Closed lsmith77 closed 1 month ago
BTW I saw this ticket here https://github.com/tiangolo/uvicorn-gunicorn-fastapi-docker/issues/47 but I think its not the same issue.
Facing the same issue. Whenever following code is executed with incorrect smtp_url, port, my worker crashes:
def validate_smtp(smtp_url: str):
try:
smtp = SMTP()
smtp.connect(smtp_url)
smtp.quit()
return True
except:
return False
There is no crash if smtp_url or port is valid. Dependencies:
I'm also facing the same issue. any workarounds?
i resolved this issue by adding worker timeout while initiating my gunicorn application.
gunicorn -k uvicorn.workers.UvicornWorker ${APP_MODULE} --bind 0.0.0.0:80 --timeout ${WORKER_TIMEOUT}
Facing the same issue when running long processes on websockets and it ends up terminating the websocket connection. Any fixes?
Facing the same issue when I use the haystack. I modifed the docker-compose.yml as following: command: "/bin/bash -c 'sleep 10 && gunicorn rest_api.application:app -b 0.0.0.0 -k uvicorn.workers.UvicornWorker -- workers 1 --timeout 600'" It can work.
Facing this issue while using docker. Working perfectly fine if run directly with gunicorn -w 1 -k uvicorn.workers.UvicornWorker --bind 0.0.0.0:8080 main:app
.
None of the following suggested solutions worked:
gevent
3.7
from 3.9
timeout
uvicorn
without gunicorn
Can someone please point me in the right direction to resolve this issue?
Facing this same issue (on both CentOS and Ubuntu VM's), it happens during typical load and all resources are not near limits.
Same here, anyone can suggest a good alternative?
In my case it seemed to happen due to request timeout to external service.
Any solution for this ? Facing the same issue when calling an endpoint which takes 1-2 min to execute.
@atTheShikhar for me - I switched to combination Flask + uWSGI.
@atTheShikhar for me - I switched to combination Flask + uWSGI.
Unfortunately i cannot change the SGI and framework, since most of the work are already done in my case. Just needed this one endpoint to work.
Can you show your launching command? what parameters do you use? I assume gunicorn should run smoothly when using 1 worker on a single process.
Can you show your launching command? what parameters do you use? I assume gunicorn should run smoothly when using 1 worker on a single process.
I am using docker so the final running command is this (the timeout part was added after reading above discussion)
CMD ["gunicorn", "-w", "4", "-k", "uvicorn.workers.UvicornWorker", "main:app", "--bind", "0.0.0.0:80", "--timeout", "300"]
Btw, this runs just fine locally. Problem only happens after i deploy it on GCR.
hmm, for me it struggles both on EC2 machine and on Fargate... but so did you try running 1 worker? just to make work,
yup i tried with 1 worker, still no luck.
anyone have any update?
I've had the same issue. BUT, it only comes up when I started using max_requests setting...
Adding some info if it helps. Gunicorn config below is run via supervisor
, and was fine for a while. Added FastAPI Cache, all was good as well but crash rate has increased dramatically in past few days.
bind = "0.0.0.0:<PORT>"
wsgi_app = "main:app"
workers = 3 # worked for a while using 1 worker
worker_class = "uvicorn.workers.UvicornWorker"
errorlog = '<LOG_FILE>'
accesslog = '<LOG_FILE>'
loglevel = 'debug'
timeout = 240 # been increasing from 30s to solve [CRITICAL] WORKER TIMEOUT, now at 240s and still crashes occassionally
Server RAM: 1.9GB
Thanks
Managed to resolve this issue, sharing in case this helps.
Our issue originated from making external API calls from within an async endpoint. These API calls did not support async, which introduced blocking calls to the event loop, resulting in the uvicorn worker timing out. Our reliance on FastAPI Cache decorators for these async endpoints prevented us from simply redefining these endpoints as sync (async def
-> def
).
To resolve, we made use of the run_in_threadpool()
utility function to ensure these sync calls are run in a separate threadpool, outside the event loop. Alongside this, we updated our gunicorn config so the workers
and threads
count was equal - setting these to 4.
from fastapi.concurrency import run_in_threadpool
@api.get('/handler')
async def handler():
...
# Slow async function
await my_async_function()
....
# Slow running sync function
await run_in_threadpool(sync_function)
We released this update over 2 weeks ago and haven't seen any worker timeouts. Hopefully this helps 🙂
@nicholasmccrea wow, should have been hard to identify it! Great news!
@nicholarmccrea your solution seems like a good solution, but i imagined removing the "async" from the endpoint functions allowed the requests to be handled in a different threadpool. Do you need to explicitly run it in a seperate function?
I am taking the knowledge from the comment on this post : https://stackoverflow.com/questions/71516140/fastapi-runs-api-calls-in-serial-instead-of-parallel-fashion
@nicholarmccrea your solution seems like a good solution, but i imagined removing the "async" from the endpoint functions allowed the requests to be handled in a different threadpool. Do you need to explicitly run it in a seperate function?
