Feature Store / Experimentation
Separate online and offline feature paths, point-in-time correctness, entity keys, and experimentation support.
On this page
Online vs offline
Show how DE supports online features, offline training, and experiments.
DE owns the freshness and correctness of feature data by separating online and offline paths while preserving point-in-time joins and experiment exposure lineage.
Quick jumps
Click to jumpMonthly users
200M-250M
Docs use both 200M+ and 250M assumptions
Daily active users
80M
Used for heartbeat and Kafka math
Peak concurrency
15M
Useful for QoE and live ops workloads
Separate low-latency features from training truth
This avoids the page feeling like an ML Engineer mode while still covering the DE responsibility well.
Offline Feature Store
S3 + Iceberg feature tables used for reproducible training and backtesting.
Stores historical training features with point-in-time aligned snapshots
Built from Silver and Gold history using Spark or dbt feature pipelines
Feeds training datasets, feature evaluation, and experimentation analysis
Online Feature Store
Redis or DynamoDB for low-latency serving features.
Stores recently watched, session context, device-time context, and hot popularity counters
Updated from streaming jobs for low-latency recommender calls
Needs explicit freshness SLAs and entity-key consistency with offline features
Feature flow
Point-in-time correctness
Correct point-in-time logic means feature_time must be less than or equal to label_event_time. Do not join a Monday training row to a Friday-computed user feature.
Show how experimentation data fits the pipeline
Assignments, exposure events, and outcome joins should feel like first-class consumers of the platform.
experiment.assigned event
recommendation.served / title.impression exposure
playback / click / retention outcome events
Silver user timeline
Gold experiment analysis mart
Describe how DE supports features and experiments
Keep the focus on data correctness and freshness.
Say This In Interview
DE owns the freshness and correctness of feature data by separating online and offline paths while preserving point-in-time joins and experiment exposure lineage.
30-second answer
FastDE owns the freshness and correctness of feature data by separating online and offline paths while preserving point-in-time joins and experiment exposure lineage.
2-minute answer
BalancedSeparate online and offline feature paths, point-in-time correctness, entity keys, and experimentation support. Focus on PRODUCTION, 8 min, Show how DE supports online features, offline training, and experiments..
Deep answer
DeepDE owns the freshness and correctness of feature data by separating online and offline paths while preserving point-in-time joins and experiment exposure lineage. Then go layer by layer through PRODUCTION, 8 min, Show how DE supports online features, offline training, and experiments., including trade-offs, failure handling, and why this shape is right for the consumer SLAs.
Last reviewed June 2026 · By Prasoon Parashar
Numbers are interview assumptions, not real Netflix internal figures.