Line data Source code
1 : //!
2 : //! WAL redo. This service runs PostgreSQL in a special wal_redo mode
3 : //! to apply given WAL records over an old page image and return new
4 : //! page image.
5 : //!
6 : //! We rely on Postgres to perform WAL redo for us. We launch a
7 : //! postgres process in special "wal redo" mode that's similar to
8 : //! single-user mode. We then pass the previous page image, if any,
9 : //! and all the WAL records we want to apply, to the postgres
10 : //! process. Then we get the page image back. Communication with the
11 : //! postgres process happens via stdin/stdout
12 : //!
13 : //! See pgxn/neon_walredo/walredoproc.c for the other side of
14 : //! this communication.
15 : //!
16 : //! The Postgres process is assumed to be secure against malicious WAL
17 : //! records. It achieves it by dropping privileges before replaying
18 : //! any WAL records, so that even if an attacker hijacks the Postgres
19 : //! process, he cannot escape out of it.
20 :
21 : /// Process lifecycle and abstracction for the IPC protocol.
22 : mod process;
23 :
24 : /// Code to apply [`NeonWalRecord`]s.
25 : pub(crate) mod apply_neon;
26 :
27 : use crate::config::PageServerConf;
28 : use crate::metrics::{
29 : WAL_REDO_BYTES_HISTOGRAM, WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM,
30 : WAL_REDO_RECORDS_HISTOGRAM, WAL_REDO_TIME,
31 : };
32 : use crate::repository::Key;
33 : use crate::walrecord::NeonWalRecord;
34 : use anyhow::Context;
35 : use bytes::{Bytes, BytesMut};
36 : use pageserver_api::models::{WalRedoManagerProcessStatus, WalRedoManagerStatus};
37 : use pageserver_api::shard::TenantShardId;
38 : use std::future::Future;
39 : use std::sync::Arc;
40 : use std::time::Duration;
41 : use std::time::Instant;
42 : use tracing::*;
43 : use utils::lsn::Lsn;
44 : use utils::sync::gate::GateError;
45 : use utils::sync::heavier_once_cell;
46 :
47 : /// The real implementation that uses a Postgres process to
48 : /// perform WAL replay.
49 : ///
50 : /// Only one thread can use the process at a time, that is controlled by the
51 : /// Mutex. In the future, we might want to launch a pool of processes to allow
52 : /// concurrent replay of multiple records.
53 : pub struct PostgresRedoManager {
54 : tenant_shard_id: TenantShardId,
55 : conf: &'static PageServerConf,
56 : last_redo_at: std::sync::Mutex<Option<Instant>>,
57 : /// We use [`heavier_once_cell`] for
58 : ///
59 : /// 1. coalescing the lazy spawning of walredo processes ([`ProcessOnceCell::Spawned`])
60 : /// 2. prevent new processes from being spawned on [`Self::shutdown`] (=> [`ProcessOnceCell::ManagerShutDown`]).
61 : ///
62 : /// # Spawning
63 : ///
64 : /// Redo requests use the once cell to coalesce onto one call to [`process::WalRedoProcess::launch`].
65 : ///
66 : /// Notably, requests don't use the [`heavier_once_cell::Guard`] to keep ahold of the
67 : /// their process object; we use [`Arc::clone`] for that.
68 : ///
69 : /// This is primarily because earlier implementations that didn't use [`heavier_once_cell`]
70 : /// had that behavior; it's probably unnecessary.
71 : /// The only merit of it is that if one walredo process encounters an error,
72 : /// it can take it out of rotation (= using [`heavier_once_cell::Guard::take_and_deinit`].
73 : /// and retry redo, thereby starting the new process, while other redo tasks might
74 : /// still be using the old redo process. But, those other tasks will most likely
75 : /// encounter an error as well, and errors are an unexpected condition anyway.
76 : /// So, probably we could get rid of the `Arc` in the future.
77 : ///
78 : /// # Shutdown
79 : ///
80 : /// See [`Self::launched_processes`].
81 : redo_process: heavier_once_cell::OnceCell<ProcessOnceCell>,
82 :
83 : /// Gate that is entered when launching a walredo process and held open
84 : /// until the process has been `kill()`ed and `wait()`ed upon.
85 : ///
86 : /// Manager shutdown waits for this gate to close after setting the
87 : /// [`ProcessOnceCell::ManagerShutDown`] state in [`Self::redo_process`].
