Line data Source code
1 : //! WAL receiver logic that ensures the pageserver gets connectected to safekeeper,
2 : //! that contains the latest WAL to stream and this connection does not go stale.
3 : //!
4 : //! To achieve that, a storage broker is used: safekepers propagate their timelines' state in it,
5 : //! the manager subscribes for changes and accumulates those to query the one with the biggest Lsn for connection.
6 : //! Current connection state is tracked too, to ensure it's not getting stale.
7 : //!
8 : //! After every connection or storage broker update fetched, the state gets updated correspondingly and rechecked for the new conneciton leader,
9 : //! then a (re)connection happens, if necessary.
10 : //! Only WAL streaming task expects to be finished, other loops (storage broker, connection management) never exit unless cancelled explicitly via the dedicated channel.
11 :
12 : use std::{collections::HashMap, num::NonZeroU64, ops::ControlFlow, sync::Arc, time::Duration};
13 :
14 : use super::{TaskStateUpdate, WalReceiverConf};
15 : use crate::context::{DownloadBehavior, RequestContext};
16 : use crate::metrics::{
17 : WALRECEIVER_ACTIVE_MANAGERS, WALRECEIVER_BROKER_UPDATES, WALRECEIVER_CANDIDATES_ADDED,
18 : WALRECEIVER_CANDIDATES_REMOVED, WALRECEIVER_SWITCHES,
19 : };
20 : use crate::task_mgr::TaskKind;
21 : use crate::tenant::{debug_assert_current_span_has_tenant_and_timeline_id, Timeline};
22 : use anyhow::Context;
23 : use chrono::{NaiveDateTime, Utc};
24 : use pageserver_api::models::TimelineState;
25 :
26 : use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
27 : use storage_broker::proto::{
28 : FilterTenantTimelineId, MessageType, SafekeeperDiscoveryRequest, SafekeeperDiscoveryResponse,
29 : SubscribeByFilterRequest, TypeSubscription, TypedMessage,
30 : };
31 : use storage_broker::{BrokerClientChannel, Code, Streaming};
32 : use tokio_util::sync::CancellationToken;
33 : use tracing::*;
34 :
35 : use postgres_connection::PgConnectionConfig;
36 : use utils::backoff::{
37 : exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
38 : };
39 : use utils::postgres_client::{
40 : wal_stream_connection_config, ConnectionConfigArgs, PostgresClientProtocol,
41 : };
42 : use utils::{
43 : id::{NodeId, TenantTimelineId},
44 : lsn::Lsn,
45 : };
46 :
47 : use super::{
48 : walreceiver_connection::WalConnectionStatus, walreceiver_connection::WalReceiverError,
49 : TaskEvent, TaskHandle,
50 : };
51 :
52 : pub(crate) struct Cancelled;
53 :
54 : /// Attempts to subscribe for timeline updates, pushed by safekeepers into the broker.
55 : /// Based on the updates, desides whether to start, keep or stop a WAL receiver task.
56 : /// If storage broker subscription is cancelled, exits.
57 : ///
58 : /// # Cancel-Safety
59 : ///
60 : /// Not cancellation-safe. Use `cancel` token to request cancellation.
61 0 : pub(super) async fn connection_manager_loop_step(
62 0 : broker_client: &mut BrokerClientChannel,
63 0 : connection_manager_state: &mut ConnectionManagerState,
64 0 : ctx: &RequestContext,
65 0 : cancel: &CancellationToken,
66 0 : manager_status: &std::sync::RwLock<Option<ConnectionManagerStatus>>,
67 0 : ) -> Result<(), Cancelled> {
68 0 : match tokio::select! {
69 0 : _ = cancel.cancelled() => { return Err(Cancelled); },
70 0 : st = connection_manager_state.timeline.wait_to_become_active(ctx) => { st }
71 : } {
72 0 : Ok(()) => {}
73 0 : Err(new_state) => {
74 0 : debug!(
75 : ?new_state,
76 0 : "state changed, stopping wal connection manager loop"
77 : );
78 0 : return Err(Cancelled);
79 : }
80 : }
81 :
82 0 : WALRECEIVER_ACTIVE_MANAGERS.inc();
83 0 : scopeguard::defer! {
84 0 : WALRECEIVER_ACTIVE_MANAGERS.dec();
85 0 : }
86 0 :
87 0 : let id = TenantTimelineId {
88 0 : tenant_id: connection_manager_state.timeline.tenant_shard_id.tenant_id,
89 0 : timeline_id: connection_manager_state.timeline.timeline_id,
90 0 : };
91 0 :
92 0 : let mut timeline_state_updates = connection_manager_state
93 0 : .timeline
94 0 : .subscribe_for_state_updates();
95 0 :
96 0 : let mut wait_lsn_status = connection_manager_state
97 0 : .timeline
98 0 : .subscribe_for_wait_lsn_updates();
99 0 :
100 0 : // TODO: create a separate config option for discovery request interval
101 0 : let discovery_request_interval = connection_manager_state.conf.lagging_wal_timeout;
102 0 : let mut last_discovery_ts: Option<std::time::Instant> = None;
103 :
104 : // Subscribe to the broker updates. Stream shares underlying TCP connection
105 : // with other streams on this client (other connection managers). When
106 : // object goes out of scope, stream finishes in drop() automatically.
107 0 : let mut broker_subscription = subscribe_for_timeline_updates(broker_client, id, cancel).await?;
108 0 : debug!("Subscribed for broker timeline updates");
109 :
110 : loop {
111 0 : let time_until_next_retry = connection_manager_state.time_until_next_retry();
112 0 : let any_activity = connection_manager_state.wal_connection.is_some()
113 0 : || !connection_manager_state.wal_stream_candidates.is_empty();
114 :
115 : // These things are happening concurrently:
116 : //
117 : // - cancellation request
118 : // - keep receiving WAL on the current connection
119 : // - if the shared state says we need to change connection, disconnect and return
120 : // - this runs in a separate task and we receive updates via a watch channel
121 : // - change connection if the rules decide so, or if the current connection dies
122 : // - receive updates from broker
123 : // - this might change the current desired connection
124 : // - timeline state changes to something that does not allow walreceiver to run concurrently
125 : // - if there's no connection and no candidates, try to send a discovery request
126 :
127 : // NB: make sure each of the select expressions are cancellation-safe
128 : // (no need for arms to be cancellation-safe).
129 0 : tokio::select! {
130 0 : _ = cancel.cancelled() => { return Err(Cancelled); }
131 0 : Some(wal_connection_update) = async {
132 0 : match connection_manager_state.wal_connection.as_mut() {
133 0 : Some(wal_connection) => Some(wal_connection.connection_task.next_task_event().await),
134 0 : None => None,
135 : }
136 0 : } => {
137 0 : let wal_connection = connection_manager_state.wal_connection.as_mut()
138 0 : .expect("Should have a connection, as checked by the corresponding select! guard");
139 0 : match wal_connection_update {
140 0 : TaskEvent::Update(TaskStateUpdate::Started) => {},
141 0 : TaskEvent::Update(TaskStateUpdate::Progress(new_status)) => {
142 0 : if new_status.has_processed_wal {
143 0 : // We have advanced last_record_lsn by processing the WAL received
144 0 : // from this safekeeper. This is good enough to clean unsuccessful
145 0 : // retries history and allow reconnecting to this safekeeper without
146 0 : // sleeping for a long time.
147 0 : connection_manager_state.wal_connection_retries.remove(&wal_connection.sk_id);
148 0 : }
149 0 : wal_connection.status = new_status;
150 : }
151 0 : TaskEvent::End(walreceiver_task_result) => {
152 0 : match walreceiver_task_result {
153 0 : Ok(()) => debug!("WAL receiving task finished"),
154 0 : Err(e) => error!("wal receiver task finished with an error: {e:?}"),
155 : }
156 0 : connection_manager_state.drop_old_connection(false).await;
157 : },
158 : }
159 : },
160 :
161 : // Got a new update from the broker
162 0 : broker_update = broker_subscription.message() /* TODO: review cancellation-safety */ => {
163 0 : match broker_update {
164 0 : Ok(Some(broker_update)) => connection_manager_state.register_timeline_update(broker_update),
165 0 : Err(status) => {
166 0 : match status.code() {
167 0 : Code::Unknown if status.message().contains("stream closed because of a broken pipe") || status.message().contains("connection reset") || status.message().contains("error reading a body from connection") => {
168 0 : // tonic's error handling doesn't provide a clear code for disconnections: we get
169 0 : // "h2 protocol error: error reading a body from connection: stream closed because of a broken pipe"
170 0 : // => https://github.com/neondatabase/neon/issues/9562
171 0 : info!("broker disconnected: {status}");
172 : },
173 : _ => {
174 0 : warn!("broker subscription failed: {status}");
175 : }
176 : }
177 0 : return Ok(());
178 : }
179 : Ok(None) => {
180 0 : error!("broker subscription stream ended"); // can't happen
181 0 : return Ok(());
182 : }
183 : }
184 : },
185 :
186 0 : new_event = async {
187 : // Reminder: this match arm needs to be cancellation-safe.
