Re: [HACKERS] SERIALIZABLE on standby servers

On Sat, Sep 22, 2018 at 12:28 AM Thomas Munro
<thomas(dot)munro(at)enterprisedb(dot)com> wrote:
> I'll add it to the next
> Commitfest so I know when to rebase it.

And cfbot immediately showed that this assertion in
OldSerXidSetActiveSerXmin() could fail in the isolation tests:

Assert(!TransactionIdIsValid(oldSerXidControl->tailXid)
|| TransactionIdFollows(xid, oldSerXidControl->tailXid));

Not sure how that ever worked or if I screwed something up while
rebasing, but the quick (and possibly wrong?) solution I found was to
exclude hypothetical SERIALIABLEXACTs when scanning for the new oldest
xid:

@@ -3181,6 +3322,7 @@ SetNewSxactGlobalXmin(void)
for (sxact = FirstPredXact(); sxact != NULL; sxact =
NextPredXact(sxact))
{
if (!SxactIsRolledBack(sxact)
+ && !SxactIsHypothetical(sxact)
&& !SxactIsCommitted(sxact)
&& sxact != OldCommittedSxact)

Here's a version like that in the meantime so that the CI passes. The
real solution might be to give them their own xid (right now they
"borrow" one, see PreCommit_CheckForSerializationFailure()... now that
I think about it, that must be wrong), when I have more time for this
project.

--
Thomas Munro
http://www.enterprisedb.com