Kube Architect

This week on Learn Kubernetes Weekly 202:

🔥 We Replaced etcd with Google Cloud Spanner
😌 How We Made Deploying a New Service Boring
🐘 Running Zookeeper on GKE with Local SSD (Z4D)
🚀 Kubernetes v1.36: Mixed Version Proxy Graduates to Beta
🌍 Building a Multi-Region EKS Platform with Crossplane, FluxCD, and GitOps

Read it now: https://kube.today/issues/202

⭐️ This newsletter is brought to you by Buoyant — The Buoyant Enterprise for Linkerd service mesh runs in production at Xbox (22,000 pods), Imagine Learning (40% cross-zone cost cut), and IntelliGRC 4× MRR after FedRAMP https://ku.bz/BwZYjDryv

This tutorial shows how to chain KEDA cron scaling with Karpenter so non-prod EKS clusters drop to zero nodes at night and on weekends, cutting the compute bill by 65 to 85%.

More: https://ku.bz/J18t6k5N5

This case study traces a 503 at the ingress back to a broken CNI overlay between nodes, even though every pod and endpoint looked healthy, and fixes it with topology spread constraints.

More: https://ku.bz/M-PjSd6Tl

This article explains how to make Kubernetes spend traceable to each team, using a consistent label taxonomy, Prometheus recording rules on resource requests, policy enforcement and chargeback reporting.

More: https://ku.bz/w2TsnQRVb

Rune is a fast native macOS app for people who debug clusters every day, with unified pod logs, YAML editing with server dry-run, port-forwarding and a command palette.

More: https://ku.bz/zRfTn2ZwJ

This case study shares a production Strimzi Kafka setup that survives broker and zone failures, using KRaft instead of ZooKeeper, per-zone node pools, Cruise Control rebalancing and careful JVM tuning.

More: https://ku.bz/Jmj90fchK

Frédéric Gaudet, Senior SRE @ BlaBlaCar, explains his testing approach for validating Istio's pod warm-up feature before production deployment.

He reveals a critical insight about the interaction between pod warm-up and Horizontal Pod Autoscaler (HPA). When a new pod warms up and receives reduced traffic, that traffic gets redistributed to existing pods, potentially triggering CPU spikes. However, he discovered that keeping warm-up duration under the HPA trigger time (2 minutes warm-up vs 5 minutes HPA trigger) prevents instability.

Watch the full episode: https://ku.bz/grxcypt9j

We just published Kubernetes Architecture in Financial Services, a free technical book about platform design and operations.

Drawing on public engineering talks and case studies from seven banks, it examines:

- Tenant boundaries beyond namespaces.
- Shared delivery controls and application ownership.
- Policy maintenance, exceptions, adoption, and reporting.
- Cross-cluster identity, reliability, external dependencies, and cluster replacement.

Each chapter starts with a real platform problem, compares different boundaries, and examines the operational cost of each option.

Thank you to Buoyant, Sysdig, and Nirmata for supporting the book and its research.

Download it for free:
https://learnkube.com/kubernetes-architecture-financial-services

This article uses a food delivery startup story to explain how a .NET API on one EC2 box moves onto EKS, with Docker, Kubernetes objects and Terraform each given a plain-language role.

More: https://ku.bz/pZdyNK2Fj

This case study shows what happened when a team swapped etcd for Google Cloud Spanner, using commit timestamps instead of a single revision counter, so writes no longer queue behind each other.

More: https://ku.bz/gPW89HJPw

This article explains why the default Kubernetes scheduler struggles with AI training jobs, and how Volcano adds gang scheduling, queues and GPU topology awareness on top of it.

More: https://ku.bz/JVDvcTt12

Kogaro continuously validates Kubernetes config with 60+ checks across reference, resource, security, image, and network domains, catching silent failures before they impact production.

More: https://ku.bz/SWl3-LNty

This tutorial shows how to run Bifrost, an open-source AI gateway, on Kubernetes with Helm, cluster mode and autoscaling so thousands of LLM requests stay fast under load.

More: https://ku.bz/Rss1qf4Sj

This case study shows how Zencity replaced the copy-pasted Helm chart in every service repo with a shared monochart, so that adding a new service stopped being a project in its own right.

More: https://ku.bz/N3Rf2lFBF

This article walks through a multi-region EKS platform where Crossplane provisions the AWS side, FluxCD reconciles every cluster and Git stays the single source of truth.

More: https://ku.bz/GlHf0gcwW

Kubernetes problems often hide in controller timing, implicit defaults, and capacity assumptions.

This week in Learn Kubernetes Weekly:

🔎 A production race condition left orphaned pods blocking new deployments.
📦 Source Hydrated Infrastructure Models commit rendered manifests for explicit GitOps audits.
🖥 Headlamp replaces the archived Kubernetes Dashboard and maps familiar workflows to a maintained UI.
⚙️ Kubernetes 1.36 adds pod-level CPU and memory management.
📊 The k8s-overcommit operator reclaims idle capacity according to priority classes.

Read issue 201: https://kube.today/issues/201

This issue is brought to you by LearnKube — understand how Kubernetes works, and what to do when it breaks. Live training with 60% hands-on labs: https://ku.bz/hypSbyc-V

Managed Kubernetes makes cluster operations easier, but service gaps still show up in real environments.

Referencing a point raised in the Grzegorz Głąb episode, Christopher Tineo describes one of the practical tradeoffs he has seen in Google Cloud: some services are not equally available across regulated environments, which means teams still need to look beyond managed offerings and evaluate open-source options.

The lesson is that managed services reduce toil, but they do not remove architectural constraints.



Watch the full interview: https://ku.bz/5f9sFC6CG

This interview is a reaction to Grzegorz Głąb's episode https://ku.bz/yg_fkP0LN.

This tutorial walks you through a full GitOps setup on minikube, where Argo CD deploys a Go service straight from Git and an HPA scales the pods under load.

More: https://ku.bz/4lY9D3HgR

Danyl Novhorodov, Software Engineer & Architect @ Eneco, explains that Kubernetes can be clean, resilient, and efficient, but only when managed by teams with deep platform understanding, clear infrastructure-application boundaries, and strong operational discipline around security and CI/CD.

Without these foundations, he compares it to "giving a Formula One car to someone who just got their driving license" - the tool isn't broken, it's simply misapplied.

Watch the full episode: https://ku.bz/BYhFw8RwW

This case study shows how a team kept delivering WebSocket events after growing to dozens of pods, using Redis to track which pod holds each user's connection and relay messages to it.

More: https://ku.bz/ftpTJ9Pqh