> If the table is dropped, there are no stats to update. right?

Ops, right. I focused too much on "all warnings should be visible in
the statistic, so the sum of warnings and statistics should match",
but of course that's not the case.

> However, I'm not entirely sure whether this behavior is always guaranteed.
> Could anyone clarify this?

There's another different corner-case if I move the injection point
inside pgstat_report_skipped_vacuum_analyze, after releasing the
syscache.

If the drop happens at that point, we insert an orphaned record into
the statistics, and it will be visible with the internal functions
(e.g. pg_stat_get_skipped_autoanalyze_count).

It is still invisible in the pg_stat_all_tables view, but now that
I've looked more at the code, I think internally it will stay there
permanently, even surviving pg_stat_reset?

> RangeVarGetRelid() still returns the table's OID

Yes, I also reached the same conclusion, I started testing because I
tried to see if I could break the double relid retrieval by some
scenarios (alter rename + create, drop-create etc), but it's not
possible. The ereports can execute CHECK_FOR_INTERRUPTS, but that
never calls ProcessCatchupInterrupt, and because of that it never runs
AccceptInvalidationMessages. And that's when I noticed that the
warning isn't visible in the statistics at all, and got distracted...


Reply via email to