[Python] How to profile a FastAPI application running with Hypercorn (handling forked...

I am trying to profile my Python module to find performance bottlenecks.

The module starts a Hypercorn server, on top of which an asynchronous FastAPI application runs.

Here’s a rough idea of how I launch it in my main.py:

hypercorn_config = Config()
hypercorn_config.bind = "localhost:8000"
hypercorn_config.workers = 1 #I tried for both 1 and multiple workers
hypercorn_config.application_path = f"service:create_app({config_str})"
run(hypercorn_config)


The create_app() method creates and returns the FastAPI app instance.

Problem I tried several profilers:

• 
  yappi
  
  
• 
  scalene
  
  
• 
  cProfile
  
  
• 
  pyinstrument
  
  
• 
  py-spy
  
  
• 
  memray (for memory profiling)
  

But I am running into problems:

1. 
   Hypercorn forks the process, so many tools (like yappi, scalene) can't correctly track the child processes.
   
   
2. 
   pyinstrument can only generate one profile per request, but I need an aggregate profile across multiple requests.
   
   
3. 
   py-spy can attach to the process but suffers from undersampling in async-heavy workloads.
   
   
4. 
   memray also struggles to track memory usage correctly across forked workers.
   

How can I properly profile (both CPU and memory) my FastAPI app running under Hypercorn, accounting for the forking behavior? Ideally, I want to capture an end-to-end performance profile over multiple requests. Is there any standard tools, tricks, or best practices to go about it?

Continue reading...