Donne Martin

A cool blog post by Qi Zhu, Jigao Luo and Andrew Lamb on embedding custom indices in Parquet files while staying compatible with the standard. This gives you a ton of flexibility to build proprietary features and stay compatible with the broader ecosystem. Now you can have your cake and eat it too! TL;DR: Parquet allows gaps in between columns and row groups in which you can put anything you want. The parquet metadata allows arbitrary key values in the footer that you can use to add custom metadata about what you put in between columns. https://lnkd.in/gWiwrQ23

70

1 Comment

When Steve Huynh was a principal engineer at Amazon, meetings began with a "study hall." Amazon had a "reading culture" even among engineers, Huynh recently told the Pragmatic Engineer podcast, speaking of his time at the tech giant. Employees frequently drafted six-page memos, which they shared with the company to update progress and demonstrate new projects. Huynh, who said the company's embrace of writing and reading the memos was part of its "secret sauce," said employees' writing was often constrained to the format, whether it was a business strategy or press release. Huynh started at Amazon in 2006, only a few years after the company turned its first profit and while Jeff Bezos was at the helm. Read more about how Bezos instilled this culture of memo-writing from the top down on Business Insider: https://lnkd.in/gBqbEf_j (Credit: Daniel Berman) #amazon #reading #jeffbezos

215

2 Comments

Clay doubling down on increasing action costs while a lot of gtm-e world is laser focused on BYOK workflows with Claude Code is an interesting tension. In the last ~week Cargo shipped a CLI, Apollo.io released an MCP connector for Claude & most of my feed is Claude Code. The ecosystem energy feels entirely focused on terminal-first. I’ve loved Clay for orchestration but have increasingly felt that the UI/UX is slowing me down as I ramp up my terminal-first workflows. That said, terminal-first isn’t multiplayer yet and isn’t in it's final form - esp for larger companies where multiple humans benefit from a shared UI. What’s interesting is this is coming from the people who coined “GTM Engineer” in the first place. The early adopters Clay built the category for are the ones feeling the pull away from the orchestration layer Clay monetizes. The dispersion is probably temporary. But if Clay’s pricing move take them further from the early adopters who built the category, that’s a deliberate bet to cross the chasm to enterprise, while some subset of builders build elsewhere. Interesting to watch.

17

3 Comments

What I'm Reading This Week 📚 ⚡ Diffusion Coming to LLMs https://lnkd.in/gMrhRsvj Lots of announcements at Google IO this week. One that caught my eye is Gemini's new Diffusion model—applying image diffusion algorithms to text generation for faster speeds. This could be a game changer for latency-sensitive applications like IDE auto-completion and real-time code editing. 🔄 Real-Time Collaboration Architecture https://lnkd.in/ghwrBAxT Matthew Weidner delivers a brilliant deep dive into synchronizing state between central servers and multiple clients. If you've ever wondered how Google Docs manages concurrent edits without chaos, this is your answer. His exploration of operational transformation and server reconciliation patterns gives me new appreciation for the complexity behind seamless collaboration. 📈 Economic Reality Check https://lnkd.in/gfTf2TG9 Thomas Klitgaard from the NY Fed provides a down-to-earth explanation of why the US consistently runs trade deficits. While the current administration's trade war rhetoric continues, one thing is clear: rebalancing trade won't happen without painful adjustments for American consumers. The macroeconomic forces at play are more complex than political soundbites suggest. 🎯 Algebraic Effects Explained https://lnkd.in/g8zeFzWX A fascinating introduction to algebraic effects—essentially resumable exceptions that enable elegant control flow. While this cutting-edge programming language feature might be hard to wrap your head around, the Ante team does an excellent job explaining when and why you'd need such abstractions. 🤖 Google's Coding Agent https://jules.google/ Google launches Jules, their take on cloud-hosted agentic coding loops. Between AI-powered IDEs, local agents like Claude Code, and these cloud solutions, we're clearly still figuring out the right form factor for agentic programming. The overloaded "Codex" branding from OpenAI isn't helping clarity either, but the underlying trend toward autonomous code generation is undeniable. What's your take on the current state of AI coding assistants? Are you finding them more helpful or distracting in your development workflow? #TechArchitecture #AICodeGeneration #DistributedSystems #EconomicPolicy #FutureOfProgramming #AgenticDevelopment #DeveloperTools

