Workers own asynchronous background processing. Kafka partitions provide distribution, while domain services remain responsible for authoritative state changes.
Worker inventory
| Worker | Responsibility |
|---|---|
| Scheduling worker | Finds due workflows and creates dispatch work |
| Workflow worker | Validates container payloads, resolves image digests, and updates build state |
| Execution worker | Claims jobs, routes containers to the assigned runtime endpoint, renews leases, and recovers abandoned work |
| Runtime agent | Registers a Docker-capable runtime node and heartbeats endpoint capacity |
| Job logs processor | Writes retained logs to ClickHouse and Meilisearch |
| Analytics processor | Applies workflow, job, and log events to aggregates |
| Outbox relay | Publishes committed outbox rows to Kafka |
| Database migration | Applies PostgreSQL and ClickHouse migrations and Meilisearch setup |
Processing contract
Workers process records per Kafka partition. Commit policy distinguishes successful work, retryable failures, and terminal handling so offsets are not advanced before the selected policy permits it.
Scaling
Replica count cannot create useful parallelism beyond the relevant Kafka topic partition count. Execution capacity is also bounded by registered runtime nodes, Docker hosts, job lease settings, downstream service latency, and log pipeline capacity.
State ownership
Workers call gRPC services to mutate domain state. They do not become an alternate source of truth for workflows or jobs.
For CONTAINER jobs, jobs-service is also the runtime ownership boundary. It chooses a fresh READY runtime node while claiming the job, stores the runtime owner on the job, and returns the Docker endpoint that the execution worker should use.