I’m currently deploying a kvm/qemu virtual machine with mesos/marathon. In marathon, I’m using the built in mesos command executor and running the script.
virsh start centos7.0; while true; do echo 'centos 7.0 guest is running'; sleep 5; done
Note the while loop is there only to keep the task running. My issue is that I cannot get mesos to monitor the resource usage of the virtual machine.
When marathon deploys this task on a mesos-agent, it is creating a container that uses the memory and cpu cgroups.
/sys/fs/cgroup/cpu/mesos/31b48dc3-6f09-4b5a-8964-b82d711bb895
/sys/fs/cgroup/memory/mesos/31b48dc3-6f09-4b5a-8964-b82d711bb895
When the virtual machine is being kicked off, the virsh start command is sending a request to libvirtd. Libvirtd then reads the guest.xml file located in /etc/libvirt/qemu/ and then sends a request to the qemu/kvm driver to deploy it.
In my guest.xml file I’m using a custom partition cgroup slice to monitor my virtual machine usage.
https://libvirt.org/cgroups.html
(for each cgroup)
/sys/fs/cgroup/???/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope
What I have tried.
I tried deleting my memory / cpu cgroup from this slice by doing
cgdelete -r cpu,memory:vmHolder.slice
and then adding my qemu guest process to the mesos controllers
cgclassify -g cpu,memory:mesos/31b48dc3-6f09-4b5a-8964-b82d711bb895 GUEST-PID
When I run the command cat /proc/5531/cgroup
11:perf_event:/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope
10:pids:/
9:devices:/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope
8:cpuset:/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope/emulator
7:net_prio,net_cls:/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope
6:freezer:/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope
5:blkio:/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope
4:hugetlb:/
3:cpuacct,cpu:/mesos/31b48dc3-6f09-4b5a-8964-b82d711bb895
2:memory:/mesos/31b48dc3-6f09-4b5a-8964-b82d711bb895
1:name=systemd:/vmHolder.slice/machine-qemu\x2d1\x2dcentos7.0\x2dclone.scope
It shows that I’m using those controllers, but when I run systemd-cgtop it's not adding the memory usage of the VM. I'm not sure what to do next. Any suggestions?
I am programatically launching a new instance using the Compute Engine API for Go [1], and a tool I made called vmproxy [2].
The problem I have is that if I launch a preemptible VM using a startup-script-url pointing to https://bitbucket.org/ronoaldo/debian-custom/raw/tip/tools/autobuild, the build script fails to download. I can see in the serial console output that the the startup script metadata is there, and that it attempts to be downloaded with curl, but that part fails.
However, if I reset the instance via the developers console, the script is properly downloaded and runs nicelly.
The code I am using to setup the instance is:
// Ronolinux is a VM Proxy that runs an live systems build on Compute Engine
var (
Ronolinux = &vmproxy.VM{
Path: "/",
Instance: vmproxy.Instance{
Name: "ronolinux-buildd",
Zone: "us-central1-f",
Image: vmproxy.ResourcePrefix + "/debian-cloud/global/images/debian-8-jessie-v20150915",
MachineType: "n1-standard-1",
Metadata: map[string]string{
"startup-script-url": "https://bitbucket.org/ronoaldo/debian-custom/raw/tip/tools/autobuild",
"shutdown-script": `!#/bin/bash
gsutil cp /var/log/startupscript.log gs://ronoaldo/ronolinux/build-$(date +%Y%m%d%H%M%S).log
`,
},
Scopes: []string{ storageReadWrite },
},
}
)
[1] https://godoc.org/google.golang.org/api/compute/v1
[2] https://godoc.org/ronoaldo.gopkg.net/aetools/vmproxy
If your startup script is not hosted on Cloud Storage, there is a random chance the download will fail. If you look at the serial console output, make sure to scroll horizontally, as it will not wrap long lines. In my case, the error line was very long, and this hidded the real end of the message:
(... long curl on-line progress output )
curl: (7) Failed to connect to bitbucket.org port 443: Connection timed out
(...)
Your host must respond within a 10s timeout. In my case, the first boot usually failed to contact Bitbucket, hence failing to download the script; a VM reset also made things work, as the network latency outside Google Cloud were probably better.
I ended up moving to host the script on cloud storage to avoid these issues.
I found there are still failed request when the traffic is high using command like this
haproxy -f /etc/haproxy.cfg -p /var/run/haproxy.pid -sf $(cat /var/run/haproxy.pid)
to hot reload the updated config file.
Here below is the presure testing result using webbench :
/usr/local/bin/webbench -c 10 -t 30 targetHProxyIP:1080
Webbench – Simple Web Benchmark 1.5
Copyright (c) Radim Kolar 1997-2004, GPL Open Source Software.
Benchmarking: GET targetHProxyIP:1080
10 clients, running 30 sec.
Speed=70586 pages/min, 13372974 bytes/sec.
