Files
aevolution/docs/adr/0170-dev-worker-placement.md
T
dingo fad79973e2
Foundation / checks (checks) (push) Successful in 4s
Foundation / checks (client) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 5s
Foundation / checks (network) (push) Successful in 4s
Foundation / checks (postgres) (push) Successful in 5s
Foundation / checks (tsan) (push) Successful in 4s
Foundation / checks (deploy) (push) Successful in 4m24s
Foundation / postgres-tools (push) Successful in 6s
Foundation / server-image (push) Successful in 5s
Foundation / image (push) Successful in 6s
fix(deploy): constrain DEV workloads to worker nodes
2026-10-19 15:08:00 -04:00

44 lines
2.3 KiB
Markdown

# ADR-0170: Explicit DEV worker placement
## Status and scope
Accepted operational clarification, 2026-10-19, from the maintainer's requirement
that DEV workloads run in `evolution-dev` and never on the control plane. Extends
ADR-0076/0088 without changing biological rules, node labels, storage placement,
database endpoints, credentials, URLs or World0 policy.
## Decision
Apply a reusable Kustomize component to the full DEV application and its
standalone region/store overlays. Every current DEV Deployment and Job requires
nodes without either `node-role.kubernetes.io/control-plane` or the older
`node-role.kubernetes.io/master` label. Both exclusions belong to the same
required selector term. Existing store selectors, volumes and resource settings
remain. No control-plane toleration or node selector override is introduced.
The rendered deployment policy requires both exclusions in **every** required
node-affinity term, since terms are alternatives. It also refuses explicit
`nodeName`, which bypasses scheduler selection. Future DEV workload kinds must
provide the same restriction before passing the policy. Shared bases, stage,
monitoring and operator-only inventory examples keep their existing scope.
## Alternatives and consequences
Relying only on common control-plane taints is weaker than the explicit
requirement: taints are external cluster configuration. Replacing the store
selector with a particular hostname would couple the application to an unknown
node and could strand its node-local volume. Required negative role affinity
preserves current placement constraints while preventing scheduling onto a
correctly role-labelled control plane.
If no eligible worker owns the existing volume, pods remain Pending. Do not
relax the rule or move retained storage merely to obtain readiness. Cluster role
labels remain an infrastructure responsibility; offline rendering cannot verify
their truth or prove a running pod's placement.
Delivery uses the existing Git Actions -> Harbor -> ArgoCD -> Kubernetes flow.
Qualify effective rendered objects and adversarial patches (removed affinity,
missing legacy-role exclusion, unrestricted OR term and direct `nodeName`).
Record actual rollout, placement and target-host measurements separately. No
direct deployment, live reachability or M12 completion is inferred from a push.