Saving a workflow with "unpickle-able" outputs.

[HaithamGaafer]

I had a workflow which had one output which is not "pickle-able", so I wasn't able to save the workflow as I can see the folder where the saved pickled workflow would appear and then disappear almost instantly.

Pyiron_workflow version = '0.12.0'

If the workflow has some output that is not 'pickle-able', a message should appear notifying the user of the issue.

Me and @pmrv discovered this bug.

[liamhuber]

Hi @HaithamGaafer, could you supply a minimum working example? The current save function should be falling back on cloudpickle where necessary, so I agree this is buggy behaviour.

[HaithamGaafer]

It happened with fitting a linear ACE potential using python-ace package where the training dataset or testing dataset object from the LinearACEDataset (defined here) is linked as an input or output.

For reference, this is an old version of the notebook that I was using and in cell 13 is where I make all the connections.

[liamhuber]

Here is an example of what I meant by a minimal working example:

import pathlib
import pickle

import pint
import pyiron_workflow as pwf

# Sanity check that we have something unpickleble
ureg = pint.UnitRegistry()
try:
    pickle.loads(pickle.dumps(ureg))
except AttributeError:
    print("Sanity check: Can't pickle a ureg")


# Node that easily switches between un/pickleable output
@pwf.as_function_node
def QuestionableOutput(pickleable):
    if pickleable:
        ureg = "ureg"
    else:
        ureg = pint.UnitRegistry()
    return ureg

wf = pwf.Workflow("canary")
wf.n = QuestionableOutput(False)
wf()

wf.save(filename="coalmine")

# See if it's there
pckl = pathlib.Path("coalmine.pckl")
cpkl = pathlib.Path("coalmine.cpkl")
print("Picklable output?", wf.n.inputs.pickleable.value)
print("File exists?", pckl.exists() or cpkl.exists())
# Clean up
pckl.unlink(missing_ok=True)
cpkl.unlink(missing_ok=True)

In this case it's minimal because I'm only using stuff already in the pyiron_workflow dependencies, and not-minimal in the sense that I added all the bells and whistles of printing outputs, doing sanity checks, making it easy to switch between pickleable and unpicklable outputs, and clean up the files afterwards. I'm willing to install python-ace locally to run an example, but you'll need to build something similarly short and self-contained for me rather than an old notebook link.

Returning to the example above, it reproduced the behaviour you talked about, and I still agree that's a bug. The relevant portion of the codebase is here:

pyiron_workflow/pyiron_workflow/storage.py

Lines 234 to 244 in edc4245

	e = None
	for suffix, save_method in attacks:
	p = filename.with_suffix(suffix)
	try:
	with open(p, "wb") as filehandle:
	save_method(node, filehandle)
	return
	except Exception:
	p.unlink(missing_ok=True)
	if e is not None:
	raise e

Where we first try to pickle the workflow, and then as a fallback try to cloudpickle it.

Now, as it turns out, you can't even cloudpickle the UnitRegistry in my example!

import cloudpickle
import pint

try:
    cloudpickle.dumps(pint.UnitRegistry())
except TypeError:
    print("You can't even cloudpickle this thing!")

But the .save() call was still passing without an exception 🔥 That is 100% a bug -- good find! 🙏 -- so I patched it here: #615.

My suspicion is that your output data is not only not picklable, but also not cloudpickleable. If that is the case, there's nothing we can do about it here; you'll need to either adjust your node to return some other representation of your data, or go into the upstream library to adjust their objects so they can be serialized.