When someone first asks to run an LLM on Kubernetes, the instinct is reasonable: it's a containerised workload, it exposes an HTTP API, it needs to scale. Deploy it like anything else — a `Deployment`, a `Service`, maybe an `HPA` on CPU. That instinct gets you surprisingly far. Until it doesn't
What separates a good internal developer platform from an expensive Kubernetes wrapper—and how to build the former.
A migration story from Airflow on VMs to cloud-native data pipelines using Argo Workflows and what we'd do differently.
Implementing canary deployments and automated rollbacks using Argo Rollouts with Prometheus analysis templates.
Running large language model inference workloads on Kubernetes at scale—the infrastructure problems that emerge past the prototype stage.
Lessons from building and running three custom operators in production, including the parts the tutorials skip.
How we standardized cluster configuration across a multi-cloud estate using Flux v2 and a monorepo approach.