Question

I already saw this post here (http://stackoverflow.com/questions/1398113/sql-select-one-row-randomly-but-taking-into-account-a-weight), but couldnt work it out. Where do I put the ‘Stuff’ table? Why don’t they use NEWID() instead of RND()?

Table Stuff

id     item       weight       location
1      ball       1            Wyoming
2      cup        2            Alaska
3      sock       1            Idaho
4      car        3            Miami
5      hot girl   5            Brazil

Now according to that article referenced above I’m supposed to do this

SELECT      TOP 1 t.*
FROM        @Table t
INNER JOIN (SELECT t.id, sum(tt.weight) AS cum_weight
            FROM        @Table t
            INNER JOIN  @Table tt ON  tt.id <= t.id
            GROUP BY    t.id) tc
        ON  tc.id = t.id,
           (SELECT  SUM(weight) AS total_weight FROM @Table) tt,
           (SELECT  RAND() AS rnd) r
WHERE       r.rnd * tt.total_weight <= tc.cum_weight
ORDER BY    t.id ASC

I want to do the above, but in this fashion:

SELECT TOP (1) from stuff WHERE blahblahblah AND (location='Brazil' OR location='Wyoming' OR location='Brazil') AND (weight <= cum_weight) ORDER BY NEWID()

I’m only guessing I can use NEWID() and not compelled to use RND()

Answer 1

You could accomplish this by sampling from the cumulative sum, instead of from the records. The idea is to take a cumulative sum of the weights, then to take a random value up to the maximum weight, and finally to see which record has a cumulative sum around that random value. The SQL looks like:

select top 1 t.*
from (select t.*, cumulative_sum(weight) as cumweight,
             sum(weight) over (partition by NULL) as totalweight
      from t
     ) t
where rand()*(totalweight+1) < cumweight
order by cumweight desc

What this does is create a cumulative weight and then create a random variable up to the total sum of the weights. The last record whose cum weight is less than the sumweight is chosen. The “+1” is just to be sure that any record could be chosen, even the last.

In SQL Server 2012, you can calculate the cumulative SUM using SUM() over (partition by NULL order by ).

In SQL Server 2012, you might use:
select top 1 t.*
from (select t., sum(weight) over (partition by NULL order by weight) as cumweight,
sum(weight) over (partition by NULL) as totalweight
from t
) t
where rand()(totalweight+1) < cumweight
order by cumweight desc

Unfortunately, this syntax is not support in SQL Server 2008. In that database, you need to do a self-join, which is the query that you pulled from the original article.