Обсуждение: Implementing product-aggregate

Hi,

we need a product aggregate and used to implement this as

exp(sum(ln([COLUMN])))

While using the sum of logarithms is working RDBMS-independently, we'd like to switch to a more PostgreSQL native way of doing this and implement an aggregate to be used. Currently the implementation is

create aggregate prod(numeric) (sfunc = numeric_mul, stype = numeric)

This is simply calling the implementation funtion of the *-Operator for the numeric datatype. Since I could not find any documentation of this implementation function, I am wondering, if using a possibly internal function might be a bad idea.

Are there any recommendations on this?

Thanks for any input!

--

MfG Jan

Hi again,

Am Montag, 11. März 2024, 10:16:33 CET schrieb Jan Kohnert:

> create aggregate prod(numeric) (sfunc = numeric_mul, stype = numeric)

my basic idea was creating a function

create function multiply(arg1 numeric, arg2 numeric)

    returns numeric

    language sql

    immutable

    returns null on null input

    return arg1 * arg2;

and use that function instead of the undocumented numeric_mul as the sfunc in the aggregate definition.

Then again, this seems odd, too, since we're only reimplementing basic stuff that's already there.

I'm still undecided...

--

MfG Jan

Jan Kohnert <nospam001-lists@jan-kohnert.de> writes:
> Am Montag, 11. März 2024, 10:16:33 CET schrieb Jan Kohnert:
>> create aggregate prod(numeric) (sfunc = numeric_mul, stype = numeric)
> ...
> Then again, this seems odd, too, since we're only reimplementing basic stuff that's
> already there.

I wouldn't be concerned about relying on numeric_mul (or any of the
other functions underlying standard operators).  They're undocumented
only because documenting both the functions and the operators would
bloat the documentation to little purpose.  Using one makes your code
not so portable to non-Postgres DBMSes, but you already crossed that
bridge by deciding to use a custom aggregate.

A bigger question is whether this implementation actually has the
properties you want --- notably, maybe you should be using type
float8 not numeric.  Numeric would get pretty slow and be carrying
an awful lot of decimal places by the end of the query, I fear.

            regards, tom lane

Hello,

Am Donnerstag, 14. März 2024, 15:17:58 CET schrieb Tom Lane:

> I wouldn't be concerned about relying on numeric_mul (or any of the

> other functions underlying standard operators).  They're undocumented

> only because documenting both the functions and the operators would

> bloat the documentation to little purpose.  Using one makes your code

> not so portable to non-Postgres DBMSes, but you already crossed that

> bridge by deciding to use a custom aggregate.

thank you for clearifying this. We're not too concerned about portability. Let's face the facts: Porting a reasonably complex database and the application using it from one DBMS to another will almost certainly introduce an awful lot of portability issues (f.e. the pseudo-types (big)serial, upserts, differences in merge implementations, progammability, and so on). My main concern was, that undocumented features sometimes tend to change without notice, since users are not expected to use them..

> A bigger question is whether this implementation actually has the

> properties you want --- notably, maybe you should be using type

> float8 not numeric.  Numeric would get pretty slow and be carrying

> an awful lot of decimal places by the end of the query, I fear.

This needs to be checked on our side. I was expecting, that using an aggregate this way would be significantly faster than using exp(sum(log())). Though we're not multiplying too many lines in a statement, if using the aggregate slows down performance, we should propably stick the old way doing it.

Best regards!

--

MfG Jan