network · deep · draft

Edge / CDN request path

A comparison of published edge request paths — Cloudflare anycast + Unimog, Fastly PoP/routing docs, Netflix Open Connect BGP-informed steering, and AWS CloudFront delivery — not an invented “generic CDN” topology.

In one sentence. Engineers often draw one unlabeled CDN box. Real public designs differ: anycast multi-service edges, DNS/PoP routing products, embedded ISP caches with BGP prefix steering, and managed CloudFront-style distribution each publish different control vs byte planes.

Why it exists

Do not invent a generic CDN. Compare primaries: Cloudflare anycast edge with in-PoP ECMP → XDP l4drop → Unimog (`/architectures/cloudflare/`); Fastly PoP + routing-traffic docs and Autopilot/CPR research; Netflix Open Connect split control plane in AWS vs OCA bytes (`/architectures/netflix/`); AWS CloudFront how-it-works / features (`/architectures/amazon-aws/`). Cross-link `/architectures/apis-edge/` for API termination vs byte delivery, and `/architectures/anycast-traffic-steering/` for steering-mechanism contrast.

Visual walkthrough

Walkthrough

Cloudflare anycast → Unimog

Published Cloudflare path: anycast to a nearby PoP, then ECMP → XDP l4drop → Unimog → service/origin. Cross-link `/architectures/cloudflare/`.

Step 1 / 3
anycast VIPClientCF PoP (anycast)UnimogXDP l4dropOrigin (proxied)

Same service IPs announced from many cities; BGP delivers the client to a nearby PoP. Not a claim about every CDN.

Walkthrough

Netflix Open Connect split plane

Compare: control/API in AWS (Playback + Steering) vs bytes from OCA. Embedded OCAs serve only BGP-announced prefixes — not classic anycast CDN language.

Step 1 / 3
authOCA URLsPlayerPlayback Apps (AWS)Steeringcontrol planeEmbedded OCAEncoded media

Player talks to Playback Apps for auth/licensing; Steering returns ranked OCA URLs. Control plane ≠ byte plane.

Walkthrough

CloudFront product path vs Fastly PoP

Side-by-side product lenses: AWS CloudFront viewer→edge→origin docs vs Fastly PoP/routing (+ CPR path reselection research). No invented mash-up topology.

Step 1 / 3
to edgemiss → originViewerCloudFront edgeAWS productFastly PoPOrigin

CloudFront teaching path: viewer hits an edge location; cache hit serves locally; miss fetches origin per AWS how-it-works.

Control vs data plane

Control plane

Cloudflare: BGP anycast announcements + in-PoP Unimog health. Netflix: Playback Apps / Steering / cache-control in AWS ranking OCA URLs from health, BGP prefixes, title availability. Fastly: published routing-to-Fastly and Autopilot/CPR themes for path intelligence — cite those docs, not invented controllers. CloudFront: AWS distribution configuration (origins, behaviors, edge locations as product surface). Exact multi-CDN orchestration for unnamed companies is unknown.

Data plane

Cloudflare: client → anycast VIP → nearby PoP → Unimog-selected server → origin (proxied) or local service. Netflix: client fetches HTTP/HTTPS bytes from steered OCA URLs — not from Playback Apps hosts; embedded OCAs serve only BGP-announced customer prefixes. CloudFront: viewer → edge location → (cache hit) or origin fetch per AWS docs. Fastly: viewer → PoP per routing docs — do not invent unpublished backbone TE.

Request / packet path

North–south

Viewer/player → edge/PoP/OCA → origin or encoded media. Whether the first hop is anycast IP, DNS/product routing, or control-plane URL steering is the published differentiator — walkthroughs compare them side by side.

East–west

Inter-PoP failover: Cloudflare emphasizes anycast/BGP withdraw; Netflix prefers embedded AS40027 short path vs PoP peering AS2906 backup/fill; CloudFront/Fastly publish product edge maps and routing concepts without claiming identical TE. Corporate fabrics behind edges remain company-lens known unknowns.

Scaling & math

Peering policy and marketing footprints age — re-verify living numbers. Open Connect preference rules are qualitative (embedded preferred when prefix heard). CloudFront/Fastly do not publish a universal hit-ratio formula usable as capacity planning here. No invented PoP inventories.

When it breaks

  • Users pin to a farther Cloudflare city after PoP withdraw.

    Cause. BGP withdraw of anycast prefixes for a stressed/failed datacenter.

    Mitigation. Treat anycast failover as the published inter-PoP recovery theme — see Cloudflare lens.

  • Players get poor OCA choices; quality drops despite healthy byte caches.

    Cause. Steering/cache-control inputs stale or control-plane outage.

    Mitigation. Separate control SLOs from byte-delivery SLOs on the Open Connect split plane.

