Possible new percentile function?

Question

Hi,
I'm running "Oracle Database 11g Enterprise Edition Release 11.2.0.3.0 - 64bit Production".
I need to calculate percentiles, so I believe I have two choices: PERCENTILE_CONT() or PERCENTILE_DISC().
I also need to get the same results as the SPSS and SAS tools.
Comparing the results, neither of the Oracle versions give the same result as SPSS/SAS. I think this is due to how Oracle determines what row numbers to use.
Oracle takes row RN = (1+ (P*(N-1)) where P is percentile and N is number of ordered rows.
For PERCENTILE_CONT(0.1) from a 16 row set that would give row 2.5 as the starting point.
http://docs.oracle.com/cd/B28359_01/server.111/b28286/functions115.htm
SPSS and SAS uses the formula RN = P*(N+1) which gives row 1.7 for a 16 row set.
Now, would it be possible for me to write a PERCENTILE_SPSS() function?
The samples of analytical/pipelined functions I have seen (and written) are happy to use the rows in the set one by one, to calculate some special min/max/avg/etc.
For this function I would need either
1 - to know the number of rows in the set, from the first call of my function, and calculate the percentile as my function "passes" the desired rows
or
2 - save all values in the set and calculate the percentile at the end
The first sounds impossible, the second sounds expensive for a large table.
Any suggestions?
Kind regards
Tomas

Answer 1

You'll be more successful in the forums if you focus on the issues, questions and suggestions rather than make assumptions about why those questions or suggestions are being asked or made. Your unwarranted personal comments are unprofessional and indicate a lack of maturity.
My questions were, and are, solely to elicit information about your actual issue so that suitable solutions or workarounds can be determined. There was no 'lecture' anywhere in any of the information that I provided; just statements and quotes from experts.
The articles and quotes I provided were intended to show you that there isn't a universal agreement on a definition for those 'percentile' calculations. That would signal at least a 'caution' flag.
>
I'm asking if it's possible to write a custom-made analytical function when the formula needs to know the record count.
>
Yes - it is possible. In my opinion it isn't feasible. Any attempt to reproduce the exact proprietary algorithm that any vendor uses would require knowledge about the internals of that vendor's code that jsut aren't likely to be known or available: what precision is being used for intermediate results? What rounding methods are being used? What is the exact order of operations. Were any of those factors changed between versions. Did the vendor make any bug fixes?
Any testing you do with any manageable set of test data won't ensure that there aren't data sets that will produce slightly different results.
So writing such a function is one thing; trying to duplicate existing functionality is something else. The vendor's own implementation may have changed due to bug fixes so the results from that vendor may depend on the actual version being used.
>
Oracle does not have to match, I've never said that.
My report, running in Oracle Reports, do need to match the results of an old report.
>
Well I read that as saying that Oracle does need to match. Your second sentence pretty much says exactly that.
Again, my question was to try to understand why Oracle needed to produce a NEW report from OLD data that matches information in a PRINTED report from the past. That is a very unusual requirement. I've never run across such a requirement in more than 30 years and I have written thousands of reports using Crystal Reports (now BO), Oracle Reports, RPG and many other tools.
It is often not possible to reproduce ANY report from the past due to the ever-changing nature of the data in the database. Data gets updated, deleted, inserted even when it shouldn't.
>
I have considered typing in the old values and selecting them.
>
That is the only method guaranteed to produce the same results. And that is about the only method that will pass the stringent auditing and Sarbanes-Oxley requirements that many organizations have to abide by. Those auditors won't allow you to just change the data to make things come out the way you want. You have to PROVE the provenance of the data: where did every piece of data come from, what changes were made to it (and when and by whome) and how were any accumulations performed.
Using a custom function and merely showing that the results, on paper, match would not sway any auditor I've ever had to deal with. You may not have an issue of having to prove the data. That's why we are asking questions. To try to understand what options are viable for you.
>
That would mean I'd have to use SPSS to produce future values to store.
>
Yep! That's one of the main drivers for a lot of the ETL processes that exist. There are multiple systems of record. A PeopleSoft system might be the SOR for Inventory while an Oracle financials system might be the SOR for the financial and payroll data. ETL tries to merge and match the data. Trying to keep those ETL processes up to date with the changes going on in all of the source systems is a MAJOR headache.
The standard solutiion for that use case is to select an SOR for reporting and then create ETL processes that bring the SOR data from each source system to the reporting SOR. For your use case that might mean using SPSS to produce data in report-ready tables and then use ETL to bring that data to the report server where it can be merged with data from other systems.
>
That's just too much manual labour for my taste.
>
Welcome to the real world!