How ClickHouse Cloud's two-window recommender and target-tracking CPU algorithm cut scale-down latency from 30 hours to 3 hours while   eliminating oscillations and reducing infrastructure costs.

ClickHouse is a leading platform for open-source columnar database technology and real-time analytics. With ClickHouse, developers can delve into documentation, case studies, and performance benchmarks to learn how to build highly scalable and performant analytical systems. It provides tips about ClickHouse's storage engine, query execution, and distributed architecture. Developers can gain insights into their data in real-time, enabling them to make informed decisions and extract  business insights.

ClickHouse

ClickHouse Cloud's auto-scaling system previously used a single 30-hour lookback window, causing clusters to remain over-provisioned for up to 30 hours after traffic dropped. The solution introduces a two-window recommender: a 3-hour small window for fast scale-downs and a 30-hour large window for stable single-step scale-ups. A merging algorithm decides which window's recommendation to follow based on scaling direction, with a hunting check to prevent oscillations when windows disagree. The old fixed-factor CPU algorithm (doubling/halving allocations) was replaced with target-tracking, which calculates exact resources needed to hit a target utilization using the geometric mean of watermarks to ensure reversible, drift-free scaling. A 10-minute rolling median smooths transient spikes. The result: scale-down latency dropped from 30 hours to 3 hours, oscillations eliminated, and infrastructure costs reduced for variable workloads.

Smarter Auto-Scaling for ClickHouse: The Two-Window Approach