GPU cost optimization starts with useful work, latency and paid capacity. A measurement-led review of data stalls, inference sizing, sharing and cost attribution separates removable spend from capacity gains.
Spark sets executor CPU requests equal to limits, so a Kubernetes cluster reserves twice the CPU it uses and refuses to schedule pending pods. Kubernetes has no native overcommit. Here is the mutating-webhook operator I use to fix it.
Two MinIO platforms, same root cause: used as a NoSQL store. Field notes on LIST IOPS, XFS directory limits, scanner & heal SLAs, the erasure-coding storage-efficiency inversion on small objects, and why Apache Cassandra (or Ceph) is the right answer on-prem in 2026.