We were just at FOSDEM and apart from technical discussions (thanks to everyone for great questions and discussions) and eating delicious waffles (turns out I was wrong that I don't like Belgian waffles) we had the opportunity to give four different talks across the Databases, Rust and eBPF devrooms. In case you weren't able to attend in person or just want to catch up on what we're working on this recap is for you.
In this session, I talked about the massive architectural shift our database underwent over the last year. We reached a point where legacy designs were limiting our scale—recovery times were in the tens of minutes, and operational costs were climbing. By leveraging modern open-source technologies like Vortex, DataFusion, and Delta Lake, we rebuilt a cloud-native database that dropped costs by 50% and improved query latency by 200%. I walked through the process of how we came to terms that our system wasn't working and moved from a disk-centric model to a truly distributed, high-performance system.
How do you profile production Rust applications without the massive overhead of copying stack traces? Brennan introduced the FOSDEM audience to Parca, our open-source continuous profiler. He highlighted Parca’s unique ability to perform unwinding directly within eBPF programs, making it low-overhead enough to run 24/7. The talk specifically covered Rust-centric features like jemalloc integration and using "custom labels" to filter profiles by specific application tags or trace IDs. He was even brave enough to perform a live demo of all of this in action!
Building distributed systems is hard; testing them is harder. Frederic explored how we ensure the correctness of our storage system using Deterministic Simulation Testing (DST). By modeling core components as state machines rather than relying on standard async runtimes, we can explore countless random execution paths and inject failures—all while maintaining the ability to perfectly reproduce any bug using a single random seed. If you’re building resilient software in Rust this approach may be worth looking into.
Out-of-memory (OOM) kills are the bane of any SRE’s existence, largely because the process is gone before you can see what went wrong. Tommy presented OOMProf, a Go library that uses eBPF to "catch" the kernel’s OOM signal. It records a final memory profile just before the program is terminated and sends it to a Parca server. This deep-dive into kernel tracepoints showed how we can finally get visibility into the "deathbed" of a Go process.
We want to thank the FOSDEM organizers and the volunteers who make this event possible every year. There were so many great talks, and inspiring ideas there we're grateful to be apart of.
If you missed the opportunity to talk to us at FOSDEM you can jump into the Parca or Polar Signals Discord, we'd love to hear from you!
See you next year, Brussels!