Netflix improved Apache Druid query performance by implementing an interval-aware caching strategy that decomposes rolling window queries into fixed time-aligned segments. Instead of caching full query outputs, the system stores intermediate aggregates per interval, reusing cached historical segments and only recomputing the most recent interval. Operating as an external proxy with granularity-aligned buckets and exponential TTL policies, the approach achieves an 84% cache hit rate, reduces Druid query load by 33%, improves P90 latency by 66%, and delivers up to a 14x reduction in result bytes at Netflix's scale of over 10 trillion rows. Future plans include support for templated SQL queries and tighter integration directly into Apache Druid.
Sort: