Salvatore Vadacca

Kubernetes is no longer just a tactical tool; it’s a critical foundation for innovation. That’s why our 2025 State of Production Kubernetes matters. Now in its fifth year, this report has become a trusted resource for understanding how real organizations scale, modernize, and operate Kubernetes in production. It’s based on independent research with 455 enterprise and public sector users, and the insights are grounded, strategic, and incredibly relevant. If you’re making decisions around AI, platform engineering, or infrastructure complexity, I highly recommend giving this a read: https://lnkd.in/gjNWvp39 #Kubernetes #KubernetesManagement #AI #PlatformEngineering #VirtualMachines

80

2 commenti

The future is here - Kafka is cheap to run. It just isn't evenly distributed yet. Case studies with Diskless Kafka architectures are starting to pop up now. The cost savings are as shocking as marketed: • WarpStream 🤝 Pixel Federation - $60k/yr AWS MSK spend down to <$10k/yr with WarpStream • Aiven 🤝 OpsHelm - $50k/yr AWS MSK spend down to <$10k/yr with Diskless Kafka BYOC Both are 5-6x cost decreases. This is what I mean when I say the Kafka market is seeing massive cost deflation through commoditization and innovation. Watch this space.

162

16 commenti

If you’ve worked with Apache Kafka, you know the trade-offs: Blazing-fast local disk writes but costly replication across zones. Tiered storage for long retention but complexity in managing hot vs cold data. The Kafka community has been asking: Can we make object storage the first-class durable store without losing Kafka’s semantics? That’s where Diskless Topics (KIP-1150 & KIP-1163) by Aiven started the journey — rethinking Kafka’s durability around object storage as the source of truth. Now comes Diskless 2.0 — a unified, zero-copy approach that: ✅ Removes parallel “disk vs tiered” estates ✅ Simplifies the produce/consume path ✅ Enables zero-copy migration (flip a topic from classic to diskless without reprocessing) ✅ Cuts costs by eliminating cross-AZ replication and reducing broker disk reliance This isn’t just an incremental patch. It’s a paradigm shift in how we think about Kafka in the cloud era — where durability and economics favor object storage, and Kafka brokers can finally be “lighter.” Of course, challenges remain: -Latency sensitivity (classic still wins for ultra-low-latency cases) -Compaction and deletion semantics (work in progress) -Operational readiness for large clusters But that’s the beauty of open source — progress through community iteration. And Diskless 2.0 feels like a community win, not just a vendor push. 👉 Dive deeper: Read the KIP-1150 proposal here: fandf.co/3VlQxW1 Explore Aiven’s take on Diskless 2.0: fandf.co/46pczw0 What do you think? Could Diskless 2.0 become the default way to run Kafka in the next 3–5 years, or will classic/tiered still dominate latency-critical workloads? Kudos to Aiven for releasing Diskless 2.0 and sponsoring this post. #Kafka #Sponsored #Aiven

266

49 commenti

This is without a doubt the best #Kubernetes write-up of features in v1.34 that I've come across. Awesome work, Nicolas Vermandé! I especially agree with the intro: "Kubernetes has evolved through three eras. The first was about survival, just getting containers scheduled and running. The second era was scale, stretching clusters to thousands of nodes. But scale exposed cracks. Abstractions broke down against messy infrastructure. In 2025, with AI workloads everywhere, those cracks show up at 3 AM when an ML job fails because the scheduler doesn’t understand GPU topology. Or in finance reviews, when inter-zone transfer fees quietly spike the cloud bill. Meanwhile, autoscalers thrash on noisy metrics. This is the third era of Kubernetes: production reality." The article goes into great detail and shares insights about the following topics, taken directly from the table of contents: 1. Dynamic Resource Allocation: The End of the “GPU Lottery” (GA)  2. Swap Memory Support (Beta)  3. Pressure Stall Information (PSI): Your True Performance Indicator (Beta) 4. Mutating Admissions Policies: Faster, Simpler, Safer (Beta) 5. Selector-Based Authorization: True Least Privilege (GA) 6. ServiceAccount Token Image Pulls: Short-Lived Secrets by Default (Beta) 7. Anonymous Auth Restrictions: Tighter by Default, Flexible Where Needed (GA) 8. VolumeAttributesClass: Live Volume Tuning, No Re-Mount Required (GA) 9. Job PodReplacementPolicy: No More Double Scheduling (GA) https://lnkd.in/dsv3hHXs

65

1 commento

🔍 Why the new Humanitec Orchestrator is a game changer for Platform Engineers Last week, we launched the all-new Humanitec Orchestrator. Today, we want to go deeper and show you why it matters. Most orchestration tools are tied to a single paradigm. They work if you’re all-in on Kubernetes or locked into a single vendor. But modern enterprises don’t look like that. They span Kubernetes clusters, VMs, bare metal servers, functions, and hybrid setups—all subject to strict compliance and security requirements. The new Humanitec Orchestrator is built for this reality: 👉 Multi-compute by design: K8s, VMs, servers, functions — unified in one control plane. 👉 Deterministic and compliant: policies enforced everywhere, no exceptions. 👉 API-first: equally consumable by humans, portals, or AI agents. In this video, we’ll explain the architecture and then demo how the Orchestrator makes complex setups simple, compliant, and reliable. Try it today on Humanitec.com or reach out and we'll show you!

