Can immutable dataclasses be hashed as dictionary keys? about mashumaro HOT 4 CLOSED

Hi @user293811

mashumaro.exceptions.UnserializableDataError: dict with dataclasses as keys is not supported by mashumaro

Let me explain why you see this exception. All the supported serialization formats (json, msgpack, yaml, toml) are kind of dictionary based and allow only simple immutable types to be keys. By default dataclasses are serialized into dicts which can't be dictionary keys because they are mutable. NamedTuples are serialized into lists or dicts depending on the engine you choose (default is lists), so they can't be dictionary keys too.

Taking all above into account, if you want to use dataclasses or named tuples as keys in your structure, you should provide a custom serialization (and deserialization for the reverse) method that will convert them into simple immutable object like string. As you have noticed, SerializableType is one of the available approaches to providing a custom serialization method. However, in the current implementation of mappings serialization there is still a code checking that the key is not a dataclass:

This is why you were forced to remove the dataclass decorator. To workaround this problem we can extend this condition by checking SerializableType subclassing. Serialization will fail with TypeError: unhashable type: 'xxx' if dataclass is intentionally converted to unhashable type, but I think we can leave it.

The proposed fix looks trivial. I can do it but if you're willing to contribute yourself, PR is welcome.

Edit: Oh, I just noticed that using forward references breaks this fix on python 3.8. It would be better if I looked at it myself.

from mashumaro.

I modified the process of checking to check if the key type is Hashable. This alteration will also trigger an UnserializableField exception for other types that are not hashable.

If you remove frozen=True parameter from NodeIdentifier the following exception will now be raised:

mashumaro.exceptions.UnserializableField: Field "network" of type dict[NodeIdentifier, NodeConfig] in NetworkConfig is not serializable: NodeIdentifier is unhashable and can not be used as a key

from mashumaro.

Ok, I kind of gotten it to work with SerializableType if I drop the dataclass decorator on the NodeIdentifier, but printing it out isn't pretty. It would be really great dataclass serialization as keys was supported.

from mashumaro.

I tried subclassing NamedTuple and got:

mashumaro.exceptions.UnserializableDataError: ForwardRef('str') as a field type is not supported by mashumaro

This looks like another problem related to ForwardRef support. I managed to support forward references to TypedDicts here but this is something new. Could you provide a reproducible example with details in a separate issue? I'd like to cover as many cases as possible.

from mashumaro.