I am taking the knowledge from the comment on this post : https://stackoverflow.com/questions/71516140/fastapi-runs-api-calls-in-serial-instead-of-parallel-fashion
@mcazim98 in our case we were not able to redefine our endpoints as sync due to our reliance on FastAPI cache decorators. The FastAPI cache version we were using was 0.1.8
, which did not support sync functions, therefore we needed to use the run_in_threadpool
utility function as a workaround.
Thankfully, FastAPI cache now support sync functions as of version 0.2.0
, which means we can now redefine our endpoints as sync and move away from using the run_in_threadpool
function 🙂
Yes I was having similar issues where endpoint was requesting a response from an ML model running within the saem fastapi application. Redefining the endpoints to sync versions fixed the issue. Since this was a CPU bound task I do not think async was really necessary here, but I still do not understand why it was causing the workers to constantly crash although I experimented with different timeout settings.
I encountered a similar issue, running FastAPI backend, where endpoint seemed to randomly ignore my requests sometimes, but sometimes 'fixes itself' and works. After spending 2 weeks on this, I realized this had to do with ports. My fix was to:
-p 8000:8000
in the docker run command. docker run -dit --gpus all -p 8000:8000 my_image
--host 0.0.0.0
to my uvicorn command. uvicorn src:app --host 0.0.0.0
I experienced a similar issue with the following (slightly different) symptoms:
backend-1 | [2024-08-02 14:46:36 +0000] [1] [ERROR] Worker (pid:648) was sent SIGABRT!
backend-1 | [2024-08-02 14:46:36 +0000] [692] [INFO] Booting worker with pid: 692
backend-1 | [2024-08-02 14:46:37 +0000] [692] [INFO] Started server process [692]
backend-1 | [2024-08-02 14:46:37 +0000] [692] [INFO] Waiting for application startup.
backend-1 | [2024-08-02 14:46:37 +0000] [692] [INFO] Application startup complete.
backend-1 | 172.19.0.1:48566 - "GET /v1/projects HTTP/1.1" 200
backend-1 | [2024-08-02 14:46:38 +0000] [1] [CRITICAL] WORKER TIMEOUT (pid:670)
storage-1 | 2024-08-02 14:46:38.790 UTC [106] LOG: could not send data to client: Broken pipe
storage-1 | 2024-08-02 14:46:38.790 UTC [106] STATEMENT: SELECT ...
storage-1 | 2024-08-02 14:46:38.791 UTC [106] FATAL: connection to client lost
storage-1 | 2024-08-02 14:46:38.791 UTC [106] STATEMENT: SELECT ...
I was able to resolve the issue by updating my connection pool to utilize asynchronous connections. Here's the difference between the two connection pools:
# ./core/database.py
from sqlalchemy.orm import sessionmaker
from sqlalchemy import create_engine
from sqlalchemy.ext.asyncio import create_async_engine, AsyncSession
from core.settings import dbconn_config
from sqlalchemy.exc import OperationalError
from fastapi import HTTPException, status
class UnreachableDatabase(Exception):
def __init__(self, message="The database is currently unreachable. Please try again later."):
self.message = message
super().__init__(self.message)
# Database session
SQLALCHEMY_DATABASE_URL = 'postgresql://{username}:{password}@{hostname}:{port}/{database}'.format(**dbconn_config)
engine = create_engine(SQLALCHEMY_DATABASE_URL)
SessionLocal = sessionmaker(autocommit=False, autoflush=False, bind=engine)
# Database async session
SQLALCHEMY_DATABASE_ASYNC_URL = 'postgresql+asyncpg://{username}:{password}@{hostname}:{port}/{database}'.format(**dbconn_config)
async_engine = create_async_engine(SQLALCHEMY_DATABASE_ASYNC_URL)
AsyncSessionLocal = sessionmaker(
bind=async_engine,
class_=AsyncSession,
expire_on_commit=False
)
# Database dependency yielder
def get_db():
db = SessionLocal()
try:
yield db
except OperationalError as e:
raise UnreachableDatabase() from e
finally:
db.close()
# Database async dependency yielder
async def get_async_db():
db = AsyncSessionLocal()
try:
yield db
except OperationalError as e:
raise UnreachableDatabase() from e
finally:
await db.close()
In use, it looks something like this:
# ./v1/endpoints/space.py
from core.models.space import Space
from core.schemas.space import SpaceResponse
from core.database import get_async_db
from fastapi import Depends, APIRouter
from sqlalchemy.ext.asyncio import AsyncSession
from sqlalchemy.future import select
router = APIRouter()
@router.get("/space", response_model=SpaceResponse)
async def pull(space_id: int, limit: int, db: AsyncSession = Depends(get_async_db)):
# Initialize query with the primary filter
stmt = (
select(Space)
.filter(Space.id > space_id)
.order_by(Space.id, Space.updated_at)
.limit(limit)
)
# Collect results
result = await db.execute(stmt)
documents = result.scalars().all()
return {"documents": documents}
This change resolved my issue.
I am struggling to know which layer is the root cause here.
My app runs fine, but then suddenly it is unable to serve requests for a while and then "fixes itself". While it's unable to serve requests my logs show:
Initially, I thought it was related to load and resource limits, but it seems to also happen during "typical load" and when resources are nowhere near their limits.