Line data Source code
1 : use crate::pageserver_client::PageserverClient;
2 : use crate::persistence::Persistence;
3 : use crate::service;
4 : use hyper::StatusCode;
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 std::collections::HashMap;
11 : use std::sync::Arc;
12 : use std::time::{Duration, Instant};
13 : use tokio_util::sync::CancellationToken;
14 : use utils::generation::Generation;
15 : use utils::id::{NodeId, TimelineId};
16 : use utils::lsn::Lsn;
17 : use utils::sync::gate::GateGuard;
18 :
19 : use crate::compute_hook::{ComputeHook, NotifyError};
20 : use crate::node::Node;
21 : use crate::tenant_shard::{IntentState, ObservedState, ObservedStateLocation};
22 :
23 : const DEFAULT_HEATMAP_PERIOD: &str = "60s";
24 :
25 : /// Object with the lifetime of the background reconcile task that is created
26 : /// for tenants which have a difference between their intent and observed states.
27 : pub(super) struct Reconciler {
28 : /// See [`crate::tenant_shard::TenantShard`] for the meanings of these fields: they are a snapshot
29 : /// of a tenant's state from when we spawned a reconcile task.
30 : pub(super) tenant_shard_id: TenantShardId,
31 : pub(crate) shard: ShardIdentity,
32 : pub(crate) generation: Option<Generation>,
33 : pub(crate) intent: TargetState,
34 :
35 : /// Nodes not referenced by [`Self::intent`], from which we should try
36 : /// to detach this tenant shard.
37 : pub(crate) detach: Vec<Node>,
38 :
39 : pub(crate) config: TenantConfig,
40 : pub(crate) observed: ObservedState,
41 :
42 : pub(crate) service_config: service::Config,
43 :
44 : /// A hook to notify the running postgres instances when we change the location
45 : /// of a tenant. Use this via [`Self::compute_notify`] to update our failure flag
46 : /// and guarantee eventual retries.
47 : pub(crate) compute_hook: Arc<ComputeHook>,
48 :
49 : /// To avoid stalling if the cloud control plane is unavailable, we may proceed
50 : /// past failures in [`ComputeHook::notify`], but we _must_ remember that we failed
51 : /// so that we can set [`crate::tenant_shard::TenantShard::pending_compute_notification`] to ensure a later retry.
52 : pub(crate) compute_notify_failure: bool,
53 :
54 : /// A means to abort background reconciliation: it is essential to
55 : /// call this when something changes in the original TenantShard that
56 : /// will make this reconciliation impossible or unnecessary, for
57 : /// example when a pageserver node goes offline, or the PlacementPolicy for
58 : /// the tenant is changed.
59 : pub(crate) cancel: CancellationToken,
60 :
61 : /// Reconcilers are registered with a Gate so that during a graceful shutdown we
62 : /// can wait for all the reconcilers to respond to their cancellation tokens.
63 : pub(crate) _gate_guard: GateGuard,
64 :
65 : /// Access to persistent storage for updating generation numbers
66 : pub(crate) persistence: Arc<Persistence>,
67 : }
68 :
69 : /// This is a snapshot of [`crate::tenant_shard::IntentState`], but it does not do any
70 : /// reference counting for Scheduler. The IntentState is what the scheduler works with,
71 : /// and the TargetState is just the instruction for a particular Reconciler run.
