I'm starting to use claude code differently as I run into the same issues again and again. The biggest one is file size bloat... It's totally fixable by hand but The AI can't deal with 6k lines of code... I'm calling it a Smithers loop, as I'm giving it a different context each time we loop

Writing a sql query engine is a compiler. Except the language is mostly pre-defined, the users are mostly other tools and everyone thinks it's a solved problem (use postgres)

Today's challenge : Teach an AI how to use kcache grind. So far my AI Dev tools can use dtrace, valgrind and perf. But we need to go deeper.... Performance is one of the few ways to reduce TCO cleanly so performance is everyone's problem.

There's a new dresden files book. It's one of those: "am I still reading it because it's good, or because I'm invested". I enjoyed the earlier books at the time as a fun ride through a magical power fantasy, but on reflection I do wonder a bit. Still bought it, but this is the last auto buy

I've been writing a database from scratch over the last few months with claude code and it's definitely an experience. The most telling thing is that it really doesn't know anything you don't tell it. It's a good faster keyboard but it's even remotely good at stuff with lots of moving parts

An enduring CS memory: a uni computer lab, ~2007. I was failing at printf and realising how weird it is. Nearby, students a year ahead were learning Java. I overhear: “wait… that’s printf.” Turns out one of them had implemented ALL of printf in Java—rather than just learning how println works...

Claude code makes CMake only the worst thing in my day once a week rather than every day I work on a Cpp project.

It's weird what software tests can and can't catch. Especially with AI. Unit coverage always feels like a vanity metric, as compilers and types should do most of that for you. When it comes to feature level testing, you can end up catching all kinds of odd behaviour with randomised inputs

First trip of the new year! I'm in #NYC for the week, talking data, AI and all things distributed! We're also hosting a meetup (link in next post) on Thursday so please come and say hello, or message for a coffee!

The new year started out on a deep freeze and then getting sick. It's been very un fun.

Had some fun this month making #clojure work with you apache flink: github.com/Ugbot/flink-...

I'm in the wierd space of evaluating the costs of mutexs vs CAS. Most of my time is spent removing mutexs, with CAS being the replacement. But you really shouldn't have too many spin locks... Currently working back through my assumptions of what and where to be fast....

The main reason I actively hate c++ is dispite all the issues with it, there's no limit to perf you can get out of it. Even with the horrible ergonomics, the foot guns, the bonkers approach of the committee and the bazantye over engineering everyone seems to commit to, you can go so fast.

With uber publishing cool ideas around moving to flink and IBM playing catchup by scooping up confluent, here's my take : www.linkedin.com/pulse/being-...

With the CFPs for current london and bengaluru running, wondering what to submit: Would ppl prefer to hear about AI & memory, consensus and replication, performance and compute, apache Fluss and data lakes, a how to on common usecases, or something entirely different?

Yeay... oomed my macbook again... Seriously need to upgrade at some point 😅

Some plane rides later, I'm working with our partners @factorhouse.io in #melbourne! Time to talk about Streaming data and drink good coffee!

The killer phrase in dangerous data systems : Business critical spreadsheets 😅

After another dig through an #agentic framework without a memory manager, here's my blog explaining why you need one, with links to an open source one I wrote : www.linkedin.com/pulse/memory...

Claude opus 4.5 is out? Not with a bang but with a whimper I think we're deep into diminishing returns of brute frontier model performance at the moment, but we'll see... The gemini release was wildly underwhelming so I've got minimum expectations

After reading a couple of recent papers I wrote an article: The Fine Art of Doing Nothing (In Distributed Systems) linkedin.com/pulse/fine-art… via

Another day, another airport. And today I'm using the layover to fix up the flink sql #DBT adapter. It's proving to be a fun exercise as dbt is not really a tool I've used before, but the concept is simple enough. The details however are not 😅

I'm giving a talk at a meetup on Thursday and the building is called The Bower, and ngl I'm a little disappointed in how unromantic the building is

This event is not to be missed, three awesome speakers!

