Line data Source code
1 : use std::collections::HashMap;
2 : use std::fmt::Display;
3 : use std::sync::Arc;
4 : use std::time::Duration;
5 :
6 : use anyhow::{Context, anyhow};
7 : use futures::StreamExt;
8 : use futures::future::Either;
9 : use pageserver_api::shard::ShardIdentity;
10 : use postgres_backend::{CopyStreamHandlerEnd, PostgresBackend};
11 : use postgres_ffi::get_current_timestamp;
12 : use postgres_ffi::waldecoder::{WalDecodeError, WalStreamDecoder};
13 : use pq_proto::{BeMessage, InterpretedWalRecordsBody, WalSndKeepAlive};
14 : use tokio::io::{AsyncRead, AsyncWrite};
15 : use tokio::sync::mpsc::error::SendError;
16 : use tokio::task::JoinHandle;
17 : use tokio::time::MissedTickBehavior;
18 : use tracing::{Instrument, error, info, info_span};
19 : use utils::critical;
20 : use utils::lsn::Lsn;
21 : use utils::postgres_client::{Compression, InterpretedFormat};
22 : use wal_decoder::models::{InterpretedWalRecord, InterpretedWalRecords};
23 : use wal_decoder::wire_format::ToWireFormat;
24 :
25 : use crate::metrics::WAL_READERS;
26 : use crate::send_wal::{EndWatchView, WalSenderGuard};
27 : use crate::timeline::WalResidentTimeline;
28 : use crate::wal_reader_stream::{StreamingWalReader, WalBytes};
29 :
30 : /// Identifier used to differentiate between senders of the same
31 : /// shard.
32 : ///
33 : /// In the steady state there's only one, but two pageservers may
34 : /// temporarily have the same shard attached and attempt to ingest
35 : /// WAL for it. See also [`ShardSenderId`].
36 : #[derive(Hash, Eq, PartialEq, Copy, Clone)]
37 : struct SenderId(u8);
38 :
39 : impl SenderId {
40 5 : fn first() -> Self {
41 5 : SenderId(0)
42 5 : }
43 :
44 2 : fn next(&self) -> Self {
45 2 : SenderId(self.0.checked_add(1).expect("few senders"))
46 2 : }
47 : }
48 :
49 : #[derive(Hash, Eq, PartialEq)]
50 : struct ShardSenderId {
51 : shard: ShardIdentity,
52 : sender_id: SenderId,
53 : }
54 :
55 : impl Display for ShardSenderId {
56 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
57 0 : write!(f, "{}{}", self.sender_id.0, self.shard.shard_slug())
58 0 : }
59 : }
60 :
61 : impl ShardSenderId {
62 867 : fn new(shard: ShardIdentity, sender_id: SenderId) -> Self {
63 867 : ShardSenderId { shard, sender_id }
64 867 : }
65 :
66 0 : fn shard(&self) -> ShardIdentity {
67 0 : self.shard
68 0 : }
69 : }
70 :
71 : /// Shard-aware fan-out interpreted record reader.
72 : /// Reads WAL from disk, decodes it, intepretets it, and sends
73 : /// it to any [`InterpretedWalSender`] connected to it.
74 : /// Each [`InterpretedWalSender`] corresponds to one shard
75 : /// and gets interpreted records concerning that shard only.
76 : pub(crate) struct InterpretedWalReader {
77 : wal_stream: StreamingWalReader,
78 : shard_senders: HashMap<ShardIdentity, smallvec::SmallVec<[ShardSenderState; 1]>>,
79 : shard_notification_rx: Option<tokio::sync::mpsc::UnboundedReceiver<AttachShardNotification>>,
80 : state: Arc<std::sync::RwLock<InterpretedWalReaderState>>,
81 : pg_version: u32,
82 : }
83 :
84 : /// A handle for [`InterpretedWalReader`] which allows for interacting with it
85 : /// when it runs as a separate tokio task.
86 : #[derive(Debug)]
87 : pub(crate) struct InterpretedWalReaderHandle {
88 : join_handle: JoinHandle<Result<(), InterpretedWalReaderError>>,
89 : state: Arc<std::sync::RwLock<InterpretedWalReaderState>>,
90 : shard_notification_tx: tokio::sync::mpsc::UnboundedSender<AttachShardNotification>,
91 : }
92 :
93 : struct ShardSenderState {
94 : sender_id: SenderId,
95 : tx: tokio::sync::mpsc::Sender<Batch>,
96 : next_record_lsn: Lsn,
97 : }
98 :
99 : /// State of [`InterpretedWalReader`] visible outside of the task running it.
100 : #[derive(Debug)]
101 : pub(crate) enum InterpretedWalReaderState {
102 : Running {
103 : current_position: Lsn,
104 : /// Tracks the start of the PG WAL LSN from which the current batch of
105 : /// interpreted records originated.
106 : current_batch_wal_start: Option<Lsn>,
107 : },
108 : Done,
109 : }
110 :
111 : pub(crate) struct Batch {
112 : wal_end_lsn: Lsn,
113 : available_wal_end_lsn: Lsn,
114 : records: InterpretedWalRecords,
115 : }
116 :
117 : #[derive(thiserror::Error, Debug)]
118 : pub enum InterpretedWalReaderError {
119 : /// Handler initiates the end of streaming.
120 : #[error("decode error: {0}")]
121 : Decode(#[from] WalDecodeError),
122 : #[error("read or interpret error: {0}")]
123 : ReadOrInterpret(#[from] anyhow::Error),
124 : #[error("wal stream closed")]
125 : WalStreamClosed,
126 : }
127 :
128 : enum CurrentPositionUpdate {
129 : Reset { from: Lsn, to: Lsn },
130 : NotReset(Lsn),
131 : }
132 :
133 : impl CurrentPositionUpdate {
134 0 : fn current_position(&self) -> Lsn {
135 0 : match self {
136 0 : CurrentPositionUpdate::Reset { from: _, to } => *to,
137 0 : CurrentPositionUpdate::NotReset(lsn) => *lsn,
138 : }
139 0 : }
140 :
141 0 : fn previous_position(&self) -> Lsn {
142 0 : match self {
143 0 : CurrentPositionUpdate::Reset { from, to: _ } => *from,
144 0 : CurrentPositionUpdate::NotReset(lsn) => *lsn,
145 : }
146 0 : }
147 : }
148 :
149 : impl InterpretedWalReaderState {
150 4 : fn current_position(&self) -> Option<Lsn> {
151 4 : match self {
152 : InterpretedWalReaderState::Running {
153 2 : current_position, ..
