Design CDN Routing Optimization - System Design | WinJob

Problem Statement and Video CDN Routing Context

Netflix/Akamai/Cloudflare interviews ask how you steer HLS/DASH segments to the best edge—not just cache static files. Routing optimizes QoE (segment TTFB, rebuffer) and cost across multi-CDN footprints.

Key points

Steering optimizes QoE first, egress cost second, with hard caps per CDN contract.
Manifest rewrite is the fastest lever; DNS/Anycast changes follow on slower TTL.
RUM closes the loop—without measurement, routing is guesswork.

Deep dive (understanding)

Separate steering control plane (policy, RUM aggregates, GeoDNS maps) from segment data plane (HTTP GET of .ts/.m4s). Players request manifests first; manifest URLs encode which CDN/PoP serves bytes. Open Connect-style private edges change economics but not the routing math.

Capacity and SLO anchors

200M concurrent streams peak, 6s segments → ~33M segment RPS; p95 segment TTFB < 80ms on warm colo; rebuffer ratio < 0.5%.

Failure and degradation (unique to Problem Statement and Video CDN Routing Context)

If a PoP shows elevated segment TTFB for 3 consecutive windows, drain it gradually: halve weight, wait 2 minutes, verify rebuffer SLO, then zero. If manifest API fails, serve last signed snapshot from regional cache—never block playback on control-plane outage.

Cost and multi-CDN lens

Weighted steering respects per-CDN max share (e.g., premium network ≤30%). Shifting 1% of 200 Tbps saves millions monthly—tie routing to finance dashboards, not only QoE.

Security and abuse

Sign manifests; rate-limit RUM ingest; validate session HMAC to prevent score poisoning. TLS terminates at edge; colo IDs must not leak internal topology in client-visible errors.

javaOne Dark Pro

1 public final class ColoScore {
2   private final String popId;
3   private final double rumP95Ms;
4   private final double capacityHeadroom;
5  
6   public double weightedScore(double wLatency, double wCapacity) {
7     return wLatency * (1000.0 / Math.max(rumP95Ms, 1.0)) + wCapacity * capacityHeadroom;
8   }
9 }

pythonOne Dark Pro

1 from dataclasses import dataclass
2  
3 @dataclass(frozen=True)
4 class RumSample:
5     pop_id: str
6     segment_ttfb_ms: int
7     rebuffer: bool
8  
9 def ewma(prev: float, sample: float, alpha: float = 0.2) -> float:
10     return alpha * sample + (1 - alpha) * prev

typescriptOne Dark Pro

1 interface SteeringPolicy {
2   version: number;
3   cdns: { id: string; weight: number; maxShare: number }[];
4   rumMinSamples: number;
5 }
6  
7 export function pickCdn(policy: SteeringPolicy, scores: Record<string, number>): string {
8   const ranked = Object.entries(scores).sort((a, b) => b[1] - a[1]);
9   return ranked[0]?.[0] ?? policy.cdns[0].id;
10 }

Why interviewers care

CDN Routing Optimization interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem Statement and Video CDN Routing Context that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Netflix/Akamai/Cloudflare interviews ask how you steer HLS/DASH segments to the best edge—
•Separate **steering control plane** (policy, RUM aggregates, GeoDNS maps) from **segment d
•Close the RUM feedback loop with hysteresis and CDN caps.

1	public final class ColoScore {
2	private final String popId;
3	private final double rumP95Ms;
4	private final double capacityHeadroom;
5
6	public double weightedScore(double wLatency, double wCapacity) {
7	return wLatency * (1000.0 / Math.max(rumP95Ms, 1.0)) + wCapacity * capacityHeadroom;
8	}
9	}

1	from dataclasses import dataclass
2
3	@dataclass(frozen=True)
4	class RumSample:
5	pop_id: str
6	segment_ttfb_ms: int
7	rebuffer: bool
8
9	def ewma(prev: float, sample: float, alpha: float = 0.2) -> float:
10	return alpha * sample + (1 - alpha) * prev

1	interface SteeringPolicy {
2	version: number;
3	cdns: { id: string; weight: number; maxShare: number }[];
4	rumMinSamples: number;
5	}
6
7	export function pickCdn(policy: SteeringPolicy, scores: Record<string, number>): string {
8	const ranked = Object.entries(scores).sort((a, b) => b[1] - a[1]);
9	return ranked[0]?.[0] ?? policy.cdns[0].id;
10	}