Fix truncateLsn update (#101)

truncateLsn is now advanced to `Min(walkeeper[i].feedback.flushLsn)` with taking epochs into account.

Author: petuhovskiy

src/backend/replication/walproposer.c

@@ -81,7 +81,6 @@ static AppendResponse lastFeedback;
  *  record-aligned (first record which might not yet received by someone).
  */
 static XLogRecPtr truncateLsn;
-static XLogRecPtr candidateTruncateLsn;
 static VoteRequest voteRequest; /* Vote request for walkeeper */
 static term_t propTerm;			/* term of the proposer */
 static XLogRecPtr propEpochStartLsn;	/* epoch start lsn of the proposer */
@@ -150,6 +149,26 @@ CalculateDiskConsistentLsn(void)
 	return lsn;
 }
 
+/*
+ * Get minimum of flushed LSNs of all safekeepers, which is the LSN of the
+ * last WAL record that can be safely discarded.
+ */
+static XLogRecPtr
+CalculateMinFlushLsn(void)
+{
+	XLogRecPtr lsn = UnknownXLogRecPtr;
+	for (int i = 0; i < n_walkeepers; i++)
+	{
+		/* We can't rely on safekeeper flushLsn if it has wrong epoch */
+		if (walkeeper[i].feedback.epoch != propTerm)
+			return 0;
+
+		if (walkeeper[i].feedback.flushLsn < lsn)
+			lsn = walkeeper[i].feedback.flushLsn;
+	}
+	return lsn;
+}
+
 /* Initializes the internal event set, provided that it is currently null */
 static void
 InitEventSet(void)
@@ -357,7 +376,7 @@ HandleWalKeeperResponse(void)
 	HotStandbyFeedback hsFeedback;
 	XLogRecPtr	minQuorumLsn;
 	XLogRecPtr	diskConsistentLsn;
-	WalMessage *msgQueueAck;
+	XLogRecPtr  minFlushLsn;
 
 	minQuorumLsn = GetAcknowledgedByQuorumWALPosition();
 	diskConsistentLsn = CalculateDiskConsistentLsn();
@@ -389,39 +408,24 @@ HandleWalKeeperResponse(void)
 									 EpochFromFullTransactionId(hsFeedback.catalog_xmin));
 	}
 
-	/* Advance truncateLsn */
-	msgQueueAck = msgQueueHead;
-	while (msgQueueAck != NULL && msgQueueAck->ackMask == ((1 << n_walkeepers) - 1))
-	{
-		/*
-		 * This piece is received by everyone; try to advance truncateLsn, but
-		 * hold it back to nearest commitLsn. Thus we will always start
-		 * streaming from the beginning of the record, which simplifies
-		 * decoding on the far end.
-		 *
-		 * This also prevents surprising violation of truncateLsn <= commitLsn
-		 * invariant which might occur because 1) truncateLsn can be advanced
-		 * immediately once chunk is broadcast to all safekeepers, and
-		 * commitLsn generally can't be advanced based on feedback from
-		 * safekeeper who is still in the previous epoch (similar to 'leader
-		 * can't commit entries from previous term' in Raft); 2) chunks we
-		 * read from WAL and send are plain sheets of bytes, but safekeepers
-		 * ack only on commit boundaries.
-		 */
-		if (msgQueueAck->req.endLsn >= minQuorumLsn && minQuorumLsn != InvalidXLogRecPtr)
-		{
-			truncateLsn = minQuorumLsn;
-			candidateTruncateLsn = InvalidXLogRecPtr;
-		}
-		else if (msgQueueAck->req.endLsn >= candidateTruncateLsn &&
-				 candidateTruncateLsn != InvalidXLogRecPtr)
-		{
-			truncateLsn = candidateTruncateLsn;
-			candidateTruncateLsn = InvalidXLogRecPtr;
-		}
-
-		msgQueueAck = msgQueueAck->next;
-	}
+	/*
+	 * Try to advance truncateLsn to minFlushLsn, which is the last record
+	 * flushed to all safekeepers. We must always start streaming from the 
+	 * beginning of the record, which simplifies decoding on the far end.
+	 *
+	 * Advanced truncateLsn should be not further than nearest commitLsn.
+	 * This prevents surprising violation of truncateLsn <= commitLsn
+	 * invariant which might occur because 1) truncateLsn can be advanced
+	 * immediately once chunk is broadcast to all safekeepers, and
+	 * commitLsn generally can't be advanced based on feedback from
+	 * safekeeper who is still in the previous epoch (similar to 'leader
+	 * can't commit entries from previous term' in Raft); 2) chunks we
+	 * read from WAL and send are plain sheets of bytes, but safekeepers
+	 * ack only on record boundaries.
+	 */
+	minFlushLsn = CalculateMinFlushLsn();
+	if (minFlushLsn > truncateLsn)
+		truncateLsn = minFlushLsn;
 
 	/* Cleanup message queue up to truncateLsn, but only messages received by everyone */
 	while (msgQueueHead != NULL && msgQueueHead->ackMask == ((1 << n_walkeepers) - 1) && msgQueueHead->req.endLsn <= truncateLsn)
@@ -1604,17 +1608,6 @@ AdvancePollState(int i, uint32 events)
 					if (minQuorumLsn > lastSentCommitLsn)
 					{
 						BroadcastMessage(CreateMessageCommitLsnOnly(lastSentLsn));
-
-						/*
-						 * commitLsn is always the record boundary; remember
-						 * it so we can advance truncateLsn there. But do so
-						 * only if previous value is applied, otherwise it
-						 * might never catch up.
-						 */
-						if (candidateTruncateLsn == InvalidXLogRecPtr)
-						{
-							candidateTruncateLsn = minQuorumLsn;
-						}
 						lastSentCommitLsn = minQuorumLsn;
 					}
 					break;