I had a use-case for a persistent dictionary and decided to over do it a bit and built this in the last few days. It has the option to either use local SSD (via BadgerDB) or S3-alike (via SlateDB). S3 sounds like it might be slow but using Tigris which in-lines small writes with metadata and because Slate batches writes, it's actually pretty reasonable. Not quite SSD but maybe close to spinning disk.

It has secondary indexes and Pydantic support so if you felt so inclined, drop the database and live life without object-relational impedance mismatch. That said, while I'm using this in an application I'm building - it's probably not production-ready quite yet. Let me know what else it might benefit from.