Starting to map the graph for a founding engineer at Perfloop.
Looking for real systems depth - perf, infra, observability, AI infra, devtools, databases, runtimes - plus extreme agency, product/architecture judgment, and native agentic building.
Europe / near-Europe overlap preferred.
@tsenart@AWS they launch during reInvent, so it's either Dec 26 or Dec 27... my guess would be the latter -whoever does it sooner will most likely run it on AWS anyway, so not like they needed to hurry
As a software engineer, your job is not primarily to write code. Much of your work is more like that of a detective. « Why is this not work ? » « Why is this slow ? » « Why is this sometimes incorrect ? »
If you have done any kind of real-world engineering, you know how true it is. We have all spent entire days testing and reading code to figure out why we get an unexpected result.
And, you know what? Sometimes we never find out. If your system is complex enough that you cannot read and understand everything... and that's most of what happens these days, then you have to rely on pattern matching and heuristics.
Of course, AI (copilot) can help. But only up to a point. For hard problems, you need to run experiments, many different experiments, to figure out the truth. You need to read and understand a lot of material.
And, sorry but your super smart AI can make things worse. Here is a true story from my youngest son who studies programming in college...
Peer: Can you help me with this code? [shows massive AI-generated codebase]
My son: Your problem is not that it does not work, your problem is a lack of understanding of what is happening. Start there.
Every other evals product is built by ML people with python brain poisoning and no idea what good DX looks like.
This is different.
A few highlights in the thread ⬇️
From 0 to 178,600,000,000 rows/second in 30 days:
Introducing Haydex - our hyper-fast filtering system that makes trillion-row needle-in-the-haystack queries interactive.
Read @tsenart's account of the journey below ↓, from a failed V0 → complete redesign → 12.7x speedups → production ⚡️
@aantn Funnily enough I was digging into something similar yesterday, it was also using the topk construct - I believe it's there to ensure you have a 1-to-1 mapping for the grouping. (can't multiply series with label=foo with possibly multiple series with label=foo)
@atcb It was a super light paper-like extension for the ISS manufactured in Indonesia, which unfortunately failed and was blown away by the solar wind.
@aantn I assume those spikes are counted into the average? Would still set the request few percent above the average, so that latency isn't too bad during those spikes... Say 2.5
If you get a kick out of Distributed systems, algorithms and data structures... We are hiring @AxiomFM
https://t.co/1Xu56umqS5
apply directly then hit me up :)