Bruce Ashford

Providers like AWS give you the tools needed to ensure your apps can withstand almost any kind of failure. The problem is that many of these situations almost never happen so it's hard to test. Using tools like the Fault Injection Simulator (FIS) can help. There are various FIS tools available for different services. They allow you to make it seem like very rare failure scenarios are happening now. This is super valuable to ensure your code to handle these cases actually works. I recommend setting up tests that run in automation using these tools. This article from Tolgahan Demirbaş shows how to use Terraform to setup Node and Availability Zone failure simulations with an Elastic Kubernetes Service (EKS) cluster. https://lnkd.in/e-rgY3sp

65

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

Agents can now manually prioritize the next important task, email, or chat in a queue. For example, when a customer calls in to enquire about their previously submitted refund request, an agent can search for any pending tickets related to the case, assign it to themselves, and resolve it immediately. https://lnkd.in/gEvgZ8zd

24

Anyone who has built more than one computer‑vision project knows the pain of writing the same boilerplate over and over - loading datasets, converting labels, drawing bounding boxes and counting objects. These tasks aren’t glamorous, but they’re essential for turning raw model output into something useful. Supervision, an open‑source Python toolkit from Roboflow, was created to tackle exactly those headaches. The library doesn’t run models itself; instead, it connects to whatever model you’re using and provides reusable utilities around it. It’s model‑agnostic, so you can plug in any classifier, object detector or segmentation model. They’ve even built connectors for popular frameworks like Ultralytics YOLO, Hugging Face Transformers and MMDetection. Here are a few of the things it does: - Visualizing detections with minimal code: The library includes highly customizable “annotators.” Instead of manually drawing rectangles or polygons on frames, you can call classes like BoxAnnotator, PolygonAnnotator, or more specialised versions such as percentage bars and oriented boxes. These handle styling, text, colours and even rounded corners. - Handling datasets efficiently: There are utilities to load data from common formats (YOLO, Pascal VOC, COCO), split them into training/validation/test sets, merge multiple datasets, and convert between formats. These functions are designed to load images on demand, so they scale to large datasets. - Zone based counting and tracking. Supervision comes with ready made components to track objects across frames and count how many cross a virtual line or enter a polygon. A recent update even lets you specify which part of a bounding box must cross the line - such as the bottom center of the box - to register a count. This is handy for traffic monitoring or analyzing foot traffic through retail entrances. Because it’s MIT‑licensed, Supervision is free for commercial use, and the team encourages community contributions. The library has been under active development since 2023; recent releases have added support for oriented bounding boxes, detection smoothing, color‑palette customization, and other enhancements. For folks who aren’t deep into computer vision jargon: “object detection” refers to algorithms that identify and locate objects in an image (e.g., finding all the cars in a photo). “Segmentation” goes a step further by outlining the precise shape of objects. “Annotation” is simply the process of drawing these results on an image or video to make them understandable to humans. No tool is a silver bullet - you still need to choose or train the right model for your task, and in some cases you might want custom visualization logic. But for common tasks like loading data, visualizing detections and counting objects in zones, having a well‑maintained open‑source toolbox saves time and reduces errors. #innovation #technology #future #management #startups

30

There have been three distinct phases of disaggregated networking history. And we're now in the "cross the chasm" stage for enterprise adoption. As I've dug into network disaggregation and digested much of what has happened, it has been helpful for me to think about its history in three phases: 1. Emergence This is where things started, in the early 2010s, with the founding of the Open Compute Project (OCP) led by Facebook, the contribution of open switch designs, all carrying the intent of hyperscalers for open networking. In addition, you had the emergence of startups like Pica8, Cumulus Networks, Big Switch Networks, and Pluribus, all contributing software to the ecosystem. And finally, this emergence stage overlapped and coincided with the work of the Open Networking Foundation (ONF), which was progressing from focusing on OpenFlow and was leading towards the broader SDN/NFV work (and has since merged its projects into the Linux Foundation in 2023). This represented the solution space where telcos wanted to use disaggregated networking hardware or "white boxes." 2. Convergence The mid 2010s' saw a convergence stage emerge. The most important event was the contribution by Microsoft of #SONiC to the OCP in 2016, as a community-led Network Operating System (NOS). SONiC rapidly gained traction and has become the driving force for the growth of disaggregated networking. As a result (IMO), the early startups who were focused on proprietary NOS and other software components started to move towards a different type of convergence--acquisitions by OEMs and other players like Nvidia. By 2020 many of these early players were no longer independent. In addition, telcos started to adopt disaggregated networking for SDN, NFV, Open RAN, and Open 5G Core use cases. 3. Crossing the Chasm The third phase is where disaggregated networking starts to become safe for enterprises. The signal event was the transfer of SONiC to the The Linux Foundation Foundation in 2022, which created a "safer" governance structure. There are a number of other factors that have made disaggregated networking ready for enterprise adoption: ➕ OCP's initiation of validated hardware and validated reference designs ➕ The growth of an automation and orchestration ecosystem based on open source (Ansible by Red Hat, NetBox Labs, Network to Code, OpsMill), and vendors (Aviz Networks, Dorado Software, BE Networks), ➕ The rise of OEM-like hardware vendors like Celestica that offer robust service and support packages that go beyond what "whitebox" vendors typically do. I think this is a fascinating time for disaggregated networking. I noticed that at #AutoCon3 (Network Automation Forum), there was a lightning presentation focused on SONIC network automation. I think we'll see more going forward. Anyway, it was helpful for me to think of things this way. I'd love to hear your thoughts. And what did I miss? #disaggregatednetworking #whiteboxswitch #SONiC #OCP

20

8 Comments