Rely on cuDF's Serializable in CumlArray

[jakirkham]

Depends on PR ( rapidsai/cudf#5139 ) and nightlies with that change.

Update CumlArray to rely on cuDF's Serializable, which cuDF's Buffer inherits from. In particular Serializable already provides pickling support through a __reduce_ex__ method of its own, which just builds off of the serialize and deserialize methods that all Serializable subclasses must implement and CumlArray already does (either by directly implementing them or relying on cuDF Buffer's implementations). So no extra work is needed to just drop this.

[jakirkham]

cc @dantegd @cjnolet

[dantegd]

@jakirkham the PR fails seem to indicate me that an issue with serializing/pickling the data inheritted as is from Serializable is serializing the data as a Buffer and losing the datatype information (which causes all that havoc), perhaps we do need to override methods? The errors in CI are very similar to the ones I saw when the new reduce_ex in Buffer broke CumlArray

[jakirkham]

rerun tests

(cudf nightlies with the change are now available)

[jakirkham]

Sorry for the confusion. I think the failure is was just because we didn't have the Conda packages with the change. Rerunning CI to confirms.

Should add here's Serializable's __reduce_ex__ method for context. We've dropped all other __reduce__/__reduce_ex__ methods of cuDF objects inheriting from Serializable. There are already quite a few tests of pickling in cuDF, which were happy with the change. So I think we should be good.

That all being said, if I'm wrong about this, happy to take a closer look tomorrow 🙂