DBMS bugs can cause serious consequences, posing severe security and privacy concerns. This paper works towards the detection of memory bugs and logic bugs in DBMSs, and aims to solve the two innate challenges, including how to generate semantically correct SQL queries in a test case, and how to propose effective oracles to capture logic bugs. To this end, our system proposes two key techniques. The first key technique is called context-sensitive instantiation, which considers all static semantic requirements (including but not limited to the identifier type used by existing systems) to generate semantically valid SQL queries. The second key technique is called multi-plan execution, which can effectively capture logic bugs. Given a test case, multi-plan execution makes the DBMS execute all query plans instead of the default optimal one, and compares the results. A logic bug is detected if a difference is found among the execution results of the executed query plans. We have implemented a prototype system called Kangaroo and applied it to three widely used and well-tested DBMSs, including SQLite, PostgreSQL, and MySQL. Our system successfully detected 50 new bugs. The comparison between our system with the state-of-the-art systems shows that our system outperforms them in terms of the number of generated semantically valid SQL queries, the explored code paths during testing, and the detected bugs.