Shitanshu Sharma

Dealing with AWS networking can be complicated. There are many options and services you can use. One more recent choice that simplifies a lot of the management is using the Cloud WAN managed service. The article below can help if you would like to transition. If you're using multiple regions, each requires its own transit gateway. Also there a number of peering and routing policies that will need to be setup. If you use the Cloud WAN approach most of this goes away and you can let AWS deal with it all. There is an extra cost but it is likely worth it in saved headaches. Sergio Cambelo provides excellent insights on this topic and this article is worth a read even if you have no current plans to make this change. https://lnkd.in/enZFxpEJ

47

3 Comments

Let me simplify what is happening around the "𝗣𝗿𝗶𝘃𝗮𝘁𝗲 𝗔𝗜 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲" landscape. Today, if an org wants to build an AI data center for GenAI applications at production scale, they have the following challenges to solve - • Infrastructure cost - Which compute (GPU/TPU/accelerator) to choose from? • Operational complexities - Which LLM and how to deploy? Size, tuning, compression, APIs, etc • Deployment constraints - How to have a similar runtime for LLMs across hybrid platforms? One option is to choose all the commoditised open-source components and stitch them together, but still, no guarantee of a 𝘤𝘰𝘯𝘴𝘪𝘴𝘵𝘦𝘯𝘵, 𝘧𝘢𝘴𝘵, 𝘢𝘯𝘥 𝘤𝘰𝘴𝘵-𝘦𝘧𝘧𝘪𝘤𝘪𝘦𝘯𝘵 𝘪𝘯𝘧𝘦𝘳𝘦𝘯𝘤𝘦 𝘢𝘵 𝘴𝘤𝘢𝘭𝘦? Here is when Red Hat is offering its AI services in its decade-old signature style, i.e., building a service around open-source components with Red Hat's guarantee & credibility 𝗣𝗶𝗰𝘁𝘂𝗿𝗲 𝟭 - shows three AI services for different use cases • RH AI inference server - Only serving ready LLMs, out-of-the-box (Consistent, fast, and cost-effective inference) • RH Enterprise Linux AI - Tuning and serving LLMs via a single machine i.e. (Single Linux Server + AI inference) • RH OpenShift AI - (AI inference + multi-machine OpenShift) with complete setup for GenAI and Predictive AI solution, including MLops. 𝗤𝘂𝗲𝘀𝘁𝗶𝗼𝗻: What is cooking inside Red Hat AI? 𝗣𝗶𝗰𝘁𝘂𝗿𝗲 𝟮 gives us the answer. RedHat acquired a firm in 2024 called 𝗡𝗲𝘂𝗿𝗮𝗹 𝗠𝗮𝗴𝗶𝗰--> Neural magic built 𝘃𝗟𝗟𝗠-->vLLM is the core of 𝗥𝗛 𝗔𝗜 "𝘃𝗟𝗟𝗠" - is a cheap and fast LLM serving engine that connects model creators to accelerated hardware providers In simple terms, vLLM is the open-source equivalent of AWS Sagemaker & Google Vertex AI platforms in the context of private cloud, but with the freedom to choose your own model & underlying infrastructure In cloud-native terminology, if - Hugging Face is the 𝘎𝘪𝘵𝘏𝘶𝘣 𝘰𝘧 𝘈𝘐 Ollama is the 𝘋𝘰𝘤𝘬𝘦𝘳 𝘰𝘧 𝘈𝘐, then vLLM is the 𝙆𝟪𝙨 𝙤𝙛 𝘼𝙄. That makes it possible to deploy any model on any infrastructure with a performance guarantee. 𝙎𝙤𝙪𝙧𝙘𝙚 - Official Red Hat sources. Link in comments

36

3 Comments

Having problems with BGP? First, you need to identify where the problem is: 1. BGP neighbor establishment (flapping or failing to peer) 2. Routes missing from the routing table 3. BGP route advertisement issues 3. Multihoming inbound 4. Multihoming outbound Second, dive deeper into the specific section. For example, if routes are missing from the routing table, you may need to use a few commands to validate: 1. Verify BGP peer establishment using the commands: show ip bgp summary show ip bgp neighbor 2. Verify routes learned from peers using the command: show ip bgp neighbor a.b.c.d routes 3. Verify filters using the command: show running-config | include route-map|prefix You can find additional strategies to troubleshoot BGP by following the link below. https://lnkd.in/dX-uf7nf #binaryavenue #bgp

29

3 Comments

Just because AI can generate network automation scripts doesn’t mean you should run them in production. There’s been a big shift lately. Engineers are using large language models to build their first scripts, often without fully understanding what the code does. And I get it. When you’re under pressure to automate, and tools like ChatGPT promise working solutions in seconds, it’s easy to take the shortcut. But infrastructure is not a sandbox. Network automation touches the systems we rely on every day. It connects services, businesses, and controls how data flows securely across organizations. One wrong line of code can take down more than just a lab environment. Out of curiosity, I’ve tested this myself. I took some of the workbook tasks from my Masterclass and asked an LLM to solve them. What happened: → Some scripts looked okay at first glance, but missed key context. → Others skipped important validation steps or made incorrect assumptions. → Most gave answers that were confidently wrong (with zero indication anything was off). None of them would pass the DevNet Expert Lab exam. And more importantly, I wouldn’t run any of them in production. AI will play a growing role in how we work. It already is. But if you can’t evaluate the code it gives you (if you don’t understand what it does and why) then you’re not automating. You’re just delegating blindly. Code will be written once and read countless times. Make sure you can read and not just write. Automation is a skill and like all skills, it’s only valuable if you understand the consequences of what you’re doing. Especially when the network is involved. If you're aiming to sharpen both your tooling and your thinking, the next Masterclass starts August 19.

58

9 Comments

Most enterprise systems don’t fail because of technology. They fail because of boundaries. 1) Modules are synchronous and local. 2) Hosts mark execution boundaries — threads, processes, machines, networks. Cross them, and the rules change. 3) Services are hosts with ownership. 4) Bounded Contexts go further, defining asynchronous contracts through events and notifications. ❌ Confuse these lines, and you get distributed monoliths, fragile chains, and shared cores that choke autonomy. ✅ Respect them, and you get resilient systems that scale. The physics of distributed systems won’t bend to your abstractions. Pretend the network is just a method call, and reality will remind you otherwise — usually in production. 👉 Stop simulating local calls across hosts. Start treating boundaries as the foundation of your architecture. 💡 Read my full article here: https://lnkd.in/eBRBCjt2 #softwarearchitecture #domaindrivendesign #ddd

26

8 Comments