Roll Forward, Not Back: A Case for Deterministic Conflict Resolution

Enabling applications to execute various tasks in parallel is difficult if those tasks exhibit read and write conflicts. In recent work, we developed a programming model based on concurrent revisions that addresses this challenge: each forked task gets a conceptual copy of all locations that are declared to be shared. Each such location has a specific isolation type; on joins, state changes to each location are merged deterministically based on its isolation type. In this paper, we study how to specify isolation types abstractly using operation-based compensation functions rather than state based merge functions. Using several examples including a list with insert, delete and modify operations, we propose compensation tables as a concise, general and intuitively accessible mechanism for determining how to merge arbitrary operation sequences. Finally, we provide sufficient conditions to verify that a state-based merge function correctly implements a compensation table.