How does data skew interact with LDV to compound sharing calculation performance issues, and what combined architectural approaches address both problems simultaneously?