Technical writing

Moving Open WebUI from SQLite to PostgreSQL can be done with no data loss and only a few minutes of downtime. This post walks through the process on a Kubernetes deployment, the problems you’re…

Ray's dashboard shows cluster health, but not the latency of the model being served. The post brings time to first token, token rate, KV cache utilization, and Ray Serve replica counts into one Grafana dashboard. It also explains why panels stay empty when Prometheus labels do not match.

Kubernetes gives a pod the whole GPU by default. The post partitions the card into isolated MIG slices, runs the device plugin in mixed mode, and schedules vLLM workloads on those slices. Capacity is calculated per slice because MIG slice memory is not pooled.

The post places Apache APISIX behind Nginx so teams can reach only the models they are allowed to use. `key-auth` validates the token and `consumer-restriction` checks access to the route. It also covers buffering that breaks streaming, timeouts across the request path, and plugins missing from Helm values.

MIG cannot be enabled on an H200 NVL node, GPU reset returns "In use by another client", and `nvidia-smi` shows no process. The card is held by the kernel framebuffer created during boot. The post explains the use of `nomodeset` on headless GPU nodes and the diagnostic steps that led there.

Label Studio stores file paths in each task's JSON, so changing storage breaks existing tasks. The post uses two scripts against the Label Studio API to rewrite paths in bulk and move tasks to S3. It also covers matching S3 keys to relative paths, presign settings for internal S3, proxy variables, and ReadWriteOnce volume constraints.

Kubeflow pods report Running while the UI returns 504. RKE2 runs Nginx Ingress in `kube-system`, but the existing NetworkPolicies do not allow traffic from there into `istio-system`. The post identifies the missing rule and shows the NetworkPolicy that fixes it.

Airflow can run on a separate server while notebook jobs run in Kubernetes through KubernetesPodOperator. The post sets up a namespace-scoped ServiceAccount, a kubeconfig generated from it, Python base images, and a DAG that removes the pod after the job. It also notes that long-lived ServiceAccount tokens need regular rotation.

The post adds an H200 NVL server to a bare-metal RKE2 cluster as a GPU node. It covers disabling nouveau, selecting the NVIDIA data-center driver, using RKE2's containerd template, and configuring MIG profiles for multi-tenant scheduling. It also documents RKE2-specific details missing from generic Kubernetes guides.

The post connects KEDA to an Azure DevOps queue so self-hosted agent pods start when work arrives and scale to zero when idle. Each runtime pool scales independently in the Helm chart, and capability tags route jobs to the right pool. It also explains the behavior difference between ScaledObject and ScaledJob.

A step-by-step guide to wiring Prometheus, Grafana, PodMonitor, and ServiceMonitor into your KubeRay cluster — so you finally see what’s…

What happens when your head node crashes mid-job — and how to make sure your workers keep running. A practical guide to GCS fault tolerance…

In the first part of this series, we built a production-grade Ray Cluster using KubeRay. The cluster was healthy, scalable, and ready — but…

Helm-based installation, custom GPU images, NFS model storage, and Prometheus/Grafana observability on a real production cluster.