network · deep · draft

Anycast & global traffic steering

Compare published steering mechanisms: BGP anycast (Cloudflare), AWS Global Accelerator anycast product path, Azure Front Door traffic acceleration, Fastly routing-to-edge docs, and Netflix BGP-informed Open Connect steering — not one universal “global load balancer.”

In one sentence. “Steer users to the best edge” is underspecified. DNS steering, anycast, managed anycast accelerators, and control-plane URL selection with BGP inputs are different mechanisms with different failure domains.

Why it exists

Phase 8 comparison page. Cloudflare publishes anycast multi-city service IPs (`/architectures/cloudflare/`). AWS documents Global Accelerator’s anycast routing to regional endpoints. Azure Front Door documents traffic acceleration / edge entry. Fastly documents how traffic is routed to its network. Netflix Open Connect steers via control-plane URLs informed by BGP prefixes — explicitly not “classic DNS anycast CDN” language (`/architectures/netflix/`, `/architectures/edge-cdn-request-path/`). Cross-link `/architectures/amazon-aws/` and `/architectures/microsoft/` for product surfaces.

Visual walkthrough

Walkthrough

BGP anycast to a nearby PoP

Cloudflare-class teaching: same service IPs from many cities; BGP picks entry; withdraw fails over. Cross-link `/architectures/cloudflare/`.

Step 1 / 3
nearbyClientPoP A (anycast)PoP B (anycast)Edge service

Identical anycast prefixes are announced from multiple cities. The client lands on a nearby PoP via BGP.

Walkthrough

Global Accelerator vs Front Door

Product-surface comparison: AWS Global Accelerator anycast to regional endpoints vs Azure Front Door edge entry / traffic acceleration — not identical mechanisms.

Step 1 / 3
anycast entryregional endpointClientGlobal AcceleratorAWS productAzure Front DoorRegional endpointOrigin / app

Global Accelerator documents anycast entry that maps to regional endpoints with product health checks.

Walkthrough

Anycast vs DNS vs BGP-informed steering

Three mechanisms: BGP anycast PoP entry, DNS/product routing (Fastly routing docs as example), Netflix Steering URLs informed by BGP prefixes.

Step 1 / 3
ClientAnycast PoPsame IP · BGPDNS / product routingNetflix SteeringOCA (BGP prefixes)CDN edge

Anycast: client packets hit a BGP-chosen PoP using the same service IP. Failover is withdraw/convergence — not URL reissue.

Control vs data plane

Control plane

Anycast: identical prefixes announced from many sites; BGP decides entry PoP; withdraw fails traffic over. Global Accelerator / Front Door: AWS/Azure control planes map anycast or edge entry to backends per product docs. Netflix: Steering service ranks OCAs using health + BGP-learned prefixes + availability. DNS-only steering (when used) depends on resolver geography and TTLs — do not invent a company DNS graph without sources.

Data plane

Packets (anycast/GA) or HTTP fetches (OCA URLs / CDN PoPs) follow the chosen entry. Failure modes differ: anycast withdraw vs DNS TTL stickiness vs stale Steering manifests vs product health-check drain.

Request / packet path

North–south

Client → steered entry (anycast IP, accelerator, Front Door, Fastly PoP, or OCA URL) → service/bytes. Walkthroughs contrast mechanism, not invent a merged topology.

East–west

Inter-site recovery is mechanism-specific: BGP convergence for anycast; product failover for GA/Front Door; Netflix embedded vs peering preference. Unpublished private WAN TE remains unknown.

Scaling & math

No universal “best steering” latency formula across vendors. Living marketing footprints age. OMP/SD-WAN pathing is a different domain (`/architectures/sd-wan-overlay/`).

When it breaks

  • Clients shift to a farther anycast site after withdraw.

    Cause. BGP withdraw of anycast prefixes.

    Mitigation. Expect BGP-driven failover; monitor announcement health per city/PoP role.

  • Global Accelerator traffic misses a regional endpoint.

    Cause. Endpoint unhealthy or removed from the accelerator configuration.

    Mitigation. Use product health checks / endpoint groups per AWS docs — not invented TE schedules.

  • Users hit a suboptimal Front Door entry or origin path.

    Cause. Edge entry / origin config or regional issue on the product path.

    Mitigation. Debug Front Door traffic acceleration and origin health per Microsoft docs; Front Door PoP counts conflict across living docs → known unknowns on the Microsoft lens.

