Setting work_directory in abstract class JobExecutorConfig breaks SlurmExecutor with no attribute 'initial_queue_polling_delay'

[tesujimath]

Running with this executor doesn't work:

    JobExecutor.get_instance(
        "slurm",
        config=JobExecutorConfig(
            launcher_log_file=Path(launcher_log_file),
            work_directory=Path(work_directory),
        ),
    )

failing with this error:

AttributeError: 'JobExecutorConfig' object has no attribute 'initial_queue_polling_delay'

I can fix it by using a BatchSchedulerExecutorConfig, but I'm pretty sure this isn't the intended use of the API, because of the obscure import path

from psij.executors.batch.batch_scheduler_executor import BatchSchedulerExecutorConfig

    JobExecutor.get_instance(
        "slurm",
        config=BatchSchedulerExecutorConfig(
            launcher_log_file=Path(launcher_log_file),
            work_directory=Path(work_directory),
        ),
    )

For now I am indeed using BatchSchedulerExecutorConfig, but I feel like I'm going off-piste here.

[hategan]

Thank you for reporting this.

We do need a better way of dealing with the covariant configurations.

If I remember correctly, the BatchSchedulerExecutorConfig would be the right choice here, but:

• we should fix that import to be less obscure
• we should consider allowing JobExecutorConfig to also work and use relevant defaults when uses as you did in the first example.

[hategan]

There is a PR (#512) for the second bullet point.

After some further consideration (or rather said, recollection), I'm somewhat inclined to not address the first bullet point. The reason is that we've followed the philosophy that the abstract classes are directly represented in the psij module, while implementation details are to be accessed explicitly.