I have seen instances where the s3 reader will start a slice but not complete it causing the job to run perpetually.
In one case there was a ceph issue at the time a slice wasn't completed. We think the reader could be hiding errors and not retrying or failing the slice on these occasions.