Greg Stark <stark@mit.edu> writes:
> On Wed, Apr 25, 2012 at 5:47 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Also, as was pointed out upthread, the underlying data in shared memory
>> is almost certainly never going to be infinite-precision; so using
>> numeric in the API seems to me to be more likely to convey a false
>> impression of exactness than to do anything useful.
> I don't think that follows. The underlyng data will be measured in
> some metric unit of time like microsecond or nanosecond or something
> like that. So a base-10 representation will show exactly the precision
> that the underlying data has. On the other hand a floating point
> number will show a base-2 approximation that may in fact display with
> more digits than the underlying data representation has.
My point is that the underlying data is going to be stored in a
fixed-width representation, and therefore it will have accuracy and/or
range limitations that are considerably more severe than use of
"numeric" for output might suggest to the user. In the current
pg_stat_statements code, timings are in fact accumulated in float8,
and emitting them as something other than float8 is just plain
misleading IMHO.
regards, tom lane
