What happened
Google Cloud has announced CPU startup boost for GKE, now in preview. It is built into the Vertical Pod Autoscaler (VPA). A container gets extra CPU while it starts, and the allocation drops back to its normal level once the pod is ready. The container is not restarted.
Under the hood it relies on Kubernetes In-place Pod Resize. Per Google, that feature reached Beta in v1.33 and is due to reach GA in v1.35. The preview needs GKE 1.36.0-gke.4447000 or later, on Standard or Autopilot.
Why it matters
Many apps need far more CPU in their first seconds than they do afterwards. Java apps using Spring Boot, Node.js servers and Python services that import heavy libraries are the examples Google gives. If you set CPU requests for steady state, startup gets throttled. That means slow cold starts and readiness probes that time out, which hurts most when you scale out under load or roll out a release.
The usual workaround is to raise CPU requests for everything. That works, but you then pay for idle headroom on every replica, all the time. For a small company with no platform team, this is often a quiet source of cloud spend that nobody has had time to look at.
Startup boost lets you size requests for real runtime needs. Google says it can cut initialisation time by up to 2x. Treat that as a ceiling, not a promise. Your own results will depend on your workloads.
It is a preview feature, so think carefully before putting it on critical production services.
What to do this week
- Find services where requests were raised “just to make it start”. Check deployment history and ask the team who set them and why.
- Look at readiness probe failures and slow rollouts. These show which workloads are actually throttled at startup.
- Check your GKE version against the 1.36.0-gke.4447000 minimum, and whether VPA is enabled on Standard clusters. Autopilot has VPA on by default.
- Try it in a non-production cluster first. Google’s configuration is a startupBoost section in the VPA manifest, and updateMode “Off” keeps your steady-state requests as they are.
- Measure start time and cost per replica before and after, so you can see whether it pays off.
Need a hand?
If you are unsure whether your services are over-provisioned or throttled at startup, our team can look at it with you. We help small teams tune Kubernetes resource settings, autoscaling and rollouts. You can read more on our Kubernetes consulting page. The first call is free and is with an engineer.
Source: GKE CPU startup boost: Accelerate app starts without over-provisioning
Related service: Kubernetes Consulting