Technical Perspective: When is it safe to run a transactional workload under Read Committed?

Abstract

A data management platform provides many capabilities to assist the data owner, application coder, or end-user. For example, it should support an expressive query language, schema definition, and sophisticated access control. Another way many platforms add value is through a transaction mechanism, which allows the application programmer to indicate that a stretch of code, including multiple accesses to data, represents a single real-world activity and so all these steps should happen as if a single step, despite really being interleaved with other programs, or perhaps cancelled after partial execution. If the platform perfectly hides interleaving of different activities, the execution is called serializable, and this is a great aid to protecting data quality. Any integrity constraint over the data (whether explicitly declared in schema or not) which is preserved by each transaction running alone, is also valid at the end of any serializable execution of several transactions.