Intermittent httpx.ReadError with high concurrency

[mmeendez8]

Issue Description

I've encountered intermittent httpx.ReadError exceptions when using high concurrency values with httpx.AsyncClient. This issue does not occur when using aiohttp with similar concurrency levels.

Reproduction Steps

1. Set up a simple FastAPI server with a 3-second delay (code provided below).
2. Run the httpx client script (provided below) with high concurrency (CONCURRENT_REQUESTS = 300, TOTAL_REQUESTS = 1000).
3. Observe intermittent httpx.ReadError exceptions.
4. Replace the httpx client with the aiohttp client (code provided) and run with the same concurrency settings.
5. Observe that the aiohttp client completes without errors.

Environment

aiohttp                           3.9.5
httpcore                          1.0.5
httpx                             0.27.0

Server code

from fastapi import FastAPI
from pydantic import BaseModel, Field
import asyncio
import uvicorn

app = FastAPI()

class ImagePayload(BaseModel):
    image: str = Field(..., description="Base64 encoded image")

@app.post("/process_image")
async def process_image(payload: ImagePayload):
    await asyncio.sleep(3)

    return {
        "message": "Image processed successfully after a delay"
    }

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)

Httpx client code

import asyncio
import httpx
import base64
import numpy as np

SERVER_URL = "http://localhost:8000/process_image"
CONCURRENT_REQUESTS = 300
TOTAL_REQUESTS = 1000
TIMEOUT = 30

async def send_request(client, image_data):
    response = await client.post(SERVER_URL, json={"image": image_data})
    return response.json()

async def main():
    image = np.random.randint(0, 255, (100, 100, 3), dtype=np.uint8)
    image_data = base64.b64encode(image).decode()
    
    async with httpx.AsyncClient(limits=httpx.Limits(max_connections=CONCURRENT_REQUESTS), verify=False, timeout=TIMEOUT) as client:
        tasks = [send_request(client, image_data) for _ in range(TOTAL_REQUESTS)]
        results = await asyncio.gather(*tasks)

    for i, result in enumerate(results, 1):
        print(f"Response {i}:", result)

if __name__ == "__main__":
    asyncio.run(main())

AIOHttp client code

import asyncio
import aiohttp
import base64
import numpy as np

SERVER_URL = "http://localhost:8000/process_image"
CONCURRENT_REQUESTS = 300
TOTAL_REQUESTS = 1000
TIMEOUT = 30

async def send_request(session, image_data):
    async with session.post(SERVER_URL, json={"image": image_data}) as response:
        return await response.json()

async def main():
    image = np.random.randint(0, 255, (100, 100, 3), dtype=np.uint8)
    image_data = base64.b64encode(image.tobytes()).decode()
    
    timeout = aiohttp.ClientTimeout(total=TIMEOUT)
    connector = aiohttp.TCPConnector(limit=CONCURRENT_REQUESTS)
    
    async with aiohttp.ClientSession(timeout=timeout, connector=connector) as session:
        tasks = [send_request(session, image_data) for _ in range(TOTAL_REQUESTS)]
        
        results = await asyncio.gather(*tasks)

    for i, result in enumerate(results, 1):
        print(f"Response {i}:", result)

if __name__ == "__main__":
    asyncio.run(main())

[mhdzumair]

+1 i observed similar random httpx.ReadError

[1138559250]

+1 i observed similar random httpx.ReadError

[adriangb]

Also have run into this. Same scenario has no issues with requests. It happens both with sync and async httpx.

[conradogarciaberrotaran]

+1, this plus aiohttp better performance made me migrate from httpx to aiohttp.
Also, this issue is solved if using aiohttp's transport layer: #3215 (comment)