Skip to content

Redis

Single-instance Redis per environment, deployed from the Bitnami Helm chart through Argo CD's multi-source pattern. Used by the Tomoda backend for the Asynq job queues, sessions, rate limits and short-lived caches; not a store of record.

Argo CD Application

k8s/envs/<env>/redis/application.yaml declares two sources:

  • Chart: registry-1.docker.io/bitnamicharts/redis at chart version 28.3.1.
  • Values: the same devops Git repo, referenced as $values, pointing at k8s/envs/<env>/redis/values.yaml.

The split lets the chart upgrade independently of the values file. Both envs sync into the data namespace.

Values (both envs)

image:
  registry: docker.io
  repository: bitnami/redis
  tag: latest
  digest: "sha256:33dbb7a4…"   # redis 8.10.2

architecture: standalone
replicaCount: 1
auth:
  enabled: true
  sentinel: false
  existingSecret: redis-<env>-password-secret
  existingSecretPasswordKey: redis-password

The image tracks the maintained bitnami/redis (not the frozen bitnamilegacy/* namespace) and pins the digest for reproducibility. Bitnami's free tier publishes only the latest tag, so the digest is the version pin: re-pin it from bitnami/redis:latest to pick up newer CVE fixes, and check the image's org.opencontainers.image.version label first, since latest can move to a new Redis major. The exporter sidecar (bitnami/redis-exporter) is digest-pinned the same way under metrics.image.digest. standalone is a single Redis pod, no failover (lean): the data is recoverable (no source-of-truth state) and the restart blast radius is small.

Node placement

No node selectors; both envs run on the server alongside everything else, on local-path storage.

HA switch (resilient flavour)

Lean is architecture: standalone, one pod. The resilient upgrade switches to Sentinel: architecture: replication + sentinel.enabled: true, giving 1 master + 2 replicas + 3 sentinels for automatic failover. Applications keep connecting through the Sentinel-aware Service; the DSN changes but the REDIS_* env contract does not. Turn this on together with the Postgres HA switch when a second server joins.

Auth

auth.enabled: true in both envs; Redis requires a password on every connection. The chart reads it from an existing Secret: auth.existingSecret: redis-<env>-password-secret, auth.existingSecretPasswordKey: redis-password (k8s/envs/{dev,prod}/redis/values.yaml).

That Secret is projected by an ExternalSecret (k8s/envs/{dev,prod}/redis/external-secret.yaml) from the per-env GCP Secret Manager key tomoda-redis-password-<env> into redis-password. The backend reads the same value as REDIS_PASSWORD, so pod and clients stay in sync.

Network reach is constrained independently: the Redis Service is reachable only inside the data namespace, gated by consumer-side NetworkPolicies and the absence of any Ingress or LoadBalancer.

Persistence

Env PVC size Notes
dev 1Gi RDB/AOF snapshots survive pod restart
prod 10Gi Larger room for prod working set

Both use the cluster default StorageClass.

Resources

Env Requests Limits maxmemory
dev 25m CPU, 64Mi memory 256Mi memory 200mb
prod 50m CPU, 256Mi memory 2Gi memory 1600mb

maxmemory sits below the container limit with maxmemory-policy noeviction: at the cap Redis rejects writes (the backend sees errors) instead of being OOMKilled, and queued jobs are never evicted. RedisMemoryNearMax warns at 80% of maxmemory, RedisHighMemory fires at 90%, RedisRejectingWrites on the first rejected write, and RedisMaxmemoryUnset if the cap is ever 0.

Metrics

Both envs set metrics.enabled: true, running the Redis exporter sidecar and registering a ServiceMonitor in monitoring under release: monitoring, which kube-prometheus-stack picks up automatically. Every Redis alert (k8s/envs/platform/manifests/redis-alerts.yaml, Data alert runbook) reads these series by job, redis-<env>-metrics.

DSNs (application-facing)

The Bitnami chart names its Service <release>-master. Applications connect to:

Env Hostname Port
dev redis-dev-master.data.svc.cluster.local 6379
prod redis-prod-master.data.svc.cluster.local 6379

These hostnames are written into the Tomoda backend-config-<env> ConfigMap. The prod release is redis-prod (application.yaml metadata), so redis-prod-master is the rendered Service name.

Operations

  • Bounce: kubectl -n data rollout restart statefulset/<release>-master. With architecture: standalone there is no quorum to worry about; persistence survives the restart.
  • Console: use Redis Insight for a GUI, or kubectl -n data exec -it <release>-master-0 -- redis-cli for ad-hoc commands.
  • Storage: local-path doesn't enforce sizes; Redis can use the server's free disk. Watch host disk instead.
  • Chart upgrade: change targetRevision on the chart source. Review the chart release notes; values keys drift between major versions.