Hi,
On 2020-10-28 21:00:30 -0700, Andres Freund wrote:
> On 2020-10-28 19:09:14 -0700, Andres Freund wrote:
> > On 2020-10-28 18:13:44 -0700, Andres Freund wrote:
> > > Just pushed this. Let's see what the BF says...
> >
> > It says that apparently something is unstable about my new test. It
> > first passed on a few animals, but then failed a lot in a row. Looking.
>
> The differentiating factor is force_parallel_mode=regress.
>
> Ugh, this is nasty: The problem is that we can end up computing the
> horizons the first time before MyDatabaseId is even set. Which leads us
> to compute a too aggressive horizon for plain tables, because we skip
> over them, as MyDatabaseId still is InvalidOid:
>
> /*
> * Normally queries in other databases are ignored for anything but
> * the shared horizon. But in recovery we cannot compute an accurate
> * per-database horizon as all xids are managed via the
> * KnownAssignedXids machinery.
> */
> if (in_recovery ||
> proc->databaseId == MyDatabaseId ||
> proc->databaseId == 0) /* always include WalSender */
> h->data_oldest_nonremovable =
> TransactionIdOlder(h->data_oldest_nonremovable, xmin);
>
> That then subsequently leads us consider a row fully dead in
> heap_hot_search_buffers(). Triggering the killtuples logic. Causing the
> test to fail.
>
> With force_parallel_mode=regress we constantly start parallel workers,
> which makes it much more likely that this case is hit.
>
> It's trivial to fix luckily...
Pushed that fix, hopefully that calms the BF.
Greetings,
Andres Freund