What We're Reading

I love competitions where constraints are one of the principles. I used to participate in size coding challenges where you had to deliver a graphics animation in a few KB (4 or 64). It was challenging, you had to learn every single detail to get most of it.

There are competitions like this in the real world, one of them is the DARPA Heavy Lift Challenge. If you forget about the final objective, the challenge itself is super interesting. There were super creative solutions but the winners were classic designs like a helicopter and a drone.

But it does not mean the process is not interesting. This video from the guys in second place shows the actual painful process to get to the best lift ratio. It looks easy but you see how every small detail is important when you stress the system.

Pablo talks about the version control history from the perspective of someone that was part of it. I still think version control should not exist and should be something that happens under the hood (especially the local copy) and it might be real with AI. TBH, not sure if version control makes sense as we know it if AI is the one that codes and manages complexity (which software is all about)

I think these guys got it. The future is not to use LLMs to generate things in the old way, that's faster horses, it's to enable exploring new ways to do software. In this case it's a point-and-click game but it can be extrapolated to any software discipline.

ClickHouse released 26.8 and I think it's a good release. It fixes major flaws in ClickHouse that have been there forever.

• Atomic populate is killer, without that you are going to miss or duplicate records if you want to generate a new mat view
• Adaptive codecs: compressing each block with the compression algorithm that suits it best means you don't need to manually pick one. It uses the one that compresses into less bytes but sometimes you want codecs that decompress/compress faster so be careful. An interesting thing, some codecs don't require to compress the data to know the final size.
• Parquet improvements (data skipping filters mostly). This makes querying parquet files closer to the native ClickHouse format in speed.
• Column statistics: before that you had to know the shape of the data beforehand to design your query to be optimal. Now CH does that for you. It'll fail for sure as stats are, well, stats, but most people don't understand the basics of working data so it's hard for them to pick the right query.
• And then a lot of performance improvements for common use cases people are not aware of. When working with data doing massive uniqs, group bys are part of the day to day