Hello everybody!
I have found a performance issue with 2 equivalent queries stably
taking different (~x2) time to finish. In just a few words it can be
described like this: if you have a lot of values in an IN() statement,
you should put most heavy (specifying most number of rows) ids first.
This is mostly just a bug submit, than looking for help.
So this is what I have:
- RHEL
- PostgreSQL 8.3.1
- A table ext_feeder_item with ~4.6M records.
kia=# \d+ ext_feeder_item;
Table
"public.ext_feeder_item"
Column | Type | Modifiers | Description
----------+--------------------------+--------------------------------------------------------------+-------------
id | bigint | not null default
nextval('ext_feeder_item_id_seq'::regclass) |
feed_id | bigint | not
null |
pub_date | timestamp with time zone
| |
Indexes:
"ext_feeder_item_pkey" PRIMARY KEY, btree (id)
"ext_feeder_item_feed_id_pub_date_idx" btree (feed_id, pub_date)
"ext_feeder_item_idx" btree (feed_id)
Triggers:
....
Has OIDs: no
- Statistics for the fields feed_id and pub_date are set to 1000;
- The table have just been vacuumed and analyzed.
- A simple query to the table:
SELECT
id
FROM
ext_feeder_item AS i
WHERE
i.feed_id IN (...)
ORDER BY pub_date DESC, id DESC
LIMIT 11 OFFSET 0;
with many (~1200) ids in the IN() statement.
- The count of rows distribution for these ids (may be thought of
as foreign keys in this table) is the following:
id = 54461: ~180000 - actually the most heavy id in the whole table.
other ids: a single id at most specifies 2032 rows; 6036 rows total.
- If I perform a query with
IN(54461, ...)
it stably (5 attempts) takes 4.5..4.7 secs. to perform.
QUERY PLAN
Limit (cost=1463104.22..1463104.25 rows=11 width=16) (actual
time=4585.420..4585.452 rows=11 loops=1)
-> Sort (cost=1463104.22..1464647.29 rows=617228 width=16)
(actual time=4585.415..4585.425 rows=11 loops=1)
Sort Key: pub_date, id
Sort Method: top-N heapsort Memory: 17kB
-> Bitmap Heap Scan on ext_feeder_item i
(cost=13832.40..1449341.79 rows=617228 width=16) (actual
time=894.622..4260.441 rows=185625 loops=1)
Recheck Cond: (feed_id = ANY ('{54461, ...}'::bigint[]))
-> Bitmap Index Scan on ext_feeder_item_idx
(cost=0.00..13678.10 rows=617228 width=0) (actual time=884.686..884.686
rows=185625 loops=1)
Index Cond: (feed_id = ANY ('{54461,
...}'::bigint[]))
Total runtime: 4585.852 ms
- If I perform a query with
IN(..., 54461)
it stably (5 attempts) takes 9.3..9.5 secs. to perform.
QUERY PLAN
Limit (cost=1463104.22..1463104.25 rows=11 width=16) (actual
time=9330.267..9330.298 rows=11 loops=1)
-> Sort (cost=1463104.22..1464647.29 rows=617228 width=16)
(actual time=9330.263..9330.273 rows=11 loops=1)
Sort Key: pub_date, id
Sort Method: top-N heapsort Memory: 17kB
-> Bitmap Heap Scan on ext_feeder_item i
(cost=13832.40..1449341.79 rows=617228 width=16) (actual
time=1018.401..8971.029 rows=185625 loops=1)
Recheck Cond: (feed_id = ANY ('{... ,54461}'::bigint[]))
-> Bitmap Index Scan on ext_feeder_item_idx
(cost=0.00..13678.10 rows=617228 width=0) (actual
time=1008.791..1008.791 rows=185625 loops=1)
Index Cond: (feed_id = ANY ('{...
,54461}'::bigint[]))
Total runtime: 9330.729 ms
I don't know what are the roots of the problem, but I think that some
symptomatic healing could be applied: the PostgreSQL could sort the IDs
due to the statistics.
So currently I tend to select the IDs from another table ordering them
due to their weights: it's easy for me thanks to denormalization.
Also I would expect from PostgreSQL that it sorted the values to make
index scan more sequential, but this expectation already conflicts with
the bug described above :)
|
