Design Video Upload System - System Design | WinJob

Video Upload Context and Platform Goals

Frame the upload plane as a reliability-first ingress contract: creators on mobile LTE must finish multi-gigabyte mezzanine transfers without re-uploading from byte zero.

Problem framing

YouTube optimizes for watch-time; the upload path optimizes for completion rate and time-to-transcode-ready.
Vimeo skews toward pro creators with longer mezzanine masters and stricter format validation.
TikTok biases smaller vertical clips but still needs MPU for flaky networks and background upload.

Design choices

Treat upload as three planes: session control, byte storage, and post-upload validation/transcode handoff.
Anchor SLOs on upload completion %, p95 time-to-accepted, and dollars per ingested GB—not API QPS alone.
Keep mezzanine immutable; all downstream ladders reference content hash + asset version.

Deep dive

Walk the interviewer through why proxying bytes through stateless API pods collapses under 100+ Mbps per connection. Mention legal hold on raw staging blobs and regional ingress caps.

javaOne Dark Pro

1 public enum UploadPlane { SESSION, BYTES, VALIDATION }
2 public record UploadSlo(double completionRate, Duration p95Accepted) {}

pythonOne Dark Pro

1 from enum import Enum
2 class UploadPlane(Enum):
3     SESSION = "session"
4     BYTES = "bytes"
5     VALIDATION = "validation"

typescriptOne Dark Pro

1 type UploadPlane = "session" | "bytes" | "validation";
2 interface UploadSlo { completionRate: number; p95AcceptedMs: number; }

Interviewer positioning

Section 1 ties decisions to upload completion rate, cost per ingested GB, and safe handoff to transcoding—not generic storage platitudes.

Operational notes

Dashboard upload funnel: initiated → first part → 50% bytes → complete → scan pass → transcode queued.
Alert when completion rate drops 0.5% week-over-week in a region.

Scale reference

2M creator DAU; 800K finished uploads/day; ~420MB average mezzanine; 99.95% completion target.
Peak ingress planning: 180 Gbps regional; multipart 8MB parts, 10 parallel.

Why interviewers care

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

Interview checkpoint

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

Key Highlights

•Separate upload session, mezzanine blob, and transcode handoff boundaries
•Direct-to-storage multipart beats proxy upload past ~100Mbps ingress
•Idempotent completeUpload with ETag manifest prevents double-publish

1	public enum UploadPlane { SESSION, BYTES, VALIDATION }
2	public record UploadSlo(double completionRate, Duration p95Accepted) {}

1	from enum import Enum
2	class UploadPlane(Enum):
3	SESSION = "session"
4	BYTES = "bytes"
5	VALIDATION = "validation"

1	type UploadPlane = "session" \| "bytes" \| "validation";
2	interface UploadSlo { completionRate: number; p95AcceptedMs: number; }