Correct and Efficient Timestamping of Temporal Data

Previous approaches to timestamping temporal data have implicitly assumed that transactions have no duration. In this paper we identify several situations where a sequence of operations over time within a single transaction can violate ACID properties. It has been previously shown that the transaction-time dimension must be timestamped after commit. This time is not known within the transaction. We describe how to correctly implement most queries that make explicit reference to this (unknown) transaction time, and argue that the rest, which can be syntactically identified, can only be answered with an approximation of the correct value. The drawback of timestamping after commit is that it requires revisiting tuples. We show that this expensive revisiting step is required only before any queries or modifications in subsequent transactions that access prior states; in most cases, revisiting tuples can be postponed, and when to revisit can be syntactically determined. We propose several strategies for revisiting tuples, and we empirically evaluate these strategies in order to determine under which circumstances each is best.