Line data Source code
1 : use crate::pageserver_client::PageserverClient;
2 : use crate::persistence::Persistence;
3 : use crate::service;
4 : use pageserver_api::controller_api::PlacementPolicy;
5 : use pageserver_api::models::{
6 : LocationConfig, LocationConfigMode, LocationConfigSecondary, TenantConfig,
7 : };
8 : use pageserver_api::shard::{ShardIdentity, TenantShardId};
9 : use pageserver_client::mgmt_api;
10 : use reqwest::StatusCode;
11 : use std::collections::HashMap;
12 : use std::sync::Arc;
13 : use std::time::{Duration, Instant};
14 : use tokio_util::sync::CancellationToken;
15 : use utils::backoff::exponential_backoff;
16 : use utils::failpoint_support;
17 : use utils::generation::Generation;
18 : use utils::id::{NodeId, TimelineId};
19 : use utils::lsn::Lsn;
20 : use utils::pausable_failpoint;
21 : use utils::sync::gate::GateGuard;
22 :
23 : use crate::compute_hook::{ComputeHook, NotifyError};
24 : use crate::node::Node;
25 : use crate::tenant_shard::{IntentState, ObservedState, ObservedStateDelta, ObservedStateLocation};
26 :
27 : const DEFAULT_HEATMAP_PERIOD: &str = "60s";
28 :
29 : /// Object with the lifetime of the background reconcile task that is created
30 : /// for tenants which have a difference between their intent and observed states.
31 : pub(super) struct Reconciler {
32 : /// See [`crate::tenant_shard::TenantShard`] for the meanings of these fields: they are a snapshot
33 : /// of a tenant's state from when we spawned a reconcile task.
34 : pub(super) tenant_shard_id: TenantShardId,
35 : pub(crate) shard: ShardIdentity,
36 : pub(crate) placement_policy: PlacementPolicy,
37 : pub(crate) generation: Option<Generation>,
38 : pub(crate) intent: TargetState,
39 :
40 : /// Nodes not referenced by [`Self::intent`], from which we should try
41 : /// to detach this tenant shard.
42 : pub(crate) detach: Vec<Node>,
43 :
44 : /// Configuration specific to this reconciler
45 : pub(crate) reconciler_config: ReconcilerConfig,
46 :
47 : pub(crate) config: TenantConfig,
48 :
49 : /// Observed state from the point of view of the reconciler.
50 : /// This gets updated as the reconciliation makes progress.
51 : pub(crate) observed: ObservedState,
52 :
53 : /// Snapshot of the observed state at the point when the reconciler
54 : /// was spawned.
55 : pub(crate) original_observed: ObservedState,
56 :
57 : pub(crate) service_config: service::Config,
58 :
59 : /// A hook to notify the running postgres instances when we change the location
60 : /// of a tenant. Use this via [`Self::compute_notify`] to update our failure flag
61 : /// and guarantee eventual retries.
62 : pub(crate) compute_hook: Arc<ComputeHook>,
63 :
64 : /// To avoid stalling if the cloud control plane is unavailable, we may proceed
65 : /// past failures in [`ComputeHook::notify`], but we _must_ remember that we failed
66 : /// so that we can set [`crate::tenant_shard::TenantShard::pending_compute_notification`] to ensure a later retry.
67 : pub(crate) compute_notify_failure: bool,
68 :
69 : /// Reconciler is responsible for keeping alive semaphore units that limit concurrency on how many
70 : /// we will spawn.
71 : pub(crate) _resource_units: ReconcileUnits,
72 :
73 : /// A means to abort background reconciliation: it is essential to
74 : /// call this when something changes in the original TenantShard that
75 : /// will make this reconciliation impossible or unnecessary, for
76 : /// example when a pageserver node goes offline, or the PlacementPolicy for
77 : /// the tenant is changed.
78 : pub(crate) cancel: CancellationToken,
79 :
80 : /// Reconcilers are registered with a Gate so that during a graceful shutdown we
81 : /// can wait for all the reconcilers to respond to their cancellation tokens.
82 : pub(crate) _gate_guard: GateGuard,
83 :
84 : /// Access to persistent storage for updating generation numbers
85 : pub(crate) persistence: Arc<Persistence>,
86 : }
87 :
88 : pub(crate) struct ReconcilerConfigBuilder {
89 : config: ReconcilerConfig,
90 : }
91 :
92 : impl ReconcilerConfigBuilder {
93 0 : pub(crate) fn new() -> Self {
94 0 : Self {
95 0 : config: ReconcilerConfig::default(),
96 0 : }
97 0 : }
98 :
99 0 : pub(crate) fn secondary_warmup_timeout(self, value: Duration) -> Self {
100 0 : Self {
101 0 : config: ReconcilerConfig {
102 0 : secondary_warmup_timeout: Some(value),
103 0 : ..self.config
104 0 : },
105 0 : }
106 0 : }
107 :
108 0 : pub(crate) fn secondary_download_request_timeout(self, value: Duration) -> Self {
109 0 : Self {
110 0 : config: ReconcilerConfig {
111 0 : secondary_download_request_timeout: Some(value),
112 0 : ..self.config
113 0 : },
114 0 : }
115 0 : }
116 :
117 0 : pub(crate) fn build(self) -> ReconcilerConfig {
118 0 : self.config
119 0 : }
120 : }
121 :
122 : #[derive(Default, Debug, Copy, Clone)]
123 : pub(crate) struct ReconcilerConfig {
124 : // During live migration give up on warming-up the secondary
125 : // after this timeout.
