We hit a hang bug when testing S3A with parallel renames.
I narrowed this down to the newer AWS Java SDK. It only happens under load, and appears to be a failure to wake up a waiting thread on timeout/error.
I've created a github issue here:
I can post a Hadoop scale test which reliably reproduces this after some cleanup. I have posted an SDK-only test here which reproduces the issue without Hadoop:
I have a support ticket open and am working with Amazon on this bug so I'll take this issue.