Alvaro Herrera <alvherre@commandprompt.com> writes:
> Tom Lane escribi�:
>> Yeah, it could definitely run slower than the existing code --- in
>> particular the combination of all three (FOR UPDATE ORDER BY LIMIT)
>> would tend to become a seqscan-and-sort rather than possibly just
>> reading one end of an index. However, I quote the old aphorism that
>> it can be made indefinitely fast if it doesn't have to give the right
>> answer. The reason the current behavior is fast is it's giving the
>> wrong answer :-(
> So this probably merits a warning in the release notes for people to
> check that their queries continue to run with the performance they
> expect.
One problem with this is that there isn't any good way for someone to
get back the old behavior if they want to. Which might be a perfectly
reasonable thing, eg if they know that no concurrent update is supposed
to change the sort-key column. The obvious thing would be to allow
select * from (select * from foo order by col limit 10) ss for update;
to apply the FOR UPDATE last. Unfortunately, that's not how it works
now because the FOR UPDATE will get pushed down into the subquery.
I was shot down when I proposed a related change, a couple weeks ago
http://archives.postgresql.org/message-id/7741.1255278907@sss.pgh.pa.us
but it seems like we might want to reconsider.
regards, tom lane