Introduction
FastAPI is a well-known and beloved (almost 70k⭐️ on GitHub) modern, fast, asynchronous web framework for building APIs with Python. It is designed for ease of use and high performance, while being robust and production-ready.
My team quite enjoyed the process of developing apps with FastAPI, finding it a rewarding learning experience. However, one learning was a bit hard: when it came to performance testing, we faced unexpectedly low Requests Per Second (RPS) values, which prompted a deeper investigation, that in turn lead us to learn more about FastAPI Middlewares.
The purpose of this note is to explore the impact of BaseHTTPMiddleware and ASGIMiddleware on FastAPI application
performance and to share our learnings.
Spoiler: this is how the performance of real-life production application may change depending on the type of middlewares
we use to implement some common logic:
Background
In FastAPI, middleware plays a crucial role in the request-response lifecycle, acting as a layer that processes requests and responses before they reach the core application logic, or after they leave it. This functionality is pivotal for implementing various features like authentication, data processing, request logging, and error handling.
Whenever you want to apply some common logic to request processing in more than one endpoint, you might consider using middlewares. FastAPI (more precisely - Starlette on which FastAPI is grounded) offers two types of middleware:
BaseHTTPMiddleware: This type of middleware is designed around a simple and straightforward interface. When using BaseHTTPMiddleware, you define an async function that takes the request, calls the next item in the middleware chain, and then you can modify the response.
ASGIMiddleware: ASGI (Asynchronous Server Gateway
Interface) specification states:
It is possible to have ASGI “middleware” - code that plays the role of both server and application, taking in a scope and the send/receive awaitable callables, potentially modifying them,and then calling an inner application.ASGIMiddleware functions within ASGI application, which is a core components of frameworks like FastAPI and Starlette. Moreover, such middleware shares the samecoroutine application(scope, receive, send)signature with ASGI applications. Generally, implementing and supporting ASGIMiddleware is more complex than BaseHTTPMiddleware due to its lower level of abstraction, offering more direct interaction with the ASGI protocol. It's important to note that ASGIMiddleware is not subject to some of the important limitations of BaseHTTPMiddleware.
As far as middlewares act as a layer in request-response lifecycle, they are expected to have an impact on application performance. The question is - how much? And what is the difference between two types of middlewares supported within FastAPI?
The Discovery Process
During performance load testing with ApacheBench (ab) we found that although the application can handle the required RPS it still seems quite low for an async application which does not contain any significant CPU-bound calls. Another point of worry, RPS didn't scale linearly with the number of application instances. We decided to investigate the issue further.
After intensive research, we came down to reviewing our middlewares and found performance caveats that I will illustrate with experiments below.
Experimentation and Analysis
Here are five experiments that nicely represent the investigation process our team went through. I will be running ApacheBench (ab) requests against each experimental case with concurrency 100 to measure RPS of our sample application with different middlewares. I will run each case five times and average the results.
>> ab -c 100 -n 1000 http://127.0.0.1:8000/
Experiment Setup
Case 0: Baseline (No Middlewares)
Let's start with a simple FastAPI application that contains a single endpoint and no middlewares.
from fastapi import FastAPI
app = FastAPI()
@app.get("/")
def read_root():
return {"Hello": "World"}
Case 1: Single Simple Middleware
Here are the implementations of both types of middlewares. Both are effectively doing nothing.
# BaseHTTPMiddleware
from urllib.request import Request
from starlette.middleware.base import BaseHTTPMiddleware
class HTTPMiddleware:
async def __call__(self, request: Request, call_next):
return await call_next(request)
app.add_middleware(BaseHTTPMiddleware, dispatch=HTTPMiddleware())
# ASGIMiddleware
from starlette.types import ASGIApp, Send, Scope, Receive
class ASGIMiddleware:
def __init__(self, app: ASGIApp):
self.app = app
async def __call__(self, scope: Scope, receive: Receive, send: Send):
await self.app(scope, receive, send)
app.add_middleware(ASGIMiddleware)
Case 2: Five Stacked Simple Middlewares
Middlewares are still doing nothing, but let's stack them up to find the impact.
# BaseHTTPMiddleware
for _ in range(5):
app.add_middleware(BaseHTTPMiddleware, dispatch=HTTPMiddleware())
# ASGIMiddleware
for _ in range(5):
app.add_middleware(ASGIMiddleware)
Case 3: Five Stacked Blocking Middlewares
So far our tests were very synthetic. Let's try to make them more realistic by adding some logic to our middlewares. In asynchronous application it's very important to avoid synchronous calls that block the entire event loop and may have dramatic performance impact, but some blocking calls are unavoidable - like logging (off-topic: offloading logging to a separate thread may help minimize impact).
Let's add some massive blocking calls to our middlewares.
