Hi,
On 2023-04-17 12:13:41 -0400, Tom Lane wrote:
> Matthias van de Meent <boekewurm+postgres@gmail.com> writes:
> > Looking at log_newpage_range, it seems like we're always trying to log
> > a record if startblk < endblk; but don't register the PageIsNew()
> > buffers in the range. That means that if the last buffers in the range
> > are new, this can result in no buffers being registered in the last
> > iteration of the main loop (if the number of non-new buffers in the
> > range is 0 (mod 32)).
>
> Yeah, I just came to the same conclusion. One thing I don't understand
> yet: log_newpage_range is old (it looks like this back to v12), and
> that Assert is older, so why doesn't this reproduce further back?
> Maybe the state where all the pages are new didn't happen before?
> Is that telling us there's a bug somewhere else? Seems like a job
> for git bisect.
One plausible explanation is that bulk relation extension has made it more
likely to encounter this scenario. We had some bulk extension code before, but
it was triggered purely based on contention - quite unlikely in simple test
scenarios - but now we also bulk extend if we know that we'll insert multiple
pages (when coming from heap_multi_insert(), with sufficient data).
Greetings,
Andres Freund