> is quite likely going to be based on Arrow

I don't see the connection.

Apache Arrow isn't going to make a b-tree or LSM faster or more efficient. It's not going to make a point look-up query faster against columnar storage, or a range scan faster against row-based storage. It doesn't make distributed quorum faster, or otherwise impact consistency and fault tolerance.

Removing or reducing SerDe overhead is great, and for analytical workloads where SerDe can be 30-50% of clock time then something like Apache Arrow is pure magic. For the remaining 9X% of use cases it's not adding any more value then you'd see from protobuf.