Thread: Query is not using index when it should
I tried to subscribe to pgsql-performance, but there seems to be
something wrong with the majordomo, so I'm sending to general too,
where I'm already subscribed.
My problem is this, using PostgreSQL 7.4.6:
I have a table that looks like this:
Table "public.cjm_object"
Column | Type | Modifiers
-----------+-------------------+-----------
timestamp | bigint | not null
jobid | bigint | not null
objectid | bigint | not null
class | integer | not null
field | character varying | not null
data | bytea |
Indexes:
"cjm_object_pkey" primary key, btree ("timestamp", jobid, objectid, "class", field)
"idx_cjm_object1" btree (objectid, "class", field)
The table has 283465 rows, and the column combination
(objectid,class,field) can occur several times.
Doing a search with all columns in the pkey works, it uses the index:
db=# explain analyze select * from cjm_object where timestamp=1102497954815296 and jobid=9 and objectid=4534 and
class=12and field='paroid';
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------
Index Scan using cjm_object_pkey on cjm_object (cost=0.00..32.75 rows=1 width=54) (actual time=0.169..0.172 rows=1
loops=1)
Index Cond: ("timestamp" = 1102497954815296::bigint)
Filter: ((jobid = 9) AND (objectid = 4534) AND ("class" = 12) AND ((field)::text = 'paroid'::text))
Total runtime: 0.381 ms
(4 rows)
But when doing a search with objectid, class and field, it doesn't use
the idx_cjm_object1 index.
db=# explain analyze select * from cjm_object where objectid=4534 and class=12 and field='paroid';
QUERY PLAN
-----------------------------------------------------------------------------------------------------------
Seq Scan on cjm_object (cost=0.00..7987.83 rows=2 width=54) (actual time=21.660..475.664 rows=1 loops=1)
Filter: ((objectid = 4534) AND ("class" = 12) AND ((field)::text = 'paroid'::text))
Total runtime: 475.815 ms
(3 rows)
I have tried to set enable_seqscan to false, but it gives the same
result, except that the estimated cost is higher.
I have also done a vacuum full analyze, and I have reindexed the
database, the table and the index. I have dropped the index and
recreated it, but it still gives the same result.
Please, could someone give me a clue to this?
Tomas
On Fri, 10 Dec 2004, Tomas [iso-8859-1] Sk�re wrote: > I have a table that looks like this: > > Table "public.cjm_object" > Column | Type | Modifiers > -----------+-------------------+----------- > timestamp | bigint | not null > jobid | bigint | not null > objectid | bigint | not null > class | integer | not null > field | character varying | not null In 7.4.x and earlier, you need to cast the value you're comparing to into a bigint in order to make sure the indexes are used (in your timestamp case it appears to work because the value doesn't fit in a plain integer). 8.0 should handle this better. > But when doing a search with objectid, class and field, it doesn't use > the idx_cjm_object1 index. > db=# explain analyze select * from cjm_object where objectid=4534 and class=12 and field='paroid'; Using one of objectid=4534::bigint objectid='4534' objectid=CAST(4534 as bigint) rather than objectid=4534 should make this indexable in 7.4.x.
Stephan Szabo <sszabo@megazone.bigpanda.com> writes:
> On Fri, 10 Dec 2004, Tomas [iso-8859-1] Skäre wrote:
>
> > I have a table that looks like this:
> >
> > Table "public.cjm_object"
> > Column | Type | Modifiers
> > -----------+-------------------+-----------
> > timestamp | bigint | not null
> > jobid | bigint | not null
> > objectid | bigint | not null
> > class | integer | not null
> > field | character varying | not null
>
> In 7.4.x and earlier, you need to cast the value you're comparing to into
> a bigint in order to make sure the indexes are used (in your timestamp
> case it appears to work because the value doesn't fit in a plain integer).
> 8.0 should handle this better.
Thanks, casting worked well for that query. Now, could someone please
help me to get this query faster? With the 283465 rows, it takes far
too long time, I think. This is on a 2GHz Celeron running Linux 2.6.
shared_buffers=1000, sort_mem=1024.
select c.* from cjm_object c
inner join
(select max(timestamp) as timestamp,objectid,field from cjm_object
group by objectid,field) t
using(timestamp,objectid,field)
where 1=1 and data is not null
order by objectid,field;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------
Merge Join (cost=145511.85..150759.75 rows=1 width=54) (actual time=17036.147..20968.811 rows=208246 loops=1)
Merge Cond: (("outer".objectid = "inner".objectid) AND ("outer"."?column7?" = "inner"."?column4?") AND
("outer"."timestamp"= "inner"."timestamp"))
-> Sort (cost=47007.75..47611.06 rows=241324 width=54) (actual time=5113.099..5586.094 rows=236710 loops=1)
Sort Key: c.objectid, (c.field)::text, c."timestamp"
-> Seq Scan on cjm_object c (cost=0.00..5862.65 rows=241324 width=54) (actual time=0.129..1788.125
rows=236710loops=1)
Filter: (data IS NOT NULL)
-> Sort (cost=98504.09..99212.75 rows=283465 width=48) (actual time=11922.081..12427.683 rows=255001 loops=1)
Sort Key: t.objectid, (t.field)::text, t."timestamp"
-> Subquery Scan t (cost=45534.39..51912.35 rows=283465 width=48) (actual time=5484.943..9289.061
rows=255001loops=1)
-> GroupAggregate (cost=45534.39..49077.70 rows=283465 width=25) (actual time=5484.925..8178.531
rows=255001loops=1)
-> Sort (cost=45534.39..46243.05 rows=283465 width=25) (actual time=5484.285..6324.067
rows=283465loops=1)
Sort Key: objectid, field
-> Seq Scan on cjm_object (cost=0.00..5862.65 rows=283465 width=25) (actual
time=0.124..852.749rows=283465 loops=1)
Total runtime: 21161.144 ms
Quick explanation of the query:
Each row in the table is a field, which is part of an object. Ex:
timestamp objectid field data
1 1 name test
1 1 type something
1 2 name test2
1 2 type whatever
Timestamp is when the entry was inserted in the databas. When updating
a single field for an object, a new line with the new value is added,
data set to NULL if the field is deleted. So the above content could
now be:
timestamp objectid field data
1 1 name test
1 1 type something
1 2 name test2
1 2 type whatever
2 1 name newname
2 1 type <NULL>
Now, the query picks out the highest timestamp for each
(objectid,field) and then selects all columns for each match,
filtering out NULL data and ordering per objectid.
Is there any way to make this query faster? I've tried rewriting it,
putting the subquery as EXISTS condition, but it doesn't make it
faster. I've tried to create different indices, but they don't seem to
be used in this query.
Greetings,
Tomas