Dataset.copy() drops encoding

[crusaderky]

ds = Dataset()
ds.encoding = {"unlimited_dims": 'x'}
ds.copy().encoding
{}

By looking at dataset.py, there's a lot of calls to Dataset._construct_direct that omit the encoding. Is it correct to add it in all cases?

[crusaderky]

By looking at the code, copy() is the only case where it is proper to pass the encoding to _construct_direct, but somebody please confirm

[shoyer]

I think this was intentional at one point, but to be honest we never carefully defined the semantics for preserving encoding are.

My original thought (along the lines of some of my comments in #1297), was that we should not propagate encoding in cases where it might no longer be valid, so it should be dropped from most operations. Essentially, encoding should only stay with original files loaded from disk. Hence why it wasn't copied in .copy().

That said, I can see why this rule would be confusing and we haven't done a good job of enforcing it. Possibly a better policy would be "encoding is copied whenever attrs is copied".

[jhamman]

I agree that we haven't done a good job defining how encoding should behave. I agree that any modification of the shape/type of an xarray object should probably drop the original encoding. I don't know if a copy should drop the encoding though. I don't see why we shouldn't be able rountrip datasets via a open/load/copy/write workflow.

[crusaderky]

Can we reach a resolution on this? It's blocking #1551...

[jhamman]

@crusaderky - After thinking about it, I'm still a 👍 on copying the encoding in this case.

[shoyer]

I'm OK copying encoding, but we do still need to figure out general rules for propagating it.