154 2 : } => Some(*current_position),
155 2 : InterpretedWalReaderState::Done => None,
156 : }
157 4 : }
158 :
159 : #[cfg(test)]
160 1380 : fn current_batch_wal_start(&self) -> Option<Lsn> {
161 1380 : match self {
162 : InterpretedWalReaderState::Running {
163 1380 : current_batch_wal_start,
164 1380 : ..
165 1380 : } => *current_batch_wal_start,
166 0 : InterpretedWalReaderState::Done => None,
167 : }
168 1380 : }
169 :
170 : // Reset the current position of the WAL reader if the requested starting position
171 : // of the new shard is smaller than the current value.
172 4 : fn maybe_reset(&mut self, new_shard_start_pos: Lsn) -> CurrentPositionUpdate {
173 4 : match self {
174 : InterpretedWalReaderState::Running {
175 4 : current_position,
176 4 : current_batch_wal_start,
177 4 : } => {
178 4 : if new_shard_start_pos < *current_position {
179 3 : let from = *current_position;
180 3 : *current_position = new_shard_start_pos;
181 3 : *current_batch_wal_start = None;
182 3 : CurrentPositionUpdate::Reset {
183 3 : from,
184 3 : to: *current_position,
185 3 : }
186 : } else {
187 : // Edge case: The new shard is at the same current position as
188 : // the reader. Note that the current position is WAL record aligned,
189 : // so the reader might have done some partial reads and updated the
190 : // batch start. If that's the case, adjust the batch start to match
191 : // starting position of the new shard. It can lead to some shards
192 : // seeing overlaps, but in that case the actual record LSNs are checked
193 : // which should be fine based on the filtering logic.
194 1 : if let Some(start) = current_batch_wal_start {
195 0 : *start = std::cmp::min(*start, new_shard_start_pos);
196 1 : }
197 1 : CurrentPositionUpdate::NotReset(*current_position)
198 : }
199 : }
200 : InterpretedWalReaderState::Done => {
201 0 : panic!("maybe_reset called on finished reader")
202 : }
203 : }
204 4 : }
205 :
206 50 : fn update_current_batch_wal_start(&mut self, lsn: Lsn) {
207 50 : match self {
208 : InterpretedWalReaderState::Running {
209 50 : current_batch_wal_start,
210 50 : ..
211 50 : } => {
212 50 : if current_batch_wal_start.is_none() {
213 41 : *current_batch_wal_start = Some(lsn);
214 41 : }
215 : }
216 : InterpretedWalReaderState::Done => {
217 0 : panic!("update_current_batch_wal_start called on finished reader")
218 : }
219 : }
220 50 : }
221 :
222 40 : fn take_current_batch_wal_start(&mut self) -> Lsn {
223 40 : match self {
224 : InterpretedWalReaderState::Running {
225 40 : current_batch_wal_start,
226 40 : ..
227 40 : } => current_batch_wal_start.take().unwrap(),
228 : InterpretedWalReaderState::Done => {
229 0 : panic!("take_current_batch_wal_start called on finished reader")
230 : }
231 : }
232 40 : }
233 :
234 40 : fn update_current_position(&mut self, lsn: Lsn) {
235 40 : match self {
236 : InterpretedWalReaderState::Running {
237 40 : current_position, ..
238 40 : } => {
239 40 : *current_position = lsn;
240 40 : }
241 : InterpretedWalReaderState::Done => {
242 0 : panic!("update_current_position called on finished reader")
243 : }
244 : }
245 40 : }
246 : }
247 :
248 : pub(crate) struct AttachShardNotification {
249 : shard_id: ShardIdentity,
250 : sender: tokio::sync::mpsc::Sender<Batch>,
251 : start_pos: Lsn,
252 : }
253 :
254 : impl InterpretedWalReader {
255 : /// Spawn the reader in a separate tokio task and return a handle
256 2 : pub(crate) fn spawn(
257 2 : wal_stream: StreamingWalReader,
258 2 : start_pos: Lsn,
259 2 : tx: tokio::sync::mpsc::Sender<Batch>,
260 2 : shard: ShardIdentity,
261 2 : pg_version: u32,
262 2 : appname: &Option<String>,
263 2 : ) -> InterpretedWalReaderHandle {
264 2 : let state = Arc::new(std::sync::RwLock::new(InterpretedWalReaderState::Running {
265 2 : current_position: start_pos,
266 2 : current_batch_wal_start: None,
267 2 : }));
268 2 :
269 2 : let (shard_notification_tx, shard_notification_rx) = tokio::sync::mpsc::unbounded_channel();
270 :
271 2 : let reader = InterpretedWalReader {
272 2 : wal_stream,
273 2 : shard_senders: HashMap::from([(
274 2 : shard,
275 2 : smallvec::smallvec![ShardSenderState {
276 0 : sender_id: SenderId::first(),
277 0 : tx,
278 0 : next_record_lsn: start_pos,
279 0 : }],
280 : )]),
281 2 : shard_notification_rx: Some(shard_notification_rx),
282 2 : state: state.clone(),
283 2 : pg_version,
284 2 : };
285 2 :
286 2 : let metric = WAL_READERS
287 2 : .get_metric_with_label_values(&["task", appname.as_deref().unwrap_or("safekeeper")])
288 2 : .unwrap();
289 :
290 2 : let join_handle = tokio::task::spawn(
291 2 : async move {
292 2 : metric.inc();
293 2 : scopeguard::defer! {
294 2 : metric.dec();
295 2 : }
296 2 :
297 2 : reader
298 2 : .run_impl(start_pos)
299 2 : .await
300 0 : .inspect_err(|err| match err {
301 : // TODO: we may want to differentiate these errors further.
