[ MLflow Roadmap ] Call for content - official MLflow Community Blog

[BenWilson2]

Summary

We're currently in the process of overhauling both our main site and the documentation for MLflow. As part of this effort, we're going to be adding an official community blog!

Do you have an interesting use case that you'd like to share?

Is there a way that you're using MLflow to simplify your MLOps needs?

Would you like to show how you're leveraging MLflow to solve a problem?

We'd love to get your story featured on the new site!

Respond to this issue and we'd love to start having a dialogue about the process, the content, and how we can help to get it published on the site!

Notes

• Make sure to open a PR from a non-master branch.

• Sign off the commit using the -s flag when making a commit:

  git commit -s -m "..."
           # ^^ make sure to use this
  

• Include #{issue_number} (e.g. #123) in the PR description when opening a PR.

[SDonkelaarGDD]

Hey Ben,

We are using Mlflow on our on-premise Kubernetes cluster. We leveraged MLflow plugins to enhance it with extra functionality such that it can be deployed securely in an enterprise environment (we protected it using Oauth2 proxy). Think about:

• Adding STS session token support for S3 Artefact Store backends.
• Adding automatic Keycloak authentication by creating a custom request_header_provider. This works similar to DefaultAzureCredential in the sense that it enables token-based authentication by checking multiple authentication flows from within its runtime environment. For example, by using credentials in the environment, or by using the browser to interactively authenticate a user. This enables seamless authentication which really benefits the user experience.
• We created a used custom mlflow server flask app such that we have proper RBAC that is fully connected with our existing (custom) IAM solutions. This enables fine-grained RBAC controls for the enterprise environment where its permissions are automatically inherited from existing IAM solutions. Hence MLflow becomes really integrated with other solutions that we have.

In our understanding, MLflow is awesome if you want to incorporate MLOps best practices on k8s. By using MLflow plugins we could tailer the solution specifically to our needs and use it apply to manage the ML lifecycle in an on-prem enterprise environment.

[github-actions[bot]]

@mlflow/mlflow-team Please assign a maintainer and start triaging this issue.

[cdreetz]

Hey Ben, I was planning on putting together some content for using MLFlow for a couple different LLMops cases. I have no preferred format for content if you guys are looking for specific kinds. Let me know

[BenWilson2]

@SDonkelaarGDD sorry that I missed your message! I think this is an awesome story to tell and we'd love to feature it as a blog post! Are you up for doing a full write-up?

[BenWilson2]

@cdreetz sounds great! A simple google doc shared with the MLflow maintainers would suffice :) we can use that as a common working document for reviews and edits.

[SDonkelaarGDD]

@SDonkelaarGDD sorry that I missed your message! I think this is an awesome story to tell and we'd love to feature it as a blog post! Are you up for doing a full write-up?

For sure, but I first need to get formal approval (we're hosting it in a strict regulatory environment). I'll get back to you once I have an answer.

[cdreetz]

@BenWilson2 I went ahead and made a google doc, whats the best way to share? Heres link

[rddefauw]

I'm building a reference implementation of a "full stack" build for GenAI workloads. It uses MLFlow for things like experiment tracking, along with other tools like OpenLLMetry for tracing. If it seems useful, I could write an article about this, including the sample code to deploy the solution.

[saradindusengupta]

Hi @BenWilson2 , At my current organisation @Nunam-Energy, we are using MLFlow in dev and post-production monitoring scenarios. The primary use case is in the energy domain and focuses a good amount of work on time-series forecasting.

We are using multiple custom mlflow.pyfunc for custom models/libraries since the use cases are very different. I would be interested to write up an article focusing on multiple different models/libraries and building a ground estimation module integration with MLFlow. Currently, I don't have a concrete structure of the article yet but would be quiet interested.