Add new replication mode synchronous_commit = 'remote_apply'.

In this mode, the master waits for the transaction to be applied on
the remote side, not just written to disk.  That means that you can
count on a transaction started on the standby to see all commits
previously acknowledged by the master.

To make this work, the standby sends a reply after replaying each
commit record generated with synchronous_commit >= 'remote_apply'.
This introduces a small inefficiency: the extra replies will be sent
even by standbys that aren't the current synchronous standby.  But
previously-existing synchronous_commit levels make no attempt at all
to optimize which replies are sent based on what the primary cares
about, so this is no worse, and at least avoids any extra replies for
people not using the feature at all.

Thomas Munro, reviewed by Michael Paquier and by me.  Some additional
tweaks by me.

Author: robertmhaas

doc/src/sgml/config.sgml

@@ -2143,8 +2143,8 @@ include_dir 'conf.d'
         Specifies whether transaction commit will wait for WAL records
         to be written to disk before the command returns a <quote>success</>
         indication to the client.  Valid values are <literal>on</>,
-        <literal>remote_write</>, <literal>local</>, and <literal>off</>.
-        The default, and safe, setting
+        <literal>remote_write</>, <literal>remote_apply</>, <literal>local</>,
+        and <literal>off</>.  The default, and safe, setting
         is <literal>on</>.  When <literal>off</>, there can be a delay between
         when success is reported to the client and when the transaction is
         really guaranteed to be safe against a server crash.  (The maximum
@@ -2169,6 +2169,10 @@ include_dir 'conf.d'
         the commit record of the transaction and flushed it to disk.  This
         ensures the transaction will not be lost unless both primary and
         standby suffer corruption of their database storage.
+        When set to <literal>remote_apply</>, commits will wait until a reply
+        from the current synchronous standby indicates it has received the
+        commit record of the transaction and applied it, so that it has become
+        visible to queries.
         When set to <literal>remote_write</>, commits will wait
         until a reply from the current synchronous standby indicates it has
         received the commit record of the transaction and written it out to
@@ -2186,9 +2190,9 @@ include_dir 'conf.d'
         setting <literal>local</> is available for transactions that
         wish to wait for local flush to disk, but not synchronous replication.
         If <varname>synchronous_standby_names</> is not set, the settings
-        <literal>on</>, <literal>remote_write</> and <literal>local</> all
-        provide the same synchronization level: transaction commits only wait
-        for local flush to disk.
+        <literal>on</>, <literal>remote_apply</>, <literal>remote_write</>
+        and <literal>local</> all provide the same synchronization level:
+        transaction commits only wait for local flush to disk.
        </para>
        <para>
         This parameter can be changed at any time; the behavior for any

doc/src/sgml/high-availability.sgml

@@ -1081,6 +1081,9 @@ primary_slot_name = 'node_a_slot'
     WAL record is then sent to the standby. The standby sends reply
     messages each time a new batch of WAL data is written to disk, unless
     <varname>wal_receiver_status_interval</> is set to zero on the standby.
+    In the case that <varname>synchronous_commit</> is set to
+    <literal>remote_apply</>, the standby sends reply messages when the commit
+    record is replayed, making the transaction visible.
     If the standby is the first matching standby, as specified in
     <varname>synchronous_standby_names</> on the primary, the reply
     messages from that standby will be used to wake users waiting for
@@ -1106,6 +1109,14 @@ primary_slot_name = 'node_a_slot'
     the database of the primary gets corrupted at the same time.
    </para>
 
+   <para>
+    Setting <varname>synchronous_commit</> to <literal>remote_apply</> will
+    cause each commit to wait until the current synchronous standby reports
+    that it has replayed the transaction, making it visible to user queries.
+    In simple cases, this allows for load balancing with causal consistency
+    on a single hot standby.
+   </para>
+
    <para>
     Users will stop waiting if a fast shutdown is requested.  However, as
     when using asynchronous replication, the server will not fully
@@ -1160,9 +1171,10 @@ primary_slot_name = 'node_a_slot'
     <title>Planning for High Availability</title>
 
    <para>
-    Commits made when <varname>synchronous_commit</> is set to <literal>on</>
-    or <literal>remote_write</> will wait until the synchronous standby responds. The response
-    may never occur if the last, or only, standby should crash.
+    Commits made when <varname>synchronous_commit</> is set to <literal>on</>,
+    <literal>remote_apply</> or <literal>remote_write</> will wait until the
+    synchronous standby responds. The response may never occur if the last, or
+    only, standby should crash.
    </para>
 
