Py-spy defaults to blocking because the results can be pretty wrong otherwise: https://github.com/benfred/py-spy/issues/56 . You can see this problem profiling a program like https://github.com/benfred/py-spy/blob/master/tests/scripts/... with or without the nonblocking flag in py-spy - the nonblocking version produces garbage output.
Somewhat interestingly, this problem doesn't seem to occur with Ruby - and rbspy can get away without pausing the target program with only minor errors seen when profiling a similar function. I suspect this is because of differences between how the Ruby and Python interpreters store call stack information, but haven't had a chance to dig into the specifics.
But it's not right for every use case: by design austin/py-spy can only really profile the whole program, and if you want to profile a specific function or endpoint in your program, something like PyInstrument https://github.com/joerick/pyinstrument (which includes Django middlewares & Flask decorators) is a lot more useful.
https://github.com/P403n1x87/austin-python
An early attempt at an APM that was specifically designed to measure endpoint response times was here
https://github.com/P403n1x87/austin/blob/apm/austin/apm.py
The problem with statistical profilers is that they might fail to catch every single call, especially when they are fast.
As for the race conditions, Austin does not introduce any pauses. Even if it did, there would be no guarantee that it paused at a "good" point, so there are no real benefits in terms of accuracy in pausing. Error rates are quite low anyway, so the actual benefit comes from not pausing at all.
Heh, so the only reason there's "no performance cost" is the enormous performance cost of using Python in the first place?
Or, to put it another way, taking Python from 500x SlowerThanCee to 501x is "negligible", but taking C from 1x to 2x slower isn't.
- py-spy: https://github.com/benfred/py-spy (written in Rust)
- pyflame: https://github.com/uber-archive/pyflame (C++, seems to be not maintained anymore)
The "no performance cost" thing is interesting: my experience writing a similar profiler is that there are a couple of things that can affect performance a little bit:
1. You have to make a lot of system calls to read the memory of the target process, and if you want to sample at a high rate then that does use some CPU. This can be an issue if you only have 1 CPU.
2. you have two choices when reading memory from a process: you can either race with the program and hope that you read its memory to get the function stack before it changes what function it's running (and you're likely to win the race, because C is faster than Python), or you can pause the program briefly while taking a sample. py-spy has an option to choose which one you want to do: https://github.com/benfred/py-spy#how-can-i-avoid-pausing-th...
Definitely this method is a lot lower overhead than a tracing profiler that instruments every single function call, and in practice it works well.
One thing I think is nice about this kind of profiler is that reading memory from the target process sounds like a complicated thing, but it's not: you can see austin's code for reading memory here, and it's implemented for 3 platforms in just 130 lines of C: https://github.com/P403n1x87/austin/blob/877e2ff946ea5313e47...