Line data Source code
1 : //! Actual Postgres connection handler to stream WAL to the server.
2 :
3 : use std::{
4 : error::Error,
5 : pin::pin,
6 : str::FromStr,
7 : sync::Arc,
8 : time::{Duration, SystemTime},
9 : };
10 :
11 : use anyhow::{anyhow, Context};
12 : use bytes::BytesMut;
13 : use chrono::{NaiveDateTime, Utc};
14 : use fail::fail_point;
15 : use futures::StreamExt;
16 : use postgres::{error::SqlState, SimpleQueryMessage, SimpleQueryRow};
17 : use postgres_ffi::WAL_SEGMENT_SIZE;
18 : use postgres_ffi::{v14::xlog_utils::normalize_lsn, waldecoder::WalDecodeError};
19 : use postgres_protocol::message::backend::ReplicationMessage;
20 : use postgres_types::PgLsn;
21 : use tokio::{select, sync::watch, time};
22 : use tokio_postgres::{replication::ReplicationStream, Client};
23 : use tokio_util::sync::CancellationToken;
24 : use tracing::{debug, error, info, trace, warn, Instrument};
25 : use wal_decoder::models::{FlushUncommittedRecords, InterpretedWalRecord};
26 :
27 : use super::TaskStateUpdate;
28 : use crate::{
29 : context::RequestContext,
30 : metrics::{LIVE_CONNECTIONS, WALRECEIVER_STARTED_CONNECTIONS, WAL_INGEST},
31 : pgdatadir_mapping::DatadirModification,
32 : task_mgr::{TaskKind, WALRECEIVER_RUNTIME},
33 : tenant::{debug_assert_current_span_has_tenant_and_timeline_id, Timeline, WalReceiverInfo},
34 : walingest::WalIngest,
35 : };
36 : use postgres_backend::is_expected_io_error;
37 : use postgres_connection::PgConnectionConfig;
38 : use postgres_ffi::waldecoder::WalStreamDecoder;
39 : use utils::{bin_ser::BeSer, id::NodeId, lsn::Lsn};
40 : use utils::{pageserver_feedback::PageserverFeedback, sync::gate::GateError};
41 :
42 : /// Status of the connection.
43 : #[derive(Debug, Clone, Copy)]
44 : pub(super) struct WalConnectionStatus {
45 : /// If we were able to initiate a postgres connection, this means that safekeeper process is at least running.
46 : pub is_connected: bool,
47 : /// Defines a healthy connection as one on which pageserver received WAL from safekeeper
48 : /// and is able to process it in walingest without errors.
49 : pub has_processed_wal: bool,
50 : /// Connection establishment time or the timestamp of a latest connection message received.
51 : pub latest_connection_update: NaiveDateTime,
52 : /// Time of the latest WAL message received.
53 : pub latest_wal_update: NaiveDateTime,
54 : /// Latest WAL update contained WAL up to this LSN. Next WAL message with start from that LSN.
55 : pub streaming_lsn: Option<Lsn>,
56 : /// Latest commit_lsn received from the safekeeper. Can be zero if no message has been received yet.
57 : pub commit_lsn: Option<Lsn>,
58 : /// The node it is connected to
59 : pub node: NodeId,
60 : }
61 :
62 : pub(super) enum WalReceiverError {
63 : /// An error of a type that does not indicate an issue, e.g. a connection closing
64 : ExpectedSafekeeperError(postgres::Error),
65 : /// An "error" message that carries a SUCCESSFUL_COMPLETION status code. Carries
66 : /// the message part of the original postgres error
67 : SuccessfulCompletion(String),
68 : /// Generic error
69 : Other(anyhow::Error),
70 : ClosedGate,
71 : }
72 :
73 : impl From<tokio_postgres::Error> for WalReceiverError {
74 0 : fn from(err: tokio_postgres::Error) -> Self {
75 0 : if let Some(dberror) = err.as_db_error().filter(|db_error| {
76 0 : db_error.code() == &SqlState::SUCCESSFUL_COMPLETION
77 0 : && db_error.message().contains("ending streaming")
78 0 : }) {
79 : // Strip the outer DbError, which carries a misleading "error" severity
80 0 : Self::SuccessfulCompletion(dberror.message().to_string())
81 0 : } else if err.is_closed()
82 0 : || err
83 0 : .source()
84 0 : .and_then(|source| source.downcast_ref::<std::io::Error>())
85 0 : .map(is_expected_io_error)
86 0 : .unwrap_or(false)
87 : {
88 0 : Self::ExpectedSafekeeperError(err)
89 : } else {
90 0 : Self::Other(anyhow::Error::new(err))
91 : }
92 0 : }
93 : }
94 :
95 : impl From<anyhow::Error> for WalReceiverError {
96 0 : fn from(err: anyhow::Error) -> Self {
97 0 : Self::Other(err)
98 0 : }
99 : }
100 :
101 : impl From<WalDecodeError> for WalReceiverError {
102 0 : fn from(err: WalDecodeError) -> Self {
103 0 : Self::Other(anyhow::Error::new(err))
104 0 : }
105 : }
106 :
107 : /// Open a connection to the given safekeeper and receive WAL, sending back progress
108 : /// messages as we go.
