On Tue, 18 Aug 2009, Tom Lane wrote:
> Matthew Wakeling <matthew@flymine.org> writes:
>> I'm seeing some interesting behaviour. I'm executing a query where I
>> perform a merge join between two copies of the same table, completely
>> symmetrically, and the two sides of the merge are sourced differently.
>
> This is not as surprising as you think. A mergejoin is *not*
> symmetrical between its two inputs: the inner side is subject to being
> partially rewound and rescanned when the outer side is advanced to a new
> row with the same merge key. This means there is a premium on cheap
> rescan for the inner side that doesn't exist for the outer ... and a
> sort node is cheaper to rescan than a generic indexscan.
Very clever. Yes, that is what is happening. I'm surprised that the system
doesn't buffer the inner side to avoid having to rescan each time, but
then I guess you would have problems if the buffer grew larger than
memory.
> It's impossible to tell from the data you provided whether the planner
> was correct to pick a sort over an indexscan for the inner side, but the
> fact that it did so is not prima facie evidence of a bug. You could
> force choice of the other plan via enable_sort = off and then compare
> estimated and actual runtimes to see if the planner got it right.
Yes, it does get an almost unmeasureable amount slower if I force sorts
off and nested loop (its next choice) off.
Matthew
--
$ rm core
Segmentation Fault (core dumped)
