From Modeling to Systematic Theory: Tracing a Decade of Database Consistency Evolution (2020‑2025)

In 2020 the speaker introduced a multi‑level consistency model for distributed transactions, categorizing consistency from weak to strong and linking each level to concrete problems, required costs, and suitable scenarios. He likened this to learning the addition commutative law after mastering simple addition, turning a set of ad‑hoc solutions into a reusable model.

In 2021 he shifted focus to data anomalies, arguing that they follow a unified "data‑state model". He defined more than twenty new anomalies, illustrated their relationships with a directed graph, and compared his work to earlier standards: ANSI‑SQL listed only four anomalies, Jim Gray’s classic texts covered eight, and Adya’s definitions tied anomalies to isolation levels without seeking underlying regularities.

2022’s talk coined the phrase "no anomaly = consistency", critiquing the vague notion of a "legal" state in traditional definitions. He argued that defining consistency by the absence of anomalies leaves the term "legal" undefined, making the definition a tautology. The 2023 presentation expanded this critique into the "Consistency Eight Immortals" diagram, mapping the consistency implementations of major databases (Oracle, MySQL, PostgreSQL, SQL Server, etc.) and exposing how many claimed "serializable" guarantees (e.g., Oracle’s MVCC‑based serializability) actually fail to prevent write‑skew anomalies.

The analysis highlighted three layers of problems with ANSI‑SQL isolation levels: (1) ambiguous natural‑language definitions (citing Berenson et al., 1995), (2) reliance on lock‑based implementations that ignore newer snapshot‑isolation techniques, and (3) the mismatch between claimed serializability and real‑world behavior, illustrated by Oracle’s long‑standing mis‑labeling.

Building on this, the 2024 talk introduced the third‑generation distributed database CCDB, grounded in the "Consistency and Concurrency" theory. CCDB decouples transaction processing from storage, achieving two "100 %" goals: independence from storage systems and a clear, unified solution to historic anomalies. The speaker also presented the zero‑time unlock (ZU) technique, which relocates lock‑release and dead‑lock detection work from the commit path to the lock‑acquisition phase, reducing both computational overhead and latency without eliminating the locking mechanism.

In 2025 he distilled twelve years of "tearing apart" into three design principles for concurrent systems: priority (address the most critical concerns first), decoupling (break tightly‑coupled components into minimal units), and independence (ensure each unit operates without interfering with others). He also introduced a quantitative metric for data‑anomaly prevalence, turning vague severity judgments into concrete percentages that guide debugging and optimization decisions.

Overall, the narrative demonstrates a progression from solving isolated problems to constructing systematic models, then to unifying the entire consistency landscape across databases and broader concurrent systems, ultimately proposing a theory‑driven, quantifiable approach to database design and evaluation.