard.number.0,
1463 2 : shard_count: tenant_shard.shard.count.literal(),
1464 2 : shard_stripe_size: tenant_shard.shard.stripe_size.0,
1465 2 : tenant_conf: TenantConfig::default(),
1466 2 : }),
1467 2 : },
1468 2 : );
1469 2 :
1470 2 : tenant_shard.intent_from_observed(&mut scheduler);
1471 2 :
1472 2 : // The highest generationed attached location gets used as attached
1473 2 : assert_eq!(tenant_shard.intent.attached, Some(NodeId(3)));
1474 : // Other locations get used as secondary
1475 2 : assert_eq!(tenant_shard.intent.secondary, vec![NodeId(2)]);
1476 :
1477 2 : scheduler.consistency_check(nodes.values(), [&tenant_shard].into_iter())?;
1478 :
1479 2 : tenant_shard.intent.clear(&mut scheduler);
1480 2 : Ok(())
1481 2 : }
1482 :
1483 : #[test]
1484 2 : fn scheduling_mode() -> anyhow::Result<()> {
1485 2 : let nodes = make_test_nodes(3);
1486 2 : let mut scheduler = Scheduler::new(nodes.values());
1487 2 :
1488 2 : let mut tenant_shard = make_test_tenant_shard(PlacementPolicy::Attached(1));
1489 2 :
1490 2 : // In pause mode, schedule() shouldn't do anything
1491 2 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Pause;
1492 2 : assert!(tenant_shard
1493 2 : .schedule(&mut scheduler, &mut ScheduleContext::default())
1494 2 : .is_ok());
1495 2 : assert!(tenant_shard.intent.all_pageservers().is_empty());
1496 :
1497 : // In active mode, schedule() works
1498 2 : tenant_shard.scheduling_policy = ShardSchedulingPolicy::Active;
1499 2 : assert!(tenant_shard
1500 2 : .schedule(&mut scheduler, &mut ScheduleContext::default())
1501 2 : .is_ok());
1502 2 : assert!(!tenant_shard.intent.all_pageservers().is_empty());
1503 :
1504 2 : tenant_shard.intent.clear(&mut scheduler);
1505 2 : Ok(())
1506 2 : }
1507 :
1508 : #[test]
1509 2 : fn optimize_attachment() -> anyhow::Result<()> {
1510 2 : let nodes = make_test_nodes(3);
1511 2 : let mut scheduler = Scheduler::new(nodes.values());
1512 2 :
1513 2 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
1514 2 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
1515 2 :
1516 2 : // Initially: both nodes attached on shard 1, and both have secondary locations
1517 2 : // on different nodes.
1518 2 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
1519 2 : shard_a.intent.push_secondary(&mut scheduler, NodeId(2));
1520 2 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(1)));
1521 2 : shard_b.intent.push_secondary(&mut scheduler, NodeId(3));
1522 2 :
1523 2 : let mut schedule_context = ScheduleContext::default();
1524 2 : schedule_context.avoid(&shard_a.intent.all_pageservers());
1525 2 : schedule_context.push_attached(shard_a.intent.get_attached().unwrap());
1526 2 : schedule_context.avoid(&shard_b.intent.all_pageservers());
1527 2 : schedule_context.push_attached(shard_b.intent.get_attached().unwrap());
1528 2 :
1529 2 : let optimization_a = shard_a.optimize_attachment(&nodes, &schedule_context);
1530 2 :
1531 2 : // Either shard should recognize that it has the option to switch to a secondary location where there
1532 2 : // would be no other shards from the same tenant, and request to do so.