88 : ///
89 : /// This type of usage is a bit unusual because gates usually keep track of
90 : /// concurrent operations, e.g., every [`Self::request_redo`] that is inflight.
91 : /// But we use it here to keep track of the _processes_ that we have launched,
92 : /// which may outlive any individual redo request because
93 : /// - we keep walredo process around until its quiesced to amortize spawn cost and
94 : /// - the Arc may be held by multiple concurrent redo requests, so, just because
95 : /// you replace the [`Self::redo_process`] cell's content doesn't mean the
96 : /// process gets killed immediately.
97 : ///
98 : /// We could simplify this by getting rid of the [`Arc`].
99 : /// See the comment on [`Self::redo_process`] for more details.
100 : launched_processes: utils::sync::gate::Gate,
101 : }
102 :
103 : /// See [`PostgresRedoManager::redo_process`].
104 : enum ProcessOnceCell {
105 : Spawned(Arc<Process>),
106 : ManagerShutDown,
107 : }
108 :
109 : struct Process {
110 : process: process::WalRedoProcess,
111 : /// This field is last in this struct so the guard gets dropped _after_ [`Self::process`].
112 : /// (Reminder: dropping [`Self::process`] synchronously sends SIGKILL and then `wait()`s for it to exit).
113 : _launched_processes_guard: utils::sync::gate::GateGuard,
114 : }
115 :
116 : impl std::ops::Deref for Process {
117 : type Target = process::WalRedoProcess;
118 :
119 48 : fn deref(&self) -> &Self::Target {
120 48 : &self.process
121 48 : }
122 : }
123 :
124 0 : #[derive(Debug, thiserror::Error)]
125 : pub enum Error {
126 : #[error("cancelled")]
127 : Cancelled,
128 : #[error(transparent)]
129 : Other(#[from] anyhow::Error),
130 : }
131 :
132 : macro_rules! bail {
133 : ($($arg:tt)*) => {
134 : return Err($crate::walredo::Error::Other(::anyhow::anyhow!($($arg)*)));
135 : }
136 : }
137 :
138 : ///
139 : /// Public interface of WAL redo manager
140 : ///
141 : impl PostgresRedoManager {
142 : ///
143 : /// Request the WAL redo manager to apply some WAL records
144 : ///
145 : /// The WAL redo is handled by a separate thread, so this just sends a request
146 : /// to the thread and waits for response.
147 : ///
148 : /// # Cancel-Safety
149 : ///
150 : /// This method is cancellation-safe.
151 18 : pub async fn request_redo(
152 18 : &self,
153 18 : key: Key,
154 18 : lsn: Lsn,
155 18 : base_img: Option<(Lsn, Bytes)>,
156 18 : records: Vec<(Lsn, NeonWalRecord)>,
157 18 : pg_version: u32,
158 18 : ) -> Result<Bytes, Error> {
159 18 : if records.is_empty() {
160 0 : bail!("invalid WAL redo request with no records");
161 18 : }
162 18 :
163 18 : let base_img_lsn = base_img.as_ref().map(|p| p.0).unwrap_or(Lsn::INVALID);
164 18 : let mut img = base_img.map(|p| p.1);
165 18 : let mut batch_neon = apply_neon::can_apply_in_neon(&records[0].1);
166 18 : let mut batch_start = 0;
167 18 : for (i, record) in records.iter().enumerate().skip(1) {
168 18 : let rec_neon = apply_neon::can_apply_in_neon(&record.1);
169 18 :
170 18 : if rec_neon != batch_neon {
171 0 : let result = if batch_neon {
172 0 : self.apply_batch_neon(key, lsn, img, &records[batch_start..i])
173 : } else {
174 0 : self.apply_batch_postgres(
175 0 : key,
176 0 : lsn,
177 0 : img,
178 0 : base_img_lsn,
179 0 : &records[batch_start..i],
180 0 : self.conf.wal_redo_timeout,
181 0 : pg_version,
182 0 : )
183 0 : .await
184 : };
185 0 : img = Some(result?);
186 :
187 0 : batch_neon = rec_neon;
188 0 : batch_start = i;
189 18 : }
190 : }
191 : // last batch
192 18 : if batch_neon {
193 0 : self.apply_batch_neon(key, lsn, img, &records[batch_start..])