72 : #[derive(Debug)]
73 : pub(crate) struct TargetState {
74 : pub(crate) attached: Option<Node>,
75 : pub(crate) secondary: Vec<Node>,
76 : }
77 :
78 : impl TargetState {
79 0 : pub(crate) fn from_intent(nodes: &HashMap<NodeId, Node>, intent: &IntentState) -> Self {
80 0 : Self {
81 0 : attached: intent.get_attached().map(|n| {
82 0 : nodes
83 0 : .get(&n)
84 0 : .expect("Intent attached referenced non-existent node")
85 0 : .clone()
86 0 : }),
87 0 : secondary: intent
88 0 : .get_secondary()
89 0 : .iter()
90 0 : .map(|n| {
91 0 : nodes
92 0 : .get(n)
93 0 : .expect("Intent secondary referenced non-existent node")
94 0 : .clone()
95 0 : })
96 0 : .collect(),
97 0 : }
98 0 : }
99 : }
100 :
101 0 : #[derive(thiserror::Error, Debug)]
102 : pub(crate) enum ReconcileError {
103 : #[error(transparent)]
104 : Remote(#[from] mgmt_api::Error),
105 : #[error(transparent)]
106 : Notify(#[from] NotifyError),
107 : #[error("Cancelled")]
108 : Cancel,
109 : #[error(transparent)]
110 : Other(#[from] anyhow::Error),
111 : }
112 :
113 : impl Reconciler {
114 0 : async fn location_config(
115 0 : &mut self,
116 0 : node: &Node,
117 0 : config: LocationConfig,
118 0 : flush_ms: Option<Duration>,
119 0 : lazy: bool,
120 0 : ) -> Result<(), ReconcileError> {
121 0 : if !node.is_available() && config.mode == LocationConfigMode::Detached {
122 : // Attempts to detach from offline nodes may be imitated without doing I/O: a node which is offline
123 : // will get fully reconciled wrt the shard's intent state when it is reactivated, irrespective of
124 : // what we put into `observed`, in [`crate::service::Service::node_activate_reconcile`]
125 0 : tracing::info!("Node {node} is unavailable during detach: proceeding anyway, it will be detached on next activation");
126 0 : self.observed.locations.remove(&node.get_id());
127 0 : return Ok(());
128 0 : }
129 0 :
130 0 : self.observed
131 0 : .locations
132 0 : .insert(node.get_id(), ObservedStateLocation { conf: None });
133 0 :
134 0 : // TODO: amend locations that use long-polling: they will hit this timeout.
135 0 : let timeout = Duration::from_secs(25);
136 0 :
137 0 : tracing::info!("location_config({node}) calling: {:?}", config);
138 0 : let tenant_shard_id = self.tenant_shard_id;
139 0 : let config_ref = &config;
140 0 : match node
141 0 : .with_client_retries(
142 0 : |client| async move {
143 0 : let config = config_ref.clone();
144 0 : client
145 0 : .location_config(tenant_shard_id, config.clone(), flush_ms, lazy)
146 0 : .await
147 0 : },
148 0 : &self.service_config.jwt_token,
149 0 : 1,
150 0 : 3,
151 0 : timeout,
152 0 : &self.cancel,
153 0 : )
154 0 : .await
155 : {
156 0 : Some(Ok(_)) => {}
157 0 : Some(Err(e)) => return Err(e.into()),
158 0 : None => return Err(ReconcileError::Cancel),
159 : };
160 0 : tracing::info!("location_config({node}) complete: {:?}", config);
161 :
162 0 : match config.mode {
163 0 : LocationConfigMode::Detached => {
164 0 : self.observed.locations.remove(&node.get_id());
165 0 : }
166 0 : _ => {
167 0 : self.observed
168 0 : .locations
169 0 : .insert(node.get_id(), ObservedStateLocation { conf: Some(config) });
170 0 : }
171 : }
172 :
173 0 : Ok(())
174 0 : }
175 :
176 0 : fn get_node(&self, node_id: &NodeId) -> Option<&Node> {
177 0 : if let Some(node) = self.intent.attached.as_ref() {
178 0 : if node.get_id() == *node_id {
179 0 : return Some(node);
180 0 : }
181 0 : }
182 :
183 0 : if let Some(node) = self
184 0 : .intent
185 0 : .secondary
186 0 : .iter()
187 0 : .find(|n| n.get_id() == *node_id)
188 : {
189 0 : return Some(node);
190 0 : }
191 :
192 0 : if let Some(node) = self.detach.iter().find(|n| n.get_id() == *node_id) {
193 0 : return Some(node);
194 0 : }
195 0 :
196 0 : None
197 0 : }
198 :
199 0 : async fn maybe_live_migrate(&mut self) -> Result<(), ReconcileError> {
200 0 : let destination = if let Some(node) = &self.intent.attached {
201 0 : match self.observed.locations.get(&node.get_id()) {
202 0 : Some(conf) => {
203 : // We will do a live migration only if the intended destination is not
204 : // currently in an attached state.
