A lot less than you'd think - the fundamentals of HBase are excellent. File formats, compacting, etc is all solid and the same.

The issues arise during fail over and the architecture of how HBase and HDFS interact. That is where the most changes are, and while it seems scary, I think people will find the changes most refreshing, and cut out literally hundreds of thousands of lines of code out of the write and read paths!

Fair enough. Plenty of new code. However, the design isn't a research project. It's actually pretty conservative.