194 : } else {
195 18 : self.apply_batch_postgres(
196 18 : key,
197 18 : lsn,
198 18 : img,
199 18 : base_img_lsn,
200 18 : &records[batch_start..],
201 18 : self.conf.wal_redo_timeout,
202 18 : pg_version,
203 18 : )
204 48 : .await
205 : }
206 18 : }
207 :
208 0 : pub fn status(&self) -> WalRedoManagerStatus {
209 0 : WalRedoManagerStatus {
210 0 : last_redo_at: {
211 0 : let at = *self.last_redo_at.lock().unwrap();
212 0 : at.and_then(|at| {
213 0 : let age = at.elapsed();
214 0 : // map any chrono errors silently to None here
215 0 : chrono::Utc::now().checked_sub_signed(chrono::Duration::from_std(age).ok()?)
216 0 : })
217 0 : },
218 0 : process: self.redo_process.get().and_then(|p| match &*p {
219 0 : ProcessOnceCell::Spawned(p) => Some(WalRedoManagerProcessStatus { pid: p.id() }),
220 0 : ProcessOnceCell::ManagerShutDown => None,
221 0 : }),
222 0 : }
223 0 : }
224 : }
225 :
226 : impl PostgresRedoManager {
227 : ///
228 : /// Create a new PostgresRedoManager.
229 : ///
230 18 : pub fn new(
231 18 : conf: &'static PageServerConf,
232 18 : tenant_shard_id: TenantShardId,
233 18 : ) -> PostgresRedoManager {
234 18 : // The actual process is launched lazily, on first request.
235 18 : PostgresRedoManager {
236 18 : tenant_shard_id,
237 18 : conf,
238 18 : last_redo_at: std::sync::Mutex::default(),
239 18 : redo_process: heavier_once_cell::OnceCell::default(),
240 18 : launched_processes: utils::sync::gate::Gate::default(),
241 18 : }
242 18 : }
243 :
244 : /// Shut down the WAL redo manager.
245 : ///
246 : /// Returns `true` if this call was the one that initiated shutdown.
247 : /// `true` may be observed by no caller if the first caller stops polling.
248 : ///
249 : /// After this future completes
250 : /// - no redo process is running
251 : /// - no new redo process will be spawned
252 : /// - redo requests that need walredo process will fail with [`Error::Cancelled`]
253 : /// - [`apply_neon`]-only redo requests may still work, but this may change in the future
254 : ///
255 : /// # Cancel-Safety
256 : ///
257 : /// This method is cancellation-safe.
258 0 : pub async fn shutdown(&self) -> bool {
259 : // prevent new processes from being spawned
260 0 : let maybe_permit = match self.redo_process.get_or_init_detached().await {
261 0 : Ok(guard) => {
262 0 : if matches!(&*guard, ProcessOnceCell::ManagerShutDown) {
263 0 : None
264 : } else {
265 0 : let (proc, permit) = guard.take_and_deinit();
266 0 : drop(proc); // this just drops the Arc, its refcount may not be zero yet
267 0 : Some(permit)
268 : }
269 : }
270 0 : Err(permit) => Some(permit),
271 : };
272 0 : let it_was_us = if let Some(permit) = maybe_permit {
273 0 : self.redo_process
274 0 : .set(ProcessOnceCell::ManagerShutDown, permit);
275 0 : true
276 : } else {
277 0 : false
278 : };
279 : // wait for ongoing requests to drain and the refcounts of all Arc<WalRedoProcess> that
280 : // we ever launched to drop to zero, which when it happens synchronously kill()s & wait()s
281 : // for the underlying process.
282 0 : self.launched_processes.close().await;
283 0 : it_was_us
284 0 : }
285 :
286 : /// This type doesn't have its own background task to check for idleness: we
287 : /// rely on our owner calling this function periodically in its own housekeeping
288 : /// loops.
289 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
290 0 : if let Ok(g) = self.last_redo_at.try_lock() {
291 0 : if let Some(last_redo_at) = *g {
292 0 : if last_redo_at.elapsed() >= idle_timeout {
293 0 : drop(g);
294 0 : drop(self.redo_process.get().map(|guard| guard.take_and_deinit()));
295 0 : }
296 0 : }
297 0 : }
298 0 : }
299 :
300 24 : async fn do_with_walredo_process<
301 24 : F: FnOnce(Arc<Process>) -> Fut,
302 24 : Fut: Future<Output = Result<O, Error>>,
303 24 : O,
304 24 : >(
305 24 : &self,
306 24 : pg_version: u32,
307 24 : closure: F,
308 24 : ) -> Result<O, Error> {
309 24 : let proc: Arc<Process> = match self.redo_process.get_or_init_detached().await {
310 0 : Ok(guard) => match &*guard {
311 0 : ProcessOnceCell::Spawned(proc) => Arc::clone(proc),
312 : ProcessOnceCell::ManagerShutDown => {
313 0 : return Err(Error::Cancelled);
314 : }
315 : },
316 24 : Err(permit) => {
317 24 : let start = Instant::now();
318 : // acquire guard before spawning process, so that we don't spawn new processes
319 : // if the gate is already closed.
