haven’t written anything up on it, but I wouldn’t use my anecdote as an indication we aren’t off the deep end together!

At a high level, we use DuckDB as an in-memory OLAP cache that’s invalidated via Postgres logical replication.

We compute a lot of derived data to offer a read-heavy workload as a product.

Possibly the most dubious choice I tried out recently was letting the front end just execute SQL directly against DuckDB with a thin serialization layer in the backend.

The best apart about it is the hot reload feedback loop. The backend doesn’t have to rebuild to iterate quickly.