Re: Race conditions in logical decoding

On 2026-Jan-23, Antonin Houska wrote:

> > This way you wait repeatedly for one transaction until it is marked
> > committed; and once it does, you don't test it again.
>
> Sure, that's much beter. Thanks.

Actually, I wonder if it would make sense to sleep just once after
testing all the transactions for whether they are marked committed (not
once per transaction); and after sleeping, we only test again those that
were not marked committed in the previous iteration. I think you would
end up doing less tests overall. Something like this

/*
* Although it's very unlikely, it's possible that a commit WAL record was
* decoded but CLOG is not aware of the commit yet. Should the CLOG update
* be delayed even more, visibility checks that use this snapshot could
* work incorrectly. Therefore we check the CLOG status here.
*/
{
TransactionId *stillrunning;
int nstillrunning = builder->committed.xcnt;

stillrunning = palloc(sizeof(TransactionId) * builder->committed.xcnt);
nstillrunning = builder->committed.xcnt;
memcpy(stillrunning, builder->committed.xip, sizeof(TransactionId) * nstillrunning);

for (;;)
{
int next = 0;

if (nstillrunning == 0)
break;
for (int i = 0; i < nstillrunning; i++)
{
if (!TransactionIdDidCommit(stillrunning[i]))
stillrunning[next++] = stillrunning[i];
}
if (next == 0)
break;
nstillrunning = next;
(void) WaitLatch(MyLatch,
WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
10L,
WAIT_EVENT_SNAPBUILD_CLOG);
ResetLatch(MyLatch);
CHECK_FOR_INTERRUPTS();
}
pfree(stillrunning);
}

--
Álvaro Herrera Breisgau, Deutschland — https://www.EnterpriseDB.com/
"Use it up, wear it out, make it do, or do without"