1533 2 : assert_eq!(
1534 2 : optimization_a,
1535 2 : Some(ScheduleOptimization {
1536 2 : sequence: shard_a.sequence,
1537 2 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1538 2 : old_attached_node_id: NodeId(1),
1539 2 : new_attached_node_id: NodeId(2)
1540 2 : })
1541 2 : })
1542 2 : );
1543 :
1544 : // Note that these optimizing two shards in the same tenant with the same ScheduleContext is
1545 : // mutually exclusive (the optimization of one invalidates the stats) -- it is the responsibility
1546 : // of [`Service::optimize_all`] to avoid trying
1547 : // to do optimizations for multiple shards in the same tenant at the same time. Generating
1548 : // both optimizations is just done for test purposes
1549 2 : let optimization_b = shard_b.optimize_attachment(&nodes, &schedule_context);
1550 2 : assert_eq!(
1551 2 : optimization_b,
1552 2 : Some(ScheduleOptimization {
1553 2 : sequence: shard_b.sequence,
1554 2 : action: ScheduleOptimizationAction::MigrateAttachment(MigrateAttachment {
1555 2 : old_attached_node_id: NodeId(1),
1556 2 : new_attached_node_id: NodeId(3)
1557 2 : })
1558 2 : })
1559 2 : );
1560 :
1561 : // Applying these optimizations should result in the end state proposed
1562 2 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
1563 2 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(2)));
1564 2 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(1)]);
1565 2 : shard_b.apply_optimization(&mut scheduler, optimization_b.unwrap());
1566 2 : assert_eq!(shard_b.intent.get_attached(), &Some(NodeId(3)));
1567 2 : assert_eq!(shard_b.intent.get_secondary(), &vec![NodeId(1)]);
1568 :
1569 2 : shard_a.intent.clear(&mut scheduler);
1570 2 : shard_b.intent.clear(&mut scheduler);
1571 2 :
1572 2 : Ok(())
1573 2 : }
1574 :
1575 : #[test]
1576 2 : fn optimize_secondary() -> anyhow::Result<()> {
1577 2 : let nodes = make_test_nodes(4);
1578 2 : let mut scheduler = Scheduler::new(nodes.values());
1579 2 :
1580 2 : let mut shard_a = make_test_tenant_shard(PlacementPolicy::Attached(1));
1581 2 : let mut shard_b = make_test_tenant_shard(PlacementPolicy::Attached(1));
1582 2 :
1583 2 : // Initially: both nodes attached on shard 1, and both have secondary locations
1584 2 : // on different nodes.
1585 2 : shard_a.intent.set_attached(&mut scheduler, Some(NodeId(1)));
1586 2 : shard_a.intent.push_secondary(&mut scheduler, NodeId(3));
1587 2 : shard_b.intent.set_attached(&mut scheduler, Some(NodeId(2)));
1588 2 : shard_b.intent.push_secondary(&mut scheduler, NodeId(3));
1589 2 :
1590 2 : let mut schedule_context = ScheduleContext::default();
1591 2 : schedule_context.avoid(&shard_a.intent.all_pageservers());
1592 2 : schedule_context.push_attached(shard_a.intent.get_attached().unwrap());
1593 2 : schedule_context.avoid(&shard_b.intent.all_pageservers());
1594 2 : schedule_context.push_attached(shard_b.intent.get_attached().unwrap());
1595 2 :
1596 2 : let optimization_a = shard_a.optimize_secondary(&scheduler, &schedule_context);
1597 2 :
1598 2 : // Since there is a node with no locations available, the node with two locations for the
1599 2 : // same tenant should generate an optimization to move one away
1600 2 : assert_eq!(
1601 2 : optimization_a,
1602 2 : Some(ScheduleOptimization {
1603 2 : sequence: shard_a.sequence,
1604 2 : action: ScheduleOptimizationAction::ReplaceSecondary(ReplaceSecondary {
1605 2 : old_node_id: NodeId(3),
1606 2 : new_node_id: NodeId(4)
1607 2 : })
1608 2 : })
1609 2 : );
1610 :
1611 2 : shard_a.apply_optimization(&mut scheduler, optimization_a.unwrap());
1612 2 : assert_eq!(shard_a.intent.get_attached(), &Some(NodeId(1)));
1613 2 : assert_eq!(shard_a.intent.get_secondary(), &vec![NodeId(4)]);
1614 :
1615 2 : shard_a.intent.clear(&mut scheduler);
1616 2 : shard_b.intent.clear(&mut scheduler);
1617 2 :
1618 2 : Ok(())
1619 2 : }
1620 :
1621 : // Optimize til quiescent: this emulates what Service::optimize_all does, when
1622 : // called repeatedly in the background.
