Can window function LAG reference the column which value is being calculated?

Question

I need to calculate value of some column X based on some other columns of the current record and the value of X for the previous record (using some partition and order). Basically I need to implement query in the form

SELECT <some fields>, 
  <some expression using LAG(X) OVER(PARTITION BY ... ORDER BY ...) AS X
FROM <table>

This is not possible because only existing columns can be used in window function so I'm looking way how to overcome this.

Here is an example. I have a table with events. Each event has type and time_stamp.

create table event (id serial, type integer, time_stamp integer);

I wan't to find "duplicate" events (to skip them). By duplicate I mean the following. Let's order all events for given type by time_stamp ascending. Then

1. the first event is not a duplicate
2. all events that follow non duplicate and are within some time frame after it (that is their time_stamp is not greater then time_stamp of the previous non duplicate plus some constant TIMEFRAME) are duplicates
3. the next event which time_stamp is greater than previous non duplicate by more than TIMEFRAME is not duplicate
4. and so on

For this data

insert into event (type, time_stamp) 
 values 
  (1, 1), (1, 2), (2, 2), (1,3), (1, 10), (2,10), 
  (1,15), (1, 21), (2,13), 
  (1, 40);

and TIMEFRAME=10 result should be

time_stamp | type | duplicate
-----------------------------
        1  |    1 | false
        2  |    1 | true     
        3  |    1 | true 
       10  |    1 | true 
       15  |    1 | false 
       21  |    1 | true
       40  |    1 | false
        2  |    2 | false
       10  |    2 | true
       13  |    2 | false

I could calculate the value of duplicate field based on current time_stamp and time_stamp of the previous non-duplicate event like this:

WITH evt AS (
  SELECT 
    time_stamp, 
    CASE WHEN 
      time_stamp - LAG(current_non_dupl_time_stamp) OVER w >= TIMEFRAME
    THEN 
      time_stamp
    ELSE
      LAG(current_non_dupl_time_stamp) OVER w
    END AS current_non_dupl_time_stamp
  FROM event
  WINDOW w AS (PARTITION BY type ORDER BY time_stamp ASC)
)
SELECT time_stamp, time_stamp != current_non_dupl_time_stamp AS duplicate

But this does not work because the field which is calculated cannot be referenced in LAG:

ERROR:  column "current_non_dupl_time_stamp" does not exist.

So the question: can I rewrite this query to achieve the effect I need?

Erst

I couldn't understand the timeframe part. specially this part: the next event which time_stamp if greater than previous non duplicate by more than TIMEFRAME is not duplicate. is timeframe a constant, a field, a calculation ? – Incurvate

TIMEFRAME is some constant. The rationale is I want to skip event if it comes within given time frame after previous event which was not skipped. – Erst

Your desired output contains timestamp 40, but your example data set does not? Could you clarify? – Nahshun

You are right, it was a mistake. – Erst

Answer 1

An alternative to a recursive approach is a custom aggregate. Once you master the technique of writing your own aggregates, creating transition and final functions is easy and logical.

State transition function:

create or replace function is_duplicate(st int[], time_stamp int, timeframe int)
returns int[] language plpgsql as $$
begin
    if st is null or st[1] + timeframe <= time_stamp
    then 
        st[1] := time_stamp;
    end if;
    st[2] := time_stamp;
    return st;
end $$;

Final function:

create or replace function is_duplicate_final(st int[])
returns boolean language sql as $$
    select st[1] <> st[2];
$$;

Aggregate:

create aggregate is_duplicate_agg(time_stamp int, timeframe int)
(
    sfunc = is_duplicate,
    stype = int[],
    finalfunc = is_duplicate_final
);

Query:

select *, is_duplicate_agg(time_stamp, 10) over w
from event
window w as (partition by type order by time_stamp asc)
order by type, time_stamp;

 id | type | time_stamp | is_duplicate_agg 
----+------+------------+------------------
  1 |    1 |          1 | f
  2 |    1 |          2 | t
  4 |    1 |          3 | t
  5 |    1 |         10 | t
  7 |    1 |         15 | f
  8 |    1 |         21 | t
 10 |    1 |         40 | f
  3 |    2 |          2 | f
  6 |    2 |         10 | t
  9 |    2 |         13 | f
(10 rows)   

Read in the documentation: 37.10. User-defined Aggregates and CREATE AGGREGATE.

