BEAM distribution + consistent hashing. Entities are assigned to nodes by their ID. If a node goes down, the hash ring updates and surviving nodes pick up its entities. Litestream continuously streams SQLite WAL changes to S3 so you can restore from the last segment.
The clustering and rebalance coordinator are implemented and tested with multi-node BEAM peers, but I'll be honest, I haven't run it in a multi-server production deployment yet. Single-node with Litestream backup is what I'm running in production right now. Recovery from backup takes seconds, not minutes.
True multi-node HA is architecturally ready but not battle-tested at scale yet. v0.1.