320 24 : let _launched_processes_guard = match self.launched_processes.enter() {
321 24 : Ok(guard) => guard,
322 0 : Err(GateError::GateClosed) => unreachable!(
323 0 : "shutdown sets the once cell to `ManagerShutDown` state before closing the gate"
324 0 : ),
325 : };
326 24 : let proc = Arc::new(Process {
327 24 : process: process::WalRedoProcess::launch(
328 24 : self.conf,
329 24 : self.tenant_shard_id,
330 24 : pg_version,
331 24 : )
332 24 : .context("launch walredo process")?,
333 24 : _launched_processes_guard,
334 24 : });
335 24 : let duration = start.elapsed();
336 24 : WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM.observe(duration.as_secs_f64());
337 24 : info!(
338 0 : elapsed_ms = duration.as_millis(),
339 0 : pid = proc.id(),
340 0 : "launched walredo process"
341 : );
342 24 : self.redo_process
343 24 : .set(ProcessOnceCell::Spawned(Arc::clone(&proc)), permit);
344 24 : proc
345 : }
346 : };
347 :
348 : // async closures are unstable, would support &Process
349 48 : let result = closure(proc.clone()).await;
350 :
351 24 : if result.is_err() {
352 : // Avoid concurrent callers hitting the same issue by taking `proc` out of the rotation.
353 : // Note that there may be other tasks concurrent with us that also hold `proc`.
354 : // We have to deal with that here.
355 : // Also read the doc comment on field `self.redo_process`.
356 : //
357 : // NB: there may still be other concurrent threads using `proc`.
358 : // The last one will send SIGKILL when the underlying Arc reaches refcount 0.
359 : //
360 : // NB: the drop impl blocks the dropping thread with a wait() system call for
361 : // the child process. In some ways the blocking is actually good: if we
362 : // deferred the waiting into the background / to tokio if we used `tokio::process`,
363 : // it could happen that if walredo always fails immediately, we spawn processes faster
364 : // than we can SIGKILL & `wait` for them to exit. By doing it the way we do here,
365 : // we limit this risk of run-away to at most $num_runtimes * $num_executor_threads.
366 : // This probably needs revisiting at some later point.
367 12 : match self.redo_process.get() {
368 0 : None => (),
369 12 : Some(guard) => {
370 12 : match &*guard {
371 0 : ProcessOnceCell::ManagerShutDown => {}
372 12 : ProcessOnceCell::Spawned(guard_proc) => {
373 12 : if Arc::ptr_eq(&proc, guard_proc) {
374 12 : // We're the first to observe an error from `proc`, it's our job to take it out of rotation.
375 12 : guard.take_and_deinit();
376 12 : } else {
377 0 : // Another task already spawned another redo process (further up in this method)
378 0 : // and put it into `redo_process`. Do nothing, our view of the world is behind.
379 0 : }
380 : }
381 : }
382 : }
383 : }
384 : // The last task that does this `drop()` of `proc` will do a blocking `wait()` syscall.
385 12 : drop(proc);
386 12 : }
387 :
388 24 : result
389 24 : }
390 :
391 : ///
392 : /// Process one request for WAL redo using wal-redo postgres
393 : ///
394 : /// # Cancel-Safety
395 : ///
396 : /// Cancellation safe.
