On 14.8.2013 20:35, Tom Lane wrote:
> "Tomas Vondra" <tv@fuzzy.cz> writes:
>> I've run into a strange plan difference on 9.1.9 - the first query
>> does "DISTINCT" by doing a GROUP BY on the columns (both INT). ...
>> Now, this takes ~45 seconds to execute, but after rewriting the
>> query to use the regular DISTINCT it suddenly switches to
>> HashAggregate with ~1/3 the cost (although it produces the same
>> output, AFAIK), and it executes in ~15 seconds.
>
> [ scratches head... ] I guess you're running into some corner case
> where choose_hashed_grouping and choose_hashed_distinct make
> different choices. It's going to be tough to debug without a test
> case though. I couldn't reproduce the behavior in a few tries here.
>
>> BTW I can't test this on 9.2 or 9.3 easily, as this is our
>> production environment and I can't just export the data. I've tried
>> to simulate this but so far no luck.
>
> I suppose they won't yet you step through those two functions with a
> debugger either ...
I've managed to get the data to a different machine, and I've spent some
time on debugging it. It seems that the difference is in evaluating
hashentrysize - while