Add docs for lakeFS<> mlflow integration #8599
Conversation
talSofer
added
docs
|
♻️ PR Preview 7d9d51f has been successfully destroyed since this PR has been closed.
🤖 By surge-preview |
E2E Test Results - DynamoDB Local - Local Block Adapter |
E2E Test Results - Quickstart |
| 1. **Experiment Reproducibility**: By leveraging MLflow's [input logging](https://mlflow.org/docs/latest/python_api/mlflow.html#mlflow.log_input) | ||
| capabilities alongside lakeFS's data versioning, you can precisely track the specific dataset version used in each experiment | ||
| run. This ensures that experiments remain reproducible over time, even as datasets evolve. |
There was a problem hiding this comment.
I'm unsure whether using LakeFS as a versioning engine for logging would be beneficial.
Assume mlflow tracking server where runs, metrics params and tags are store in postgres while the models, images, configs and logs are on lakeFS
There was a problem hiding this comment.
talked with tal, I got it wrong and the data that mlflow log is the data source location which is not the logging data I thought the experiment produce.
can ignore my comment.
| To determine whether two distinct MLflow runs utilized the same input dataset, you can compare specific attributes of | ||
| their logged Dataset objects. The source attribute, which contains the versioned dataset's URI, is a common choice for | ||
| this comparison. Here's an example: |
There was a problem hiding this comment.
how the following approach works with how mlflow compare the results? https://mlflow.org/docs/latest/getting-started/quickstart-2/index.html#compare-the-results
There was a problem hiding this comment.
This is orthogonal to their guide, in my opinion. We let MLflow users drill down into additional run details, which they can use to reproduce experiment results or troubleshoot.
Would it add clarity if I explained why to compare runs input?
There was a problem hiding this comment.
talked with tal and discuss the option to use the SDK to compare the actual data or provide more information about the data source from the commit.
| 1. **Experiment Reproducibility**: By leveraging MLflow's [input logging](https://mlflow.org/docs/latest/python_api/mlflow.html#mlflow.log_input) | ||
| capabilities alongside lakeFS's data versioning, you can precisely track the specific dataset version used in each experiment | ||
| run. This ensures that experiments remain reproducible over time, even as datasets evolve. |
There was a problem hiding this comment.
talked with tal, I got it wrong and the data that mlflow log is the data source location which is not the logging data I thought the experiment produce.
can ignore my comment.
Closes #8598