Actors, supervision, lifecycle
The classic actor model in idiomatic TypeScript. Each actor owns its state and processes messages one at a time, so you stop reasoning about locks and race conditions. When a child fails, its parent gets to decide whether to restart, escalate, or stop it — supervision trees give you fault tolerance as a structural property, not as a try/catch afterthought.
Cluster + sharding
Run the same actor system across many nodes. Members discover each other via gossip, the failure detector flags unreachable peers, and split-brain resolvers decide who survives a partition. Sharding distributes long-lived stateful entities across the cluster, rebalances when nodes join or leave, and remembers which entity lives where so a restart picks up cleanly.
Event sourcing + persistence
Make actor state survive crashes by recording the events that
led to it instead of the state itself. PersistentActor replays
its event log on startup; DurableState writes a single snapshot
when full history isn’t useful. Journals plug in — in-memory for
tests, SQLite for single-node prod, Cassandra or S3 for scale —
with first-class snapshotting + projections for read-side views.
Distributed data (CRDTs)
Shared state across the cluster without a coordinator. Eight conflict-free replicated data types — counters, registers, sets, maps — that always converge to the same value no matter the order of updates. Pick quorum reads + writes when you want strong-ish consistency, or accept gossip propagation when you don’t. Durable backend keeps the state across restarts.
Integrations
The world outside the actor system, behind one consistent shape.
HTTP with a directive-style route DSL on Fastify, Express, or
Hono backends. Message brokers — Kafka, MQTT, AMQP, NATS, Redis
Streams, gRPC, SSE, WebSocket, raw sockets — all wrapped by a
common BrokerActor base with reconnect, buffering, and fan-out
baked in. Cache + serialization are pluggable abstractions.
Observability + testing
Production-grade visibility from day one. Stock metrics for every actor, mailbox depth, gossip lag, restart counts — exported via Prometheus or OpenTelemetry. Per-message tracer spans, so a request flows through your cluster with one trace ID. Management HTTP endpoints expose cluster state. TestKit gives deterministic multi-node tests with a manual scheduler for time control.
Two end-to-end sample apps live in the repo, each with six
interchangeable frontends (Plain HTML, Lit, Angular, React, Next.js,
SvelteKit) talking the same WebSocket protocol to a clustered
backend. Clone, bun examples/<sample>/backend/main.ts, open the
frontend in a browser, and poke at a real running actor system.
actor-ts is pre-1.0. The API surface is broad and battle-tested
in internal use, but breaking changes can land between minor versions
until 1.0 is cut. Pinning a specific minor version in your
package.json and reading the
CHANGELOG
before upgrading is the safe play during this phase.
