sre（斯雷）

Cluster access (--context patterns) and internal service URLs are in the k8s skill.

Debugging Kubernetes Incidents

Core Principles

• 5 Whys Analysis — NEVER stop at symptoms. Ask "why" until you reach the root cause.
• Multi-Source Correlation — Combine logs, events, metrics for a complete picture.
• Zero Alert Tolerance — Every firing alert must be addressed: fix the root cause, or as a last resort, create a declarative Silence CR with justification. Never ignore or defer.

The 5 Whys Analysis (CRITICAL)

Apply 5 Whys before concluding any investigation. Stopping at symptoms leads to ineffective fixes.

Example:

Symptom: Helm install failed with "context deadline exceeded"

Why #1: Pods never became Ready
Why #2: Pods stuck in Pending state
Why #3: PVCs couldn't bind (StorageClass "fast" not found)
Why #4: longhorn-storage Kustomization failed to apply
Why #5: numberOfReplicas was integer instead of string

ROOT CAUSE: YAML type coercion issue
FIX: Use properly typed variable for StorageClass parameters

See investigation-guide.md for red flags that you haven't reached root cause.

Investigation Phases

Phase 1 — Triage: Confirm cluster (ask user: dev/integration/live) → assess severity (P1 down / P2 degraded / P3 minor) → identify scope.

Phase 2 — Data Collection: Use scripts/cluster-health.sh [namespace] for a quick snapshot. For targeted collection:

# Pod status and events
kubectl get pods -n <namespace>
kubectl describe pod <pod> -n <namespace>

# Logs (current and previous)
kubectl logs <pod> -n <namespace> --tail=100
kubectl logs <pod> -n <namespace> --previous

# Events timeline
kubectl get events -n <namespace> --sort-by='.lastTimestamp'

# Resource usage
kubectl top pods -n <namespace>

Metrics and alerts (Prometheus is behind OAuth2 Proxy — DNS URLs won't work for API queries):

# Check firing alerts
kubectl --context <cluster> exec -n monitoring prometheus-kube-prometheus-stack-0 -c prometheus -- \
  wget -qO- 'http://localhost:9090/api/v1/alerts' | jq '.data.alerts[] | select(.state == "firing")'

Phase 3 — Correlation: Extract timestamps from logs, events, metrics → identify what happened FIRST → trace cascade.

Phase 4 — Root Cause: Apply 5 Whys. Validate: temporal (before symptom?), causal (logically explains it?), evidence (supporting data?), complete (asked "why" enough times?).

Phase 5 — Remediation: Use AskUserQuestion when multiple valid approaches exist. Provide recommendations only (read-only on integration/live):

• Immediate: rollback, scale, restart
• Permanent: code/config fixes
• Prevention: alerts, quotas, tests

For symptom → first check → common cause mapping, see investigation-guide.md.

Network Policy Debugging (Cilium + Hubble)

All traffic is implicitly denied. Missing labels are the most common cause of blocked traffic.

# Setup Hubble access (run once per session)
kubectl --context <cluster> port-forward -n kube-system svc/hubble-relay 4245:80 &

# See dropped traffic in a namespace
hubble observe --verdict DROPPED --namespace <namespace> --since 5m

# Check specific traffic flow
hubble observe --from-namespace <source> --to-namespace <dest> --since 5m

Check namespace labels:

kubectl --context <cluster> get ns <namespace> -o jsonpath='{.metadata.labels}' | jq
# Required: network-policy.homelab/profile: standard|internal|internal-egress|isolated
# Optional: access.network-policy.homelab/postgres|garage-s3|kube-api: "true"

Emergency escape hatch (triggers alert after 5m):

kubectl --context <cluster> label namespace <ns> network-policy.homelab/enforcement=disabled
# Re-enable after fixing:
kubectl --context <cluster> label namespace <ns> network-policy.homelab/enforcement-

See docs/runbooks/network-policy-escape-hatch.md for full procedure.

Kickstarting Stalled HelmReleases

HelmReleases can get stuck in Stalled/RetriesExceeded even after the underlying issue is resolved. Suspend and resume to reset the failure counter:

flux --context <cluster> suspend helmrelease <name> -n flux-system
flux --context <cluster> resume helmrelease <name> -n flux-system

Common self-healed causes: missing Secret/ConfigMap (ExternalSecret eventually created it), missing CRD, transient image pull failure, temporary resource quota exceeded. Ensure proper dependsOn ordering to prevent recurrence.

Promotion Pipeline Debugging

Symptom: "Live cluster not updating after merge"

Walk through each stage in order — see investigation-guide.md for the failure mode table. Quick diagnostic flow:

1. PR merged → did build-platform-artifact.yaml trigger?
   └─ If not: was kubernetes/ modified? (paths filter)

2. OCI artifact in GHCR?
   └─ flux list artifact oci://ghcr.io/<repo>/platform | grep integration

3. Integration OCIRepository seeing new version?
   └─ kubectl --context integration get ocirepository -n flux-system
   └─ Semver constraint must be ">= 0.0.0-0" to accept RCs

4. Integration Kustomization healthy?
   └─ flux --context integration get kustomizations -n flux-system

5. Flux Alert fired repository_dispatch?
   └─ kubectl --context integration describe alert validation-success -n flux-system

6. tag-validated-artifact.yaml ran?
   └─ GitHub Actions → "Tag Validated Artifact" workflow

7. Live OCIRepository seeing stable semver?
   └─ kubectl --context live get ocirepository -n flux-system
   └─ Semver constraint must be ">= 0.0.0" (stable only, no RCs)

See .github/CLAUDE.md for full pipeline architecture and rollback procedures.

Keywords

kubernetes, debugging, crashloopbackoff, oomkilled, pending, root cause analysis, 5 whys, incident investigation, pod logs, events, troubleshooting, network policy, hubble, stalled helmrelease, promotion pipeline, live not updating