On Mon, Mar 17, 2014 at 11:15 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> David Johnston <polobo@yahoo.com> writes:
>> Need to discuss the general "why" before any meaningful help on the "how" is
>> going to be considered by hackers.
>
> Possibly worth noting is that in past discussions, we've concluded that
> the most sensible type of hint would not be "use this plan" at all, but
> "here's what to assume about the selectivity of this WHERE clause".
> That seems considerably less likely to break than any attempt to directly
> specify plan details.
Yeah -- the most common case I see is outlier culling where several
repeated low non-deterministic selectivity quals stack reducing the
row count estimate to 1. For example:
SELECT * FROM foo WHERE length(bar) <= 1000 AND length(bar) >= 2;
The user may have special knowledge that the above is very (or very
un-) selective that is difficult or not cost effective to gather in
the general case. IIRC in the archives (heh) there is a special
workaround using indexes and some discussion regarding how a
hypothetical feature involving user input selectivity estimates might
look. I don't think that discussion is complete: the syntax for user
input selectivity is an unsolved problem.
There's a big difference between saying to the planner, "Use plan X"
vs "Here's some information describing the data supporting choosing
plan X intelligently". The latter allows for better plans in the face
of varied/changing data, integrates with the planner in natural way,
and encourages users to understand how the planner works.
merlin