Design Video Encoding Farm - System Design | WinJob

Encoding Farm Context and Platform Goals

A video encoding farm is the compute layer that turns mezzanine masters into adaptive bitrate (ABR) renditions. Netflix, YouTube, and Encoding.com-style platforms treat this as a batch HPC problem with product SLOs—not a single FFmpeg box.

Scale anchors for this design: ~500K UGC uploads/day, ~8 minutes average duration, evening peak ~3× daily average. Each asset may spawn 8–12 renditions (144p through 2160p) across H.264, HEVC, and optional AV1 tiers.

Headline SLIs: p95 time-to-first-playable under 3 minutes for 1080p; encode job success ≥ 99.9%; VMAF regression rate under 0.1% of published assets per week.

Three planes to name early: (1) ingest and metadata control, (2) GPU/CPU encode fleet with scheduling, (3) packaging and CDN handoff. Interviewers reward separating mezzanine immutability from mutable rendition objects.

Why "farm" not "service": Workers are fungible, jobs are parallel segments, and capacity is bin-packed across codec families. Failure domain is a segment task, not an entire asset—unless orchestration loses idempotency.

javaOne Dark Pro

1 public final class EncodeJobKey {
2   private final String assetId;
3   private final String profileId;
4   private final int segmentIndex;
5   public String dedupeKey() {
6     return assetId + ":" + profileId + ":" + segmentIndex;
7   }
8 }

pythonOne Dark Pro

1 from dataclasses import dataclass
2  
3 @dataclass(frozen=True)
4 class SegmentLease:
5     job_id: str
6     worker_id: str
7     expires_at_epoch: int

typescriptOne Dark Pro

1 interface PublishGate {
2   assetId: string;
3   vmafScore: number;
4   minVmaf: number;
5 }
6 export function canPublish(g: PublishGate): boolean {
7   return g.vmafScore >= g.minVmaf;
8 }

Why interviewers care

Video Encoding Farm interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Encoding Farm Context and Platform Goals that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Separate mezzanine, encode, and packaging boundaries explicitly
•Tie decisions to time-to-playable, quality, and cost SLOs
•Design for segment retry, re-encode, and profile drift from day one

1	public final class EncodeJobKey {
2	private final String assetId;
3	private final String profileId;
4	private final int segmentIndex;
5	public String dedupeKey() {
6	return assetId + ":" + profileId + ":" + segmentIndex;
7	}
8	}

1	from dataclasses import dataclass
2
3	@dataclass(frozen=True)
4	class SegmentLease:
5	job_id: str
6	worker_id: str
7	expires_at_epoch: int

1	interface PublishGate {
2	assetId: string;
3	vmafScore: number;
4	minVmaf: number;
5	}
6	export function canPublish(g: PublishGate): boolean {
7	return g.vmafScore >= g.minVmaf;
8	}

Encoding Farm Context and Platform Goals

Why interviewers care

Interview checkpoint

Key Highlights

Section Rescue Kit