HBase's use of this API is indeed what brought the lack of support to my attention. I wouldn't say I'm specifically trying to support HBase-on-S3, though - just trying to improve support in general where we can. If there's no reasonable solution, there's no reasonable solution. I certainly don't want to be increasing the potential for inconsistency if that's what it would take.
On top of what Steve said, though, I think it's well-understood that when using Hadoop-on-S3 in general there's been the potential for consistency issues. That's improved dramatically on S3's end over time, of course. In this case all nodes will currently always fail because of an unsupported API. Due to the potential for inconsistency, if this API was supported, it would be possible for some nodes to fail because of a FileAlreadyExistsException, and the operation could be retried. That's certainly an improvement, and well within the expectations set by the S3 filesystems, IMO. I would still vote to move in this direction.
Sean: fancy taking that on?
Yeah I think that'd be good to do. I'll wait until we reach some consensus on the consistency issue, though.