205 0 : match &conf.conf {
206 0 : Some(conf) if conf.mode == LocationConfigMode::Secondary => {
207 0 : // Fall through to do a live migration
208 0 : node
209 : }
210 : None | Some(_) => {
211 : // Attached or uncertain: don't do a live migration, proceed
212 : // with a general-case reconciliation
213 0 : tracing::info!("maybe_live_migrate: destination is None or attached");
214 0 : return Ok(());
215 : }
216 : }
217 : }
218 : None => {
219 : // Our destination is not attached: maybe live migrate if some other
220 : // node is currently attached. Fall through.
221 0 : node
222 : }
223 : }
224 : } else {
225 : // No intent to be attached
226 0 : tracing::info!("maybe_live_migrate: no attached intent");
227 0 : return Ok(());
228 : };
229 :
230 0 : let mut origin = None;
231 0 : for (node_id, state) in &self.observed.locations {
232 0 : if let Some(observed_conf) = &state.conf {
233 0 : if observed_conf.mode == LocationConfigMode::AttachedSingle {
234 : // We will only attempt live migration if the origin is not offline: this
235 : // avoids trying to do it while reconciling after responding to an HA failover.
236 0 : if let Some(node) = self.get_node(node_id) {
237 0 : if node.is_available() {
238 0 : origin = Some(node.clone());
239 0 : break;
240 0 : }
241 0 : }
242 0 : }
243 0 : }
244 : }
245 :
246 0 : let Some(origin) = origin else {
247 0 : tracing::info!("maybe_live_migrate: no origin found");
248 0 : return Ok(());
249 : };
250 :
251 : // We have an origin and a destination: proceed to do the live migration
252 0 : tracing::info!("Live migrating {}->{}", origin, destination);
253 0 : self.live_migrate(origin, destination.clone()).await?;
254 :
255 0 : Ok(())
256 0 : }
257 :
258 0 : async fn get_lsns(
259 0 : &self,
260 0 : tenant_shard_id: TenantShardId,
261 0 : node: &Node,
262 0 : ) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
263 0 : let client = PageserverClient::new(
264 0 : node.get_id(),
265 0 : node.base_url(),
266 0 : self.service_config.jwt_token.as_deref(),
267 0 : );
268 :
269 0 : let timelines = client.timeline_list(&tenant_shard_id).await?;
270 0 : Ok(timelines
271 0 : .into_iter()
272 0 : .map(|t| (t.timeline_id, t.last_record_lsn))
273 0 : .collect())
274 0 : }
275 :
276 0 : async fn secondary_download(
277 0 : &self,
278 0 : tenant_shard_id: TenantShardId,
279 0 : node: &Node,
280 0 : ) -> Result<(), ReconcileError> {
281 0 : // This is not the timeout for a request, but the total amount of time we're willing to wait
282 0 : // for a secondary location to get up to date before
283 0 : const TOTAL_DOWNLOAD_TIMEOUT: Duration = Duration::from_secs(300);
284 0 :
285 0 : // This the long-polling interval for the secondary download requests we send to destination pageserver
286 0 : // during a migration.
287 0 : const REQUEST_DOWNLOAD_TIMEOUT: Duration = Duration::from_secs(20);
288 0 :
289 0 : let started_at = Instant::now();
290 :
291 : loop {
292 0 : let (status, progress) = match node
293 0 : .with_client_retries(
294 0 : |client| async move {
295 0 : client
296 0 : .tenant_secondary_download(
297 0 : tenant_shard_id,
298 0 : Some(REQUEST_DOWNLOAD_TIMEOUT),
299 0 : )
300 0 : .await
301 0 : },
302 0 : &self.service_config.jwt_token,
303 0 : 1,
304 0 : 3,
305 0 : REQUEST_DOWNLOAD_TIMEOUT * 2,
306 0 : &self.cancel,
307 0 : )
308 0 : .await
309 : {
310 0 : None => Err(ReconcileError::Cancel),
311 0 : Some(Ok(v)) => Ok(v),
312 0 : Some(Err(e)) => {
313 0 : // Give up, but proceed: it's unfortunate if we couldn't freshen the destination before
314 0 : // attaching, but we should not let an issue with a secondary location stop us proceeding
315 0 : // with a live migration.
