Mark Farrow’s Post

Bcachefs 1.38 has been released by Kent Overstreet, introducing a range of performance, scalability, and stability improvements to this modern copy-on-write filesystem designed to compete with Btrfs and ZFS.

Loading every MCP server upfront is the lazy approach and breaks the model’s awareness of what’s actually relevant. But the opposite fails too. If the schema isn’t there the model doesn’t know the tool exists and can’t choose to call it. What I found works best is loading a lightweight index, tool names plus a one line description of when each should be called (around 50 tokens per tool). Full schemas only get pulled when the tool is actually invoked. Same principle applies to memory. Presence in context should be earned by relevance, not loaded “just in case.”

MindStudio measured the MCP tax many developers carry blind. Every connected server injects its full tool schema into every message … not at session start, every turn. Four servers burn 15,000–20,000 tokens per turn … 9% of a 200k context window before the first real prompt. Estimation: tokens ≈ (tools × 200) + (description chars ÷ 4). Project-scoped configs and pruning mitigate it … something like lazy tool loading at the protocol level could fix it. #MCP #contextEngineering #tokenOptimization #claudeCode

Many teams pursue WHQL certification not as a “nice to have,” but because something is already blocking their progress. The issues can range from Windows displaying warning messages ⚠️ during driver installation to drivers not appearing in Windows Update or the Microsoft Compatibility Center. After working with device vendors and driver teams 🛠️ for years, we gathered the most common recurring questions into a practical FAQ that explains the business value, risks, and key technical steps behind WHQL certification ✔️ And there’s much more to explore. The full article includes additional FAQs and real questions from clients, you’ll find the link in the comments 📄 #WHQL #DriverDevelopment #WindowsDrivers #MicrosoftCertification #DriverTesting

Today the team upgraded our TestAD to 2025 functional level. No drama so far but tomorrow I will verify a few more things to make sure. Next we will enable full 32K DB paging to see the effects this has. So far we are glad our 8K conversion to 32K worked with no faults. Something we didn’t want to do but due to our Prod setup we needed to test. We also have backups to verify… fun! 🤩 Adding to this… Yesterday, one of our remote companies did some fault finding which they presented to us yesterday. They found that their Windows Server 2025 had a bug towards older OS DCs when rotating passwords. Similar to the major Linus domain join bug that Microsoft still won talk to us about!!! The difference… the rotation completes but the destination host never gets the message to say it was completed. Strange huh!? 🤔 Good news they also found a work around they could apply to their machines. This is awesome because our logs showed they were fine still confirming we wouldn’t have know without their amazing troubleshooting over the last few months (huge respect!). 🫡 This gives us the reason to push forward with our upgrades that we paused last year. One step at a time since we can’t trust Microsoft yet! (They are still pumping out problems these days) 🤪

I wish MCP was gRPC I've been building tools that interact with MCP servers, and I keep wishing I had all of the things gRPC would have given me out of the box. Static types chief among them, gRPC also has a wider toolchain that is battle tested in production for over 10 years by thousands of businesses. But I see a world where gRPC service definitions could generate an MCP server itself. The annotations system of protobufs could act as the MCP definition. This is how JSON/REST transcoders work, and it should work for MCP. Large language models that can make usage of tools don't care about the transport, so how we have three separate ways of talking to MCP servers (in less than two years!) is crazy to me. Not the end of the world, but could've been so much better.

A simple Linux command that reminded me why storage and system fundamentals matter So recently, I was using a server and needed to extract a .deb file using the dpkg-deb command to run an application. Simple enough, right? But even after running the command for 3+ hours, it still didn’t complete. I checked online (including ChatGPT and other resources), and most sources suggested it should take just 2–3 minutes on a typical system. So something was clearly off. After digging deeper, I checked my working directory and realized it was mounted over NFS (Network File System). That’s when things started to make sense. Extracting a .deb file involves many small file operations (reads, writes, metadata updates). While this is fast on a local disk, it becomes significantly slower over NFS due to: network latency repeated round trips metadata operation overhead So the issue wasn’t the command itself — it was where it was being executed. This was a great reminder that performance bottlenecks often come from system design choices, not just code or commands. It reinforced for me the importance of understanding: storage systems networking behavior underlying system architecture If you’re curious about how NFS works and why it can impact performance in such cases, this is a great resource: https://lnkd.in/ddD3Xarw #Linux #NFS #SystemDesign #BackendEngineering #PerformanceEngineering #DistributedSystems #DevOps #SoftwareEngineering #Networking #StorageSystems #Debugging #LearningInPublic

From Microsoft Community Hub, Upcoming changes to driver metadata and publishing validations in Hardware Dev Center, by Cymoki "We're making improvements to how the Windows Hardware Dev Center (HDC) handles driver OS compatibility. These changes will..." https://lnkd.in/eHem7_7f

At what point does a tool become an MCP server? And what do modern agent harness like hermes or claude code prefer? Tools or MCPs?

Integrating Claude and NotebookLM is a true game changer. After going through a bunch of tutorials on how to do the integration through an MCP server - this one worked for me without any glitches: https://lnkd.in/eMz4FSiX