188 : loop {
189 0 : if connection_manager_state.timeline.current_state() == TimelineState::Loading {
190 0 : warn!("wal connection manager should only be launched after timeline has become active");
191 0 : }
192 0 : match timeline_state_updates.changed().await {
193 : Ok(()) => {
194 0 : let new_state = connection_manager_state.timeline.current_state();
195 0 : match new_state {
196 : // we're already active as walreceiver, no need to reactivate
197 0 : TimelineState::Active => continue,
198 : TimelineState::Broken { .. } | TimelineState::Stopping => {
199 0 : debug!("timeline entered terminal state {new_state:?}, stopping wal connection manager loop");
200 0 : return ControlFlow::Break(());
201 : }
202 : TimelineState::Loading => {
203 0 : warn!("timeline transitioned back to Loading state, that should not happen");
204 0 : return ControlFlow::Continue(());
205 : }
206 : }
207 : }
208 0 : Err(_sender_dropped_error) => return ControlFlow::Break(()),
209 : }
210 : }
211 0 : } => match new_event {
212 : ControlFlow::Continue(()) => {
213 0 : return Ok(());
214 : }
215 : ControlFlow::Break(()) => {
216 0 : debug!("Timeline is no longer active, stopping wal connection manager loop");
217 0 : return Err(Cancelled);
218 : }
219 : },
220 :
221 0 : Some(()) = async {
222 0 : match time_until_next_retry {
223 0 : Some(sleep_time) => {
224 0 : tokio::time::sleep(sleep_time).await;
225 0 : Some(())
226 : },
227 : None => {
228 0 : debug!("No candidates to retry, waiting indefinitely for the broker events");
229 0 : None
230 : }
231 : }
232 0 : } => debug!("Waking up for the next retry after waiting for {time_until_next_retry:?}"),
233 :
234 0 : Some(()) = async {
235 0 : // Reminder: this match arm needs to be cancellation-safe.
236 0 : // Calculating time needed to wait until sending the next discovery request.
237 0 : // Current implementation is conservative and sends discovery requests only when there are no candidates.
238 0 :
239 0 : if any_activity {
240 : // No need to send discovery requests if there is an active connection or candidates.
241 0 : return None;
242 0 : }
243 :
244 : // Waiting for an active wait_lsn request.
245 0 : while wait_lsn_status.borrow().is_none() {
246 0 : if wait_lsn_status.changed().await.is_err() {
247 : // wait_lsn_status channel was closed, exiting
248 0 : warn!("wait_lsn_status channel was closed in connection_manager_loop_step");
249 0 : return None;
250 0 : }
251 : }
252 :
253 : // All preconditions met, preparing to send a discovery request.
254 0 : let now = std::time::Instant::now();
255 0 : let next_discovery_ts = last_discovery_ts
256 0 : .map(|ts| ts + discovery_request_interval)
257 0 : .unwrap_or_else(|| now);
258 0 :
259 0 : if next_discovery_ts > now {
260 : // Prevent sending discovery requests too frequently.
261 0 : tokio::time::sleep(next_discovery_ts - now).await;
262 0 : }
263 :
264 0 : let tenant_timeline_id = Some(ProtoTenantTimelineId {
265 0 : tenant_id: id.tenant_id.as_ref().to_owned(),
266 0 : timeline_id: id.timeline_id.as_ref().to_owned(),
267 0 : });
268 0 : let request = SafekeeperDiscoveryRequest { tenant_timeline_id };
269 0 : let msg = TypedMessage {
270 0 : r#type: MessageType::SafekeeperDiscoveryRequest as i32,
271 0 : safekeeper_timeline_info: None,
272 0 : safekeeper_discovery_request: Some(request),
273 0 : safekeeper_discovery_response: None,
274 0 : };
275 0 :
276 0 : last_discovery_ts = Some(std::time::Instant::now());
277 0 : info!("No active connection and no candidates, sending discovery request to the broker");
278 :
279 : // Cancellation safety: we want to send a message to the broker, but publish_one()
280 : // function can get cancelled by the other select! arm. This is absolutely fine, because
281 : // we just want to receive broker updates and discovery is not important if we already
282 : // receive updates.
283 : //
284 : // It is possible that `last_discovery_ts` will be updated, but the message will not be sent.
285 : // This is totally fine because of the reason above.
286 :
287 : // This is a fire-and-forget request, we don't care about the response
288 0 : let _ = broker_client.publish_one(msg).await;
289 0 : debug!("Discovery request sent to the broker");
290 0 : None
291 0 : } => {}
292 : }
293 :
294 0 : if let Some(new_candidate) = connection_manager_state.next_connection_candidate() {
295 0 : info!("Switching to new connection candidate: {new_candidate:?}");
296 0 : connection_manager_state
297 0 : .change_connection(new_candidate, ctx)
298 0 : .await
299 0 : }
300 0 : *manager_status.write().unwrap() = Some(connection_manager_state.manager_status());
301 : }
302 0 : }
303 :
304 : /// Endlessly try to subscribe for broker updates for a given timeline.
305 0 : async fn subscribe_for_timeline_updates(
306 0 : broker_client: &mut BrokerClientChannel,
307 0 : id: TenantTimelineId,
308 0 : cancel: &CancellationToken,
309 0 : ) -> Result<Streaming<TypedMessage>, Cancelled> {
310 0 : let mut attempt = 0;
311 : loop {
312 0 : exponential_backoff(
313 0 : attempt,
314 0 : DEFAULT_BASE_BACKOFF_SECONDS,
315 0 : DEFAULT_MAX_BACKOFF_SECONDS,
316 0 : cancel,
317 0 : )
318 0 : .await;
319 0 : attempt += 1;
320 0 :
321 0 : // subscribe to the specific timeline
322 0 : let request = SubscribeByFilterRequest {
323 0 : types: vec![
324 0 : TypeSubscription {
325 0 : r#type: MessageType::SafekeeperTimelineInfo as i32,
326 0 : },
327 0 : TypeSubscription {
328 0 : r#type: MessageType::SafekeeperDiscoveryResponse as i32,
329 0 : },
330 0 : ],
331 0 : tenant_timeline_id: Some(FilterTenantTimelineId {
332 0 : enabled: true,
333 0 : tenant_timeline_id: Some(ProtoTenantTimelineId {
334 0 : tenant_id: id.tenant_id.as_ref().to_owned(),
335 0 : timeline_id: id.timeline_id.as_ref().to_owned(),
336 0 : }),
337 0 : }),
338 0 : };
339 0 :
340 0 : match {
341 0 : tokio::select! {
342 0 : r = broker_client.subscribe_by_filter(request) => { r }
343 0 : _ = cancel.cancelled() => { return Err(Cancelled); }
344 : }
345 : } {
346 0 : Ok(resp) => {
347 0 : return Ok(resp.into_inner());
348 : }
349 0 : Err(e) => {
350 0 : // Safekeeper nodes can stop pushing timeline updates to the broker, when no new writes happen and
351 0 : // entire WAL is streamed. Keep this noticeable with logging, but do not warn/error.
352 0 : info!("Attempt #{attempt}, failed to subscribe for timeline {id} updates in broker: {e:#}");
353 0 : continue;
354 : }
355 : }
356 : }
357 0 : }
358 :
359 : const WALCONNECTION_RETRY_MIN_BACKOFF_SECONDS: f64 = 0.1;
360 : const WALCONNECTION_RETRY_MAX_BACKOFF_SECONDS: f64 = 15.0;
361 : const WALCONNECTION_RETRY_BACKOFF_MULTIPLIER: f64 = 1.5;
362 :
363 : /// All data that's needed to run endless broker loop and keep the WAL streaming connection alive, if possible.
364 : pub(super) struct ConnectionManagerState {
365 : id: TenantTimelineId,
366 : /// Use pageserver data about the timeline to filter out some of the safekeepers.
367 : timeline: Arc<Timeline>,
368 : /// Child token of [`super::WalReceiver::cancel`], inherited to all tasks we spawn.
369 : cancel: CancellationToken,
370 : conf: WalReceiverConf,
371 : /// Current connection to safekeeper for WAL streaming.
372 : wal_connection: Option<WalConnection>,
373 : /// Info about retries and unsuccessful attempts to connect to safekeepers.
374 : wal_connection_retries: HashMap<NodeId, RetryInfo>,
375 : /// Data about all timelines, available for connection, fetched from storage broker, grouped by their corresponding safekeeper node id.
376 : wal_stream_candidates: HashMap<NodeId, BrokerSkTimeline>,
377 : }
378 :
379 : /// An information about connection manager's current connection and connection candidates.
380 : #[derive(Debug, Clone)]
381 : pub struct ConnectionManagerStatus {
382 : existing_connection: Option<WalConnectionStatus>,
383 : wal_stream_candidates: HashMap<NodeId, BrokerSkTimeline>,
384 : }
385 :
386 : impl ConnectionManagerStatus {
387 : /// Generates a string, describing current connection status in a form, suitable for logging.