**Requests: 35289 susceed, 4 failed.**
I run command
haproxy -f /etc/haproxy.cfg -p /var/run/haproxy.pid -sf $(cat /var/run/haproxy.pid)
several times during the pressure testing.
In the haproxy documentation, it mentioned
They will receive the SIGTTOU
611 signal to ask them to temporarily stop listening to the ports so that the new
612 process can grab them
so there is a time period that the old process is not listening on the PORT(say 80) and the new process haven’t start to listen to the PORT (say 80), and during this specific time period, it will cause the NEW connections failed, make sense?
So is there any approach that makes the configuration reload of haproxy that will not impact both existing connections and new connections?
On recent kernels where SO_REUSEPORT is finally implemented (3.9+), this dead period does not exist anymore. While a patch has been available for older kernels for something like 10 years, it's obvious that many users cannot patch their kernels. If your system is more recent, then the new process will succeed its attempt to bind() before asking the previous one to release the port, then there's a period where both processes are bound to the port instead of no process.
There is still a very tiny possibility that a connection arrived in the leaving process' queue at the moment it closes it. There is no reliable way to stop this from happening though.
I have setup gunicorn with 3 workers, 30 worker connections and using eventlet worker class. It is set up behind Nginx. After every few requests, I see this in the logs.
[ERROR] gunicorn.error: WORKER TIMEOUT (pid:23475)
None
[INFO] gunicorn.error: Booting worker with pid: 23514
Why is this happening? How can I figure out what's going wrong?
We had the same problem using Django+nginx+gunicorn. From Gunicorn documentation we have configured the graceful-timeout that made almost no difference.
After some testings, we found the solution, the parameter to configure is: timeout (And not graceful timeout). It works like a clock..
So, Do:
1) open the gunicorn configuration file
2) set the TIMEOUT to what ever you need - the value is in seconds
NUM_WORKERS=3
TIMEOUT=120
exec gunicorn ${DJANGO_WSGI_MODULE}:application \
--name $NAME \
--workers $NUM_WORKERS \
--timeout $TIMEOUT \
--log-level=debug \
--bind=127.0.0.1:9000 \
--pid=$PIDFILE
On Google Cloud
Just add --timeout 90 to entrypoint in app.yaml
entrypoint: gunicorn -b :$PORT main:app --timeout 90
Run Gunicorn with --log-level debug.
It should give you an app stack trace.
Is this endpoint taking too many time?
Maybe you are using flask without asynchronous support, so every request will block the call. To create async support without make difficult, add the gevent worker.
With gevent, a new call will spawn a new thread, and you app will be able to receive more requests
pip install gevent
gunicon .... --worker-class gevent
The Microsoft Azure official documentation for running Flask Apps on Azure App Services (Linux App) states the use of timeout as 600
gunicorn --bind=0.0.0.0 --timeout 600 application:app
https://learn.microsoft.com/en-us/azure/app-service/configure-language-python#flask-app
WORKER TIMEOUT means your application cannot response to the request in a defined amount of time. You can set this using gunicorn timeout settings. Some application need more time to response than another.
Another thing that may affect this is choosing the worker type
The default synchronous workers assume that your application is resource-bound in terms of CPU and network bandwidth. Generally this means that your application shouldn’t do anything that takes an undefined amount of time. An example of something that takes an undefined amount of time is a request to the internet. At some point the external network will fail in such a way that clients will pile up on your servers. So, in this sense, any web application which makes outgoing requests to APIs will benefit from an asynchronous worker.
When I got the same problem as yours (I was trying to deploy my application using Docker Swarm), I've tried to increase the timeout and using another type of worker class. But all failed.
And then I suddenly realised I was limitting my resource too low for the service inside my compose file. This is the thing slowed down the application in my case
deploy:
replicas: 5
resources:
limits:
cpus: "0.1"
memory: 50M
restart_policy:
condition: on-failure
So I suggest you to check what thing slowing down your application in the first place
Could it be this?
http://docs.gunicorn.org/en/latest/settings.html#timeout
Other possibilities could be your response is taking too long or is stuck waiting.
This worked for me:
gunicorn app:app -b :8080 --timeout 120 --workers=3 --threads=3 --worker-connections=1000
If you have eventlet add:
--worker-class=eventlet
If you have gevent add:
--worker-class=gevent
I've got the same problem in Docker.
In Docker I keep trained LightGBM model + Flask serving requests. As HTTP server I used gunicorn 19.9.0. When I run my code locally on my Mac laptop everything worked just perfect, but when I ran the app in Docker my POST JSON requests were freezing for some time, then gunicorn worker had been failing with [CRITICAL] WORKER TIMEOUT exception.
I tried tons of different approaches, but the only one solved my issue was adding worker_class=gthread.
