Mlos_interactive: User friendly graphs and data exploration

[yshady-acheev]

I put https://github.com/microsoft/MLOS/tree/4f6f5b025a33ae61ece70442a62bf825aeb9bad0/mlos_vizpanel into one of my https://github.com/Microsoft-CISL/sqlite-autotuning/ forks

Microsoft-CISL/sqlite-autotuning#53
Microsoft-CISL/sqlite-autotuning#51

Based on this there are two key objectives to ensure a preformance tuner (the end user/validator) would be pleased whilst using MLOS

1. Making notebooks interactive
2. Using the same code to create a no-code exploration GUI

Either it is put in mlos_viz or we create a new effort for this as discussed previously #824 and #838. Im fine with either

[yshady-acheev]

I also think there should be an assumtion that we will only focus on viewing published data such as in https://github.com/Microsoft-CISL/sqlite-autotuning/pull/51/files

A preformance tuner researcher can then view diffrent dropdowns of benchmarks with param values as a guide to tune based on their workloads

[yshady-acheev]

@bpkroth @motus @eujing check out Microsoft-CISL/sqlite-autotuning#53 on codespaces, should work out of the box (except for dependencies: pip install may be req)

1. Navigate to the project directory.
2. Start the FastAPI server:

   uvicorn backend:app --reload

Running the Frontend

1. Navigate to the project directory.
2. Start the Streamlit application:

   streamlit run frontend.py

[yshady-acheev]

[yshady-acheev]

[yshady-acheev]

@eujing had the idea for the parallel plot over summer I think it's super cool, never heard of it before that

[yshady-acheev]

I guess the Main Issue is the decision on how to integrate the GUI into mlos upstream, it works fine on codespaces

I think leaving the GUI code unexposed for a dashboard is a good idea to not over engineer but for the notebook I do see benifit in making it an api in mlos_viz/mlos_interactive