[valkmit]

This looks to be a really cool project. A lot of existing profilers all seem to have different sets of limitations. The last cool profiler I looked at was Google orbit, but it's probably tangential to this project given the breadth of different application times they purport to support.

This is question probably has a very nuanced answer, but here goes anyway - broadly, what kind of performance impact can we expect when running this in an always-online manner?

[tdullien]

We are generally aiming for ~1% of CPU, but on very fork-intensive workloads we exceed this (in order to do symbol-less native code stack unwinding, we need to do some housekeeping for each new process). There are also approx 250m of memory needed for the various pieces of housekeeping data.