Some issues spotted from looking at code and/or testing regression suite with all indexes forced to be row-compressed:
- Loading a blob will crash
- The needCopy flag is not properly passed in meaning memory-mapped indexes will still use the node buffers.
- Better to use a derived class for such things as compareRowAt, getValueAt etc to avoid conditional tests in the normal cases (which need to be fast). No need to test rowExp in variable node case - we don't support rowdiff on variable-size rows.