Nick Nikolakakis

// Karmada v1.19 can place a whole PyTorchJob, master and workers, into one member cluster. Four kind clusters with fake GPUs show the four conditions it needs, and what happens without each one: the job runs three times, the scheduler finds no room with 16 free GPUs, and a job lands where its workers can never start.

// Kubernetes is retiring the SPDY streaming transport behind kubectl exec, attach, cp and port-forward in favour of WebSockets. This is an audit you can run against your clusters this week to find the hop that will break, before the upgrade finds it for you.

// What actually changes in how you build, roll out and roll back a Kubernetes node when the node OS ships as an OCI image instead of a package set, using bootc and a small worked example called Corium.

// KEP-4671 landed gang scheduling in Kubernetes 1.37 beta, behind a feature gate that is off by default. A four-node Kind cluster with six fake DRA GPUs shows the partial-placement deadlock, the single field that fixes it, and the two limits nobody documents: the Job field is pruned silently without a second alpha gate, and workload-aware preemption cannot reclaim DRA devices.

// Part two of the Modelplane fleet inference series, covering the layer above the scheduler. ModelService is a routing policy rather than a load balancer, spec.stack Dynamo swaps the whole gang-scheduling machinery, and four documented gotchas will cost you an afternoon each.

// A hands-on tour of Modelplane, the Crossplane-based control plane for AI inference across a fleet of GPU clusters. Two regions on Kind with fake DRA GPUs, the platform/developer contract written in DRA's own vocabulary, fleet-level drain, and the release trap that costs you two APIs.

// A hands-on tour of agent-sandbox, the Kubernetes SIG Apps project for isolated, stateful, singleton workloads built for AI agent runtimes. Includes a Go agent with sandboxed tools on Kind, warm pools claiming in 53ms, Flux GitOps, and the gotchas the docs don't mention.

// A hands-on tour of llm-d, the CNCF Sandbox framework for distributed LLM inference on Kubernetes - inference-aware routing, prefill/decode disaggregation, and KV-cache offload. Includes a GPU-free demo on Kind using the vLLM simulator, wired with Flux GitOps.

// Choragos runs a team of AI coding agents with a real division of labour: an orchestrator that plans and delegates, workers with their own context, model and credentials, and a delegate/work-done protocol carrying work between them. A walkthrough of v0.11.2 with config recipes for real teams.

// How I built and shipped Pinax, an LLM-buildable dashboard framework for Obsidian, using Claude Code: the guardrails that made agentic development work, the automated review battles, and why machine-checkable success criteria beat vibe coding.