diff options
| author | Matti Picus <matti.picus@gmail.com> | 2018-05-25 02:54:02 +0200 |
|---|---|---|
| committer | Nathaniel J. Smith <njs@pobox.com> | 2018-05-24 17:54:02 -0700 |
| commit | 055620ccd669ea56d83391e107a467824cd5b60b (patch) | |
| tree | 26d8442e343abbc37e4e2b26538d2521e7e5dac3 /numpy/testing/_private/utils.py | |
| parent | 4d02fb11352a45b05e5d50db6e71153b2aec7527 (diff) | |
| download | numpy-055620ccd669ea56d83391e107a467824cd5b60b.tar.gz | |
TST: disable gc in refcount test (#11158)
The vectorize version of this test was failing consistently on several of the Appveyor builds, ever since a recent pytest upgrade.
Our theory is that by random chance, things changed so that during the call to vectorize(op).__call__, python started running a garbage collection, which perturbed the refcounts that this test is checking. (Specifically this test is doing a weird thing and checking that the refcount of the object 1 doesn't decrease, and it's very plausible that some random bit of garbage was holding a reference to this object.)
Disabling the gc during the test makes this kind of refcount assertion more reliable, and seems to have fixed the appveyor builds, so I guess it's good.
Diffstat (limited to 'numpy/testing/_private/utils.py')
| -rw-r--r-- | numpy/testing/_private/utils.py | 14 |
1 files changed, 9 insertions, 5 deletions
diff --git a/numpy/testing/_private/utils.py b/numpy/testing/_private/utils.py index c420e1fb5..a7935f175 100644 --- a/numpy/testing/_private/utils.py +++ b/numpy/testing/_private/utils.py @@ -1373,16 +1373,20 @@ def _assert_valid_refcount(op): """ if not HAS_REFCOUNT: return True - import numpy as np + import numpy as np, gc b = np.arange(100*100).reshape(100, 100) c = b i = 1 - rc = sys.getrefcount(i) - for j in range(15): - d = op(b, c) - assert_(sys.getrefcount(i) >= rc) + gc.disable() + try: + rc = sys.getrefcount(i) + for j in range(15): + d = op(b, c) + assert_(sys.getrefcount(i) >= rc) + finally: + gc.enable() del d # for pyflakes |