316 0 : tracing::warn!("Failed to prepare by downloading layers on node {node}: {e})");
317 0 : return Ok(());
318 : }
319 0 : }?;
320 :
321 0 : if status == StatusCode::OK {
322 0 : tracing::info!(
323 0 : "Downloads to {} complete: {}/{} layers, {}/{} bytes",
324 0 : node,
325 0 : progress.layers_downloaded,
326 0 : progress.layers_total,
327 0 : progress.bytes_downloaded,
328 0 : progress.bytes_total
329 0 : );
330 0 : return Ok(());
331 0 : } else if status == StatusCode::ACCEPTED {
332 0 : let total_runtime = started_at.elapsed();
333 0 : if total_runtime > TOTAL_DOWNLOAD_TIMEOUT {
334 0 : tracing::warn!("Timed out after {}ms downloading layers to {node}. Progress so far: {}/{} layers, {}/{} bytes",
335 0 : total_runtime.as_millis(),
336 0 : progress.layers_downloaded,
337 0 : progress.layers_total,
338 0 : progress.bytes_downloaded,
339 0 : progress.bytes_total
340 0 : );
341 : // Give up, but proceed: an incompletely warmed destination doesn't prevent migration working,
342 : // it just makes the I/O performance for users less good.
343 0 : return Ok(());
344 0 : }
345 0 :
346 0 : // Log and proceed around the loop to retry. We don't sleep between requests, because our HTTP call
347 0 : // to the pageserver is a long-poll.
348 0 : tracing::info!(
349 0 : "Downloads to {} not yet complete: {}/{} layers, {}/{} bytes",
350 0 : node,
351 0 : progress.layers_downloaded,
352 0 : progress.layers_total,
353 0 : progress.bytes_downloaded,
354 0 : progress.bytes_total
355 0 : );
356 0 : }
357 : }
358 0 : }
359 :
360 0 : async fn await_lsn(
361 0 : &self,
362 0 : tenant_shard_id: TenantShardId,
363 0 : node: &Node,
364 0 : baseline: HashMap<TimelineId, Lsn>,
365 0 : ) -> anyhow::Result<()> {
366 : loop {
367 0 : let latest = match self.get_lsns(tenant_shard_id, node).await {
368 0 : Ok(l) => l,
369 0 : Err(e) => {
370 0 : tracing::info!("🕑 Can't get LSNs on node {node} yet, waiting ({e})",);
371 0 : std::thread::sleep(Duration::from_millis(500));
372 0 : continue;
373 : }
374 : };
375 :
376 0 : let mut any_behind: bool = false;
377 0 : for (timeline_id, baseline_lsn) in &baseline {
378 0 : match latest.get(timeline_id) {
379 0 : Some(latest_lsn) => {
380 0 : tracing::info!("🕑 LSN origin {baseline_lsn} vs destination {latest_lsn}");
381 0 : if latest_lsn < baseline_lsn {
382 0 : any_behind = true;
383 0 : }
384 : }
385 0 : None => {
386 0 : // Expected timeline isn't yet visible on migration destination.
