Design Twitch - System Design | WinJob

Problem Statement and Live Streaming Context

Twitch is a live interactive broadcast platform: one creator can go live to millions of viewers while thousands of messages per second flow through the same channel. Unlike VOD, there is no finished file to cache ahead of time—every second of video and chat is produced in real time.

Interviewers probe whether you understand glass-to-glass latency (camera → encoder → CDN → player), hot-channel skew (0.1% of streams hold most CCU), and social monetization (subs, bits, raids) that must not block playback.

Stakeholder	Primary pain	Architectural lever
Broadcaster	Stable ingest, instant "live" badge	Regional RTMP/WebRTC ingest + health checks
Viewer	Low delay, smooth ABR	LL-HLS (~2s parts) + CDN edge
Trust & Safety	Spam, hate raids	Async moderation bus off chat Kafka topic

Anchor scale verbally: ~35M DAU, ~6.5M peak platform CCU, single esports finals channels above 500K CCU, chat bursts beyond 100K msgs/sec on mega-events.

javaOne Dark Pro

1 public final class IngestSession {
2   private final String channelId;
3   private final String streamKey;
4   private volatile long sequence;
5  
6   public IngestSession(String channelId, String streamKey) {
7     this.channelId = channelId;
8     this.streamKey = streamKey;
9   }
10  
11   public long nextChunkSequence() {
12     return ++sequence;
13   }
14 }

pythonOne Dark Pro

1 from dataclasses import dataclass
2  
3 @dataclass(frozen=True)
4 class ChatEnvelope:
5     channel_id: str
6     user_id: str
7     body: str
8     server_seq: int
9  
10 def is_spam_burst(rate_per_sec: float, ceiling: float = 25.0) -> bool:
11     return rate_per_sec > ceiling

typescriptOne Dark Pro

1 interface PlaybackTicket {
2   channelId: string;
3   rendition: "1080p60" | "720p60" | "480p30";
4   edgePop: string;
5   expiresAtMs: number;
6 }
7  
8 export function isTicketValid(ticket: PlaybackTicket, nowMs: number): boolean {
9   return nowMs < ticket.expiresAtMs;
10 }

Close this section by stating you will size ingest Mbps, transcode fleet, CDN egress, and chat gateway pools separately—never as one generic "streaming service."

Why interviewers care

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

Interview checkpoint

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

Key Highlights

•Live + chat are separate failure domains
•Quantify CCU before naming services
•Hot channels need shield + dedicated transcode

1	public final class IngestSession {
2	private final String channelId;
3	private final String streamKey;
4	private volatile long sequence;
5
6	public IngestSession(String channelId, String streamKey) {
7	this.channelId = channelId;
8	this.streamKey = streamKey;
9	}
10
11	public long nextChunkSequence() {
12	return ++sequence;
13	}
14	}

1	from dataclasses import dataclass
2
3	@dataclass(frozen=True)
4	class ChatEnvelope:
5	channel_id: str
6	user_id: str
7	body: str
8	server_seq: int
9
10	def is_spam_burst(rate_per_sec: float, ceiling: float = 25.0) -> bool:
11	return rate_per_sec > ceiling

1	interface PlaybackTicket {
2	channelId: string;
3	rendition: "1080p60" \| "720p60" \| "480p30";
4	edgePop: string;
5	expiresAtMs: number;
6	}
7
8	export function isTicketValid(ticket: PlaybackTicket, nowMs: number): boolean {
9	return nowMs < ticket.expiresAtMs;
10	}

Problem Statement and Live Streaming Context

Why interviewers care

Interview checkpoint

Key Highlights

Section Rescue Kit