Question on Randomized allocation

Database
Enthusiast

Question on Randomized allocation

Hello,

I have a question concerning Randimized allocation:

I would like to split a customers' selection in two parts  (50% and 50%) with the condition that the two parts has to contain the same number of customers which belongs to the same Abglkey 480, 481 and 482.

Example 600 customers --->100 have 480 as Abglkey, 200 have 481 as Abglkey, 300 have 482 as Abglkey. The results must be:

            Table 1     Table 2

480      50             50

481      100           100

482      150           150

I wrote the following SQL

ins into BP_ATV.einsatz_115_S3

sel  REFNR, case when REFNR is not null then ' ' end Gutscheinnummer  , case when REFNR is not null then 'AKTIV3' end Kennzeichen

from BP_ATV.AKQ_CLEANS_TEST_115

where ABGLKEY in (480, 481, 482)

randomized allocation.5

order by Abglkey

The problem is that order by is not allowed in subqueries. I also found the statement randomized allocation when... but it's not perfect and when I have 100 Abglkey it's not easy to do.

Do you have an idea how can I do that?

Thank you,

Valeroa

5 REPLIES
Junior Contributor

Re: Question on Randomized allocation

Hi Valeria,

RANDOMIZED ALLOCATION is an option of the SAMPLE clause, but sampling only works on all rows. Plus it's random, not repeatable, so you can't do two Inserts and still get distinct sets.

if the number of ABGLKESs is small and known (like in your example) you can do

select ..., SAMPLEID as sid
from BP_ATV.AKQ_CLEANS_TEST_115
where ABGLKEY in (480, 481, 482)
SAMPLE RANDOMIZED ALLOCATION
WHEN ABGLKEY = 480 THEN 0.5,0.5
WHEN ABGLKEY = 481 THEN 0.5,0.5
WHEN ABGLKEY = 482 THEN 0.5,0.5
END

This will result in two 50% samples for each ABGLKEY, numbered from 1 to 6.

You must materialize this result in a table and when you need to select a single sample filter using 

WHERE sid MOD 1 = 0 (or 1)

For a larger number of ABGLKEYs (or unknown values) this will work:

SELECT ... ,
CASE WHEN PERCENT_RANK()
OVER (PARTITION BY ABGLKEY
ORDER BY rnd) < 0.5
THEN 0
ELSE 1
END AS sid
FROM
(
SELECT ... ,
RANDOM(-2000000000,2000000000) AS
from BP_ATV.AKQ_CLEANS_TEST_115
where ABGLKEY in (480, 481, 482)
) AS dt

This also creates two random groups each 50%.

Enthusiast

Re: Question on Randomized allocation

Hello Dieter,

thank you very much it works very well. The only thing that I noticed is that Teradata that in the end I have some customers more in the group 1 than in the group 0 because Teradata. This is due to Abglkey that contains odd numbers of customers. for eg:

Abglkey  sid 1         sid   0

1            10                 9

2            7                   6

3            5                   4

Is there a way that allow this 1 more in group 1 is assigned sometimes to 0 instead of to 1?

I hope it is clear. This is only for being more precise :) But it works very well you helped me a lot!

Thank you,

Valeria

Junior Contributor

Re: Question on Randomized allocation

Hi Valeria,

you could switch 0 and 1 like this (not random, but based on odd/even ABGLKEYs):

SELECT ... ,
(CASE WHEN PERCENT_RANK()
OVER (PARTITION BY ABGLKEY
ORDER BY rnd) < 0.5
THEN 0
ELSE 1
END + (ABGLKEY MOD 2)) MOD 2 AS sid
FROM
(
SELECT ... ,
RANDOM(-2000000000,2000000000) AS
from BP_ATV.AKQ_CLEANS_TEST_115
where ABGLKEY in (480, 481, 482)
) AS dt
Enthusiast

Re: Question on Randomized allocation

Hello,

I have a question concerning this skript. I elaborated this skirpt and adaped it to an excel file that I uploaded in teradata:

sel U_KDNR, U_KEY, CANTONE, Ort, PLZ, Strasse,  (CASE WHEN PERCENT_RANK()
             OVER (PARTITION BY U_KEY
                   ORDER BY  U_KEY, Ort ,CANTONE, PLZ, Strasse) < 0.5
                THEN 0
        ELSE 1
         END + (U_KEY MOD 2)) MOD 2 AS sid
     from (
  sel a.U_KEY,
CANTONE,
a.U_KDNR,
dec_postadr(mandant_id,U_ORT) Ort,
dec_postadr(mandant_id,U_PLZ ) PLZ,
dec_postadr(mandant_id,U_STRASSE) Strasse,
mandant_id,
TEST_KONTRAST_CHF_M07_M09,
      RANDOM(-2000000000,2000000000) as sid2
  from sch.BASE_TEST_KONTRAST_CHF_SK10_CRIP a
join sch.PLZ_CANTONI b on dec_postadr(mandant_id, a.U_PLZ )= b.PLZ
where TEST_KONTRAST_CHF_M07_M09  is null
 ) AS dt;

 

This skript gives me as results a split between sid 1 and 0 but if I let it run twice or theree times it gives me the same results. It woulb be great if you could help me to modify this skript so that, if we let it run many times the split between 1 and 0 changes BUT the changes should be slight because the split should respect the split according to U_KEY, Ort ,CANTONE, PLZ, Strasse. Do you think it is possible?

 

Thank you and best regards,

Valeria

Teradata Employee

Re: Question on Randomized allocation

The only thing I see that can change the split is the addition or deletion of rows in  sch.BASE_TEST_KONTRAST_CHF_SK10_CRIP.  You use a Random function in the derived table but the number it generates ("sid2") is not used anywhere else in the query.

 

Also, by the way, it is wasteful and distracting to include U_KEY in the ORDER BY clause of the Percent_Rank() function because you are already PARTITIONing by it - all the rows in any partition will have the same U_KEY!