126 : secondary_warmup_timeout: Option<Duration>,
127 :
128 : // During live migrations this is the amount of time that
129 : // the pagserver will hold our poll.
130 : secondary_download_request_timeout: Option<Duration>,
131 : }
132 :
133 : impl ReconcilerConfig {
134 0 : pub(crate) fn get_secondary_warmup_timeout(&self) -> Duration {
135 : const SECONDARY_WARMUP_TIMEOUT_DEFAULT: Duration = Duration::from_secs(300);
136 0 : self.secondary_warmup_timeout
137 0 : .unwrap_or(SECONDARY_WARMUP_TIMEOUT_DEFAULT)
138 0 : }
139 :
140 0 : pub(crate) fn get_secondary_download_request_timeout(&self) -> Duration {
141 : const SECONDARY_DOWNLOAD_REQUEST_TIMEOUT_DEFAULT: Duration = Duration::from_secs(20);
142 0 : self.secondary_download_request_timeout
143 0 : .unwrap_or(SECONDARY_DOWNLOAD_REQUEST_TIMEOUT_DEFAULT)
144 0 : }
145 : }
146 :
147 : /// RAII resource units granted to a Reconciler, which it should keep alive until it finishes doing I/O
148 : pub(crate) struct ReconcileUnits {
149 : _sem_units: tokio::sync::OwnedSemaphorePermit,
150 : }
151 :
152 : impl ReconcileUnits {
153 0 : pub(crate) fn new(sem_units: tokio::sync::OwnedSemaphorePermit) -> Self {
154 0 : Self {
155 0 : _sem_units: sem_units,
156 0 : }
157 0 : }
158 : }
159 :
160 : /// This is a snapshot of [`crate::tenant_shard::IntentState`], but it does not do any
161 : /// reference counting for Scheduler. The IntentState is what the scheduler works with,
162 : /// and the TargetState is just the instruction for a particular Reconciler run.
163 : #[derive(Debug)]
164 : pub(crate) struct TargetState {
165 : pub(crate) attached: Option<Node>,
166 : pub(crate) secondary: Vec<Node>,
167 : }
168 :
169 : impl TargetState {
170 0 : pub(crate) fn from_intent(nodes: &HashMap<NodeId, Node>, intent: &IntentState) -> Self {
171 0 : Self {
172 0 : attached: intent.get_attached().map(|n| {
173 0 : nodes
174 0 : .get(&n)
175 0 : .expect("Intent attached referenced non-existent node")
176 0 : .clone()
177 0 : }),
178 0 : secondary: intent
179 0 : .get_secondary()
180 0 : .iter()
181 0 : .map(|n| {
182 0 : nodes
183 0 : .get(n)
184 0 : .expect("Intent secondary referenced non-existent node")
185 0 : .clone()
186 0 : })
187 0 : .collect(),
188 0 : }
189 0 : }
190 : }
191 :
192 0 : #[derive(thiserror::Error, Debug)]
193 : pub(crate) enum ReconcileError {
194 : #[error(transparent)]
195 : Remote(#[from] mgmt_api::Error),
196 : #[error(transparent)]
197 : Notify(#[from] NotifyError),
198 : #[error("Cancelled")]
199 : Cancel,
200 : #[error(transparent)]
201 : Other(#[from] anyhow::Error),
202 : }
203 :
204 : impl Reconciler {
205 0 : async fn location_config(
206 0 : &mut self,
207 0 : node: &Node,
208 0 : config: LocationConfig,
209 0 : flush_ms: Option<Duration>,
210 0 : lazy: bool,
211 0 : ) -> Result<(), ReconcileError> {
212 0 : if !node.is_available() && config.mode == LocationConfigMode::Detached {
213 : // Attempts to detach from offline nodes may be imitated without doing I/O: a node which is offline
214 : // will get fully reconciled wrt the shard's intent state when it is reactivated, irrespective of
215 : // what we put into `observed`, in [`crate::service::Service::node_activate_reconcile`]
216 0 : tracing::info!("Node {node} is unavailable during detach: proceeding anyway, it will be detached on next activation");
217 0 : self.observed.locations.remove(&node.get_id());
218 0 : return Ok(());
219 0 : }
220 0 :
221 0 : self.observed
222 0 : .locations
223 0 : .insert(node.get_id(), ObservedStateLocation { conf: None });
224 0 :
225 0 : // TODO: amend locations that use long-polling: they will hit this timeout.
