Stop Answering the Same Question Twice: Interval-Aware Caching for Druid at Netflix Scale

This post describes an interval-aware caching layer Netflix built on top of Apache Druid to eliminate redundant query load at trillion-row scale.

• •Netflix's Druid cluster holds over 10 trillion rows and ingests up to 15 million events per second, making query efficiency critical
• •A single popular dashboard with 26 charts generates 64 queries per load; with 30 viewers refreshing every 10 seconds, that produces 192 queries/second from one dashboard alone
• •The cache uses a map-of-maps keyed by query hash (excluding time interval) and time-bucketed timestamps, enabling efficient range scans so overlapping rolling windows reuse cached buckets
• •Exponential TTLs are assigned based on data age: minimum 5 seconds for data under 2 minutes old, doubling per additional minute, up to 1 hour max, to account for late-arriving events