This looks good to me, but I wonder: should this be done along with the addition of a FileIO API? Maybe you should supply the FileIO API and use that in HadoopTableOperations to avoid the need for this. If we do it at the same time, that's a bit less API churn.

Should the FileIO API support file rename as well though? The FileSystem here is used for things other than fs.create, fs.delete, and fs.open. It also does stuff like exists.

I don't think that FileIO should handle other operations like rename. We could add exists to InputFile and OutputFile.

This is the only use of rename in the project. It is used to ensure that metadata commits are atomic when the underlying file system supports atomic rename. Since that's required for this class, then I think we should go ahead with this PR and add FileIO separately.