Veer: Verifying Equivalence of Dataflow Versions in Iterative Data Analytics (Extended Version)

TL;DR

Veer is a tool that efficiently verifies the equivalence of different dataflow versions in iterative data analytics, enabling performance optimization by reusing previous results, especially when dataflows are large and complex.

Contribution

Veer introduces a window-based approach that decomposes dataflow pairs and leverages existing equivalence verifiers to handle complex, large dataflows effectively.

Findings

Successfully verifies equivalence of complex dataflows

Outperforms existing EVs in efficiency and capability

Supports large-scale real-world dataflows

Abstract

Data analytics using GUI-based dataflows is an iterative process in which an analyst makes many iterations of changes to refine the dataflow, generating a different version at each iteration. In many cases, the result of executing a dataflow version is equivalent to a result of a prior executed version. Identifying such equivalence between the execution results of different dataflow versions is important for optimizing the performance of a dataflow by reusing results from a previous run. The size of the dataflows and the complexity of their operators often make existing equivalence verifiers (EVs) not able to solve the problem. In this paper, we present "Veer," which leverages the fact that two dataflow versions can be very similar except for a few changes. The solution divides the dataflow version pair into small parts, called windows, and verifies the equivalence within each window by using an existing EV as a black box. We develop solutions to efficiently generate windows and verify the equivalence within each window. Our thorough experiments on real dataflows show that Veer is able to not only verify the equivalence of dataflows that cannot be supported by existing EVs but also do the verification efficiently.