company · deep · draft

Microsoft / Azure (edge & host SDN)

Microsoft’s public networking story centers on Azure Front Door (anycast edge + private WAN), host SDN (VFP / AccelNet), Internet peering (AS8075 policy), and Cascara edge TE research — plus Service Fabric as a historical microservices platform paper. Living Front Door PoP counts disagree; treat them as known unknowns.

In one sentence. Teaching Azure’s published edge and host-SDN designs without inventing DC Clos inventories or freezing conflicting living-doc PoP counts as fact.

Why it exists

Azure Front Door terminates HTTP(S) near users via Anycast PoPs connected over Microsoft’s private WAN (Learn overview + traffic-acceleration docs; 2022 Standard/Premium blog). Host SDN (VFP NSDI 2017; AccelNet NSDI 2018) implements most Azure virtual networking in the host / SmartNIC path. Cascara (NSDI 2021) studies cloud-edge TE under bandwidth pricing. Cross-links: `/architectures/apis-edge/` (edge termination), `/architectures/spine-leaf-clos/` (Clos teaching — not Azure fabric disclosure).

Visual walkthrough

Walkthrough

Front Door Anycast → private WAN → origin

Published Front Door path: Anycast to a nearby PoP, edge processing / split TCP, then Microsoft private WAN toward healthy origins. Not a PoP count claim.

Step 1 / 3
Anycast HTTP(S)Internet clientFront Door PoPMicrosoft private WANHealthy origin

DNS/HTTP traffic uses Anycast so the client reaches a nearby Front Door PoP. Primary and fallback VIP rings handle preference and overflow as documented.

Walkthrough

Host SDN: VFP → AccelNet SmartNIC

Teaching path from guest VM through VFP host virtual switch with AccelNet SmartNIC offload. Fabric beyond the host remains undisclosed.

Step 1 / 3
Guest VMVFP (host vSwitch)tenant policySmartNIC (AccelNet)Azure fabric (undisclosed)

Almost all Azure virtual networking features are implemented in host virtual switches via VFP layered match-action tables (NSDI 2017).

Walkthrough

Service Fabric platform (2018 paper era)

EuroSys 2018 Service Fabric as a dense microservices platform on shared machine pools. Historical — not a live 2026 Azure scheduler map.

Step 1 / 2
Microservice AMicroservice BService Fabric (2018 paper)era-labeledShared machine pool

Service Fabric provides a distributed platform for building microservices on shared machine pools, with design lineage since the early 2000s and production use described since ~2007 in the paper.

Control vs data plane

Control plane

Front Door steers via Anycast VIP rings (outer preferred, inner overflow/failover) and health-aware origin selection after edge processing (traffic-acceleration). Cascara is research TE at cloud PoPs — not a public B4-equivalent WAN controller dump. Service Fabric (EuroSys 2018) describes a historical dense microservices platform underpinning named Azure services as of the paper — treat as era-labeled, not a 2026 scheduler inventory.

Data plane

Client → Front Door PoP (Anycast) → Microsoft private WAN → healthy origins (including Private Link origins on Premium per product docs). Inside the Azure host: VFP layered match-action tables; AccelNet offloads VFP to FPGA SmartNICs for near-native performance. Do not invent regional Clos port counts.

Request / packet path

North–south

Internet client hits Front Door Anycast; split TCP at the edge; then private-WAN forwarding toward origins. Internet peering at Microsoft edge is governed by the published AS8075 / AS-SET policy (peers must filter AS12076 ExpressRoute). ExpressRoute is a separate private-connectivity product.

East–west

Tenant east–west is mediated by host SDN policy (VFP/AccelNet papers). Full Azure datacenter fabric generations and oversubscription remain unpublished → known unknowns. Cascara addresses inter-domain bandwidth cost at the edge, not intra-DC Clos math.

Scaling & math

Living Front Door docs disagree on edge footprint (“more than 118 edge locations across 100 metro areas” vs “over 150 edge locations” on traffic acceleration) — do not pick a single number as authoritative. Cascara and AccelNet paper figures are era-labeled research/product claims, not live rack inventories.

When it breaks

  • Users reach a Front Door PoP but origins fail after edge processing.

    Cause. Origin health probes fail or Private Link / backend path degraded.

    Mitigation. Treat origin health and Premium Private Link paths as hard dependencies per Front Door docs; separate edge SLO from origin SLO.

  • Traffic shifts to farther PoPs or inner VIP ring under PoP stress.

    Cause. Outer Anycast VIP ring withdraw/failover themes as published.

    Mitigation. Expect primary/fallback VIP rings; do not invent unpublished TE objectives.

  • Tenant VNET isolation or ACL behavior unexpected on a VM path.

    Cause. Host virtual-switch (VFP) policy misconfiguration or incomplete offload.

