Postgresql – CTE where to place the where clause to filter rows sooner (in postgresql)

cteoptimizationpostgresql

I have function with CTE which looks like this:

function param(p_userId, p_templateId)

    with a(select...),
         b (select..), 
         c (insert..returning..), 
         d (update..returning..),
         ...k(select..)
    insert ... 
    select * from k 
    where  userId = p_userId 
    and    templateId = p_templateId

I'm able to put the where clause right at the begining in the a CTE. If I place the condition at the end like I have it now, will it filtered the rows out at the earliest or at the latest time? What is faster?

Basically what I do is, in the CTE's which appear sooner, like the a or b, (rather then k), I'm prefetching data or precalculating columns in order to use them later in the following CTEs.

Hope this question is still not too vague.

Best Answer

CTEs pose as optimization fences in Postgres.

Each CTE is planned, executed and materialized separately. The predicates of a later CTE are not visible in earlier CTEs. Place conditions to eliminate rows as early as possible (a or b in your example).

Or rewrite with subqueries to allow more aggressive optimizations in Postgres. (But you should still eliminate irrelevant rows as early as possible, optimization only goes so far ...)

Details in Craig's blog post:

https://blog.2ndquadrant.com/postgresql-ctes-are-optimization-fences/

Get qualifying rows only

One way ..

WITH x AS (
   SELECT *
         ,row_number() OVER (PARTITION BY reference ORDER BY date_out DESC) AS rn
   FROM   tbl
   )
, y AS (
   SELECT *
   FROM   x
   WHERE  x.rn = 1
   AND    date_out >= '2012-10-01'
   AND    date_out <  '2012-11-01'
   )
, z AS (
   SELECT x.*
   FROM   x
   JOIN   y USING (reference)
   WHERE  x.rn = 2
   AND    x.centreid = 1
   )
SELECT y.taskid, y.reference, y.centreid, y.date_out
FROM   y
JOIN   z USING (reference)

UNION  ALL
SELECT taskid, reference, centreid, date_out
FROM   z
ORDER  BY reference, date_out;

Another way:

WITH x AS (
   SELECT *
         ,row_number() OVER (PARTITION BY reference ORDER BY date_out DESC) AS rn
   FROM   tbl
   )
,y AS (
    SELECT x.*, y.taskid AS taskid2, y.centreid AS centreid2, y.date_out AS date_out2
    FROM   x
    JOIN   x y USING (reference)
    WHERE  x.rn = 1
    AND    x.date_out >= '2012-10-01'
    AND    x.date_out <  '2012-11-01'
    AND    y.rn = 2
    AND    y.centreid = 1
    )
SELECT y.taskid, y.reference, y.centreid, y.date_out
FROM   y

UNION  ALL
SELECT y.taskid2, y.reference, y.centreid2, y.date_out2
FROM   y
ORDER  BY reference, date_out;

I'd expect the second one to be faster. Depends on your data distribution. Test with EXPLAIN ANALYZE.

Get all rows for qualifying references

WITH x AS (
   SELECT *
         ,row_number() OVER (PARTITION BY reference ORDER BY date_out DESC) AS rn
   FROM   tbl
   )
,y AS (
    SELECT reference
    FROM   x
    JOIN   x y USING (reference)
    WHERE  x.rn = 1
    AND    x.date_out >= '2012-10-01'
    AND    x.date_out <  '2012-11-01'
    AND    y.rn = 2
    AND    y.centreid = 1
    )
SELECT *
FROM   tbl
JOIN   y USING (reference)
ORDER  BY reference, date_out;

-> sqlfiddle

Answer to follow-up in comment

Separate groups if more than 30 days between entries.

WITH a AS (
   SELECT *
         ,lag(date_out) OVER (PARTITION BY reference ORDER BY date_out DESC) AS last_date
         ,CASE WHEN date_out > 
                   (lag(date_out) OVER (PARTITION BY reference ORDER BY date_out DESC) - 30)
               THEN 0 ELSE 1
          END AS step
   FROM   tbl
   )
,b AS (
   SELECT *
         ,sum(step) OVER (PARTITION BY reference ORDER BY date_out DESC) AS grp
   FROM a
   )
 ,c AS (
   SELECT *
         ,row_number() OVER (PARTITION BY reference, grp ORDER BY date_out DESC) AS rn
   FROM b
   )
,d AS (
   SELECT reference, grp
   FROM   c
   JOIN   c d USING (reference, grp)
   WHERE  c.rn = 1
   AND    c.date_out >= '2012-10-01'
   AND    c.date_out <  '2012-11-01'
   AND    d.rn = 2
   AND    d.centreid = 1
   )
SELECT b.taskid, b.reference, b.centreid, b.date_out
FROM   b
JOIN   d USING (reference, grp)
ORDER  BY reference, date_out

-> sqlfiddle

But while pure SQL is a beaut .. I would solve this procedurally in a plpgsql function. Very similar to this recent answer on SO. Would be faster, because it can be done in a single table scan.

Postgresql – Updating table of versioned rows with historical records in PostgreSQL

1st case

You seem to forget the valid_during range. As your third case suggests, there can be multiple entries per (rec_id, val), so you must select the right one:

UPDATE master m
SET    valid_on = f_array_sort(m.valid_on || u.valid_on) -- sorted array, see below
FROM   updates u
WHERE  m.rec_id = u.rec_id 
AND    m.valid_during @> u.valid_on  -- additional check
AND    m.val = u.val 
AND    NOT m.valid_on @> ARRAY[u.valid_on];

I assume the whole possible date range is always covered for each existing rec_id and valid_during shall not overlap per rec_id, or you'd have to do more.

