Oracle - grouping between pairs of records

Question

I have a logging table that contains data that looks like this:

ID     MSG                DATE
---------------------------------------------
1      TEst               2010-01-01 09:00:00
2      Job Start          2010-01-01 09:03:00
3      Do something       2010-01-01 09:03:10
4      Do something else  2010-01-01 09:03:12
5      Do something       2010-01-01 09:04:19
6      Job End            2010-01-01 09:06:30
7      Job Start          2010-01-01 09:18:03
8      Do something       2010-01-01 09:18:17
9      Do other thing     2010-01-01 09:19:48
10     Job End            2010-01-01 09:20:27

It contains (among other things) messags written by the application. I am interested in producing a report of everything that gets written between a "Job Start" record and a "Job End" record, for all such pairs of "Job Start" and "Job End". Ideally the report would look like this:

BATCH_NUM   ID     MSG                DATE
---------------------------------------------------------         
1           3      Do something       2010-01-01 09:03:10
1           4      Do something else  2010-01-01 09:03:12
1           5      Do something       2010-01-01 09:04:19

2           8      Do something       2010-01-01 09:18:17
2           9      Do other thing     2010-01-01 09:19:48

(line break between batches added to make it easier to read)

The output report omits the "Job STart and "Job End" messages, as well as the "TEst" message (which exists outside a pair of "Job Start" and "Job End".

I'm not even sure where to begin writing this kind of query, or if it's better to go with PL/SQL for this. Heck, I'm not even sure if there's a proper technical term for what I'm trying to do. ;)

(oracle version is 10g)

Answer 1

I'm sure there's a more analytic-ish way to do this, but I'm cheating with scalar subqueries to build the windows.

SQL> select * from logging_table;

        ID MSG                            LOG_DT
---------- ------------------------------ -------------------
         1 TEst                           2010-01-01 09:00:00
         2 Job Start                      2010-01-01 09:03:00
         3 Do something                   2010-01-01 09:03:10
         4 Do something else              2010-01-01 09:03:12
         5 Do something                   2010-01-01 09:04:19
         6 Job End                        2010-01-01 09:06:30
         7 Job Start                      2010-01-01 09:18:03
         8 Do something                   2010-01-01 09:18:17
         9 Do other thing                 2010-01-01 09:19:48
        10 Job End                        2010-01-01 09:20:27

SQL> l
  1      select dense_rank() over (order by job_start_id) as batch, 
         -- ^-- this part gets the batch
  2             job_step_id, msg, log_dt
  3             -- nested select to filter out rows outside of the boundaries
  4        from (select *
  5                from (select id as job_step_id, msg, log_dt,
  6                             -- scalar subquery to get start of "window"
  7                             (select max(id)
  8                                from logging_table
  9                               where msg = 'Job Start'
 10                                 and id < log.id) as job_start_id,
 11                             -- scalar subquery to get end of "window"
 12                             (select min(id)
 13                                from logging_table
 14                               where msg = 'Job End'
 15                                 and id > log.id) as job_end_id
 16                       from logging_table log
 17                      -- filter out the "window" rows themselves
 18                      where msg not in ('Job Start', 'Job End')
 19                     )
 20               -- the filtering out of "unbounded" records
 21               where job_start_id is not null
 22                 and job_end_id is not null
 23             )
 24*      order by job_step_id
SQL> /

 BATCH JOB_STEP_ID MSG                            LOG_DT

---

     1           3 Do something                   2010-01-01 09:03:10
     1           4 Do something else              2010-01-01 09:03:12
     1           5 Do something                   2010-01-01 09:04:19
     2           8 Do something                   2010-01-01 09:18:17
     2           9 Do other thing                 2010-01-01 09:19:48

Answer 2

The following sample uses much the same idea as Adam Musch's answer, but gets the logging sets' start and end in a single select which is joined to the logging statement, rather than using scalar sub-queries.

I don't think DENSE_RANK() can be used any more directly because there is no column available to partition on.

Also, either of these solutions assume that the will never be overlapping sets of logs. If a second set starts before the first one ends, it's a whole new problem...

WITH logging_sets AS
     (SELECT DENSE_RANK () OVER (ORDER BY start_date) AS set_rank, start_date, end_date
        FROM (SELECT CASE msg
                        WHEN 'Job End'
                           THEN NULL
                        ELSE LEAD (log_dt, 1, NULL) OVER (ORDER BY log_dt)
                     END AS end_date, log_dt AS start_date, msg
                FROM logging_table lt
               WHERE msg IN ('Job Start', 'Job End') )
       WHERE msg = 'Job Start')
SELECT ls.set_rank, lt.ID, lt.msg, lt.log_dt
  FROM logging_table lt, logging_sets ls
 WHERE lt.log_dt > ls.start_date AND lt.log_dt < ls.end_date 
   AND msg NOT IN ('Job Start', 'Job End')
ORDER BY ls.set_rank, lt.log_dt;