109 : #[allow(clippy::too_many_arguments)]
110 0 : pub(super) async fn handle_walreceiver_connection(
111 0 : timeline: Arc<Timeline>,
112 0 : wal_source_connconf: PgConnectionConfig,
113 0 : events_sender: watch::Sender<TaskStateUpdate<WalConnectionStatus>>,
114 0 : cancellation: CancellationToken,
115 0 : connect_timeout: Duration,
116 0 : ctx: RequestContext,
117 0 : node: NodeId,
118 0 : ingest_batch_size: u64,
119 0 : ) -> Result<(), WalReceiverError> {
120 0 : debug_assert_current_span_has_tenant_and_timeline_id();
121 :
122 : // prevent timeline shutdown from finishing until we have exited
123 0 : let _guard = timeline.gate.enter().map_err(|e| match e {
124 0 : GateError::GateClosed => WalReceiverError::ClosedGate,
125 0 : })?;
126 : // This function spawns a side-car task (WalReceiverConnectionPoller).
127 : // Get its gate guard now as well.
128 0 : let poller_guard = timeline.gate.enter().map_err(|e| match e {
129 0 : GateError::GateClosed => WalReceiverError::ClosedGate,
130 0 : })?;
131 :
132 0 : WALRECEIVER_STARTED_CONNECTIONS.inc();
133 0 :
134 0 : // Connect to the database in replication mode.
135 0 : info!("connecting to {wal_source_connconf:?}");
136 :
137 0 : let (replication_client, connection) = {
138 0 : let mut config = wal_source_connconf.to_tokio_postgres_config();
139 0 : config.application_name("pageserver");
140 0 : config.replication_mode(tokio_postgres::config::ReplicationMode::Physical);
141 0 : match time::timeout(connect_timeout, config.connect(postgres::NoTls)).await {
142 0 : Ok(client_and_conn) => client_and_conn?,
143 0 : Err(_elapsed) => {
144 0 : // Timing out to connect to a safekeeper node could happen long time, due to
145 0 : // many reasons that pageserver cannot control.
146 0 : // Do not produce an error, but make it visible, that timeouts happen by logging the `event.
147 0 : info!("Timed out while waiting {connect_timeout:?} for walreceiver connection to open");
148 0 : return Ok(());
149 : }
150 : }
151 : };
152 :
153 0 : debug!("connected!");
154 0 : let mut connection_status = WalConnectionStatus {
155 0 : is_connected: true,
156 0 : has_processed_wal: false,
157 0 : latest_connection_update: Utc::now().naive_utc(),
158 0 : latest_wal_update: Utc::now().naive_utc(),
159 0 : streaming_lsn: None,
160 0 : commit_lsn: None,
161 0 : node,
162 0 : };
163 0 : if let Err(e) = events_sender.send(TaskStateUpdate::Progress(connection_status)) {
164 0 : warn!("Wal connection event listener dropped right after connection init, aborting the connection: {e}");
165 0 : return Ok(());
166 0 : }
167 0 :
168 0 : // The connection object performs the actual communication with the database,
169 0 : // so spawn it off to run on its own. It shouldn't outlive this function, but,
170 0 : // due to lack of async drop, we can't enforce that. However, we ensure that
171 0 : // 1. it is sensitive to `cancellation` and
172 0 : // 2. holds the Timeline gate open so that after timeline shutdown,
173 0 : // we know this task is gone.
