Workload planning desk / Reading notes
Plan a recoverable batch job
Resource requests, checkpoints and a clean result boundary.
Request what is needed
Specify CPU, host memory, accelerators and temporary storage. Over-requesting reduces scheduling flexibility; under-requesting can cause avoidable termination.
Choose checkpoints
Save progress at a consistent boundary and estimate checkpoint cost. Confirm that a checkpoint can be read in the intended recovery environment.
Publish only complete results
Write partial output separately and mark completion after validation. Downstream jobs should not consume a directory merely because it has appeared.