Container-Optimized Compute Platform for GKE Autopilot

Container Optimized Compute for GKE Autopilot: Faster Scaling

Key Concepts:

• GKE Autopilot: Fully managed Kubernetes service.
• Container Optimized Compute: Redesigned compute stack for faster scaling.
• Horizontal Pod Autoscaler (HPA): Automatically scales the number of pods in a deployment.
• In-place Pod Resize: Resizing pods without disruption (Kubernetes 1.33).
• High Performance HPA Profile: Low-latency autoscaling with faster HPA calculation.
• General Purpose Compute Class: Workload configuration for optimal performance with Container Optimized Compute.

1. The Challenge of Scaling in GKE Autopilot:

• Historically, scaling in Kubernetes has been a challenge.
• In GKE Autopilot, scaling requires creating new nodes before applications can scale onto them.
• This node provisioning delay can be problematic for applications needing rapid scaling.
• Users previously used "balloon pods" (dummy pods) to reserve nodes, which was costly and difficult to maintain.

2. Container Optimized Compute: A Solution for Faster Scaling:

• Mission: Provide near real-time, vertically and horizontally scalable compute.
• Goal: Deliver capacity when needed at the best price and performance.
• Redesigned compute stack in GKE Autopilot.
• Provides flexible compute on demand.
• Results in up to 7x faster pod scheduling runtime.

3. Speeding Up Horizontal Pod Autoscaler (HPA):

• Container Optimized Compute speeds up the HPA.
• Leverages in-place pod resize (Kubernetes 1.33) for non-disruptive scaling.
• All features are available out-of-the-box in GKE Autopilot.

4. High Performance HPA Profile:

• Introduced for low-latency autoscaling reaction time.
• Provides consistent horizontal scaling reaction time.
• Offers up to 3x faster HPA calculation.
• Uses higher resolution metrics for improved scheduling decisions.
• Supports scaling up to 1000 HPA objects with predictive latency.

5. Demo of Container Optimized Compute:

• Demonstrates rapid scaling by changing the replica count from 1 to 10.
• Shows how quickly new pods are scheduled.

6. How to Use Container Optimized Compute:

• Create a new GKE Autopilot cluster with the "HPA profile performance" enabled.
• Ensure workloads use the "general purpose compute class."
• Best suited for services that need to scale gradually.
• Most improvement seen in workloads with smaller resource requests.

7. Limitations:

• Not suitable for one pod per node deployments (anti-affinity situations).
• Not suitable for batch workloads.

8. Conclusion:

• Container Optimized Compute platform improves application autoscaling in GKE Autopilot.
• Encourages users to try it out and provide feedback.