If many stages can run, the server will launch a job array with a vmem request based on the largest amount needed, which is quite wasteful (and in some cases may cause the scheduler to reject the request).
Seems somewhat important to fix (e.g., I have a custom pipeline with 1000+ 1MB stages runnable and 1 80GB stage, and all executors are now running with 80GB requests).
If many stages can run, the server will launch a job array with a vmem request based on the largest amount needed, which is quite wasteful (and in some cases may cause the scheduler to reject the request).