Right now, users can use pandas series/dataframes in state and serialize them with a JsonPlusSerializer that has pickle_fallback enabled.

It'd be great to have these as first class citizens, able to be serialized via msgpack like numpy arrays.

This is a bit of a tricky task, as dataframes have lots of nuanced features like:

• multiindexes (both row and column)
• dtypes by column
• opportunity for arbitrary objects in table cells

We want to preserve df structure during serialization and deserialization.

There are a few options here, assuming we continue to use msgpack:

• Dump pickled content (this is a bit redundant, both are serialization protocols). One benefit here is that pandas x pickle work well together with all of the above nuances
• Dump bytes directly, though custom logic will have to be written for the above pandas features
• Use arrow - this is the most efficient storage wise, though there are some type inconsistencies (like with object dtype) that will need to be considered.

A PR addressing this should have thorough testing, perhaps mimicking many of the conditions tested for in #5057.

You might want to reference #5035 as a reference for how to add logic for new types to JsonPlusSerializer.