Re: No index only scan on md5 index

From: Tom Lane
Subject: Re: No index only scan on md5 index
Date: ,
Msg-id: 31103.1448499661@sss.pgh.pa.us
(view: Whole thread, Raw)
In response to: No index only scan on md5 index  (Adam Brusselback)
Responses: Re: No index only scan on md5 index  (Adam Brusselback)
List: pgsql-performance

Tree view

No index only scan on md5 index  (Adam Brusselback, )
 Re: No index only scan on md5 index  ("David G. Johnston", )
  Re: No index only scan on md5 index  (Adam Brusselback, )
 Re: No index only scan on md5 index  (Tom Lane, )
  Re: No index only scan on md5 index  (Adam Brusselback, )
   Re: No index only scan on md5 index  (Albe Laurenz, )

Adam Brusselback <> writes:
> CREATE TABLE attachment
> (
>   attachment_id uuid NOT NULL DEFAULT gen_random_uuid(),
>   attachment_name character varying NOT NULL,
>   attachment_bytes_size integer NOT NULL,
>   attachment_bytes bytea NOT NULL,
>   CONSTRAINT attachment_pkey PRIMARY KEY (attachment_id)
> );
> CREATE INDEX idx_attachment_bytes_md5 ON attachment
> ((md5(attachment_bytes)::uuid));

> But if I wanted to return the md5 value, it seems to be totally unable to
> use an index only scan:
> SELECT md5(attachment_bytes)::uuid
> FROM attachment;

Nope, sorry, you're out of luck on that, because the check for whether an
index-only scan is feasible checks whether all the variables used in the
query are available from the index.  (Testing whether an index expression
could match everything asked for would greatly slow down planning, whether
or not the index turned out to be applicable, so we don't try.  I have
some rough ideas about making that better, but don't hold your breath.)

IIRC, it does actually get it right in terms of constructing the
finished plan, if you can get past the index-only-scan-is-feasible test.
So some people have done something like this to avoid recalculations of
expensive functions:

create table ff(f1 float8);
create index on ff(sin(f1), f1);
select sin(f1) from ff;  -- can generate IOS and not re-evaluate sin()

But if I'm right in guessing that attachment_bytes can be large,
that's not going to be a workable hack for your case.

Probably the only thing that's going to work for you is to store
md5(attachment_bytes) in its own plain column (you can use a trigger
to compute it for you), and then build a regular index on that,
and query for that column not the md5() expression.

            regards, tom lane



pgsql-performance by date:

From: Adam Brusselback
Date:
Subject: Re: No index only scan on md5 index
From: Albe Laurenz
Date:
Subject: Re: No index only scan on md5 index