302 : InterpretedWalReaderError::Decode(_) => {
303 0 : critical!("failed to decode WAL record: {err:?}");
304 : }
305 0 : err => error!("failed to read WAL record: {err}"),
306 0 : })
307 0 : }
308 2 : .instrument(info_span!("interpreted wal reader")),
309 : );
310 :
311 2 : InterpretedWalReaderHandle {
312 2 : join_handle,
313 2 : state,
314 2 : shard_notification_tx,
315 2 : }
316 2 : }
317 :
318 : /// Construct the reader without spawning anything
319 : /// Callers should drive the future returned by [`Self::run`].
320 1 : pub(crate) fn new(
321 1 : wal_stream: StreamingWalReader,
322 1 : start_pos: Lsn,
323 1 : tx: tokio::sync::mpsc::Sender<Batch>,
324 1 : shard: ShardIdentity,
325 1 : pg_version: u32,
326 1 : shard_notification_rx: Option<
327 1 : tokio::sync::mpsc::UnboundedReceiver<AttachShardNotification>,
328 1 : >,
329 1 : ) -> InterpretedWalReader {
330 1 : let state = Arc::new(std::sync::RwLock::new(InterpretedWalReaderState::Running {
331 1 : current_position: start_pos,
332 1 : current_batch_wal_start: None,
333 1 : }));
334 1 :
335 1 : InterpretedWalReader {
336 1 : wal_stream,
337 1 : shard_senders: HashMap::from([(
338 1 : shard,
339 1 : smallvec::smallvec![ShardSenderState {
340 0 : sender_id: SenderId::first(),
341 0 : tx,
342 0 : next_record_lsn: start_pos,
343 0 : }],
344 : )]),
345 1 : shard_notification_rx,
346 1 : state: state.clone(),
347 1 : pg_version,
348 1 : }
349 1 : }
350 :
351 : /// Entry point for future (polling) based wal reader.
352 1 : pub(crate) async fn run(
353 1 : self,
354 1 : start_pos: Lsn,
355 1 : appname: &Option<String>,
356 1 : ) -> Result<(), CopyStreamHandlerEnd> {
357 1 : let metric = WAL_READERS
358 1 : .get_metric_with_label_values(&["future", appname.as_deref().unwrap_or("safekeeper")])
359 1 : .unwrap();
360 1 :
361 1 : metric.inc();
362 1 : scopeguard::defer! {
363 1 : metric.dec();
364 1 : }
365 1 :
366 1 : match self.run_impl(start_pos).await {
367 0 : Err(err @ InterpretedWalReaderError::Decode(_)) => {
368 0 : critical!("failed to decode WAL record: {err:?}");
369 : }
370 0 : Err(err) => error!("failed to read WAL record: {err}"),
371 0 : Ok(()) => info!("interpreted wal reader exiting"),
372 : }
373 :
374 0 : Err(CopyStreamHandlerEnd::Other(anyhow!(
375 0 : "interpreted wal reader finished"
376 0 : )))
377 0 : }
378 :
379 : /// Send interpreted WAL to one or more [`InterpretedWalSender`]s
380 : /// Stops when an error is encountered or when the [`InterpretedWalReaderHandle`]
381 : /// goes out of scope.
382 3 : async fn run_impl(mut self, start_pos: Lsn) -> Result<(), InterpretedWalReaderError> {
383 3 : let defer_state = self.state.clone();
384 3 : scopeguard::defer! {
385 3 : *defer_state.write().unwrap() = InterpretedWalReaderState::Done;
386 3 : }
387 3 :
388 3 : let mut wal_decoder = WalStreamDecoder::new(start_pos, self.pg_version);
389 :
390 : loop {
391 57 : tokio::select! {
392 : // Main branch for reading WAL and forwarding it
393 57 : wal_or_reset = self.wal_stream.next() => {
394 50 : let wal = wal_or_reset.map(|wor| wor.get_wal().expect("reset handled in select branch below"));
395 : let WalBytes {
396 50 : wal,
397 50 : wal_start_lsn,
398 50 : wal_end_lsn,
399 50 : available_wal_end_lsn,
400 50 : } = match wal {
401 50 : Some(some) => some.map_err(InterpretedWalReaderError::ReadOrInterpret)?,
402 : None => {
403 : // [`StreamingWalReader::next`] is an endless stream of WAL.
404 : // It shouldn't ever finish unless it panicked or became internally
405 : // inconsistent.
406 0 : return Result::Err(InterpretedWalReaderError::WalStreamClosed);
407 : }
408 : };
409 :
410 50 : self.state.write().unwrap().update_current_batch_wal_start(wal_start_lsn);
411 50 :
412 50 : wal_decoder.feed_bytes(&wal);
413 50 :
414 50 : // Deserialize and interpret WAL records from this batch of WAL.
415 50 : // Interpreted records for each shard are collected separately.
416 50 : let shard_ids = self.shard_senders.keys().copied().collect::<Vec<_>>();
417 50 : let mut records_by_sender: HashMap<ShardSenderId, Vec<InterpretedWalRecord>> = HashMap::new();
418 50 : let mut max_next_record_lsn = None;
419 647 : while let Some((next_record_lsn, recdata)) = wal_decoder.poll_decode()?
