Hey HN! A few months ago I announced that I founded Polar Signals, and today we are super excited to announce our first product (in private beta) as well as a first publicly available feature! Polar Signals Continuous Profiler is a tool we wish we had in the past whenever we optimized code. We are longtime members and maintainers of the Prometheus [1], Thanos [2] and Cortex [3] projects, so we have optimized a thing or two, and we hope to share our experience of doing so through this product. This product is built on top of the open source Conprof [4] project.
Continuous Profiling at heart just means taking CPU/memory/IO profiles periodically and storing them to be queried over time. We have effectively created a purpose built time-series database to store profiling samples, that is based on the Prometheus time-series database. It allows us to efficiently scan over large volumes of data, to extract interesting insights from the profiling data. A common first question at this point is: How much overhead does continuous profiling have on the process that is being profiled? Our experiments have shown no more than 5% CPU/memory overhead, but typically within the 1000ths of CPU cores per process and negligible in memory.
We have many more ideas to explore, but we wanted to get this private beta out to test it with you, and get your feedback! If you want to try the private beta submit a request on our website, or send me an email (email is in my profile)!
The first publicly available feature we have is uploading and sharing pprof profiles for free! No more screenshots of profiles, without the ability to inspect the surrounding context when you have been sent a profile. Share them on GitHub PRs, on slack with your team or however you see fit!
For the Continuous Profiling features just send me an email (my email is in my profile) or fill out the form on the website and we'll invite you to the private beta!
Continuous Profiling at heart just means taking CPU/memory/IO profiles periodically and storing them to be queried over time. We have effectively created a purpose built time-series database to store profiling samples, that is based on the Prometheus time-series database. It allows us to efficiently scan over large volumes of data, to extract interesting insights from the profiling data. A common first question at this point is: How much overhead does continuous profiling have on the process that is being profiled? Our experiments have shown no more than 5% CPU/memory overhead, but typically within the 1000ths of CPU cores per process and negligible in memory.
We have many more ideas to explore, but we wanted to get this private beta out to test it with you, and get your feedback! If you want to try the private beta submit a request on our website, or send me an email (email is in my profile)!
The first publicly available feature we have is uploading and sharing pprof profiles for free! No more screenshots of profiles, without the ability to inspect the surrounding context when you have been sent a profile. Share them on GitHub PRs, on slack with your team or however you see fit!
Thank you for reading! - Polar Signals Team
[1] https://prometheus.io/ [2] https://thanos.io/ [3] https://cortexmetrics.io/ [4] https://github.com/conprof/conprof