Skip to content

Encryption at Rest

What encrypts data of record where it is stored, and what makes it recoverable. In transit is covered in TLS, Postgres client TLS and flannel's WireGuard backend (OVH).

Controls

Data Where Encrypted by Keys held by
Postgres data directory, Redis, Prometheus, OTA databases, Photon index, every other local-path PVC RAID1 NVMe on the single OVH server Nothing. The server is installed from the OVH template with no LUKS n/a
etcd (all cluster objects) Same disk Kubernetes Secrets only: k3s secrets-encryption: true (scripts/ovh/k3s-server.yaml). Other objects are plaintext k3s, on the same disk
CNPG WAL and daily base backups R2 tomoda-db-backups/prod/, gzip-compressed by the Barman Cloud plugin (ObjectStore in k8s/envs/prod/postgres/manifests/cluster.yaml) R2 provider-side AES-256-GCM Cloudflare
etcd snapshots R2 tomoda-db-backups/k3s-etcd/ and local disk R2 provider-side on R2; Secrets inside are also k3s-encrypted Cloudflare; k3s
User media, OTA bundles and dumps, logs and traces, seed sets R2 buckets (Cloudflare) R2 provider-side AES-256-GCM Cloudflare
Secret values GCP Secret Manager Google-managed encryption Google

Barman adds no encryption of its own. CNPG's encryption field asks for S3 or KMS server-side encryption, which R2's S3 API does not support, so it is unset. R2 supports customer-provided keys (SSE-C), but the Barman Cloud plugin that prod backs up through does not expose them yet (plugin-barman-cloud issue #646, PR #1017) (finding in Data Plane).

Recoverability

Prod Postgres is a single CNPG instance (instances: 1) on one node. Recovery from data loss or a lost server relies on the off-node backups:

  • Continuous WAL archiving plus a daily base backup to R2 give point-in-time recovery to any second in the 30-day retention (Disaster Recovery).
  • task dr:drill restores the latest backup, or a point in time, into a separate cluster and checks it against prod. It runs quarterly (Restore drill).

Last restore drill: 2026-10-19, pass (pre-launch data volume)

Accepted risks

Both are Accepted in the findings register:

Risk Rationale Revisit when
No disk encryption on the OVH NVMe Pre-launch, single node; OVHcloud erases a server's drives after the service ends and before reuse (NIST SP 800-88 Purge, Clear where Purge is not possible), and we Purge the NVMe ourselves before releasing a server (release procedure); backups are off-node and encrypted at rest in R2, with PITR and a passing restore drill The node is reprovisioned or its OS reinstalled, a node is added, or a customer or contract requires it
Single-instance Postgres Pre-launch, single node, so a standby would share the same hardware; PITR from R2 and a passing restore drill cover data loss Same triggers