Open question: Should we support s3 paths?

In early versions of AutoGluon I supported s3 paths for TabularPredictor. I eventually found that once I mixed in ray and started using optimized model saving techniques that it became hard to ensure the code worked properly in both scenarios.

Because of this difficulty, I eventually stopped supporting S3 paths for predictor artifacts.

One work-around could be to use a local directory for all training / prediction, and then upload the local files to the s3 location after fitting is complete. Similarly, download the artifact if called via .load. I haven't implemented this approach but it could work.

If you have found a way to make s3 paths work with relative ease, I'd be interested in knowing how you did it.

@Innixma I assumed that TabularPredictor already supported S3 paths since it worked fine with medium_quality presets, but now I understand that this might not be the case for high_quality / best_quality, where ray is used.

I will check how much work it would take to make S3 paths work in all scenarios. At the very least, we should raise some informative error message explaining that only local paths are supported. Currently for the TimeSeriesPredictor we just observe some model failures in the middle of the training process, which is a bad customer experience.