Significant slowdown since version 0.1.10

[martinstein]

When I was running py-spy 0.1.8, there was no noticeable slowdown of the program that I was profiling (rate=700). In my case this was a web request in Pyramid that returns in 6 seconds.

After upgrading to py-spy 0.1.10 (still using rate=700), the performance of the profiled program drops significantly: the same web request now takes 18-20 seconds! When I use --nonblocking, the performance becomes fast again.

So it seems that between 0.1.8 to 0.1.10, a significant slowdown of the profiled program was introduced for the default blocking mode.

System: Windows
Python: 3.7

But anyway, thanks a lot for such a great tool! Py-spy has been immensely helpful.

[benfred]

Hmm - that doesn't seem ideal. When I tested this I was seeing a much much smaller slowdown - on the order of ~2% instead.

How many threads are in your target process? There is a potential performance optimization that we could do: instead of locking the whole process it might be possible to just lock one thread at a time, which might help alleviate this.

Also - it seems like you should be able to drop your sampling rate down to have less of an impact.

[martinstein]

In my local dev setup where I ran the profiling, I used the waitress server that is the default for Pyramid. I think by default waitress spawns 4 threads on startup (https://docs.pylonsproject.org/projects/waitress/en/stable/arguments.html), but I only sent 1 request so only 1 thread was actively running.

I just ran a quick test where I reduced the rate from 700 to 100. The slowdown of the request went down from 300% to roughly 10% (much better, but still seems higher than previously).

[martinstein]

On a completely unrelated note, just a quick thought: Py-spy is amazing and it gives me the best results out of all the Python-profilers that I've used so far. Right now, my profiling workflow for specific backend functionality is:

1. Prepare the console-command that executes py-spy (figure out the correct PID first), but don't run it yet!
2. Fire off an HTTP-request to my web application in Postman.
3. Quickly switch to the console-window and hit Enter to start py-spy.
4. Wait until the request has returned in Postman, then quickly hit Ctrl-C to stop py-spy.

If py-spy also had a programmatic API to be able to use it from Python directly, that would make the process way easier. Something along the lines of this, maybe (similar to cProfile and others):

from pyspy import Profiler

profiler = Profiler(rate=..., non_blocking=False)
profiler.start()

# Run whatever code you care about
...

profiler.stop()
profiler.write_results(format='flamegraph', output='my_profile.svg')