[TIP] Randomizing test order with nose
Scott David Daniels
Scott.Daniels at Acm.Org
Thu Apr 23 08:32:33 PDT 2009
Douglas Philips wrote:
> On or about 2009 Apr 22, at 6:36 PM, Scott David Daniels indited:
>> There, the simple expedient of subclassing Random by attaching a
>> counter to the root pseudo-random number generator meant that you
>> could characterize your position in the random number sequence as
>> (_seed, _calls).
> ... in previous versions of our attempts to make the PRNG reproducible
> we used a counter (and manually called random() to catch up). It was
> very tricky to make sure everyone had pulled in the right version of
> the instrumented random.
> Had we planned for that up front we could have forced the bottleneck
> necessary. Instead we've found it was just must simpler to seed the
> PRNG to captured and known value, when needed....
OK, I agree the choke point is a good idea. However, I ran off
and read the random module, and if you can accept a large state,
you can certainly use .getstate() and .setstate(). The states are
large, so you really want to store them somewhere, rather than retype
them (the advantage of the counter -- it can actually be retyped).
We might ask for a couple of methods on Random: _ss_ = .get_small_state()
and .set_small_state(_ss_), or even .setstate(_ss) where the consideration
is not speed of extraction/setting, but compactness of state. Such a method
might work as the counter does, but would allow the module to do other
things
like a mixed counter systems which counts block-recalcs and advances-within-
block.
I am thinking this use case has not been considered for the random module;
perhaps I should bring it up on python-dev. If anyone else thinks so,
let me
know and I'll bring it up there. Otherwise I'll figure (as I have before)
that it is just my itch, and I might as well scratch it myself.
> Test reproducibility is crucial when debugging large systems. I'm not
> sure it is of much, if any, use for strict unit-test scope testing.
As long as your unit tests do not random sample an input universe, you
are correct. My former lead and I would argue about this for fun. Of
course, you and he are both wrong, and I am right :-). I know, for example,
that Xerox distributes its employee directory under control of a random
number generator, and, by that expedient, manages to keep the data more
effectively up to date. Similarly, ethernet cannot effectively work
without
randomization.
--Scott David Daniels
Scott.Daniels at Acm.Org
More information about the testing-in-python
mailing list