Netflix moved live-ad enrichment off the serving path with a Flink rewrite
Published August 17th, the redesign replaces registry writes with Kafka and Flink joins; Netflix says a load test handled hundreds of thousands of events per second.
By Ryan Merket · Published
Primary source: Netflix TechBlog
Why it matters
Live inventory turns ad telemetry into a serving dependency: late or lost events can affect frequency caps, measurement and billing. Netflix's migration removes a synchronous database write from that path, while its own caveats show the remaining tradeoff: in-stream deduplication is best effort, with late events recovered separately.
Yogesh Nagarur authored Netflix's account of an infrastructure change built for a specific pressure point: live ads can make a database dependency on the ad-serving path a liability just when millions of viewers arrive at once. Netflix replaced per-impression writes to a globally replicated metadata registry with a stateful Apache Flink job that joins ad-serving records with events from viewers' devices.…