387 0 : // (IRL we would have to account for timeline deletion, but this
388 0 : // is just test helper)
389 0 : any_behind = true;
390 0 : }
391 : }
392 : }
393 :
394 0 : if !any_behind {
395 0 : tracing::info!("✅ LSN caught up. Proceeding...");
396 0 : break;
397 0 : } else {
398 0 : std::thread::sleep(Duration::from_millis(500));
399 0 : }
400 : }
401 :
402 0 : Ok(())
403 0 : }
404 :
405 0 : pub async fn live_migrate(
406 0 : &mut self,
407 0 : origin_ps: Node,
408 0 : dest_ps: Node,
409 0 : ) -> Result<(), ReconcileError> {
410 0 : // `maybe_live_migrate` is responsibble for sanity of inputs
411 0 : assert!(origin_ps.get_id() != dest_ps.get_id());
412 :
413 0 : fn build_location_config(
414 0 : shard: &ShardIdentity,
415 0 : config: &TenantConfig,
416 0 : mode: LocationConfigMode,
417 0 : generation: Option<Generation>,
418 0 : secondary_conf: Option<LocationConfigSecondary>,
419 0 : ) -> LocationConfig {
420 0 : LocationConfig {
421 0 : mode,
422 0 : generation: generation.map(|g| g.into().unwrap()),
423 0 : secondary_conf,
424 0 : tenant_conf: config.clone(),
425 0 : shard_number: shard.number.0,
426 0 : shard_count: shard.count.literal(),
427 0 : shard_stripe_size: shard.stripe_size.0,
428 0 : }
429 0 : }
430 :
431 0 : tracing::info!("🔁 Switching origin node {origin_ps} to stale mode",);
432 :
433 : // FIXME: it is incorrect to use self.generation here, we should use the generation
434 : // from the ObservedState of the origin pageserver (it might be older than self.generation)
435 0 : let stale_conf = build_location_config(
436 0 : &self.shard,
437 0 : &self.config,
438 0 : LocationConfigMode::AttachedStale,
439 0 : self.generation,
440 0 : None,
441 0 : );
442 0 : self.location_config(&origin_ps, stale_conf, Some(Duration::from_secs(10)), false)
443 0 : .await?;
444 :
445 0 : let baseline_lsns = Some(self.get_lsns(self.tenant_shard_id, &origin_ps).await?);
446 :
447 : // If we are migrating to a destination that has a secondary location, warm it up first
448 0 : if let Some(destination_conf) = self.observed.locations.get(&dest_ps.get_id()) {
449 0 : if let Some(destination_conf) = &destination_conf.conf {
450 0 : if destination_conf.mode == LocationConfigMode::Secondary {
451 0 : tracing::info!("🔁 Downloading latest layers to destination node {dest_ps}",);
452 0 : self.secondary_download(self.tenant_shard_id, &dest_ps)
453 0 : .await?;
454 0 : }
455 0 : }
456 0 : }
457 :
458 : // Increment generation before attaching to new pageserver
459 : self.generation = Some(
460 0 : self.persistence
461 0 : .increment_generation(self.tenant_shard_id, dest_ps.get_id())
462 0 : .await?,
463 : );
464 :
465 0 : let dest_conf = build_location_config(
466 0 : &self.shard,
467 0 : &self.config,
468 0 : LocationConfigMode::AttachedMulti,
469 0 : self.generation,
470 0 : None,
471 0 : );
472 0 :
473 0 : tracing::info!("🔁 Attaching to pageserver {dest_ps}");
474 0 : self.location_config(&dest_ps, dest_conf, None, false)
475 0 : .await?;
476 :
477 0 : if let Some(baseline) = baseline_lsns {
478 0 : tracing::info!("🕑 Waiting for LSN to catch up...");
479 0 : self.await_lsn(self.tenant_shard_id, &dest_ps, baseline)
480 0 : .await?;
481 0 : }
482 :
483 0 : tracing::info!("🔁 Notifying compute to use pageserver {dest_ps}");
484 :
485 : // During a live migration it is unhelpful to proceed if we couldn't notify compute: if we detach
486 : // the origin without notifying compute, we will render the tenant unavailable.
487 0 : while let Err(e) = self.compute_notify().await {
488 0 : match e {
489 0 : NotifyError::Fatal(_) => return Err(ReconcileError::Notify(e)),
490 0 : NotifyError::ShuttingDown => return Err(ReconcileError::Cancel),
491 : _ => {
492 0 : tracing::warn!(
493 0 : "Live migration blocked by compute notification error, retrying: {e}"
494 0 : );
495 : }
496 : }
497 : }
498 :
499 : // Downgrade the origin to secondary. If the tenant's policy is PlacementPolicy::Attached(0), then
500 : // this location will be deleted in the general case reconciliation that runs after this.
501 0 : let origin_secondary_conf = build_location_config(
502 0 : &self.shard,
503 0 : &self.config,
504 0 : LocationConfigMode::Secondary,
505 0 : None,
506 0 : Some(LocationConfigSecondary { warm: true }),
507 0 : );
508 0 : self.location_config(&origin_ps, origin_secondary_conf.clone(), None, false)
509 0 : .await?;
510 : // TODO: we should also be setting the ObservedState on earlier API calls, in case we fail
511 : // partway through. In fact, all location conf API calls should be in a wrapper that sets
512 : // the observed state to None, then runs, then sets it to what we wrote.
