On Wed, Jul 1, 2015 at 1:17 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Petr Jelinek <petr@2ndquadrant.com> writes:
>> On 2015-06-25 10:01, Michael Paquier wrote:
>>> I think that we should change the returned double to be (0.0,1.0]
>
>> Agreed.
>
> I find this to be a pretty bad idea. That definition is simply weird;
> where else in the world will you find a random number generator that does
> that? What are the odds that any callers are actually designed for that
> behavior?
>
> Another problem is that we consider anl_random_fract() to be an exported
> API, and the very longstanding definition of that is that the result is
> in (0,1), excluding both endpoints. Whatever we do with
> sampler_random_fract(), we'd better make sure that anl_random_fract()
> preserves that behavior, else we are likely to break third-party modules.
Wait a minute... Yes you are right I clearly missed the fact that in
~9.4 the range of values returned by anl_random_fract() does not
include 1. I thought it did, visibly I misread the code...
> A simple fix would be to adjust sampler_random_fract to disallow 0
> as result, say by repeating the pg_erand48 call if it produces 0.
> I'm not sure if that would throw off any of the math in the new
> tablesample-related callers. If it would, I'm inclined to fix the
> problem call-site-by-call-site, rather than inventing a definition
> of sampler_random_fract() that fails to satisfy the POLA.
Agreed. Disallowing 0 in sampler_random_fract looks like a good answer
to that. Looking at the tablesample code, for the bernouilli trial I
recall that the range of probability success and failure is actually
(0,1), (if p is a success rate, the failure is 1 - p). I am not sure
for Knuth Algo S though for the system sampling but that looks OK from
a pure logical viewpoint.
--
Michael