Federating Clusters for Zero-Downtime Kubernetes

4 days ago

Linkerd's multicluster extension enables multiple clusters to present a service as a single, load-balanced endpoint, with three non-exclusive modes: hierarchical (gateway), flat (pod-to-pod), and federated, selectable per service via labels.
A demonstration setup involves three GKE clusters across regions with full-mesh links, VPC peering for flat networking, and services using different modes: federated (frontend), flat-mirrored (api), and gateway-mirrored (analytics).
Implementation steps include provisioning clusters with non-overlapping CIDRs, installing Linkerd with a shared trust anchor, setting up multicluster links, deploying labeled services, verifying traffic patterns, and testing failover via chaos tests.
Key lessons highlight that federation provides automatic failover for cluster-agnostic clients, mirroring offers explicit cluster selection, and mixing modes on the same links is supported; common pitfalls include VPC peering route configuration and overlapping CIDRs.
Production recommendations include using bidirectional full-mesh links, cert-manager for certificates, network policies, monitoring, and regular failover testing to reduce operational toil in multi-region deployments.

Hasty Briefsbeta