9.2. Scaling
Direction, not decided
One Postgres store should handle thousands to tens of thousands of appends per second (an estimate, not measured). The bottlenecks are the global position sequence and busy groups (every write to a group locks its counter).
Next step: shard by key. Rules that read only their own key stay fully transactional. Only cross-key rules need coordination: grouped invariants and preconditions with in. The concurrency footprints say exactly which keys and groups a batch touches. Sharding is when a separate engine-set event id may be needed, since one global position becomes hard.