420 : {
421 597 : assert!(next_record_lsn.is_aligned());
422 597 : max_next_record_lsn = Some(next_record_lsn);
423 :
424 597 : let interpreted = InterpretedWalRecord::from_bytes_filtered(
425 597 : recdata,
426 597 : &shard_ids,
427 597 : next_record_lsn,
428 597 : self.pg_version,
429 597 : )
430 597 : .with_context(|| "Failed to interpret WAL")?;
431 :
432 1394 : for (shard, record) in interpreted {
433 797 : if record.is_empty() {
434 200 : continue;
435 597 : }
436 597 :
437 597 : let mut states_iter = self.shard_senders
438 597 : .get(&shard)
439 597 : .expect("keys collected above")
440 597 : .iter()
441 993 : .filter(|state| record.next_record_lsn > state.next_record_lsn)
442 597 : .peekable();
443 987 : while let Some(state) = states_iter.next() {
444 787 : let shard_sender_id = ShardSenderId::new(shard, state.sender_id);
445 787 :
446 787 : // The most commont case is one sender per shard. Peek and break to avoid the
447 787 : // clone in that situation.
448 787 : if states_iter.peek().is_none() {
449 397 : records_by_sender.entry(shard_sender_id).or_default().push(record);
450 397 : break;
451 390 : } else {
452 390 : records_by_sender.entry(shard_sender_id).or_default().push(record.clone());
453 390 : }
454 : }
455 : }
456 : }
457 :
458 50 : let max_next_record_lsn = match max_next_record_lsn {
459 40 : Some(lsn) => lsn,
460 : None => {
461 10 : continue;
462 : }
463 : };
464 :
465 : // Update the current position such that new receivers can decide
466 : // whether to attach to us or spawn a new WAL reader.
467 40 : let batch_wal_start_lsn = {
468 40 : let mut guard = self.state.write().unwrap();
469 40 : guard.update_current_position(max_next_record_lsn);
470 40 : guard.take_current_batch_wal_start()
471 40 : };
472 40 :
473 40 : // Send interpreted records downstream. Anything that has already been seen
474 40 : // by a shard is filtered out.
475 40 : let mut shard_senders_to_remove = Vec::new();
476 94 : for (shard, states) in &mut self.shard_senders {
477 134 : for state in states {
478 80 : let shard_sender_id = ShardSenderId::new(*shard, state.sender_id);
479 :
480 80 : let batch = if max_next_record_lsn > state.next_record_lsn {
481 : // This batch contains at least one record that this shard has not
482 : // seen yet.
483 66 : let records = records_by_sender.remove(&shard_sender_id).unwrap_or_default();
484 66 :
485 66 : InterpretedWalRecords {
486 66 : records,
487 66 : next_record_lsn: max_next_record_lsn,
488 66 : raw_wal_start_lsn: Some(batch_wal_start_lsn),
489 66 : }
490 14 : } else if wal_end_lsn > state.next_record_lsn {
491 : // All the records in this batch were seen by the shard
492 : // However, the batch maps to a chunk of WAL that the
493 : // shard has not yet seen. Notify it of the start LSN
494 : // of the PG WAL chunk such that it doesn't look like a gap.
495 0 : InterpretedWalRecords {
496 0 : records: Vec::default(),
497 0 : next_record_lsn: state.next_record_lsn,
498 0 : raw_wal_start_lsn: Some(batch_wal_start_lsn),
499 0 : }
500 : } else {
501 : // The shard has seen this chunk of WAL before. Skip it.
502 14 : continue;
503 : };
504 :
505 66 : let res = state.tx.send(Batch {
506 66 : wal_end_lsn,
507 66 : available_wal_end_lsn,
508 66 : records: batch,
509 66 : }).await;
510 :
511 66 : if res.is_err() {
512 0 : shard_senders_to_remove.push(shard_sender_id);
513 66 : } else {
514 66 : state.next_record_lsn = std::cmp::max(state.next_record_lsn, max_next_record_lsn);
515 66 : }
516 : }
517 : }
518 :
519 : // Clean up any shard senders that have dropped out.
520 : // This is inefficient, but such events are rare (connection to PS termination)
521 : // and the number of subscriptions on the same shards very small (only one
522 : // for the steady state).
523 40 : for to_remove in shard_senders_to_remove {
524 0 : let shard_senders = self.shard_senders.get_mut(&to_remove.shard()).expect("saw it above");
525 0 : if let Some(idx) = shard_senders.iter().position(|s| s.sender_id == to_remove.sender_id) {
526 0 : shard_senders.remove(idx);
527 0 : tracing::info!("Removed shard sender {}", to_remove);
528 0 : }
529 :
530 0 : if shard_senders.is_empty() {
531 0 : self.shard_senders.remove(&to_remove.shard());
532 0 : }
533 : }
534 : },
535 : // Listen for new shards that want to attach to this reader.
536 : // If the reader is not running as a task, then this is not supported
537 : // (see the pending branch below).
538 57 : notification = match self.shard_notification_rx.as_mut() {
539 57 : Some(rx) => Either::Left(rx.recv()),
540 0 : None => Either::Right(std::future::pending())
541 : } => {
542 4 : if let Some(n) = notification {
543 4 : let AttachShardNotification { shard_id, sender, start_pos } = n;
544 4 :
545 4 : // Update internal and external state, then reset the WAL stream
546 4 : // if required.
547 4 : let senders = self.shard_senders.entry(shard_id).or_default();
548 4 : let new_sender_id = match senders.last() {
549 2 : Some(sender) => sender.sender_id.next(),
550 2 : None => SenderId::first()
551 : };
552 :
553 4 : senders.push(ShardSenderState { sender_id: new_sender_id, tx: sender, next_record_lsn: start_pos});
554 4 :
555 4 : // If the shard is subscribing below the current position the we need
556 4 : // to update the cursor that tracks where we are at in the WAL
557 4 : // ([`Self::state`]) and reset the WAL stream itself
558 4 : // (`[Self::wal_stream`]). This must be done atomically from the POV of
559 4 : // anything outside the select statement.
