A Tekton dind sidecar rewrites its cgroup path so step and build share one memory ceiling, enforced down to a real OOM kill. The same controller leaves the build container with no memory metrics.
A Go daemon maps pod priority to cgroup v2 io.weight, tested against real I/O contention on a kind cluster. io.latency wasn't available, so this measures what proportional I/O weighting delivers.
A design thesis for closing the pod-startup gap: combine the Kubernetes API's phase boundaries with eBPF's syscall-level evidence, correlated by pod identity and persisted before the pod disappears.
A survey of what eBPF observability tools, continuous profilers, and generic APM platforms actually do today, and why none of them attribute pod startup latency to a phase or a syscall.
What a slow Kubernetes pod startup actually costs in autoscaling, CI/CD, and on-call time, and why kubectl events and kubelet's own latency metric can't answer which pod, which second, or why.