Sum Interval of Dates within same column

Question

How do you best sum the differences of a range of dates in the same column between interleaving rows? I have a Datetime column and want to calculate the difference between Rows. I want the difference in Seconds. This questions is not about how to get a difference between 2 timestamps, but is more focused on how to most efficiently calculate between rows on the same table. In my case each row has a datetime eventype that links 2 rows together logically.

Details Related to how to group the eventtypes of start and end. ( Andriy M's Question) Starts and Ends "should" be consecutive. If a Start does not have a subsequent end it should be left out of the sum. Moving to the next Start to see if it has a end. Only consecutive Start - End Pairs should be added to the sum of the total seconds.

Working in postgresql 9.x...

Example data in table;

eventtype, eventdate
START, 2015-01-01 14:00
END, 2015-01-01 14:25
START, 2015-01-01 14:30
END, 2015-01-01 14:43
START, 2015-01-01 14:45
END, 2015-01-01 14:49
START, 2015-01-01 14:52
END, 2015-01-01 14:55

Note, All Start and End Dates will be sequential.

Here's my first attempt. Seems to be working.

SELECT 
-- starts.*
SUM(EXTRACT(EPOCH FROM (eventdate_next - eventdate))) AS duration_seconds
FROM
( 
    WITH x AS (
        SELECT *, dense_rank() OVER (ORDER BY eventdate) AS rnk
        FROM   table
        AND eventdate > '2015-01-01 00:00:00.00'
        AND eventdate < '2016-01-01 23:59:59.59' 
        )
    SELECT x.eventdate, x.eventtype, y.eventdate AS eventdate_next,  y.eventtype AS eventtype_next
    FROM   x
    LEFT   JOIN (SELECT DISTINCT eventdate, eventtype, rnk FROM x) y ON y.rnk = (x.rnk + 1)
    ORDER  BY x.eventdate
) starts
WHERE
eventtype = 'START'   
GROUP BY eventtype 

My first attempt is based on a great example from stackoverflow Postgres 9.1 - Getting the next value

Note; You can comment the GROUP BY and the SUM and un-comment the starts.* to get a record for each individual duration going into the sum.

Answer 1

You can use the LEAD analytic function to obtain the next row's eventtype and eventdate alongside the current row's data:

SELECT
  eventtype,
  eventdate,
  LEAD(eventtype) OVER (ORDER BY eventdate) AS nexttype,
  LEAD(eventdate) OVER (ORDER BY eventdate) AS nextdate
FROM
  atable
WHERE
      eventdate >= '2015-01-01 00:00:00.00'
  AND eventdate <  '2016-01-01 23:59:59.59'

Using the above query as a derived table, you can filter the output further on eventtype = 'START' AND nexttype = 'END' and get the difference total:

SELECT
  SUM(EXTRACT(EPOCH FROM (nextdate - eventdate))) AS duration_seconds
FROM
  (
    SELECT
      eventtype,
      eventdate,
      LEAD(eventtype) OVER (ORDER BY eventdate) AS nexttype,
      LEAD(eventdate) OVER (ORDER BY eventdate) AS nextdate
    FROM
      atable
    WHERE
          eventdate >= '2015-01-01 00:00:00.00'
      AND eventdate <  '2016-01-01 23:59:59.59'
  ) AS s
WHERE
      eventtype = 'START'
  AND nexttype  = 'END'
;

As a slight variation, you can implement the subquery as a CTE:

WITH cte AS
  (
    SELECT
      eventtype,
      eventdate,
      LEAD(eventtype) OVER (ORDER BY eventdate) AS nexttype,
      LEAD(eventdate) OVER (ORDER BY eventdate) AS nextdate
    FROM
      atable
    WHERE
          eventdate >= '2015-01-01 00:00:00.00'
      AND eventdate <  '2016-01-01 23:59:59.59'
  )
SELECT
  SUM(EXTRACT(EPOCH FROM (nextdate - eventdate))) AS duration_seconds
FROM
  cte
WHERE
      eventtype = 'START'
  AND nexttype  = 'END'
;

This rewrite can have implications for performance, because unlike a derived table, a CTE is materialised in PostgreSQL. Testing should reveal if there is a difference and, if so, which option is better for you.