513 0 : self.observed.locations.insert(
514 0 : origin_ps.get_id(),
515 0 : ObservedStateLocation {
516 0 : conf: Some(origin_secondary_conf),
517 0 : },
518 0 : );
519 0 :
520 0 : tracing::info!("🔁 Switching to AttachedSingle mode on node {dest_ps}",);
521 0 : let dest_final_conf = build_location_config(
522 0 : &self.shard,
523 0 : &self.config,
524 0 : LocationConfigMode::AttachedSingle,
525 0 : self.generation,
526 0 : None,
527 0 : );
528 0 : self.location_config(&dest_ps, dest_final_conf.clone(), None, false)
529 0 : .await?;
530 0 : self.observed.locations.insert(
531 0 : dest_ps.get_id(),
532 0 : ObservedStateLocation {
533 0 : conf: Some(dest_final_conf),
534 0 : },
535 0 : );
536 0 :
537 0 : tracing::info!("✅ Migration complete");
538 :
539 0 : Ok(())
540 0 : }
541 :
542 0 : async fn maybe_refresh_observed(&mut self) -> Result<(), ReconcileError> {
543 : // If the attached node has uncertain state, read it from the pageserver before proceeding: this
544 : // is important to avoid spurious generation increments.
545 : //
546 : // We don't need to do this for secondary/detach locations because it's harmless to just PUT their
547 : // location conf, whereas for attached locations it can interrupt clients if we spuriously destroy/recreate
548 : // the `Timeline` object in the pageserver.
549 :
550 0 : let Some(attached_node) = self.intent.attached.as_ref() else {
551 : // Nothing to do
552 0 : return Ok(());
553 : };
554 :
555 0 : if matches!(
556 0 : self.observed.locations.get(&attached_node.get_id()),
557 : Some(ObservedStateLocation { conf: None })
558 : ) {
559 0 : let tenant_shard_id = self.tenant_shard_id;
560 0 : let observed_conf = match attached_node
561 0 : .with_client_retries(
562 0 : |client| async move { client.get_location_config(tenant_shard_id).await },
563 0 : &self.service_config.jwt_token,
564 0 : 1,
565 0 : 1,
566 0 : Duration::from_secs(5),
567 0 : &self.cancel,
568 0 : )
569 0 : .await
570 : {
571 0 : Some(Ok(observed)) => Some(observed),
572 0 : Some(Err(mgmt_api::Error::ApiError(status, _msg)))
573 0 : if status == StatusCode::NOT_FOUND =>
574 0 : {
575 0 : None
576 : }
577 0 : Some(Err(e)) => return Err(e.into()),
578 0 : None => return Err(ReconcileError::Cancel),
579 : };
580 0 : tracing::info!("Scanned location configuration on {attached_node}: {observed_conf:?}");
581 0 : match observed_conf {
582 0 : Some(conf) => {
583 0 : // Pageserver returned a state: update it in observed. This may still be an indeterminate (None) state,
584 0 : // if internally the pageserver's TenantSlot was being mutated (e.g. some long running API call is still running)
585 0 : self.observed
586 0 : .locations
587 0 : .insert(attached_node.get_id(), ObservedStateLocation { conf });
588 0 : }
589 0 : None => {
590 0 : // Pageserver returned 404: we have confirmation that there is no state for this shard on that pageserver.
591 0 : self.observed.locations.remove(&attached_node.get_id());
592 0 : }
593 : }
594 0 : }
595 :
596 0 : Ok(())
597 0 : }
598 :
599 : /// Reconciling a tenant makes API calls to pageservers until the observed state
600 : /// matches the intended state.
601 : ///
602 : /// First we apply special case handling (e.g. for live migrations), and then a
603 : /// general case reconciliation where we walk through the intent by pageserver
604 : /// and call out to the pageserver to apply the desired state.
