Recent Parquet Improvements in Apache Spark

Apache Parquet is a very popular columnar file format supported by Apache Spark. In a typical Spark job, scanning Parquet files is sometimes one of the most time consuming steps, as it incurs high CPU and IO overhead. Therefore, optimizing Parquet scan performance is crucial to job latency and cost efficiency.

Spark currently have two Parquet reader implementations: a vectorized one and a non-vectorized one. The former was implemented from scratch and offers much better performance than the latter. However, it currently doesn’t support complex types (e.g., array, list, map) at the moment and will fallback to the latter when encountering them. In addition to the reader implementation, predicate pushdown is also crucial to Parquet scan performance as it enables Spark to skip those data that do not satisfy the predicates, before the scan. Currently, Spark constructs predicates itself and rely on Parquet-MR to do the heavy lifting, which does the filtering based on various information such as statistics, dictionary, bloom filter or column index.

This talk will go through two recent improvements for Parquet scan performance: 1) vectorized read support for complex types, which allows Spark to achieve 10x+ improvement when reading Parquet data of complex types, and 2) Parquet column index support, which enables Spark to leverage Parquet column index feature during predicate pushdown. Last but not least, Chao go over some future work items that can further enhance Parquet read performance.

Connect with us:
Website: https://databricks.com
Facebook:   / databricksinc  
Twitter:   / databricks  
LinkedIn:   / data.  .
Instagram:   / databricksinc