One useful clue: when I run these tests on OS X, I get the following warnings: ``` ========================================= warnings summary ========================================== xarray/tests/test_distributed.py::test_dask_distributed_integration_test[scipy] /Users/shoyer/conda/envs/xarray-py36/lib/python3.6/site-packages/distributed/utils_test.py:528: UserWarning: This test leaked 8 file descriptors warnings.warn("This test leaked %d file descriptors" % diff) /Users/shoyer/conda/envs/xarray-py36/lib/python3.6/site-packages/distributed/utils_test.py:538: UserWarning: This test leaked 15 MB of memory warnings.warn("This test leaked %d MB of memory" % diff)

xarray/tests/test_distributed.py::test_dask_distributed_integration_test[netcdf4] /Users/shoyer/conda/envs/xarray-py36/lib/python3.6/site-packages/distributed/utils_test.py:528: UserWarning: This test leaked 6 file descriptors warnings.warn("This test leaked %d file descriptors" % diff)

-- Docs: http://doc.pytest.org/en/latest/warnings.html ```

This is Python 2.7.14.

We do have a build against git master for dask, but only on Linux (Appveyor only gives us one free simultaneous build).

It might be useful in the future to change one of XArray's test builds to be against git-master rather than latest release. This would help us to identify issues like this closer to when they occurred.

On Thu, Nov 23, 2017 at 12:20 PM, Stephan Hoyer notifications@github.com wrote:

Comparing a failed https://ci.appveyor.com/project/shoyer/xray/build/1.0.3015 build to the last successful https://ci.appveyor.com/project/shoyer/xray/build/1.0.3012 build on master, I find the following difference in the dependencies:

passed, just before the v0.10.0 release

ca-certificates 2017.7.27.1 0 conda-forge dask 0.15.4 py_0 conda-forge dask-core 0.15.4 py_0 conda-forge distributed 1.19.3 py27_0 conda-forge setuptools 36.6.0 py27_1 conda-forge

failed, at the v0.10.0 release

ca-certificates 2017.11.5 0 conda-forge dask 0.16.0 py_0 conda-forge dask-core 0.16.0 py_0 conda-forge distributed 1.20.0 py27_0 conda-forge setuptools 36.7.2 py27_0 conda-forge

It looks like dask 0.16 or distributed 0.1.20 release is the most likely culprit. There were no Python changes to xarray in the v0.10.0 release commit (only setup.py and documentation changes).

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/issues/1738#issuecomment-346670059, or mute the thread https://github.com/notifications/unsubscribe-auth/AASszEI5Wou1G91aMSsN7lCuovp2cDSAks5s5alwgaJpZM4QoKTw .

Comparing a failed build to the last successful build on master, I find the following difference in the dependencies: ```

passed, just before the v0.10.0 release

ca-certificates 2017.7.27.1 0 conda-forge dask 0.15.4 py_0 conda-forge dask-core 0.15.4 py_0 conda-forge distributed 1.19.3 py27_0 conda-forge setuptools 36.6.0 py27_1 conda-forge

failed, at the v0.10.0 release

ca-certificates 2017.11.5 0 conda-forge dask 0.16.0 py_0 conda-forge dask-core 0.16.0 py_0 conda-forge distributed 1.20.0 py27_0 conda-forge setuptools 36.7.2 py27_0 conda-forge `` It looks like dask 0.16 or distributed 0.1.20 release is the most likely culprit. There were no Python changes to xarray in the v0.10.0 release commit (onlysetup.py` and documentation changes).

cc @pitrou in case this reminds him of anything

On Thu, Nov 23, 2017 at 12:54 AM, Spencer Hill notifications@github.com wrote:

FWIW aospy is having similar failures starting roughly at the same time: spencerahill/aospy#238 https://github.com/spencerahill/aospy/issues/238

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/issues/1738#issuecomment-346536629, or mute the thread https://github.com/notifications/unsubscribe-auth/AASszD_B8JyfRw354324H3J0l2XeDHBcks5s5QitgaJpZM4QoKTw .

It definitely started recently (past week). I don't think we changed anything relevant in xarray so my guess is something to do with a dependency. Hopefully it will be clear from the appveyor logs! On Wed, Nov 22, 2017 at 4:49 PM Matthew Rocklin notifications@github.com wrote:

If I may generalize the question, did this start happening recently? Do we know what triggered this change?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/issues/1738#issuecomment-346510516, or mute the thread https://github.com/notifications/unsubscribe-auth/ABKS1hI27jKOwumHya6BpXxDcQNa_Hn1ks5s5MEXgaJpZM4QoKTw .

If I may generalize the question, did this start happening recently? Do we know what triggered this change?

It looks like it, I'll double check.

On Wed, Nov 22, 2017 at 4:46 PM Matthew Rocklin notifications@github.com wrote:

At first glance, no. Does this happen both on latest release and on git-master?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/pydata/xarray/issues/1738#issuecomment-346510216, or mute the thread https://github.com/notifications/unsubscribe-auth/ABKS1vBBuWl_-ilWNa7yrd2bIiI6c8vkks5s5MBmgaJpZM4QoKTw .

At first glance, no. Does this happen both on latest release and on git-master?