On Mon, Mar 27, 2023 at 2:34 PM Peter Geoghegan <pg@bowt.ie> wrote:
> > Since this was back-patched, I think it's probably better to just
> > remove the error. We can introduce new validation if we want, but that
> > should probably be master-only.
>
> That makes sense.
Patch attached.
> I don't think that it's particularly likely that having refined
> aborted speculative insertion amcheck coverage will make a critical
> difference to any user, at any time. But "amcheck as documentation of
> the on-disk format" is reason enough to have it.
Sure, if someone feels like writing the code. I have to admit that I
have mixed feelings about this whole direction. In concept, I agree
with you entirely: a fringe benefit of having checks that tell us
whether or not a page is valid is that it helps to make clear what
page states we think are valid. In practice, however, the point you
raise in your first sentence weighs awfully heavily with me. Spending
a lot of energy on checks that are unlikely to catch practical
problems feels like it may not be the best use of time. I'm not sure
exactly where to draw the line, but it seems highly likely to be that
there are things we could deduce about the page that wouldn't be worth
the effort. For example, would we bother checking that a tuple with an
in-progress xmin does not have a smaller natts value than a tuple with
a committed xmin? Or that natts values are non-decreasing across a HOT
chain? I suspect there are even more obscure examples of things that
should be true but might not really be worth worrying about in the
code.
--
Robert Haas
EDB: http://www.enterprisedb.com