How to start a mlflow remote server?

[HongHuangNeu]

System information

• python 2.7.5
• MLflow installed from pip
• MLflow version 0.1.0

Describe the problem

How to start a mlflow remote server? By default "mlflow ui" starts a mlflow server on 127.0.0.1:5000, which is hard coded in the code of the mlflow server. I cannot access the server from outside. Examples and documentations shows that server only runs in localhost. How could I configure mlflow server so that I can access the server remotely?

[Cs4r]

I don't think this is a real issue.

However, here's one possible solution to your problem:

1. Get a machine where to run mlflow (It can be your local machine or an EC2/GCloud/Azure instance)
2. Run mlflow on the aforementioned machine
3. Make the running machine accept Internet traffic to port 5000
4. Get to know the public IP of the running machine
5. Use the public IP to connect to the mlflow server e.g: 10.20.30.40:5000

If you want to expose a ml server running on your local machine, services like https://ngrok.com/ might help.

Bear in mind that exposing your local machine to the Internet involves taking some security risks.

[bohea]

mlflow uses flask to build server
command 'server' equals flask app.run

according to flask's doc, flask app.run is:

a very simple builtin server, which is good enough for testing but probably not what you want to use in production

if you want a remote server, try some thing like uwsgi, gunicorn

[jeanineharb]

I set up my machine the same way as suggested by @Cs4r. However, when attempting to connect to the MLflow UI using the public IP, I get a 404 page.

Any idea if this is a configuration problem or a bug? Thanks.

[Cs4r]

It might be related to your machine's firewall configuration. I tested what I suggested above and worked fine on my home computer.

Bear in mind that your company might be imposing a strict firewall policy that you are not aware of.

I believe this doesn't have to do with mlflow.

[Sufod]

I'm encountering the same problem, the server is running and listening, but i can't get anything else than a 404 error...

I can see the requests on the server console.

Dockerfile is not functional either...

[jeanineharb]

Undoing the commit for #16 resolved the error locally. This is a known Flask-related issue: https://stackoverflow.com/questions/29375903/flask-app-only-returns-404s-regardless-of-how-it-is-run

In order to expose the service from a remote machine, my guess is that the host in server.app.run("0.0.0.0") should point to the machine's public IP, not 0.0.0.0.

[sanzgiri]

One way around this problem is to simply specify the public IP of the the remote mlflow server in cli.py in the call to server.app.run() e.g. server.app.run("xxx.yyy.zzz.www")

[mateiz]

FYI this should be fixed in the master branch now. Try mlflow server --host 0.0.0.0. We plan to push a release to PyPI soon to have this in there by default (probably in the next 2-3 days).

[nipunsadvilkar]

@mateiz : Is it released? I tried building from source code but couldn't get around with mlflow server --host 0.0.0.0. command

[mateiz]

It should be in the code. Did you follow the instructions in https://github.com/databricks/mlflow/blob/master/CONTRIBUTING.rst for how to run mlflow from source? Maybe you have another one on your PATH.

[giordyb]

Hi,

I'm having a similar issue: I have installed mlflow with pip pointing at the source (pip3.6 install git+https://github.com/databricks/mlflow.git) and I can run the tracking server with "mlflow server -h myip -p 5000".
Data is being recorded in the mlruns folder so it looks like the rest api is actually working.
Unfortunately when I try to access the web interface I also get a 404.
I also tried to run "mlflow ui-h myip -p 5000" but there is no difference.
any ideas?

[mateiz]

Ah, I'm not sure the web UI works when you install it that way. I believe you have to run the JS part of the web UI manually with npm as described at https://github.com/databricks/mlflow/blob/master/CONTRIBUTING.rst. We just published a 0.2.0 release btw so that will have the updated server and working UI if you just pip install mlflow.

[giordyb]

thanks, updating to 0.2.0 did the trick!

[andrewmchen]

Fixed with mlflow server --host 0.0.0.0 in release 0.2.0.

[wedesoft]

I had to allow access to the port to get it working with the host option.

sudo iptables -I INPUT -p tcp -m tcp --dport 5000 -j ACCEPT
poetry run mlflow ui --host 0.0.0.0