Real-time, Media & AI: Live Streaming (Hotstar/Twitch), Video Conferencing (Zoom/Meet), ChatGPT-style LLM Service & Google Docs

Four designs interviewers added to their rotation in 2024–2026: one-to-50-million video fan-out, sub-150 ms group calls, GPU-bound token streaming, and convergent real-time co-editing.

About 170 minutes · 20 exam questions · 14 interview questions · 2 code labs

What you will learn

  1. Live streaming at IPL scale (1/3): requirements, estimation & how it differs from VOD
  2. Live streaming (2/3): API, data model and the ingest → transcode → package → CDN pipeline
  3. Live streaming (3/3): latency, origin protection, surviving the toss, and live comments
  4. Video conferencing (Zoom/Meet) (1/3): requirements, estimation & mesh vs MCU vs SFU
  5. Video conferencing (2/3): signaling API, NAT traversal and the meeting architecture
  6. Video conferencing (3/3): simulcast, bandwidth estimation, big rooms, recording and E2EE
  7. ChatGPT-style LLM chat service (1/3): requirements, estimation & why GPUs change everything
  8. LLM chat service (2/3): API, conversation storage and the serving architecture
  9. LLM chat service (3/3): continuous batching, KV memory, routing, token rate limits and overload
  10. Google Docs collaborative editing (1/3): requirements, estimation, API & data model
  11. Google Docs (2/3): OT vs CRDT — the transform function and when to choose which
  12. Google Docs (3/3): architecture — per-doc ownership, presence, snapshots and failover
  13. Bottlenecks, trade-offs & wrap-up across the four real-time designs
  14. Level bar: what mid, senior and staff answers include
  15. Deep-dive ladders: bad → good → great

Previous: Advanced: Leaderboard, Stock Exchange, Distributed Cache, Global Rate Limiter, Search Engine/Typeahead, Ad Serving, Recommendation System & Multi-tenant SaaS · Next: More Classics: Web Crawler, Top-K / Trending, Distributed Counters (Likes & Views), Online Judge (LeetCode) & Online Auction (eBay)