Cloudpickle fails on boost-histogram

[paulgessinger]

I'm using boost histograms inside Coffea accumulators. I ran into the issue now that boost histograms appear not to be pickleable by cloudpickle (anymore? I think this used to work).

When running something like this:

import cloudpickle
h = bh.Histogram(bh.axis.Regular(50, 0, 20))
h.fill([1 ,2 , 3, 4, 5])
cloudpickle.dumps(h)

I get an error like

AttributeError: module 'boost_histogram._core' has no attribute '__path__'

Should that even work? If so, any pointers how I can resolve or circumvent this issue? This is especially unfortunate since standard pickle appears to work fine, but Coffea uses cloudpickle.

[henryiii]

This should be doable, as long as you can cloudpickle numpy arrays and things like that. I thought cloudpickle was supposed to simply pickle any normally picklable objects. I'll check, and maybe see if I can add it to the tests.

[paulgessinger]

I think it's failing while trying to figure out if the module is a dynamic one or not, where cloudpickle advertises being able to also serialize dynamic ones.

[henryiii]

The following workaround solves this for now:

bh._core.hist.__file__ = bh._core.__file__

I think PyBind11 should probably do this by default. Manually doing this at the C++ level seems to cause a segfault.

[paulgessinger]

The workaround does fix it it for me! Thanks for the help!