Hi,
On 2019-05-20 23:56:33 +0100, Andrew Gierth wrote:
> [ To: header pruned ]
>
> >>>>> "Andres" == Andres Freund <andres@anarazel.de> writes:
>
> Andres> <para>
> Andres> Avoid performing unnecessary rounding of <link
> Andres> linkend="datatype-float"><type>REAL</type></link> and <type>DOUBLE
> Andres> PRECISION</type> values (Andrew Gierth)
> Andres> </para>
>
> Andres> <para>
> Andres> This dramatically speeds up processing of floating-point
> Andres> values but causes additional trailing digits to
> Andres> potentially be displayed. Users wishing to have output
> Andres> that is rounded to match the previous behavior can set <link
> Andres> linkend="guc-extra-float-digits"><literal>extra_float_digits=0</literal></link>,
> Andres> which is no longer the default.
> Andres> </para>
> Andres> </listitem>
>
> Andres> Isn't it exactly the *other* way round? *Previously* we'd
> Andres> output additional trailing digits. The new algorithm instead
> Andres> will instead have *exactly* the required number of digits?
>
> Yeah, this wording is not right. But your description is also wrong.
>
> Previously we output values rounded to 6+d or 15+d digits where
> d=extra_float_digits, with d=0 being the default. Only clients that
> wanted exact results would set that to 3 instead.
>
> Now we output the minimum digits to get an exact result, which is
> usually 8 or 17 digits (sometimes less depending on the value, or 9 for
> the relatively rare float4 values that need it) for any
> extra_float_digits value > 0. Clients that set d=3 will therefore
> usually get one less digit than before, and the value they get will
> usually be slightly different (i.e. not the same value that they would
> have seen with d=2), but it should give them the same binary value after
> going through strtod() or strtof().
Any chance for you to propose a text?
Greetings,
Andres Freund