226 0 : let timeout = Duration::from_secs(25);
227 0 :
228 0 : tracing::info!("location_config({node}) calling: {:?}", config);
229 0 : let tenant_shard_id = self.tenant_shard_id;
230 0 : let config_ref = &config;
231 0 : match node
232 0 : .with_client_retries(
233 0 : |client| async move {
234 0 : let config = config_ref.clone();
235 0 : client
236 0 : .location_config(tenant_shard_id, config.clone(), flush_ms, lazy)
237 0 : .await
238 0 : },
239 0 : &self.service_config.jwt_token,
240 0 : 1,
241 0 : 3,
242 0 : timeout,
243 0 : &self.cancel,
244 0 : )
245 0 : .await
246 : {
247 0 : Some(Ok(_)) => {}
248 0 : Some(Err(e)) => return Err(e.into()),
249 0 : None => return Err(ReconcileError::Cancel),
250 : };
251 0 : tracing::info!("location_config({node}) complete: {:?}", config);
252 :
253 0 : match config.mode {
254 0 : LocationConfigMode::Detached => {
255 0 : self.observed.locations.remove(&node.get_id());
256 0 : }
257 0 : _ => {
258 0 : self.observed
259 0 : .locations
260 0 : .insert(node.get_id(), ObservedStateLocation { conf: Some(config) });
261 0 : }
262 : }
263 :
264 0 : Ok(())
265 0 : }
266 :
267 0 : fn get_node(&self, node_id: &NodeId) -> Option<&Node> {
268 0 : if let Some(node) = self.intent.attached.as_ref() {
269 0 : if node.get_id() == *node_id {
270 0 : return Some(node);
271 0 : }
272 0 : }
273 :
274 0 : if let Some(node) = self
275 0 : .intent
276 0 : .secondary
277 0 : .iter()
278 0 : .find(|n| n.get_id() == *node_id)
279 : {
280 0 : return Some(node);
281 0 : }
282 :
283 0 : if let Some(node) = self.detach.iter().find(|n| n.get_id() == *node_id) {
284 0 : return Some(node);
285 0 : }
286 0 :
287 0 : None
288 0 : }
289 :
290 0 : async fn maybe_live_migrate(&mut self) -> Result<(), ReconcileError> {
291 0 : let destination = if let Some(node) = &self.intent.attached {
292 0 : match self.observed.locations.get(&node.get_id()) {
293 0 : Some(conf) => {
294 : // We will do a live migration only if the intended destination is not
295 : // currently in an attached state.
296 0 : match &conf.conf {
297 0 : Some(conf) if conf.mode == LocationConfigMode::Secondary => {
298 0 : // Fall through to do a live migration
299 0 : node
300 : }
301 : None | Some(_) => {
302 : // Attached or uncertain: don't do a live migration, proceed
303 : // with a general-case reconciliation
304 0 : tracing::info!("maybe_live_migrate: destination is None or attached");
305 0 : return Ok(());
306 : }
307 : }
308 : }
309 : None => {
310 : // Our destination is not attached: maybe live migrate if some other
311 : // node is currently attached. Fall through.
312 0 : node
313 : }
314 : }
315 : } else {
316 : // No intent to be attached
317 0 : tracing::info!("maybe_live_migrate: no attached intent");
318 0 : return Ok(());
319 : };
320 :
321 0 : let mut origin = None;
322 0 : for (node_id, state) in &self.observed.locations {
323 0 : if let Some(observed_conf) = &state.conf {
324 0 : if observed_conf.mode == LocationConfigMode::AttachedSingle {
325 : // We will only attempt live migration if the origin is not offline: this
326 : // avoids trying to do it while reconciling after responding to an HA failover.
327 0 : if let Some(node) = self.get_node(node_id) {
328 0 : if node.is_available() {
329 0 : origin = Some(node.clone());
330 0 : break;
331 0 : }
332 0 : }
333 0 : }
334 0 : }
335 : }
336 :
337 0 : let Some(origin) = origin else {
338 0 : tracing::info!("maybe_live_migrate: no origin found");
339 0 : return Ok(());
340 : };
341 :
342 : // We have an origin and a destination: proceed to do the live migration
343 0 : tracing::info!("Live migrating {}->{}", origin, destination);
344 0 : self.live_migrate(origin, destination.clone()).await?;
345 :
346 0 : Ok(())
347 0 : }
348 :
349 0 : async fn get_lsns(
350 0 : &self,
351 0 : tenant_shard_id: TenantShardId,
352 0 : node: &Node,
353 0 : ) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
354 0 : let client = PageserverClient::new(
355 0 : node.get_id(),
356 0 : node.base_url(),
357 0 : self.service_config.jwt_token.as_deref(),
358 0 : );
359 :
360 0 : let timelines = client.timeline_list(&tenant_shard_id).await?;
361 0 : Ok(timelines
362 0 : .into_iter()
363 0 : .map(|t| (t.timeline_id, t.last_record_lsn))
364 0 : .collect())
365 0 : }
366 :
367 0 : async fn secondary_download(
368 0 : &self,
369 0 : tenant_shard_id: TenantShardId,
370 0 : node: &Node,
371 0 : ) -> Result<(), ReconcileError> {
372 0 : // This is not the timeout for a request, but the total amount of time we're willing to wait
373 0 : // for a secondary location to get up to date before
374 0 : let total_download_timeout = self.reconciler_config.get_secondary_warmup_timeout();
375 0 :
376 0 : // This the long-polling interval for the secondary download requests we send to destination pageserver
377 0 : // during a migration.
378 0 : let request_download_timeout = self
379 0 : .reconciler_config
380 0 : .get_secondary_download_request_timeout();
381 0 :
382 0 : let started_at = Instant::now();
383 :
384 : loop {
385 0 : let (status, progress) = match node
386 0 : .with_client_retries(
387 0 : |client| async move {
388 0 : client
389 0 : .tenant_secondary_download(
390 0 : tenant_shard_id,
391 0 : Some(request_download_timeout),
392 0 : )
393 0 : .await
394 0 : },
395 0 : &self.service_config.jwt_token,
396 0 : 1,
397 0 : 3,
398 0 : request_download_timeout * 2,
399 0 : &self.cancel,
400 0 : )
401 0 : .await
402 : {
403 0 : None => Err(ReconcileError::Cancel),
404 0 : Some(Ok(v)) => Ok(v),
405 0 : Some(Err(e)) => {
406 0 : // Give up, but proceed: it's unfortunate if we couldn't freshen the destination before
407 0 : // attaching, but we should not let an issue with a secondary location stop us proceeding
408 0 : // with a live migration.
