Add http-service chart

[tasuku43]

Summary

Add http-service chart — an opinionated, language-agnostic Helm chart for HTTP services with sensible defaults.

Designed to absorb common boilerplate currently hand-written across Helmfile repositories using the slime chart.

Key features

• Opinionated defaults: RollingUpdate strategy, PDB enabled, Reloader, Datadog unified service tags, required graceful shutdown and health probes
• Type-based autoscaling (none / hpa / keda) with mutual exclusivity validation against replicas
• Argo Rollouts support via workloadRef with automatic HPA/KEDA scaleTargetRef switching
• Graceful shutdown: trafficDrainSeconds (preStop sleep for LB deregistration) + appShutdownTimeoutSeconds (time after SIGTERM) = auto-calculated terminationGracePeriodSeconds

Testing

helm lint --strict and helm template pass for all 5 example patterns (basic, HPA, Ingress, KEDA, Rollout).

$ for f in examples/*.yaml; do echo "=> $f"; helm lint --strict -f "$f" .; done
=> examples/deployment-hpa.yaml
1 chart(s) linted, 0 chart(s) failed
=> examples/deployment-ingress.yaml
1 chart(s) linted, 0 chart(s) failed
=> examples/deployment-keda.yaml
1 chart(s) linted, 0 chart(s) failed
=> examples/deployment-rollout.yaml
1 chart(s) linted, 0 chart(s) failed
=> examples/deployment.yaml
1 chart(s) linted, 0 chart(s) failed

[tasuku43]

Design decision: single container structure

Replaced slime's containers[] array with a single main container defined at the values root (image, containerPort, resources, etc.). HTTP services have one primary container; sidecars go in extraContainers.

[tasuku43]

Design decision: separate drain vs app shutdown

Split into trafficDrainSeconds (preStop sleep waiting for LB deregistration) and appShutdownTimeoutSeconds (time the app gets after SIGTERM). The chart auto-calculates terminationGracePeriodSeconds = drain + app.

A single gracefulShutdownSeconds value would silently reduce the app's actual shutdown time by the preStop sleep duration, which is not the user's intent.

[tasuku43]

Design decision: replicas vs autoscaling guard

Setting replicas when autoscaling is active (hpa or keda) causes a template error. This prevents a common misconfiguration where a fixed replica count conflicts with an autoscaler.

[tasuku43]

Design decision: Rollout uses workloadRef

The Rollout references the Deployment via workloadRef instead of embedding its own Pod template. This avoids duplicating the Pod spec and keeps the Deployment as the single source of truth.

[tasuku43]

Design decision: affinity as passthrough

No auto-generated anti-affinity or node affinity rules. The chart passes values through as-is. Environment-specific settings (e.g. ARM node affinity) are injected at the Helmfile layer, keeping the chart vendor-neutral.

[tasuku43]

Design decision: pod. namespace convention*

All Pod-level settings grouped under pod. (annotations, labels, securityContext, disruptionBudget). Mirrors the deployment. namespace for Deployment-level settings.

[tasuku43]

Design decision: Datadog enabled toggle

Defaults to true with required validation on env/service/version to prevent tag omission. Can be set to false for services using alternative observability (e.g. OpenTelemetry).

[tasuku43]

Design decision: escape hatches and no tpl

extraContainers for sidecars and extraPodSpec for arbitrary Pod spec fields. These cover edge cases without adding complexity to the main template.

Also note: this chart does not use tpl anywhere — all values are rendered with plain toYaml to avoid the debugging complexity of nested Go template evaluation.

[tasuku43]

Design decision: type-based autoscaling

Using type: none | hpa | keda instead of a boolean enabled. HPA and KEDA are mutually exclusive (KEDA creates its own HPA), so an enum makes invalid states unrepresentable. Invalid type values produce a template error.

[cw-atkhry]

LGTM!