Problem Statement

Today, OpenShift’s kubelet collects pod and container metrics (CPU, memory, network I/O, etc.) via cAdvisor.
However:

• Both kubelet (cAdvisor) and CRI-O gather similar runtime stats, creating duplication.

• High-cardinality metrics, especially container_network_* for host-networked pods, significantly inflate Prometheus storage and scrape costs.

• Maintaining and evolving metrics across two code paths (cAdvisor + CRI-O) is cumbersome and error-prone.

• Kubernetes upstream is deprecating cAdvisor in favor of KEP-2371 (CRI Stats), where the runtime becomes the single source of truth for metrics. https://github.com/kubernetes/enhancements/tree/master/keps/sig-node/2371-cri-pod-container-stats 

To remain upstream-aligned and reduce duplication, OpenShift needs to begin adopting the CRI Stats architecture.

Goal

Introduce a Technology Preview (TP) in OpenShift 4.21 that allows kubelet to consume pod and container stats directly from CRI-O via the CRI Stats API (PodStats, ContainerStats, NetworkStats).

The feature will:

• Shift metric collection responsibility to CRI-O.

• Simplify metric aggregation and reduce duplication.

• Provide an early opt-in path for testing performance, accuracy, and backward compatibility.

• Lay groundwork for high-cardinality reduction in OpenShift 4.22 when this becomes the default.