Someone proposed to use the new tracemalloc module in test.regrtest, I don't remember who. Maybe Nick Coghlan?

I tried an hack dash_R(), but the result was not reliable. I should try harder :-)

I tried different options to show the memory leaks found by tracemalloc, but I'm not satisfied by any option. I get a lot of noise, the output is almost useless. Randomly, between 1 and 1000 KB are allocated or released in random files: in unittest or linecache modules for example.

It looks like the major issue is that the unittest leaks memory. For example, there are still 4 TestCase instances alive after the execution of test_sys. But later, these instances are deleted.

It looks like it creates reference leaks and so memory is only released late. Calling gc.collect() doesn't help. I remember that I already saw something strange with the _Outcome class used in unittest.TestCase.run(). See the issue #19880 (changeset 09658ea0b93d).

The asyncio module has a similar issue: it stores an exception which indirectly contains a reference cycle in the Exception.__traceback__ attribute.
http://bugs.python.org/issue17911
https://code.google.com/p/tulip/issues/detail?id=155

regrtest.patch is a work-in-progress patch. It shows the "top 10" when the -R option is used, ex: "python -m test -R 3:3 test_sys".

The problem is snapshots will by themselves create allocation noise: you cannot really use tracemalloc to detect leaks, only to diagnose the leaks you have detected.

> The problem is snapshots will by themselves create allocation noise: you cannot really use tracemalloc to detect leaks, only to diagnose the leaks you have detected.

It's trivial to ignore allocations done in the tracemalloc module. My work-in-progress patch uses for example these filters:

tracemalloc_filters = [
   tracemalloc.Filter(False, '<frozen importlib._bootstrap>', all_frames=True),
   tracemalloc.Filter(False, tracemalloc.__file__, all_frames=True),
]

> My work-in-progress patch uses for example these filters

If you use filters, you need to add filters for fnmatch and a couple other things :)

2014-07-10 22:55 GMT+02:00 Antoine Pitrou <report@bugs.python.org>:
> If you use filters, you need to add filters for fnmatch and a couple other things :)

tracemalloc.Filter(False, tracemalloc.__file__, all_frames=True)
ignores all memory allocated directly or indirectly by the tracemalloc
module: notice the all_frames=True parameter. I'm saving at least 5
frames in tracemalloc when I work on this issue.

Some of the changes to the way regrtest handles test cases has shown a weird cycle in test_codecs - see issue #22166 for details.

I'll try Victor's patch to see if it's more enlightening than my hacked together attempt.

On second thoughts, what Victor is doing isn't that different from my existing hack - I'll lift the display code, but I'm happy with my data collection code.

One thing I did that seemed to reduce the noise substantially was to leave tracing turned off most of the time, and have a start/stop pair around the actual execution of the test and the subsequent cache cleanup.

Slightly unrelated, but in numba I've rewritten parallel testing and refleak testing as regular unittest extensions (TestRunner / TestResult subclasses). It also means you have per-test method refleak results, instead of per-test file. See e.g. RefleakTestResult in https://github.com/numba/numba/blob/master/numba/tests/__init__.py

I'm no more interested to work on this issue. regrtest.patch is unstable, and I don't know how to make it more reliable. Antoine's allocation counter is enough right now.