Train a model to classify Azure ML "log lines" as failure vs. non-failure

[dkmiller]

Log files are in the workspaceblobstore, container azureml, path ExperimentRun/dcid.{run_id}/.

Scrape the run history APIs to get a list of:

• run ID
• run type
• end state (failed, cancelled, etc.)

join that with log files to get a dataset containing

• (run info minus end state)
• log file name
• log line

Then create a model to predict end state from that data.

Possible workspaces

• Get started in amlworkspace6upghgpmdpxoq, can easily work from personal machine.
• Transition to Alexander256V100, many runs and robust GPU there.

[dkmiller]

May need credential passthrough (OBO), since the container azureml doesn't show up as a datastore, even though users' identity has direct access.