Is it possible to deploy centralized MLflow tracking server as an app on OOD?

Get Help
, ,
1

I want to deploy centralized mlflow tracking server for all users on our cluster, one option is to deploy it as a standalone service using LDAP auth and NGNIX proxy. As Open OnDemand already has these two, I am looking for a way to integrate it with OOD. Is it possible ? Has anyone ever attempted ?

2

Varun: Welcome and thanks for posting. A key feature of Open OnDemand is that everything launches as the authenticated client. This is something most people want in order to minimize potential security risks. Juptyer is the most popular app in the Open OnDemand community, and instead of the default Juptyer approach of standing up an instance for everyone to access, each client gets their own instance. You might want to look at our architecture diagrams a bit to learn a bit more about this: Architecture — Open OnDemand 3.1.0 documentation

As such, I’m not sure there is a good way to accomplish what you are asking, but I’d suggest maybe consider why you are looking for a centralized server versus can things be accomplished with individualize servers somehow?

3

I don’t know what this app is/does. Deploying it as a Passenger application on OOD may be an option.

However, you have to consider our Per-User-Nginx architecture, which is the fact that everyone gets their own web server stack. The work centralized throws me off here. For example, connecting to a database is going to be tricky if not plain impossible. Unless you have databases for every single user (like a sqlite3 database in their HOME directory).

https://osc.github.io/ood-documentation/latest/tutorials/tutorials-passenger-apps.html

4

Hi Alan,

Thank you for the response.
I was considering a centralized server due to the possibility of collaboration but I think individual server similar to jupyter are a better option, as it can be easily accomplished with OOD.

5

Hi Jeff,

Thanks for the quick response. Yes I realized having a centralized server has more cons than pros and it’s easier with your architecture to have per user MLflow servers, similar to jupyter notebooks.