409 0 : tracing::warn!("Failed to prepare by downloading layers on node {node}: {e})");
410 0 : return Ok(());
411 : }
412 0 : }?;
413 :
414 0 : if status == StatusCode::OK {
415 0 : tracing::info!(
416 0 : "Downloads to {} complete: {}/{} layers, {}/{} bytes",
417 : node,
418 : progress.layers_downloaded,
419 : progress.layers_total,
420 : progress.bytes_downloaded,
421 : progress.bytes_total
422 : );
423 0 : return Ok(());
424 0 : } else if status == StatusCode::ACCEPTED {
425 0 : let total_runtime = started_at.elapsed();
426 0 : if total_runtime > total_download_timeout {
427 0 : tracing::warn!("Timed out after {}ms downloading layers to {node}. Progress so far: {}/{} layers, {}/{} bytes",
428 0 : total_runtime.as_millis(),
429 : progress.layers_downloaded,
430 : progress.layers_total,
431 : progress.bytes_downloaded,
432 : progress.bytes_total
433 : );
434 : // Give up, but proceed: an incompletely warmed destination doesn't prevent migration working,
435 : // it just makes the I/O performance for users less good.
436 0 : return Ok(());
437 0 : }
438 0 :
439 0 : // Log and proceed around the loop to retry. We don't sleep between requests, because our HTTP call
440 0 : // to the pageserver is a long-poll.
441 0 : tracing::info!(
442 0 : "Downloads to {} not yet complete: {}/{} layers, {}/{} bytes",
443 : node,
444 : progress.layers_downloaded,
445 : progress.layers_total,
446 : progress.bytes_downloaded,
447 : progress.bytes_total
448 : );
449 0 : }
450 : }
451 0 : }
452 :
453 0 : async fn await_lsn(
454 0 : &self,
455 0 : tenant_shard_id: TenantShardId,
456 0 : node: &Node,
457 0 : baseline: HashMap<TimelineId, Lsn>,
458 0 : ) -> anyhow::Result<()> {
459 : loop {
460 0 : let latest = match self.get_lsns(tenant_shard_id, node).await {
461 0 : Ok(l) => l,
462 0 : Err(e) => {
463 0 : tracing::info!("🕑 Can't get LSNs on node {node} yet, waiting ({e})",);
464 0 : tokio::time::sleep(Duration::from_millis(500)).await;
465 0 : continue;
466 : }
467 : };
468 :
469 0 : let mut any_behind: bool = false;
470 0 : for (timeline_id, baseline_lsn) in &baseline {
471 0 : match latest.get(timeline_id) {
472 0 : Some(latest_lsn) => {
473 0 : tracing::info!(timeline_id = %timeline_id, "🕑 LSN origin {baseline_lsn} vs destination {latest_lsn}");
474 0 : if latest_lsn < baseline_lsn {
475 0 : any_behind = true;
476 0 : }
477 : }
478 0 : None => {
479 0 : // Timeline was deleted in the meantime - ignore it
480 0 : }
481 : }
482 : }
483 :
484 0 : if !any_behind {
485 0 : tracing::info!("✅ LSN caught up. Proceeding...");
486 0 : break;
487 : } else {
488 0 : tokio::time::sleep(Duration::from_millis(500)).await;
489 : }
490 : }
491 :
492 0 : Ok(())
493 0 : }
494 :
495 0 : pub async fn live_migrate(
496 0 : &mut self,
497 0 : origin_ps: Node,
498 0 : dest_ps: Node,
499 0 : ) -> Result<(), ReconcileError> {
500 0 : // `maybe_live_migrate` is responsibble for sanity of inputs
501 0 : assert!(origin_ps.get_id() != dest_ps.get_id());
502 :
503 0 : fn build_location_config(
504 0 : shard: &ShardIdentity,
505 0 : config: &TenantConfig,
506 0 : mode: LocationConfigMode,
507 0 : generation: Option<Generation>,
508 0 : secondary_conf: Option<LocationConfigSecondary>,
509 0 : ) -> LocationConfig {
510 0 : LocationConfig {
511 0 : mode,
512 0 : generation: generation.map(|g| g.into().unwrap()),
513 0 : secondary_conf,
514 0 : tenant_conf: config.clone(),
515 0 : shard_number: shard.number.0,
516 0 : shard_count: shard.count.literal(),
517 0 : shard_stripe_size: shard.stripe_size.0,
518 0 : }
519 0 : }
520 :
521 0 : tracing::info!("🔁 Switching origin node {origin_ps} to stale mode",);
522 :
523 : // FIXME: it is incorrect to use self.generation here, we should use the generation
524 : // from the ObservedState of the origin pageserver (it might be older than self.generation)
525 0 : let stale_conf = build_location_config(
526 0 : &self.shard,
527 0 : &self.config,
528 0 : LocationConfigMode::AttachedStale,
529 0 : self.generation,
530 0 : None,
531 0 : );
532 0 : self.location_config(&origin_ps, stale_conf, Some(Duration::from_secs(10)), false)
533 0 : .await?;
534 :