  • ISP customers miss embedded OCAs and fall back to PoP/peering.

    Cause. ISP did not announce customer prefixes to the embedded OCA over BGP.

    Mitigation. Verify BGP announcements per Open Connect network-configuration docs.

  • CloudFront/Fastly-class miss path fetches origin; latency rises.

    Cause. Cache miss or behavior/origin misconfiguration on the product path.

    Mitigation. Debug distribution behaviors/origins per CloudFront docs; do not invent hit ratios.

  • Edge connection quality degrades until path reselection recovers.

    Cause. Poor first-path choice or changing underlay conditions at the edge.

    Mitigation. Fastly CPR/Autopilot publish connection path reselection research — cite papers, not invented TE.

Misconceptions

  • “Every CDN is global anycast like Cloudflare.” — Netflix publishes BGP-informed steering + URL selection; CloudFront is an AWS product path.
  • “Playback/API hosts serve the video bytes.” — Open Connect: Steering returns OCA URLs; clients fetch bytes from appliances.
  • “One generic CDN diagram covers Cloudflare, Fastly, Netflix, and CloudFront.” — Compare published patterns; do not invent a mash-up topology.

Reference expression

Primary: Cloudflare anycast primer, architecture/BPF, Unimog; Fastly PoP docs, routing-traffic, Autopilot, CPR NSDI; Netflix Open Connect Overview + network config; AWS CloudFront how-it-works + features. review: draft. When to choose which teaching lens: anycast multi-service edge (Cloudflare), BGP-informed embedded cache (Netflix), managed distribution product (CloudFront), research-grade PoP path reselection (Fastly CPR/Autopilot). When not: draw one unlabeled “CDN” and assert it matches all four.

Standards & sources

Known unknowns

  • Living PoP/city marketing counts age — re-verify; not frozen inventory.
  • Hit ratios, fill schedules, and unpublished backbone TE are not asserted.
  • Multi-CDN orchestration policies for unnamed enterprises are unknown.

Check yourself

  1. Cloudflare’s published inter-city steering for service IPs is primarily…

    • DNS-only without BGP
    • Anycast: same IPs announced from many cities
    • Open Connect AS40027 only
    • Invented private WAN TE

    Answer: Anycast: same IPs announced from many cities. Cloudflare anycast primer / architecture posts.

  2. After Netflix Steering returns OCA URLs, video bytes come from…

    • Playback Apps hosts in AWS
    • Open Connect appliances over HTTP/HTTPS
    • Only CloudFront distributions
    • Unimog alone

    Answer: Open Connect appliances over HTTP/HTTPS. Open Connect Overview split plane.

  3. Embedded OCAs serve client IPs when…

    • Any global anycast VIP matches
    • The ISP announces those customer prefixes to the OCA over BGP
    • STP converges
    • OMP assigns a TLOC color

    Answer: The ISP announces those customer prefixes to the OCA over BGP. Open Connect network configuration.

  4. AWS CloudFront’s published teaching path is…

    • Invented Netflix corporate Clos
    • Viewer → edge location → cache hit or origin fetch (product docs)
    • Only SGT microseg
    • Only Fastly Faild

    Answer: Viewer → edge location → cache hit or origin fetch (product docs). CloudFront how-it-works.

  5. Why avoid a single unlabeled “generic CDN” diagram here?

    • Because SPEC forbids teaching edges
    • Published patterns differ (anycast vs BGP steering vs product distribution) — inventing a mash-up is dishonest
    • Because only Azure Front Door exists
    • Because CDNs never use HTTP

    Answer: Published patterns differ (anycast vs BGP steering vs product distribution) — inventing a mash-up is dishonest. Phase 8 ship rule: compare primaries; no invented generic CDN.

  6. Fastly CPR (NSDI 2021) is primary evidence for…

    • Cisco OMP
    • Connection path reselection at the edge
    • EVPN Type-5 only
    • Starlink ISL TE

    Answer: Connection path reselection at the edge. CPR paper title/theme.

  7. Cloudflare in-PoP path after ECMP includes…

    • XDP l4drop then Unimog L4 LB
    • Only MPLS L3VPN
    • Only ISE SGT
    • Only Direct Connect gateways

    Answer: XDP l4drop then Unimog L4 LB. Architecture/BPF + Unimog posts.

  8. When is CloudFront the better teaching lens than Open Connect?

    • When explaining embedded ISP BGP prefix steering for Netflix appliances
    • When explaining a managed AWS distribution product path (behaviors, origins, edge locations)
    • When teaching Unimog XDP
    • When teaching Catalyst SD-WAN VPN 0

    Answer: When explaining a managed AWS distribution product path (behaviors, origins, edge locations). When-to-choose among published patterns.