PATCH: add support for IN and @> in functional-dependency statistics use

From: Pierre Ducroquet <p(dot)psql(at)pinaraf(dot)info>
To: pgsql-hackers(at)postgresql(dot)org
Cc: Tomas Vondra <tomas(dot)vondra(at)2ndquadrant(dot)com>
Subject: PATCH: add support for IN and @> in functional-dependency statistics use
Date: 2020-02-01 07:51:04
Message-ID: 13902317.Eha0YfKkKy@pierred-pdoc
Views: Raw Message | Whole Thread | Download mbox | Resend email
Lists: pgsql-hackers


At my current job, we have a lot of multi-tenant databases, thus with tables
containing a tenant_id column. Such a column introduces a severe bias in
statistics estimation since any other FK in the next columns is very likely to
have a functional dependency on the tenant id. We found several queries where
this functional dependency messed up the estimations so much the optimizer
chose wrong plans.
When we tried to use extended statistics with CREATE STATISTIC on tenant_id,
other_id, we noticed that the current implementation for detecting functional
dependency lacks two features (at least in our use case):
- support for IN clauses
- support for the array contains operator (that could be considered as a
special case of IN)

After digging in the source code, I think the lack of support for IN clauses
is an oversight and due to the fact that IN clauses are ScalarArrayOpExpr
instead of OpExpr. The attached patch fixes this by simply copying the code-
path for OpExpr and changing the type name. It compiles and the results are
correct, with a dependency being correctly taken into consideration when
estimating rows. If you think such a copy paste is bad and should be factored
in another static bool function, please say so and I will happily provide an
updated patch.
The lack of support for @> operator, on the other hand, seems to be a decision
taken when writing the initial code, but I can not find any mathematical nor
technical reason for it. The current code restricts dependency calculation to
the = operator, obviously because inequality operators are not going to
work... but array contains is just several = operators grouped, thus the same
for the dependency calculation. The second patch refactors the operator check
in order to also include array contains.

I tested the patches on current HEAD, but I can test and provide back-ported
versions of the patch for other versions if needed (this code path hardly
changed since its introduction in 10).

Best regards

Pierre Ducroquet

Attachment Content-Type Size
0001-Add-support-for-IN-clauses-in-dependencies-check.patch text/x-patch 1.9 KB
0002-Add-support-for-array-contains-in-dependency-check.patch text/x-patch 1.4 KB


Browse pgsql-hackers by date

  From Date Subject
Next Message Fabien COELHO 2020-02-01 09:07:13 Re: pgbench - add pseudo-random permutation function
Previous Message Julien Rouhaud 2020-02-01 07:46:25 Re: unsupportable composite type partition keys