Here is my complete config:
import multiprocessing
workers = multiprocessing.cpu_count() * 2 + 1
accesslog = "-" # STDOUT
access_log_format = '%(h)s %(l)s %(u)s %(t)s "%(r)s" %(s)s %(b)s "%(q)s" "%(D)s"'
bind = "0.0.0.0:5000"
keepalive = 120
timeout = 120
worker_class = "gthread"
threads = 3
I had very similar problem, I also tried using "runserver" to see if I could find anything but all I had was a message Killed
So I thought it could be resource problem, and I went ahead to give more RAM to the instance, and it worked.
You need to used an other worker type class an async one like gevent or tornado see this for more explanation :
First explantion :
You may also want to install Eventlet or Gevent if you expect that your application code may need to pause for extended periods of time during request processing
Second one :
The default synchronous workers assume that your application is resource bound in terms of CPU and network bandwidth. Generally this means that your application shouldn’t do anything that takes an undefined amount of time. For instance, a request to the internet meets this criteria. At some point the external network will fail in such a way that clients will pile up on your servers.
If you are using GCP then you have to set workers per instance type.
Link to GCP best practices https://cloud.google.com/appengine/docs/standard/python3/runtime
timeout is a key parameter to this problem.
however it's not suit for me.
i found there is not gunicorn timeout error when i set workers=1.
when i look though my code, i found some socket connect (socket.send & socket.recv) in server init.
socket.recv will block my code and that's why it always timeout when workers>1
hope to give some ideas to the people who have some problem with me
For me, the solution was to add --timeout 90 to my entrypoint, but it wasn't working because I had TWO entrypoints defined, one in app.yaml, and another in my Dockerfile. I deleted the unused entrypoint and added --timeout 90 in the other.
For me, it was because I forgot to setup firewall rule on database server for my Django.
Frank's answer pointed me in the right direction. I have a Digital Ocean droplet accessing a managed Digital Ocean Postgresql database. All I needed to do was add my droplet to the database's "Trusted Sources".
(click on database in DO console, then click on settings. Edit Trusted Sources and select droplet name (click in editable area and it will be suggested to you)).
Check that your workers are not killed by a health check. A long request may block the health check request, and the worker gets killed by your platform because the platform thinks that the worker is unresponsive.
E.g. if you have a 25-second-long request, and a liveness check is configured to hit a different endpoint in the same service every 10 seconds, time out in 1 second, and retry 3 times, this gives 10+1*3 ~ 13 seconds, and you can see that it would trigger some times but not always.
The solution, if this is your case, is to reconfigure your liveness check (or whatever health check mechanism your platform uses) so it can wait until your typical request finishes. Or allow for more threads - something that makes sure that the health check is not blocked for long enough to trigger worker kill.
You can see that adding more workers may help with (or hide) the problem.
The easiest way that worked for me is to create a new config.py file in the same folder where your app.py exists and to put inside it the timeout and all your desired special configuration:
timeout = 999
Then just run the server while pointing to this configuration file
gunicorn -c config.py --bind 0.0.0.0:5000 wsgi:app
note that for this statement to work you need wsgi.py also in the same directory having the following
from myproject import app
if __name__ == "__main__":
app.run()
Cheers!
Apart from the gunicorn timeout settings which are already suggested, since you are using nginx in front, you can check if these 2 parameters works, proxy_connect_timeout and proxy_read_timeout which are by default 60 seconds. Can set them like this in your nginx configuration file as,
proxy_connect_timeout 120s;
proxy_read_timeout 120s;
In my case I came across this issue when sending larger(10MB) files to my server. My development server(app.run()) received them no problem but gunicorn could not handle them.
for people who come to the same problem I did. My solution was to send it in chunks like this:
ref / html example, separate large files ref
def upload_to_server():
upload_file_path = location
def read_in_chunks(file_object, chunk_size=524288):
"""Lazy function (generator) to read a file piece by piece.
Default chunk size: 1k."""
while True:
data = file_object.read(chunk_size)
if not data:
break
yield data
with open(upload_file_path, 'rb') as f:
for piece in read_in_chunks(f):
r = requests.post(
url + '/api/set-doc/stream' + '/' + server_file_name,
files={name: piece},
headers={'key': key, 'allow_all': 'true'})
my flask server:
#app.route('/api/set-doc/stream/<name>', methods=['GET', 'POST'])
def api_set_file_streamed(name):
folder = escape(name) # secure_filename(escape(name))
if 'key' in request.headers:
if request.headers['key'] != key:
return 404
else:
return 404
for fn in request.files:
file = request.files[fn]
if fn == '':
print('no file name')
flash('No selected file')
return 'fail'
if file and allowed_file(file.filename):
file_dir_path = os.path.join(app.config['UPLOAD_FOLDER'], folder)
if not os.path.exists(file_dir_path):
os.makedirs(file_dir_path)
file_path = os.path.join(file_dir_path, secure_filename(file.filename))
with open(file_path, 'ab') as f:
f.write(file.read())
return 'sucess'
return 404
in case you have changed the name of the django project you should also go to
cd /etc/systemd/system/
then
sudo nano gunicorn.service
then verify that at the end of the bind line the application name has been changed to the new application name