[Tune] [Docs] Checkpointing example resume functionality is misleading

[justinvyu]

Description

This example doesn't seem to actually resume the experiment as one of the comments suggest:

# if this experiment exists, we will resume from the last run
# as specified by
name="my-tune-exp",

Recreating a Tuner(…) rather than going through Tuner.restore(...) currently never resumes (resume=False is being passed into tune.run -> TrialRunner), and there’s no way to pass in a resume config into Tuner.

The old docs had the same example, and the code here does work because it manually passed in resume="AUTO", but this is not configured in the current broken example with the new 2.0 API.

In summary, the only way to resume a Tune experiment is through Tuner.restore right now, and the docs should not imply that there is another way. This makes it so that a change in your run script is needed to resume from an errored experiment (i.e. changing tuner = Tuner(...) to tuner = Tuner.restore(exp_path, resume_errored=True)).

Link

https://docs.ray.io/en/master/tune/tutorials/tune-checkpoints.html#a-simple-cloud-checkpointing-example