535 0 : let baseline_lsns = Some(self.get_lsns(self.tenant_shard_id, &origin_ps).await?);
536 :
537 : // If we are migrating to a destination that has a secondary location, warm it up first
538 0 : if let Some(destination_conf) = self.observed.locations.get(&dest_ps.get_id()) {
539 0 : if let Some(destination_conf) = &destination_conf.conf {
540 0 : if destination_conf.mode == LocationConfigMode::Secondary {
541 0 : tracing::info!("🔁 Downloading latest layers to destination node {dest_ps}",);
542 0 : self.secondary_download(self.tenant_shard_id, &dest_ps)
543 0 : .await?;
544 0 : }
545 0 : }
546 0 : }
547 :
548 0 : pausable_failpoint!("reconciler-live-migrate-pre-generation-inc");
549 :
550 : // Increment generation before attaching to new pageserver
551 : self.generation = Some(
552 0 : self.persistence
553 0 : .increment_generation(self.tenant_shard_id, dest_ps.get_id())
554 0 : .await?,
555 : );
556 :
557 0 : let dest_conf = build_location_config(
558 0 : &self.shard,
559 0 : &self.config,
560 0 : LocationConfigMode::AttachedMulti,
561 0 : self.generation,
562 0 : None,
563 0 : );
564 0 :
565 0 : tracing::info!("🔁 Attaching to pageserver {dest_ps}");
566 0 : self.location_config(&dest_ps, dest_conf, None, false)
567 0 : .await?;
568 :
569 0 : pausable_failpoint!("reconciler-live-migrate-pre-await-lsn");
570 :
571 0 : if let Some(baseline) = baseline_lsns {
572 0 : tracing::info!("🕑 Waiting for LSN to catch up...");
573 0 : self.await_lsn(self.tenant_shard_id, &dest_ps, baseline)
574 0 : .await?;
575 0 : }
576 :
577 0 : tracing::info!("🔁 Notifying compute to use pageserver {dest_ps}");
578 :
579 : // During a live migration it is unhelpful to proceed if we couldn't notify compute: if we detach
580 : // the origin without notifying compute, we will render the tenant unavailable.
581 0 : self.compute_notify_blocking(&origin_ps).await?;
582 0 : pausable_failpoint!("reconciler-live-migrate-post-notify");
583 :
584 : // Downgrade the origin to secondary. If the tenant's policy is PlacementPolicy::Attached(0), then
585 : // this location will be deleted in the general case reconciliation that runs after this.
586 0 : let origin_secondary_conf = build_location_config(
587 0 : &self.shard,
588 0 : &self.config,
589 0 : LocationConfigMode::Secondary,
590 0 : None,
591 0 : Some(LocationConfigSecondary { warm: true }),
592 0 : );
593 0 : self.location_config(&origin_ps, origin_secondary_conf.clone(), None, false)
594 0 : .await?;
595 : // TODO: we should also be setting the ObservedState on earlier API calls, in case we fail
596 : // partway through. In fact, all location conf API calls should be in a wrapper that sets
597 : // the observed state to None, then runs, then sets it to what we wrote.
598 0 : self.observed.locations.insert(
599 0 : origin_ps.get_id(),
600 0 : ObservedStateLocation {
601 0 : conf: Some(origin_secondary_conf),
602 0 : },
603 0 : );
604 0 :
605 0 : pausable_failpoint!("reconciler-live-migrate-post-detach");
606 :
607 0 : tracing::info!("🔁 Switching to AttachedSingle mode on node {dest_ps}",);
608 0 : let dest_final_conf = build_location_config(
609 0 : &self.shard,
610 0 : &self.config,
611 0 : LocationConfigMode::AttachedSingle,
612 0 : self.generation,
613 0 : None,
614 0 : );
615 0 : self.location_config(&dest_ps, dest_final_conf.clone(), None, false)
616 0 : .await?;
617 0 : self.observed.locations.insert(
618 0 : dest_ps.get_id(),
619 0 : ObservedStateLocation {
620 0 : conf: Some(dest_final_conf),
621 0 : },
622 0 : );
623 0 :
624 0 : tracing::info!("✅ Migration complete");
625 :
626 0 : Ok(())
627 0 : }
628 :
629 0 : async fn maybe_refresh_observed(&mut self) -> Result<(), ReconcileError> {
630 : // If the attached node has uncertain state, read it from the pageserver before proceeding: this
631 : // is important to avoid spurious generation increments.
632 : //
633 : // We don't need to do this for secondary/detach locations because it's harmless to just PUT their
634 : // location conf, whereas for attached locations it can interrupt clients if we spuriously destroy/recreate
635 : // the `Timeline` object in the pageserver.
