Howto: SQL query to obtain latest n rows

Question

Hey everyone,
I want to query a subset of a table to return, at most, a certain number of rows. The problem is that 'rowcnt' is not going to be good enough. Here's the situation:
I have a table with lots of data (15 days worth). It will accumulate over time, but I want to extract, say, only the last 10 rows that were placed into the table. That would be simple, but there is one other catch, there is a key column and I may want to select rows from each "key". So a simple row count won't work :(.
Here's a sample table:
key         varchar(10)
val1        varchar(5)
val2        varchar(5)
epoch_time  number(30)Now, this would ALMOST work:
select key, val1, val2, epoch_time from t where key='KEYVAL1' and rowcnt <= 10 order by epoch_time desc;
Because the decending sort on epoch_time guarentees that the latest 10 rows will be returned. The problem arises when I want this:
select key, val1, val2, epoch_time from t where key IN ('KEYVAL1', 'KEYVAL2') and rowcnt <= 10 order by epoch_time desc;
I intend to get 10 rows (or less) returned for each 'KEYVAL1' and 'KEYVAL2', but since rowcnt is a global count of rows returned this obviously won't work as desired.
Keep in mind that I could have 5 rows, 10 rows, 1000 rows for a key...so I could be selecting three keys, one with 5 rows, the next with 40 and the next with 7. Therefore, if I did something like assume I'll get back 30 rows (10 per key) this situtation would render that solution invalid because I'd get 5 from the first and 25 from the second and none from the 3rd.
So, the actual question is, can I make some modification to the above query such that I can get the last 10 rows from both 'KEYVAL1' and 'KEYVAL2'. Possibly changing the order by clause to order by both key and epoch_time and then somehow switching the rowcnt to a value that would be capable of returning itself to 0 upon encountering a new key value?? I think I could do this with PL/SQL, but I'd have to use a static value to hold the row count and last passed key value, and that wouldn't work if more than one query executed at once unless there were some kind of "query specific data" (analogous to thread specific data) that I could utilize.
If not, I guess I'll have to write a PL/SQL routine to simulate the entire query because I can't use unions (at least I don't think I can) because I have to be capable of using an IN list to determine the key values I desire.
If anyone has a suggestion, I'd really appreciate it. THANKS!
Brian Gregg
Computer Sciences Corp.

Answer 1

Are you running with 8.1.6 Enterprise Edition or higher? If so, you can solve your problem with Analytic functions, specifically RANK() or DENSE_RANK().
SELECT * FROM
( SELECT key, val1, val2, epoch_time,
       DENSE_RANK() OVER (PARTITION BY key ORDER BY epoch_time DESC) AS ep_rank
FROM   t
WHERE  key IN ('KEYVAL1', 'KEYVAL2')
) WHERE ep_rank <= 10;Caveat: this code is typed without benefit of database so you may need to check the syntax.
Cheers, APC