On 2023-10-19 00:12:51 -0400, Tom Lane wrote: > =?GBK?B?w6vDqw==?= <kr...@163.com> writes: > > Do we have a rule by follow which one can accurately info the output of a > > SELECT statment FROM a table with floating-point data type? > > The goal of our floating-point output functions is to produce > the shortest representation from which the stored value could > be recovered exactly. It's not unusual for that to require > two or three decimal digits more than the nominal precision.
To illustrate this:
A real has 24 bits of mantissa. 12345.123456789 is between 2**13 and
2**14, so there are 10 bits left for the fraction. Or in other words,
the number must be approximated as a multiple of 1/1024.
The closest we can get is 12345+124/1024 = 12345.123046875
12345.123 would obviously be rounded to the same number, so it's close
enough and additional digits aren't necessary.
But 12345.12 would be rounded to 12345+123/1024 = 12345.1201171875.
That's different, so 7 digits are not enough in this case.
hp
--
_ | Peter J. Holzer | Story must make more sense than reality.
|_|_) | |
| | | h...@hjp.at | -- Charles Stross, "Creative writing
__/ | http://www.hjp.at/ | challenge!"
signature.asc
Description: PGP signature
