Thanks for the patch Yuanbo. I'm still thinking about this. The patch basically changes hadoop-hdfs' dependency on hadoop-hdfs-client from compile scope to provided, then fixed up other poms to refer to hadoop-hdfs-client with provided or test scope where needed.
Overall, it seems okay. Applications shouldn't be depending on hadoop-hdfs directly, so breaking them is okay (possibly good, as a prod to move to the client deps). We also expect Hadoop deps like this to be on the classpath when running Hadoop daemons.
That said, I am not a Maven expert, or super familiar with the expectations of dist-layout-stitching.
Steve Loughran, Allen Wittenauer, any comments?