After installing the additional module btree_gist, add an exclusion constraint to rule out overlapping date ranges if you don't have one, yet:

ALTER TABLE master ADD CONSTRAINT EXCLUDE
USING gist (rec_id WITH =, valid_during WITH &&)  -- disallow overlap

The GiST index this is implemented with is also a perfect match for the query. Details:

2nd / 3rd case

Assuming that every date range starts with the smallest date in the (now sorted!) array: lower(m.valid_during) = m.valid_on[1]. I would enforce that with a CHECK constraint.

Here we need to create one or two new rows In the 2nd case it is enough to shrink the range of the old row and insert one new row In the 3rd case we update the old row with the left half of array and range, insert the new row and finally insert the with the right half of array and range.

Helper functions

To keep it simple I introduce a new constraint: every array is sorted. Use this helper function

-- sort array
CREATE OR REPLACE FUNCTION f_array_sort(anyarray) 
  RETURNS anyarray LANGUAGE sql IMMUTABLE AS
$$SELECT ARRAY (SELECT unnest($1) ORDER BY 1)$$;

I don't need your helper function arraymin() any more, but it could be simplified to:

CREATE OR REPLACE FUNCTION f_array_min(anyarray) 
  RETURNS anyelement LANGUAGE sql IMMUTABLE AS
$$SELECT min(a) FROM unnest($1) a$$;

Two more to get the left and right half of an array split at a given element:

-- split left array at given element
CREATE OR REPLACE FUNCTION f_array_left(anyarray, anyelement) 
  RETURNS anyarray LANGUAGE sql IMMUTABLE AS
$$SELECT ARRAY (SELECT * FROM unnest($1) a WHERE a < $2 ORDER BY 1)$$;

-- split right array at given element
CREATE OR REPLACE FUNCTION f_array_right(anyarray, anyelement) 
  RETURNS anyarray LANGUAGE sql IMMUTABLE AS
$$SELECT ARRAY (SELECT * FROM unnest($1) a WHERE a >= $2 ORDER BY 1)$$;

Query

This does all the rest:

WITH u AS (  -- identify candidates
   SELECT m.id, rec_id, m.val, m.valid_on, m.valid_during
        , u.val AS u_val, u.valid_on AS u_valid_on
   FROM   master  m
   JOIN   updates u USING (rec_id)
   WHERE  m.val <> u.val
   AND    m.valid_during @> u.valid_on
   FOR    UPDATE  -- lock for update
   )
, upd1 AS (  -- case 2: no overlap, no split
   UPDATE master m  -- shrink old row
   SET    valid_during = daterange(lower(u.valid_during), u.u_valid_on)
   FROM   u
   WHERE  u.id = m.id
   AND    u.u_valid_on > m.valid_on[array_upper(m.valid_on, 1)]
   RETURNING m.id
   )
, ins1 AS (  -- insert new row
   INSERT INTO master (rec_id, val, valid_on, valid_during)
   SELECT u.rec_id, u.u_val, ARRAY[u.u_valid_on]
        , daterange(u.u_valid_on, upper(u.valid_during))
   FROM   upd1
   JOIN   u USING (id)
   )
, upd2 AS (  -- case 3: overlap, need to split row
   UPDATE master m  -- shrink to first half
   SET    valid_during = daterange(lower(u.valid_during), u.u_valid_on)
        , valid_on = f_array_left(u.valid_on, u.u_valid_on)
   FROM   u
   LEFT   JOIN upd1 USING (id)
   WHERE  upd1.id IS NULL  -- all others
   AND    u.id = m.id
   RETURNING m.id, f_array_right(u.valid_on, u.u_valid_on) AS arr_right
   )
INSERT INTO master (rec_id, val, valid_on, valid_during)
          -- new row
SELECT u.rec_id, u.u_val, ARRAY[u.u_valid_on]
     , daterange(u.u_valid_on, upd2.arr_right[1])
FROM   upd2
JOIN   u USING (id)
UNION ALL  -- second half of old row
SELECT u.rec_id, u.val, upd2.arr_right
     , daterange(upd2.arr_right[1], upper(u.valid_during))
FROM   upd2
JOIN   u USING (id);

SQL Fiddle.

Notes

You need to understand the concept of data-modifying CTEs (writeable CTEs), before you touch this. Judging from the code you provided, you know your way around Postgres.
- Are SELECT type queries the only type that can be nested?
FOR UPDATE is to avoid race conditions with concurrent write access. If you are the only user writing to the tables, you don't need it.
I took a piece of paper and drew a timeline so not to get lost in all of this.
Each row is only updated / inserted once, and operations are simple and roughly optimized. No expensive window functions. This should perform well. Much faster than your previous approach in any case.
It would be a bit less confusing if you'd use distinct column names for u.valid_on and m.valid_on, which are related but different things.
I compute the right half of the split array in the RETURNING clause of CTE upd2: f_array_right(u.valid_on, u.u_valid_on) AS arr_right, because I need it several times in the next step. This is a (legal) trick to save one more CTE.
As for solutions that don't involve unnesting the master table: You have to unnest the array valid_on either way, to split it, at least as long as it's not sorted. Also, your helper function arraymin() is already unnesting it anyway.