174 0 : let _connection_ctx = ctx.detached_child(
175 0 : TaskKind::WalReceiverConnectionPoller,
176 0 : ctx.download_behavior(),
177 0 : );
178 0 : let connection_cancellation = cancellation.clone();
179 0 : WALRECEIVER_RUNTIME.spawn(
180 0 : async move {
181 0 : debug_assert_current_span_has_tenant_and_timeline_id();
182 0 : select! {
183 0 : connection_result = connection => match connection_result {
184 0 : Ok(()) => debug!("Walreceiver db connection closed"),
185 0 : Err(connection_error) => {
186 0 : match WalReceiverError::from(connection_error) {
187 0 : WalReceiverError::ExpectedSafekeeperError(_) => {
188 0 : // silence, because most likely we've already exited the outer call
189 0 : // with a similar error.
190 0 : },
191 0 : WalReceiverError::SuccessfulCompletion(_) => {}
192 0 : WalReceiverError::ClosedGate => {
193 0 : // doesn't happen at runtime
194 0 : }
195 0 : WalReceiverError::Other(err) => {
196 0 : warn!("Connection aborted: {err:#}")
197 : }
198 : }
199 : }
200 : },
201 0 : _ = connection_cancellation.cancelled() => debug!("Connection cancelled"),
202 : }
203 0 : drop(poller_guard);
204 0 : }
205 : // Enrich the log lines emitted by this closure with meaningful context.
206 : // TODO: technically, this task outlives the surrounding function, so, the
207 : // spans won't be properly nested.
208 0 : .instrument(tracing::info_span!("poller")),
209 : );
210 :
211 0 : let _guard = LIVE_CONNECTIONS
212 0 : .with_label_values(&["wal_receiver"])
213 0 : .guard();
214 :
215 0 : let identify = identify_system(&replication_client).await?;
216 0 : info!("{identify:?}");
217 :
218 0 : let end_of_wal = Lsn::from(u64::from(identify.xlogpos));
219 0 : let mut caught_up = false;
220 0 :
221 0 : connection_status.latest_connection_update = Utc::now().naive_utc();
222 0 : connection_status.latest_wal_update = Utc::now().naive_utc();
223 0 : connection_status.commit_lsn = Some(end_of_wal);
224 0 : if let Err(e) = events_sender.send(TaskStateUpdate::Progress(connection_status)) {
225 0 : warn!("Wal connection event listener dropped after IDENTIFY_SYSTEM, aborting the connection: {e}");
226 0 : return Ok(());
227 0 : }
228 0 :
229 0 : //
230 0 : // Start streaming the WAL, from where we left off previously.
231 0 : //
232 0 : // If we had previously received WAL up to some point in the middle of a WAL record, we
233 0 : // better start from the end of last full WAL record, not in the middle of one.
234 0 : let mut last_rec_lsn = timeline.get_last_record_lsn();
235 0 : let mut startpoint = last_rec_lsn;
236 0 :
237 0 : if startpoint == Lsn(0) {
238 0 : return Err(WalReceiverError::Other(anyhow!("No previous WAL position")));
239 0 : }
240 0 :
241 0 : // There might be some padding after the last full record, skip it.
242 0 : startpoint += startpoint.calc_padding(8u32);
243 0 :
244 0 : // If the starting point is at a WAL page boundary, skip past the page header. We don't need the page headers
245 0 : // for anything, and in some corner cases, the compute node might have never generated the WAL for page headers
246 0 : //. That happens if you create a branch at page boundary: the start point of the branch is at the page boundary,
247 0 : // but when the compute node first starts on the branch, we normalize the first REDO position to just after the page
248 0 : // header (see generate_pg_control()), so the WAL for the page header is never streamed from the compute node
249 0 : // to the safekeepers.
250 0 : startpoint = normalize_lsn(startpoint, WAL_SEGMENT_SIZE);
251 0 :
252 0 : info!("last_record_lsn {last_rec_lsn} starting replication from {startpoint}, safekeeper is at {end_of_wal}...");
253 :
254 0 : let query = format!("START_REPLICATION PHYSICAL {startpoint}");
255 :
256 0 : let copy_stream = replication_client.copy_both_simple(&query).await?;
257 0 : let mut physical_stream = pin!(ReplicationStream::new(copy_stream));
258 0 :
259 0 : let mut waldecoder = WalStreamDecoder::new(startpoint, timeline.pg_version);
260 :
261 0 : let mut walingest = WalIngest::new(timeline.as_ref(), startpoint, &ctx).await?;
262 :
263 0 : while let Some(replication_message) = {
264 0 : select! {
265 0 : _ = cancellation.cancelled() => {
266 0 : debug!("walreceiver interrupted");
267 0 : None
268 : }
269 0 : replication_message = physical_stream.next() => replication_message,
270 : }
271 : } {
272 0 : let replication_message = replication_message?;
273 :
274 0 : let now = Utc::now().naive_utc();
275 0 : let last_rec_lsn_before_msg = last_rec_lsn;
276 0 :
277 0 : // Update the connection status before processing the message. If the message processing
278 0 : // fails (e.g. in walingest), we still want to know latests LSNs from the safekeeper.
