On Tue, Nov 29, 2022 at 8:03 AM Chris Travers <chris@orioledata.com> wrote:
> To be clear, I never suggested shutting down the database. What I have suggested is that repurposing the current
approaching-xid-wraparoundwarnings to start complaining loudly when a threshold is exceeded would be helpful. I think
itmakes sense to make that threshold configurable especially if we eventually have people running bloat-free table
structures.
>
> There are two fundamental problems here. The first is that if, as you say, a table is progressively bloating and we
aregetting further and further behind on vacuuming and freezing, something is seriously wrong and we need to do
somethingabout it. In these cases, I my experience is that vacuuming and related tools tend to suffer degraded
performance,and determining how to solve the problem takes quite a bit more time than a routine bloat issue would. So
whatI am arguing against is treating the problem just as a bloat issue. If you get there due to vacuum being slow,
somethingelse is wrong and you are probably going to have to find and fix that as well in order to catch up. At least
that'smy experience.
>
> I don't object to the db continuing to run, allocate xids etc. What I object to is it doing so in silently where
thingsare almost certainly going very wrong.
OK. My feeling is that the set of things we can do to warn the user is
somewhat limited. I'm open to trying our best, but we need to have
reasonable expectations. Sophisticated users will be monitoring for
problems even if we do nothing to warn, and dumb ones won't look at
their logs. Any feature that proposes to warn must aim at the uses who
are smart enough to check the logs but dumb enough not to have any
more sophisticated monitoring. Such users certainly exist and are not
even uncommon, but they aren't the only kind by a long shot.
My argument is that removing xidStopLimit is totally fine, because it
only serves to stop the database. What to do about xidWarnLimit is a
slightly more complex question. Certainly it can't be left untouched,
because warning that we're about to shut down the database for lack of
allocatable XIDs is not sensible if there is no such lack and we
aren't going to shut it down. But I'm also not sure if the model is
right. Doing nothing for a long time and then warning in every
transaction when some threshold is crossed is an extreme behavior
change. Right now that's somewhat justified because we're about to hit
a brick wall at full speed, but if we remove the brick wall and
replace it with a gentle pelting with rotten eggs, it's unclear that a
similarly strenuous reaction is the right model. But that's also not
to say that we should do nothing at all.
--
Robert Haas
EDB: http://www.enterprisedb.com