560 4 : let position_reset = self.state.write().unwrap().maybe_reset(start_pos);
561 4 : match position_reset {
562 3 : CurrentPositionUpdate::Reset { from: _, to } => {
563 3 : self.wal_stream.reset(to).await;
564 3 : wal_decoder = WalStreamDecoder::new(to, self.pg_version);
565 : },
566 1 : CurrentPositionUpdate::NotReset(_) => {}
567 : };
568 :
569 4 : tracing::info!(
570 0 : "Added shard sender {} with start_pos={} previous_pos={} current_pos={}",
571 0 : ShardSenderId::new(shard_id, new_sender_id),
572 0 : start_pos,
573 0 : position_reset.previous_position(),
574 0 : position_reset.current_position(),
575 : );
576 0 : }
577 : }
578 : }
579 : }
580 0 : }
581 :
582 : #[cfg(test)]
583 1 : fn state(&self) -> Arc<std::sync::RwLock<InterpretedWalReaderState>> {
584 1 : self.state.clone()
585 1 : }
586 : }
587 :
588 : impl InterpretedWalReaderHandle {
589 : /// Fan-out the reader by attaching a new shard to it
590 3 : pub(crate) fn fanout(
591 3 : &self,
592 3 : shard_id: ShardIdentity,
593 3 : sender: tokio::sync::mpsc::Sender<Batch>,
594 3 : start_pos: Lsn,
595 3 : ) -> Result<(), SendError<AttachShardNotification>> {
596 3 : self.shard_notification_tx.send(AttachShardNotification {
597 3 : shard_id,
598 3 : sender,
599 3 : start_pos,
600 3 : })
601 3 : }
602 :
603 : /// Get the current WAL position of the reader
604 4 : pub(crate) fn current_position(&self) -> Option<Lsn> {
605 4 : self.state.read().unwrap().current_position()
606 4 : }
607 :
608 4 : pub(crate) fn abort(&self) {
609 4 : self.join_handle.abort()
610 4 : }
611 : }
612 :
613 : impl Drop for InterpretedWalReaderHandle {
614 2 : fn drop(&mut self) {
615 2 : tracing::info!("Aborting interpreted wal reader");
616 2 : self.abort()
617 2 : }
618 : }
619 :
620 : pub(crate) struct InterpretedWalSender<'a, IO> {
621 : pub(crate) format: InterpretedFormat,
622 : pub(crate) compression: Option<Compression>,
623 : pub(crate) appname: Option<String>,
624 :
625 : pub(crate) tli: WalResidentTimeline,
626 : pub(crate) start_lsn: Lsn,
627 :
628 : pub(crate) pgb: &'a mut PostgresBackend<IO>,
629 : pub(crate) end_watch_view: EndWatchView,
630 : pub(crate) wal_sender_guard: Arc<WalSenderGuard>,
631 : pub(crate) rx: tokio::sync::mpsc::Receiver<Batch>,
632 : }
633 :
634 : impl<IO: AsyncRead + AsyncWrite + Unpin> InterpretedWalSender<'_, IO> {
635 : /// Send interpreted WAL records over the network.
636 : /// Also manages keep-alives if nothing was sent for a while.
637 0 : pub(crate) async fn run(mut self) -> Result<(), CopyStreamHandlerEnd> {
638 0 : let mut keepalive_ticker = tokio::time::interval(Duration::from_secs(1));
639 0 : keepalive_ticker.set_missed_tick_behavior(MissedTickBehavior::Skip);
640 0 : keepalive_ticker.reset();
641 0 :
642 0 : let mut wal_position = self.start_lsn;
643 :
644 : loop {
645 0 : tokio::select! {
646 0 : batch = self.rx.recv() => {
647 0 : let batch = match batch {
648 0 : Some(b) => b,
649 : None => {
650 0 : return Result::Err(
651 0 : CopyStreamHandlerEnd::Other(anyhow!("Interpreted WAL reader exited early"))
652 0 : );
653 : }
654 : };
655 :
656 0 : wal_position = batch.wal_end_lsn;
657 :
658 0 : let buf = batch
659 0 : .records
660 0 : .to_wire(self.format, self.compression)
661 0 : .await
662 0 : .with_context(|| "Failed to serialize interpreted WAL")
663 0 : .map_err(CopyStreamHandlerEnd::from)?;
664 :
665 : // Reset the keep alive ticker since we are sending something
666 : // over the wire now.
667 0 : keepalive_ticker.reset();
668 0 :
669 0 : self.pgb
670 0 : .write_message(&BeMessage::InterpretedWalRecords(InterpretedWalRecordsBody {
671 0 : streaming_lsn: batch.wal_end_lsn.0,
672 0 : commit_lsn: batch.available_wal_end_lsn.0,
673 0 : data: &buf,
674 0 : })).await?;
675 : }
676 : // Send a periodic keep alive when the connection has been idle for a while.
677 : // Since we've been idle, also check if we can stop streaming.
678 0 : _ = keepalive_ticker.tick() => {
679 0 : if let Some(remote_consistent_lsn) = self.wal_sender_guard
680 0 : .walsenders()
681 0 : .get_ws_remote_consistent_lsn(self.wal_sender_guard.id())
682 : {
683 0 : if self.tli.should_walsender_stop(remote_consistent_lsn).await {
684 : // Stop streaming if the receivers are caught up and
685 : // there's no active compute. This causes the loop in
686 : // [`crate::send_interpreted_wal::InterpretedWalSender::run`]
687 : // to exit and terminate the WAL stream.