388 0 : pub fn to_human_readable_string(&self) -> String {
389 0 : let mut resulting_string = String::new();
390 0 : match &self.existing_connection {
391 0 : Some(connection) => {
392 0 : if connection.has_processed_wal {
393 0 : resulting_string.push_str(&format!(
394 0 : " (update {}): streaming WAL from node {}, ",
395 0 : connection.latest_wal_update.format("%Y-%m-%d %H:%M:%S"),
396 0 : connection.node,
397 0 : ));
398 0 :
399 0 : match (connection.streaming_lsn, connection.commit_lsn) {
400 0 : (None, None) => resulting_string.push_str("no streaming data"),
401 0 : (None, Some(commit_lsn)) => {
402 0 : resulting_string.push_str(&format!("commit Lsn: {commit_lsn}"))
403 : }
404 0 : (Some(streaming_lsn), None) => {
405 0 : resulting_string.push_str(&format!("streaming Lsn: {streaming_lsn}"))
406 : }
407 0 : (Some(streaming_lsn), Some(commit_lsn)) => resulting_string.push_str(
408 0 : &format!("commit|streaming Lsn: {commit_lsn}|{streaming_lsn}"),
409 0 : ),
410 : }
411 0 : } else if connection.is_connected {
412 0 : resulting_string.push_str(&format!(
413 0 : " (update {}): connecting to node {}",
414 0 : connection
415 0 : .latest_connection_update
416 0 : .format("%Y-%m-%d %H:%M:%S"),
417 0 : connection.node,
418 0 : ));
419 0 : } else {
420 0 : resulting_string.push_str(&format!(
421 0 : " (update {}): initializing node {} connection",
422 0 : connection
423 0 : .latest_connection_update
424 0 : .format("%Y-%m-%d %H:%M:%S"),
425 0 : connection.node,
426 0 : ));
427 0 : }
428 : }
429 0 : None => resulting_string.push_str(": disconnected"),
430 : }
431 :
432 0 : resulting_string.push_str(", safekeeper candidates (id|update_time|commit_lsn): [");
433 0 : let mut candidates = self.wal_stream_candidates.iter().peekable();
434 0 : while let Some((node_id, candidate_info)) = candidates.next() {
435 0 : resulting_string.push_str(&format!(
436 0 : "({}|{}|{})",
437 0 : node_id,
438 0 : candidate_info.latest_update.format("%H:%M:%S"),
439 0 : Lsn(candidate_info.timeline.commit_lsn)
440 0 : ));
441 0 : if candidates.peek().is_some() {
442 0 : resulting_string.push_str(", ");
443 0 : }
444 : }
445 0 : resulting_string.push(']');
446 0 :
447 0 : resulting_string
448 0 : }
449 : }
450 :
451 : /// Current connection data.
452 : #[derive(Debug)]
453 : struct WalConnection {
454 : /// Time when the connection was initiated.
455 : started_at: NaiveDateTime,
456 : /// Current safekeeper pageserver is connected to for WAL streaming.
457 : sk_id: NodeId,
458 : /// Availability zone of the safekeeper.
459 : availability_zone: Option<String>,
460 : /// Status of the connection.
461 : status: WalConnectionStatus,
462 : /// WAL streaming task handle.
463 : connection_task: TaskHandle<WalConnectionStatus>,
464 : /// Have we discovered that other safekeeper has more recent WAL than we do?
465 : discovered_new_wal: Option<NewCommittedWAL>,
466 : }
467 :
468 : /// Notion of a new committed WAL, which exists on other safekeeper.
469 : #[derive(Debug, Clone, Copy)]
470 : struct NewCommittedWAL {
471 : /// LSN of the new committed WAL.
472 : lsn: Lsn,
473 : /// When we discovered that the new committed WAL exists on other safekeeper.
474 : discovered_at: NaiveDateTime,
475 : }
476 :
477 : #[derive(Debug, Clone, Copy)]
478 : struct RetryInfo {
479 : next_retry_at: Option<NaiveDateTime>,
480 : retry_duration_seconds: f64,
481 : }
482 :
483 : /// Data about the timeline to connect to, received from the broker.
484 : #[derive(Debug, Clone)]
485 : struct BrokerSkTimeline {
486 : timeline: SafekeeperDiscoveryResponse,
487 : /// Time at which the data was fetched from the broker last time, to track the stale data.
488 : latest_update: NaiveDateTime,
489 : }
490 :
491 : impl ConnectionManagerState {
492 0 : pub(super) fn new(
493 0 : timeline: Arc<Timeline>,
494 0 : conf: WalReceiverConf,
495 0 : cancel: CancellationToken,
496 0 : ) -> Self {
497 0 : let id = TenantTimelineId {
498 0 : tenant_id: timeline.tenant_shard_id.tenant_id,
499 0 : timeline_id: timeline.timeline_id,
500 0 : };
501 0 : Self {
502 0 : id,
503 0 : timeline,
504 0 : cancel,
505 0 : conf,
506 0 : wal_connection: None,
507 0 : wal_stream_candidates: HashMap::new(),
508 0 : wal_connection_retries: HashMap::new(),
509 0 : }
510 0 : }
511 :
512 20 : fn spawn<Fut>(
513 20 : &self,
514 20 : task: impl FnOnce(
515 20 : tokio::sync::watch::Sender<TaskStateUpdate<WalConnectionStatus>>,
516 20 : CancellationToken,
517 20 : ) -> Fut
518 20 : + Send
519 20 : + 'static,
520 20 : ) -> TaskHandle<WalConnectionStatus>
521 20 : where
522 20 : Fut: std::future::Future<Output = anyhow::Result<()>> + Send,
523 20 : {
524 20 : // TODO: get rid of TaskHandle
525 20 : super::TaskHandle::spawn(&self.cancel, task)
526 20 : }
527 :
528 : /// Shuts down the current connection (if any) and immediately starts another one with the given connection string.
529 0 : async fn change_connection(&mut self, new_sk: NewWalConnectionCandidate, ctx: &RequestContext) {
530 0 : WALRECEIVER_SWITCHES
531 0 : .with_label_values(&[new_sk.reason.name()])
532 0 : .inc();
533 0 :
534 0 : self.drop_old_connection(true).await;
535 :
536 0 : let node_id = new_sk.safekeeper_id;
537 0 : let connect_timeout = self.conf.wal_connect_timeout;
538 0 : let ingest_batch_size = self.conf.ingest_batch_size;
539 0 : let protocol = self.conf.protocol;
540 0 : let timeline = Arc::clone(&self.timeline);
541 0 : let ctx = ctx.detached_child(
542 0 : TaskKind::WalReceiverConnectionHandler,
543 0 : DownloadBehavior::Download,
544 0 : );
545 :
546 0 : let span = info_span!("connection", %node_id);
547 0 : let connection_handle = self.spawn(move |events_sender, cancellation| {
548 0 : async move {
549 0 : debug_assert_current_span_has_tenant_and_timeline_id();
550 :
551 0 : let res = super::walreceiver_connection::handle_walreceiver_connection(
552 0 : timeline,
553 0 : protocol,
554 0 : new_sk.wal_source_connconf,
555 0 : events_sender,
556 0 : cancellation.clone(),
557 0 : connect_timeout,
558 0 : ctx,
559 0 : node_id,
560 0 : ingest_batch_size,
561 0 : )
562 0 : .await;
563 :
564 0 : match res {
565 0 : Ok(()) => Ok(()),
566 0 : Err(e) => {
567 0 : match e {
568 0 : WalReceiverError::SuccessfulCompletion(msg) => {
569 0 : info!("walreceiver connection handling ended with success: {msg}");
570 0 : Ok(())
571 : }
572 0 : WalReceiverError::ExpectedSafekeeperError(e) => {
573 0 : info!("walreceiver connection handling ended: {e}");
574 0 : Ok(())
575 : }
576 : WalReceiverError::ClosedGate => {
577 0 : info!(
578 0 : "walreceiver connection handling ended because of closed gate"
579 : );
580 0 : Ok(())
581 : }
582 0 : WalReceiverError::Other(e) => {
583 0 : // give out an error to have task_mgr give it a really verbose logging
584 0 : if cancellation.is_cancelled() {
585 : // Ideally we would learn about this via some path other than Other, but
586 : // that requires refactoring all the intermediate layers of ingest code
587 : // that only emit anyhow::Error
588 0 : Ok(())
589 : } else {
590 0 : Err(e).context("walreceiver connection handling failure")
591 : }
592 : }
593 : }
594 : }
595 : }
596 0 : }
597 0 : .instrument(span)
598 0 : });
599 0 :
600 0 : let now = Utc::now().naive_utc();
601 0 : self.wal_connection = Some(WalConnection {
602 0 : started_at: now,
603 0 : sk_id: new_sk.safekeeper_id,
604 0 : availability_zone: new_sk.availability_zone,
605 0 : status: WalConnectionStatus {
606 0 : is_connected: false,
607 0 : has_processed_wal: false,
608 0 : latest_connection_update: now,
609 0 : latest_wal_update: now,
610 0 : streaming_lsn: None,
611 0 : commit_lsn: None,
612 0 : node: node_id,
613 0 : },
614 0 : connection_task: connection_handle,
615 0 : discovered_new_wal: None,
616 0 : });
617 0 : }
618 :
619 : /// Drops the current connection (if any) and updates retry timeout for the next
620 : /// connection attempt to the same safekeeper.
621 : ///
622 : /// # Cancel-Safety
623 : ///
624 : /// Not cancellation-safe.
625 0 : async fn drop_old_connection(&mut self, needs_shutdown: bool) {
626 0 : let wal_connection = match self.wal_connection.take() {
627 0 : Some(wal_connection) => wal_connection,
628 0 : None => return,
629 : };
630 :
631 0 : if needs_shutdown {
632 0 : wal_connection
633 0 : .connection_task
634 0 : .shutdown()
635 0 : // This here is why this function isn't cancellation-safe.
636 0 : // If we got cancelled here, then self.wal_connection is already None and we lose track of the task.
637 0 : // Even if our caller diligently calls Self::shutdown(), it will find a self.wal_connection=None
638 0 : // and thus be ineffective.
