TikTok System Design Interview Questions

Accept uploads to edge storage, virus-scan and extract metadata, transcode to multiple bitrates and aspect ratios, generate thumbnails and previews, and publish artifacts to a CDN with a job queue that prioritizes creators awaiting immediate publishing.

Use write-optimized endpoints that append events to a durable log, fan out via pub/sub to counters and timelines, and serve reads from caches with periodic reconciliation to avoid hot keys on viral videos.

Create a notifications service that consumes engagement events, applies per-user budgets, dedupes, and targets channels (push, in-app, email) with rate limits and preference rules; store delivery receipts for retries and analytics.

Use region-aware ingest, transcode to low-latency HLS/LL-HLS, deliver via CDN, and back chat with WebSockets and sharded rooms; process gifts in a ledgered microservice with anti-fraud checks and near real-time creator payouts.

Provide low-latency online features (recent watch time, session context) and consistent offline features (creator stats, video quality) with versioned definitions, backfills, and point-in-time correctness guarantees.

For users, bootstrap with popular and diverse content plus rapid preference elicitation; for videos, guarantee exploration slots and creator-quality priors, then promote or demote based on early engagement rates.

Define immutable events with IDs, timestamps, positions, and device context; include session IDs and content IDs, and enforce contracts so both online ranking and offline analytics interpret them consistently.

Run a separate ad retrieval and quality ranker, then blend ads into the organic feed under frequency caps and pacing goals; enforce advertiser budgets and brand safety while optimizing predicted watch and click-through.

Maintain a double-entry ledger for coins and gifts, apply device and behavioral risk scoring, use velocity limits and step-up verification, and reconcile creator payouts with audit trails.

Sign deep links with short-lived tokens, restore playback position and UI state, and prefetch video metadata and the next candidates; handle expired links by redirecting to the video page safely.

Reserve a small exploration budget per session or topic, try diverse candidates, and update per-user and global priors online; monitor regret and clamp exploration if quality drops.

Serve rapidly changing signals (recent interactions, session context) from an online store with low TTLs; compute heavy features offline with scheduled backfills and publish freshness metadata so rankers can fall back if stale.

Index video and user embeddings in a sharded ANN service, retrieve top-k similar items, then filter by policy, language, and diversity before sending to ranking; monitor recall and tail latency.

Produce ladder renditions, package HLS/LL-HLS with segment sizes tuned for quick start, and let the player pick renditions based on real-time throughput with fast up/down switches.

Pre-warm hot content in edge PoPs, use origin shield and segment prefetch, coalesce requests, and cache manifests aggressively while respecting creator privacy and takedowns.

Favor HTTP/3 for better handshake and loss recovery on flaky networks, while keeping HTTP/2 fallback; measure startup time and rebuffering to decide per-region rollout.

Prefetch short previews or the first segments only when on Wi-Fi or strong signal, honor user settings and data caps, and abort prefetch on scroll direction changes to save bandwidth.

Tie SLOs to user retention and watch time deltas, publish p50/p95 startup targets per region and device class, and enforce them with canaries and automated rollback if regressions exceed thresholds.

Pick based on operational maturity and features: Kafka for ecosystem breadth and simplicity, Pulsar for tiered storage and multi-tenancy; standardize an event contract so producers and consumers are portable.

Apply client-side batching and rate caps, shed noncritical events first, and use bounded queues with load-aware throttling on the server; drop or delay low-value analytics before core engagement or safety paths.

Use WebSockets for bidirectional chat and presence; use SSE for one-way updates when simplicity and proxy-friendliness matter; in both cases, include heartbeats, resume tokens, and rate limits.

Attach unique event IDs with producer timestamps, dedupe at the edge and in stream processors, and make writes idempotent so retries do not inflate counters.

Use at-least-once with idempotent sinks for scalability; reserve exactly-once–like behavior for critical aggregations by combining idempotent keys and transactional writes where needed.

Run lightweight checks on upload to block obvious violations and run heavier models on delivery paths for context-aware decisions; cache outcomes and allow rapid policy updates.

Extract audio and frames, run modality-specific models in parallel, then fuse results with policy thresholds; route borderline cases to human review and store explanations for audits.

Combine device fingerprints, graph anomalies, and behavior models; throttle suspicious accounts, require step-up verification, and quarantine suspicious events pending review.

Blend follow-graph proximity with embedding similarity and recent interaction signals; add caps for diversity, language, and safety, and run A/B tests to validate uplift.

Maintain a curated taxonomy with community additions, tag videos via NLP and vision models, and expose discovery through search, trending, and related-topic carousels.

Maintain per-user frequency caps and quiet hours, dedupe by collapse keys, and prioritize high-signal events; back off when recent notifications were ignored.

Use short-lived signed links, device and IP reputation checks, rate limits, and anti-farm heuristics; reward genuine conversions and audit suspicious clusters before granting benefits.