gpu operator

Analysis of Kubernetes 1.37 node lifecycle conditions for GPU maintenance: five condition names, the no-op Alpha gate, drain criteria, single-writer ownership, failure tests and a lost GPU-hours formula (2026).

A 2026 decision guide for Kubernetes GPU sharing: compare isolation, memory behavior and scheduler resources, then use a workload worksheet, canary scorecard and rollback plan.

A 2026 constraint-first comparison of NVIDIA MIG, CUDA MPS, GPU time-slicing, passthrough, and dedicated GPUs across isolation, latency, scheduling, and observability.

A 2026 runbook for air-gapped NVIDIA NIM upgrades, covering release BOMs, model caches, measured capacity, offline proof, registry imports, and rollback evidence.