Why is l-diversity often described as "nice in theory, hard to achieve in practice"?

Meeting it often forces harsh generalization, and with multiple sensitive attributes the cost compounds — gutting utility.

Two practical problems:

• Harsh generalizations required. To force enough diversity into every class, you frequently have to generalize quasi-identifiers aggressively, drastically reducing the utility of the result.
• Multiple sensitive attributes. If a dataset has more than one sensitive column, l-diversity must be satisfied for each one individually. The constraints stack, compounding the utility loss dramatically.

So l-diversity is conceptually clean but operationally expensive — a recurring theme: stronger privacy models keep raising the utility bill.

Tip: Every privacy guarantee you add is another constraint the data must satisfy simultaneously. Constraints multiply; usable detail divides.

Go deeper:

• l-diversity (Wikipedia) — the utility cost and multi-attribute compounding.