IPython 3.0 added a useful feature that we ought to consider for inclusion either in timeit.repeat() or in the command-line interface:

"""Using %timeit prints warnings if there is at least a 4x difference in timings between the slowest and fastest runs, since this might meant that the multiple runs are not independent of one another."""

How it looks?

See attached

This seems like it probably will report something useless (and ultimately be disabled) on PyPy, where runs before and after the JIT will display significant variance.

May be add an option to control a warning (e.g. turn it off when Python run with -Wignore)? May be write warning to stderr?

I think for PyPI its actually important here - the JIT'd state of the code is essentially global state being mutated - you can't assess how fast the code is without first warming up the JIT, and if it warms up half way through your fastest run, you're still not actually finding out what you might want to find out.

E.g. do you want to know:
 - how fast is this unjitted [e.g. CLI's]
 - how fast will this be once its hot [e.g. services]

Personally, I think as a first approximation, warning about massive variance is a good thing. We could add an option to turn it off, and we could also look at hooking properly into the jit to allow detection of stable state and benchmark only from there on in. But those extra things don't detract from the utility of warning about suspect runs.

Reviewed on rietvald.

Here is a patch that emits a warning using the warnings module. The warning is output to stderr and can be suppressed with the -Wignore option, as all other warnings.

$ ./python -m timeit -n1 -r 10 -s "import time, random" -- "time.sleep(random.random())"
1 loops, best of 10: 79.6 msec per loop
:0: UserWarning: These test results likely aren't reliable.  The worst
time was more than four times slower than the best time.
$ ./python -Wignore -m timeit -n1 -r 10 -s "import time, random" -- "time.sleep(random.random())"
1 loops, best of 10: 16.2 msec per loop

Implemented Robert's suggestion.

$ ./python -m timeit -n1 -r 10 -s "import time, random" -- "time.sleep(random.random())"
1 loops, best of 10: 30.2 msec per loop
:0: UserWarning: The test results are likely unreliable. The worst
time (946 msec) was more than four times slower than the best time.

New changeset 2e9cf58c891d by Robert Collins in branch 'default':
Issue #23552: Timeit now warns when there is substantial (4x) variance
https://hg.python.org/cpython/rev/2e9cf58c891d