Memory error during serialization of large objects

This is an issue that I'm facing for a while.

Now we are running our pipelines in large scale experiments (several thousands of images), the list of SampleSets that we are generating during pipeline.transform are getting BIG (>1GB) and this is raising some MemoryError Exceptions during serialization (even when we have enough memory). This is very annoying, basically, I can't work with large datasets.

I managed to generate a very simple example describing this issue here: https://github.com/dask/distributed/issues/3806

I know we can change the serializer dask-distributed uses (https://distributed.dask.org/en/latest/serialization.html#use), but I'm not sure if this is the real problem.

However, I would like to propose a workaround that will slow down a bit the execution of experiments, but, at least, the code will not crash. I would like to change the serialization behavior of DelayedSamples to this.

class DelayedSample(_ReprMixin):
       def __init__(self, load, parent=None, **kwargs):
        self.load = load
        if parent is not None:
            _copy_attributes(self, parent.__dict__)
        _copy_attributes(self, kwargs)
        self._data = None

    @property
    def data(self):
        """Loads the data from the disk file."""
        if self._data is None:
            self._data = self.load()
        return self._data

    def __getstate__(self):
        self._data = None
        d = dict(self.__dict__)
        return d

What do you think? ping @andre.anjos @amohammadi

ping @ydayer

thanks