15

I've been discussing this recently with Frances Buontempo in the context of coding assistants. We hear a lot from developers saying that LLM assistants in their fave IDEs are great for generating a lot of the (boring) boilerplate code needed to stand up a non-trivial system. However, most popular IDEs have been able to do exactly this for decades--whether with "wizards", code-block and project templates, or other various forms of meta-programming. The difference is that the "old" way is much less hungry of resources, while simultaneously being much more predictable and reliable at automatically generating the "boring" stuff. Coding assistants are also good at other things, of course, but many of the popular (in my conversations with developers) tasks are those I would much prefer to do myself, such as writing tests.

15

6 Comments

You spend months building the perfect PyTorch model. Then the real nightmare begins. Porting it. - One version for your flagship mobile app. - Another for that new wearable. - A third for the tiny IoT sensor. Each one needs different optimizations, different pipelines, different frameworks. It's a fragmented, time-sucking mess that kills your time-to-market. This is the single biggest bottleneck holding back true, at-scale edge AI. *Until now.* What if you could just... stop? What if you could use one unified workflow to deploy that one model across BILLIONS of devices? From ultra-efficient microcontrollers to flagship smartphones. From Arm Cortex-M CPUs to high-performance Ethos-U NPUs and Mali GPUs. *This isn't a "what if" anymore.* Meta and Arm just made it a practical reality. Introducing the ExecuTorch 1.0 GA (General Availability) release. This is the on-device runtime for PyTorch that developers have been waiting for. It's one toolset to rule them all. Developers can now author, export, optimize, quantize, and deploy using the same end-to-end PyTorch workflow. The best part? Your apps automatically benefit from performance and efficiency gains. Backend integrations with Arm KleidiAI, TOSA, and CMSIS-NN mean you get optimized performance "for free," with no need to modify your code. This is how we get the real promise of edge AI. Not just cloud-tethered apps, but... ➡️ Private, on-device assistants that run Llama 3. ➡️ Real-time audio generation (Stable Audio in <4 secs). ➡️ Smarter, power-efficient wearables. ➡️ Gaming experiences that adapt in real-time. Meta is already using this to power features for billions of users on Instagram, WhatsApp, and Facebook. Now, it's available to all developers. The fragmented, "port-it-again" days of edge AI are over. The "build-once, deploy-everywhere" era is here. Arm and Meta have dropped the full GA release, docs, tutorials, and pre-validated models. It's all in the blog post here: https://lnkd.in/ggj2rYCT I want to hear from the builders: - How will a single, unified PyTorch workflow change the way you develop for the edge? - What's the first on-device app you're excited to build with this? Drop your thoughts below 👇 and share this with every AI developer you know. This is a big one.

73

3 Comments

A powerful architectural principle is quietly reshaping streaming and real-time analytics: Separate compute from storage. Make compute stateless. Make storage durable and cheap. We’re seeing this clearly in Apache Kafka’s KIP-1150 (Diskless Topics) — where durability moves to object storage and brokers become lightweight, elastic compute nodes. Traditional Elasticsearch tightly couples: Storage (Lucene segments) Compute (query execution) Replication (shard ownership) This coupling leads to: Heavy shard rebalancing Cross-AZ replication cost Over-provisioned nodes (CPU + disk together) Slow elasticity during traffic spikes But imagine a disaggregated model: 📦 Object storage as the source of truth ⚡ Stateless query nodes that cache hot data 🔄 Independent ingest tier flushing immutable segments 📈 Compute scales with traffic — not with data size Now scaling search traffic doesn’t require resharding terabytes of data. Across streaming, analytics, and search: Monolithic nodes → Stateless compute Local disks → Durable object storage Shard ownership → Elastic access And systems that embrace this early will win on cost, elasticity, and operational simplicity. https://lnkd.in/gBGspnpp #Architecture #DistributedSystems #Kafka #Elasticsearch #CloudNative #RealTimeAnalytics

