AI/ML Pod Cold Start Timeline

AI/ML Pod Cold Start Timeline A sequence diagram generated by Archify. Pod scheduled CRI image preparation Image layers Create container Start process readiness probe result Pod placement: measured waiting Image fetch / unpack: cache and network Container creation: runtime configuration Model loading and readiness: per workload K8s Scheduler · kube-scheduler · Sequence participant K8s Scheduler kube-scheduler Kubelet · worker node · Sequence participant Kubelet worker node Container Registry · e.g. Amazon ECR · Sequence participant Container Registry e.g. Amazon ECR Container Runtime · containerd · Sequence participant Container Runtime containerd AI/ML Application · model serving process · Sequence participant AI/ML Application model serving process Legend request return default message

Startup measurement

  • • These are logical stages; the runtime performs the actual image pull.
  • • SOCI indexing/prefetch does not guarantee80–95% improvement.
  • • Immediate reads of all weights can limit lazy-loading benefits.