397 : #[allow(clippy::too_many_arguments)]
398 18 : async fn apply_batch_postgres(
399 18 : &self,
400 18 : key: Key,
401 18 : lsn: Lsn,
402 18 : base_img: Option<Bytes>,
403 18 : base_img_lsn: Lsn,
404 18 : records: &[(Lsn, NeonWalRecord)],
405 18 : wal_redo_timeout: Duration,
406 18 : pg_version: u32,
407 18 : ) -> Result<Bytes, Error> {
408 18 : *(self.last_redo_at.lock().unwrap()) = Some(Instant::now());
409 :
410 18 : let (rel, blknum) = key.to_rel_block().context("invalid record")?;
411 : const MAX_RETRY_ATTEMPTS: u32 = 1;
412 18 : let mut n_attempts = 0u32;
413 24 : loop {
414 24 : let base_img = &base_img;
415 24 : let closure = |proc: Arc<Process>| async move {
416 24 : let started_at = std::time::Instant::now();
417 :
418 : // Relational WAL records are applied using wal-redo-postgres
419 24 : let result = proc
420 24 : .apply_wal_records(rel, blknum, base_img, records, wal_redo_timeout)
421 48 : .await
422 24 : .context("apply_wal_records");
423 24 :
424 24 : let duration = started_at.elapsed();
425 24 :
426 24 : let len = records.len();
427 48 : let nbytes = records.iter().fold(0, |acumulator, record| {
428 48 : acumulator
429 48 : + match &record.1 {
430 48 : NeonWalRecord::Postgres { rec, .. } => rec.len(),
431 0 : _ => unreachable!("Only PostgreSQL records are accepted in this batch"),
432 : }
433 48 : });
434 24 :
435 24 : WAL_REDO_TIME.observe(duration.as_secs_f64());
436 24 : WAL_REDO_RECORDS_HISTOGRAM.observe(len as f64);
437 24 : WAL_REDO_BYTES_HISTOGRAM.observe(nbytes as f64);
438 24 :
439 24 : debug!(
440 0 : "postgres applied {} WAL records ({} bytes) in {} us to reconstruct page image at LSN {}",
441 0 : len,
442 0 : nbytes,
443 0 : duration.as_micros(),
444 : lsn
445 : );
446 :
447 24 : if let Err(e) = result.as_ref() {
448 12 : error!(
449 0 : "error applying {} WAL records {}..{} ({} bytes) to key {key}, from base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
450 0 : records.len(),
451 12 : records.first().map(|p| p.0).unwrap_or(Lsn(0)),
452 12 : records.last().map(|p| p.0).unwrap_or(Lsn(0)),
453 : nbytes,
454 : base_img_lsn,
455 : lsn,
456 : n_attempts,
457 : e,
458 : );
459 12 : }
460 :
461 24 : result.map_err(Error::Other)
462 24 : };
463 48 : let result = self.do_with_walredo_process(pg_version, closure).await;
464 :
465 24 : if result.is_ok() && n_attempts != 0 {
466 0 : info!(n_attempts, "retried walredo succeeded");
467 24 : }
468 24 : n_attempts += 1;
469 24 : if n_attempts > MAX_RETRY_ATTEMPTS || result.is_ok() {
470 18 : return result;
471 6 : }
472 : }
473 18 : }
474 :
475 : ///
476 : /// Process a batch of WAL records using bespoken Neon code.
477 : ///
478 0 : fn apply_batch_neon(
479 0 : &self,
480 0 : key: Key,
481 0 : lsn: Lsn,
482 0 : base_img: Option<Bytes>,
483 0 : records: &[(Lsn, NeonWalRecord)],
484 0 : ) -> Result<Bytes, Error> {
485 0 : let start_time = Instant::now();
486 0 :
487 0 : let mut page = BytesMut::new();
488 0 : if let Some(fpi) = base_img {
489 0 : // If full-page image is provided, then use it...
490 0 : page.extend_from_slice(&fpi[..]);
491 0 : } else {
492 : // All the current WAL record types that we can handle require a base image.
493 0 : bail!("invalid neon WAL redo request with no base image");
494 : }
495 :
496 : // Apply all the WAL records in the batch
497 0 : for (record_lsn, record) in records.iter() {
498 0 : self.apply_record_neon(key, &mut page, *record_lsn, record)?;
499 : }
500 : // Success!
501 0 : let duration = start_time.elapsed();
502 0 : // FIXME: using the same metric here creates a bimodal distribution by default, and because
503 0 : // there could be multiple batch sizes this would be N+1 modal.
