[Python-Dev] Sporadic test failures in Lib/test/test_mimetools.py

[Guido van Rossum]

> choose_boundary() ends likes so:
> 
>     timestamp = '%.3f' % time.time()
>     seed = `random.randint(0, 32767)`
>     return _prefix + '.' + timestamp + '.' + seed
> 
> On Windows, time.time() updates fewer than 100 times per second (but more
> than 10), so the timestamp format is overly optimistic about how much unique
> info it can get out of time.time().  The test in question calls
> choose_boundary() 100 times as fast as it can, and on Windows probably no
> more than two distinct time.time() values are seen for the duration.  The
> chance of a boundary string duplicate is then about the chance of seeing a
> duplicate from the randint() in 100 tries, which is non-trivial.
> 
> So there are more-or-less obvious flaws in this algorithm, but I don't think
> there exist any simple methods to generate a guid that's truly gu.  Using a
> threadsafe counter instead of randomization would at least be unique across
> a program run.

A simple fix would be to change the seed to use 31 instead of 15 bits;
we now have a random generator capable of that.  Perhaps we should go
ahead and use even more bits...  Aren't we using mersenne Twister by
default?

--Guido van Rossum (home page: http://www.python.org/~guido/)