On Fri, May 27, 2022 at 10:21 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> That's possible, certainly. It's also possible that it's a real bug
> that so far has only manifested there for (say) timing reasons.
> The buildfarm is not so large that we can write off single-machine
> failures as being unlikely to hit in the real world.
>
> What I'd suggest is to promote that failure to elog(PANIC), which
> would at least give us the PID and if we're lucky a stack trace.
That proposed change is fine with me.
As to the question of whether it's a real bug, nobody can prove
anything unless we actually run it down. It's just a question of what
you think the odds are. Noah's PGCon talk a few years back on the long
tail of buildfarm failures convinced me (perhaps unintentionally) that
low-probability failures that occur only on obscure systems or
configurations are likely not worth running down, because while they
COULD be real bugs, a lot of them aren't, and the time it would take
to figure it out could be spent on other things - for instance, fixing
things that we know for certain are bugs. Spending 40 hours of
person-time on something with a 10% chance of being a bug in the
PostgreSQL code doesn't necessarily make sense to me, because while
you are correct that the buildfarm isn't that large, neither is the
developer community.
--
Robert Haas
EDB: http://www.enterprisedb.com