639 0 : .await;
640 0 : }
641 :
642 0 : let retry = self
643 0 : .wal_connection_retries
644 0 : .entry(wal_connection.sk_id)
645 0 : .or_insert(RetryInfo {
646 0 : next_retry_at: None,
647 0 : retry_duration_seconds: WALCONNECTION_RETRY_MIN_BACKOFF_SECONDS,
648 0 : });
649 0 :
650 0 : let now = Utc::now().naive_utc();
651 0 :
652 0 : // Schedule the next retry attempt. We want to have exponential backoff for connection attempts,
653 0 : // and we add backoff to the time when we started the connection attempt. If the connection
654 0 : // was active for a long time, then next_retry_at will be in the past.
655 0 : retry.next_retry_at =
656 0 : wal_connection
657 0 : .started_at
658 0 : .checked_add_signed(chrono::Duration::milliseconds(
659 0 : (retry.retry_duration_seconds * 1000.0) as i64,
660 0 : ));
661 :
662 0 : if let Some(next) = &retry.next_retry_at {
663 0 : if next > &now {
664 0 : info!(
665 0 : "Next connection retry to {:?} is at {}",
666 : wal_connection.sk_id, next
667 : );
668 0 : }
669 0 : }
670 :
671 0 : let next_retry_duration =
672 0 : retry.retry_duration_seconds * WALCONNECTION_RETRY_BACKOFF_MULTIPLIER;
673 0 : // Clamp the next retry duration to the maximum allowed.
674 0 : let next_retry_duration = next_retry_duration.min(WALCONNECTION_RETRY_MAX_BACKOFF_SECONDS);
675 0 : // Clamp the next retry duration to the minimum allowed.
676 0 : let next_retry_duration = next_retry_duration.max(WALCONNECTION_RETRY_MIN_BACKOFF_SECONDS);
677 0 :
678 0 : retry.retry_duration_seconds = next_retry_duration;
679 0 : }
680 :
681 : /// Returns time needed to wait to have a new candidate for WAL streaming.
682 0 : fn time_until_next_retry(&self) -> Option<Duration> {
683 0 : let now = Utc::now().naive_utc();
684 :
685 0 : let next_retry_at = self
686 0 : .wal_connection_retries
687 0 : .values()
688 0 : .filter_map(|retry| retry.next_retry_at)
689 0 : .filter(|next_retry_at| next_retry_at > &now)
690 0 : .min()?;
691 :
692 0 : (next_retry_at - now).to_std().ok()
693 0 : }
694 :
695 : /// Adds another broker timeline into the state, if its more recent than the one already added there for the same key.
696 0 : fn register_timeline_update(&mut self, typed_msg: TypedMessage) {
697 0 : let mut is_discovery = false;
698 0 : let timeline_update = match typed_msg.r#type() {
699 : MessageType::SafekeeperTimelineInfo => {
700 0 : let info = match typed_msg.safekeeper_timeline_info {
701 0 : Some(info) => info,
702 : None => {
703 0 : warn!("bad proto message from broker: no safekeeper_timeline_info");
704 0 : return;
705 : }
706 : };
707 0 : SafekeeperDiscoveryResponse {
708 0 : safekeeper_id: info.safekeeper_id,
709 0 : tenant_timeline_id: info.tenant_timeline_id,
710 0 : commit_lsn: info.commit_lsn,
711 0 : safekeeper_connstr: info.safekeeper_connstr,
712 0 : availability_zone: info.availability_zone,
713 0 : standby_horizon: info.standby_horizon,
714 0 : }
715 : }
716 : MessageType::SafekeeperDiscoveryResponse => {
717 0 : is_discovery = true;
718 0 : match typed_msg.safekeeper_discovery_response {
719 0 : Some(response) => response,
720 : None => {
721 0 : warn!("bad proto message from broker: no safekeeper_discovery_response");
722 0 : return;
723 : }
724 : }
725 : }
726 : _ => {
727 : // unexpected message
728 0 : return;
729 : }
730 : };
731 :
732 0 : WALRECEIVER_BROKER_UPDATES.inc();
733 0 :
734 0 : trace!(
735 0 : "safekeeper info update: standby_horizon(cutoff)={}",
736 : timeline_update.standby_horizon
737 : );
738 0 : if timeline_update.standby_horizon != 0 {
739 0 : // ignore reports from safekeepers not connected to replicas
740 0 : self.timeline
741 0 : .standby_horizon
742 0 : .store(Lsn(timeline_update.standby_horizon));
743 0 : self.timeline
744 0 : .metrics
745 0 : .standby_horizon_gauge
746 0 : .set(timeline_update.standby_horizon as i64);
747 0 : }
748 :
749 0 : let new_safekeeper_id = NodeId(timeline_update.safekeeper_id);
750 0 : let old_entry = self.wal_stream_candidates.insert(
751 0 : new_safekeeper_id,
752 0 : BrokerSkTimeline {
753 0 : timeline: timeline_update,
754 0 : latest_update: Utc::now().naive_utc(),
755 0 : },
756 0 : );
757 0 :
758 0 : if old_entry.is_none() {
759 0 : info!(
760 : ?is_discovery,
761 : %new_safekeeper_id,
762 0 : "New SK node was added",
763 : );
764 0 : WALRECEIVER_CANDIDATES_ADDED.inc();
765 0 : }
766 0 : }
767 :
768 : /// Cleans up stale broker records and checks the rest for the new connection candidate.
769 : /// Returns a new candidate, if the current state is absent or somewhat lagging, `None` otherwise.
770 : /// The current rules for approving new candidates:
771 : /// * pick a candidate different from the connected safekeeper with biggest `commit_lsn` and lowest failed connection attemps
772 : /// * if there's no such entry, no new candidate found, abort
773 : /// * otherwise check if the candidate is much better than the current one
774 : ///
775 : /// To understand exact rules for determining if the candidate is better than the current one, refer to this function's implementation.
776 : /// General rules are following:
777 : /// * if connected safekeeper is not present, pick the candidate
778 : /// * if we haven't received any updates for some time, pick the candidate
779 : /// * if the candidate commit_lsn is much higher than the current one, pick the candidate
780 : /// * if the candidate commit_lsn is same, but candidate is located in the same AZ as the pageserver, pick the candidate
781 : /// * if connected safekeeper stopped sending us new WAL which is available on other safekeeper, pick the candidate
782 : ///
783 : /// This way we ensure to keep up with the most up-to-date safekeeper and don't try to jump from one safekeeper to another too frequently.
784 : /// Both thresholds are configured per tenant.
785 36 : fn next_connection_candidate(&mut self) -> Option<NewWalConnectionCandidate> {
786 36 : self.cleanup_old_candidates();
787 36 :
788 36 : match &self.wal_connection {
789 20 : Some(existing_wal_connection) => {
790 20 : let connected_sk_node = existing_wal_connection.sk_id;
791 :
792 20 : let (new_sk_id, new_safekeeper_broker_data, new_wal_source_connconf) =
793 20 : self.select_connection_candidate(Some(connected_sk_node))?;
794 20 : let new_availability_zone = new_safekeeper_broker_data.availability_zone.clone();
795 20 :
796 20 : let now = Utc::now().naive_utc();
797 20 : if let Ok(latest_interaciton) =
798 20 : (now - existing_wal_connection.status.latest_connection_update).to_std()
799 : {
800 : // Drop connection if we haven't received keepalive message for a while.
801 20 : if latest_interaciton > self.conf.wal_connect_timeout {
802 4 : return Some(NewWalConnectionCandidate {
803 4 : safekeeper_id: new_sk_id,
804 4 : wal_source_connconf: new_wal_source_connconf,
805 4 : availability_zone: new_availability_zone,
806 4 : reason: ReconnectReason::NoKeepAlives {
807 4 : last_keep_alive: Some(
808 4 : existing_wal_connection.status.latest_connection_update,
809 4 : ),
810 4 : check_time: now,
811 4 : threshold: self.conf.wal_connect_timeout,
812 4 : },
813 4 : });
814 16 : }
815 0 : }
816 :
817 16 : if !existing_wal_connection.status.is_connected {
818 : // We haven't connected yet and we shouldn't switch until connection timeout (condition above).
819 0 : return None;
820 16 : }
821 :
822 16 : if let Some(current_commit_lsn) = existing_wal_connection.status.commit_lsn {
823 16 : let new_commit_lsn = Lsn(new_safekeeper_broker_data.commit_lsn);
824 16 : // Check if the new candidate has much more WAL than the current one.
825 16 : match new_commit_lsn.0.checked_sub(current_commit_lsn.0) {
826 16 : Some(new_sk_lsn_advantage) => {
827 16 : if new_sk_lsn_advantage >= self.conf.max_lsn_wal_lag.get() {
828 4 : return Some(NewWalConnectionCandidate {
829 4 : safekeeper_id: new_sk_id,
830 4 : wal_source_connconf: new_wal_source_connconf,
831 4 : availability_zone: new_availability_zone,
832 4 : reason: ReconnectReason::LaggingWal {
833 4 : current_commit_lsn,
834 4 : new_commit_lsn,
835 4 : threshold: self.conf.max_lsn_wal_lag,
836 4 : },
837 4 : });
838 12 : }
839 12 : // If we have a candidate with the same commit_lsn as the current one, which is in the same AZ as pageserver,
840 12 : // and the current one is not, switch to the new one.
