Sometimes users issue "bad" transactions, and these have to be rolled back after they have been committed. This creates a problem of cascading roll-backs. This paper suggests an efficient way to roll-back as little as possible while removing the effects of bad transactions.

The paper is written in the context of a traditional transaction-oriented (database) system. But I think the ideas in the paper are applicable more broadly. For example, consider a system that is gather information from new sources on the internet and performing some analyses. What happens when one of these news sources retracts a story? Can we efficiently edintify the dependent analyses and redo them? The same applies to biomedical science, and the retraction of a data set because of discovery of scientific fraud.