Global Traffic Management

Three mechanisms with materially different failover characteristics:

DNS-based returns different addresses by geography, latency or weight. Simple and provider-neutral, and failover is slow and unreliable in a specific way — resolvers, operating systems and application libraries cache beyond the TTL, so a long tail of clients keeps reaching the dead region for far longer than the TTL suggests.

Anycast advertises one address from many locations and lets routing choose. Failover is a route withdrawal, taking effect in seconds without client cooperation — much stronger. Long-lived TCP sessions can be re-routed mid-session, so stateful protocols need care.

Edge / proxy-based terminates connections at a global edge and forwards to a healthy origin. The most control — health-aware, per-request, capable of failing over mid-session — and it makes the edge provider a hard dependency.

The part that is consistently underestimated is health checking. Deciding a region is unhealthy is harder than it looks: partial degradation does not fail a simple check, and a check that is too sensitive causes flapping between regions, which is worse than staying put. Health signals should reflect real user-facing success, and failover should be damped so it does not oscillate.