Thread: Query doesn't use index on hstore column

On Fri, Dec 05, 2014 at 09:42:20AM +1300, Michael Barker wrote:
> 1)  Created table with hstore column and btree index.
>
> barkerm=# \d audit
>                                        Table "public.audit"
>     Column     |            Type             |
> Modifiers
> ---------------+-----------------------------+----------------------------------------------------
>  id            | integer                     | not null default
> nextval('audit_id_seq'::regclass)
>  principal_id  | integer                     |
>  created_at    | timestamp without time zone |
>  root          | character varying(255)      |
>  template_code | character(3)                |
>  attributes    | hstore                      |
>  args          | character varying(255)[]    |
> Indexes:
>     "audit_pkey" PRIMARY KEY, btree (id)
>     "audit_attributes_idx" btree (attributes)
>
> ...
> 5) Explain query using the attributes column in the where clause (uses Seq
> Scan).
>
> barkerm=# explain analyse select * from audit where attributes->'accountId'
> = '1879355460';
>                                                  QUERY PLAN
>
> ------------------------------------------------------------------------------------------------------------
>  Seq Scan on audit  (cost=0.00..35409.00 rows=5000 width=133) (actual
> time=114.314..218.821 rows=1 loops=1)
>    Filter: ((attributes -> 'accountId'::text) = '1879355460'::text)
>    Rows Removed by Filter: 999999
>  Planning time: 0.074 ms
>  Execution time: 218.843 ms
> (5 rows)
>
Hi Michael,

I think your index definitions need to be on the particular attribute from
attributes and not attributes itself. That works but it does not apply to
the query you show above. I think that the binary json type in 9.4 will
do what you want. I have not worked with it myself, just looked at the docs.

Regards,
Ken

Michael Barker <mikeb01@gmail.com> writes:
> I'm currently experimenting with hstore on Posgtres 9.4rc1.  I've created a
> table with an hstore column, with and index on that column (tried both gin
> and btree indexes) and the explain plan says that the index is never used
> for the lookup and falls to a sequential scan every time (table has 1 000
> 000 rows).  The query plans and execution time for btree index, gin index
> and unindexed are the same.  Is there something I'm doing wrong or missing
> in order to get indexes to work on hstore columns?

Well, first off, a btree index is fairly useless for this query,
because btree has no concept that the hstore has any sub-structure.
A GIN index or GIST index could work though.  Secondly, you have to
remember that indexable WHERE conditions in Postgres are *always* of
the form "WHERE indexed_column indexable_operator some_comparison_value".
So the trick is to recast the condition you have into something that
looks like that.  Instead of

    WHERE attributes->'accountId' = '1879355460'

you could do

    WHERE attributes @> 'accountId=>1879355460'

(@> being the hstore containment operator, ie "does attributes contain
a pair that looks like this?") or equivalently but possibly easier to
generate,

    WHERE attributes @> hstore('accountId', '1879355460')

Another possibility if you're only concerned about indexing searches
for one or a few specific keys is to use expression indexes:

    CREATE INDEX ON audit ((attributes->'accountId'));

whereupon your original query works, since the left-hand side of
the '=' operator is now the indexed expression.  (Here, since you
are testing plain equality on the indexed value, a btree works fine.)

You might care to read
http://www.postgresql.org/docs/9.4/static/indexes.html
to get a better handle on what Postgres indexes can and can't do.

            regards, tom lane