Обсуждение: BUG #17546: power() function - value is distorted via automatic type cast

The following bug has been logged on the website:

Bug reference:      17546
Logged by:          Su Sinodan
Email address:      su.sinodan@gmail.com
PostgreSQL version: 14.2
Operating system:   Windows Server 2016 (10.0.14393 Build 1439)
Description:

When a certain number (p, s) with a fractional part (specific examples -
1.11 and 69.96) is multiplied by a power (10, N), the resulting value is
distorted.
However, if the power(10, N) is manually converted to an integer, the
resulting value is correct.
It looks like there is some kind of problem with the automatic type
conversion of the power() function.

Test query:
with w_degree as (
                   select 0::integer as degree union all
                   select 1::integer as degree union all
                   select 2::integer as degree union all
                   select 3::integer as degree union all
                   select 4::integer as degree union all
                   select 5::integer as degree union all
                   select 6::integer as degree union all
                   select 7::integer as degree union all
                   select 8::integer as degree union all
                   select 9::integer as degree
                 )
select
  power(10, w.degree::integer)                                  as
"power()"
, 1.11::numeric(20, 5) * power(10, w.degree::integer)           as " 1.11 *
power()"
, 1.11::numeric(20, 5) * power(10, w.degree::integer)::integer  as " 1.11 *
power()::integer"
, 69.96::numeric(20, 5) * power(10, w.degree::integer)          as "69.96 *
power()"
, 69.96::numeric(20, 5) * power(10, w.degree::integer)::integer as "69.96 *
power()::integer"
from w_degree w;

Result:
  power()   |   1.11 * power()   |  1.11 * power()::integer |  69.96 *
power()  | 69.96 * power()::integer
------------+--------------------+--------------------------+-------------------+--------------------------
          1 |               1.11 |                  1.11000 |
69.96 |                 69.96000
         10 | 11.100000000000001 |                 11.10000 |
699.5999999999999 |                699.60000
        100 | 111.00000000000001 |                111.00000 |
6995.999999999999 |               6996.00000
       1000 |               1110 |               1110.00000 |
69960 |              69960.00000
      10000 | 11100.000000000002 |              11100.00000 |
699599.9999999999 |             699600.00000
     100000 | 111000.00000000001 |             111000.00000 |
6995999.999999999 |            6996000.00000
    1000000 |            1110000 |            1110000.00000 |
69960000 |           69960000.00000
   10000000 | 11100000.000000002 |           11100000.00000 |
699599999.9999999 |          699600000.00000
  100000000 | 111000000.00000001 |          111000000.00000 |
6995999999.999999 |         6996000000.00000
 1000000000 |         1110000000 |         1110000000.00000 |
69960000000 |        69960000000.00000
(10 rows)

PG Bug reporting form <noreply@postgresql.org> writes:
> When a certain number (p, s) with a fractional part (specific examples -
> 1.11 and 69.96) is multiplied by a power (10, N), the resulting value is
> distorted.
> However, if the power(10, N) is manually converted to an integer, the
> resulting value is correct.
> It looks like there is some kind of problem with the automatic type
> conversion of the power() function.

Your query is invoking the float8 variant of power(), which is
necessarily of limited precision.  If I change "w.degree::integer"
to "w.degree::numeric", then it invokes the numeric variant of
power(), and I get correctly rounded results (much more slowly
unfortunately :-().  So I don't see any bug here.

It might be surprising that the parser prefers float8 over numeric
when the given function arguments don't exactly match either one.
But that's a very ancient decision that we're not going to change.
There is support for it in the SQL standard, which directs
implementations to prefer inexact numeric types over exact ones
when they have to make a choice.

            regards, tom lane

=?utf-8?B?0JDQvdGC0L7QvSDQn9C+0YLQsNC/0L7Qsg==?= <su.sinodan@gmail.com> writes:
> Is there any chance that you will reflect this in the documentation?

It is documented --- see under
    https://www.postgresql.org/docs/current/datatype-numeric.html
and
    https://www.postgresql.org/docs/current/typeconv.html
(particularly example 10.3, which points out that float8 is
considered a preferred type).  We are not going to repeat all that
material under every single affected function, if that's what
you're asking for.

            regards, tom lane