Start Here
Clarify scope, show the end-to-end journey, and position the data-engineering boundary before any deep dive.
On this page
Track intent
Design the data platform, not the playback backend
Understand the full Netflix data-platform journey from events to Kafka, lakehouse, analytics, ML features, and backfill.
Data platform end-to-end flow — click any stage
Kafka / MSK
2 / 6Kafka buffers high-volume events, decouples producers and consumers, supports replay, and feeds streaming, raw lake, and feature pipelines in parallel.
Event journey
How user actions become analytics facts via Kafka, streaming jobs, and lakehouse tables.
Correctness story
Why heartbeat beats play/pause, how sessionization works, and when batch owns final truth.
Scale math
How to derive Kafka partitions, storage size, and streaming parallelism from DAU and event rate.
Failure recovery
Late data, DLQ, quarantine, replay paths, and how to protect official metrics from bad code.
What we are designing
High-volume event ingestion from mobile, web, TV, gaming consoles, CDN, and backend services
Real-time processing for dashboards, trending content, anomaly detection, and online recommendation features
Batch processing for daily metrics, retention, churn, content performance, and ML training datasets
Bronze, Silver, and Gold lakehouse layers plus warehouse, BI, feature store, and serving consumers
Data quality, governance, lineage, security, replay, backfill, correction, and late-arriving data handling
What we are not designing
Playback API orchestration, CDN request routing, DRM flow, and video content delivery path
Video upload product, transcoding internals, or microservice REST endpoint design
User login backend, billing service internals, or search ranking model internals
Recommendation model architecture in depth beyond the data and feature pipeline
Say this in interview
I would design Netflix's data platform, not the playback backend. The system ingests playback, heartbeat, search, impression, QoE, billing, and recommendation events into Kafka. From there, real-time jobs compute freshness-sensitive metrics, while raw events land in S3/Iceberg for batch processing. Cleaned Silver tables and aggregated Gold tables power dashboards, experimentation, recommendations, data quality, replay, and backfills.
The 14-chapter interview answer path
Each chapter below is a tab. Work through them in order for a complete system design answer.
Desktop journey canvas
Click any stage to jump into the deeper chapter.
Scope
Frame the platform boundary first.
Events
Map the producers and event families.
Kafka
Choose topics, keys, and replay posture.
Streaming
Turn events into trusted live facts.
Lakehouse
Keep replayable truth by layer.
Gold Metrics
Publish decision-ready outputs.
BI / ML
Serve each consumer correctly.
Replay
Protect correctness after failure.
Last reviewed June 2026 · By Prasoon Parashar
Numbers are interview assumptions, not real Netflix internal figures.