Kubernetes metrics collection

Design metrics collection for a highly elastic environment: a Kubernetes platform where 30,000 pods churn constantly (median pod lifetime 20 minutes), spread across 12 clusters in 4 regions. You need every pod's metrics scraped within 30s of it appearing, no metric gaps when pods die, and a single global query view across all regions. Decide push vs pull, how targets are discovered, and how you make collection highly available without double-counting.