Referencing current row in FILTER clause of window function

Question

In PostgreSQL 9.4 the window functions have the new option of a FILTER to select a sub-set of the window frame for processing. The documentation mentions it, but provides no sample. An online search yields some samples, including from 2ndQuadrant but all that I found were rather trivial examples with constant expressions. What I am looking for is a filter expression that includes the value of the current row.

Assume I have a table with a bunch of columns, one of which is of date type:

col1 | col2 |     dt
------------------------
  1  |  a   | 2015-07-01
  2  |  b   | 2015-07-03
  3  |  c   | 2015-07-10
  4  |  d   | 2015-07-11
  5  |  e   | 2015-07-11
  6  |  f   | 2015-07-13
...

A window definition for processing on the date over the entire table is trivially constructed: WINDOW win AS (ORDER BY dt)

I am interested in knowing how many rows are present in, say, the 4 days prior to the current row (inclusive). So I want to generate this output:

col1 | col2 |     dt     | count
--------------------------------
  1  |  a   | 2015-07-01 |   1
  2  |  b   | 2015-07-03 |   2
  3  |  c   | 2015-07-10 |   1
  4  |  d   | 2015-07-11 |   3
  5  |  e   | 2015-07-11 |   3
  6  |  f   | 2015-07-13 |   4
...

The FILTER clause of the window functions seems like the obvious choice:

count(*) FILTER (WHERE current_row.dt - dt <= 4) OVER win

But how do I specify current_row.dt (for lack of a better syntax)? Is this even possible?

If this is not possible, are there other ways of selecting date ranges in a window frame? The frame specification is no help as it is all row-based.

I am not interested in alternative solutions using sub-queries, it has to be based on window processing.

Answer 1

You are not actually aggregating rows, so the new aggregate FILTER clause is not the right tool. A window function is more like it, a problem remains, however: the frame definition of a window cannot depend on values of the current row. It can only count a given number of rows preceding or following with the ROWS clause.

To make that work, aggregate counts per day and LEFT JOIN to a full set of days in range. Then you can apply a window function:

SELECT t.*, ct.ct_last4days
FROM  (
   SELECT *, sum(ct) OVER (ORDER BY dt ROWS 3 PRECEDING) AS ct_last4days
   FROM  (
      SELECT generate_series(min(dt), max(dt), interval '1 day')::date AS dt
      FROM   tbl t1
      ) d
   LEFT   JOIN (SELECT dt, count(*) AS ct FROM tbl GROUP BY 1) t USING (dt)
   ) ct
JOIN  tbl t USING (dt);

Omitting ORDER BY dt in the widow frame definition usually works, since the order is carried over from generate_series() in the subquery. But there are no guarantees in the SQL standard without explicit ORDER BY and it might break in more complex queries.

SQL Fiddle.

Related:

• Select finishes where athlete didn't finish first for the past 3 events
• PostgreSQL: running count of rows for a query 'by minute'
• PostgreSQL unnest() with element number