Ignore NumPy warnings in compute_meta

[pentschev]

Fixes #5098

cc @mrocklin @shoyer

[mrocklin]

You might consider using the finally: block here

try:
    ...
except:
    ...
finally:
    np.seterr(**np_err)

The finally block is run regardless of whether or not there was an error.

[pentschev]

Sure, but there's more code after this block and we still want to keep warnings disabled until just before the last return statement.

[jcrist]

Instead of handling this ourselves, we could also just use:

with np.errstate(all="ignore"):
    ...

which handles setting and resetting for you.

[jrbourbeau]

+1

[pentschev]

Thanks for the suggestion @jcrist, it's indeed a much cleaner solution!

Fixed in latest commits.

[mrocklin]

Minor issue, but it looks like the black check failed. @pentschev you may want to add the pre-commit hooks to run black automatically when committing.

# from the root dask directory
pip install pre-commit
pre-commit install

[pentschev]

The error here happens because the test failing doesn't get a RuntimeWarning in the __array_function__-enabled build only, but that warning is raised and it's not coming from _meta computation. I've spent quite some time trying to figure why that happens, but I can't reproduce it locally no matter what.

[jrbourbeau]

@pentschev I wasn't able to reproduce the failure locally either, so I restarted the failing Travis build which is now passing ¯_(ツ)_/¯

[pentschev]

Damn, I've been very unlucky lately with non-reproducible behavior. Thanks for triggering the build again @jrbourbeau!

[jcrist]

LGTM, merging.

[pentschev]

Thanks fro reviews everyone, thanks for merging @jcrist!

[estebanag]

I've just noticed that the following generates a warning (using dask version 2.3.0):

import dask.array as da
import numpy as np

data0 = da.zeros((3, 10, 10), chunks=(3, 2, 2))
data1 = da.map_blocks(lambda x: np.mean(x, axis=0), data0, dtype=data0.dtype, drop_axis=0)

Output:

numpy/core/fromnumeric.py:3257: RuntimeWarning: Mean of empty slice

Was this supposed to be silenced as well? This is not a problem with version 1.2.2.

Thanks!

Details

dask==2.3.0 numpy==1.17.0 pkg-resources==0.0.0 toolz==0.10.0