605 0 : pub(crate) async fn reconcile(&mut self) -> Result<(), ReconcileError> {
606 0 : // Prepare: if we have uncertain `observed` state for our would-be attachement location, then refresh it
607 0 : self.maybe_refresh_observed().await?;
608 :
609 : // Special case: live migration
610 0 : self.maybe_live_migrate().await?;
611 :
612 : // If the attached pageserver is not attached, do so now.
613 0 : if let Some(node) = self.intent.attached.as_ref() {
614 : // If we are in an attached policy, then generation must have been set (null generations
615 : // are only present when a tenant is initially loaded with a secondary policy)
616 0 : debug_assert!(self.generation.is_some());
617 0 : let Some(generation) = self.generation else {
618 0 : return Err(ReconcileError::Other(anyhow::anyhow!(
619 0 : "Attempted to attach with NULL generation"
620 0 : )));
621 : };
622 :
623 0 : let mut wanted_conf = attached_location_conf(
624 0 : generation,
625 0 : &self.shard,
626 0 : &self.config,
627 0 : !self.intent.secondary.is_empty(),
628 0 : );
629 0 : match self.observed.locations.get(&node.get_id()) {
630 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
631 0 : // Nothing to do
632 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
633 : }
634 0 : observed => {
635 : // In all cases other than a matching observed configuration, we will
636 : // reconcile this location. This includes locations with different configurations, as well
637 : // as locations with unknown (None) observed state.
638 :
639 : // The general case is to increment the generation. However, there are cases
640 : // where this is not necessary:
641 : // - if we are only updating the TenantConf part of the location
642 : // - if we are only changing the attachment mode (e.g. going to attachedmulti or attachedstale)
643 : // and the location was already in the correct generation
644 0 : let increment_generation = match observed {
645 0 : None => true,
646 0 : Some(ObservedStateLocation { conf: None }) => true,
647 : Some(ObservedStateLocation {
648 0 : conf: Some(observed),
649 0 : }) => {
650 0 : let generations_match = observed.generation == wanted_conf.generation;
651 :
652 : use LocationConfigMode::*;
653 0 : let mode_transition_requires_gen_inc =
654 0 : match (observed.mode, wanted_conf.mode) {
655 : // Usually the short-lived attachment modes (multi and stale) are only used
656 : // in the case of [`Self::live_migrate`], but it is simple to handle them correctly
657 : // here too. Locations are allowed to go Single->Stale and Multi->Single within the same generation.
658 0 : (AttachedSingle, AttachedStale) => false,
659 0 : (AttachedMulti, AttachedSingle) => false,
660 0 : (lhs, rhs) => lhs != rhs,
661 : };
662 :
663 0 : !generations_match || mode_transition_requires_gen_inc
664 : }
665 : };
666 :
667 0 : if increment_generation {
668 0 : let generation = self
669 0 : .persistence
670 0 : .increment_generation(self.tenant_shard_id, node.get_id())
671 0 : .await?;
672 0 : self.generation = Some(generation);
673 0 : wanted_conf.generation = generation.into();
674 0 : }
675 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
676 :
677 : // Because `node` comes from a ref to &self, clone it before calling into a &mut self
678 : // function: this could be avoided by refactoring the state mutated by location_config into
679 : // a separate type to Self.
680 0 : let node = node.clone();
681 0 :
682 0 : // Use lazy=true, because we may run many of Self concurrently, and do not want to
683 0 : // overload the pageserver with logical size calculations.
684 0 : self.location_config(&node, wanted_conf, None, true).await?;
685 0 : self.compute_notify().await?;
686 : }
687 : }
688 0 : }
689 :
690 : // Configure secondary locations: if these were previously attached this
691 : // implicitly downgrades them from attached to secondary.
692 0 : let mut changes = Vec::new();
693 0 : for node in &self.intent.secondary {
694 0 : let wanted_conf = secondary_location_conf(&self.shard, &self.config);
695 0 : match self.observed.locations.get(&node.get_id()) {
696 0 : Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
697 0 : // Nothing to do
698 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration already correct.")
699 : }
700 : _ => {
701 : // In all cases other than a matching observed configuration, we will
702 : // reconcile this location.
703 0 : tracing::info!(node_id=%node.get_id(), "Observed configuration requires update.");
704 0 : changes.push((node.clone(), wanted_conf))
705 : }
706 : }
707 : }
708 :
709 : // Detach any extraneous pageservers that are no longer referenced
710 : // by our intent.
