Fit and requirements
The Executor is the right tool when you want one application to carry its runtime, its limits and its stop conditions with it, wherever it runs. This page also says plainly where it is not the right fit.
A good fit
- You ship a service or a batch job to hosts you do not fully control: customer-hosted machines, on-premise servers, cloud VMs or containers.
- You need to stop an application centrally (revocation, a maintenance cutoff, a lease) without adding an SDK to every codebase.
- You run multi-process frameworks such as
uvicorn --workers 4or Gunicorn and need every worker stopped, every time. - You want CPU and memory limits that travel with the application and resolve against the machine it lands on.
- You want a second replica under load on a single host without installing an orchestrator or a reverse proxy.
- You want to isolate one binary on a bare-metal host or VM without a privileged container daemon.
When the Executor is not the right fit
- Scaling across many hosts. The Executor's auto-scaler runs replicas on one host, behind its built-in load balancer. Spreading a service over a cluster of machines remains the job of your orchestrator.
- Runtimes outside the adapter matrix. The Executor packages native binaries, Python 3.10 to 3.13, JVM 17, 21 and 25, and Node.js, Deno and Bun. Other interpreters are not bundled.
- Load balancing beyond TCP and HTTP/1.1. The built-in load balancer balances TCP connections or HTTP/1.1 requests with
round_robinorleast_connections. It is not a general-purpose reverse proxy. - Message brokers other than Kafka, MQTT, NATS and Redis. AMQP and cloud queues are reached through the
webhookorcommandexit-event sources instead of a native subscription. - Changing a strict cap without a rebuild. A
strictcap is sealed into the signed binary; changing it means a new signed build. That is the point of a strict cap, and the wrong choice when operators must tune limits per host at run time; useflexiblethere.
Requirements by feature
| Feature | What the host must offer |
|---|---|
| Running a wrapped binary | Linux (x86_64, aarch64), Windows (x86_64) or macOS (aarch64, x86_64), or an OCI container on one of them. No interpreter, daemon or root. |
strict CPU and memory caps on Linux | A writable cgroup v2 for the user: a delegated subtree, a transient scope from the user's systemd manager, or a writable container cgroup. Without one, the binary refuses to start with exit code 69, unless the manifest opts into the governor tier with resources.allow_governor_fallback: true. |
strict caps on macOS | No kernel hard limit exists for a process tree, so a supervisor governor enforces the cap; on Linux the governor is used only with allow_governor_fallback: true. |
| Full sandbox with namespaces on Linux | Creating user, PID, mount and network namespaces without root (CLONE_NEWUSER). Inside a default Docker or Kubernetes container, container_fallback: "auto" applies Landlock, seccomp-BPF, subreaper containment and setrlimit instead; "strict" refuses to start with exit code 69. |
| Landlock filesystem rules | A Linux kernel with Landlock (ABI v1 to v5). |
| Auto-scaling | A cap on every metric the auto-scaler uses, and the extended supervisor variant, which oee build selects automatically. |
| Message-broker and webhook exit events | The extended supervisor variant, selected automatically. |
| Python, JVM and Node.js workloads | A writable cache directory for the runtime cache, by default ~/.cache/oee/runtimes/, or the directory in OEE_CACHE_DIR. |