Leonardo Francalanci <m_lists@yahoo.it> writes:
>> Cases with lots of irrelevant indexes. Zoltan's example had 4 indexes
>> per child table, only one of which was relevant to the query. In your
>> test case there are no irrelevant indexes, which is why the runtime
>> didn't change.
> Mmh... I must be doing something wrong. It looks to me it's not just
> the irrelevant indexes: it's the "order by" that counts.
Ah, I oversimplified a bit: actually, if you don't have an ORDER BY or
any mergejoinable join clauses, then the possibly_useful_pathkeys test
in find_usable_indexes figures out that we aren't interested in the sort
ordering of *any* indexes, so the whole thing gets short-circuited.
You need at least the possibility of interest in sorted output from an
indexscan before any of this code runs.
regards, tom lane