btw, are there any examples of exporting metrics using Python client? I could only find last_metrics attribute of the task

running
docker network prunebefore starting the containers kind of helped. I still see an error when I'm comparing > 20 experiments, but at least trains works okay after that, and there are no connection pool limit errors in the logs

Yeah, connections keep getting dropped since the pool is full

This shouldn't be an issue, since the server should reuse connections, but perhaps the max connections limit on your server os relatively low?

network logs

OK, on first glance ES doesn't seem to have any issue

I'll think about it and get back to you - it might be interesting to understand what causes this when comparing >20 experiments...

Can you please check the max connections setting in your OS?

okay, give me a sec

we do log a lot of the different metrics, maybe this can be part of the problem

Hi DilapidatedDucks58 , I am trying to reproduce the "Connection is full warning". Do you override any apiserver environment variables is docker compose? If yes then can you share your version of docker-compose? Do you provide a configuration file for gunicorn? Can you please share it?

some of the POST requests "tasks.get_all_ex" fail as far as I can see

Error
Failed to get Scalar Charts

Is it possible to get a longer log file for the apiserver? From what I see, there's some kind of a connection pool issue

I'll try to check it out and get back to you, seems very strange

nope, the only changes to config that we made are adding web-auth and non-responsive tasks watchdog

just in case, this warning disappeared after I https://stackoverflow.com/questions/49638699/docker-compose-restart-connection-pool-full

What is your server machine profile/spec?

Also, what's the Trains Server version?

tail of the api server log

thanks!

m5.xlarge EC2 instance (4 vCPUs, 16 GB RAM), 100GB disk

As always, the server log ( trains-apiserver for start) and more details from the browser's developer tools Network section would be appreciated 🙂

0.16.1

btw, there are "[2020-09-02 15:15:40,331] [9] [WARNING] [urllib3.connectionpool] Connection pool is full, discarding connection: elasticsearch" in the apiserver logs again

And the trains-elasticsearch log?

any suggestions on how to fix it?

Well, server seems OK, disk size might be a little on the low-end (just to be safe)

Great! What error do you still see in UI when comparing more than 20 experiments? At the time of error do you see any error response from the apiserver (in the browser network tab)? When the call to compare of 20+ task metrics succeed how much time does it usually takes in your environment?

Do you see any error in the browser network tab?

Can you share all the error info that you get in the network tab?