688 0 : break;
689 0 : }
690 0 : }
691 :
692 0 : self.pgb
693 0 : .write_message(&BeMessage::KeepAlive(WalSndKeepAlive {
694 0 : wal_end: self.end_watch_view.get().0,
695 0 : timestamp: get_current_timestamp(),
696 0 : request_reply: true,
697 0 : }))
698 0 : .await?;
699 : },
700 : }
701 : }
702 :
703 0 : Err(CopyStreamHandlerEnd::ServerInitiated(format!(
704 0 : "ending streaming to {:?} at {}, receiver is caughtup and there is no computes",
705 0 : self.appname, wal_position,
706 0 : )))
707 0 : }
708 : }
709 : #[cfg(test)]
710 : mod tests {
711 : use std::collections::HashMap;
712 : use std::str::FromStr;
713 : use std::time::Duration;
714 :
715 : use pageserver_api::shard::{ShardIdentity, ShardStripeSize};
716 : use postgres_ffi::MAX_SEND_SIZE;
717 : use tokio::sync::mpsc::error::TryRecvError;
718 : use utils::id::{NodeId, TenantTimelineId};
719 : use utils::lsn::Lsn;
720 : use utils::shard::{ShardCount, ShardNumber};
721 :
722 : use crate::send_interpreted_wal::{AttachShardNotification, Batch, InterpretedWalReader};
723 : use crate::test_utils::Env;
724 : use crate::wal_reader_stream::StreamingWalReader;
725 :
726 : #[tokio::test]
727 1 : async fn test_interpreted_wal_reader_fanout() {
728 1 : let _ = env_logger::builder().is_test(true).try_init();
729 1 :
730 1 : const SIZE: usize = 8 * 1024;
731 1 : const MSG_COUNT: usize = 200;
732 1 : const PG_VERSION: u32 = 17;
733 1 : const SHARD_COUNT: u8 = 2;
734 1 :
735 1 : let start_lsn = Lsn::from_str("0/149FD18").unwrap();
736 1 : let env = Env::new(true).unwrap();
737 1 : let tli = env
738 1 : .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn)
739 1 : .await
740 1 : .unwrap();
741 1 :
742 1 : let resident_tli = tli.wal_residence_guard().await.unwrap();
743 1 : let end_watch = Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT, None)
744 1 : .await
745 1 : .unwrap();
746 1 : let end_pos = end_watch.get();
747 1 :
748 1 : tracing::info!("Doing first round of reads ...");
749 1 :
750 1 : let streaming_wal_reader = StreamingWalReader::new(
751 1 : resident_tli,
752 1 : None,
753 1 : start_lsn,
754 1 : end_pos,
755 1 : end_watch,
756 1 : MAX_SEND_SIZE,
757 1 : );
758 1 :
759 1 : let shard_0 = ShardIdentity::new(
760 1 : ShardNumber(0),
761 1 : ShardCount(SHARD_COUNT),
762 1 : ShardStripeSize::default(),
763 1 : )
764 1 : .unwrap();
765 1 :
766 1 : let shard_1 = ShardIdentity::new(
767 1 : ShardNumber(1),
768 1 : ShardCount(SHARD_COUNT),
769 1 : ShardStripeSize::default(),
770 1 : )
771 1 : .unwrap();
772 1 :
773 1 : let mut shards = HashMap::new();
774 1 :
775 3 : for shard_number in 0..SHARD_COUNT {
776 2 : let shard_id = ShardIdentity::new(
777 2 : ShardNumber(shard_number),
778 2 : ShardCount(SHARD_COUNT),
779 2 : ShardStripeSize::default(),
780 2 : )
781 2 : .unwrap();
782 2 : let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(MSG_COUNT * 2);
783 2 : shards.insert(shard_id, (Some(tx), Some(rx)));
784 2 : }
785 1 :
786 1 : let shard_0_tx = shards.get_mut(&shard_0).unwrap().0.take().unwrap();
787 1 : let mut shard_0_rx = shards.get_mut(&shard_0).unwrap().1.take().unwrap();
788 1 :
789 1 : let handle = InterpretedWalReader::spawn(
790 1 : streaming_wal_reader,
791 1 : start_lsn,
792 1 : shard_0_tx,
793 1 : shard_0,
794 1 : PG_VERSION,
795 1 : &Some("pageserver".to_string()),
796 1 : );
797 1 :
798 1 : tracing::info!("Reading all WAL with only shard 0 attached ...");
799 1 :
800 1 : let mut shard_0_interpreted_records = Vec::new();
801 13 : while let Some(batch) = shard_0_rx.recv().await {
802 13 : shard_0_interpreted_records.push(batch.records);
803 13 : if batch.wal_end_lsn == batch.available_wal_end_lsn {
804 1 : break;
805 12 : }
806 1 : }
807 1 :
808 1 : let shard_1_tx = shards.get_mut(&shard_1).unwrap().0.take().unwrap();
809 1 : let mut shard_1_rx = shards.get_mut(&shard_1).unwrap().1.take().unwrap();
810 1 :
811 1 : tracing::info!("Attaching shard 1 to the reader at start of WAL");
812 1 : handle.fanout(shard_1, shard_1_tx, start_lsn).unwrap();
813 1 :
814 1 : tracing::info!("Reading all WAL with shard 0 and shard 1 attached ...");
815 1 :
816 1 : let mut shard_1_interpreted_records = Vec::new();
817 13 : while let Some(batch) = shard_1_rx.recv().await {
818 13 : shard_1_interpreted_records.push(batch.records);
819 13 : if batch.wal_end_lsn == batch.available_wal_end_lsn {
820 1 : break;
821 12 : }
822 1 : }
823 1 :
824 1 : // This test uses logical messages. Those only go to shard 0. Check that the
825 1 : // filtering worked and shard 1 did not get any.
826 1 : assert!(
827 1 : shard_1_interpreted_records
828 1 : .iter()
829 13 : .all(|recs| recs.records.is_empty())
830 1 : );
831 1 :
832 1 : // Shard 0 should not receive anything more since the reader is
833 1 : // going through wal that it has already processed.
834 1 : let res = shard_0_rx.try_recv();
835 1 : if let Ok(ref ok) = res {
836 1 : tracing::error!(
837 1 : "Shard 0 received batch: wal_end_lsn={} available_wal_end_lsn={}",
838 1 : ok.wal_end_lsn,
839 1 : ok.available_wal_end_lsn
840 1 : );
841 1 : }
842 1 : assert!(matches!(res, Err(TryRecvError::Empty)));
843 1 :
844 1 : // Check that the next records lsns received by the two shards match up.
