Anycast at the Edge

The Trick

 UNICAST (normal): one IP → one physical location
 ANYCAST:          one IP announced from MANY locations;
                   internet routing (BGP) delivers each user
                   to the NEAREST announcing point

 104.16.0.1 announced simultaneously from:
   Frankfurt POP ──► European users' packets arrive here
   Tokyo POP    ──► Japanese users arrive here  
   Ashburn POP  ──► US-East users arrive here

 no DNS tricks, no redirects, no client config.
 ROUTING ITSELF is the load balancer.

Why It Wins for CDNs

 ✓ PROXIMITY AUTOMATIC: BGP prefers shortest AS-path ≈ nearest POP
 ✓ FAILOVER FREE:       POP dies → its announcements withdrawn →
                        traffic reroutes to neighbors in seconds
                        (no health-check machinery at all)
 ✓ LOAD ABSORPTION:     regional surges spill to adjacent POPs
                        naturally
 ✓ SIMPLE CLIENTS:      same hostname/IP everywhere; zero logic

The Mechanics Underneath

 each POP runs BGP with upstream providers:

 POP announces:  "reach 104.16.0.0/16 through ME"
 multiple announcements exist globally;
 routers pick best path (typically fewest hops/ASNs)

 consequences and subtleties:
 - granularity = network topology, NOT geographic distance:
   a user might reach the "wrong" city if paths say so —
   usually fine (path-short = latency-short anyway)
 - route FLAPS can move users between POPs mid-session →
   edge protocols must tolerate client migration
   (that's partly why QUIC connection-IDs matter)

Anycast + Unicast Hybrid Patterns

 pure anycast has limits; production designs blend:

 - anycast for INGRESS (user→edge): the win zone
 - unicast/DNS-GEO for REGION SELECTION when state locality matters:
     user lands at Manila edge via anycast,
     edge consults geo-DNS/config to pick their DATA region
 - weighted anycast: announce with prepends to shape traffic ratios
 
 the layering: anycast answers "which edge?",
               other layers answer "which database/region?"

Failure Behavior Worth Knowing

 POP outage sequence (automatic):
 1. POP stops announcing (BGP session down)
 2. routes withdraw within seconds-minutes globally
    (BGP convergence is fast but not instant)
 3. affected users' packets flow to next-nearest POP
 4. caches there are COLDER → brief hit-rate dip → self-heals

 vs unicast failover: no LB config, no health checks,
 no DNS TTL waiting. the internet's own machinery does it.

Interview Framing

“How do users worldwide reach one service endpoint fast?” scored answer: anycast announcement from many POPs, BGP proximity routing, automatic failover via route withdrawal, plus honest limits (topology-vs-geography granularity, mid-session POP migration). Contrasting it against DNS-geo-routing (lookup-time vs routing-time) demonstrates real networking depth — interviewers probe exactly this distinction.