From 799ede2f30453aaa0ed379735125a4ab008f383e Mon Sep 17 00:00:00 2001 From: Ajin Cherian Date: Wed, 16 Mar 2022 01:09:05 -0400 Subject: [PATCH v26] Skip empty transactions for logical replication. The current logical replication behavior is to send every transaction to subscriber even if the transaction is empty (because it does not contain changes from the selected publications). It is a waste of CPU cycles and network bandwidth to build/transmit these empty transactions. This patch addresses the above problem by postponing the BEGIN message until the first change. While processing a COMMIT message, if there was no other change for that transaction, do not send the COMMIT message. This means that pgoutput will skip BEGIN/COMMIT messages for transactions that are empty. The patch also makes sure that in synchronous replication mode, when skipping empty transactions, keepalive messages are sent to keep the LSN locations updated on the standby. This patch does not skip empty transactions that are "streaming" or "two-phase". Discussion: https://postgr.es/m/CAMkU=1yohp9-dv48FLoSPrMqYEyyS5ZWkaZGD41RJr10xiNo_Q@mail.gmail.com --- src/backend/replication/logical/logical.c | 4 +- src/backend/replication/pgoutput/pgoutput.c | 142 +++++++++++++++++++++++++--- src/backend/replication/syncrep.c | 12 ++- src/backend/replication/walsender.c | 31 ++++-- src/include/replication/logical.h | 3 +- src/include/replication/output_plugin.h | 2 +- src/include/replication/syncrep.h | 1 + src/test/subscription/t/001_rep_changes.pl | 10 ++ src/test/subscription/t/020_messages.pl | 5 +- src/tools/pgindent/typedefs.list | 1 + 10 files changed, 179 insertions(+), 32 deletions(-) diff --git a/src/backend/replication/logical/logical.c b/src/backend/replication/logical/logical.c index 934aa13..99b2775 100644 --- a/src/backend/replication/logical/logical.c +++ b/src/backend/replication/logical/logical.c @@ -683,12 +683,12 @@ OutputPluginWrite(struct LogicalDecodingContext *ctx, bool last_write) * Update progress tracking (if supported). */ void -OutputPluginUpdateProgress(struct LogicalDecodingContext *ctx) +OutputPluginUpdateProgress(struct LogicalDecodingContext *ctx, bool send_keepalive) { if (!ctx->update_progress) return; - ctx->update_progress(ctx, ctx->write_location, ctx->write_xid); + ctx->update_progress(ctx, ctx->write_location, ctx->write_xid, send_keepalive); } /* diff --git a/src/backend/replication/pgoutput/pgoutput.c b/src/backend/replication/pgoutput/pgoutput.c index ea57a04..be2f438 100644 --- a/src/backend/replication/pgoutput/pgoutput.c +++ b/src/backend/replication/pgoutput/pgoutput.c @@ -166,6 +166,16 @@ typedef struct RelationSyncEntry AttrMap *attrmap; } RelationSyncEntry; +/* + * Maintain a per-transaction level variable to track whether the transaction + * has sent BEGIN. BEGIN is only sent when the first change in a transaction + * is processed. This makes it possible to skip transactions that are empty. + */ +typedef struct PGOutputTxnData { + bool sent_begin_txn; /* flag indicating whether BEGIN has + * been sent */ +} PGOutputTxnData; + /* Map used to remember which relation schemas we sent. */ static HTAB *RelationSyncCache = NULL; @@ -452,18 +462,45 @@ pgoutput_startup(LogicalDecodingContext *ctx, OutputPluginOptions *opt, } /* - * BEGIN callback + * BEGIN callback. + * + * Don't send BEGIN message here. Instead, postpone it until the first + * change. In logical replication, a common scenario is to replicate a set of + * tables (instead of all tables) and transactions whose changes were on + * table(s) that are not published will produce empty transactions. These + * empty transactions will send BEGIN and COMMIT messages to subscribers, + * using bandwidth on something with little/no use for logical replication. + */ +static void +pgoutput_begin_txn(LogicalDecodingContext * ctx, ReorderBufferTXN * txn) +{ + PGOutputTxnData *txndata = MemoryContextAllocZero(ctx->context, + sizeof(PGOutputTxnData)); + + txn->output_plugin_private = txndata; +} + +/* + * Send BEGIN. + * + * This is where the BEGIN is actually sent. This is called while processing + * the first change of the transaction. */ static void -pgoutput_begin_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn) +pgoutput_send_begin(LogicalDecodingContext * ctx, ReorderBufferTXN * txn) { bool send_replication_origin = txn->origin_id != InvalidRepOriginId; + PGOutputTxnData *txndata = (PGOutputTxnData *) txn->output_plugin_private; + + Assert(txndata); + Assert(!txndata->sent_begin_txn); OutputPluginPrepareWrite(ctx, !send_replication_origin); logicalrep_write_begin(ctx->out, txn); + txndata->sent_begin_txn = true; send_repl_origin(ctx, txn->origin_id, txn->origin_lsn, - send_replication_origin); + send_replication_origin); OutputPluginWrite(ctx, true); } @@ -472,10 +509,28 @@ pgoutput_begin_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn) * COMMIT callback */ static void -pgoutput_commit_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, - XLogRecPtr commit_lsn) +pgoutput_commit_txn(LogicalDecodingContext * ctx, ReorderBufferTXN * txn, + XLogRecPtr commit_lsn) { - OutputPluginUpdateProgress(ctx); + PGOutputTxnData *txndata = (PGOutputTxnData *) txn->output_plugin_private; + bool sent_begin_txn; + + Assert(txndata); + + /* + * If a BEGIN message was not yet sent, then it means there were no + * relevant changes encountered, so we can skip the COMMIT message + * too. + */ + sent_begin_txn = txndata->sent_begin_txn; + txn->output_plugin_private = NULL; + OutputPluginUpdateProgress(ctx, !sent_begin_txn); + + pfree(txndata); + if (!sent_begin_txn) { + elog(DEBUG1, "Skipping replication of an empty transaction in commit"); + return; + } OutputPluginPrepareWrite(ctx, true); logicalrep_write_commit(ctx->out, txn, commit_lsn); @@ -484,9 +539,22 @@ pgoutput_commit_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, /* * BEGIN PREPARE callback + * + * We don't postpone sending BEGIN PREPARE till the first change in order to + * skip empty prepared transactions like we do for regular transactions. This + * is because there could be a restart of walsender after a prepared + * transaction is sent out. After the restart, when the corresponding COMMIT + * PREPARED is decoded, we will not know if the prepared transaction was + * skipped because it was empty. This is because we would have lost the + * in-memory txndata information that was present prior to the restart. This + * might result in sending a spurious COMMIT PREPARED without a correspodning + * prepared transaction at the subscriber. + * + * This could be solved with a protocol change that allows the apply worker + * to detect spurious COMMIT PREPARED commands and drop them. */ static void -pgoutput_begin_prepare_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn) +pgoutput_begin_prepare_txn(LogicalDecodingContext * ctx, ReorderBufferTXN * txn) { bool send_replication_origin = txn->origin_id != InvalidRepOriginId; @@ -494,7 +562,7 @@ pgoutput_begin_prepare_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn) logicalrep_write_begin_prepare(ctx->out, txn); send_repl_origin(ctx, txn->origin_id, txn->origin_lsn, - send_replication_origin); + send_replication_origin); OutputPluginWrite(ctx, true); } @@ -506,7 +574,7 @@ static void pgoutput_prepare_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, XLogRecPtr prepare_lsn) { - OutputPluginUpdateProgress(ctx); + OutputPluginUpdateProgress(ctx, false); OutputPluginPrepareWrite(ctx, true); logicalrep_write_prepare(ctx->out, txn, prepare_lsn); @@ -520,7 +588,7 @@ static void pgoutput_commit_prepared_txn(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, XLogRecPtr commit_lsn) { - OutputPluginUpdateProgress(ctx); + OutputPluginUpdateProgress(ctx, false); OutputPluginPrepareWrite(ctx, true); logicalrep_write_commit_prepared(ctx->out, txn, commit_lsn); @@ -536,7 +604,7 @@ pgoutput_rollback_prepared_txn(LogicalDecodingContext *ctx, XLogRecPtr prepare_end_lsn, TimestampTz prepare_time) { - OutputPluginUpdateProgress(ctx); + OutputPluginUpdateProgress(ctx, false); OutputPluginPrepareWrite(ctx, true); logicalrep_write_rollback_prepared(ctx->out, txn, prepare_end_lsn, @@ -1141,6 +1209,7 @@ pgoutput_change(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, Relation relation, ReorderBufferChange *change) { PGOutputData *data = (PGOutputData *) ctx->output_plugin_private; + PGOutputTxnData *txndata = (PGOutputTxnData *) txn->output_plugin_private; MemoryContext old; RelationSyncEntry *relentry; TransactionId xid = InvalidTransactionId; @@ -1216,6 +1285,16 @@ pgoutput_change(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, &action)) break; + /* + * Send BEGIN if we haven't yet. + * + * The reason this is delayed till after the row_filter check + * is to make sure we don't send a begin if the change is not + * send out due to the row_filter check failing. + */ + if (txndata && !txndata->sent_begin_txn) + pgoutput_send_begin(ctx, txn); + /* * Schema should be sent using the original relation because it * also sends the ancestor's relation. @@ -1266,6 +1345,10 @@ pgoutput_change(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, relentry, &action)) break; + /* Send BEGIN if we haven't yet */ + if (txndata && !txndata->sent_begin_txn) + pgoutput_send_begin(ctx, txn); + maybe_send_schema(ctx, change, relation, relentry); OutputPluginPrepareWrite(ctx, true); @@ -1324,6 +1407,10 @@ pgoutput_change(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, relentry, &action)) break; + /* Send BEGIN if we haven't yet */ + if (txndata && !txndata->sent_begin_txn) + pgoutput_send_begin(ctx, txn); + maybe_send_schema(ctx, change, relation, relentry); OutputPluginPrepareWrite(ctx, true); @@ -1354,6 +1441,7 @@ pgoutput_truncate(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, int nrelations, Relation relations[], ReorderBufferChange *change) { PGOutputData *data = (PGOutputData *) ctx->output_plugin_private; + PGOutputTxnData *txndata; MemoryContext old; RelationSyncEntry *relentry; int i; @@ -1366,6 +1454,7 @@ pgoutput_truncate(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, xid = change->txn->xid; old = MemoryContextSwitchTo(data->context); + txndata = (PGOutputTxnData *) txn->output_plugin_private; relids = palloc0(nrelations * sizeof(Oid)); nrelids = 0; @@ -1392,11 +1481,21 @@ pgoutput_truncate(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, continue; relids[nrelids++] = relid; + + /* Send BEGIN if we haven't yet */ + if (txndata && !txndata->sent_begin_txn) + pgoutput_send_begin(ctx, txn); maybe_send_schema(ctx, change, relation, relentry); } if (nrelids > 0) { + txndata = (PGOutputTxnData *) txn->output_plugin_private; + + /* Send BEGIN if we haven't yet */ + if (txndata && !txndata->sent_begin_txn) + pgoutput_send_begin(ctx, txn); + OutputPluginPrepareWrite(ctx, true); logicalrep_write_truncate(ctx->out, xid, @@ -1429,6 +1528,18 @@ pgoutput_message(LogicalDecodingContext *ctx, ReorderBufferTXN *txn, if (in_streaming) xid = txn->xid; + /* + * Output BEGIN if we haven't yet. Avoid for non-transactional + * messages. + */ + if (transactional) { + PGOutputTxnData *txndata = (PGOutputTxnData *) txn->output_plugin_private; + + /* Send BEGIN if we haven't yet */ + if (txndata && !txndata->sent_begin_txn) + pgoutput_send_begin(ctx, txn); + } + OutputPluginPrepareWrite(ctx, true); logicalrep_write_message(ctx->out, xid, @@ -1506,6 +1617,11 @@ publication_invalidation_cb(Datum arg, int cacheid, uint32 hashvalue) /* * START STREAM callback + * + * We don't try and skip empty streams because streams might + * contain prepared transactions. Read the comments in the header of + * pgoutput_begin_prepare_txn() for details on why empty prepared + * transactions are not skipped. */ static void pgoutput_stream_start(struct LogicalDecodingContext *ctx, @@ -1598,7 +1714,7 @@ pgoutput_stream_commit(struct LogicalDecodingContext *ctx, Assert(!in_streaming); Assert(rbtxn_is_streamed(txn)); - OutputPluginUpdateProgress(ctx); + OutputPluginUpdateProgress(ctx, false); OutputPluginPrepareWrite(ctx, true); logicalrep_write_stream_commit(ctx->out, txn, commit_lsn); @@ -1619,7 +1735,7 @@ pgoutput_stream_prepare_txn(LogicalDecodingContext *ctx, { Assert(rbtxn_is_streamed(txn)); - OutputPluginUpdateProgress(ctx); + OutputPluginUpdateProgress(ctx, false); OutputPluginPrepareWrite(ctx, true); logicalrep_write_stream_prepare(ctx->out, txn, prepare_lsn); OutputPluginWrite(ctx, true); diff --git a/src/backend/replication/syncrep.c b/src/backend/replication/syncrep.c index ce163b9..11f7358 100644 --- a/src/backend/replication/syncrep.c +++ b/src/backend/replication/syncrep.c @@ -171,8 +171,7 @@ SyncRepWaitForLSN(XLogRecPtr lsn, bool commit) * described in SyncRepUpdateSyncStandbysDefined(). On the other hand, if * it's false, the lock is not necessary because we don't touch the queue. */ - if (!SyncRepRequested() || - !