Shall we change the hashing to encoding to reduce potential replicated cache files?

Hi there. For now, we are using `xxhash` to hash the transformations to fingerprint and we will save a copy of the processed dataset to disk if there is a new hash value. However, there are some transformations that are idempotent or commutative to each other. I think that encoding the transformation chain as the fingerprint may help in those cases, for example, use `base64.urlsafe_b64encode`. In this way, before we want to save a new copy, we can decode the transformation chain and normalize it to prevent omit potential reuse. As the main targets of this project are the really large datasets that cannot be loaded entirely in memory, I believe it would save a lot of time if we can avoid some write.

If you have interest in this, I'd love to help :).