Noah Misch <noah@leadboat.com> writes:
> On Fri, Feb 02, 2024 at 08:18:50PM -0500, Tom Lane wrote:
>> Noah Misch <noah@leadboat.com> writes:
>>> Shall the top of the notes advise to reindex GIN indexes?
>> I thought about that, but it's a pretty low-probability failure
>> I think, so I didn't write that advice. Maybe I misjudged it.
> I can see there being failures so low-probability to omit that text, e.g. a
> failure identified theoretically and requiring a process to lose the CPU for
> hours. For this one, the reporter seems to have arrived at it without a
> deliberate search. This one just needs a recovery at the right WAL record,
> then two processes reaching the incomplete split concurrently.
The reporter didn't exactly say, but it did seem that the initial
detection was made without any code modifications, so I take your
point. I'll add the advice. Also, I now have this text for your
CREATE DATABASE fixes:
<para>
Ensure durability of <command>CREATE DATABASE</command> (Noah Misch)
</para>
<para>
If an operating system crash occurred during or shortly
after <command>CREATE DATABASE</command>, recovery could fail, or
subsequent connections to the new database could fail. If a base
backup was taken in that window, similar problems could be observed
when trying to use the backup. The symptom would be that the
database directory, <filename>PG_VERSION</filename> file, or
<filename>pg_filenode.map</filename> file was missing or empty.
</para>
This is ignoring the point that the set of observable symptoms might
differ between the OS crash and base-backup-recovery cases, but
I'm not sure that that's actually true, and in any case I don't think
it matters for the release notes.
regards, tom lane