636 :
637 0 : let Some(attached_node) = self.intent.attached.as_ref() else {
638 : // Nothing to do
639 0 : return Ok(());
640 : };
641 :
642 0 : if matches!(
643 0 : self.observed.locations.get(&attached_node.get_id()),
644 : Some(ObservedStateLocation { conf: None })
645 : ) {
646 0 : let tenant_shard_id = self.tenant_shard_id;
647 0 : let observed_conf = match attached_node
648 0 : .with_client_retries(
649 0 : |client| async move { client.get_location_config(tenant_shard_id).await },
650 0 : &self.service_config.jwt_token,
651 0 : 1,
652 0 : 1,
653 0 : Duration::from_secs(5),
654 0 : &self.cancel,
655 0 : )
656 0 : .await
657 : {
658 0 : Some(Ok(observed)) => Some(observed),
659 0 : Some(Err(mgmt_api::Error::ApiError(status, _msg)))
660 0 : if status == StatusCode::NOT_FOUND =>
661 0 : {
662 0 : None
663 : }
664 0 : Some(Err(e)) => return Err(e.into()),
665 0 : None => return Err(ReconcileError::Cancel),
666 : };
667 0 : tracing::info!("Scanned location configuration on {attached_node}: {observed_conf:?}");
668 0 : match observed_conf {
669 0 : Some(conf) => {
670 0 : // Pageserver returned a state: update it in observed. This may still be an indeterminate (None) state,
671 0 : // if internally the pageserver's TenantSlot was being mutated (e.g. some long running API call is still running)
672 0 : self.observed
673 0 : .locations
674 0 : .insert(attached_node.get_id(), ObservedStateLocation { conf });
675 0 : }
676 0 : None => {
677 0 : // Pageserver returned 404: we have confirmation that there is no state for this shard on that pageserver.
678 0 : self.observed.locations.remove(&attached_node.get_id());
679 0 : }
680 : }
681 0 : }
682 :
683 0 : Ok(())
684 0 : }
685 :
686 : /// Reconciling a tenant makes API calls to pageservers until the observed state
687 : /// matches the intended state.
688 : ///
689 : /// First we apply special case handling (e.g. for live migrations), and then a
690 : /// general case reconciliation where we walk through the intent by pageserver
691 : /// and call out to the pageserver to apply the desired state.
692 0 : pub(crate) async fn reconcile(&mut self) -> Result<(), ReconcileError> {
693 0 : // Prepare: if we have uncertain `observed` state for our would-be attachement location, then refresh it
694 0 : self.maybe_refresh_observed().await?;
695 :
696 : // Special case: live migration
697 0 : self.maybe_live_migrate().await?;
698 :
699 : // If the attached pageserver is not attached, do so now.
700 0 : if let Some(node) = self.intent.attached.as_ref() {
701 : // If we are in an attached policy, then generation must have been set (null generations
702 : // are only present when a tenant is initially loaded with a secondary policy)
703 0 : debug_assert!(self.generation.is_some());
704 0 : let Some(generation) = self.generation else {
705 0 : return Err(ReconcileError::Other(anyhow::anyhow!(
706 0 : "Attempted to attach with NULL generation"
707 0 : )));
708 : };
709 :
710 0 : let mut wanted_conf = attached_location_conf(
711 0 : generation,
712 0 : &self.shard,
713 0 : &self.config,
714 0 : &self.placement_policy,
715 0 : );
716 0 : match self.observed.locations.get(&node.get_id()) {
717 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
718 0 : // Nothing to do
719 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
720 : }
721 0 : observed => {
722 : // In all cases other than a matching observed configuration, we will
723 : // reconcile this location. This includes locations with different configurations, as well
724 : // as locations with unknown (None) observed state.
725 :
726 : // Incrementing generation is the safe general case, but is inefficient for changes that only
727 : // modify some details (e.g. the tenant's config).
728 0 : let increment_generation = match observed {
729 0 : None => true,
730 0 : Some(ObservedStateLocation { conf: None }) => true,
731 : Some(ObservedStateLocation {
732 0 : conf: Some(observed),
733 0 : }) => {
734 0 : let generations_match = observed.generation == wanted_conf.generation;
735 0 :
736 0 : // We may skip incrementing the generation if the location is already in the expected mode and
737 0 : // generation. In principle it would also be safe to skip from certain other modes (e.g. AttachedStale),
738 0 : // but such states are handled inside `live_migrate`, and if we see that state here we're cleaning up
739 0 : // after a restart/crash, so fall back to the universally safe path of incrementing generation.
740 0 : !generations_match || (observed.mode != wanted_conf.mode)
741 : }
742 : };
743 :
744 0 : if increment_generation {
745 0 : let generation = self
746 0 : .persistence
747 0 : .increment_generation(self.tenant_shard_id, node.get_id())
748 0 : .await?;
749 0 : self.generation = Some(generation);
750 0 : wanted_conf.generation = generation.into();
751 0 : }
752 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
753 :
754 : // Because `node` comes from a ref to &self, clone it before calling into a &mut self
755 : // function: this could be avoided by refactoring the state mutated by location_config into
756 : // a separate type to Self.
757 0 : let node = node.clone();
758 0 :
759 0 : // Use lazy=true, because we may run many of Self concurrently, and do not want to
760 0 : // overload the pageserver with logical size calculations.
761 0 : self.location_config(&node, wanted_conf, None, true).await?;
762 0 : self.compute_notify().await?;
763 : }
764 : }
765 0 : }
766 :
767 : // Configure secondary locations: if these were previously attached this
768 : // implicitly downgrades them from attached to secondary.
769 0 : let mut changes = Vec::new();
770 0 : for node in &self.intent.secondary {
771 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
772 0 : match self.observed.locations.get(&node.get_id()) {
773 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
774 0 : // Nothing to do
775 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
776 : }
777 : _ => {
778 : // In all cases other than a matching observed configuration, we will
779 : // reconcile this location.
780 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
781 0 : changes.push((node.clone(), wanted_conf))
782 : }
783 : }
784 : }
785 :
786 : // Detach any extraneous pageservers that are no longer referenced
787 : // by our intent.
