withsoon
N
CDN5/14

Netflix System Design

CDN & Open Connect

~6 min read

Control Plane vs Data Plane

This is the single most important mental model for Netflix architecture. Interviewers expect you to state it clearly.

Control Plane — API Path

ClientPOST /playback/session
API Gateway (Zuul2)JWT validation, rate limiting
Playback ServiceEntitlement + concurrency + DRM fan-out
Steering ServicePick best OCA for client IP + title
ReturnsSigned manifest URL → OCA
Total: ~85ms. API tier is done.

Data Plane — After Manifest

ISP OCA (co-located)99% hit rate
Sub-ms disk read. >99% of requests for top titles.
Exchange OCA (IXP)<1% of requests
Internet Exchange Point OCA. Regional fallback on ISP OCA miss.
Netflix S3 Origin<0.1% of requests
Master copy. Rare fallback — only for long-tail titles or cache cold start.
300 Tbps of video bytes. API tier sees 0% of this.

OCA Selection Algorithm

Steering Service executes this on every playback session creation, returning a ranked OCA list.

1
Client IP → ASN lookup: Determine client's ISP and geographic region from IP address.
2
Filter by content availability: Only OCAs that already have the requested title cached are eligible. No point routing to an OCA that will miss.
3
Score by BGP hop proximity: Fewer BGP hops = lower latency. Prefer co-located OCAs inside the client's ISP.
4
Score by OCA health + load: OCAs report health and current load to Steering Service. Overloaded or unhealthy OCAs are demoted.
5
Return ranked list of 3–5 OCAs: Client receives a prioritized list. Tries #1 first, falls back to #2 on failure, etc.
6
Client-side retry down the list: If segment fetch fails on OCA #1, client automatically retries OCA #2 without re-contacting the API tier.

Key Terms

Interviewers expect you to define these without hesitation.

Manifest

A file (HLS: .m3u8 / DASH: .mpd) listing all available video segments and bitrate variants. Client downloads this first, then decides which quality tier to fetch.

Video Segment

A 2–10 second chunk of encoded video. Player downloads these sequentially. Losing one is recoverable; the player buffers ahead.

Adaptive Bitrate (ABR)

Player monitors bandwidth and switches between quality tiers automatically. If network slows, player switches to lower bitrate. Seamless to viewer.

OCA (Open Connect Appliance)

Netflix-custom hardware placed inside ISP data centers. Stores popular video content locally to eliminate transit hops between Netflix and ISP.

Signed URL

Manifest URL is HMAC-SHA256 signed with a 6-hour TTL. CDN validates signature before serving. Prevents URL sharing — the token is bound to the session.

Content Pre-positioning

Nightly algorithm predicts next-day popular titles (using recommendation signals + release schedule) and proactively fills OCAs during off-peak hours via ISP peering links.

Why Netflix Built Its Own CDN

At 300 Tbps, commercial CDN transit fees would exceed $500M/year. Netflix instead co-locates ~17,000 OCA appliances inside ISP networks, paying only hardware and negotiation costs. The ISP benefits too: local caching reduces their upstream transit. Cache hit rate for top-200 titles: >99%.

~17,000
OCA appliances globally
>99%
cache hit (top-200 titles)
Off-peak
nightly fill window

CDN Failure Handling

✓Edge OCA unhealthy → Steering Service demotes it; clients route to next OCA in list
✓Segment fetch times out → Player retries next OCA in manifest (client-side retry, no API call)
✓Bandwidth drops → ABR player switches to lower bitrate variant — no buffering
✓OCA cache cold (new title) → Falls back to Exchange OCA or S3 origin; fills cache for subsequent requests
✓Regional network partition → Route53 shifts DNS to healthy region; active-active multi-region absorbs traffic
📋 Say This In Interview
95% of Netflix traffic is video bytes flowing client-to-OCA. The control plane — auth, entitlement, session, DRM — handles <300ms of the user's click. After that, the API tier is completely out of the path. Netflix built Open Connect because at 300 Tbps, commercial CDN transit fees would cost hundreds of millions per year. OCA hardware inside ISP networks eliminates that cost entirely. The Steering Service picks the best OCA per client using ASN, title availability, health, and load — returning a ranked list so the player can retry down the list without another API call.
Last reviewed June 2026 · By Prasoon ParasharNumbers are interview assumptions, not real Netflix internal figures.
Was this tab useful for interview prep?