Coo: Consistency Check for Transactional Databases

TL;DR

Coo is a framework that efficiently checks database consistency by modeling transaction conflicts with a novel partial order graph, revealing inconsistencies and implementation gaps across various database systems.

Contribution

The paper introduces the POP graph for better conflict modeling and provides a cost-effective, SQL-based method for comprehensive consistency verification in databases.

Findings

Databases often violate ANSI SQL consistency standards.

Different databases exhibit varied behaviors under the same isolation levels.

Coo uncovers detailed inconsistent behaviors missed by traditional checks.

Abstract

In modern databases, transaction processing technology provides ACID (Atomicity, Consistency, Isolation, Durability) features. Consistency refers to the correctness of databases and is a crucial property for many applications, such as financial and banking services. However, there exist typical challenges for consistency. Theoretically, the current two definitions of consistency express quite different meanings, which are causal and sometimes controversial. Practically, it is notorious to check the consistency of databases, especially in terms of the verification cost. This paper proposes Coo, a framework to check the consistency of databases. Specifically, Coo has the following advancements. First, Coo proposes partial order pair (POP) graph, which has a better expressiveness on transaction conflicts in a schedule by considering stateful information like Commit and Abort. By POP graph with no cycle, Coo defines consistency completely. Secondly, Coo can construct inconsistent test cases based on POP cycles. These test cases can be used to check the consistency of databases in accurate (all types of anomalies), user-friendly (SQL-based test), and cost-effective (one-time checking in a few minutes) ways. We evaluate Coo with eleven databases, both centralized and distributed, under all supported isolation levels. The evaluation shows that databases did not completely follow the ANSI SQL standard (e.g., Oracle claimed to be serializable but appeared in some inconsistent cases), and have different implementation methods and behaviors for concurrent controls (e.g., PostgreSQL, MySQL, and SQL Server performed quite differently at Repeatable Read level). Coo aids to comprehend the gap between coarse levels, finding more detailed and complete inconsistent behaviors.