279 0 : match &replication_message {
280 0 : ReplicationMessage::XLogData(xlog_data) => {
281 0 : connection_status.latest_connection_update = now;
282 0 : connection_status.commit_lsn = Some(Lsn::from(xlog_data.wal_end()));
283 0 : connection_status.streaming_lsn = Some(Lsn::from(
284 0 : xlog_data.wal_start() + xlog_data.data().len() as u64,
285 0 : ));
286 0 : if !xlog_data.data().is_empty() {
287 0 : connection_status.latest_wal_update = now;
288 0 : }
289 : }
290 0 : ReplicationMessage::PrimaryKeepAlive(keepalive) => {
291 0 : connection_status.latest_connection_update = now;
292 0 : connection_status.commit_lsn = Some(Lsn::from(keepalive.wal_end()));
293 0 : }
294 0 : ReplicationMessage::RawInterpretedWalRecords(raw) => {
295 0 : connection_status.latest_connection_update = now;
296 0 : if !raw.data().is_empty() {
297 0 : connection_status.latest_wal_update = now;
298 0 : }
299 :
300 0 : connection_status.commit_lsn = Some(Lsn::from(raw.commit_lsn()));
301 0 : connection_status.streaming_lsn = Some(Lsn::from(raw.streaming_lsn()));
302 : }
303 0 : &_ => {}
304 : };
305 0 : if let Err(e) = events_sender.send(TaskStateUpdate::Progress(connection_status)) {
306 0 : warn!("Wal connection event listener dropped, aborting the connection: {e}");
307 0 : return Ok(());
308 0 : }
309 :
310 0 : async fn commit(
311 0 : modification: &mut DatadirModification<'_>,
312 0 : uncommitted: &mut u64,
313 0 : filtered: &mut u64,
314 0 : ctx: &RequestContext,
315 0 : ) -> anyhow::Result<()> {
316 0 : WAL_INGEST
317 0 : .records_committed
318 0 : .inc_by(*uncommitted - *filtered);
319 0 : modification.commit(ctx).await?;
320 0 : *uncommitted = 0;
321 0 : *filtered = 0;
322 0 : Ok(())
323 0 : }
324 :
325 0 : let status_update = match replication_message {
326 0 : ReplicationMessage::RawInterpretedWalRecords(raw) => {
327 0 : WAL_INGEST.bytes_received.inc_by(raw.data().len() as u64);
328 0 :
329 0 : let mut uncommitted_records = 0;
330 0 : let mut filtered_records = 0;
331 0 :
332 0 : // This is the end LSN of the raw WAL from which the records
333 0 : // were interpreted.
334 0 : let streaming_lsn = Lsn::from(raw.streaming_lsn());
335 0 : tracing::debug!(
336 0 : "Received WAL up to {streaming_lsn} with next_record_lsn={}",
337 0 : Lsn(raw.next_record_lsn().unwrap_or(0))
338 : );
339 :
340 0 : let records = Vec::<InterpretedWalRecord>::des(raw.data()).with_context(|| {
341 0 : anyhow::anyhow!(
342 0 : "Failed to deserialize interpreted records ending at LSN {streaming_lsn}"
343 0 : )
344 0 : })?;
345 :
346 : // We start the modification at 0 because each interpreted record
347 : // advances it to its end LSN. 0 is just an initialization placeholder.