43

4 commenti

Dear Architects, This week, we start with a great case study on how Vercel leveraged #AWS Lambda to build Fluid Compute, their new offering. I really enjoyed their novel approach to tackling this challenge. Next up, let's talk about microservices. An InfoQ article uncovers why these flexible and scalable systems often consume more energy than monoliths and shares practical design and operational strategies to build greener, more efficient microservices that balance modularity with sustainability. For those diving into #API design, Learning API styles offers clear guidance on understanding various API styles and trade-offs, helping architects make smarter integration choices that align with technical and business goals. Shifting from #architecture patterns to system thinking, a thought-provoking piece on constraints and emergence reveals how carefully crafted constraints can foster innovation, resilience, and effective collaboration within complex systems. Finally, a Thoughtworks podcast tackles the evolving role of documentation in the era of large language models, highlighting how caring, living documentation practices powered by #AI can improve developer experience and knowledge sharing. READ & SUBSCRIBE (it's FREE): https://lnkd.in/eqDjmd34 O'Reilly Lukasz Dynowski

47

4 commenti

I remember when I was learning more about software architectures; I compared a lot of solutions architecture to get a good grasp of what good software architecture should look like. And as well, the "Fundamentals of Software Architecture By Mark Richards and Neal Ford" and "Design It By Michael Keeling" were two great books that helped me solidify my understanding of software architectures better. For LLMs, a lot of models are still powered with the original Transformer architecture and most innovations have happened with component replacement and improvements for both KV cache memory utilization and inference time latency improvements--such as different attention mechanisms usage and normalization placement and feedforward layers. One resource that helped me get a good grasp of these changes is Sebastian Raschka, PhD article on comparing LLM model architectures. Really great resource on getting better understanding of what has changed and what remained unchanged in all model innovations. #artificialintelligence #deeplearning #machinelearning #llm  #startups #innovation #technology

29

Most node failures don’t happen instantly. Kubernetes tells you — if you know where to listen. Three months ago, at 2:14 AM, a production cluster went red. Dashboards were still showing CPU, memory, and pod counts in the green. But kubelet had been warning me since 11:02 PM. 𝗧𝗵𝗲 𝘀𝗶𝗴𝗻𝗮𝗹: ↳ kube_node_status_condition{condition="MemoryPressure"} = true Originates in kubelet’s /metrics. ↳ Surfaces in kube-state-metrics only if node metrics scraping is enabled. ↳ It flips when memory.available drops below --eviction-hard (ours: <500Mi, default: <100Mi). 𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱 𝗻𝗲𝘅𝘁: ↬ Eviction manager removed BestEffort pods first, then Burstable pods as pressure persisted — QoS in action. ↬ Signal flapped every 5–10 min for 3 hrs before failure. ↬ container_memory_usage_bytes looked normal — hiding the risk. ↬ At 2:13 AM, critical Burstable workloads were evicted. ↬ Re-scheduling + API object floods spiked API server CPU from 18% ↬ 94%.  ↬ p95 latency jumped from 120 ms → 2.1 s in under 60 sec. 𝗪𝗵𝗮𝘁 𝗜 𝗰𝗵𝗮𝗻𝗴𝗲𝗱: ➡ Scrape kubelet /metrics directly for node_pressure_memory & node_pressure_disk. ➡ Alert on first true in 30 min window with rate limit to avoid flapping noise. ➡ Visualize eviction frequency per node in Grafana. ➡ Tune eviction thresholds to actual workload patterns. ➡ Simulate eviction in staging to measure mitigation speed. 𝗟𝗲𝘀𝘀𝗼𝗻 𝗹𝗲𝗮𝗿𝗻𝗲𝗱: Your earliest outage warnings live in kubelet’s survival instincts — long before app metrics turn red. 📧 If you want my “Kubelet Signal Playbook” with the exact PromQL queries and Grafana panels we now use, comment 𝘀𝗶𝗴𝗻𝗮𝗹𝘀 and I’ll send it.   🗨️ 𝗕𝘂𝘁 𝗳𝗶𝗿𝘀𝘁, 𝘄𝗵𝗮𝘁’𝘀 𝘁𝗵𝗲 𝗼𝗻𝗲 𝗺𝗲𝘁𝗿𝗶𝗰 𝘆𝗼𝘂 𝘄𝗶𝘀𝗵 𝘆𝗼𝘂’𝗱 𝗰𝗮𝘂𝗴𝗵𝘁 𝗲𝗮𝗿𝗹𝗶𝗲𝗿?

30

1 commento