((volatile WalSndCtlData *) WalSndCtl)->sync_standbys_defined) + if (!SyncRepEnabled()) return; /* Cap the level for anything other than commit to remote flush only. */ @@ -539,6 +538,15 @@ SyncRepReleaseWaiters(void) } /* + * Check if synchronous replication is enabled + */ +bool +SyncRepEnabled(void) +{ + return SyncRepRequested() && ((volatile WalSndCtlData *) WalSndCtl)->sync_standbys_defined; +} + +/* * Calculate the synced Write, Flush and Apply positions among sync standbys. * * Return false if the number of sync standbys is less than diff --git a/src/backend/replication/walsender.c b/src/backend/replication/walsender.c index 2d0292a..21e7ac3 100644 --- a/src/backend/replication/walsender.c +++ b/src/backend/replication/walsender.c @@ -242,14 +242,15 @@ static void ProcessStandbyMessage(void); static void ProcessStandbyReplyMessage(void); static void ProcessStandbyHSFeedbackMessage(void); static void ProcessRepliesIfAny(void); -static void WalSndKeepalive(bool requestReply); +static void WalSndKeepalive(bool requestReply, XLogRecPtr writePtr); static void WalSndKeepaliveIfNecessary(void); static void WalSndCheckTimeOut(void); static long WalSndComputeSleeptime(TimestampTz now); static void WalSndWait(uint32 socket_events, long timeout, uint32 wait_event); static void WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write); static void WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write); -static void WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid); +static void WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, + bool send_keepalive); static XLogRecPtr WalSndWaitForWal(XLogRecPtr loc); static void LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time); static TimeOffset LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now); @@ -1449,12 +1450,20 @@ WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, * Write the current position to the lag tracker (see XLogSendPhysical). */ static void -WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid) +WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, + bool send_keepalive) { static TimestampTz sendTime = 0; TimestampTz now = GetCurrentTimestamp(); /* + * When skipping empty transactions in synchronous replication, we need + * to send a keepalive to keep the MyWalSnd locations updated. + */ + if (send_keepalive && SyncRepEnabled()) + WalSndKeepalive(true, ctx->write_location); + + /* * Track lag no more than once per WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS to * avoid flooding the lag tracker when we commit frequently. */ @@ -1550,7 +1559,7 @@ WalSndWaitForWal(XLogRecPtr loc) if (MyWalSnd->flush < sentPtr && MyWalSnd->write < sentPtr && !waiting_for_ping_response) - WalSndKeepalive(false); + WalSndKeepalive(false, InvalidXLogRecPtr); /* check whether we're done */ if (loc <= RecentFlushPtr) @@ -2068,7 +2077,7 @@ ProcessStandbyReplyMessage(void) /* Send a reply if the standby requested one. */ if (replyRequested) - WalSndKeepalive(false); + WalSndKeepalive(false, InvalidXLogRecPtr); /* * Update shared state for this WalSender process based on reply data from @@ -3074,7 +3083,7 @@ WalSndDone(WalSndSendDataCallback send_data) proc_exit(0); } if (!waiting_for_ping_response) - WalSndKeepalive(true); + WalSndKeepalive(true, InvalidXLogRecPtr); } /* @@ -3563,18 +3572,20 @@ pg_stat_get_wal_senders(PG_FUNCTION_ARGS) * * If requestReply is set, the message requests the other party to send * a message back to us, for heartbeat purposes. We also set a flag to - * let nearby code that we're waiting for that response, to avoid + * let nearby code know that we're waiting for that response, to avoid * repeated requests. + * + * If writePtr is set, mark that as the LSN processed, else use sentPtr. */ static void -WalSndKeepalive(bool requestReply) +WalSndKeepalive(bool requestReply, XLogRecPtr writePtr) { elog(DEBUG2, "sending replication keepalive"); /* construct the message... */ resetStringInfo(&output_message); pq_sendbyte(&output_message, 'k'); - pq_sendint64(&output_message, sentPtr); + pq_sendint64(&output_message, XLogRecPtrIsInvalid(writePtr) ? sentPtr : writePtr); pq_sendint64(&output_message, GetCurrentTimestamp()); pq_sendbyte(&output_message, requestReply ? 