[Feature] Catch signals for SLURM jobs?

[ltalirz]

This is just a thought triggered by an observation.

We noticed that when a SLURM job is scanceled that is currently inside a tempfile.TemporaryDirectory context manager, the context is not unwound (leaving behind a temporary directory that could potentially be large and which could potentially accumulate on nodes over time).

Apparently, this is expected behavior as the __exit__ block only runs during normal Python exception unwinding.

Now, I believe as an executorlib user one can work around this by catching the SIGTERM signal directly in the corresponding Python code and raising an exception instead, allowing the code to unwind. Or perhaps executorlib could do this for us, such that cancelling a slurm job is seen by the Python code as an exception being raised rather than a SIGTERM.

I'm not sure whether it is a good idea for executorlib to do this sort of transformation, just wanted to mention the idea

@Gitdowski

[ltalirz]

P.S. After looking into the behavior of /tmp, since 24.04 Ubuntu has a daily cleanup job that removes tmp files older than 30 days. Which perhaps makes the above use case a bit less important.

Nevertheless, there can be other situations where you may want to be able to react to the job being shut down in Python, and transforming the signal to an exception could allow certain code to benefit from existing automatisms rather than having to implement signal handling manually.