Andres Freund <andres@anarazel.de> writes:
> That seems like it'd add a good number of new wakeups, or at least
> scheduling of wakeups.
Yes, as it stands this will result in a huge increase in alarm-scheduling
kernel call traffic. I understand the issue but I do not think this is
an acceptable path to a fix.
> Or we could do nothing - I actually think that's a viable option.
I tend to agree. I'm not really sure that the presented problem is a
big deal: for it to be an issue, you have to assume that a DML operation
that takes less than PGSTAT_STAT_INTERVAL is capable of causing enough
table churn that it's a problem if autovacuum doesn't hear about that
churn promptly.
I wonder if a better answer wouldn't be to reduce PGSTAT_STAT_INTERVAL.
I don't think that value has been reconsidered since the code was written,
circa turn of the century. Maybe even make it configurable, though that
could be overkill.
regards, tom lane