cerf
Netflix

Netflix

Netflix: control plane on AWS, video on its own CDN

Why the biggest streamer runs everything on AWS except the one thing you'd expect: the video bytes.

Two networks, one product

Netflix splits its architecture along the cleanest seam in networking: control plane versus data plane. Everything that happens before you press play (sign-in, browsing, recommendations, DRM licensing, playback orchestration) runs on AWS, active across four Regions, on thousands of auto-scaling compute groups with Aurora, EMR, and EKS underneath.

The video itself never comes from AWS. It comes from Open Connect, Netflix's own CDN: purpose-built cache appliances (OCAs) that Netflix gives to ISPs for free, racked inside the ISP's own network or peered at internet exchanges.

NetflixTwo networks, one product: AWS control plane vs Open Connect data plane
Two networks, one product: AWS control plane vs Open Connect data planeAWS control planeISP network OCA embeddedInternet exchange (IXP)4 active-active Regionscontrol-plane DNSsign-in, browse, playroute requestsorchestrate playbackissue licenseread/write metadatarecommendationscontent masternightly fillnightly fillBGPvideo bytesvideo bytes (peered)Viewersweb, mobile, TV appsRoute 53control-plane DNSEdge / APIgatewayElastic Load Balanc…Microservicesthousands of auto-s…Playbackorchestrationmanifest, session s…DRM licensingcontent keysAuroramember & metadata s…EMRrecommendation pipe…S3 originmaster content storeISP edge routerBGP peer with OCAOCA (in ISP)cache appliance, BG…OCA (at IXP)peered at internet …
Trace
Everything that happens before you press play (sign-in, browsing, recommendations, DRM licensing, playback orchestration) runs on AWS across four active Regions on auto-scaling compute with Aurora, EMR and EKS. The video bytes themselves never come from AWS: they are served by Open Connect Appliances (OCAs) racked inside ISP networks and peered at internet exchanges. S3 is the master content store that feeds the appliances, but user streams terminate at the nearest OCA, not in a Region.

How traffic finds the cache

Each OCA announces routes over BGP sessions with the ISP, and Netflix's steering services in AWS decide, per client, which appliance serves the stream. Popular titles are pushed to appliances during off-peak windows (the nightly fill), so at prime time the bytes travel meters of ISP network instead of transit internet.

This is the pattern behind every 'optimize global delivery' exam scenario: terminate user traffic as close to the user as possible, keep the origin path for fills and misses, and let BGP plus health-aware steering pick the door. CloudFront plus Route 53 or Global Accelerator is the same shape inside AWS.

NetflixHow traffic finds the cache: BGP announcements plus AWS steering
How traffic finds the cache: BGP announcements plus AWS steeringAWS control planeISP networkInternet exchange (IXP)Control-plane Region (1 of 4)1. request streamresolve best OCAhealth + fill mapreport load/fill2. OCA to useBGP announcecontent master3. nightly fill (off-peak)4. video bytes (local)peered fallbackClient playerrequests a streamPlayback APIreturns steering URLOpen Connectsteeringper-client OCA choi…OCA health &fill maptelemetryS3 originmaster content storeFill managerorchestrates nightl…ISP edge routerBGP peerOCA (filled)serves this clientOCA (backup)same ISP, standbyOCA (IXP peer)internet-exchange f…
Trace
Each OCA announces routes over BGP sessions with the ISP, and Netflix's steering services in AWS decide, per client, which appliance serves the stream. Popular titles are pushed to the appliances during off-peak windows (the nightly fill) driven by fill state and appliance health, so at prime time the client's video request terminates at a filled OCA meters away instead of crossing the transit internet. This is the shape behind every 'optimize global delivery' scenario: BGP plus health-aware steering picks the door, and the origin path is kept only for fills and misses.

What this teaches for the exam

When a question separates 'API traffic' from 'bulk content delivery,' treat them as different networks with different answers: edge caching and steering for content, Regional entry points and failover for the control plane.

The multi-Region active-active control plane also matters: four Regions with health-based traffic management is the reference architecture behind Route 53 failover and Application Recovery Controller questions.

NetflixWhat this teaches: multi-Region active-active control plane and the AWS analog
What this teaches: multi-Region active-active control plane and the AWS analogAWS control planeISP / Open ConnectSame shape, AWS-nativeFour active-active Regionsus-east-1us-west-2Regions 3 & 4DNS: health + latencyreadiness + routing controlactiveactiveactivecross-Region replicationbulk video (Open Connect)content edge cacheor Global AcceleratorGlobal usersAPI + content reque…Route 53health-based / late…ApplicationRecovery…readiness + routing…Edge / ALBus-east-1Control-planeservicesus-east-1Aurora (active)us-east-1Edge / ALBus-west-2Control-planeservicesus-west-2Aurora (active)us-west-2Sameactive-active…Regions 3 & 4Open Connectbulk video, separat…CloudFrontAWS-native edge cac…GlobalAcceleratorAWS-native steering
Trace
The control plane runs active-active across four Regions with health-based traffic management: Route 53 resolves users to a healthy nearby Region, and Application Recovery Controller governs readiness and routing so traffic shifts away from an impaired Region. This is the reference architecture behind Route 53 failover and Application Recovery Controller questions. Content delivery stays a separate network (Open Connect); the AWS-native equivalent of that same shape is CloudFront plus Route 53 or Global Accelerator. When a question separates 'API traffic' from 'bulk content delivery,' treat them as two networks with different answers.

More diagrams

Netflix: AWS control plane, Open Connect data plane
Netflix: AWS control plane, Open Connect data planeAWS control planeISP network or IXP peering4 active Regionsauth, browse, playOCA URLvideo bytesnightly content fillViewersplay requestPlaybackservicesEC2, EKS fleetsData platformS3, EMR, AuroraSteeringclient to best OCAOpen ConnectOCA, BGP announced
Trace
Everything before you press play runs on AWS across four Regions. The video bytes themselves come from Open Connect Appliances embedded inside ISP networks and at internet exchanges, announced over BGP, so streaming traffic never crosses the transit internet or AWS.

Sources

Practice what you just read

6 questions from this architecture

Loading…