This is two-fold:

1. Keep users from doing things which are allowed but may come back to bite them. For example, spaces in field names, negative dataset IDs
2. Document code paths which are allowed by our library but will fail on the backend. (Discussions in Draw distinction between bytes in blob_value (datastore property type) and large objects in GCS #329, Key id / name should fail if set to 0 / '' #443, Storing an Entity as an indexed property value always fails #529, Trouble with TextProperties #579, datastore: Nested entity parsing fails #616 and DISCUSSION: Entity still not behaving like a dict #620 need to be incorporated).

Is it better to add lots of conditionals and (IMO brittle) validity checks? Or, is the pragmatic approach (less code, let backend indicate failure) better?

/cc @jgeewax