ASGI Server-Timing middleware

An ASGI middleware that wraps the excellent yappi profiler to let you measure the execution time of any function or coroutine in the context of an HTTP request, and return it as a standard Server-Timing HTTP header.

Sample configurations

Here are some example configurations for various frameworks and libraries. Feel free to combine them as needed.

FastAPI

fastapi_app.add_middleware(ServerTimingMiddleware, calls_to_track={
	"1deps": (fastapi.routing.solve_dependencies,),
	"2main": (fastapi.routing.run_endpoint_function,),
	"3valid": (pydantic.fields.ModelField.validate,),
	"4encode": (fastapi.encoders.jsonable_encoder,),
	"5render": (
		fastapi.responses.JSONResponse.render,
		fastapi.responses.ORJSONResponse.render,
	),
})

Starlette

from starlette.middleware import Middleware

middleware = [
  Middleware(ServerTimingMiddleware, calls_to_track={
	  # TODO: ...
  }),
]

starlette_app = Starlette(routes=routes, middleware=middleware)

SQLAlchemy

fastapi_app.add_middleware(ServerTimingMiddleware, calls_to_track={
	"db_exec": (sqlalchemy.engine.base.Engine.execute,),
	"db_fetch": (
		sqlalchemy.engine.ResultProxy.fetchone,
		sqlalchemy.engine.ResultProxy.fetchmany,
		sqlalchemy.engine.ResultProxy.fetchall,
	),
})

More Frameworks

Feel free to submit PRs containing examples for more libraries and ASGI frameworks.

Caveats

• Only the end-to-end time is reported for both functions and coroutines, so it's not possible to tell from the metrics when a coroutine took a long time because the event loop thread got stalled (though asyncio's debug mode can help).
• The profiler's memory is not freed over time, and only gets cleared when it exceeds a given threshold (50MB by default). When memory gets cleared, data collected for all ongoing requests is lost, so the timing for these will be incorrect.
• Executing the same task multiple times in parallel (such as with asyncio.gather()) will report the duration as if they had been executed sequentially.
• The minimum version of Python supported is 3.7, since this middleware makes use of PEP 567 context variables to track which function call belongs to which request, and the Python 3.6 backport doesn't have asyncio support.

Special Thanks

• Sümer Cip (@sumerc), for creating and maintaininng yappi, as well as being very responsive and open to adding all the new features needed to make this work.
• David Montague (@dmontagu) for his involvement in shaping this middleware at every step of the way.