"Heikki Linnakangas" <heikki@enterprisedb.com> writes:
> Hannes Dorbath wrote:
>> Though it should only have to join a few rows it seems to scan all rows.
> What makes you think that's the case?
What it looks like to me is that the range of keys present in
pk_revisions_active corresponds to just the upper end of the range of
keys present in pk_revisions (somehow not too surprising). So the
mergejoin isn't the most effective plan possible for this case --- it
has to scan through much of pk_revisions before it starts getting
matches. The planner doesn't have any model for that though, and is
costing the plan on the assumption of uniformly-distributed matches.
A nestloop plan would be faster for this specific case, but much
slower if a large number of rows were requested.
regards, tom lane