348 0 : let mut modification = timeline.begin_modification(Lsn(0));
349 :
350 0 : for interpreted in records {
351 0 : if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes)
352 0 : && uncommitted_records > 0
353 : {
354 0 : commit(
355 0 : &mut modification,
356 0 : &mut uncommitted_records,
357 0 : &mut filtered_records,
358 0 : &ctx,
359 0 : )
360 0 : .await?;
361 0 : }
362 :
363 0 : let next_record_lsn = interpreted.next_record_lsn;
364 0 : let ingested = walingest
365 0 : .ingest_record(interpreted, &mut modification, &ctx)
366 0 : .await
367 0 : .with_context(|| format!("could not ingest record at {next_record_lsn}"))?;
368 :
369 0 : if !ingested {
370 0 : tracing::debug!("ingest: filtered out record @ LSN {next_record_lsn}");
371 0 : WAL_INGEST.records_filtered.inc();
372 0 : filtered_records += 1;
373 0 : }
374 :
375 0 : uncommitted_records += 1;
376 0 :
377 0 : // FIXME: this cannot be made pausable_failpoint without fixing the
378 0 : // failpoint library; in tests, the added amount of debugging will cause us
379 0 : // to timeout the tests.
380 0 : fail_point!("walreceiver-after-ingest");
381 0 :
382 0 : // Commit every ingest_batch_size records. Even if we filtered out
383 0 : // all records, we still need to call commit to advance the LSN.
384 0 : if uncommitted_records >= ingest_batch_size
385 0 : || modification.approx_pending_bytes()
386 0 : > DatadirModification::MAX_PENDING_BYTES
387 : {
388 0 : commit(
389 0 : &mut modification,
390 0 : &mut uncommitted_records,
391 0 : &mut filtered_records,
392 0 : &ctx,
393 0 : )
394 0 : .await?;
395 0 : }
396 : }
397 :
398 : // Records might have been filtered out on the safekeeper side, but we still
399 : // need to advance last record LSN on all shards. If we've not ingested the latest
400 : // record, then set the LSN of the modification past it. This way all shards
401 : // advance their last record LSN at the same time.
402 0 : let needs_last_record_lsn_advance = match raw.next_record_lsn().map(Lsn::from) {
403 0 : Some(lsn) if lsn > modification.get_lsn() => {
404 0 : modification.set_lsn(lsn).unwrap();
405 0 : true
406 : }
407 0 : _ => false,
408 : };
409 :
410 0 : if uncommitted_records > 0 || needs_last_record_lsn_advance {
411 : // Commit any uncommitted records
412 0 : commit(
413 0 : &mut modification,
414 0 : &mut uncommitted_records,
415 0 : &mut filtered_records,
416 0 : &ctx,
417 0 : )
418 0 : .await?;
419 0 : }
420 :
421 0 : if !caught_up && streaming_lsn >= end_of_wal {
422 0 : info!("caught up at LSN {streaming_lsn}");
423 0 : caught_up = true;
424 0 : }
425 :
426 0 : tracing::debug!(
427 0 : "Ingested WAL up to {streaming_lsn}. Last record LSN is {}",
428 0 : timeline.get_last_record_lsn()
429 : );
430 :
431 0 : Some(streaming_lsn)
432 : }
433 :
434 0 : ReplicationMessage::XLogData(xlog_data) => {
435 0 : // Pass the WAL data to the decoder, and see if we can decode
436 0 : // more records as a result.
437 0 : let data = xlog_data.data();
438 0 : let startlsn = Lsn::from(xlog_data.wal_start());
439 0 : let endlsn = startlsn + data.len() as u64;
440 0 :
441 0 : trace!("received XLogData between {startlsn} and {endlsn}");
442 :
443 0 : WAL_INGEST.bytes_received.inc_by(data.len() as u64);
444 0 : waldecoder.feed_bytes(data);
445 0 :
446 0 : {
447 0 : let mut modification = timeline.begin_modification(startlsn);
448 0 : let mut uncommitted_records = 0;
449 0 : let mut filtered_records = 0;
450 :
451 0 : while let Some((next_record_lsn, recdata)) = waldecoder.poll_decode()? {
452 : // It is important to deal with the aligned records as lsn in getPage@LSN is
453 : // aligned and can be several bytes bigger. Without this alignment we are
454 : // at risk of hitting a deadlock.
455 0 : if !next_record_lsn.is_aligned() {
456 0 : return Err(WalReceiverError::Other(anyhow!("LSN not aligned")));
457 0 : }
458 :
459 : // Deserialize and interpret WAL record
460 0 : let interpreted = InterpretedWalRecord::from_bytes_filtered(
461 0 : recdata,
462 0 : modification.tline.get_shard_identity(),
463 0 : next_record_lsn,
464 0 : modification.tline.pg_version,
465 0 : )?;
466 :
467 0 : if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes)
468 0 : && uncommitted_records > 0
469 : {
470 : // Special case: legacy PG database creations operate by reading pages from a 'template' database:
471 : // these are the only kinds of WAL record that require reading data blocks while ingesting. Ensure
472 : // all earlier writes of data blocks are visible by committing any modification in flight.
