@inbook{floratou2018columnar,
author = {Floratou, Avrilia},
title = {Columnar Storage Formats},
booktitle = {Encyclopedia of Big Data Technologies},
year = {2018},
month = {May},
abstract = {Fast analytics over Hadoop data has gained significant traction over the last few years, as multiple enterprises are using Hadoop to store data coming from various sources including operational systems, sensors and mobile devices, and web applications. Various Big Data frameworks have been developed to support fast analytics on top of this data and to provide insights in near real time.

A crucial aspect in delivering high performance in such large-scale environments is the underlying data layout. Most Big Data frameworks are designed to operate on top of data stored in various formats, and they are extensible enough to incorporate new data formats. Over the years, a plethora of open-source data formats have been designed to support the needs of various applications. These formats can be row or column oriented and can support various forms of serialization and compression. The columnar data formats are a popular choice for fast analytics workloads. As opposed to row-oriented storage, columnar storage can significantly reduce the amount of data fetched from disk by allowing access to only the columns that are relevant for the particular query or workload. Moreover, columnar storage combined with efficient encoding and compression techniques can drastically reduce the storage requirements without sacrificing query performance.

Column-oriented storage has been successfully incorporated in both disk-based and memory-based relational databases that target OLAP workloads (Vertica 2017). In the context of Big Data frameworks, the first works on columnar storage for data stored in HDFS (Apache Hadoop HDFS 2017) have appeared around 2011 (He et al. 2011; Floratou et al. 2011). Over the years, multiple proposals have been made to satisfy the needs of various applications and to address the increasing data volume and complexity. These discussions resulted in the creation of two popular columnar formats, namely, the Parquet (Apache Parquet 2017) and ORC (Apache ORC 2017) file formats. These formats are both open-source and are currently supported by multiple proprietary and open-source Big Data frameworks. Apart from columnar organization, the formats provide efficient encoding and compression techniques and incorporate various statistics that enable predicate pushdown which can further improve the performance of analytics workloads.

In this article, we first present the major works in disk-based columnar storage in the context of Big Data systems and Hadoop data. We then provide a detailed description of the Parquet and ORC file formats which are the most widely adopted columnar formats in current Big Data frameworks. We conclude the article by highlighting the similarities and differences of these two formats.},
publisher = {Springer},
url = {http://approjects.co.za/?big=en-us/research/publication/columnar-storage-formats/},
}