N · cluster
cluster
Scheduler, GPU partitioning, autoscaling, node lifecycle — the layer that decides where a workload lands.
Mechanism reference for Kubernetes at the helm of AI
One article, one mechanism, with the real APIs and the real failure modes. Forty charts across four surfaces.
E · serve
serve
Inference servers, KV-cache, batching, streaming, model routing — the layer that answers requests.
S · train
train
Distributed training, NCCL, checkpointing, fault tolerance — the layer that turns GPU-hours into weights.
W · operate
operate
Observability, security, GitOps, cost — the layer that keeps the other three honest.
latest charts all charts →
How Dynamic Resource Allocation changes the GPU request contract
What Kubeflow Trainer's TrainJob moves out of the job spec
What vLLM's prefix cache can and cannot reuse
Model weights belong in object storage, not container images
Checkpoint storage patterns for distributed training