841 12 : if self.conf.availability_zone.is_some()
842 4 : && existing_wal_connection.availability_zone
843 4 : != self.conf.availability_zone
844 4 : && self.conf.availability_zone == new_availability_zone
845 : {
846 4 : return Some(NewWalConnectionCandidate {
847 4 : safekeeper_id: new_sk_id,
848 4 : availability_zone: new_availability_zone,
849 4 : wal_source_connconf: new_wal_source_connconf,
850 4 : reason: ReconnectReason::SwitchAvailabilityZone,
851 4 : });
852 8 : }
853 : }
854 0 : None => debug!(
855 0 : "Best SK candidate has its commit_lsn behind connected SK's commit_lsn"
856 : ),
857 : }
858 0 : }
859 :
860 8 : let current_lsn = match existing_wal_connection.status.streaming_lsn {
861 8 : Some(lsn) => lsn,
862 0 : None => self.timeline.get_last_record_lsn(),
863 : };
864 8 : let current_commit_lsn = existing_wal_connection
865 8 : .status
866 8 : .commit_lsn
867 8 : .unwrap_or(current_lsn);
868 8 : let candidate_commit_lsn = Lsn(new_safekeeper_broker_data.commit_lsn);
869 8 :
870 8 : // Keep discovered_new_wal only if connected safekeeper has not caught up yet.
871 8 : let mut discovered_new_wal = existing_wal_connection
872 8 : .discovered_new_wal
873 8 : .filter(|new_wal| new_wal.lsn > current_commit_lsn);
874 8 :
875 8 : if discovered_new_wal.is_none() {
876 : // Check if the new candidate has more WAL than the current one.
877 : // If the new candidate has more WAL than the current one, we consider switching to the new candidate.
878 4 : discovered_new_wal = if candidate_commit_lsn > current_commit_lsn {
879 4 : trace!(
880 0 : "New candidate has commit_lsn {}, higher than current_commit_lsn {}",
881 : candidate_commit_lsn,
882 : current_commit_lsn
883 : );
884 4 : Some(NewCommittedWAL {
885 4 : lsn: candidate_commit_lsn,
886 4 : discovered_at: Utc::now().naive_utc(),
887 4 : })
888 : } else {
889 0 : None
890 : };
891 4 : }
892 :
893 8 : let waiting_for_new_lsn_since = if current_lsn < current_commit_lsn {
894 : // Connected safekeeper has more WAL, but we haven't received updates for some time.
895 0 : trace!(
896 0 : "Connected safekeeper has more WAL, but we haven't received updates for {:?}. current_lsn: {}, current_commit_lsn: {}",
897 0 : (now - existing_wal_connection.status.latest_wal_update).to_std(),
898 : current_lsn,
899 : current_commit_lsn
900 : );
901 0 : Some(existing_wal_connection.status.latest_wal_update)
902 : } else {
903 8 : discovered_new_wal.as_ref().map(|new_wal| {
904 8 : // We know that new WAL is available on other safekeeper, but connected safekeeper don't have it.
905 8 : new_wal
906 8 : .discovered_at
907 8 : .max(existing_wal_connection.status.latest_wal_update)
908 8 : })
909 : };
910 :
911 : // If we haven't received any WAL updates for a while and candidate has more WAL, switch to it.
912 8 : if let Some(waiting_for_new_lsn_since) = waiting_for_new_lsn_since {
913 8 : if let Ok(waiting_for_new_wal) = (now - waiting_for_new_lsn_since).to_std() {
914 4 : if candidate_commit_lsn > current_commit_lsn
915 4 : && waiting_for_new_wal > self.conf.lagging_wal_timeout
916 : {
917 4 : return Some(NewWalConnectionCandidate {
918 4 : safekeeper_id: new_sk_id,
919 4 : wal_source_connconf: new_wal_source_connconf,
920 4 : availability_zone: new_availability_zone,
921 4 : reason: ReconnectReason::NoWalTimeout {
922 4 : current_lsn,
923 4 : current_commit_lsn,
924 4 : candidate_commit_lsn,
925 4 : last_wal_interaction: Some(
926 4 : existing_wal_connection.status.latest_wal_update,
927 4 : ),
928 4 : check_time: now,
929 4 : threshold: self.conf.lagging_wal_timeout,
930 4 : },
931 4 : });
932 0 : }
933 4 : }
934 0 : }
935 :
936 4 : self.wal_connection.as_mut().unwrap().discovered_new_wal = discovered_new_wal;
937 : }
938 : None => {
939 12 : let (new_sk_id, new_safekeeper_broker_data, new_wal_source_connconf) =
940 16 : self.select_connection_candidate(None)?;
941 12 : return Some(NewWalConnectionCandidate {
942 12 : safekeeper_id: new_sk_id,
943 12 : availability_zone: new_safekeeper_broker_data.availability_zone.clone(),
944 12 : wal_source_connconf: new_wal_source_connconf,
945 12 : reason: ReconnectReason::NoExistingConnection,
946 12 : });
947 : }
948 : }
949 :
950 4 : None
951 36 : }
952 :
953 : /// Selects the best possible candidate, based on the data collected from the broker updates about the safekeepers.
954 : /// Optionally, omits the given node, to support gracefully switching from a healthy safekeeper to another.
955 : ///
956 : /// The candidate that is chosen:
957 : /// * has no pending retry cooldown
958 : /// * has greatest commit_lsn among the ones that are left
959 36 : fn select_connection_candidate(
960 36 : &self,
961 36 : node_to_omit: Option<NodeId>,
962 36 : ) -> Option<(NodeId, &SafekeeperDiscoveryResponse, PgConnectionConfig)> {
963 36 : self.applicable_connection_candidates()
964 52 : .filter(|&(sk_id, _, _)| Some(sk_id) != node_to_omit)
965 40 : .max_by_key(|(_, info, _)| info.commit_lsn)
966 36 : }
967 :
968 : /// Returns a list of safekeepers that have valid info and ready for connection.
969 : /// Some safekeepers are filtered by the retry cooldown.
970 36 : fn applicable_connection_candidates(
971 36 : &self,
972 36 : ) -> impl Iterator<Item = (NodeId, &SafekeeperDiscoveryResponse, PgConnectionConfig)> {
973 36 : let now = Utc::now().naive_utc();
974 36 :
975 36 : self.wal_stream_candidates
976 36 : .iter()
977 72 : .filter(|(_, info)| Lsn(info.timeline.commit_lsn) != Lsn::INVALID)
978 64 : .filter(move |(sk_id, _)| {
979 64 : let next_retry_at = self
980 64 : .wal_connection_retries
981 64 : .get(sk_id)
982 64 : .and_then(|retry_info| {
983 4 : retry_info.next_retry_at
984 64 : });
985 64 :
986 64 : next_retry_at.is_none() || next_retry_at.unwrap() <= now
987 64 : }).filter_map(|(sk_id, broker_info)| {
988 60 : let info = &broker_info.timeline;
989 60 : if info.safekeeper_connstr.is_empty() {
990 8 : return None; // no connection string, ignore sk
991 52 : }
992 :
993 52 : let (shard_number, shard_count, shard_stripe_size) = match self.conf.protocol {
994 : PostgresClientProtocol::Vanilla => {
995 52 : (None, None, None)
996 : },
997 : PostgresClientProtocol::Interpreted { .. } => {
998 0 : let shard_identity = self.timeline.get_shard_identity();
999 0 : (
1000 0 : Some(shard_identity.number.0),
1001 0 : Some(shard_identity.count.0),
1002 0 : Some(shard_identity.stripe_size.0),
1003 0 : )
1004 : }
1005 : };
1006 :
1007 52 : let connection_conf_args = ConnectionConfigArgs {
1008 52 : protocol: self.conf.protocol,
1009 52 : ttid: self.id,
1010 52 : shard_number,
1011 52 : shard_count,
1012 52 : shard_stripe_size,
1013 52 : listen_pg_addr_str: info.safekeeper_connstr.as_ref(),
1014 52 : auth_token: self.conf.auth_token.as_ref().map(|t| t.as_str()),
1015 52 : availability_zone: self.conf.availability_zone.as_deref()
1016 52 : };
1017 52 :
1018 52 : match wal_stream_connection_config(connection_conf_args) {
1019 52 : Ok(connstr) => Some((*sk_id, info, connstr)),
1020 0 : Err(e) => {
1021 0 : error!("Failed to create wal receiver connection string from broker data of safekeeper node {}: {e:#}", sk_id);
1022 0 : None
1023 : }
1024 : }
1025 60 : })
1026 36 : }
1027 :
1028 : /// Remove candidates which haven't sent broker updates for a while.
1029 36 : fn cleanup_old_candidates(&mut self) {
1030 36 : let mut node_ids_to_remove = Vec::with_capacity(self.wal_stream_candidates.len());
1031 36 : let lagging_wal_timeout = self.conf.lagging_wal_timeout;
1032 36 :
1033 76 : self.wal_stream_candidates.retain(|node_id, broker_info| {
1034 76 : if let Ok(time_since_latest_broker_update) =
1035 76 : (Utc::now().naive_utc() - broker_info.latest_update).to_std()
1036 : {
1037 76 : let should_retain = time_since_latest_broker_update < lagging_wal_timeout;
1038 76 : if !should_retain {
1039 4 : node_ids_to_remove.push(*node_id);
1040 72 : }
1041 76 : should_retain
1042 : } else {
1043 0 : true
1044 : }
1045 76 : });
1046 36 :
1047 36 : if !node_ids_to_remove.is_empty() {
1048 8 : for node_id in node_ids_to_remove {
1049 4 : info!("Safekeeper node {node_id} did not send events for over {lagging_wal_timeout:?}, not retrying the connections");
1050 4 : self.wal_connection_retries.remove(&node_id);
1051 4 : WALRECEIVER_CANDIDATES_REMOVED.inc();
1052 : }
1053 32 : }
1054 36 : }
1055 :
1056 : /// # Cancel-Safety
1057 : ///
1058 : /// Not cancellation-safe.
