Question

Let’s assume we have a table Maintenance

Customer LastLogin ActionType
1        12/1/2007 2
1        12/2/2007 2
etc.

We want a list of all customers who at any point during a given year had one or more uninterrupted sequences, 14 days long, of login with action type 2.

I can of course easily do this with code, and even have it be fairly quick over small sets. Is there a non-cursor way to do it in SQL?

Answer 1

This will select all customers with at least two consecutive actions of the same type.

WITH    rows AS 
        (
        SELECT  customer, action,
                ROW_NUMBER() OVER (PARTITION BY customer ORDER BY lastlogin) AS rn
        FROM    mytable
        )
SELECT  DISTINCT customer
FROM    rows rp
WHERE   EXISTS
        (
        SELECT  NULL
        FROM    rows rl
        WHERE   rl.customer = rp.customer
                AND rl.rn = rp.rn + 1
                AND rl.action = rp.action
        )

Here’s the more efficient query for just action 2:

WITH    rows AS 
        (
        SELECT  customer, ROW_NUMBER() OVER (PARTITION BY customer ORDER BY lastlogin) AS rn
        FROM    mytable
        WHERE   action = 2
        )
SELECT  DISTINCT customer
FROM    rows rp
WHERE   EXISTS
        (
        SELECT  NULL
        FROM    rows rl
        WHERE   rl.customer = rp.customer
                AND rl.rn = rp.rn + 1
        )

Update 2:

To select uninterrupted ranges:

WITH    rows AS 
        (
        SELECT  customer, action, lastlogin
                ROW_NUMBER() OVER (PARTITION BY customer ORDER BY lastlogin) AS rn
                ROW_NUMBER() OVER (PARTITION BY customer, action ORDER BY lastlogin) AS series
        FROM    mytable
        )
SELECT  DISTINCT customer
FROM    (
        SELECT  customer
        FROM    rows rp
        WHERE   action
        GROUP BY
                customer, actioncode, series - rn
        HAVING
                DETEDIFF(day, MIN(lastlogin), MAX(lastlogin)) >= 14
        ) q

This query calculates two series: one returns contiguous ORDER BY lastlogin, the second one partitions by action additionally:

action  logindate rn  series diff = rn - series
1       Jan 01    1   1      0
1       Jan 02    2   2      0
2       Jan 03    3   1      2
2       Jan 04    4   2      2
1       Jan 05    5   3      2
1       Jan 06    6   4      2

As long as the difference between the two schemes is the same, the series are uninterrupted. Each interruption breaks the series.

This means that the combination of (action, diff) defines the uninterrupted groups.

We can group by action, diff, find MAX and MIN within the groups and filter on them.

If you need to select 14 rows rather than 14 consecutive days, just filter on COUNT(*) instead of the DATEDIFF.