Today in my database experiments : - Too many indexs - Too many mutexs - NVME drives are *FAST*

So mistakes were made and I've realised I'm writing a database... By virtue of AI powered prototyping, it's working 😅 I'm not sure if this is good or bad, but a transactional column store is proving to be very useful.

Halloween in new Orleans: am I from interview of a vampire or Austin and powers? 😅

On my way to #current25 in #NOLA! Time to talk about Streaming in spooky season! A quick thank you to @adipolak for adding me to the committee for this one! The talks look awesome!

I used to save CMake templates because of the pain of creating them. This is one place I am happy to give to AI.

One of the biggest outcomes of AI coding tools is that I'm finally seperating out my libraries into reusable blocks. Today it's logging in in C & C++. It needs to be able to be configured at build time, and at runtime. To have outputs as either stdio or files, and debouncing.

So I got carried away and built a vibe coding Ide... It takes preferences seriously with agents built in to do things like testing & docs and build scripts. Also runs off a local llm by default.

As much as I prefer writing other languages, it's annoying how much more productive I can be writing c++. There's a 1 million request challenge to win!

As much as I prefer writing other languages, it's annoying how productive I can be writing c++. There's a 1 million request challenge to win!

The @KuzuDB repo has been archived on github. But the entitlement never stops... Opensource doesn't mean free. Open source does not mean maintained. Things still take time and money to build and test. If you're making money from it, maybe contribute / pay? www.theregister.com/2025/10/14/k...

Flight delays on the way to flinkward means time to mess about with code!! Turns out that if you spend ages hacking it, you can run @duckdb with more than one node!😁 With only minimal slowdown! 😭

The amount of guard rails I've had to add to my environment to stop @cursor_ai just faking stuff is extreme.... There's literally rules against mocks, and I still needed to add an mcp service that checks code paths are being called... Seriously, I'm not in a hurry stop trying to rush.

I didn't think I'd find myself writing a cypher parser at almost 11pm on a Saturday just before flink forward, but that's life 🤣 Now to give the optimizer another poke...

Turns out that you really can build AND hot reload stuff with TinyCC so fast that it makes cython Dev workflows as fast as #python.... Shout out to @charshenton for the idea.

Continuing my experiments in cython, I built github.com/Ugbot/CyRedis because redis / valkey are cool but we end up not getting all the juice out of it in python land. It's wildly experimental, but I let claude code write the readme 😅 I think my Next experiment will be "fasterAPI" 😎

Today's interesting diversion was from building an operator of other operators in k8s. Firstly, as much as I like the operator pattern, the implementation leaves much to be desired. Not sure there's a better way but the current one is annoying... Especially Cascading roles through systems

In a moment of true AI hallucination, #claude decided that the best way forward was to mix macros with template metaprogramming in C++ There are a few things I really didn't want to see again in code, and this was on the list...

In a moment of true AI hallucination, #claude decided that the best way forward was to mix macros with template metaprogramming in C++ There are a few things I really didn't want to do see again in code, and this was on the list...

Well, I'm in the deep stuff now.... I've got a #postgres plugin called wal2arrow and it's very very fast. Now to see if I can make it work from Pgoutput directly... Time to channel @gunnarmorling.dev ...

I may have fallen down the rabbit hole here but... #apacheArrow had my curiosity but now had my attention... And I'm building wal2arrow for #postgres and #mysql... It's kind of awesome not to have json /avro around in the hot paths.

The vibe of today's claude code is an epilepsy warning!

There's no escaping the fact that python is slow. Slow enough that it's got a whole ecosystem of tools like numpy, numba and cython to fix that. Currently I'm working on a way to apply numba dynamically across python code via ast and byte code parsing. Wierdly, it's working 🤣

A quick PSA: The number of AI written "seo" articles is getting rather crazy, but if you have to, at least run the article through a fact check so you don't end up with content that makes you look stupid.... Also use an up to date LLM, the new ones are better...

We're having a family gathering for my dad's 70th birthday, and when I asked my nephew (16) what he was watching, he seemed surprised I knew what naruto was 🤣... Then I felt old when he didn't know full metal alchemist... But now we can watch cowboy bebop together!

I spent some time last week Playing with Apache Arrow. Sometimes the numbers are frankly crazy....