Hey everyone! Just stumbled onto something that completely changed how I interact with my experiment data and had to share.
I was doing my usual post-run analysis, trying to compare metrics across a bunch of different hyperparameter configurations, and it was getting messy. I was exporting CSV files and joining them in a notebook, which felt clunky. Then I found it buried in the docs: you can query your runs directly using SQL via their public API.
It’s a game-changer. Instead of pulling down everything and filtering locally, you can write a SQL query to fetch exactly the cross-sections you need. For example, I could ask for all runs from last week where the learning rate was above 0.001 and the final validation accuracy was > 92%, and join that with their tags and config parameters. It feels like having a supercharged experiment database.
The setup is pretty straightforward—you just need your API key and the base URL. They have a query endpoint that accepts a standard SQL `SELECT` statement. I’ve been using it from Python scripts and it integrates seamlessly into my analysis pipeline, especially when I need to feed curated results into our reporting dashboards.
Has anyone else been using this feature heavily? I'm curious about the kinds of complex joins or aggregations you're running. Also, any gotchas on performance with really large run histories? I'm still exploring the limits, but so far it's made evaluating model performance across our projects way more efficient. 😊