Velero vs etcd Snapshot: Not a Real Choice

Velero and etcd snapshots protect different layers of a cluster, not the same thing. Here's what each covers and why production DR needs both.

Frequently Asked Questions

Does an etcd snapshot back up the actual data inside a PersistentVolume?

No — an etcd snapshot only captures the PVC/PV object definitions (the references), not the bytes of data sitting on the underlying volume. Losing PV data to something like a bad Helm upgrade isn't something an etcd snapshot can recover from.

Can you take etcd snapshots on EKS, GKE, or AKS?

No — on managed Kubernetes, the cloud provider manages etcd as part of the control plane and doesn't expose it for snapshotting. Velero is effectively the only backup mechanism available for anything above the provider's own control-plane SLA.

Can a Velero backup restore a fully dead cluster whose API server won't start?

No — Velero works by calling the Kubernetes API server, so if the API server is down entirely, Velero has nothing to call. That scenario is specifically what etcd snapshots (on self-managed clusters) exist to recover from.

Do you need both etcd snapshots and Velero, or does one cover the other?

On self-managed Kubernetes, you need both — they protect different layers with almost no overlap. An etcd snapshot alone would leave you with a healthy control plane and zero recovered application data if something like a bad Helm chart deleted every PVC in production.

Is having backups enough, or does the restore process itself need to be validated?

Backups alone aren't enough — regular restore drills (monthly for Velero, quarterly is common for etcd) into isolated namespaces or a scratch cluster are what actually confirm recovery works and let you measure real recovery time, rather than assuming a successful backup implies a successful restore.

Discussion0