Closed annshress closed 1 year ago
Currently, SlurmCluster
is passed into flow as a decorator as a class.
Two arguments are required to create the log file are log_directory
and name
of the job, which are directly written into a job shell script
based on dask-jobqueue code. And presumably this is happening even before we can access prefect flow run's name (which can be accessed using prefect.context.get_run_context()
inside a flow).
Therefore, we are letting this go for now, unless we find a way to inject flow run's name into the sbatch
script.
Currently, we are accessing the callback output from the slurm logs. To make things easier, we could write it to a file instead.
Furthermore, slurm log file names could be much more accessible rather than devs solely relying on timestamps.