PostgreSQL – Improving Performance of Window Functions on Large Tables

performancepostgresqlquery-performancewindow functions

I have a PostgreSQL table constructed as

device_id | point     | dt_edit
----------+-----------+----------
100       | geometry  | timestamp
101       | geometry  | timestamp
100       | geometry  | timestamp
102       | geometry  | timestamp
102       | geometry  | timestamp
101       | geometry  | timestamp

I need to select the last 2 records ordered by dt_edit from unique device_id. This query works very slow, on 1 billion records – 500 seconds:

 SELECT dt.device_id,
    dt.point,
    dt.dt_edit
   FROM ( SELECT gps_data.device_id,
            gps_data.point,
            gps_data.dt_edit,
            rank() OVER (PARTITION BY gps_data.device_id
                         ORDER BY gps_data.dt_edit DESC) AS rank
           FROM gps_data) dt
  WHERE dt.rank <= 2

Any ideas?

Best Answer

If we can assume a device table holding all devices of interest. Example:

CREATE TABLE device (device_id int, device text);
INSERT INTO device (device_id, device) VALUES
  (100, 'a')
, (101, 'b')
, (102, 'c');

The query can be very simple:

SELECT d.device_id, g.point, g.dt_edit
FROM   device d
,      LATERAL (
   SELECT point, dt_edit
   FROM   gps_data
   WHERE  device_id = d.device_id
   ORDER  BY dt_edit DESC NULLS LAST -- more items?
   LIMIT  2
   ) g;

Assuming a much smaller cardinality in table device than in table gps_data, i.e. many rows per device in the latter.

If dt_edit is not UNIQUE, add more items to ORDER BY to disambiguate the sort order.

All you need for this to be fast is a multicolumn index;

CREATE INDEX gps_data_foo_idx
ON gps_data (device_id, dt_edit DESC NULLS LAST); -- more items?

The NULLS LAST modifier is useful if the column can be NULL, but does not hurt in any case. Either way, the query must match the index.

PostgreSQL sort by datetime asc, null first?

If you don't have a device table, I would suggest to create one (and keep it current).
But there are ways to make this fast even without device table ...

Details:

Consistent rows

The important question which does not seem to be on your radar yet:
From each set of rows for the same seriesName, do you want the columns of one row, or just any values from multiple rows (which may or may not go together)?

Your answer does the latter, you combine the maximum dbid with the maximum retreivaltime, which may come from a different row.

To get consistent rows, use DISTINCT ON and wrap it in a subquery to order the result differently:

SELECT * FROM (
   SELECT DISTINCT ON (seriesName)
          dbid, seriesName, retreivaltime
   FROM   FileItems
   WHERE  sourceSite = 'mk' 
   ORDER  BY seriesName, retreivaltime DESC NULLS LAST  -- latest retreivaltime
   ) sub
ORDER BY retreivaltime DESC NULLS LAST
LIMIT  100;

Details for DISTINCT ON:

Select first row in each GROUP BY group?

Aside: should probably be retrievalTime, or better yet: retrieval_time. Unquoted mixed case identifiers are a common source of confusion in Postgres.

Better Performance with rCTE

Since we are dealing with a big table here, we'd need a query that can use an index, which is not the case for the above query (except for WHERE sourceSite = 'mk')

On closer inspection, your problem seems to be a special case of a loose index scan. Postgres does not support loose index scans natively, but it can be emulated with a recursive CTE. There is a code example for the simple case in the Postgres Wiki.

Related answer on SO with more advanced solutions, explanation, fiddle:

Optimize GROUP BY query to retrieve latest record per user

Your case is more complex, though. But I think I found a variant to make it work for you. Building on this index (without WHERE sourceSite = 'mk')

CREATE INDEX mi_special_full_idx ON MangaItems
(retreivaltime DESC NULLS LAST, seriesName DESC NULLS LAST, dbid)

Or (with WHERE sourceSite = 'mk')

CREATE INDEX mi_special_granulated_idx ON MangaItems
(sourceSite, retreivaltime DESC NULLS LAST, seriesName DESC NULLS LAST, dbid)

The first index can be used for both queries, but is not fully efficient with the additional WHERE condition. The second index is of very limited use for the first query. Since you have both variants of the query consider creating both indexes.

I added dbid at the end to allow Index Only scans.

This query with a recursive CTE makes use of the index. I tested with Postgres 9.3 and it works for me: no sequential scan, all index-only scans:

WITH RECURSIVE cte AS (
   (
   SELECT dbid, seriesName, retreivaltime, 1 AS rn, ARRAY[seriesName] AS arr
   FROM   MangaItems
   WHERE  sourceSite = 'mk'
   ORDER  BY retreivaltime DESC NULLS LAST, seriesName DESC NULLS LAST
   LIMIT  1
   )
   UNION ALL
   SELECT i.dbid, i.seriesName, i.retreivaltime, c.rn + 1, c.arr || i.seriesName
   FROM   cte c
   ,      LATERAL (
      SELECT dbid, seriesName, retreivaltime
      FROM   MangaItems
      WHERE (retreivaltime, seriesName) < (c.retreivaltime, c.seriesName)
      AND    sourceSite = 'mk'  -- repeat condition!
      AND    seriesName <> ALL(c.arr)
      ORDER  BY retreivaltime DESC NULLS LAST, seriesName DESC NULLS LAST
      LIMIT  1
      ) i
   WHERE  c.rn < 101
   )
SELECT dbid
FROM   cte
ORDER  BY rn;

You need to include seriesName in ORDER BY, since retreivaltime is not unique. "Almost" unique is still not unique.

Explain

The non-recursive query starts with the latest row.
The recursive query adds the next-latest row with a seriesName that's not in the list, yet etc., until we have 100 rows.
Essential parts are the JOIN condition (b.retreivaltime, b.seriesName) < (c.retreivaltime, c.seriesName) and the ORDER BY clause ORDER BY retreivaltime DESC NULLS LAST, seriesName DESC NULLS LAST. Both match the sort order of the index, which allows for the magic to happen.
Collecting seriesName in an array to rule out duplicates. The cost for b.seriesName <> ALL(c.foo_arr) grows progressively with the number of rows, but for just 100 rows it is still cheap.
Just returning dbid as clarified in the comments.

Alternative with partial indexes:

We have been dealing with similar problems before. Here is a highly optimized complete solution based on partial indexes and a looping function:

Can spatial index help a "range - order by - limit" query

Probably the fastest way (except for a materialized view) if done right. But more complex.

Materialized View

Since you do not have a lot of write operations and they are not performance-critical as stated in the comments (should be in the question), save the top n pre-computed rows in a materialized view and refresh it after relevant changes to the underlying table. Base your performance-critical queries on the materialized view instead.

Could just be a "thin" mv of the latest 1000 dbid or so. In the query, join to the original table. For instance, if content is sometimes updated, but the top n rows can remain unchanged.
Or a "fat" mv with whole rows to return. Faster, yet. Needs to be refreshed more often, obviously.

Details in the manual here and here.

Best Answer

Related Solutions

Postgresql – slow queries – set weight to token type in tsquery – postgresql

PostgreSQL – GROUP BY One Column While Sorting by Another

Consistent rows

Better Performance with rCTE

Explain

Alternative with partial indexes:

Materialized View

Related Question