ModelBase._as_dict() makes assumptions about __dict__ that are changing

[Mike Bayer]

in SQLAlchemy https://gerrit.sqlalchemy.org/#/c/sqlalchemy/sqlalchemy/+/1283/ the assumptions are changing when one does this:

    obj = SomeObject()
    print(obj.some_collection)

You'll still get the usual [] for `some_collection` or whatever it's supposed to be. But it's not going to be in `__dict__` until you mutate it, which is the same behavior as when you say "obj.some_column". You get None back but that's also not in `__dict__`.

Where this breaks is here:

https://github.com/openstack/oslo.db/blob/master/oslo_db/sqlalchemy/models.py#L98

which is because nova over here is relying upon the list being implicitly present in `__dict__` in these tests, specifically for the "rules" attribute:

https://github.com/openstack/nova/blob/master/nova/tests/unit/db/test_db_api.py#L1692

their comparison thing turns the object into dictionaries using _as_dict() here and expects the blank collection to be present, which is because they are implicitly setting it up here:

https://github.com/openstack/nova/blob/master/nova/db/sqlalchemy/api.py#L3862

now, the fix may be just for them to correctly say this instead:

    security_group_ref.rules = []

i might go that route instead...but...looking in `__dict__` is kind of weird here in general considering some things are loaded, some things not. I would think it would want to include relationships the same way it gets the columns, e.g. from the mapper.relationships, but I don't really know the assumptions being made here, and it's possible we just aren't going to know them anymore.

[Mike Bayer (zzzeek)]

Nova will probably be accepting my patch for this in https://review.opendev.org/#/c/661472/1 , I don't have plans to change oslo.db right now though I'm not super enthusiastic about how that function works.