1059 0 : pub(super) async fn shutdown(mut self) {
1060 0 : if let Some(wal_connection) = self.wal_connection.take() {
1061 0 : wal_connection.connection_task.shutdown().await;
1062 0 : }
1063 0 : }
1064 :
1065 0 : fn manager_status(&self) -> ConnectionManagerStatus {
1066 0 : ConnectionManagerStatus {
1067 0 : existing_connection: self.wal_connection.as_ref().map(|conn| conn.status),
1068 0 : wal_stream_candidates: self.wal_stream_candidates.clone(),
1069 0 : }
1070 0 : }
1071 : }
1072 :
1073 : #[derive(Debug)]
1074 : struct NewWalConnectionCandidate {
1075 : safekeeper_id: NodeId,
1076 : wal_source_connconf: PgConnectionConfig,
1077 : availability_zone: Option<String>,
1078 : reason: ReconnectReason,
1079 : }
1080 :
1081 : /// Stores the reason why WAL connection was switched, for furter debugging purposes.
1082 : #[derive(Debug, PartialEq, Eq)]
1083 : enum ReconnectReason {
1084 : NoExistingConnection,
1085 : LaggingWal {
1086 : current_commit_lsn: Lsn,
1087 : new_commit_lsn: Lsn,
1088 : threshold: NonZeroU64,
1089 : },
1090 : SwitchAvailabilityZone,
1091 : NoWalTimeout {
1092 : current_lsn: Lsn,
1093 : current_commit_lsn: Lsn,
1094 : candidate_commit_lsn: Lsn,
1095 : last_wal_interaction: Option<NaiveDateTime>,
1096 : check_time: NaiveDateTime,
1097 : threshold: Duration,
1098 : },
1099 : NoKeepAlives {
1100 : last_keep_alive: Option<NaiveDateTime>,
1101 : check_time: NaiveDateTime,
1102 : threshold: Duration,
1103 : },
1104 : }
1105 :
1106 : impl ReconnectReason {
1107 0 : fn name(&self) -> &str {
1108 0 : match self {
1109 0 : ReconnectReason::NoExistingConnection => "NoExistingConnection",
1110 0 : ReconnectReason::LaggingWal { .. } => "LaggingWal",
1111 0 : ReconnectReason::SwitchAvailabilityZone => "SwitchAvailabilityZone",
1112 0 : ReconnectReason::NoWalTimeout { .. } => "NoWalTimeout",
1113 0 : ReconnectReason::NoKeepAlives { .. } => "NoKeepAlives",
1114 : }
1115 0 : }
1116 : }
1117 :
1118 : #[cfg(test)]
1119 : mod tests {
1120 : use super::*;
1121 : use crate::tenant::harness::{TenantHarness, TIMELINE_ID};
1122 : use pageserver_api::config::defaults::DEFAULT_WAL_RECEIVER_PROTOCOL;
1123 : use url::Host;
1124 :
1125 76 : fn dummy_broker_sk_timeline(
1126 76 : commit_lsn: u64,
1127 76 : safekeeper_connstr: &str,
1128 76 : latest_update: NaiveDateTime,
1129 76 : ) -> BrokerSkTimeline {
1130 76 : BrokerSkTimeline {
1131 76 : timeline: SafekeeperDiscoveryResponse {
1132 76 : safekeeper_id: 0,
1133 76 : tenant_timeline_id: None,
1134 76 : commit_lsn,
1135 76 : safekeeper_connstr: safekeeper_connstr.to_owned(),
1136 76 : availability_zone: None,
1137 76 : standby_horizon: 0,
1138 76 : },
1139 76 : latest_update,
1140 76 : }
1141 76 : }
1142 :
1143 : #[tokio::test]
1144 4 : async fn no_connection_no_candidate() -> anyhow::Result<()> {
1145 4 : let harness = TenantHarness::create("no_connection_no_candidate").await?;
1146 4 : let mut state = dummy_state(&harness).await;
1147 4 : let now = Utc::now().naive_utc();
1148 4 :
1149 4 : let lagging_wal_timeout = chrono::Duration::from_std(state.conf.lagging_wal_timeout)?;
1150 4 : let delay_over_threshold = now - lagging_wal_timeout - lagging_wal_timeout;
1151 4 :
1152 4 : state.wal_connection = None;
1153 4 : state.wal_stream_candidates = HashMap::from([
1154 4 : (NodeId(0), dummy_broker_sk_timeline(1, "", now)),
1155 4 : (NodeId(1), dummy_broker_sk_timeline(0, "no_commit_lsn", now)),
1156 4 : (NodeId(2), dummy_broker_sk_timeline(0, "no_commit_lsn", now)),
1157 4 : (
1158 4 : NodeId(3),
1159 4 : dummy_broker_sk_timeline(
1160 4 : 1 + state.conf.max_lsn_wal_lag.get(),
1161 4 : "delay_over_threshold",
1162 4 : delay_over_threshold,
1163 4 : ),
1164 4 : ),
1165 4 : ]);
1166 4 :
1167 4 : let no_candidate = state.next_connection_candidate();
1168 4 : assert!(
1169 4 : no_candidate.is_none(),
1170 4 : "Expected no candidate selected out of non full data options, but got {no_candidate:?}"
1171 4 : );
1172 4 :
1173 4 : Ok(())
1174 4 : }
1175 :
1176 : #[tokio::test]
1177 4 : async fn connection_no_candidate() -> anyhow::Result<()> {
1178 4 : let harness = TenantHarness::create("connection_no_candidate").await?;
1179 4 : let mut state = dummy_state(&harness).await;
1180 4 : let now = Utc::now().naive_utc();
1181 4 :
1182 4 : let connected_sk_id = NodeId(0);
1183 4 : let current_lsn = 100_000;
1184 4 :
1185 4 : let connection_status = WalConnectionStatus {
1186 4 : is_connected: true,
1187 4 : has_processed_wal: true,
1188 4 : latest_connection_update: now,
1189 4 : latest_wal_update: now,
1190 4 : commit_lsn: Some(Lsn(current_lsn)),
1191 4 : streaming_lsn: Some(Lsn(current_lsn)),
1192 4 : node: NodeId(1),
1193 4 : };
1194 4 :
1195 4 : state.conf.max_lsn_wal_lag = NonZeroU64::new(100).unwrap();
1196 4 : state.wal_connection = Some(WalConnection {
1197 4 : started_at: now,
1198 4 : sk_id: connected_sk_id,
1199 4 : availability_zone: None,
1200 4 : status: connection_status,
1201 4 : connection_task: state.spawn(move |sender, _| async move {
1202 4 : sender
1203 4 : .send(TaskStateUpdate::Progress(connection_status))
1204 4 : .ok();
1205 4 : Ok(())
1206 4 : }),
1207 4 : discovered_new_wal: None,
1208 4 : });
1209 4 : state.wal_stream_candidates = HashMap::from([
1210 4 : (
1211 4 : connected_sk_id,
1212 4 : dummy_broker_sk_timeline(
1213 4 : current_lsn + state.conf.max_lsn_wal_lag.get() * 2,
1214 4 : DUMMY_SAFEKEEPER_HOST,
1215 4 : now,
1216 4 : ),
1217 4 : ),
1218 4 : (
1219 4 : NodeId(1),
1220 4 : dummy_broker_sk_timeline(current_lsn, "not_advanced_lsn", now),
1221 4 : ),
1222 4 : (
1223 4 : NodeId(2),
1224 4 : dummy_broker_sk_timeline(
1225 4 : current_lsn + state.conf.max_lsn_wal_lag.get() / 2,
1226 4 : "not_enough_advanced_lsn",
1227 4 : now,
1228 4 : ),
1229 4 : ),
1230 4 : ]);
1231 4 :
1232 4 : let no_candidate = state.next_connection_candidate();
1233 4 : assert!(
1234 4 : no_candidate.is_none(),
1235 4 : "Expected no candidate selected out of valid options since candidate Lsn data is ignored and others' was not advanced enough, but got {no_candidate:?}"
1236 4 : );
1237 4 :
1238 4 : Ok(())
1239 4 : }
1240 :
1241 : #[tokio::test]
1242 4 : async fn no_connection_candidate() -> anyhow::Result<()> {
1243 4 : let harness = TenantHarness::create("no_connection_candidate").await?;
1244 4 : let mut state = dummy_state(&harness).await;
1245 4 : let now = Utc::now().naive_utc();
1246 4 :
1247 4 : state.wal_connection = None;
1248 4 : state.wal_stream_candidates = HashMap::from([(
1249 4 : NodeId(0),
1250 4 : dummy_broker_sk_timeline(
1251 4 : 1 + state.conf.max_lsn_wal_lag.get(),
1252 4 : DUMMY_SAFEKEEPER_HOST,
1253 4 : now,
1254 4 : ),
1255 4 : )]);
1256 4 :
1257 4 : let only_candidate = state
1258 4 : .next_connection_candidate()
1259 4 : .expect("Expected one candidate selected out of the only data option, but got none");
1260 4 : assert_eq!(only_candidate.safekeeper_id, NodeId(0));
1261 4 : assert_eq!(
1262 4 : only_candidate.reason,
1263 4 : ReconnectReason::NoExistingConnection,
1264 4 : "Should select new safekeeper due to missing connection, even if there's also a lag in the wal over the threshold"