1 : 0); @@ -3613,7 +3624,7 @@ WalSndKeepaliveIfNecessary(void) wal_sender_timeout / 2); if (last_processing >= ping_time) { - WalSndKeepalive(true); + WalSndKeepalive(true, InvalidXLogRecPtr); /* Try to flush pending output to the client */ if (pq_flush_if_writable() != 0) diff --git a/src/include/replication/logical.h b/src/include/replication/logical.h index 1097cc9..9f59855 100644 --- a/src/include/replication/logical.h +++ b/src/include/replication/logical.h @@ -26,7 +26,8 @@ typedef LogicalOutputPluginWriterWrite LogicalOutputPluginWriterPrepareWrite; typedef void (*LogicalOutputPluginWriterUpdateProgress) (struct LogicalDecodingContext *lr, XLogRecPtr Ptr, - TransactionId xid + TransactionId xid, + bool send_keepalive ); typedef struct LogicalDecodingContext diff --git a/src/include/replication/output_plugin.h b/src/include/replication/output_plugin.h index a16bebf..eb91d17 100644 --- a/src/include/replication/output_plugin.h +++ b/src/include/replication/output_plugin.h @@ -270,6 +270,6 @@ typedef struct OutputPluginCallbacks /* Functions in replication/logical/logical.c */ extern void OutputPluginPrepareWrite(struct LogicalDecodingContext *ctx, bool last_write); extern void OutputPluginWrite(struct LogicalDecodingContext *ctx, bool last_write); -extern void OutputPluginUpdateProgress(struct LogicalDecodingContext *ctx); +extern void OutputPluginUpdateProgress(struct LogicalDecodingContext *ctx, bool send_keepalive); #endif /* OUTPUT_PLUGIN_H */ diff --git a/src/include/replication/syncrep.h b/src/include/replication/syncrep.h index 27be230..7086532 100644 --- a/src/include/replication/syncrep.h +++ b/src/include/replication/syncrep.h @@ -90,6 +90,7 @@ extern void SyncRepCleanupAtProcExit(void); /* called by wal sender */ extern void SyncRepInitConfig(void); extern void SyncRepReleaseWaiters(void); +extern bool SyncRepEnabled(void); /* called by wal sender and user backend */ extern int SyncRepGetCandidateStandbys(SyncRepStandbyData **standbys); diff --git a/src/test/subscription/t/001_rep_changes.pl b/src/test/subscription/t/001_rep_changes.pl index eca1c63..40c510b 100644 --- a/src/test/subscription/t/001_rep_changes.pl +++ b/src/test/subscription/t/001_rep_changes.pl @@ -473,6 +473,16 @@ $node_publisher->safe_psql('postgres', "INSERT INTO tab_full VALUES(0)"); $node_publisher->wait_for_catchup('tap_sub'); +#create a new table which is not published, any transaction on this table +#should be skipped. +$node_publisher->safe_psql('postgres', + "CREATE TABLE tab_empty (a int)"); +$node_publisher->safe_psql('postgres', + "INSERT INTO tab_empty SELECT generate_series(1,50)"); + +$node_publisher->wait_for_catchup('tap_sub'); + +$node_publisher->wait_for_catchup('tap_sub'); # note that data are different on provider and subscriber $result = $node_subscriber->safe_psql('postgres', "SELECT count(*), min(a), max(a) FROM tab_ins"); diff --git a/src/test/subscription/t/020_messages.pl b/src/test/subscription/t/020_messages.pl index b5045ff..d21d929 100644 --- a/src/test/subscription/t/020_messages.pl +++ b/src/test/subscription/t/020_messages.pl @@ -87,9 +87,8 @@ $result = $node_publisher->safe_psql( 'publication_names', 'tap_pub') )); -# 66 67 == B C == BEGIN COMMIT -is( $result, qq(66 -67), +# no message and no BEGIN and COMMIT because of empty transaction optimization +is($result, qq(), 'option messages defaults to false so message (M) is not available on slot' ); diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list index eaf3e7a..4c575f3 100644 --- a/src/tools/pgindent/typedefs.list +++ b/src/tools/pgindent/typedefs.list @@ -1608,6 +1608,7 @@ PGMessageField PGModuleMagicFunction PGNoticeHooks PGOutputData +PGOutputTxnData PGPROC PGP_CFB PGP_Context -- 1.8.3.1