The problem

Reactive autoscaling adds capacity after latency has already risen, so teams keep big safety margins running all day. Proactive scaling can size ahead of demand, and it doesn't need a heavy model.

The fix you can use today

• Forecast CPU from request volume with a lightweight model.
• Try exponential smoothing before anything neural.
• Scale ahead of predicted peaks and shrink the standing safety margin.

GreenCode researchers found that a simple maths model predicted load better than neural networks.

Tools that help

• KEDA: event-driven autoscaling for Kubernetes, including scaling a workload to zero when there is nothing to do
• Kepler: exports energy estimates per pod and container on Kubernetes, using eBPF and hardware counters

How GreenCode helps

GreenCode helps tackle this by mapping your running systems and tracing their energy and CO2e to the services that cause it. It shows services running hot, and services barely running at all. More on this.

Tags

#kubernetes#autoscaling#cloud

Sources

• Comparing neural and statistical time-series models for proactive auto-scaling in Kubernetes (SOSE 2025)

All GreenCode Tech Tips