# BaseHTTPMiddleware
class HTTPMiddleware:
async def __call__(self, request: Request, call_next):
with open('/tmp/middleware.log', 'a') as f:
f.write('HTTPMiddleware\n' * 1000)
return await call_next(request)
# ASGIMiddleware
class ASGIMiddleware:
def __init__(self, app: ASGIApp):
self.app = app
async def __call__(self, scope: Scope, receive: Receive, send: Send):
with open('/tmp/middleware.log', 'a') as f:
f.write('ASGIMiddleware\n' * 1000)
await self.app(scope, receive, send)
Case 4: Five Stacked Non-Blocking Middlewares
Let's see whether adding async calls to our middlewares have different impact. Here I'm including an asyncio.sleep for 10ms call to our logic.
# BaseHTTPMiddleware
class HTTPMiddleware:
async def __call__(self, request: Request, call_next):
await asyncio.sleep(0.01)
return await call_next(request)
# ASGIMiddleware
class ASGIMiddleware:
def __init__(self, app: ASGIApp):
self.app = app
async def __call__(self, scope: Scope, receive: Receive, send: Send):
await asyncio.sleep(0.01)
await self.app(scope, receive, send)
Case 5: Five Stacked Non-Blocking Middlewares with 100ms sleep
Same as previous case but with increased sleep time to 100ms - take it as a representation of a few async database or services discovery calls.
Results and Observations
Here is the summary of the results from the experiments above.
Case 1: Single Simple Middleware (doing nothing)
- Impact from adding a single middleware is pretty visible, telling us there is a noticeable overhead of running a middleware.
- BaseHTTPMiddleware is ~15% slower.
Case 2: Five Stacked Simple Middlewares (doing nothing)
- The impact of executing five BaseHTTPMiddlewares is significant. RPS drops by an order of magnitude, comparing to Baseline.
- ASGIMiddleware virtually not affected by stacking up.
Case 3: Five Stacked Blocking Middlewares (writing to a file)
- Heavy blocking call is not adding much impact to BaseHTTPMiddleware performance, telling us that response time is probably already saturated by middleware overhead.
- Impact on ASGIMiddleware is significant, which is expected.
Case 4: Five Stacked Non-Blocking Middlewares (asyncio.sleep(0.01))
- BaseHTTPMiddleware with
sleepperforms just a tiny bit slower than in case 2 where middleware is doing nothing, this somewhat supports the idea that we maxed-out middleware overhead impact. - ASGIMiddleware still faster but to a lesser extent.
Case 5: Five Stacked Non-Blocking Middlewares with 100ms sleep
- Finally, middlewares of different types have gotten on par with each other.
Findings and Recommendations
Based on experiments above BaseHTTPMiddleware has a measurable performance overhead comparing to ASGIMiddleware. The shorter the processing time of a middleware logic, the more significant the difference in performance between the two types of middlewares will be.
It looks, this overhead comes from anyio task management routines used by Starlette to handle middlewares.
BaseHTTPMiddleware class provides a nice and simple abstraction to user. By giving the immediate access to the Request it makes development and support of middlewares effortless and cost-effective. At the same time, abstraction comes at a cost of overhead (task execution, exception and stream management through anyio library).
ASGIMiddleware in turn is way more involving to implement and maintain. Actually it forces you drop down abstraction to the level of ASGI protocol. It steepens the learning curve and raises the requirements to testing and quality assurance, but may bring reasonable performance benefit.
Here is how a real-life impact of migration from BaseHTTPMiddleware to ASGIMiddleware looks like (optimized middlewares are marked with arrows) - 20%-30% improvement in request processing time in a single endpoint.
Even the performance benefit looks impressive, as it happens often - it's a tradeoff with complexity. Unless you heavily relay on multiple stacked middlewares for you business logic, I would not suggest to use ASGIMiddleware before RPS and/or request processing time requirements will force you to. But it's important to keep this option under the belt and use it when other more obvious optimizations opportunities are exhausted.
Conclusion
FastAPI middlewares are a powerful tool for implementing various features like authentication, CORS handling, logging, adding custom metrics and many more. However, overhead of middleware execution may be significant and should be taken into account while sizing an application. I strongly suggest avoiding the premature optimization and keeping the design as simple as possible which helps to increase Time to Market and reduce maintenance costs. While, if requirements force you to optimize - migrating to ASGIMiddleware might be a good option to consider.
References and Further Reading
- FastAPI documentation on middleware: https://fastapi.tiangolo.com/tutorial/middleware/ and ASGI Middlewares https://fastapi.tiangolo.com/advanced/middleware/
- Starlette documentation on middleware: https://www.starlette.io/middleware/
- Discussion at FastAPI GitHub: https://github.com/tiangolo/fastapi/issues/2696
- Another experiment on the topic https://kisspeter.github.io/fastapi-performance-optimization/middleware
Top comments (0)