iiinova — where your digital self keeps evolving

runbook. Check this out from pradip Kumar pandey and Pratap Kumar Nanda.

000

That can take days and the right people. Letting AI do the first pass and having humans verify it can save a lot. The example here builds a Strands agent on AgentCore Runtime that reads source code and manifests, scores migration readiness from 0 to 100, ranks blockers by severity, and drafts a

100

easily tell how hard the move will be. Before an app moves to Amazon EKS, someone has to read the code for hardcoded IPs, NFS mounts, and assumptions about the environment.

100

Once you separate the actor (scheduler vs kubelet vs API server) from the reason, a lot of the usual confusion goes away. Check it out!

000

One area like this is scheduling, preemption, eviction, QoS, PriorityClass, and PDB. The article below attempts to pull them apart and show where each one actually applies. Tanat Lokejaroenlarb framed it as: for any Pod being removed or placed, ask who is making the decision and why.

100

lckhd.eu/yhOJ1v #Kubernetes #scheduling #pdb Kubernetes has many concepts to understand. Many of these make sense alone but get tangled the moment you put them together.

100

Once you separate the actor (scheduler vs kubelet vs API server) from the reason, a lot of the usual confusion goes away. Check it out!

000

100

lckhd.eu/0YCY7t #Kubernetes #scheduling #pdb Kubernetes has many concepts to understand. Many of these make sense alone but get tangled the moment you put them together.

101

The article below shows an approach that moves those conventions into the repo. It has an 80-line CLAUDE.md, on-demand skills and subagents, and hooks that validate Terraform files. Filipe Motta includes full AWS and GCP setups in a companion repo. Check it out!

000

If you use a coding agent to build a Terraform stack on two different days, you will most likely get two different approaches. For example, your VPC may be called "main" one day and "vpc_primary" the next, and many other small details will change.

110

Subscriber can start from a past timestamp. Check this out from Nahid Karimaghalou, Jamie Dool and Tolga Orhon and give the new custom event buses a try.

000

across AWS accounts. Figuring out which buses are used for what can becomes difficult. The new custom event bus lets you share a single bus across your entire org, with each team owning a Subscriber for filters, retries, and a Dead Letter Queue (DLQ). It keeps events for up to 1 year, so a new

100

Nothing here is EKS-specific. The descheduler is a Kubernetes SIG project. Check this out from Ramya D, Himanshu Bansal, and Tushar Mishra.

000

In the example, the team cordoned one of three AZs running ~1,000 pods on EKS. It went to a ~500/500/0 AZ split in 90 seconds, and the skew held at 227 for nearly 17 hours after the AZ returned. Two descheduler passes fixed it with 183 evictions, against a minimum of about 180.

100

constraint in the manifest. This is a realistic scenario and the solution below is really interesting.

100

and ctrl+a lists the namespaces. Harish Thangadurai's article is a good starting point to use K9s.

000

can work better. This is where K9s can help. Using kubectl tends to become repetitive including getting pods, copying the pod name, describe the pod, pulling the logs and rinse and repeat. K9s helps to put all that in a live terminal view, where :po lists pods, l shows logs, s puts you in a shell,

100

lckhd.eu/UYBeFf #Kubernetes #K9s #kubectl I have spent years looking at Kubernetes clusters with kubectl, and I know the common resource types and flags by heart. This works really good for me, but it can take a long time to get there, and for many a higher level view of what is running

100

Netflix built a tool called the Stratum Resource Tuner that right-sizes these configs automatically: CPU and network are tuned to p99 usage with no headroom, and memory and disk to p99.9 plus a buffer. Check out this writeup from Violetta Pidvolotska and Naveen Mareddy.

000

for extreme cases. Most never change until they fail one day, and the fix is usually more memory.

100

lckhd.eu/p9Xyh5 #Serverless #FinOps #Containers #PlatformEngineering I really like reading engineering blogs from big orgs like Netflix. Real-world stories from large systems are always interesting. Many function resource configs start as a copy of some other function, with extra padding

200

versions, mTLS verification, and service-to-service communication. Rasanpreet provides a ready-to-use Terraform config below with the full project structure and a GitHub repo is included.

000

handles it at the infrastructure level with YAML config instead. This guide walks through deploying Istio on an existing EKS cluster using Terraform. It covers the architecture including Istiod control plane, Envoy sidecars, Ingress Gateway and then demonstrates traffic splitting between service

100

lckhd.eu/J9AbQL #Istio #ServiceMesh #Terraform #EKS Managing observability, security, and service-to-service communication gets complex as microservices scale. Traffic routing, mTLS, and circuit breaking are important. Implementing these typically means modifying application code. Istio

100

legacy ConfigMap. The key is verification before deletion. Check this out from M Bilal if you're still using the configmap.

000

now. The article below shows how to migrate to EKS Access Entries using a phased hybrid approach rather than flipping a switch in production. It covers inventorying your current state, running both methods in parallel, validating each principal category, and only then cutting the cord on the

100

lckhd.eu/AtQoY3 #EKS #Auth #AccessEntries For anyone still managing EKS access through the aws-auth ConfigMap, you really should move on. A single YAML indentation slip can lock out your admins and CI/CD pipelines at the worst possible moment. There's been a better path for a long time

100

The article below walks through how the Vertical Pod Autoscaler (VPA) Recommender, Updater and Admission Controller work together, and shows how the recently GA'd in-place resize works in Kubernetes 1.35+. Check this out from Arush Kumar.

000

Most values get set when you first build the app, often copied from a similar service, and never touched again. What ends up happening is pods with values set too low that get OOMKilled, or over-provisioned pods that waste CPU and memory.

100

lckhd.eu/j3b5Ha #Kubernetes #VPA #Autoscaling #PlatformEngineering #FinOps Getting resource requests right for a pod in Kubernetes is difficult a lot of the time.

200

Max Levin breaks down the practical differences between these workloads. Worth a read if you're working with Kubernetes and want clarity on when to use each approach.

000

Deployment is what you want. The real mistake is overthinking it or forcing a stateless app into a StatefulSet setup. I keep almost everything stateless with Kubernetes, but there are exceptions. You can always migrate a Deployment to a StatefulSet later if requirements change. This article from

100