Detecting Transactional Bugs in Database Engines via Graph-Based Oracle Construction

Transactions are an important feature of database management systems (DBMSs), as they provide the ACID guarantees for a sequence of database operations. Consequently, approaches have been proposed to automatically find transactional bugs in DBMSs. However, they cannot handle complex operations and predicates common in real-world database queries, and thus miss bugs.

This paper introduces a general, effective technique for finding transactional bugs in DBMSs that supports complex SQL queries and predicates. At the conceptual level, we address the test-oracle problem by constructing semanticallyequivalent test cases based on fine-grained statement-level dependencies in transactions. At the technical level, we introduce (1) statement-dependency graphs to describe dependencies among SQL statements in transactions, (2) SQLlevel instrumentation to capture possible statement-level dependencies, and (3) transactional oracle construction to generate semantically-equivalent test cases using statementdependency graphs. We also establish the correctness of our approach in generating semantically-equivalent test cases. We have realized our technique as a tool, TxCheck, and evaluated it on three widely-used and well-tested DBMSs, namely TiDB, MySQL, and MariaDB. In total, TxCheck found 56 unique bugs, 52 of which have been confirmed and 18 already fixed. We believe that TxCheck can help solidify DBMSs' support for transactions thanks to its generality and effectiveness.