A Model of Independence and Overlap for Transactions on Database Schemata

Traditional models of support for concurrent transactions invariably rely upon a notion of serializability, which involves not only complex scheduling, but also primitives (such as locks) for requiring transactions to wait, as well for aborting a transaction and forcing it to re-run. For batch transactions, this approach is often the most reasonable. On the other hand, for interactive transactions, only a very limited amount of waiting and aborting is tolerable, and so minimizing their occurrence, even at the cost of increased analysis of the transactions themselves, is warranted. In this work, a systematic study of independence for transactions, without any explicit serialization, is initiated. Each transaction operates on a view of the main schema, and each such view is partitioned into a write region and a read-only region. For a set of transactions to run concurrently, their views may overlap only on their read-only regions. These regions need not be specified explicitly; rather, they are defined naturally using a component-based model of the main schema. Furthermore, when two transactions do conflict, because their views overlap on write regions, the precise point of conflict is immediately identified. To illustrate the utility of the framework, the case of relational schemata governed by the most common types of constraints in practice — functional and foreign-key dependencies — is developed in detail.