1265 4 : );
1266 4 : assert_eq!(
1267 4 : only_candidate.wal_source_connconf.host(),
1268 4 : &Host::Domain(DUMMY_SAFEKEEPER_HOST.to_owned())
1269 4 : );
1270 4 :
1271 4 : let selected_lsn = 100_000;
1272 4 : state.wal_stream_candidates = HashMap::from([
1273 4 : (
1274 4 : NodeId(0),
1275 4 : dummy_broker_sk_timeline(selected_lsn - 100, "smaller_commit_lsn", now),
1276 4 : ),
1277 4 : (
1278 4 : NodeId(1),
1279 4 : dummy_broker_sk_timeline(selected_lsn, DUMMY_SAFEKEEPER_HOST, now),
1280 4 : ),
1281 4 : (
1282 4 : NodeId(2),
1283 4 : dummy_broker_sk_timeline(selected_lsn + 100, "", now),
1284 4 : ),
1285 4 : ]);
1286 4 : let biggest_wal_candidate = state.next_connection_candidate().expect(
1287 4 : "Expected one candidate selected out of multiple valid data options, but got none",
1288 4 : );
1289 4 :
1290 4 : assert_eq!(biggest_wal_candidate.safekeeper_id, NodeId(1));
1291 4 : assert_eq!(
1292 4 : biggest_wal_candidate.reason,
1293 4 : ReconnectReason::NoExistingConnection,
1294 4 : "Should select new safekeeper due to missing connection, even if there's also a lag in the wal over the threshold"
1295 4 : );
1296 4 : assert_eq!(
1297 4 : biggest_wal_candidate.wal_source_connconf.host(),
1298 4 : &Host::Domain(DUMMY_SAFEKEEPER_HOST.to_owned())
1299 4 : );
1300 4 :
1301 4 : Ok(())
1302 4 : }
1303 :
1304 : #[tokio::test]
1305 4 : async fn candidate_with_many_connection_failures() -> anyhow::Result<()> {
1306 4 : let harness = TenantHarness::create("candidate_with_many_connection_failures").await?;
1307 4 : let mut state = dummy_state(&harness).await;
1308 4 : let now = Utc::now().naive_utc();
1309 4 :
1310 4 : let current_lsn = Lsn(100_000).align();
1311 4 : let bigger_lsn = Lsn(current_lsn.0 + 100).align();
1312 4 :
1313 4 : state.wal_connection = None;
1314 4 : state.wal_stream_candidates = HashMap::from([
1315 4 : (
1316 4 : NodeId(0),
1317 4 : dummy_broker_sk_timeline(bigger_lsn.0, DUMMY_SAFEKEEPER_HOST, now),
1318 4 : ),
1319 4 : (
1320 4 : NodeId(1),
1321 4 : dummy_broker_sk_timeline(current_lsn.0, DUMMY_SAFEKEEPER_HOST, now),
1322 4 : ),
1323 4 : ]);
1324 4 : state.wal_connection_retries = HashMap::from([(
1325 4 : NodeId(0),
1326 4 : RetryInfo {
1327 4 : next_retry_at: now.checked_add_signed(chrono::Duration::hours(1)),
1328 4 : retry_duration_seconds: WALCONNECTION_RETRY_MAX_BACKOFF_SECONDS,
1329 4 : },
1330 4 : )]);
1331 4 :
1332 4 : let candidate_with_less_errors = state
1333 4 : .next_connection_candidate()
1334 4 : .expect("Expected one candidate selected, but got none");
1335 4 : assert_eq!(
1336 4 : candidate_with_less_errors.safekeeper_id,
1337 4 : NodeId(1),
1338 4 : "Should select the node with no pending retry cooldown"
1339 4 : );
1340 4 :
1341 4 : Ok(())
1342 4 : }
1343 :
1344 : #[tokio::test]
1345 4 : async fn lsn_wal_over_threshold_current_candidate() -> anyhow::Result<()> {
1346 4 : let harness = TenantHarness::create("lsn_wal_over_threshcurrent_candidate").await?;
1347 4 : let mut state = dummy_state(&harness).await;
1348 4 : let current_lsn = Lsn(100_000).align();
1349 4 : let now = Utc::now().naive_utc();
1350 4 :
1351 4 : let connected_sk_id = NodeId(0);
1352 4 : let new_lsn = Lsn(current_lsn.0 + state.conf.max_lsn_wal_lag.get() + 1);
1353 4 :
1354 4 : let connection_status = WalConnectionStatus {
1355 4 : is_connected: true,
1356 4 : has_processed_wal: true,
1357 4 : latest_connection_update: now,
1358 4 : latest_wal_update: now,
1359 4 : commit_lsn: Some(current_lsn),
1360 4 : streaming_lsn: Some(current_lsn),
1361 4 : node: connected_sk_id,
1362 4 : };
1363 4 :
1364 4 : state.wal_connection = Some(WalConnection {
1365 4 : started_at: now,
1366 4 : sk_id: connected_sk_id,
1367 4 : availability_zone: None,
1368 4 : status: connection_status,
1369 4 : connection_task: state.spawn(move |sender, _| async move {
1370 4 : sender
1371 4 : .send(TaskStateUpdate::Progress(connection_status))
1372 4 : .ok();
1373 4 : Ok(())
1374 4 : }),
1375 4 : discovered_new_wal: None,
1376 4 : });
1377 4 : state.wal_stream_candidates = HashMap::from([
1378 4 : (
1379 4 : connected_sk_id,
1380 4 : dummy_broker_sk_timeline(current_lsn.0, DUMMY_SAFEKEEPER_HOST, now),
1381 4 : ),
1382 4 : (
1383 4 : NodeId(1),
1384 4 : dummy_broker_sk_timeline(new_lsn.0, "advanced_by_lsn_safekeeper", now),
1385 4 : ),
1386 4 : ]);
1387 4 :
1388 4 : let over_threshcurrent_candidate = state.next_connection_candidate().expect(
1389 4 : "Expected one candidate selected out of multiple valid data options, but got none",
1390 4 : );
1391 4 :
1392 4 : assert_eq!(over_threshcurrent_candidate.safekeeper_id, NodeId(1));
1393 4 : assert_eq!(
1394 4 : over_threshcurrent_candidate.reason,
1395 4 : ReconnectReason::LaggingWal {
1396 4 : current_commit_lsn: current_lsn,
1397 4 : new_commit_lsn: new_lsn,
1398 4 : threshold: state.conf.max_lsn_wal_lag
1399 4 : },
1400 4 : "Should select bigger WAL safekeeper if it starts to lag enough"
1401 4 : );
1402 4 : assert_eq!(
1403 4 : over_threshcurrent_candidate.wal_source_connconf.host(),
1404 4 : &Host::Domain("advanced_by_lsn_safekeeper".to_owned())
1405 4 : );
1406 4 :
1407 4 : Ok(())
1408 4 : }
1409 :
1410 : #[tokio::test]
1411 4 : async fn timeout_connection_threshold_current_candidate() -> anyhow::Result<()> {
1412 4 : let harness =
1413 4 : TenantHarness::create("timeout_connection_threshold_current_candidate").await?;
1414 4 : let mut state = dummy_state(&harness).await;
1415 4 : let current_lsn = Lsn(100_000).align();
1416 4 : let now = Utc::now().naive_utc();
1417 4 :
1418 4 : let wal_connect_timeout = chrono::Duration::from_std(state.conf.wal_connect_timeout)?;
1419 4 : let time_over_threshold =
1420 4 : Utc::now().naive_utc() - wal_connect_timeout - wal_connect_timeout;
1421 4 :
1422 4 : let connection_status = WalConnectionStatus {
1423 4 : is_connected: true,
1424 4 : has_processed_wal: true,
1425 4 : latest_connection_update: time_over_threshold,
1426 4 : latest_wal_update: time_over_threshold,
1427 4 : commit_lsn: Some(current_lsn),
1428 4 : streaming_lsn: Some(current_lsn),
1429 4 : node: NodeId(1),
1430 4 : };
1431 4 :
1432 4 : state.wal_connection = Some(WalConnection {
1433 4 : started_at: now,
1434 4 : sk_id: NodeId(1),
1435 4 : availability_zone: None,
1436 4 : status: connection_status,
1437 4 : connection_task: state.spawn(move |sender, _| async move {
1438 4 : sender
1439 4 : .send(TaskStateUpdate::Progress(connection_status))
1440 4 : .ok();
1441 4 : Ok(())
1442 4 : }),
1443 4 : discovered_new_wal: None,
1444 4 : });
1445 4 : state.wal_stream_candidates = HashMap::from([(
1446 4 : NodeId(0),
1447 4 : dummy_broker_sk_timeline(current_lsn.0, DUMMY_SAFEKEEPER_HOST, now),
1448 4 : )]);
1449 4 :
1450 4 : let over_threshcurrent_candidate = state.next_connection_candidate().expect(
1451 4 : "Expected one candidate selected out of multiple valid data options, but got none",
1452 4 : );
1453 4 :
1454 4 : assert_eq!(over_threshcurrent_candidate.safekeeper_id, NodeId(0));
1455 4 : match over_threshcurrent_candidate.reason {
1456 4 : ReconnectReason::NoKeepAlives {
1457 4 : last_keep_alive,
1458 4 : threshold,
1459 4 : ..