Economic

Answer 2

Naive recursive chain knitter:

---

        -- temp view to avoid nested CTE
CREATE TEMP VIEW drag AS
        SELECT e.type,e.time_stamp
        , ROW_NUMBER() OVER www as rn                   -- number the records
        , FIRST_VALUE(e.time_stamp) OVER www as fst     -- the "group leader"
        , EXISTS (SELECT * FROM event x
                WHERE x.type = e.type
                AND x.time_stamp < e.time_stamp) AS is_dup
        FROM event e
        WINDOW www AS (PARTITION BY type ORDER BY time_stamp)
        ;

WITH RECURSIVE ttt AS (
        SELECT d0.*
        FROM drag d0 WHERE d0.is_dup = False -- only the "group leaders"
    UNION ALL
        SELECT d1.type, d1.time_stamp, d1.rn
          , CASE WHEN d1.time_stamp - ttt.fst > 20 THEN d1.time_stamp
                 ELSE ttt.fst END AS fst   -- new "group leader"
          , CASE WHEN d1.time_stamp - ttt.fst > 20 THEN False
                 ELSE True END AS is_dup
        FROM drag d1
        JOIN ttt ON d1.type = ttt.type AND d1.rn = ttt.rn+1
        )
SELECT * FROM ttt
ORDER BY type, time_stamp
        ;

---

Results:

---

CREATE TABLE
INSERT 0 10
CREATE VIEW
 type | time_stamp | rn | fst | is_dup 
------+------------+----+-----+--------
    1 |          1 |  1 |   1 | f
    1 |          2 |  2 |   1 | t
    1 |          3 |  3 |   1 | t
    1 |         10 |  4 |   1 | t
    1 |         15 |  5 |   1 | t
    1 |         21 |  6 |   1 | t
    1 |         40 |  7 |  40 | f
    2 |          2 |  1 |   2 | f
    2 |         10 |  2 |   2 | t
    2 |         13 |  3 |   2 | t
(10 rows)

Strikebound

Answer 3

This feels more like a recursive problem than windowing function. The following query obtained the desired results:

WITH RECURSIVE base(type, time_stamp) AS (

  -- 3. base of recursive query
  SELECT x.type, x.time_stamp, y.next_time_stamp
    FROM 
         -- 1. start with the initial records of each type   
         ( SELECT type, min(time_stamp) AS time_stamp
             FROM event
             GROUP BY type
         ) x
         LEFT JOIN LATERAL
         -- 2. for each of the initial records, find the next TIMEFRAME (10) in the future
         ( SELECT MIN(time_stamp) next_time_stamp
             FROM event
             WHERE type = x.type
               AND time_stamp > (x.time_stamp + 10)
         ) y ON true

  UNION ALL

  -- 4. recursive join, same logic as base
  SELECT e.type, e.time_stamp, z.next_time_stamp
    FROM event e
    JOIN base b ON (e.type = b.type AND e.time_stamp = b.next_time_stamp)
    LEFT JOIN LATERAL
    ( SELECT MIN(time_stamp) next_time_stamp
       FROM event
       WHERE type = e.type
         AND time_stamp > (e.time_stamp + 10)
    ) z ON true

)

-- The actual query:

-- 5a. All records from base are not duplicates
SELECT time_stamp, type, false
  FROM base

UNION

-- 5b. All records from event that are not in base are duplicates
SELECT time_stamp, type, true
  FROM event
  WHERE (type, time_stamp) NOT IN (SELECT type, time_stamp FROM base) 

ORDER BY type, time_stamp

There are a lot of caveats with this. It assumes no duplicate time_stamp for a given type. Really the joins should be based on a unique id rather than type and time_stamp. I didn't test this much, but it may at least suggest an approach.

This is my first time to try a LATERAL join. So there may be a way to simplify that moe. Really what I wanted to do was a recursive CTE with the recursive part using MIN(time_stamp) based on time_stamp > (x.time_stamp + 10), but aggregate functions are not allowed in CTEs in that manner. But it seems the lateral join can be used in the CTE.

Alkanet