5

The Claude ecosystem is exploding faster than I expected. My newest favorite plugins: - Get Shit Done (orchestration layer that plans and builds apps phase by phase) - Ralph Wiggum Plugin ( iterative, self-referential AI development loops) - n8n-MCP (direct Claude integration with 1,000+ automation nodes) What's interesting is how quickly we've moved from simple AI coding assistants to full orchestration platforms. The shift feels similar to what happened with CI/CD tools a few years ago. First you had basic scripts, then comprehensive pipelines, now full DevOps orchestration. We're seeing the same evolution with AI coding tools, just compressed into months instead of years.  #AI #Claude #SoftwareDevelopment #AIAgents #DevTools #Automation

10

I built two versions of the same conversation Agent using LangChain and LangGraph. LangChain version worked fine, until the agent started to loop endlessly. It was difficult to debug or even understand the why's. So I rebuilt it in LangGraph by expressing the flow as a state graph. It immediately felt more structured, deterministic, and transparent. But it also immediately raised a question - how is this different from traditional state machine? The concept of planner node made lot of sense, where it helped in dynamic transition using the state (input, context, tools etc) and LLM own reasoning. The planner (+ other capabilities) is what turns LangGraph based Agents from a workflow engine into a thinking system.

11

I deployed a single Redis instance on one CPU core and pushed heavy traffic against it. Even under high load, Redis handled massive throughput. Out of curiosity, I increased client threads and expected better parallelism — but throughput stayed almost the same. That’s when the confusion hit: If Redis is single-threaded, why doesn’t it become a bottleneck? What is it actually doing while “waiting” for I/O? The key is that Redis is single-threaded for execution, not blocking. Redis uses an event-driven loop with non-blocking sockets (epoll). It doesn’t sit idle waiting for clients. Instead, it only wakes up when there is real work to do. In practice, Redis works like this: epoll tells Redis a socket is ready Redis reads the request executes the command fully in memory writes the response immediately moves on to the next ready client Each command usually completes in microseconds. Because execution is fast and I/O is non-blocking: Redis never wastes CPU on idle connections There are no locks or context switches Adding more client threads doesn’t help — Redis is already processing requests as fast as the single core allows This is also why Redis doesn’t need multiple threads to scale throughput. One core can easily handle hundreds of thousands of requests per second when commands are small and memory-bound. Redis only starts to struggle when: commands are slow or blocking network bandwidth becomes the limit or that single core is fully saturated So the takeaway is simple: Redis isn’t fast despite being single-threaded. It’s fast because it avoids blocking, locks, and unnecessary concurrency. Minimal design. Maximum efficiency.

1

1 Comment

Rust needs to start pitching itself as an ecosystem rather than just a language. Here's our list of things to talk A LOT more about: - 1st class tooling (cargo, clippy, error messages, etc.) - 1st class platform support (Rust runs everywhere) What else should be on the list?

1

I've been using Claude Code for Rust development and noticed it sometimes generates structurally identical functions with different names. Existing tools either work at the text level or aren't integrated into the Rust toolchain. So I built cargo-dupes — a cargo subcommand that parses Rust into ASTs using syn, normalizes identifiers into positional placeholders, erases literal values, and fingerprints the structural skeleton. Two functions that do the same thing but have different variable names? Same fingerprint. Part 1 covers the normalization approach. Part 2 will cover near-duplicate detection using the Sørensen-Dice coefficient. https://lnkd.in/dPce5Uxm

10

2 Comments

This is fantastic news. The number of mysteries and performance problems caused by poor allocators (glibc malloc) is mind boggling. So in practice, many systems critically rely on jemalloc or tcmalloc and it was very sad to hear about meta abandoning the open development of jemalloc. It's great to hear that this decision has been reverted, and that there will be continued competition in the production-hardened allocator space 🎉.

60

2 Comments

𝗧𝗵𝗲 𝗯𝗲𝘀𝘁 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗹𝗲𝘀𝘀𝗼𝗻 𝗰𝗼𝘀𝘁 𝗼𝘂𝗿 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿 $𝟲𝟬𝟬 𝗮𝗻𝗱 𝗮 𝗺𝗮𝗿𝗿𝗶𝗮𝗴𝗲 𝗮𝗿𝗴𝘂𝗺𝗲𝗻𝘁. Our VP of Engineering, Xiaofan(James) Luan, was supposed to buy his wife a Dior bag for their anniversary.  Instead, he bought three Claude Code subscriptions and spent the holiday trying to cross-compile 2 million lines of C++. Every fix on one platform broke two others.  $600 later, the only output was "git reset --hard" — and a very cold dinner table.😂 "Make it compile on Windows" is a trap. The real goal was "compile everywhere without hacks" — no AI is going to figure that out for you at 2 am. What worked: constraints before code, review tests not code, bottom-up, one layer at a time. Same task, two days. Then he ran six parallel Claude sessions across three machines with git worktree. The bottleneck stopped being intelligence and started being how fast one person can alt-tab. AI solves exactly the problem you give it. Engineering is in knowing which one to give. His wife is still waiting for that bag. Full story: https://lnkd.in/gtsW_Wvk ——— Follow Milvus, created by Zilliz, for everything related to unstructured data

