cpu blocked for ~10s before first request on complex Flask app with gevent on 8.10.1 release

[ddorian]

Hi,

I have a pretty complex flask app locally that hangs for about ~10s ONLY before the first request when using 8.10.1 release (doesn't happen on 8.10.0). Still happens on 9.0.0 though. The CPU usage is 100% for those seconds. All the other requests work normally after.

How do you want me to profile it? Using ubuntu 22.04 & Python 3.11 and many libraries.

[ddorian]

The first request takes 0.6s on 8.10.0, while it takes ~8.8s on 8.10.1. (makes PostgreSQL connections, compiles jinja2 etc)

The subsequent requests take 10ms.

I tried with Pycharm profiler but it only showed "gevent-wait" time, and couldn't find it.

mytest.pstat.zip

Tried with flask-debugtoolbar but it only profiles inside the request (it's ~100ms for both versions).

So, even on 8.10.0, the first request takes 100ms on flask-side + 500ms on new-relic side.

Which should also be unacceptable IMHO. (see screenshot)

I'm running newrelic with manually like this:

import gevent.monkey

# had problems monkey-patching inside uwsgi so doing it here
gevent.monkey.patch_all()

import os
import sys
from pathlib import Path

# start newrelic manually, so we get no errors from gevent
# https://discuss.newrelic.com/t/found-problems-running-newrelic-admin-with-python-3-6-gevent-and-gunicorn/70583/2

newrelic_path = Path("deploy/files/newrelic.ini")
environment = os.environ.get("FLASK_ENV")
if newrelic_path.exists():
    import newrelic.agent

    print("starting newrelic!")
    newrelic.agent.initialize(newrelic_path.as_posix(), environment=environment)

[mthipparthi]

We have the same issue. It has killed all our applications and performance. High time it is resolved. BTW we are customer.

[TimPansino]

Can we get some more info on this issue? A pip freeze would be very helpful if you're comfortable providing it. Or at the very least gevent and greenlet versions used.

This may be related to setting startup_timeout. We do not recommend setting startup_timeout for web applications, as this blocks completion of requests to the application for the initial request(s) until registration with the backend can complete (up to the time set, in seconds). Not setting it or setting to 0 means that initial requests will not be monitored while the agent starts on a background thread. Once it completes its connect cycle, it will begin monitoring requests.

I was unable to reproduce the issue on a machine running the following app. I saw no real difference in startup time. After examining the changes between 8.10.0 and 8.10.1 the only interesting change I see is a change to prefer importlib.metadata calls over builtin version attributes on packages for environment reporting. This could have potential performance implications, but I don't see any in my initial testing. I'd like to ensure my environment is identical before I do any larger scale performance profiling sessions.

newrelic (8.10.0 vs 8.10.0)
Flask==2.3.3
gevent==23.9.1
greenlet==3.0.0rc3

# Patch with gevent
from gevent import monkey
monkey.patch_all()

# Initialize agent
import newrelic.agent
newrelic.agent.initialize("newrelic.ini")

# Create application
from gevent.pywsgi import WSGIServer
from flask import Flask

app = Flask(__name__)

@app.route('/')
def hello():
    return "Hello from Flask!"

if __name__ == '__main__':
    # Start server
    print("Starting gevent server...")
    server = WSGIServer(("0.0.0.0", 8000), app)
    server.serve_forever()

[ddorian]

@TimPansino another alternative would be to tell me how you want me to profile it (what tool etc) and I'll do it. Might be faster than doing a lot of tests locally. Example, I can't replicate it with the minimal code above.

I saw no real difference in startup time.

The delay is ONLY before the FIRST flask request.

Attached is the pip freeze.

freeze-requirements.txt

[mthipparthi]

we run using - newrelic-admin run-program gunicorn --workers 3 --timeout 65 --graceful-timeout 65 --max-requests 2000 --max-requests-jitter 200 --bind 0.0.0.0:8088 app.wsgi. same issue. Not only for gevent, even for sync workers.

[natemoser]

We're seeing a similar jump in service time, but chiefly manifesting in our interactions with Kafka. We first noticed it when going from newrelic==8.8.0 to 9.0.0, but we see the same behavior when going from 8.10.0 to 8.10.1 (8.8.0 and 8.10.0 are at parity). The service time difference can be seen in New Relic, but is confirmed by our application logs.

• Heroku stack-22
• python-3.10.13
• Django==3.2.20
• gunicorn==20.1.0
• kafka-helper==0.2
• kafka-python==2.0.2

This shows the drop in Kafka message message consuming transaction time when rolling back from newrelic==8.10.1 to newrelic=8.8.0

Ironically but very helpfully, it was New Relic profiling that showed the increase in importlib.metadata usage, which pointed us toward the newrelic lib update:

v8.10.0

v8.10.1

[ddorian]

cc @umaannamalai

[hmstepanek]

Hey all we have a fix coming for this to be released on Monday Nov 13th.

[hmstepanek]

This fix was released in v9.1.2.