Peter Geoghegan <pg@bowt.ie> writes:
> On Fri, Feb 8, 2019 at 6:00 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Yeah, I saw that in the commit message, but didn't really think that
>> the release note entry needed to explain it that way. Could be
>> argued differently though.
> I'm pretty confident that somebody is going to miss this
> functionality, if this account of how the patch helped Yandex is
> anything to go by:
> https://www.postgresql.org/message-id/7B44397E-5E0A-462F-8148-1C444640FA0B%40simply.name
Ugh. Well, hopefully somebody will find a less buggy solution
in the future.
>>> We may even need to revise the v10 release notes.
>> Perhaps just remove that item from the 10.0 notes?
> The wording could be changed to reflect the current reality within
> GIN. It's still useful that posting trees are only locked when there
> are pages to be deleted.
The v10 release notes just say
Reduce page locking during vacuuming of <acronym>GIN</acronym> indexes
(Andrey Borodin)
so it doesn't seem like there's any difference at that level of detail.
But I'll expand the new release note, say
This change partially reverts a performance improvement, introduced
in version 10, that attempted to reduce the number of index pages
locked during deletion of a GIN posting tree page. That's now been
found to lead to deadlocks, so we've removed it pending closer
analysis.
regards, tom lane