> At a minimum, it should at least be using it’s own maven module for a > lot of the bits that generates it’s own maven jars so that we can split this > functionality up at build/test time.
I expected this to be the case, but looks like it isn't. There's lot of value in splitting the HDFS code into smaller modules. Definitely newer code like Ozone. When we did this for YARN, initially there were concerns about module proliferation, but looking back, my observations have been that it has done us far more good than expected. Starting with the fact that we had clients and servers modularized independently, as well as servers from other servers, with far cleaner contracts than what we had in Hadoop 1 world. Thanks +Vinod