473 0 : commit(
474 0 : &mut modification,
475 0 : &mut uncommitted_records,
476 0 : &mut filtered_records,
477 0 : &ctx,
478 0 : )
479 0 : .await?;
480 0 : }
481 :
482 : // Ingest the records without immediately committing them.
483 0 : let ingested = walingest
484 0 : .ingest_record(interpreted, &mut modification, &ctx)
485 0 : .await
486 0 : .with_context(|| {
487 0 : format!("could not ingest record at {next_record_lsn}")
488 0 : })?;
489 0 : if !ingested {
490 0 : tracing::debug!("ingest: filtered out record @ LSN {next_record_lsn}");
491 0 : WAL_INGEST.records_filtered.inc();
492 0 : filtered_records += 1;
493 0 : }
494 :
495 : // FIXME: this cannot be made pausable_failpoint without fixing the
496 : // failpoint library; in tests, the added amount of debugging will cause us
497 : // to timeout the tests.
498 0 : fail_point!("walreceiver-after-ingest");
499 0 :
500 0 : last_rec_lsn = next_record_lsn;
501 0 :
502 0 : // Commit every ingest_batch_size records. Even if we filtered out
503 0 : // all records, we still need to call commit to advance the LSN.
504 0 : uncommitted_records += 1;
505 0 : if uncommitted_records >= ingest_batch_size
506 0 : || modification.approx_pending_bytes()
507 0 : > DatadirModification::MAX_PENDING_BYTES
508 : {
509 0 : commit(
510 0 : &mut modification,
511 0 : &mut uncommitted_records,
512 0 : &mut filtered_records,
513 0 : &ctx,
514 0 : )
515 0 : .await?;
516 0 : }
517 : }
518 :
519 : // Commit the remaining records.
520 0 : if uncommitted_records > 0 {
521 0 : commit(
522 0 : &mut modification,
523 0 : &mut uncommitted_records,
524 0 : &mut filtered_records,
525 0 : &ctx,
526 0 : )
527 0 : .await?;
528 0 : }
529 : }
530 :
531 0 : if !caught_up && endlsn >= end_of_wal {
532 0 : info!("caught up at LSN {endlsn}");
533 0 : caught_up = true;
534 0 : }
535 :
536 0 : Some(endlsn)
537 : }
538 :
539 0 : ReplicationMessage::PrimaryKeepAlive(keepalive) => {
540 0 : let wal_end = keepalive.wal_end();
541 0 : let timestamp = keepalive.timestamp();
542 0 : let reply_requested = keepalive.reply() != 0;
543 0 :
544 0 : trace!("received PrimaryKeepAlive(wal_end: {wal_end}, timestamp: {timestamp:?} reply: {reply_requested})");
545 :
546 0 : if reply_requested {
547 0 : Some(last_rec_lsn)
548 : } else {
549 0 : None
550 : }
551 : }
552 :
553 0 : _ => None,
554 : };
555 :
556 0 : if !connection_status.has_processed_wal && last_rec_lsn > last_rec_lsn_before_msg {
557 : // We have successfully processed at least one WAL record.
558 0 : connection_status.has_processed_wal = true;
559 0 : if let Err(e) = events_sender.send(TaskStateUpdate::Progress(connection_status)) {
560 0 : warn!("Wal connection event listener dropped, aborting the connection: {e}");
561 0 : return Ok(());
562 0 : }
563 0 : }
564 :
565 0 : if let Some(last_lsn) = status_update {
566 0 : let timeline_remote_consistent_lsn = timeline
567 0 : .get_remote_consistent_lsn_visible()
568 0 : .unwrap_or(Lsn(0));
569 0 :
570 0 : // The last LSN we processed. It is not guaranteed to survive pageserver crash.
571 0 : let last_received_lsn = last_lsn;
572 0 : // `disk_consistent_lsn` is the LSN at which page server guarantees local persistence of all received data
573 0 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
574 0 : // The last LSN that is synced to remote storage and is guaranteed to survive pageserver crash
575 0 : // Used by safekeepers to remove WAL preceding `remote_consistent_lsn`.
576 0 : let remote_consistent_lsn = timeline_remote_consistent_lsn;
577 0 : let ts = SystemTime::now();
578 0 :
579 0 : // Update the status about what we just received. This is shown in the mgmt API.
