Reversible watermarking is an important method of information hiding, which has been widely used in copyright protection of relational data. Reversible watermarking is more advanced than its predecessors in many ways. It can be used not only to claim copyright but also to recover the original data. However, existing schemes do not allow control of the extent of data recovery. Watermarked data are either completely restored to the original version or kept unchanged. After analyzing the current problems, a graded reversible watermarking scheme for relational data is proposed here. By removing the arbitrary portion of the watermark, data quality can be enhanced incrementally. The notion of data quality grade is defined to describe the impact of watermark embedding on the usability of data. Four fundamental algorithms are designed to facilitate the processes of watermark embedding, data quality grade detection, watermark detection, and data quality grade enhancement. Before data distribution, numbers of data quality grades can be predefined. Graded reversibility can be achieved by upgrading watermarked data from low to higher data quality grades. A watermark with any data quality grade is enough to claim copyright. With watermarks embedded into different data partitions, flexible watermark reversion can be achieved via partitioned auxiliary data design. A more practical mechanism is devised to efficiently handle hash table collisions and reduce both computational and storage overheads. Tests of the computational performance of the algorithms and their response to various attacks showed that the proposed scheme is feasible and robust.