    <para>

src/backend/access/transam/twophase.c

@@ -1107,7 +1107,7 @@ EndPrepare(GlobalTransaction gxact)
 	 * Note that at this stage we have marked the prepare, but still show as
 	 * running in the procarray (twice!) and continue to hold locks.
 	 */
-	SyncRepWaitForLSN(gxact->prepare_end_lsn);
+	SyncRepWaitForLSN(gxact->prepare_end_lsn, false);
 
 	records.tail = records.head = NULL;
 	records.num_chunks = 0;
@@ -2103,7 +2103,7 @@ RecordTransactionCommitPrepared(TransactionId xid,
 	 * Note that at this stage we have marked clog, but still show as running
 	 * in the procarray and continue to hold locks.
 	 */
-	SyncRepWaitForLSN(recptr);
+	SyncRepWaitForLSN(recptr, true);
 }
 
 /*
@@ -2156,5 +2156,5 @@ RecordTransactionAbortPrepared(TransactionId xid,
 	 * Note that at this stage we have marked clog, but still show as running
 	 * in the procarray and continue to hold locks.
 	 */
-	SyncRepWaitForLSN(recptr);
+	SyncRepWaitForLSN(recptr, false);
 }

src/backend/access/transam/xact.c

@@ -1324,7 +1324,7 @@ RecordTransactionCommit(void)
 	 * in the procarray and continue to hold locks.
 	 */
 	if (wrote_xlog && markXidCommitted)
-		SyncRepWaitForLSN(XactLastRecEnd);
+		SyncRepWaitForLSN(XactLastRecEnd, true);
 
 	/* remember end of last commit record */
 	XactLastCommitEnd = XactLastRecEnd;
@@ -5122,6 +5122,13 @@ XactLogCommitRecord(TimestampTz commit_time,
 	if (forceSyncCommit)
 		xl_xinfo.xinfo |= XACT_COMPLETION_FORCE_SYNC_COMMIT;
 
+	/*
+	 * Check if the caller would like to ask standbys for immediate feedback
+	 * once this commit is applied.
+	 */
+	if (synchronous_commit >= SYNCHRONOUS_COMMIT_REMOTE_APPLY)
+		xl_xinfo.xinfo |= XACT_COMPLETION_APPLY_FEEDBACK;
+
 	/*
 	 * Relcache invalidations requires information about the current database
 	 * and so does logical decoding.
@@ -5459,6 +5466,13 @@ xact_redo_commit(xl_xact_parsed_commit *parsed,
 	if (XactCompletionForceSyncCommit(parsed->xinfo))
 		XLogFlush(lsn);
 
+	/*
+	 * If asked by the primary (because someone is waiting for a synchronous
+	 * commit = remote_apply), we will need to ask walreceiver to send a
+	 * reply immediately.
+	 */
+	if (XactCompletionApplyFeedback(parsed->xinfo))
+		XLogRequestWalReceiverReply();
 }
 
 /*

src/backend/access/transam/xlog.c

@@ -345,6 +345,9 @@ static XLogRecPtr RedoRecPtr;
  */
 static bool doPageWrites;
 
+/* Has the recovery code requested a walreceiver wakeup? */
+static bool doRequestWalReceiverReply;
+
 /*
  * RedoStartLSN points to the checkpoint's REDO location which is specified
  * in a backup label file, backup history file or control file. In standby
@@ -6879,6 +6882,17 @@ StartupXLOG(void)
 				XLogCtl->lastReplayedTLI = ThisTimeLineID;
 				SpinLockRelease(&XLogCtl->info_lck);
 
+				/*
+				 * If rm_redo called XLogRequestWalReceiverReply, then we
+				 * wake up the receiver so that it notices the updated
+				 * lastReplayedEndRecPtr and sends a reply to the master.
+				 */
+				if (doRequestWalReceiverReply)
+				{
+					doRequestWalReceiverReply = false;
+					WalRcvForceReply();
+				}
+
 				/* Remember this record as the last-applied one */
 				LastRec = ReadRecPtr;
 
@@ -11594,3 +11608,12 @@ SetWalWriterSleeping(bool sleeping)
 	XLogCtl->WalWriterSleeping = sleeping;
 	SpinLockRelease(&XLogCtl->info_lck);
 }
+
+/*
+ * Schedule a walreceiver wakeup in the main recovery loop.
+ */
+void
+XLogRequestWalReceiverReply(void)
+{
+	doRequestWalReceiverReply = true;
+}

src/backend/replication/README

@@ -16,14 +16,16 @@ bool walrcv_connect(char *conninfo, XLogRecPtr startpoint)
 Establish connection to the primary, and starts streaming from 'startpoint'.
 Returns true on success.
 