845 1 : let shard_0_next_lsns = shard_0_interpreted_records
846 1 : .iter()
847 13 : .map(|recs| recs.next_record_lsn)
848 1 : .collect::<Vec<_>>();
849 1 : let shard_1_next_lsns = shard_1_interpreted_records
850 1 : .iter()
851 13 : .map(|recs| recs.next_record_lsn)
852 1 : .collect::<Vec<_>>();
853 1 : assert_eq!(shard_0_next_lsns, shard_1_next_lsns);
854 1 :
855 1 : handle.abort();
856 1 : let mut done = false;
857 2 : for _ in 0..5 {
858 2 : if handle.current_position().is_none() {
859 1 : done = true;
860 1 : break;
861 1 : }
862 1 : tokio::time::sleep(Duration::from_millis(1)).await;
863 1 : }
864 1 :
865 1 : assert!(done);
866 1 : }
867 :
868 : #[tokio::test]
869 1 : async fn test_interpreted_wal_reader_same_shard_fanout() {
870 1 : let _ = env_logger::builder().is_test(true).try_init();
871 1 :
872 1 : const SIZE: usize = 8 * 1024;
873 1 : const MSG_COUNT: usize = 200;
874 1 : const PG_VERSION: u32 = 17;
875 1 : const SHARD_COUNT: u8 = 2;
876 1 :
877 1 : let start_lsn = Lsn::from_str("0/149FD18").unwrap();
878 1 : let env = Env::new(true).unwrap();
879 1 : let tli = env
880 1 : .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn)
881 1 : .await
882 1 : .unwrap();
883 1 :
884 1 : let resident_tli = tli.wal_residence_guard().await.unwrap();
885 1 : let mut next_record_lsns = Vec::default();
886 1 : let end_watch =
887 1 : Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT, Some(&mut next_record_lsns))
888 1 : .await
889 1 : .unwrap();
890 1 : let end_pos = end_watch.get();
891 1 :
892 1 : let streaming_wal_reader = StreamingWalReader::new(
893 1 : resident_tli,
894 1 : None,
895 1 : start_lsn,
896 1 : end_pos,
897 1 : end_watch,
898 1 : MAX_SEND_SIZE,
899 1 : );
900 1 :
901 1 : let shard_0 = ShardIdentity::new(
902 1 : ShardNumber(0),
903 1 : ShardCount(SHARD_COUNT),
904 1 : ShardStripeSize::default(),
905 1 : )
906 1 : .unwrap();
907 1 :
908 1 : struct Sender {
909 1 : tx: Option<tokio::sync::mpsc::Sender<Batch>>,
910 1 : rx: tokio::sync::mpsc::Receiver<Batch>,
911 1 : shard: ShardIdentity,
912 1 : start_lsn: Lsn,
913 1 : received_next_record_lsns: Vec<Lsn>,
914 1 : }
915 1 :
916 1 : impl Sender {
917 3 : fn new(start_lsn: Lsn, shard: ShardIdentity) -> Self {
918 3 : let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(MSG_COUNT * 2);
919 3 : Self {
920 3 : tx: Some(tx),
921 3 : rx,
922 3 : shard,
923 3 : start_lsn,
924 3 : received_next_record_lsns: Vec::default(),
925 3 : }
926 3 : }
927 1 : }
928 1 :
929 1 : assert!(next_record_lsns.len() > 7);
930 1 : let start_lsns = vec![
931 1 : next_record_lsns[5],
932 1 : next_record_lsns[1],
933 1 : next_record_lsns[3],
934 1 : ];
935 1 : let mut senders = start_lsns
936 1 : .into_iter()
937 3 : .map(|lsn| Sender::new(lsn, shard_0))
938 1 : .collect::<Vec<_>>();
939 1 :
940 1 : let first_sender = senders.first_mut().unwrap();
941 1 : let handle = InterpretedWalReader::spawn(
942 1 : streaming_wal_reader,
943 1 : first_sender.start_lsn,
944 1 : first_sender.tx.take().unwrap(),
945 1 : first_sender.shard,
946 1 : PG_VERSION,
947 1 : &Some("pageserver".to_string()),
948 1 : );
949 1 :
950 2 : for sender in senders.iter_mut().skip(1) {
951 2 : handle
952 2 : .fanout(sender.shard, sender.tx.take().unwrap(), sender.start_lsn)
953 2 : .unwrap();
954 2 : }
955 1 :
956 3 : for sender in senders.iter_mut() {
957 1 : loop {
958 39 : let batch = sender.rx.recv().await.unwrap();
959 39 : tracing::info!(
960 1 : "Sender with start_lsn={} received batch ending at {} with {} records",
961 0 : sender.start_lsn,
962 0 : batch.wal_end_lsn,
963 0 : batch.records.records.len()
964 1 : );
965 1 :
966 627 : for rec in batch.records.records {
967 588 : sender.received_next_record_lsns.push(rec.next_record_lsn);
968 588 : }
969 1 :
970 39 : if batch.wal_end_lsn == batch.available_wal_end_lsn {
971 3 : break;
972 36 : }
973 1 : }
974 1 : }
975 1 :
976 1 : handle.abort();
977 1 : let mut done = false;
978 2 : for _ in 0..5 {
979 2 : if handle.current_position().is_none() {
980 1 : done = true;
981 1 : break;
982 1 : }
983 1 : tokio::time::sleep(Duration::from_millis(1)).await;
984 1 : }
985 1 :
986 1 : assert!(done);
987 1 :
988 4 : for sender in senders {
989 3 : tracing::info!(
990 1 : "Validating records received by sender with start_lsn={}",
991 1 : sender.start_lsn
992 1 : );
993 1 :
994 3 : assert!(sender.received_next_record_lsns.is_sorted());
995 3 : let expected = next_record_lsns
996 3 : .iter()
997 600 : .filter(|lsn| **lsn > sender.start_lsn)
998 3 : .copied()
999 3 : .collect::<Vec<_>>();
1000 3 : assert_eq!(sender.received_next_record_lsns, expected);
1001 1 : }
1002 1 : }
1003 :
1004 : #[tokio::test]
1005 1 : async fn test_batch_start_tracking_on_reset() {
1006 1 : // When the WAL stream is reset to an older LSN,
1007 1 : // the current batch start LSN should be invalidated.