504 0 : WAL_REDO_TIME.observe(duration.as_secs_f64());
505 0 :
506 0 : debug!(
507 0 : "neon applied {} WAL records in {} us to reconstruct page image at LSN {}",
508 0 : records.len(),
509 0 : duration.as_micros(),
510 : lsn
511 : );
512 :
513 0 : Ok(page.freeze())
514 0 : }
515 :
516 0 : fn apply_record_neon(
517 0 : &self,
518 0 : key: Key,
519 0 : page: &mut BytesMut,
520 0 : record_lsn: Lsn,
521 0 : record: &NeonWalRecord,
522 0 : ) -> anyhow::Result<()> {
523 0 : apply_neon::apply_in_neon(record, record_lsn, key, page)?;
524 :
525 0 : Ok(())
526 0 : }
527 : }
528 :
529 : #[cfg(test)]
530 : mod tests {
531 : use super::PostgresRedoManager;
532 : use crate::repository::Key;
533 : use crate::{config::PageServerConf, walrecord::NeonWalRecord};
534 : use bytes::Bytes;
535 : use pageserver_api::shard::TenantShardId;
536 : use std::str::FromStr;
537 : use tracing::Instrument;
538 : use utils::{id::TenantId, lsn::Lsn};
539 :
540 : #[tokio::test]
541 6 : async fn short_v14_redo() {
542 6 : let expected = std::fs::read("test_data/short_v14_redo.page").unwrap();
543 6 :
544 6 : let h = RedoHarness::new().unwrap();
545 6 :
546 6 : let page = h
547 6 : .manager
548 6 : .request_redo(
549 6 : Key {
550 6 : field1: 0,
551 6 : field2: 1663,
552 6 : field3: 13010,
553 6 : field4: 1259,
554 6 : field5: 0,
555 6 : field6: 0,
556 6 : },
557 6 : Lsn::from_str("0/16E2408").unwrap(),
558 6 : None,
559 6 : short_records(),
560 6 : 14,
561 6 : )
562 6 : .instrument(h.span())
563 12 : .await
564 6 : .unwrap();
565 6 :
566 6 : assert_eq!(&expected, &*page);
567 6 : }
568 :
569 : #[tokio::test]
570 6 : async fn short_v14_fails_for_wrong_key_but_returns_zero_page() {
571 6 : let h = RedoHarness::new().unwrap();
572 6 :
573 6 : let page = h
574 6 : .manager
575 6 : .request_redo(
576 6 : Key {
577 6 : field1: 0,
578 6 : field2: 1663,
579 6 : // key should be 13010
580 6 : field3: 13130,
581 6 : field4: 1259,
582 6 : field5: 0,
583 6 : field6: 0,
584 6 : },
585 6 : Lsn::from_str("0/16E2408").unwrap(),
586 6 : None,
587 6 : short_records(),
588 6 : 14,
589 6 : )
590 6 : .instrument(h.span())
591 12 : .await
592 6 : .unwrap();
593 6 :
594 6 : // TODO: there will be some stderr printout, which is forwarded to tracing that could
595 6 : // perhaps be captured as long as it's in the same thread.
596 6 : assert_eq!(page, crate::ZERO_PAGE);
597 6 : }
598 :
599 : #[tokio::test]
600 6 : async fn test_stderr() {
601 6 : let h = RedoHarness::new().unwrap();
602 6 : h
603 6 : .manager
604 6 : .request_redo(
605 6 : Key::from_i128(0),
606 6 : Lsn::INVALID,
607 6 : None,
608 6 : short_records(),
609 6 : 16, /* 16 currently produces stderr output on startup, which adds a nice extra edge */
610 6 : )
611 6 : .instrument(h.span())
612 24 : .await
613 6 : .unwrap_err();
614 6 : }
615 :
616 : #[allow(clippy::octal_escapes)]
617 18 : fn short_records() -> Vec<(Lsn, NeonWalRecord)> {
618 18 : vec![
619 18 : (
620 18 : Lsn::from_str("0/16A9388").unwrap(),
621 18 : NeonWalRecord::Postgres {
622 18 : will_init: true,
