Scale to zero: what carbon-aware autoscaling really buys you

Most clusters are sized by anxiety. Teams provision for the worst hour of the worst day and then leave that capacity running around the clock, paying for it in both euros and emissions. Carbon-aware autoscaling is the antidote — but it takes more than turning on the HPA.

Right-size before you autoscale

Autoscaling a bloated workload just scales the waste. We start by profiling what a single replica actually needs, trimming resource requests, and killing background jobs nobody remembers scheduling. Only then does elasticity pay off.

Scale to zero, not to "one, just in case"

The biggest wins come from workloads that can genuinely drop to zero replicas between bursts — batch pipelines, internal tools, staging environments. Keeping a warm instance "just in case" quietly doubles a service's annual footprint for a few seconds of saved cold-start.

Every idle replica is a decision to spend carbon on latency you may never need.

Follow the grid, not just the clock

Carbon intensity swings through the day as renewables ebb and flow. Deferring flexible work to greener hours — or shifting it to greener regions — can cut a workload's emissions without changing a line of business logic. The trick is knowing which work is truly flexible.

• Batch and reporting jobs: highly shiftable
• CI pipelines: often shiftable by minutes
• User-facing requests: not shiftable — make them efficient instead

Done well, carbon-aware autoscaling is invisible to users and obvious on the bill.