1623 2 : fn optimize_til_idle(
1624 2 : nodes: &HashMap<NodeId, Node>,
1625 2 : scheduler: &mut Scheduler,
1626 2 : shards: &mut [TenantShard],
1627 2 : ) {
1628 2 : let mut loop_n = 0;
1629 : loop {
1630 14 : let mut schedule_context = ScheduleContext::default();
1631 14 : let mut any_changed = false;
1632 :
1633 56 : for shard in shards.iter() {
1634 56 : schedule_context.avoid(&shard.intent.all_pageservers());
1635 56 : if let Some(attached) = shard.intent.get_attached() {
1636 56 : schedule_context.push_attached(*attached);
1637 56 : }
1638 : }
1639 :
1640 26 : for shard in shards.iter_mut() {
1641 26 : let optimization = shard.optimize_attachment(nodes, &schedule_context);
1642 26 : if let Some(optimization) = optimization {
1643 4 : shard.apply_optimization(scheduler, optimization);
1644 4 : any_changed = true;
1645 4 : break;
1646 22 : }
1647 22 :
1648 22 : let optimization = shard.optimize_secondary(scheduler, &schedule_context);
1649 22 : if let Some(optimization) = optimization {
1650 8 : shard.apply_optimization(scheduler, optimization);
1651 8 : any_changed = true;
1652 8 : break;
1653 14 : }
1654 : }
1655 :
1656 14 : if !any_changed {
1657 2 : break;
1658 12 : }
1659 12 :
1660 12 : // Assert no infinite loop
1661 12 : loop_n += 1;
1662 12 : assert!(loop_n < 1000);
1663 : }
1664 2 : }
1665 :
1666 : /// Test the balancing behavior of shard scheduling: that it achieves a balance, and
1667 : /// that it converges.
1668 : #[test]
1669 2 : fn optimize_add_nodes() -> anyhow::Result<()> {
1670 2 : let nodes = make_test_nodes(4);
1671 2 :
1672 2 : // Only show the scheduler a couple of nodes
1673 2 : let mut scheduler = Scheduler::new([].iter());
1674 2 : scheduler.node_upsert(nodes.get(&NodeId(1)).unwrap());
1675 2 : scheduler.node_upsert(nodes.get(&NodeId(2)).unwrap());
1676 2 :
1677 2 : let mut shards = make_test_tenant(PlacementPolicy::Attached(1), ShardCount::new(4));
1678 2 : let mut schedule_context = ScheduleContext::default();
1679 10 : for shard in &mut shards {
1680 8 : assert!(shard
1681 8 : .schedule(&mut scheduler, &mut schedule_context)
1682 8 : .is_ok());
1683 : }
1684 :
1685 : // We should see equal number of locations on the two nodes.
1686 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 4);
1687 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 2);
1688 :
1689 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 4);
1690 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 2);
1691 :
1692 : // Add another two nodes: we should see the shards spread out when their optimize
1693 : // methods are called
1694 2 : scheduler.node_upsert(nodes.get(&NodeId(3)).unwrap());
1695 2 : scheduler.node_upsert(nodes.get(&NodeId(4)).unwrap());
1696 2 : optimize_til_idle(&nodes, &mut scheduler, &mut shards);
1697 2 :
1698 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(1)), 2);
1699 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(1)), 1);
1700 :
1701 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(2)), 2);
1702 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(2)), 1);
1703 :
1704 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(3)), 2);
1705 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(3)), 1);
1706 :
1707 2 : assert_eq!(scheduler.get_node_shard_count(NodeId(4)), 2);
1708 2 : assert_eq!(scheduler.get_node_attached_shard_count(NodeId(4)), 1);
1709 :
1710 8 : for shard in shards.iter_mut() {
1711 8 : shard.intent.clear(&mut scheduler);
1712 8 : }
1713 :
1714 2 : Ok(())
1715 2 : }
1716 : }
|
