How to Efficiently Sample a Fixed Number of Rows

Question

Good afternoon. I need to select a specific number of random rows from a table, and while I believe my logic is right it's taking too long, 30 minutes for a routine data size. Hopefully someone here can show me a more efficient query. I've seen the SAMPLE function, but it just randomly selects rows on a one-by-one basis, without a guaranteed total count.
This is the idea:
INSERT INTO Tmp_Table (Value, Sequence) SELECT Value FROM Perm_Table, DBMS_RANDOM.VALUE;
SELECT Value FROM Tmp_Table WHERE ROWNUM <= 1234 ORDER BY Sequence;I'd need to put the ORDER BY in a subselect for ROWNUM to work correctly, but anyway that's just an illustration. My actual need is a little more complicated. I have many sets of data; each set has many rows; and for each set I need to return a specific (different) number of rows. Perhaps project A has three rows in this table, and I want to keep two of them; project B has two rows, and I want to keep one of them. So I need to identify, for each row, whether it's valid for that project. This is what my data looks like:
Project Person  Sequence Position Keeper
A       Bill    1234     1        Yes
A       Fred    5678     3        No
A       George  1927     2        Yes
B       April   5784     2        No
B       Janice  2691     1        YesI populate Sequence with random values, then calculate the position of each person within their project, and finally discard people who's Position is greater than Max_Targets for the Project. Fred and April have the highest random numbers, so they're cut. It's not the case that I'm just trimming one person from each project; the actual percentage kept will range from zero to 100.
Populating the list with random values is not time-consuming, but calculating Position is. This is my code:
UPDATE Tmp_Targets T1
SET Position =
  SELECT
   COUNT(*)
  FROM
   Perm_Targets PT1
   INNER JOIN Perm_Targets PT2 ON PT1.Project = PT2.Project
   INNER JOIN Tmp_Targets T2 ON PT2.Target = T2.Target
  WHERE
       T1.Target = PT1.Target
   AND T2.Sequence <= T1.Sequence
  );The Target fields are PKs, and the Project and Sequence fields are indexed. Is there a better way to approach this? I could write a cursor that pulls out project codes and performs the above operations for each project in turn; that would be logically simpler and possibly faster. Has anyone here addressed a similar problem before? I'd appreciate any ideas.
This is on 9.2, in case it matters. Thank you,
Jonathan

Answer 1

You've not given any indication of how max targets for a given project is determined, so for my example I'm using the ceiling of 1/2 of the number of records in each project which gives the same number of yes and no responses per project as you had:
with dta as (
  select 'A' project, 'Bill' person from dual union all
  select 'A', 'Fred' from dual union all
  select 'A', 'George' from dual union all
  select 'B', 'April' from dual union all
  select 'B', 'Janice' from dual
), t1 as (
  select project
       , person
       , row_number() over (partition by project order by dbms_random.value) ord
       , count(*) over (partition by project) cnt
       , rownum rn
    from dta
select project
     , person
     , ord
     , cnt
     , case when ord <= ceil(cnt/2) then 'Yes' else 'No' end keep
  from t1
  order by rn
PROJECT PERSON ORD                    CNT                    KEEP
A       Bill   2                      3                      Yes 
A       Fred   3                      3                      No  
A       George 1                      3                      Yes 
B       April  1                      2                      Yes 
B       Janice 2                      2                      No  
5 rows selectedIn this example I use an analytic function to assign a random ordering for each record within a project in the middle query, in the final output query I am determining the yes no status based on the order within a project and the count of records in the project. If you had a table of projects indicating the thresh hold you could join to that and use the thresh hold in place of the ceil(cnt/2) portion of my inequality in the case statement.