  • OCA URLs poor despite healthy appliances.

    Cause. Steering inputs (health, BGP prefixes, titles) stale.

    Mitigation. Treat Steering as control-plane dependency; see Netflix lens.

  • Clients stick to a bad answer longer than anycast would.

    Cause. DNS-based steering with long TTLs (general mechanism contrast — not a company claim).

    Mitigation. Understand TTL blast radius when DNS is the steering plane; prefer citing product DNS docs when asserting a vendor’s design.

Misconceptions

  • “Open Connect is a global anycast CDN like classic DNS anycast CDNs.” — Partner docs emphasize BGP-informed steering + URL selection.
  • “Global Accelerator, Front Door, and Cloudflare anycast are identical mechanisms.” — Compare product docs; do not merge.
  • “Any steering problem is solved by lowering DNS TTL alone.” — Anycast and control-plane URL steering have different failure domains.

Reference expression

Primary: Cloudflare anycast primer + architecture; AWS Global Accelerator how-it-works; Azure Front Door overview + traffic acceleration; Fastly routing-traffic; Netflix Open Connect overview + network config. review: draft. When to choose anycast teaching: same service IP, BGP-driven PoP entry. When to choose BGP-informed CDN steering: embedded caches with ISP prefix control. When not: collapse all into one unlabeled GSLB box.

Standards & sources

Known unknowns

  • Exact DNS TTL policies for most companies are unpublished.
  • Azure Front Door PoP/count figures conflict across living docs — see Microsoft lens known unknowns.
  • Private WAN TE between anycast sites is generally unpublished.

Check yourself

  1. Cloudflare’s published city entry for service IPs relies primarily on…

    • Only Netflix Steering manifests
    • BGP anycast of the same service IPs from many cities
    • Only Azure Private Link
    • Only Catalyst SD-WAN OMP

    Answer: BGP anycast of the same service IPs from many cities. Anycast primer.

  2. AWS Global Accelerator how-it-works documents…

    • Anycast routing to regional endpoints as a product path
    • Cisco SGT microseg
    • Starlink ISL TE
    • EVPN Type-2 only

    Answer: Anycast routing to regional endpoints as a product path. GA product documentation.

  3. Netflix embedded OCA steering is primarily…

    • Identical to Cloudflare Unimog
    • Control-plane URL selection informed by BGP-announced prefixes
    • Only CloudFront behaviors
    • Only Front Door WAF

    Answer: Control-plane URL selection informed by BGP-announced prefixes. Open Connect network configuration + overview.

  4. When is DNS TTL stickiness a distinct failure mode vs anycast withdraw?

    • Never — they are identical
    • When steering depends on DNS answers that caches hold, unlike BGP-driven anycast entry shift
    • Only inside a single rack
    • Only for VPN 0

    Answer: When steering depends on DNS answers that caches hold, unlike BGP-driven anycast entry shift. Mechanism contrast for teaching.

  5. Azure Front Door traffic acceleration docs are best used to teach…

    • Invented Microsoft DC Clos
    • Product edge entry / acceleration — with PoP count conflicts left as known unknowns
    • OMP TLOC colors
    • TrustSec SGACL syntax only

    Answer: Product edge entry / acceleration — with PoP count conflicts left as known unknowns. Product surface + honesty about living docs.

  6. When should you pick BGP-informed CDN steering as the teaching lens?

    • When explaining embedded ISP caches with prefix control (Open Connect class)
    • When explaining only Unimog XDP
    • When explaining collapsed core FHRP
    • When explaining Spanner TrueTime

    Answer: When explaining embedded ISP caches with prefix control (Open Connect class). When-to-choose among steering mechanisms.

  7. Collapsing GA, Front Door, Cloudflare anycast, and Open Connect into one GSLB box is…

    • Recommended pedagogy
    • Dishonest — mechanisms and failure domains differ across primaries
    • Required by SPEC
    • How Fastly Faild works

    Answer: Dishonest — mechanisms and failure domains differ across primaries. Comparison gate: no invented universal GSLB.

  8. Fastly routing-traffic documentation helps contrast…

    • How clients are directed to Fastly’s network vs pure anycast narratives
    • Cisco Validator bring-up only
    • Jupiter Clos stages
    • S3 erasure coding

    Answer: How clients are directed to Fastly’s network vs pure anycast narratives. Routing-to-Fastly primary.