That and significant differences in performance for more complex operations like the aggregation framework

Curious about that. I recently made something similar, but for SQLite w/JSON. The backend can use a local sqlite3 file instead of mongodb, while using the same queries. It can only translate the queries we actually use, but it's quite doable to translate $lookup, $group and $project to their SQL equivalents (which were the inspiration for the aggregation in the first place, IIRC). Initial evaluation shows that the performance is quite good, and the translations look decent, so it might be possible to get some mileage out of ferretdb.