Question

A team runs a stateless microservice on GKE. Engineers set CPU and memory requests conservatively high 'to be safe,' and the cluster autoscaler keeps adding nodes even though average pod utilization is around 20%. Monthly compute cost has doubled. The team wants data-driven request values without risking disruption to running production pods while they validate the recommendations. What should you do first?

Accepted Answer

Deploy Vertical Pod Autoscaler in 'Off' (recommendation) mode and use its recommended CPU/memory values to adjust the Deployment's requests during the next planned rollout.

Answer

Deploy Vertical Pod Autoscaler in 'Auto' mode so it immediately resizes and evicts running pods to the optimal requests.

Answer

Enable Horizontal Pod Autoscaler on CPU to reduce the number of pods and lower cost.

Answer

Switch the node pool to larger machine types so fewer nodes are needed to schedule the over-provisioned pods.