Kube Builders

Migratowl is an AI-powered dependency migration analyzer that upgrades dependencies in isolated sandboxes, runs tests, reads changelogs, explains breakages, and suggests fixes. More: https://ku.bz/5Ddt6Pjj1

Vitalii Horbachov, Staff Software Engineer at Agoda, explains the massive scale challenges his team faced when running iOS CI/CD infrastructure. With 20,000 iOS tests requiring 200 Mac Minis and pipelines taking hours to complete, Agoda turned to Kubernetes as a job scheduler to manage workloads at scale. However, this approach introduced significant problems that ultimately led them to develop a custom solution for integrating native macOS workloads with Kubernetes - setting the stage for their innovative macOS VZ Kubelet project. Watch the full episode: https://ku.bz/q_JS76SvM

dt is a set of utilities available as a standalone or Helm Plugin for making offline work with Helm charts easier. You can create reproducible and relocatable packages for Helm charts that can be easily moved across registries without hassles. More: https://ku.bz/bSVHSvXqj

Cardamon audits Prometheus metrics against query logs, rules, and Grafana dashboards, then finds unused series and generates drop relabeling rules to reduce TSDB storage and cardinality waste. More: https://ku.bz/Hg6vtcg09

"He'd been looking forward to in-place pod resizing for years." Pedro Ignácio discusses Karpenter's limitations and the need for multi-cloud support. While Karpenter works well for AWS, Pedro emphasizes that global enterprises need cloud-agnostic tools that work across Azure, Google, and other platforms to standardize workflows and avoid vendor lock-in. Watch the full interview: https://ku.bz/515x11H-T This interview is a reaction to Kensei Nakada's episode https://ku.bz/bRd0243xQ

Three tools Rodrigo Bersa watches closely at AWS: 1. Karpenter — node autoscaler that consolidates capacity fast and spins up nodes without waiting. 2. KRO (Kubernetes Resource Orchestrator) — define resource graph definitions that provision Kubernetes and infrastructure resources together. Pair it with ACK for AWS-managed infra in the same lifecycle. 3. Cedar — an upcoming RBAC extension for Kubernetes 1.36/1.37 that adds deny and conditional permissions. Today, RBAC is additive-only. Cedar changes that. Each solves a different layer of the platform problem. Watch the full interview: https://ku.bz/dB7PDNt0v

This tutorial explains how to install Node Problem Detector in a home Kubernetes cluster using Helm and Argo CD. It shows how node-level issues like disk I/O errors, filesystem corruption, and kernel panics become Kubernetes node conditions. More: https://ku.bz/TM-1_pM3L

Jorrick Stempher explains the Node Ranking Index (NRI), a custom metric his team developed to solve a fundamental challenge in Kubernetes resource management. Discovered halfway through their 10-week project, the NRI combines CPU, RAM, and request data into a single comparable number, enabling efficient scaling decisions that would otherwise be impossible when dealing with different measurement types. Watch the full episode: https://ku.bz/clbDWqPYp

This week on Learn Kubernetes Weekly 192: 🔧 Our Kubernetes Operator Didn't Scale, So We Rebuilt It 🔀 ClickHouse Shard Rebalancing on Kubernetes: From Talk to Operator 💥 Invisible OOMkill: Java Pods Crashing in Kubernetes 🗂️ Using Kubernetes ConfigMaps as a Real-Time State Store 🚨 From Container Escape to Cloud Takeover: A Real-World Cloud Security Assessment Read it now: https://kube.today/issues/192 ⭐️ This newsletter is brought to you by Buoyant, the creators of Linkerd https://ku.bz/BB-RtVFWs

SloK lets teams define SLOs as Kubernetes custom resources, generate Prometheus rules, track error budgets, backtest SLO YAML, and view reliability trends in a dashboard. More: https://ku.bz/XwCd1mhXc

Radar provides Kubernetes cluster visibility through topology graphs, event timelines, and service traffic visualization running as a single binary that connects directly to the Kubernetes API without cluster-side installation. More: https://ku.bz/rq6l_SmGY

This tutorial shows how to run Hermes Agent on MicroK8s as a stateful gateway, move scheduled briefings into Kubernetes CronJobs, and use PVCs, ConfigMaps, retry jobs, and Helm for a small agent system. More: https://ku.bz/7ymZBxBMr

Molly Sheets, Director of Engineering, Kubernetes at Zynga, explains the strategic decision points for when teams should create separate clusters instead of continuing with multi-tenant approaches. She outlines two primary scenarios: geographic and latency requirements for global gaming deployments, and critical workload isolation during major incidents. Using Battle Royale games as an example, Molly demonstrates how different game components have varying latency needs - competitive gameplay requires ultra-low latency through regional clusters, while transactional features like purchasing can tolerate higher latency. She also discusses how separate clusters enable prioritized incident response, allowing teams to handle critical workloads first when cloud providers experience outages or availability zone failures. Watch the full episode: https://ku.bz/Rmpl8948_

KubePlumber tests Kubernetes networking from inside the cluster by checking: - internal DNS, - pod-to-pod traffic, - external DNS, - and bandwidth between nodes. More: https://ku.bz/nTkFwCt_Y

Ben, Senior Staff Platform Engineer @ Faire, discusses the specific technical limitations they encountered when scaling Jenkins with Kubernetes. He explains how their single Jenkins controller hit a critical breaking point in October 2021, forcing the rapid deployment of additional controllers that eventually grew to over 20 instances. The Jenkins Kubernetes plugin consistently failed around 1500-2000 pods, experiencing progressive slowdown that couldn't be resolved by adding more CPU or memory. Watch the full episode: https://ku.bz/klBmzMY5-

This case study shows how CoreDNS became an EKS bottleneck under heavy DNS traffic and uses kube-burner test results to explain why NodeLocal DNSCache helped. More: https://ku.bz/cs-fvstLZ

"I assumed because it's working in one place, it should work in the other place." That assumption broke Reza Ramezanpour's cluster. He'd configured IPIP encapsulation on AWS and expected the same setup to work on Azure — but Azure doesn't support IPIP. The underlying networking rules are different, and the CNI doesn't paper over that. His takeaway: Kubernetes networking still follows the OSI model from layer 1 to layer 7. Cloud abstractions don't change the fundamentals — you still need to know what your provider permits before choosing an encapsulation method. Watch the full interview: https://ku.bz/SjDx5CqXJ This interview is a reaction to Amos Wenger's episode https://ku.bz/6Ll_7slr9

KubeShark gives Claude and Codex Kubernetes guardrails for safer manifests, covering deprecated APIs, RBAC, security contexts, probes, resource limits, Helm, Kustomize, and GitOps patterns. More: https://ku.bz/sPF_MC1LY

This week on Learn Kubernetes Weekly 191: 🔥 What Does 4.4% GPU Utilization Actually Mean? 🛠️ GKE IP Exhaustion Fixed: The Class E Migration Guide 🧹 Evicting MCP Tool Calls from Your Kubernetes Cluster 🔄 The Feedback Loops Behind Kubernetes 🧠 You Don't Have a GIL Problem — You Have a CPU Problem Read it now: https://kube.today/issues/191 ⭐️ This newsletter is brought to you by LearnKube — master Kubernetes with hands-on training designed for engineers who want to learn the smart way https://ku.bz/hypSbyc-V

This article explains why low GPU utilization during LLM inference can be normal by breaking down prefill, decode, memory bandwidth limits, batching, and GKE B200 benchmark numbers. More: https://ku.bz/zyv99_RgD