TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

Kubernetes CPU Allocation for Go Applications

1 post · May 2023

What this trail covers

Grab's real-time data platform team, Coban, has been optimizing the performance of Go applications running on Kubernetes. This post details the challenges encountered with CPU throttling and Vertical Pod Autoscaler (VPA) when using non-integer CPU allocations. The team discovered that the `AUTO-GOMAXPROCS` setting in Go rounds down non-integer CPU values, leading to performance degradation and consumer lag. By ensuring VPA proposes integer CPU values (especially for high-resource pipelines) and by setting a minimum integer CPU value for critical pipelines, they improved performance and resource utilization.

All Grab trails
  1. Performance bottlenecks of Go application on Kubernetes with non-integer (floating) CPU allocationThis post details the investigation and resolution of CPU throttling and consumer lag issues in Go applications on Kubernetes. The core problem was identified as the interaction between Kubernetes CPU allocation, the `AUTO-GOMAXPROCS` setting in Go (which rounds down non-integer values), and the Vertical Pod Autoscaler (VPA). The solution involved ensuring VPA proposes integer CPU values and setting a minimum integer CPU limit for resource-intensive pipelines to prevent performance degradation.