Обсуждение: BUG #17329: Aggregate Functions Precision Error

The following bug has been logged on the website:

Bug reference:      17329
Logged by:          Max Neverov
Email address:      neverov.max@gmail.com
PostgreSQL version: 13.3
Operating system:   Alpine 10.3.1_git20210424
Description:

Aggregate functions (described here
https://www.postgresql.org/docs/current/functions-aggregate.html#FUNCTIONS-AGGREGATE-STATISTICS-TABLE)
that are defined for double precision type suffer from loss of
significance.
Corresponding code see

https://github.com/postgres/postgres/blob/49407dc32a2931550e4ff1dea314b6a25afdfc35/src/backend/utils/adt/float.c#L3401.

Consider the following:
drop table if exists test;
create table test(y numeric, x numeric);
insert into test values
(1, 3),
(2, 3.0e+22),
(3, -3);
select covar_pop(y, x) from test;
     covar_pop
-------------------
 699050.6666666666
(1 row)

truncate table test;

insert into test values
(1, 3),
(3, -3),
(2, 3.0e+22);
select covar_pop(y, x) from test;
 covar_pop
-----------
        -2
(1 row)

truncate table test;

insert into test values
(2, 3.0e+22),
(3, -3),
(1, 3);
select covar_pop(y, x) from test;
     covar_pop
--------------------
 -699050.6666666666
(1 row)

The expected result is -2. 
The result depends on the order of values although it shouldn't. This
happens because operations with 3.0e+22 lead to the loss of precision since
the type can hold only 15 decimal digits precision. 
Even if the functions defined for double precision type I would expect
Postgres either to report an error or to return the correct result.

PG Bug reporting form <noreply@postgresql.org> writes:
> Aggregate functions (described here
> https://www.postgresql.org/docs/current/functions-aggregate.html#FUNCTIONS-AGGREGATE-STATISTICS-TABLE)
> that are defined for double precision type suffer from loss of
> significance.

This is pretty much inherent in all uses of float arithmetic.
You might be happier using the numeric type (of course, that's
much slower).

Another possibility, for some aggregates, is to order the inputs
in a way that minimizes error accumulation.  For example,

select sum(f1 order by abs(f1)) from ...

I don't know offhand what the best such incantation is for covar_pop;
it might depend on the problem.

            regards, tom lane