This enables a practical strategy: start with minimal requests and let VPA right-size them dynamically.
Anup Dubey shows the setup below and explains how VPA makes its recommendations.
Too high and you waste capacity. The Vertical Pod Autoscaler helps, but traditionally required pod restarts to apply changes.
#VPA v1.5.0 with Kubernetes 1.33 introduces InPlaceOrRecreate mode, which can adjust resource requests on running containers without disrupting the pod.
https://lckhd.eu/InKKCB
Setting static CPU and memory requests in #Kubernetes is often guesswork. Too low and you get throttling or OOMKills.
Sardarnoorulhassan has put together a good read including the choice of EC2 instance for the load generator to reading CloudWatch alongside k6 output. Check it out if you run anything on Fargate with RDS behind it or are curious about this kind of testing.
The results are a system that handled 500 gradual users with zero errors collapsed to an 89% failure rate when the same traffic arrived as a spike. The database hit its ceiling long before the compute layer did, and the auto-scaling cooldown turned into a 2 min exposure.
#Fargate API holds up and where it quietly falls apart.
https://lckhd.eu/1UUTts
#LoadTesting often gets glossed over before going live, but the article below shows how much signal you actually get when you treat it as a proper investigation. It describes four scripts, four questions, and a clear picture of where an AWS (Elastic Container Service) #ECS
The article from the Pinterest Engineering team goes over perf, mpstat, and Flamescope-based temporal profiling. The full investigation is an interesting read and includes a lot more than just the tidy conclusion. (4/4)
on individual cores, and eventually to zombie memory cgroups piling up from a crashlooping Elastic Container Service (ECS) agent that shipped by default in their base AMI. (3/4)
deep in the stack, somewhere nobody was actively watching. The article below discusses an example of that pattern in action.
The team involved chased intermittent Ray training crashes for three months, tracing them through AWS ENA network driver resets, CPU starvation (2/4)
https://lckhd.eu/4kjeNv
Debugging distributed Machine Learning (ML) training failures is rarely about the ML itself. More often it comes down to an unexpected interaction 🧵
Seeing examples always helps for these things and this one can help if you're putting together a lightweight ingestion layer or just want a clearer picture of how these services connect in practice. Check this out from Drishi Gupta. (3/3)
The article moves through the setup, covering the S3 bucket, the IAM role, a Python Lambda that timestamps files and routes them into folders based on the resource path, and the API Gateway method with a mapping template to pass the body through cleanly to Lambda. (2/3)
medium.com/@drishigupta...
Building apps with #serverless and managed services on #AWS is mostly about wiring the right pieces together correctly. The example below shows how #APIGateway, #Lambda, and #S3 can work together to accept JSON payloads and land them in storage without servers. 🧵
https://lckhd.eu/2Lwwcs
And yet another new model. Just when you were getting used to Claude Opus 4.6, we have Claude Opus 4.7 available on Amazon Bedrock! Can't wait to try it out.
It covers provider aliases, route tables, security groups, and the two-step peering acceptance process that AWS requires for cross-region setups.
Adarsh Gupta put together a good walkthrough below of #VPCPeering setup. Check it out!
(#IaC) tools like #Terraform make these setups easier. The article below shows a Terraform implementation for cross-region VPC peering.
https://lckhd.eu/wXy3wm
VPC peering is one of those foundational #AWS networking concepts that's worth understanding well. It enables private communication between VPCs without routing traffic over the public internet, which matters for both security and latency.
Using Infrastructure as Code
fire on a schedule, respond to GitHub events, or trigger from an API call, and they keep working whether you're at your desk or not.
So many things to check out - and so little time!!
https://lckhd.eu/CvRAQ2
The speed we're moving at is crazy! Anthropic seems to put out a new feature every week (or day). This week it's "routines".
#ClaudeCode #routines let you package a prompt, repositories, and connectors into a configuration that runs autonomously in the cloud. They can
https://lckhd.eu/v7A9yZ
🆕 My latest blog post! Three decades on the #terminal, and I still think faster there than anywhere else.
Sarah Wanjiru documents it all, including the part where she ignored her budget alerts. #finops #devops
https://lckhd.eu/vzYnKr
is practical and worth adopting early.
The article below walks through what can go wrong and a layered cleanup approach that can be followed.
Automated infrastructure cleanup needs to be explicit and unconditional. Relying on a single mechanism like defer is not enough when CI runners can die mid-job. The three-layer fix outlined here
Getting a $20 AWS bill from orphaned EC2 instances running overnight or some NAT Gateways you forgot about. These things are not a disaster, but the kind of mistake that teaches you more than most tutorials will.
single-container pods through multi-container configurations, showing the specific #kubectl commands you need at each stage. Khandokar Ismail has put together a good guide worth checking out.
https://lckhd.eu/IueShO
Working with #Kubernetes Pods is a important skill, but it helps to have a clear, step-by-step reference when you need to create pods, exec into containers, or #troubleshoot multi-container setups. The article below covers how to do these.
The example goes from
This covers the daily workflow, the pipelines that replace 20-line scripts (grep | awk | sort | uniq -c | sort -rn | head - you know the one), the macOS vs #Linux issues that trip everyone up, and the modern alternatives worth adopting (ripgrep, fd, jq, yq, ncdu).
I wrote down the #Bash commands I actually reach for every day. It's not an exhaustive reference, but the patterns that genuinely save me time with navigating, searching, processing, and more.