1008 1 : // This test constructs such a scenario:
1009 1 : // 1. Shard 0 is reading somewhere ahead
1010 1 : // 2. Reader reads some WAL, but does not decode a full record (partial read)
1011 1 : // 3. Shard 1 attaches to the reader and resets it to an older LSN
1012 1 : // 4. Shard 1 should get the correct batch WAL start LSN
1013 1 : let _ = env_logger::builder().is_test(true).try_init();
1014 1 :
1015 1 : const SIZE: usize = 64 * 1024;
1016 1 : const MSG_COUNT: usize = 10;
1017 1 : const PG_VERSION: u32 = 17;
1018 1 : const SHARD_COUNT: u8 = 2;
1019 1 : const WAL_READER_BATCH_SIZE: usize = 8192;
1020 1 :
1021 1 : let start_lsn = Lsn::from_str("0/149FD18").unwrap();
1022 1 : let env = Env::new(true).unwrap();
1023 1 : let mut next_record_lsns = Vec::default();
1024 1 : let tli = env
1025 1 : .make_timeline(NodeId(1), TenantTimelineId::generate(), start_lsn)
1026 1 : .await
1027 1 : .unwrap();
1028 1 :
1029 1 : let resident_tli = tli.wal_residence_guard().await.unwrap();
1030 1 : let end_watch =
1031 1 : Env::write_wal(tli, start_lsn, SIZE, MSG_COUNT, Some(&mut next_record_lsns))
1032 1 : .await
1033 1 : .unwrap();
1034 1 :
1035 1 : assert!(next_record_lsns.len() > 3);
1036 1 : let shard_0_start_lsn = next_record_lsns[3];
1037 1 :
1038 1 : let end_pos = end_watch.get();
1039 1 :
1040 1 : let streaming_wal_reader = StreamingWalReader::new(
1041 1 : resident_tli,
1042 1 : None,
1043 1 : shard_0_start_lsn,
1044 1 : end_pos,
1045 1 : end_watch,
1046 1 : WAL_READER_BATCH_SIZE,
1047 1 : );
1048 1 :
1049 1 : let shard_0 = ShardIdentity::new(
1050 1 : ShardNumber(0),
1051 1 : ShardCount(SHARD_COUNT),
1052 1 : ShardStripeSize::default(),
1053 1 : )
1054 1 : .unwrap();
1055 1 :
1056 1 : let shard_1 = ShardIdentity::new(
1057 1 : ShardNumber(1),
1058 1 : ShardCount(SHARD_COUNT),
1059 1 : ShardStripeSize::default(),
1060 1 : )
1061 1 : .unwrap();
1062 1 :
1063 1 : let mut shards = HashMap::new();
1064 1 :
1065 3 : for shard_number in 0..SHARD_COUNT {
1066 2 : let shard_id = ShardIdentity::new(
1067 2 : ShardNumber(shard_number),
1068 2 : ShardCount(SHARD_COUNT),
1069 2 : ShardStripeSize::default(),
1070 2 : )
1071 2 : .unwrap();
1072 2 : let (tx, rx) = tokio::sync::mpsc::channel::<Batch>(MSG_COUNT * 2);
1073 2 : shards.insert(shard_id, (Some(tx), Some(rx)));
1074 2 : }
1075 1 :
1076 1 : let shard_0_tx = shards.get_mut(&shard_0).unwrap().0.take().unwrap();
1077 1 :
1078 1 : let (shard_notification_tx, shard_notification_rx) = tokio::sync::mpsc::unbounded_channel();
1079 1 :
1080 1 : let reader = InterpretedWalReader::new(
1081 1 : streaming_wal_reader,
1082 1 : shard_0_start_lsn,
1083 1 : shard_0_tx,
1084 1 : shard_0,
1085 1 : PG_VERSION,
1086 1 : Some(shard_notification_rx),
1087 1 : );
1088 1 :
1089 1 : let reader_state = reader.state();
1090 1 : let mut reader_fut = std::pin::pin!(reader.run(shard_0_start_lsn, &None));
1091 1 : loop {
1092 1380 : let poll = futures::poll!(reader_fut.as_mut());
1093 1380 : assert!(poll.is_pending());
1094 1 :
1095 1380 : let guard = reader_state.read().unwrap();
1096 1380 : if guard.current_batch_wal_start().is_some() {
1097 1 : break;
1098 1379 : }
1099 1 : }
1100 1 :
1101 1 : shard_notification_tx
1102 1 : .send(AttachShardNotification {
1103 1 : shard_id: shard_1,
1104 1 : sender: shards.get_mut(&shard_1).unwrap().0.take().unwrap(),
1105 1 : start_pos: start_lsn,
1106 1 : })
1107 1 : .unwrap();
1108 1 :
1109 1 : let mut shard_1_rx = shards.get_mut(&shard_1).unwrap().1.take().unwrap();
1110 1 : loop {
1111 66 : let poll = futures::poll!(reader_fut.as_mut());
1112 66 : assert!(poll.is_pending());
1113 1 :
1114 66 : let try_recv_res = shard_1_rx.try_recv();
1115 65 : match try_recv_res {
1116 1 : Ok(batch) => {
1117 1 : assert_eq!(batch.records.raw_wal_start_lsn.unwrap(), start_lsn);
1118 1 : break;
1119 1 : }
1120 65 : Err(tokio::sync::mpsc::error::TryRecvError::Empty) => {}
1121 1 : Err(tokio::sync::mpsc::error::TryRecvError::Disconnected) => {
1122 1 : unreachable!();
1123 1 : }
1124 1 : }
1125 1 : }
1126 1 : }
1127 : }
|
