If your $MCR_CACHE_ROOT on $scratch_dir is shared across multiple jobs, you may run into difficulties where two jobs start at roughly the same time, causing one to delete the cache that the other just unpacked. Ideally you would seed time or jobID into the scratch_dir name to avoid collisions, but 350MB x n jobs may be more scratch than your node has available. It's an unfortunately tricky balance inherent to deployed matlab applications.

Our nodes don't have enough scratch to avoid overlaps, so we have this in a wrapper script:

while ! stat "$MCR_CACHE_ROOT" > /dev/null 2>&1; do
    echo "Waiting for $MCR_CACHE_ROOT to be accessible..."
    sleep 1  # Wait 1 second before checking again
done