    Mitigation. Debug as host SDN / SmartNIC path per VFP and AccelNet papers — not as invented ToR Clos.

  • VM networking falls back toward software VFP performance under offload issues.

    Cause. SmartNIC / AccelNet offload path unavailable for VFP tables.

    Mitigation. AccelNet paper: FPGA SmartNIC offload for near-native rates; treat offload health as a host data-plane concern.

  • Unexpected ExpressRoute (AS12076) routes appear via Internet peering.

    Cause. Peer failed to filter AS12076 as required by Microsoft peering policy.

    Mitigation. Peers must filter AS12076; Internet peering ≠ ExpressRoute product path.

Misconceptions

  • “Front Door Learn pages agree on a single PoP count.” — Living docs conflict (118+ vs 150+); record the conflict in Known unknowns.
  • “Service Fabric 2018 figures are today’s live Azure scheduler map.” — Era-label the paper; current SF vs AKS split is unknown.
  • “VFP/AccelNet papers disclose Azure regional Clos port counts.” — They document host SDN / SmartNIC policy, not building floorplans.

Reference expression

Primary: Front Door overview + traffic acceleration + 2022 blog; Internet peering policy; VFP NSDI 2017; AccelNet NSDI 2018; Cascara NSDI 2021; Service Fabric EuroSys 2018 (historical). review: draft. Thin honesty on PoP counts and DC fabric beats invented topology.

Standards & sources

Known unknowns

  • Authoritative single Front Door PoP / edge-location count — living Learn docs conflict (118+ vs 150+ language).
  • Full Azure datacenter fabric topology (spine-leaf generations, oversubscription) beyond host SDN papers.
  • Detailed Microsoft WAN TE comparable to Google B4 (objectives, utilization, site inventory).
  • Whether every named Service Fabric consumer from the 2018 paper still runs on SF vs AKS or other substrates.
  • Physical interconnect capacity and peering graphs per metro beyond policy + secondary operator data.
  • How Bing / M365 / LinkedIn traffic share Front Door vs private stacks beyond high-level blog claims.

Check yourself

  1. Azure Front Door as published primarily uses…

    • Only DNS steering with no Anycast
    • Anycast edge PoPs plus Microsoft private WAN toward origins
    • Invented Azure Clos port counts
    • Open Connect OCAs

    Answer: Anycast edge PoPs plus Microsoft private WAN toward origins. Learn docs: Anycast PoPs connected via Microsoft private WAN.

  2. VFP (NSDI 2017) primarily documents…

    • Bing private WAN TE
    • Host virtual-switch layered match-action for Azure SDN
    • Starlink laser mesh
    • Cloudflare Unimog

    Answer: Host virtual-switch layered match-action for Azure SDN. VFP is the Azure host SDN virtual switch paper.

  3. AccelNet (NSDI 2018) offloads VFP to…

    • Customer laptops only
    • FPGA SmartNICs for near-native VM networking performance
    • Netflix OCAs
    • PeeringDB ASNs

    Answer: FPGA SmartNICs for near-native VM networking performance. AccelNet: SmartNIC offload of host SDN tables.

  4. Microsoft Internet peering policy asks peers to filter…

    • AS13335 only
    • AS12076 (ExpressRoute) routes
    • All IPv6
    • Nothing — open by default

    Answer: AS12076 (ExpressRoute) routes. Published peering policy: filter AS12076.

  5. How should Front Door PoP counts be treated?

    • Pick 118 or 150 and freeze forever
    • As living-doc drift / known unknowns when docs conflict
    • Invent a third unpublished number
    • Ignore Learn docs entirely

    Answer: As living-doc drift / known unknowns when docs conflict. Phase 7 ship rule: reconcile conflicts into knownUnknowns.

  6. Service Fabric EuroSys 2018 should be presented as…

    • Live 2026 scheduler inventory for all Azure
    • Historical platform paper with era-labeled claims
    • Proof of Azure Clos radix
    • Cloudflare anycast primer

    Answer: Historical platform paper with era-labeled claims. Backlog: SF paper is historical; current split is unknown.

  7. Cascara (NSDI 2021) studies…

    • Starlink ISL routing
    • Cloud-edge traffic engineering under bandwidth cost / latency bounds
    • Airbnb SmartStack
    • S3 erasure coding recipes

    Answer: Cloud-edge traffic engineering under bandwidth cost / latency bounds. Cascara is Microsoft cloud-edge TE research.

  8. What belongs in Known unknowns for this lens?

    • That Front Door uses Anycast (already published)
    • Authoritative single PoP count, full Azure Clos generations, B4-like WAN TE dump
    • AS8075 existence in the peering policy
    • That VFP is a host virtual switch

    Answer: Authoritative single PoP count, full Azure Clos generations, B4-like WAN TE dump. Thick edge/host SDN; thin inventable DC/WAN inventory.