Design TikTok

Interview-Ready Approach

1) Clarify Scope and SLOs

• •Problem statement: Design TikTok's short-video platform - video creation, For You feed, content moderation, and viral distribution for 1 B+ users.
• •Design for a peak load target around 80,000 RPS (including burst headroom).
• •Monthly active users: 1,500,000,000
• •Avg session time: 95 minutes/day
• •Video uploads/day: ~30,000,000
• •Videos served/day (views): ~50,000,000,000
• •Moderation latency: < 5 minutes per video

2) Capacity Planning Method

• •Convert traffic and growth constraints into request rate, storage growth, and concurrency budgets.
• •Keep at least 2-3x safety margin per tier (ingress, compute, storage, async workers).
• •Reserve explicit latency budgets per hop so p95 can be defended in review.

3) Architecture Decisions

• •CDN: Serve static and cacheable content from edge and keep origin strictly for misses and dynamic requests.
• •Media Processing: Split ingest, transform, and delivery into independent stages with async orchestration.
• •Databases: Define a clear system-of-record and design read/write paths separately before adding optimizations.
• •Caching: Put cache on hot read paths first and pick cache-aside or write-through explicitly.
• •Analytics: Maintain separate OLTP and analytics paths; stream events into a warehouse/time-series layer.
• •Message Queues: Move non-blocking and retry-heavy work to async consumers with explicit retry and DLQ policies.

4) Reliability and Failure Strategy

• •Define cache keys and purge workflows before launch to avoid stale/global outages.
• •Store original media durably and make transforms replayable.
• •Use strong write constraints (transactions or conditional writes) and explicit backup/restore strategy.
• •Bound staleness with TTL + invalidation hooks for critical entities.
• •Version event schemas and monitor drop/late-event rates.

5) Validation Plan

• •Run one peak-load test, one dependency-degradation test, and one failover test.
• •Verify idempotency for all retried writes and async consumers.
• •Track user-facing SLOs first: p95 latency, error rate, and successful throughput.

6) Trade-offs to Call Out in Interviews

• •CDN: Long TTL improves latency/cost; short TTL improves freshness.
• •Media Processing: Pre-processing improves playback UX, but requires substantial compute/storage budget.
• •Databases: SQL gives stronger transactional guarantees; NoSQL often gives better write scaling and flexibility.
• •Caching: Higher hit rate cuts latency/cost, but stale data and invalidation bugs become primary risks.
• •Analytics: Analytics pipeline unlocks insights, but adds eventual consistency and governance overhead.

Practical Notes

• •Two-stage recommendation: candidate retrieval (from millions → thousands using approximate nearest neighbor) → ranking (ML model scores top few hundred → return top 30).
• •Feedback loop: log every user interaction (video_id, watch_duration, like, share, skip) into a streaming pipeline for near-real-time model feature updates.
• •CDN pre-warming: when a video's view velocity exceeds a threshold, proactively push it to all edge locations before it goes fully viral.

Why This Solution Works

Request path: The solution keeps ingress, service logic, and stateful dependencies separated so each layer can scale independently.

Reference flow: Mobile Clients -> DNS -> CDN Edge -> Load Balancer -> Core Service -> Primary NoSQL DB -> Redis Cache -> Event Bus

Design strengths

• •Cache sits on the read path to absorb repeated queries and keep DB pressure stable.
• •Async queue/event bus isolates bursty workloads and supports retries without blocking synchronous requests.
• •Media processing is handled by background workers so user-facing latency stays low.
• •Analytics pipeline is separated from OLTP path to avoid reporting workloads impacting transactions.

Interview defense

• •This design makes bottlenecks explicit (ingress, core compute, persistence, async workers).
• •It supports progressive scaling without re-architecting the core request path.
• •It keeps correctness-sensitive state changes in durable systems while offloading background work asynchronously.