High-Availability Registry

A single registry is a SPOF. If it dies, no service can find any other.

Replication options:

Consensus-replicated (Raft/Paxos):

• 3-5 servers
• Strongly consistent
• Lower throughput
• Used by: Consul, etcd, ZooKeeper

Peer-to-peer (Eureka):

• Many servers; gossip-replicated
• Eventually consistent
• High throughput
• Higher staleness; AP per CAP

Federation: multiple datacenters, each with its own cluster. WAN replication for cross-DC discovery.

Choosing CP vs AP:

• CP (Consul, etcd): registry data is always consistent. Lookup may fail during partition (some clients can't reach quorum).
• AP (Eureka): registry always responds; data may be stale. Better for "best-effort" discovery.

Eureka's reasoning: it's better to have some stale results than NO results. A stale instance address might still work; an empty result definitely doesn't.

In practice: most modern stacks use CP (Kubernetes uses etcd; Consul is CP). The trade-off: deploy carefully so you never lose quorum.

Caching: clients cache last-known good results. Survive registry outages.