Re: Performance improvement for joins where outer side is unique

On 7 April 2016 at 08:01, David Rowley <david.rowley@2ndquadrant.com> wrote:
> On 7 April 2016 at 04:05, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Starting to look at this again.  I wonder, now that you have the generic
>> caching mechanism for remembering whether join inner sides have been
>> proven unique, is it still worth having the is_unique_join field in
>> SpecialJoinInfo?  It seems like that's creating a separate code path
>> for special joins vs. inner joins that may not be buying us much.
>> It does potentially save lookups in the unique_rels cache, if you already
>> have the SpecialJoinInfo at hand, but I'm not sure what that's worth.
>
> I quite like that field where it is, as it should make
> remove_useless_joins() a bit more efficient, as after a LEFT JOIN is
> removed, the previous code would go off and try to make sure all the
> joins are unique again, but now we cache that, and save it from having
> to bother doing that again, on joins already marked as unique.
>
> Certainly changing that would mean one less special case in
> joinpath.c, as the JOIN_LEFT case can be handle the same as the other
> cases, although it looks like probably, if I do change that, then I'd
> probably move is_innerrel_unique_for() into analyzejoins.c, and put
> the special case for JOIN_LEFT in that function, so that it calls
> specialjoin_is_unique_join(), then cache the sjinfo->min_righthand in
> the unique_rels cache if the result comes back positive, and in the
> non_unique_rels cache if negative... But it seems a bit crazy to go to
> the trouble or all that caching, when we can just throw the result in
> a struct field in the case of Special Joins.  Maybe we could just hide
> both the new joinpath.c functions in analyzejoins.c and call it quits.
> It's not as if there's no special cases for JOIN_LEFT in that file.

We could also get rid of the SpecialJoinInfo.is_unique_join and just
store this as optimal_jointype, where this would be initialised to
jointype in make_outerjoininfo(), and then set in mark_unique_joins().
This would simplify the test in get_optimal_jointype(), perhaps if
(IS_OUTER_JOIN(jointype)) return sjinfo->optimal_jointype;

-- David Rowley                   http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services