788 0 : for node in &self.detach {
789 0 : changes.push((
790 0 : node.clone(),
791 0 : LocationConfig {
792 0 : mode: LocationConfigMode::Detached,
793 0 : generation: None,
794 0 : secondary_conf: None,
795 0 : shard_number: self.shard.number.0,
796 0 : shard_count: self.shard.count.literal(),
797 0 : shard_stripe_size: self.shard.stripe_size.0,
798 0 : tenant_conf: self.config.clone(),
799 0 : },
800 0 : ));
801 0 : }
802 :
803 0 : for (node, conf) in changes {
804 0 : if self.cancel.is_cancelled() {
805 0 : return Err(ReconcileError::Cancel);
806 0 : }
807 0 : self.location_config(&node, conf, None, false).await?;
808 : }
809 :
810 : // The condition below identifies a detach. We must have no attached intent and
811 : // must have been attached to something previously. Pass this information to
812 : // the [`ComputeHook`] such that it can update its tenant-wide state.
813 0 : if self.intent.attached.is_none() && !self.detach.is_empty() {
814 0 : // TODO: Consider notifying control plane about detaches. This would avoid situations
815 0 : // where the compute tries to start-up with a stale set of pageservers.
816 0 : self.compute_hook
817 0 : .handle_detach(self.tenant_shard_id, self.shard.stripe_size);
818 0 : }
819 :
820 0 : failpoint_support::sleep_millis_async!("sleep-on-reconcile-epilogue");
821 :
822 0 : Ok(())
823 0 : }
824 :
825 0 : pub(crate) async fn compute_notify(&mut self) -> Result<(), NotifyError> {
826 : // Whenever a particular Reconciler emits a notification, it is always notifying for the intended
827 : // destination.
828 0 : if let Some(node) = &self.intent.attached {
829 0 : let result = self
830 0 : .compute_hook
831 0 : .notify(
832 0 : self.tenant_shard_id,
833 0 : node.get_id(),
834 0 : self.shard.stripe_size,
835 0 : &self.cancel,
836 0 : )
837 0 : .await;
838 0 : if let Err(e) = &result {
839 : // It is up to the caller whether they want to drop out on this error, but they don't have to:
840 : // in general we should avoid letting unavailability of the cloud control plane stop us from
841 : // making progress.
842 0 : if !matches!(e, NotifyError::ShuttingDown) {
843 0 : tracing::warn!("Failed to notify compute of attached pageserver {node}: {e}");
844 0 : }
845 :
846 : // Set this flag so that in our ReconcileResult we will set the flag on the shard that it
847 : // needs to retry at some point.
848 0 : self.compute_notify_failure = true;
849 0 : }
850 0 : result
851 : } else {
852 0 : Ok(())
853 : }
854 0 : }
855 :
856 : /// Compare the observed state snapshot from when the reconcile was created
857 : /// with the final observed state in order to generate observed state deltas.
858 0 : pub(crate) fn observed_deltas(&self) -> Vec<ObservedStateDelta> {
859 0 : let mut deltas = Vec::default();
860 :
861 0 : for (node_id, location) in &self.observed.locations {
862 0 : let previous_location = self.original_observed.locations.get(node_id);
863 0 : let do_upsert = match previous_location {
864 : // Location config changed for node
865 0 : Some(prev) if location.conf != prev.conf => true,
866 : // New location config for node
867 0 : None => true,
868 : // Location config has not changed for node
869 0 : _ => false,
870 : };
871 :
872 0 : if do_upsert {
873 0 : deltas.push(ObservedStateDelta::Upsert(Box::new((
874 0 : *node_id,
875 0 : location.clone(),
876 0 : ))));
877 0 : }
878 : }
879 :
880 0 : for node_id in self.original_observed.locations.keys() {
881 0 : if !self.observed.locations.contains_key(node_id) {
882 0 : deltas.push(ObservedStateDelta::Delete(*node_id));
883 0 : }
884 : }
885 :
886 0 : deltas
887 0 : }
888 :
889 : /// Keep trying to notify the compute indefinitely, only dropping out if:
890 : /// - the node `origin` becomes unavailable -> Ok(())
891 : /// - the node `origin` no longer has our tenant shard attached -> Ok(())
892 : /// - our cancellation token fires -> Err(ReconcileError::Cancelled)
893 : ///
894 : /// This is used during live migration, where we do not wish to detach
895 : /// an origin location until the compute definitely knows about the new
896 : /// location.
897 : ///
898 : /// In cases where the origin node becomes unavailable, we return success, indicating
899 : /// to the caller that they should continue irrespective of whether the compute was notified,
900 : /// because the origin node is unusable anyway. Notification will be retried later via the
901 : /// [`Self::compute_notify_failure`] flag.
902 0 : async fn compute_notify_blocking(&mut self, origin: &Node) -> Result<(), ReconcileError> {
903 0 : let mut notify_attempts = 0;
904 0 : while let Err(e) = self.compute_notify().await {
905 0 : match e {
906 0 : NotifyError::Fatal(_) => return Err(ReconcileError::Notify(e)),
907 0 : NotifyError::ShuttingDown => return Err(ReconcileError::Cancel),
908 : _ => {
909 0 : tracing::warn!(
910 0 : "Live migration blocked by compute notification error, retrying: {e}"
911 : );
912 : }
913 : }
914 :
915 : // Did the origin pageserver become unavailable?