623 18 : rec: Bytes::from_static(b"j\x03\0\0\0\x04\0\0\xe8\x7fj\x01\0\0\0\0\0\n\0\0\xd0\x16\x13Y\0\x10\0\04\x03\xd4\0\x05\x7f\x06\0\0\xd22\0\0\xeb\x04\0\0\0\0\0\0\xff\x03\0\0\0\0\x80\xeca\x01\0\0\x01\0\xd4\0\xa0\x1d\0 \x04 \0\0\0\0/\0\x01\0\xa0\x9dX\x01\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0.\0\x01\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\00\x9f\x9a\x01P\x9e\xb2\x01\0\x04\0\0\0\0\0\0\0\0\0\0\0\0\0\0\x02\0!\0\x01\x08 \xff\xff\xff?\0\0\0\0\0\0@\0\0another_table\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\x98\x08\0\0\x02@\0\0\0\0\0\0\n\0\0\0\x02\0\0\0\0@\0\0\0\0\0\0\0\0\0\0\0\0\x80\xbf\0\0\0\0\0\0\0\0\0\0pr\x01\0\0\0\0\0\0\0\0\x01d\0\0\0\0\0\0\x04\0\0\x01\0\0\0\0\0\0\0\x0c\x02\0\0\0\0\0\0\0\0\0\0\0\0\0\0/\0!\x80\x03+ \xff\xff\xff\x7f\0\0\0\0\0\xdf\x04\0\0pg_type\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\x0b\0\0\0G\0\0\0\0\0\0\0\n\0\0\0\x02\0\0\0\0\0\0\0\0\0\0\0\x0e\0\0\0\0@\x16D\x0e\0\0\0K\x10\0\0\x01\0pr \0\0\0\0\0\0\0\0\x01n\0\0\0\0\0\xd6\x02\0\0\x01\0\0\0[\x01\0\0\0\0\0\0\0\t\x04\0\0\x02\0\0\0\x01\0\0\0\n\0\0\0\n\0\0\0\x7f\0\0\0\0\0\0\0\n\0\0\0\x02\0\0\0\0\0\0C\x01\0\0\x15\x01\0\0\0\0\0\0\0\0\0\0\0\0\0\0.\0!\x80\x03+ \xff\xff\xff\x7f\0\0\0\0\0;\n\0\0pg_statistic\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\x0b\0\0\0\xfd.\0\0\0\0\0\0\n\0\0\0\x02\0\0\0;\n\0\0\0\0\0\0\x13\0\0\0\0\0\xcbC\x13\0\0\0\x18\x0b\0\0\x01\0pr\x1f\0\0\0\0\0\0\0\0\x01n\0\0\0\0\0\xd6\x02\0\0\x01\0\0\0C\x01\0\0\0\0\0\0\0\t\x04\0\0\x01\0\0\0\x01\0\0\0\n\0\0\0\n\0\0\0\x7f\0\0\0\0\0\0\x02\0\x01")
624 18 : }
625 18 : ),
626 18 : (
627 18 : Lsn::from_str("0/16D4080").unwrap(),
628 18 : NeonWalRecord::Postgres {
629 18 : will_init: false,
630 18 : rec: Bytes::from_static(b"\xbc\0\0\0\0\0\0\0h?m\x01\0\0\0\0p\n\0\09\x08\xa3\xea\0 \x8c\0\x7f\x06\0\0\xd22\0\0\xeb\x04\0\0\0\0\0\0\xff\x02\0@\0\0another_table\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\x98\x08\0\0\x02@\0\0\0\0\0\0\n\0\0\0\x02\0\0\0\0@\0\0\0\0\0\0\x05\0\0\0\0@zD\x05\0\0\0\0\0\0\0\0\0pr\x01\0\0\0\0\0\0\0\0\x01d\0\0\0\0\0\0\x04\0\0\x01\0\0\0\x02\0")
631 18 : }
632 18 : )
633 18 : ]
634 18 : }
635 :
636 : struct RedoHarness {
637 : // underscored because unused, except for removal at drop
638 : _repo_dir: camino_tempfile::Utf8TempDir,
639 : manager: PostgresRedoManager,
640 : tenant_shard_id: TenantShardId,
641 : }
642 :
643 : impl RedoHarness {
644 18 : fn new() -> anyhow::Result<Self> {
645 18 : crate::tenant::harness::setup_logging();
646 :
647 18 : let repo_dir = camino_tempfile::tempdir()?;
648 18 : let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
649 18 : let conf = Box::leak(Box::new(conf));
650 18 : let tenant_shard_id = TenantShardId::unsharded(TenantId::generate());
651 18 :
652 18 : let manager = PostgresRedoManager::new(conf, tenant_shard_id);
653 18 :
654 18 : Ok(RedoHarness {
655 18 : _repo_dir: repo_dir,
656 18 : manager,
657 18 : tenant_shard_id,
658 18 : })
659 18 : }
660 18 : fn span(&self) -> tracing::Span {
661 18 : tracing::info_span!("RedoHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
662 18 : }
663 : }
664 : }
|
