Re: Hot Standby and prepared transactions

Simon Riggs wrote:
> On Thu, 2009-12-17 at 13:24 +0200, Heikki Linnakangas wrote:
> 
>> Hmm, looking at the code, I think Simon threw that baby with the
>> bathwater when he removed support for starting standby from a shutdown
>> checkpoint.
> 
> Hmm, I think that code was just for starting points only. It would not
> have been executed in normal running of the standby, so it appears the
> bug was older than that. Absence of baby now appears inconclusive.

This is the piece of code we're talking about, in xlog_redo():

--- a/src/backend/access/transam/xlog.c
+++ b/src/backend/access/transam/xlog.c
@@ -7340,41 +7340,6 @@ xlog_redo(XLogRecPtr lsn, XLogRecord *record)
 		ShmemVariableCache->oldestXid = checkPoint.oldestXid;
 		ShmemVariableCache->oldestXidDB = checkPoint.oldestXidDB;

-		/*
-		 * We know nothing was running on the master at this point, though
-		 * we can only use it as a starting point iff wal_standby_info was
-		 * enabled, otherwise we may not get further information about changes
-		 * from the master.
-		 */
-		if (standbyState >= STANDBY_UNINITIALIZED &&
checkPoint.XLogStandbyInfoMode)
-		{
-			RunningTransactionsData running;
-			TransactionId oldestRunningXid;
-			TransactionId *xids;
-			int nxids;
-
-			oldestRunningXid = PrescanPreparedTransactions(&xids, &nxids);
-
-			/*
-			 * Construct a RunningTransactions snapshot representing a shut
-			 * down server, with only prepared transactions still alive.
-			 *
-			 * StandbyRecoverPreparedTransactions will call SubTransSetParent
-			 * for any subtransactions, so we consider this a non-overflowed
-			 * snapshot.
-			 */
-			running.xcnt = nxids;
-			running.subxid_overflow = false;
-			running.nextXid = checkPoint.nextXid;
-			running.oldestRunningXid = oldestRunningXid;
-			running.xids = xids;
-
-			ProcArrayInitRecoveryInfo(oldestRunningXid);
-			ProcArrayApplyRecoveryInfo(&running);
-
-			StandbyRecoverPreparedTransactions();
-		}
-
 		/* Check to see if any changes to max_connections give problems */
 		if (standbyState != STANDBY_DISABLED)
 			CheckRequiredParameterValues(checkPoint);


That removed piece of code was executed in the standby whenever we saw a
shutdown checkpoint. It calls ProcArrayApplyRecoveryInfo(), which calls
ExpireOldKnownAssignedTransactionIds() and StandbyReleaseOldLocks() to
clean up known-assigned-xid entries and locks of the implicitly-aborted
transactions.

I see now that in the presence of prepared transactions, we would fail
to clean up failed transations with XID > the oldest prepared
transaction, but other than that it looks fine to me.

-- 
  Heikki Linnakangas
  EnterpriseDB   http://www.enterprisedb.com