916 0 : if !origin.is_available() {
917 0 : tracing::info!("Giving up on compute notification because {origin} is unavailable");
918 0 : break;
919 0 : }
920 0 :
921 0 : // Does the origin pageserver still host the shard we are interested in? We should only
922 0 : // continue waiting for compute notification to be acked if the old location is still usable.
923 0 : let tenant_shard_id = self.tenant_shard_id;
924 0 : match origin
925 0 : .with_client_retries(
926 0 : |client| async move { client.get_location_config(tenant_shard_id).await },
927 0 : &self.service_config.jwt_token,
928 0 : 1,
929 0 : 3,
930 0 : Duration::from_secs(5),
931 0 : &self.cancel,
932 0 : )
933 0 : .await
934 : {
935 0 : Some(Ok(Some(location_conf))) => {
936 0 : if matches!(
937 0 : location_conf.mode,
938 : LocationConfigMode::AttachedMulti
939 : | LocationConfigMode::AttachedSingle
940 : | LocationConfigMode::AttachedStale
941 : ) {
942 0 : tracing::debug!(
943 0 : "Still attached to {origin}, will wait & retry compute notification"
944 : );
945 : } else {
946 0 : tracing::info!(
947 0 : "Giving up on compute notification because {origin} is in state {:?}",
948 : location_conf.mode
949 : );
950 0 : return Ok(());
951 : }
952 : // Fall through
953 : }
954 : Some(Ok(None)) => {
955 0 : tracing::info!(
956 0 : "No longer attached to {origin}, giving up on compute notification"
957 : );
958 0 : return Ok(());
959 : }
960 0 : Some(Err(e)) => {
961 0 : match e {
962 : mgmt_api::Error::Cancelled => {
963 0 : tracing::info!(
964 0 : "Giving up on compute notification because {origin} is unavailable"
965 : );
966 0 : return Ok(());
967 : }
968 : mgmt_api::Error::ApiError(StatusCode::NOT_FOUND, _) => {
969 0 : tracing::info!(
970 0 : "No longer attached to {origin}, giving up on compute notification"
971 : );
972 0 : return Ok(());
973 : }
974 0 : e => {
975 0 : // Other API errors are unexpected here.
976 0 : tracing::warn!("Unexpected error checking location on {origin}: {e}");
977 :
978 : // Fall through, we will retry compute notification.
979 : }
980 : }
981 : }
982 0 : None => return Err(ReconcileError::Cancel),
983 : };
984 :
985 0 : exponential_backoff(
986 0 : notify_attempts,
987 0 : // Generous waits: control plane operations which might be blocking us usually complete on the order
988 0 : // of hundreds to thousands of milliseconds, so no point busy polling.
989 0 : 1.0,
990 0 : 10.0,
991 0 : &self.cancel,
992 0 : )
993 0 : .await;
994 0 : notify_attempts += 1;
995 : }
996 :
997 0 : Ok(())
998 0 : }
999 : }
1000 :
1001 : /// We tweak the externally-set TenantConfig while configuring
1002 : /// locations, using our awareness of whether secondary locations
1003 : /// are in use to automatically enable/disable heatmap uploads.
1004 0 : fn ha_aware_config(config: &TenantConfig, has_secondaries: bool) -> TenantConfig {
1005 0 : let mut config = config.clone();
1006 0 : if has_secondaries {
1007 0 : if config.heatmap_period.is_none() {
1008 0 : config.heatmap_period = Some(DEFAULT_HEATMAP_PERIOD.to_string());
1009 0 : }
1010 0 : } else {
1011 0 : config.heatmap_period = None;
1012 0 : }
1013 0 : config
1014 0 : }
1015 :
1016 0 : pub(crate) fn attached_location_conf(
1017 0 : generation: Generation,
1018 0 : shard: &ShardIdentity,
1019 0 : config: &TenantConfig,
1020 0 : policy: &PlacementPolicy,
1021 0 : ) -> LocationConfig {
1022 0 : let has_secondaries = match policy {
1023 : PlacementPolicy::Attached(0) | PlacementPolicy::Detached | PlacementPolicy::Secondary => {
1024 0 : false
1025 : }
1026 0 : PlacementPolicy::Attached(_) => true,
1027 : };
1028 :
1029 0 : LocationConfig {
1030 0 : mode: LocationConfigMode::AttachedSingle,
1031 0 : generation: generation.into(),
1032 0 : secondary_conf: None,
1033 0 : shard_number: shard.number.0,
1034 0 : shard_count: shard.count.literal(),
1035 0 : shard_stripe_size: shard.stripe_size.0,
1036 0 : tenant_conf: ha_aware_config(config, has_secondaries),
1037 0 : }
1038 0 : }
1039 :
1040 0 : pub(crate) fn secondary_location_conf(
1041 0 : shard: &ShardIdentity,
1042 0 : config: &TenantConfig,
1043 0 : ) -> LocationConfig {
1044 0 : LocationConfig {
1045 0 : mode: LocationConfigMode::Secondary,
1046 0 : generation: None,
1047 0 : secondary_conf: Some(LocationConfigSecondary { warm: true }),
1048 0 : shard_number: shard.number.0,
1049 0 : shard_count: shard.count.literal(),
1050 0 : shard_stripe_size: shard.stripe_size.0,
1051 0 : tenant_conf: ha_aware_config(config, true),
1052 0 : }
1053 0 : }
|
