persistence

State is there when the process comes back.

Every write goes to the backing store. Restart, redeploy, crash and recover. The values are still there in the same slots you declared.

scalability

Shard the storage. Refs do not notice.

The same Ref API runs against a laptop directory and a terabyte-scale store. Partition by key, split across disks, add read-only tailers. The Shape code does not change.

reactivity

Any Ref emits on change.

Subscribe with on_change and wake code on write. The default observer fires in-process. Swap in the Redis observer and the same subscription fires across the cluster.

durability

Snapshots and transactions built in.

Wrap reads in Snapshot for one consistent view. Wrap writes in Transaction and the whole body commits or none of it does.

rocksdb

LSM store for terabytes and up.

Billions of keys, range scans, snapshots, WAL. Primary writer plus read-only tailers in other processes. The default when data outlives the process.

lmdb

Memory-mapped, single-writer ACID.

Zero-copy reads, one writer, real transactions in a single directory. Reach for it when one process owns writes and readers want speed.

in-memory

ACID in RAM for tests and drafts.

Same Ref API, same brackets, no disk. Real snapshots and transactions, so tests exercise the production code path.

observers

Redis or in-memory pub/sub.

The observer decides who wakes up on write. In-memory for one process, Redis for many. on_change stays the same call.