0036. Local execution: a non-production dev/CLI backend (amends 0005)

• Status: Accepted
• Date: 2026-07-13
• Deciders: thulasi.ram (architect)

Context

0005 states Kubernetes is the ONLY execution substrate (explicit non-goal: Docker-socket, local, SSH). Yet a scarab-executor-local crate exists and the roadmap (CONTEXT §9.7, slice 7) lists “local exec.” Read literally these conflict. The Executor port (0004, 0016) makes alternative backends possible — the durable brain (scarab-engine + the Postgres outbox + the API) never touches Kubernetes — so the question is one of product stance, not feasibility.

Decision

scarab-executor-local is a non-production developer / CLI / test backend, not a supported deployment mode. [0005]’s “Kubernetes-only” governs how Scarab is deployed and run as a service; it is unchanged. Local execution exists only to:

• scarab run on a laptop — execute a pipeline as host child processes (tokio::process, no Docker, no cluster) for a fast inner loop before pushing.
• Test the real engine end-to-end without a cluster — drive the actual scheduler/state-machine against a process-backed executor in CI of Scarab itself. (This is distinct from the fake executor, which scripts outcomes.)

It is never offered as a hosted/self-hosted execution backend, and carries no multi-backend abstraction beyond the single Executor port that already exists. “Local dev against a cluster” remains kind (Kubernetes-in-Docker) on the k8s adapter — the real substrate.

Consequences for its semantics (a dev tool, so deliberately narrower than k8s):

• Idempotent-on-fence within a process (re-launching the same {run, step, attempt} re-attaches rather than double-spawning), matching the k8s adapter’s contract in-process. Surviving a control-plane restart mid-run is the k8s adapter’s job (durable Pods); after a local restart an in-flight step reports Lost and the orchestrator relaunches (0020).
• No content-addressed workspace / output in v1 — returns None; workspace CAS is the k8s post-step path. So restart skip-if-unchanged ([0027]) and outputs: ([0035]) don’t engage under local exec (safe cascade), same as the k8s adapter until its post-step snapshot lands.
• Steps must carry a command (there is no image to default to).

Consequences

• Fast local iteration and cluster-free engine tests, without weakening the k8s-only product stance.
• The non-goal in [0005] is clarified, not reversed: “no local backend” means “no local production backend.”

Alternatives considered

• Drop host-process exec entirely; local == kind — strictly honors [0005]’s letter, but forfeits a cluster-free inner loop and Docker-free engine tests.
• Promote local to a real backend — reintroduces the multi-backend baggage [0005] exists to shed. Rejected.