Grow your own way: Introducing native support for custom metrics in GKE

When platform engineers, AI Infrastructure leads and developers think about autoscaling workloads running on Kubernetes, their goal is straightforward: get the capacity they need, when they need it, at the best price. However, while scaling on CPU and memory is simple enough, scaling on application signals like queue depth or active requests is not. Historically, it’s been achieved via a complex sequence of different steps involving monitoring, IAM and specific agent configuration, adding significant operational overhead.