EFFICIENT LAYOUTS AND ALGORITHMS FOR MANAGING VERSIONED DATASETS

Loading...
Thumbnail Image

Publication or External Link

Date

2018

Citation

Abstract

Version Control Systems were primarily designed to keep track of and provide control over changes to source code and have since provided an excellent way to combat the problem of sharing and editing files in a collaborative setting. The recent surge in data-driven decision making has resulted in a proliferation of datasets elevating them to the level of source code which in turn has led the data analysts to resort to version control systems for the purpose of storing and managing datasets and their versions over time. Unfortunately existing version control systems are poor at handling large datasets primarily due to the underlying assumption that the stored files are relatively small text files with localized changes. Moreover the algorithms used by these systems tend to be fairly simple leading to suboptimal performance when applied to large datasets. In order to address the shortcomings, a key requirement here is to have a Dataset Version Control System (DVCS) that will serve as a common platform to enable data analysts to efficiently store and query dataset versions, track changes to datasets and share datasets between users at ease.

Towards this goal, we address the fundamental problem of designing storage layouts for a wide range of datasets to serve as the primary building block for an efficient and scalable DVCS. The key problem in this setting is to compactly store a large number of dataset versions and efficiently retrieve any specific version (or a collection of partial versions). We initiate our study by considering storage-retrieval trade-offs for versions of unstructured dataset such as text files, blobs, etc. where the notion of a partial version is not well-defined. Next, we consider array datasets, i.e., a collection of temporal snapshots (or versions) of multi-dimensional arrays, where the data is predominantly represented in single precision or double precision format. The primary challenge here is to develop efficient compression techniques for the hard-to-compress floating point data due to the high degree of entropy. We observe that the underlying techniques developed for unstructured or array datasets are not well suited for more structured dataset versions -- a version in this setting is defined by a collection of records each of which is uniquely addressable. We carefully explore the design space for building such a system and the various storage-retrieval trade-offs, and discuss how different storage layouts influence those trade-offs. Next, we formulate several problems trading off the version storage and retrieval cost in various ways and design several offline storage layout algorithms that effectively minimize the storage costs while keeping the retrieval costs low. In addition to version retrieval queries, our system also provides support for record provenance queries. Through extensive experiments on large datasets, we demonstrate that our proposed designs can operate at the scale required in most practical scenarios.

Notes

Rights