Kubernetes Upgrade

Kubernetes Upgrade

Orchestrate safe Kubernetes version upgrades on Talos Linux. Primary value: comprehensive pre-flight safety before talosctl upgrade-k8s.

Quick Reference

Workflow

Phase 1: Input Parsing

• Parse target version from user message; prompt if missing
• Read current version from talos/talconfig.yaml (kubernetesVersion)
• Classify: minor (1.34→1.35, higher risk) or patch (1.35.0→1.35.1)

Phase 2: Create GitHub Issue

Create a GitHub issue using the infra template so SRE agents see maintenance context. Include:

• Title: infra(k8s): upgrade Kubernetes to v<version>
• Label: infra
• Summary, motivation (Renovate PR ref if applicable), planned changes, rollback plan, risk level (patch vs minor)
• Track issue number for commit references (Ref #<number>)

Phase 3: Breaking Changes Research

Consult references/breaking-changes-lookup.md for procedures.

HARD GATE: Present findings. Wait for user acknowledgment. If removed APIs match cluster resources, recommend aborting.

Phase 4: Cluster API Compatibility Scan

Consult references/api-deprecation-scanning.md for procedures.

HARD GATE: If removed APIs are in active use, BLOCK. List resources requiring migration.

Phase 5: Talos Compatibility Check

1. Read Talos version from talos/talconfig.yaml (talosVersion)
2. Query Context7: query-docs(libraryId: "/siderolabs/talos", query: "supported kubernetes versions for Talos v<current>")

HARD GATE: If incompatible, BLOCK. Recommend talos-upgrade agent first.

Phase 6: Cluster Health Gate

Discover CP node IPs dynamically:

CP_NODES=$(kubectl get nodes -l node-role.kubernetes.io/control-plane \
  -o jsonpath='{.items[*].status.addresses[?(@.type=="InternalIP")].address}')

HARD GATE: All checks must pass. Report specific failures.

Phase 7: etcd Backup

CP_NODE=$(kubectl get nodes -l node-role.kubernetes.io/control-plane \
  -o jsonpath='{.items[0].status.addresses[?(@.type=="InternalIP")].address}')
talosctl -n $CP_NODE etcd snapshot /tmp/etcd-backup-$(date +%Y%m%d-%H%M%S).snapshot

Phase 8: Dry Run

talosctl upgrade-k8s -n <cp-node-ip> --to v<version> --dry-run

HARD GATE: Must succeed.

Phase 9: Execute Upgrade

talosctl upgrade-k8s -n <cp-node-ip> --to v<version>

Wait for all nodes to show new version (kubectl get nodes). If no progress after 20 min, investigate with talosctl dmesg and kubectl describe nodes.

Phase 10: Post-Upgrade Validation

Re-run Phase 6 health checks plus kubectl version and kubectl get nodes -o wide. Confirm all nodes report new version.

Phase 11: Roll Stale Secret Volume Mounts

K8s upgrades restart all kubelets. After restart, kubelet's Watch-based secret manager fails to re-establish watches for pre-existing pods' volume mounts. Secrets mounted as volumes become frozen at the pre-upgrade state.

Detection: For every Running pod with a secret volume (non-subPath), exec stat -c %Y <mount>/..data and compare against the kubelet restart time. If the ..data symlink timestamp predates the kubelet restart, the mount is stale.

Roll order (sequential, not parallel):

1. Safe tier — stateless or independently restartable. Roll all at once:

   • kubectl rollout restart for Deployments
   • kubectl delete pod for StatefulSets (one at a time, wait for Ready)
   • Namespaces: observability, authentik, cnpg plugins, firefly-iii, nexus, app workloads

2. DNS tier — technitium primary + secondary:

   • Roll secondary first, wait for Ready
   • Roll primary, wait for Ready
   • Verify: dig @<technitium-ip> <any-internal-record>

3. Storage tier — rook-ceph (mons, mgrs, rgws, crashcollectors, exporters, tools):

   • Before each restart: ceph status must show HEALTH_OK (HEALTH_WARN acceptable only for expected warnings)
   • Roll one pod at a time, wait for Ready + Ceph health between each
   • Order: tools → crashcollectors → exporters → rgw → mgr-b → mgr-a → mons (one at a time)
   • STOP if Ceph goes HEALTH_ERR — investigate before continuing

4. Skip: Pods where stat failed (subPath mounts) — these read at pod start and don't use the ..data symlink mechanism

Report: List what was rolled and current health after completion.

Phase 12: Update Files & Report

1. Update kubernetesVersion in talos/talconfig.yaml
2. Search for all old version references:
   • Grep tool: search <old-version> (no v prefix) in talos/*.yaml, docs/*.md, cluster/*.yaml
   • Grep may miss hookify-blocked files. Fallback: grep -r "v<old-version>" cluster/ --include="*.yaml" -l 2>/dev/null. Files found only by bash need sed -i instead of Edit tool.
3. Common locations: talos/talconfig.yaml, talos/README.md, cluster/flux/meta/cluster-settings.yaml, kubernetes-json-schema URLs in 30+ manifest files
4. Update all references; verify zero remain
5. Present final report: version change, node status, health results, files changed

Rollback

• talosctl upgrade-k8s is idempotent — re-run if it fails partway
• etcd backup from Phase 7 is primary recovery (WARNING: restore is destructive, resets to snapshot point)
• Debug: talosctl -n <ip> logs kubelet, talosctl -n <ip> dmesg
• Context7: query-docs(libraryId: "/siderolabs/talos", query: "kubernetes upgrade rollback recovery")

Commit Pattern

infra(k8s): upgrade Kubernetes to v<version>

Ref #<issue-number>

Source: anthony-spruyt/spruyt-labs — distributed by TomeVault.