Design Video Storage System - System Design | WinJob

Video Storage Context and Platform Goals

Frame petabyte video storage as a durability-first bytes plane plus a queryable metadata catalog—distinct from upload, transcode, and CDN playback.

Problem framing

YouTube stores mezzanine separately from packaged HLS; Netflix optimizes for origin pull during peak primetime; AWS S3 Glacier Deep Archive targets compliance archives with hour-scale restore.
Companies like YouTube, Netflix, and AWS ask this to test bytes-vs-metadata thinking, not generic CRUD APIs.
Section 1 anchors the narrative before APIs and shard math.

Design choices

Keep catalog strongly consistent; move bytes asynchronously across hot, warm, and cold pools.
Content-hash keys with version lineage; never mutate mezzanine bytes in place.
Credential broker issues object-scoped pre-signed URLs—no bucket admin keys on clients.
Lifecycle driven by age, access heat, and contract tier with explicit transition metrics.

Deep dive

Walk through failure modes relevant here: degraded erasure shard, restore backlog, catalog replica lag, orphan multipart, and legal hold blocking GC. Mention reconciliation jobs proving catalog pointers match physical inventory.

javaOne Dark Pro

1 // sec-001 — storage domain types
2 public record BlobPointer(String assetId, int version, String pool, String contentHash, long byteSize) {}
3 public enum StorageClass { HOT, WARM, COLD }

pythonOne Dark Pro

1 # sec-001 — tier transition guard
2 from enum import Enum
3  
4 class StorageClass(Enum):
5     HOT = "hot"
6     WARM = "warm"
7     COLD = "cold"
8  
9 def can_transition(current: StorageClass, target: StorageClass, legal_hold: bool) -> bool:
10     return not legal_hold and current != target

typescriptOne Dark Pro

1 // sec-001 — signed read contract
2 export interface BlobPointer {
3   assetId: string;
4   version: number;
5   pool: "hot" | "warm" | "cold";
6   contentHash: string;
7   byteSize: number;
8 }
9  
10 export function signedGetUrl(ptr: BlobPointer, region: string, ttlSec: number): string {
11   return `https://${region}.origin.example/v1/blobs/${ptr.assetId}/v/${ptr.version}?ttl=${ttlSec}`;
12 }

Interviewer positioning

Section 1 must cite measurable SLOs (hot p95 GET, restore p99, $/GB-month) and explicit boundaries with upload, transcode, and CDN—avoid vague scalability claims.

Operational notes

Runbook: spike 403 on signed GET → rotate broker signing keys.

Scale reference

Restore queue SLA: 10k jobs/hour/worker; 1 workers baseline.

Why interviewers care

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

Interview checkpoint

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

Key Highlights

•Split catalog metadata from immutable content-addressed blobs
•Catalog metadata never stores multi-GB payloads inline
•Tier transitions are async jobs with rate limits and metrics

1	// sec-001 — storage domain types
2	public record BlobPointer(String assetId, int version, String pool, String contentHash, long byteSize) {}
3	public enum StorageClass { HOT, WARM, COLD }

1	# sec-001 — tier transition guard
2	from enum import Enum
3
4	class StorageClass(Enum):
5	HOT = "hot"
6	WARM = "warm"
7	COLD = "cold"
8
9	def can_transition(current: StorageClass, target: StorageClass, legal_hold: bool) -> bool:
10	return not legal_hold and current != target

1	// sec-001 — signed read contract
2	export interface BlobPointer {
3	assetId: string;
4	version: number;
5	pool: "hot" \| "warm" \| "cold";
6	contentHash: string;
7	byteSize: number;
8	}
9
10	export function signedGetUrl(ptr: BlobPointer, region: string, ttlSec: number): string {
11	return `https://${region}.origin.example/v1/blobs/${ptr.assetId}/v/${ptr.version}?ttl=${ttlSec}`;
12	}