08.12.2023 02:02, Nathan Bossart wrote:
> On Fri, Dec 08, 2023 at 09:55:58AM +1300, Thomas Munro wrote:
>> Now we have the question of whether to go forwards (commit the "socket
>> table" thing), or backwards (revert 04a09ee for now to clear the CI
>> failures). I don't love the hidden complexity of the socket table and
>> am not in a hurry to commit it, but I don't currently see another
>> way... on the other hand we have other CI flapping due to that problem
>> too so reverting 04a09ee would be sweeping problems under the carpet.
>> I still need to process your feedback/discoveries on that other thread
>> and it may take a few weeks for me to get to it.
> I don't think we need to revert 04a09ee provided the issue is unrelated and
> a fix is in development.
I've reviewed the links posted upthread and analyzed statistics of such
failures:
yes, it happens rather frequently in Cirrus CI, but there might be dozens
of successful runs, for example:
https://cirrus-ci.com/github/postgresql-cfbot/postgresql/commitfest%2F45%2F3686
has 1 postgres_fdw failure on Windows per 32 runs.
And there is only one such failure for 90 days in the buildfarm.
(Perhaps the probability of the failure depend on external factors, such as
concurrent activity.)
So I would not say that it's a dominant failure for now, and given that
04a09ee lives in master only, maybe we can save two commits (Revert +
Revert of revert) while moving to a more persistent solution.
Best regards,
Alexander