Re: RV: bad result in a query!! :-(

Richard Huxton <dev(at)archonet(dot)com> writes:
> On Tuesday 15 Oct 2002 12:12 pm, Nigel J. Andrews wrote:
>> On Tue, 15 Oct 2002, Jose Antonio Leo wrote:
>>> -> Sort (cost=10821.77..10821.77 rows=1485
> width=118) (actual time=16453.64..199329.55 rows=49801240 loops=1)
>>
>> What is this all about, the seqscan only returns 75918 rows?

> Yep - very strange. I'm not sure where the 4 million comes from - I can't see
> any relationship with the 75918.

I think what is happening is that there are many equal keys in the
relations being joined. If you think about how a mergejoin works,
it has to back up and rescan a segment of the inner relation each
time it advances to a new outer tuple that has a key matching the
prior key. I believe that the EXPLAIN ANALYZE machinery counts each
row fetched from the inner relation afresh, even if it's a re-fetch
of a row already fetched.

There is not currently any code in the planner to try to account
for this effect; if there were, it might choose a different plan.
(Not that I'm sure a hash join would be much better.)

Jose, how many distinct cod_ae1 values have you actually got in
each table? Can you use additional join conditions (perhaps
cod_ae2, cod_ae3) to improve the specificity of the match between
the tables?

regards, tom lane