Re: INDEX ONLY scan with expression index - Mailing list pgsql-general

From Tom Lane
Subject Re: INDEX ONLY scan with expression index
Date
Msg-id 14366.1457418800@sss.pgh.pa.us
Whole thread Raw
In response to INDEX ONLY scan with expression index  (James Sewell <james.sewell@lisasoft.com>)
Responses Re: INDEX ONLY scan with expression index  ("David G. Johnston" <david.g.johnston@gmail.com>)
List pgsql-general
James Sewell <james.sewell@lisasoft.com> writes:
> Would anyone be able to shed some light on why expression based indexes
> can't be used for an index only scan?
> I've found a few comments saying this is the case, and I've proven it is
> the case in reality - but I can't seem to find the why.

Well, it would help if you posted a concrete example ... but there's
at least one known limitation: the planner's rule for whether an
index can be used for an index-only scan is that all variables needed
by the query be available from the index.  So if you have an index
on f(x), it might be useful for a query that needs f(x), but you won't
get an index-only scan for it because the planner fails to notice that
the query has no references to bare "x" but just "f(x)".  (This is
something that could be fixed, but it's not clear how to do so without
imposing considerable cost on queries that get no benefit because they
have no interest in f(x).)

The recommended workaround at the moment is to create a two-column index
on "f(x), x".  The second index column has no great value in reality,
but it lets the planner accept the index as usable for an IOS.  As a
small consolation prize, it might let you get an IOS on cases where you
*do* need x as well.

            regards, tom lane


pgsql-general by date:

Previous
From: James Sewell
Date:
Subject: INDEX ONLY scan with expression index
Next
From: "David G. Johnston"
Date:
Subject: Re: INDEX ONLY scan with expression index