On Mon, 06 Feb 2012 00:07:04 -0500, Terry Reedy wrote: > On 2/5/2012 11:01 PM, Steven D'Aprano wrote: > >> Reading the docs, I would expect that when using an int as seed, you >> should get identical results. > > That is similar to expecting hash to be consistent from version to > version.
No. hash is not intended to be consistent across versions, or even across runs of the interpreter. Of course it may be, but that's not an implicit or explicit promise. Seeding a pseudo-random number generator, on the other hand, is explicitly for generating the same repeatable, consistent set of results. That's what seed is *for*. It is even documented that way: http://docs.python.org/py3k/library/random.html#notes-on-reproducibility although the docs weasel out of promising anything other than random.random() will be predictable. When the Mersenne Twister was introduced, the old Wichman-Hill PRNG was provided for those who needed repeatability. (I see it's gone now, but if people haven't migrated their code from 2.3 yet, shame on them.) >> There is no mention that the PRNG has changed between 2.6 and 3.2; > > There is at best an informal policy. This was discussed in > http://bugs.python.org/issue9025 > Antoine argued that if there were a written policy, it should be limited > to bug-fix releases within a version. I agree. I think this thread demonstrates that there are people who depend on repeatability of the random number routines, and not just for random.random(). I think it is ironic (and annoying) that the same release of Python that introduced a version argument to seed() to provide a backward compatible seeding algorithm, also introduced a backward incompatible change to choice(). This, plus Raymond Hettinger's comments on the bug report, make me think that the change in behaviour of randrange and choice is not deliberate and should be treated as a bug. Raymond made a strong case arguing for repeatability, and then approved a bug fix that broke repeatability. I doubt that was deliberate. -- Steven -- http://mail.python.org/mailman/listinfo/python-list