711 0 : for node in &self.detach {
712 0 : changes.push((
713 0 : node.clone(),
714 0 : LocationConfig {
715 0 : mode: LocationConfigMode::Detached,
716 0 : generation: None,
717 0 : secondary_conf: None,
718 0 : shard_number: self.shard.number.0,
719 0 : shard_count: self.shard.count.literal(),
720 0 : shard_stripe_size: self.shard.stripe_size.0,
721 0 : tenant_conf: self.config.clone(),
722 0 : },
723 0 : ));
724 0 : }
725 :
726 0 : for (node, conf) in changes {
727 0 : if self.cancel.is_cancelled() {
728 0 : return Err(ReconcileError::Cancel);
729 0 : }
730 0 : self.location_config(&node, conf, None, false).await?;
731 : }
732 :
733 0 : Ok(())
734 0 : }
735 :
736 0 : pub(crate) async fn compute_notify(&mut self) -> Result<(), NotifyError> {
737 : // Whenever a particular Reconciler emits a notification, it is always notifying for the intended
738 : // destination.
739 0 : if let Some(node) = &self.intent.attached {
740 0 : let result = self
741 0 : .compute_hook
742 0 : .notify(
743 0 : self.tenant_shard_id,
744 0 : node.get_id(),
745 0 : self.shard.stripe_size,
746 0 : &self.cancel,
747 0 : )
748 0 : .await;
749 0 : if let Err(e) = &result {
750 : // It is up to the caller whether they want to drop out on this error, but they don't have to:
751 : // in general we should avoid letting unavailability of the cloud control plane stop us from
752 : // making progress.
753 0 : tracing::warn!("Failed to notify compute of attached pageserver {node}: {e}");
754 : // Set this flag so that in our ReconcileResult we will set the flag on the shard that it
755 : // needs to retry at some point.
756 0 : self.compute_notify_failure = true;
757 0 : }
758 0 : result
759 : } else {
760 0 : Ok(())
761 : }
762 0 : }
763 : }
764 :
765 : /// We tweak the externally-set TenantConfig while configuring
766 : /// locations, using our awareness of whether secondary locations
767 : /// are in use to automatically enable/disable heatmap uploads.
768 0 : fn ha_aware_config(config: &TenantConfig, has_secondaries: bool) -> TenantConfig {
769 0 : let mut config = config.clone();
770 0 : if has_secondaries {
771 0 : if config.heatmap_period.is_none() {
772 0 : config.heatmap_period = Some(DEFAULT_HEATMAP_PERIOD.to_string());
773 0 : }
774 0 : } else {
775 0 : config.heatmap_period = None;
776 0 : }
777 0 : config
778 0 : }
779 :
780 0 : pub(crate) fn attached_location_conf(
781 0 : generation: Generation,
782 0 : shard: &ShardIdentity,
783 0 : config: &TenantConfig,
784 0 : has_secondaries: bool,
785 0 : ) -> LocationConfig {
786 0 : LocationConfig {
787 0 : mode: LocationConfigMode::AttachedSingle,
788 0 : generation: generation.into(),
789 0 : secondary_conf: None,
790 0 : shard_number: shard.number.0,
791 0 : shard_count: shard.count.literal(),
792 0 : shard_stripe_size: shard.stripe_size.0,
793 0 : tenant_conf: ha_aware_config(config, has_secondaries),
794 0 : }
795 0 : }
796 :
797 0 : pub(crate) fn secondary_location_conf(
798 0 : shard: &ShardIdentity,
799 0 : config: &TenantConfig,
800 0 : ) -> LocationConfig {
801 0 : LocationConfig {
802 0 : mode: LocationConfigMode::Secondary,
803 0 : generation: None,
804 0 : secondary_conf: Some(LocationConfigSecondary { warm: true }),
805 0 : shard_number: shard.number.0,
806 0 : shard_count: shard.count.literal(),
807 0 : shard_stripe_size: shard.stripe_size.0,
808 0 : tenant_conf: ha_aware_config(config, true),
809 0 : }
810 0 : }
|