-bool walrcv_receive(int timeout, unsigned char *type, char **buffer, int *len)
-
-Retrieve any message available through the connection, blocking for
-maximum of 'timeout' ms. If a message was successfully read, returns true,
-otherwise false. On success, a pointer to the message payload is stored in
-*buffer, length in *len, and the type of message received in *type. The
-returned buffer is valid until the next call to walrcv_* functions, the
-caller should not attempt freeing it.
+int walrcv_receive(char **buffer, int *wait_fd)
+
+Retrieve any message available without blocking through the
+connection.  If a message was successfully read, returns its
+length. If the connection is closed, returns -1.  Otherwise returns 0
+to indicate that no data is available, and sets *wait_fd to a file
+descriptor which can be waited on before trying again.  On success, a
+pointer to the message payload is stored in *buffer. The returned
+buffer is valid until the next call to walrcv_* functions, and the
+caller should not attempt to free it.
 
 void walrcv_send(const char *buffer, int nbytes)
 

src/backend/replication/libpqwalreceiver/libpqwalreceiver.c

@@ -52,7 +52,7 @@ static void libpqrcv_readtimelinehistoryfile(TimeLineID tli, char **filename, ch
 static bool libpqrcv_startstreaming(TimeLineID tli, XLogRecPtr startpoint,
 						char *slotname);
 static void libpqrcv_endstreaming(TimeLineID *next_tli);
-static int	libpqrcv_receive(int timeout, char **buffer);
+static int	libpqrcv_receive(char **buffer, int *wait_fd);
 static void libpqrcv_send(const char *buffer, int nbytes);
 static void libpqrcv_disconnect(void);
 
@@ -463,24 +463,23 @@ libpqrcv_disconnect(void)
 }
 
 /*
- * Receive a message available from XLOG stream, blocking for
- * maximum of 'timeout' ms.
+ * Receive a message available from XLOG stream.
  *
  * Returns:
  *
  *	 If data was received, returns the length of the data. *buffer is set to
  *	 point to a buffer holding the received message. The buffer is only valid
  *	 until the next libpqrcv_* call.
  *
- *	 0 if no data was available within timeout, or wait was interrupted
- *	 by signal.
+ *	 If no data was available immediately, returns 0, and *wait_fd is set to a
+ *	 file descriptor which can be waited on before trying again.
  *
  *	 -1 if the server ended the COPY.
  *
  * ereports on error.
  */
 static int
-libpqrcv_receive(int timeout, char **buffer)
+libpqrcv_receive(char **buffer, int *wait_fd)
 {
 	int			rawlen;
 
@@ -492,16 +491,7 @@ libpqrcv_receive(int timeout, char **buffer)
 	rawlen = PQgetCopyData(streamConn, &recvBuf, 1);
 	if (rawlen == 0)
 	{
-		/*
-		 * No data available yet. If the caller requested to block, wait for
-		 * more data to arrive.
-		 */
-		if (timeout > 0)
-		{
-			if (!libpq_select(timeout))
-				return 0;
-		}
-
+		/* Try consuming some data. */
 		if (PQconsumeInput(streamConn) == 0)
 			ereport(ERROR,
 					(errmsg("could not receive data from WAL stream: %s",
@@ -510,7 +500,11 @@ libpqrcv_receive(int timeout, char **buffer)
 		/* Now that we've consumed some input, try again */
 		rawlen = PQgetCopyData(streamConn, &recvBuf, 1);
 		if (rawlen == 0)
+		{
+			/* Tell caller to try again when our socket is ready. */
+			*wait_fd = PQsocket(streamConn);
 			return 0;
+		}
 	}
 	if (rawlen == -1)			/* end-of-streaming or error */
 	{

src/backend/replication/syncrep.c

@@ -91,13 +91,24 @@ static bool SyncRepQueueIsOrderedByLSN(int mode);
  * to the wait queue. During SyncRepWakeQueue() a WALSender changes
  * the state to SYNC_REP_WAIT_COMPLETE once replication is confirmed.
  * This backend then resets its state to SYNC_REP_NOT_WAITING.
+ *
+ * 'lsn' represents the LSN to wait for.  'commit' indicates whether this LSN
+ * represents a commit record.  If it doesn't, then we wait only for the WAL
+ * to be flushed if synchronous_commit is set to the higher level of
+ * remote_apply, because only commit records provide apply feedback.
  */
 void
-SyncRepWaitForLSN(XLogRecPtr XactCommitLSN)
+SyncRepWaitForLSN(XLogRecPtr lsn, bool commit)
 {
 	char	   *new_status = NULL;
 	const char *old_status;
-	int			mode = SyncRepWaitMode;
+	int			mode;
+
+	/* Cap the level for anything other than commit to remote flush only. */
+	if (commit)
+		mode = SyncRepWaitMode;
+	else
+		mode = Min(SyncRepWaitMode, SYNC_REP_WAIT_FLUSH);
 
