Skip to content

feat(vault): 3-node raft HA on single-replica storage (JDWLABS-25) [DRAFT — window-gated]#10

Draft
jdwillmsen wants to merge 1 commit into
mainfrom
feat/JDWLABS-25-vault-ha
Draft

feat(vault): 3-node raft HA on single-replica storage (JDWLABS-25) [DRAFT — window-gated]#10
jdwillmsen wants to merge 1 commit into
mainfrom
feat/JDWLABS-25-vault-ha

Conversation

@jdwillmsen
Copy link
Copy Markdown
Member

⚠️ DRAFT — DO NOT MERGE until a maintenance window

Merging triggers ArgoCD to redeploy Vault as 3-node raft, which re-inits Vault and wipes current data. Execute only via the JDWLABS-25 cutover runbook (pre-flight seed-value check → phase 3 init → phase 4 re-seed). Design + runbook live on JDWLABS-25.

Summary

Removes Vault as a single point of failure. On 2026-06-06 a Longhorn instance-manager fault sealed the single standalone Vault, downing the cluster ClusterSecretStore. Part of epic JDWLABS-22.

Closes JDWLABS-25, JDWLABS-30 (Vault PDB is auto-managed by the chart in HA mode).

Changes

  • New longhorn-single StorageClass — 1 replica, best-effort, Retain. Raft replicates across the 3 pods, so one Longhorn fault takes at most one Vault node; quorum (2/3) survives. This is the SPOF removal.
  • vault/values.yamlha.raft enabled, 3 replicas, host anti-affinity, retry_join (follower auto-join), data on longhorn-single.
  • vault-unseal-cronjob.yaml — unseal all 3 pods by DNS name (the VIP only reaches one, leaving followers sealed).

Verification (pre-merge, non-destructive)

  • helm template (chart 0.30.1): renders 3-replica StatefulSet, raft config w/ 3 retry_join, longhorn-single PVCs, host anti-affinity, auto PDB platform-vault.
  • kubectl apply --dry-run=server on SC + CronJob.
  • ✅ All files valid YAML.
  • ⬜ Post-cutover (in window): raft list-peers=3; kill-one-pod keeps secrets served; ExternalSecrets re-sync.

Cutover runbook

On JDWLABS-25. Pre-flight: confirm all seed values on hand (porkbun, grafana, longhorn, alertmanager, usersrole, argocd-dex, per-tenant github-app/ai-keys/discord) — re-init wipes current secrets.

🤖 Generated with Claude Code

Removes Vault as a single point of failure. On 2026-06-06 a Longhorn
instance-manager fault sealed the single standalone Vault, taking down
the cluster ClusterSecretStore.

- New longhorn-single StorageClass (1 replica, best-effort, Retain):
  raft replicates across pods at the app layer, so one Longhorn fault
  takes at most one Vault node and quorum (2/3) survives.
- vault values: ha.raft enabled, 3 replicas, host anti-affinity,
  retry_join for follower auto-join, data on longhorn-single.
- unseal cron: unseal all three pods by name (the service VIP only
  reaches one, leaving followers sealed).

Verified: helm template (chart 0.30.1) renders a 3-replica StatefulSet
with the raft config, longhorn-single PVCs, host anti-affinity, and an
auto-managed PodDisruptionBudget; SC and CronJob pass --dry-run=server;
all files parse as valid YAML.

DO NOT MERGE until the maintenance window: applying this re-inits Vault
(wipes current data). Follow the JDWLABS-25 cutover runbook (pre-flight
seed-value check, phase 3 init, phase 4 re-seed).

Closes JDWLABS-25. Closes JDWLABS-30 (Vault PDB auto-managed by chart).
Refs JDWLABS-22.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant