Resilient store: a heuristic-based data format selector for intermediate results

Abstract

Large-scale data analysis is an important activity in many organizations that typically requires the deployment of data-intensive workflows. As data is processed these workflows generate large intermediate results, which are typically pipelined from one operator to the following. However, if materialized, these results become reusable, hence, subsequent workflows need not recompute them. There are already many solutions that materialize intermediate results but all of them assume a fixed data format. A fixed format, however, may not be the optimal one for every situation. For example, it is well-known that different data fragmentation strategies (e.g., horizontal and vertical) behave better or worse according to the access patterns of the subsequent operations. In this paper, we present ResilientStore, which assists on selecting the most appropriate data format for materializing intermediate results. Given a workflow and a set of materialization points, it uses rule-based heuristics to choose the best storage data format based on subsequent access patterns.We have implemented ResilientStore for HDFS and three different data formats: SequenceFile, Parquet and Avro. Experimental results show that our solution gives 18% better performance than any solution based on a single fixed format.