 	/*
 	 * Fast exit if user has not requested sync replication, or there are no
@@ -122,7 +133,7 @@ SyncRepWaitForLSN(XLogRecPtr XactCommitLSN)
 	 * to be a low cost check.
 	 */
 	if (!WalSndCtl->sync_standbys_defined ||
-		XactCommitLSN <= WalSndCtl->lsn[mode])
+		lsn <= WalSndCtl->lsn[mode])
 	{
 		LWLockRelease(SyncRepLock);
 		return;
@@ -132,7 +143,7 @@ SyncRepWaitForLSN(XLogRecPtr XactCommitLSN)
 	 * Set our waitLSN so WALSender will know when to wake us, and add
 	 * ourselves to the queue.
 	 */
-	MyProc->waitLSN = XactCommitLSN;
+	MyProc->waitLSN = lsn;
 	MyProc->syncRepState = SYNC_REP_WAITING;
 	SyncRepQueueInsert(mode);
 	Assert(SyncRepQueueIsOrderedByLSN(mode));
@@ -147,7 +158,7 @@ SyncRepWaitForLSN(XLogRecPtr XactCommitLSN)
 		new_status = (char *) palloc(len + 32 + 1);
 		memcpy(new_status, old_status, len);
 		sprintf(new_status + len, " waiting for %X/%X",
-				(uint32) (XactCommitLSN >> 32), (uint32) XactCommitLSN);
+				(uint32) (lsn >> 32), (uint32) lsn);
 		set_ps_display(new_status, false);
 		new_status[len] = '\0'; /* truncate off " waiting ..." */
 	}
@@ -416,6 +427,7 @@ SyncRepReleaseWaiters(void)
 	WalSnd	   *syncWalSnd;
 	int			numwrite = 0;
 	int			numflush = 0;
+	int			numapply = 0;
 
 	/*
 	 * If this WALSender is serving a standby that is not on the list of
@@ -462,12 +474,18 @@ SyncRepReleaseWaiters(void)
 		walsndctl->lsn[SYNC_REP_WAIT_FLUSH] = MyWalSnd->flush;
 		numflush = SyncRepWakeQueue(false, SYNC_REP_WAIT_FLUSH);
 	}
+	if (walsndctl->lsn[SYNC_REP_WAIT_APPLY] < MyWalSnd->apply)
+	{
+		walsndctl->lsn[SYNC_REP_WAIT_APPLY] = MyWalSnd->apply;
+		numapply = SyncRepWakeQueue(false, SYNC_REP_WAIT_APPLY);
+	}
 
 	LWLockRelease(SyncRepLock);
 
-	elog(DEBUG3, "released %d procs up to write %X/%X, %d procs up to flush %X/%X",
+	elog(DEBUG3, "released %d procs up to write %X/%X, %d procs up to flush %X/%X, %d procs up to apply %X/%x",
 		 numwrite, (uint32) (MyWalSnd->write >> 32), (uint32) MyWalSnd->write,
-	   numflush, (uint32) (MyWalSnd->flush >> 32), (uint32) MyWalSnd->flush);
+		 numflush, (uint32) (MyWalSnd->flush >> 32), (uint32) MyWalSnd->flush,
+		 numapply, (uint32) (MyWalSnd->apply >> 32), (uint32) MyWalSnd->apply);
 
 	/*
 	 * If we are managing the highest priority standby, though we weren't
@@ -728,6 +746,9 @@ assign_synchronous_commit(int newval, void *extra)
 		case SYNCHRONOUS_COMMIT_REMOTE_FLUSH:
 			SyncRepWaitMode = SYNC_REP_WAIT_FLUSH;
 			break;
+		case SYNCHRONOUS_COMMIT_REMOTE_APPLY:
+			SyncRepWaitMode = SYNC_REP_WAIT_APPLY;
+			break;
 		default:
 			SyncRepWaitMode = SYNC_REP_NO_WAIT;
 			break;