Kube Builders

This article opens up Xatastor, the storage system Xata built for millions of Postgres databases, explaining why they chose ZFS zvols over NVMe-oF instead of Ceph or Longhorn, and how their Kubernetes operator drives it. More: https://ku.bz/LcBp9Pl1h

Molly Sheets, Director of Engineering, Kubernetes at Zynga, challenges conventional approaches to measuring Kubernetes reliability in multi-tenant environments. She explains why traditional metrics like control plane uptime and node availability aren't useful when running managed Kubernetes services like Amazon EKS, where the cloud provider handles much of the infrastructure layer. Instead, Molly advocates for focusing on SLOs for individual applications and add-ons. She provides specific examples from Zynga's infrastructure, emphasizing the importance of monitoring critical networking components like Cilium agents and external DNS, since failures in these areas cascade to affect all workloads in the cluster. Watch the full episode: https://ku.bz/Rmpl8948_

ModelExpress is an NVIDIA service that manages model weights across a cluster, caching them and moving them GPU-to-GPU over RDMA instead of reloading from storage, so new vLLM or Dynamo pods start serving much sooner. More: https://ku.bz/RyyNYRgJs

Not every organization can move to the cloud immediately — data residency, sovereignty, and compliance requirements are real constraints. Raglin Anthony walks through the options: self-managed clusters with Kubeadm, vendor-managed solutions, AWS Outposts, EKS Hybrid Nodes, and EKS Anywhere for fully air-gapped environments. Each trade controls for operational burden differently. Watch the full interview: https://ku.bz/2XqMJnLVx

This article describes the architecture behind a self-hosted cloud IDE, where each developer workspace is an isolated container on Kubernetes, requests are routed to small or large models by task, and clusters are spread across regions. More: https://ku.bz/429V7pdXG

Ben Poland, Senior Staff Platform Engineer @ Faire, breaks down why their massive monolithic CI pipeline became a major bottleneck for developer productivity. He explains how 4-5 thousand lines of Groovy code with complex pipeline libraries created a system so difficult to understand that only a few platform engineers could make changes, forcing teams to rely on centralized support instead of self-service. Watch the full episode: https://ku.bz/klBmzMY5-

This week on Learn Kubernetes Weekly 196: 🛠️ How I Rebuilt YouTube’s Load Balancing Algorithm in Go 💾 When ETCD Crashes, Check Your Disks First: A Pod CrashLoopBack Debugging Story 🧩 We Blamed CoreDNS for Weeks. The Real Culprit Was a Default We Never Questioned 🚚 Zero-Downtime EKS Migration: Production-Grade Kubernetes at Scale ⚡ Deterministic Routing: The Hidden Key to Low Latency Read it now: https://kube.today/issues/196 ⭐️ 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 presents a three-layer tenant isolation design where each tenant gets its own control plane, VM nodes and isolated network via KubeFlex, KubeVirt and OVN-Kubernetes, with latency measurements. More: https://ku.bz/YRcVzxByx

IncidentFox automates incident investigation with AI agents using 178+ tools for Kubernetes, AWS, and Grafana, featuring RAPTOR knowledge base for runbooks, alert correlation reducing noise by 85-95%, and Slack/GitHub/PagerDuty integrations. More: https://ku.bz/r_Q9FFtc0

IncidentFox automates incident investigation with AI agents using 178+ tools for Kubernetes, AWS, and Grafana, featuring RAPTOR knowledge base for runbooks, alert correlation reducing noise by 85-95%, and Slack/GitHub/PagerDuty integrations. More: https://ku.bz/r_Q9FFtc0

IncidentFox automates incident investigation with AI agents using 178+ tools for Kubernetes, AWS, and Grafana, featuring RAPTOR knowledge base for runbooks, alert correlation reducing noise by 85-95%, and Slack/GitHub/PagerDuty integrations. More: https://ku.bz/r_Q9FFtc0

How do you know when your control plane is struggling? Alex Kestner shares three signals: API server request latency is the most critical — if it's growing, pay attention. Next, watch the API priority and fairness metrics to see if requests are failing to complete. Finally, keep an eye on etcd database size — if it's filling up, you're heading for trouble. Watch the full interview: https://ku.bz/blMzG_XM7

New from LearnKube: Setting the right requests and limits in Kubernetes. CPU and memory requests and limits appear together in a Pod manifest, but Kubernetes and Linux use them at different stages. You will learn: • How the scheduler compares requests with node allocatable capacity • Why actual usage can exceed a request • How CPU quotas cause throttling even when average CPU looks low • Why an OOM kill is different from a node-pressure eviction • How requests and limits determine Kubernetes QoS classes The article is the first chapter of The Technical Guide to Kubernetes Rightsizing, co-authored by Gulcan and Daniele Polencic. Read the chapter: https://learnkube.com/setting-cpu-memory-limits-requests Download the free technical guide: https://learnkube.com/kubernetes-rightsizing This book was made possible by CloudBolt Software.

This case study shows how etcd CrashLoopBack pods on a Karmada/k3s demo cluster turned out to be a ZFS I/O latency problem, and how four ZFS tuning settings fixed it — including the exact etcd Prometheus metrics to watch. More: https://ku.bz/p1hBF3Jdq

This case study explains how an ECS workload was migrated to EKS across two AWS regions with zero production downtime. It covers KEDA autoscaling, HashiCorp Vault, IRSA, disaster recovery, and a coordinated production cutover. More: https://ku.bz/VDHmS189t

Karpenter can consolidate nodes and reduce cloud costs. But without Pod Disruption Budgets, those savings can expose workloads to avoidable downtime. Ahmad Asmar from Zencity explains how they use Kyverno to generate PDBs automatically instead of relying on every developer to remember the required configuration. You will learn: - Why Karpenter consolidation and Spot instances make PDB coverage critical - How Kyverno detects existing PDBs through workload labels - Why percentage-based budgets work better for scaling workloads - How ClusterRole aggregation extends Kyverno without modifying its Helm chart Watch: https://ku.bz/xrlPJg54D 🌟 This episode is sponsored by LearnKube. Download the free book, The Technical Guide to Kubernetes Rightsizing, to understand what Prometheus and Grafana cannot tell you about safely reducing requests and limits. https://learnkube.com/kubernetes-rightsizing With @Birthmarkb

MAAS (Metal-as-a-Service) is a production-grade tool that automates provisioning, network config, and OS installs (Ubuntu, CentOS, Windows, ESXi) across physical servers using PXE, IPAM, and full API-driven workflows. More: https://ku.bz/9z6rg3_cZ

When one pull request spans app code, config, and CI, the weakest point is usually not the app. Artem Lajko explains that ownership is split between developers and platform teams, so config changes lack true end-to-end accountability. That makes breakage harder to spot before merge and easier to discover only when CI or infrastructure behavior changes.

This article explains how Kubernetes pod DNS failures can come from the default ndots:5 setting and a Linux conntrack issue, and shows how ndots:2 plus single-request-reopen can reduce DNS noise and fix intermittent resolution errors. More: https://ku.bz/pZG7kVlFk

K8up is a Kubernetes Operator that helps you: - Backup all PVCs marked as ReadWriteMany or with a specific label. - Perform individual, on-demand backups. - Schedule backups to be executed regularly. And more. More: https://ku.bz/1ynMNZrK9