1460 4 : } => {
1461 4 : assert_eq!(last_keep_alive, Some(time_over_threshold));
1462 4 : assert_eq!(threshold, state.conf.lagging_wal_timeout);
1463 4 : }
1464 4 : unexpected => panic!("Unexpected reason: {unexpected:?}"),
1465 4 : }
1466 4 : assert_eq!(
1467 4 : over_threshcurrent_candidate.wal_source_connconf.host(),
1468 4 : &Host::Domain(DUMMY_SAFEKEEPER_HOST.to_owned())
1469 4 : );
1470 4 :
1471 4 : Ok(())
1472 4 : }
1473 :
1474 : #[tokio::test]
1475 4 : async fn timeout_wal_over_threshold_current_candidate() -> anyhow::Result<()> {
1476 4 : let harness = TenantHarness::create("timeout_wal_over_threshold_current_candidate").await?;
1477 4 : let mut state = dummy_state(&harness).await;
1478 4 : let current_lsn = Lsn(100_000).align();
1479 4 : let new_lsn = Lsn(100_100).align();
1480 4 : let now = Utc::now().naive_utc();
1481 4 :
1482 4 : let lagging_wal_timeout = chrono::Duration::from_std(state.conf.lagging_wal_timeout)?;
1483 4 : let time_over_threshold =
1484 4 : Utc::now().naive_utc() - lagging_wal_timeout - lagging_wal_timeout;
1485 4 :
1486 4 : let connection_status = WalConnectionStatus {
1487 4 : is_connected: true,
1488 4 : has_processed_wal: true,
1489 4 : latest_connection_update: now,
1490 4 : latest_wal_update: time_over_threshold,
1491 4 : commit_lsn: Some(current_lsn),
1492 4 : streaming_lsn: Some(current_lsn),
1493 4 : node: NodeId(1),
1494 4 : };
1495 4 :
1496 4 : state.wal_connection = Some(WalConnection {
1497 4 : started_at: now,
1498 4 : sk_id: NodeId(1),
1499 4 : availability_zone: None,
1500 4 : status: connection_status,
1501 4 : connection_task: state.spawn(move |_, _| async move { Ok(()) }),
1502 4 : discovered_new_wal: Some(NewCommittedWAL {
1503 4 : discovered_at: time_over_threshold,
1504 4 : lsn: new_lsn,
1505 4 : }),
1506 4 : });
1507 4 : state.wal_stream_candidates = HashMap::from([(
1508 4 : NodeId(0),
1509 4 : dummy_broker_sk_timeline(new_lsn.0, DUMMY_SAFEKEEPER_HOST, now),
1510 4 : )]);
1511 4 :
1512 4 : let over_threshcurrent_candidate = state.next_connection_candidate().expect(
1513 4 : "Expected one candidate selected out of multiple valid data options, but got none",
1514 4 : );
1515 4 :
1516 4 : assert_eq!(over_threshcurrent_candidate.safekeeper_id, NodeId(0));
1517 4 : match over_threshcurrent_candidate.reason {
1518 4 : ReconnectReason::NoWalTimeout {
1519 4 : current_lsn,
1520 4 : current_commit_lsn,
1521 4 : candidate_commit_lsn,
1522 4 : last_wal_interaction,
1523 4 : threshold,
1524 4 : ..
1525 4 : } => {
1526 4 : assert_eq!(current_lsn, current_lsn);
1527 4 : assert_eq!(current_commit_lsn, current_lsn);
1528 4 : assert_eq!(candidate_commit_lsn, new_lsn);
1529 4 : assert_eq!(last_wal_interaction, Some(time_over_threshold));
1530 4 : assert_eq!(threshold, state.conf.lagging_wal_timeout);
1531 4 : }
1532 4 : unexpected => panic!("Unexpected reason: {unexpected:?}"),
1533 4 : }
1534 4 : assert_eq!(
1535 4 : over_threshcurrent_candidate.wal_source_connconf.host(),
1536 4 : &Host::Domain(DUMMY_SAFEKEEPER_HOST.to_owned())
1537 4 : );
1538 4 :
1539 4 : Ok(())
1540 4 : }
1541 :
1542 : const DUMMY_SAFEKEEPER_HOST: &str = "safekeeper_connstr";
1543 :
1544 32 : async fn dummy_state(harness: &TenantHarness) -> ConnectionManagerState {
1545 32 : let (tenant, ctx) = harness.load().await;
1546 32 : let timeline = tenant
1547 32 : .create_test_timeline(TIMELINE_ID, Lsn(0x8), crate::DEFAULT_PG_VERSION, &ctx)
1548 32 : .await
1549 32 : .expect("Failed to create an empty timeline for dummy wal connection manager");
1550 32 :
1551 32 : ConnectionManagerState {
1552 32 : id: TenantTimelineId {
1553 32 : tenant_id: harness.tenant_shard_id.tenant_id,
1554 32 : timeline_id: TIMELINE_ID,
1555 32 : },
1556 32 : timeline,
1557 32 : cancel: CancellationToken::new(),
1558 32 : conf: WalReceiverConf {
1559 32 : protocol: DEFAULT_WAL_RECEIVER_PROTOCOL,
1560 32 : wal_connect_timeout: Duration::from_secs(1),
1561 32 : lagging_wal_timeout: Duration::from_secs(1),
1562 32 : max_lsn_wal_lag: NonZeroU64::new(1024 * 1024).unwrap(),
1563 32 : auth_token: None,
1564 32 : availability_zone: None,
1565 32 : ingest_batch_size: 1,
1566 32 : },
1567 32 : wal_connection: None,
1568 32 : wal_stream_candidates: HashMap::new(),
1569 32 : wal_connection_retries: HashMap::new(),
1570 32 : }
1571 32 : }
1572 :
1573 : #[tokio::test]
1574 4 : async fn switch_to_same_availability_zone() -> anyhow::Result<()> {
1575 4 : // Pageserver and one of safekeepers will be in the same availability zone
1576 4 : // and pageserver should prefer to connect to it.
1577 4 : let test_az = Some("test_az".to_owned());
1578 4 :
1579 4 : let harness = TenantHarness::create("switch_to_same_availability_zone").await?;
1580 4 : let mut state = dummy_state(&harness).await;
1581 4 : state.conf.availability_zone.clone_from(&test_az);
1582 4 : let current_lsn = Lsn(100_000).align();
1583 4 : let now = Utc::now().naive_utc();
1584 4 :
1585 4 : let connected_sk_id = NodeId(0);
1586 4 :
1587 4 : let connection_status = WalConnectionStatus {
1588 4 : is_connected: true,
1589 4 : has_processed_wal: true,
1590 4 : latest_connection_update: now,
1591 4 : latest_wal_update: now,
1592 4 : commit_lsn: Some(current_lsn),
1593 4 : streaming_lsn: Some(current_lsn),
1594 4 : node: connected_sk_id,
1595 4 : };
1596 4 :
1597 4 : state.wal_connection = Some(WalConnection {
1598 4 : started_at: now,
1599 4 : sk_id: connected_sk_id,
1600 4 : availability_zone: None,
1601 4 : status: connection_status,
1602 4 : connection_task: state.spawn(move |sender, _| async move {
1603 4 : sender
1604 4 : .send(TaskStateUpdate::Progress(connection_status))
1605 4 : .ok();
1606 4 : Ok(())
1607 4 : }),
1608 4 : discovered_new_wal: None,
1609 4 : });
1610 4 :
1611 4 : // We have another safekeeper with the same commit_lsn, and it have the same availability zone as
1612 4 : // the current pageserver.
1613 4 : let mut same_az_sk = dummy_broker_sk_timeline(current_lsn.0, "same_az", now);
1614 4 : same_az_sk.timeline.availability_zone.clone_from(&test_az);
1615 4 :
1616 4 : state.wal_stream_candidates = HashMap::from([
1617 4 : (
1618 4 : connected_sk_id,
1619 4 : dummy_broker_sk_timeline(current_lsn.0, DUMMY_SAFEKEEPER_HOST, now),
1620 4 : ),
1621 4 : (NodeId(1), same_az_sk),
1622 4 : ]);
1623 4 :
1624 4 : // We expect that pageserver will switch to the safekeeper in the same availability zone,
1625 4 : // even if it has the same commit_lsn.
1626 4 : let next_candidate = state.next_connection_candidate().expect(
1627 4 : "Expected one candidate selected out of multiple valid data options, but got none",
1628 4 : );
1629 4 :
1630 4 : assert_eq!(next_candidate.safekeeper_id, NodeId(1));
1631 4 : assert_eq!(
1632 4 : next_candidate.reason,
1633 4 : ReconnectReason::SwitchAvailabilityZone,
1634 4 : "Should switch to the safekeeper in the same availability zone, if it has the same commit_lsn"
1635 4 : );
1636 4 : assert_eq!(
1637 4 : next_candidate.wal_source_connconf.host(),
1638 4 : &Host::Domain("same_az".to_owned())
1639 4 : );
1640 4 :
1641 4 : Ok(())
1642 4 : }
1643 : }
|