6

𝗧𝗵𝗲 𝗯𝗲𝘀𝘁 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗹𝗲𝘀𝘀𝗼𝗻 𝗰𝗼𝘀𝘁 𝗼𝘂𝗿 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿 $𝟲𝟬𝟬 𝗮𝗻𝗱 𝗮 𝗺𝗮𝗿𝗿𝗶𝗮𝗴𝗲 𝗮𝗿𝗴𝘂𝗺𝗲𝗻𝘁. Our VP of Engineering, Xiaofan(James) Luan, was supposed to buy his wife a Dior bag for their anniversary.  Instead, he bought three Claude Code subscriptions and spent the holiday trying to cross-compile 2 million lines of C++. Every fix on one platform broke two others.  $600 later, the only output was "git reset --hard" — and a very cold dinner table.😂 "Make it compile on Windows" is a trap. The real goal was "compile everywhere without hacks" — no AI is going to figure that out for you at 2 am. What worked: constraints before code, review tests not code, bottom-up, one layer at a time. Same task, two days. Then he ran six parallel Claude sessions across three machines with git worktree. The bottleneck stopped being intelligence and started being how fast one person can alt-tab. AI solves exactly the problem you give it. Engineering is in knowing which one to give. His wife is still waiting for that bag. Full story: https://lnkd.in/gtsW_Wvk ——— Follow Milvus, created by Zilliz, for everything related to unstructured data

9

[My Claude Code setup changelog] Sharing what’s been working well for me lately, in case it’s helpful to others: 1. Agent Teams - Multiple Claude instances that talk to each other. One session is the team lead, spawning teammates with their own context windows and peer-to-peer messaging through a shared task list. Speeds up problem-solving by letting multiple agents work in parallel and compare notes. Leads to faster debugging, broader exploration, and higher-quality output than a single agent. - Example use: "It's taking forever for X page to load for users after we merged our last PR. Spin up a team of agents to figure out root causes (each agent exploring a different hypothesis), ranked from most to least impactful, across every tab." - Link: https://lnkd.in/grtS9qxs 2. Automated demo GIF creation of new features based off a PR - Built a skill where Sonnet reads a git diff, plans a storyboard, launches a browser, navigates the feature, takes screenshots, and stitches them into a GIF. I don't have to record Loom videos or take a bunch of screenshots to attach to my Slack messages explaining new features anymore. - Example prompt: "/pr-demo-gif-agent-browser to showcase the new feature". - Made possible by Chris Tate's Agent Browser CLI: https://lnkd.in/gTey6_N2 3. Reflection + self-improving Skill - After any session where I learn something reusable, a hook prompts Siqi Chen's /claudeception skill to evaluate the session. It analyzes the session, extracts patterns, and creates or updates skills automatically. Combined with Anthropic's /skill-creator. My setup gets smarter every session. - Example use: "I know I'm going to set up Google OAuth for every application I deploy like we did in this session. [if not automatically triggered] Use /claudeception and /skill-creator to create the optimal skill for future sessions when I need to do the same thing?" - Link: https://lnkd.in/gDnCkqwX 4. Select Star + Snowflake MCP servers together for data analysis - We have a lot of data in Snowflake. I usually have a team of agents or a sub-agent use Select Star (data catalog) as the map (knows what tables exist and what columns mean) and then another agent use the Snowflake MCP server to executes queries. Way fewer hallucinations. - Example use: "The metric in this dashboard is wrong. Spin up a sub-agent that uses the Select Star MCP to find the right table, then the Snowflake MCP to figure out why." - Link: https://lnkd.in/g5gpi9cy | https://lnkd.in/g5HafvDd

34

3 Comments