580 0 : let last_received_wal = WalReceiverInfo {
581 0 : wal_source_connconf: wal_source_connconf.clone(),
582 0 : last_received_msg_lsn: last_lsn,
583 0 : last_received_msg_ts: ts
584 0 : .duration_since(SystemTime::UNIX_EPOCH)
585 0 : .expect("Received message time should be before UNIX EPOCH!")
586 0 : .as_micros(),
587 0 : };
588 0 : *timeline.last_received_wal.lock().unwrap() = Some(last_received_wal);
589 :
590 : // Send the replication feedback message.
591 : // Regular standby_status_update fields are put into this message.
592 0 : let current_timeline_size = if timeline.tenant_shard_id.is_shard_zero() {
593 0 : timeline
594 0 : .get_current_logical_size(
595 0 : crate::tenant::timeline::GetLogicalSizePriority::User,
596 0 : &ctx,
597 0 : )
598 0 : // FIXME: https://github.com/neondatabase/neon/issues/5963
599 0 : .size_dont_care_about_accuracy()
600 : } else {
601 : // Non-zero shards send zero for logical size. The safekeeper will ignore
602 : // this number. This is because in a sharded tenant, only shard zero maintains
603 : // accurate logical size.
604 0 : 0
605 : };
606 :
607 0 : let status_update = PageserverFeedback {
608 0 : current_timeline_size,
609 0 : last_received_lsn,
610 0 : disk_consistent_lsn,
611 0 : remote_consistent_lsn,
612 0 : replytime: ts,
613 0 : shard_number: timeline.tenant_shard_id.shard_number.0 as u32,
614 0 : };
615 0 :
616 0 : debug!("neon_status_update {status_update:?}");
617 :
618 0 : let mut data = BytesMut::new();
619 0 : status_update.serialize(&mut data);
620 0 : physical_stream
621 0 : .as_mut()
622 0 : .zenith_status_update(data.len() as u64, &data)
623 0 : .await?;
624 0 : }
625 : }
626 :
627 0 : Ok(())
628 0 : }
629 :
630 : /// Data returned from the postgres `IDENTIFY_SYSTEM` command
631 : ///
632 : /// See the [postgres docs] for more details.
633 : ///
634 : /// [postgres docs]: https://www.postgresql.org/docs/current/protocol-replication.html
635 : #[derive(Debug)]
636 : // As of nightly 2021-09-11, fields that are only read by the type's `Debug` impl still count as
637 : // unused. Relevant issue: https://github.com/rust-lang/rust/issues/88900
638 : #[allow(dead_code)]
639 : struct IdentifySystem {
640 : systemid: u64,
641 : timeline: u32,
642 : xlogpos: PgLsn,
643 : dbname: Option<String>,
644 : }
645 :
646 : /// There was a problem parsing the response to
647 : /// a postgres IDENTIFY_SYSTEM command.
648 0 : #[derive(Debug, thiserror::Error)]
649 : #[error("IDENTIFY_SYSTEM parse error")]
650 : struct IdentifyError;
651 :
652 : /// Run the postgres `IDENTIFY_SYSTEM` command
653 0 : async fn identify_system(client: &Client) -> anyhow::Result<IdentifySystem> {
654 0 : let query_str = "IDENTIFY_SYSTEM";
655 0 : let response = client.simple_query(query_str).await?;
656 :
657 : // get(N) from row, then parse it as some destination type.
658 0 : fn get_parse<T>(row: &SimpleQueryRow, idx: usize) -> Result<T, IdentifyError>
659 0 : where
660 0 : T: FromStr,
661 0 : {
662 0 : let val = row.get(idx).ok_or(IdentifyError)?;
663 0 : val.parse::<T>().or(Err(IdentifyError))
664 0 : }
665 :
666 : // extract the row contents into an IdentifySystem struct.
667 : // written as a closure so I can use ? for Option here.
668 0 : if let Some(SimpleQueryMessage::Row(first_row)) = response.first() {
669 : Ok(IdentifySystem {
670 0 : systemid: get_parse(first_row, 0)?,
671 0 : timeline: get_parse(first_row, 1)?,
672 0 : xlogpos: get_parse(first_row, 2)?,
673 0 : dbname: get_parse(first_row, 3).ok(),
674 : })
675 : } else {
676 0 : Err(IdentifyError.into())
677 : }
678 0 : }
|
