Documentation
¶
Overview ¶
Command split-executors runs one registry across two executors: a long-running worker process and a bounded, invocation-scoped burst process.
It is one program with three modes so the pieces stay side by side, but the point is that they are separate deployments sharing one database and one registry:
go run ./examples/split-executors enqueue # produce some work go run ./examples/split-executors worker # the long-running pool (Ctrl+C to stop) go run ./examples/split-executors burst # the bounded pool, one invocation
The shape it demonstrates:
- One Registry, built once by newRegistry and compiled into both binaries. Both processes know every kind; ExecutorClass decides who runs what.
- The worker process runs a Node (runners + scheduler) until its context is cancelled. It claims the "worker" class.
- The burst process builds a bare Runner claiming the "burst" class, derives a deadline that reserves a teardown margin below its own budget, and calls RunUntilIdle. That is the shape of a Lambda or a Kubernetes Job.
- The trap: the scheduler runs on exactly one process. Two schedulers means two sweeps.
It uses a shared SQLite file so both modes run on one machine with no setup. A real split deployment uses PostgreSQL, whose SKIP LOCKED claim is what makes two pools safe at concurrency above 1.
Click to show internal directories.
Click to hide internal directories.