0026. Resource & placement model

• Status: Accepted
• Date: 2026-07-12
• Deciders: thulasi.ram (architect)

Context

Tension: the IR should stay portable/agnostic (0009), but placement is inherently k8s-ish (nodeSelector, tolerations, runtimeClass, resource requests, GPU). Letting pipelines write raw k8s couples the IR and leaks infra into pipeline files; over-abstracting blocks power users (a specific taint, a GPU class).

Decision

Abstracted labels in the IR + admin-configured mapping + a raw escape hatch:

• Pipelines say runs-on: linux/arm64, resources: { cpu, mem }, optional size: large.
• A cluster-admin config maps those labels to k8s specifics (nodeSelector / tolerations / runtimeClass / requests).
• A raw pod-spec overlay is available for power users who truly need it.

Portable by default; no ceiling. Infra lives in admin config, not pipeline files.

Consequences

• IR stays agnostic and multi-frontend-friendly; pipelines are clean.
• Admins own the label→k8s mapping (one place).
• The raw overlay is an explicit, discouraged-by-default escape hatch.

Alternatives considered

• Raw k8s in the IR — max power, but IR k8s-coupled; infra leaks into every pipeline.
• Fixed sizes only — cleanest, but no escape hatch for niche needs.