Pure SQL to partition date-time occurrences into non-overlapping windows?

Question

i've a question that so far, i've never been able to solve via a pure SQL option.
it's hard to explain in words, but it's something like this:
given a set of date-time, i would like to partition the data into non-overlapping windows of 30 minutes each.
the data is supposed to be partitioned into windows of 30 minutes, meaning when the data is within 30 minutes of the first occurrence, only the first occurrence will be returned. in the next second after the 30th minute, the record will be considered as the start of a new window and is also returned. so those data that occurs within the window period are suppressed. the first occurrence is not necessarily occurring on the 00th minute, so the window start will never be constant.
run the below query to look at the dummy data.
SELECT     'A' AS ID
          , TRUNC(SYSDATE) + 7 / 24 + 1 *(ROWNUM - 1) / 1440 AS datetime
      FROM DUAL
CONNECT BY ROWNUM <= 50
UNION ALL
SELECT     'A' AS ID
          , TRUNC(SYSDATE) + 9 / 24 + 8 / 1440 + 1 *(ROWNUM - 1) / 1440 AS datetime
      FROM DUAL
CONNECT BY ROWNUM <= 35
UNION ALL
SELECT     'B' AS ID
          , TRUNC(SYSDATE) + 7 / 24 + 5 *(ROWNUM - 1) / 1440 AS datetime
      FROM DUAL
CONNECT BY ROWNUM <= 15this is supposed to be the output.
ID     DATETIME
A     5/19/2010 07:00:00
A     5/19/2010 07:30:00
A     5/19/2010 09:08:00
A     5/19/2010 09:38:00
B     5/19/2010 07:00:00
B     5/19/2010 07:30:00
B     5/19/2010 08:00:00so far, i'm using a PL/SQL to pipe the records. but i would like to know if this is achievable via SQL or not.
i've tried looking at analytics, width_bucket, ntile and alll options i can think of, but i just can't solve this at all.

Answer 1

hey Bob,
your answer is most definitely correct and does what i want. i've verified it again my data set and it returns the results as required!
you've definitely proven me wrong. i was always under the impression that this wasn't possible. thanks!
just a small note:
i need the windows to be binned by seconds, so have changed the numtodsinterval to raw numbers.
WITH t AS
     (SELECT 'A' AS ID
            , TRUNC(SYSDATE) +(6.75 / 24) AS datetime
        FROM DUAL
      UNION ALL
      SELECT 'A' AS ID
            , TRUNC(SYSDATE) +(6.75 / 24) AS datetime
        FROM DUAL
      UNION ALL
      SELECT     'A' AS ID
                , TRUNC(SYSDATE) + 7 / 24 + 1 *(ROWNUM - 1) / 1440 AS datetime
            FROM DUAL
      CONNECT BY ROWNUM <= 50
      UNION ALL
      SELECT     'A' AS ID
                , TRUNC(SYSDATE) + 9 / 24 + 8 / 1440 + 1 *(ROWNUM - 1) / 1440 AS datetime
            FROM DUAL
      CONNECT BY ROWNUM <= 35
      UNION ALL
      SELECT     'B' AS ID
                , TRUNC(SYSDATE) + 7 / 24 + 5 *(ROWNUM - 1) / 1440 AS datetime
            FROM DUAL
      CONNECT BY ROWNUM <= 15)
    ,a AS
     (SELECT ID
            ,datetime
            ,LAG(datetime) OVER(PARTITION BY ID ORDER BY datetime) AS prevtime
            ,LAST_VALUE(datetime) OVER(PARTITION BY ID ORDER BY datetime RANGE BETWEEN CURRENT ROW AND 30 / 1440 + 1 / 86400 FOLLOWING) AS interval_end
        FROM t)
    ,b AS
     (SELECT ID
            ,datetime
            ,LEAD(datetime) OVER(PARTITION BY ID ORDER BY datetime) AS nexttime
        FROM t)
    ,ab AS
     (SELECT a.ID
            ,a.datetime
            ,a.prevtime
               ,a.interval_end
               ,b.datetime as b_datetime
            ,b.nexttime
        FROM a JOIN b ON(a.ID = b.ID
                         AND a.interval_end = b.datetime)
SELECT     ID
          ,datetime
      FROM ab
START WITH prevtime IS NULL
CONNECT BY ID = PRIOR ID
       AND datetime = PRIOR nexttime
  ORDER BY ID
          ,datetime;this most definitely proves that i'm still not sure of how to use hierarchy queries.
Edited by: casey on May 20, 2010 11:20 AM