AutoShard -- Declaratively Managing Hot Spot Data Objects in NoSQL Document Stores

TL;DR

AutoShard is a Java-based object mapper that simplifies managing hot spot data objects in NoSQL document stores by automatically sharding them to improve write throughput and reduce contention.

Contribution

It introduces a declarative approach for handling hot spot data objects in NoSQL stores, enabling easy integration with minimal code changes.

Findings

Sharding hot spot objects significantly improves write throughput.

AutoShard reduces write contention in NoSQL document stores.

Experiments demonstrate notable performance gains with AutoShard.

Abstract

NoSQL document stores are becoming increasingly popular as backends in web development. Not only do they scale out to large volumes of data, many systems are even custom-tailored for this domain: NoSQL document stores like Google Cloud Datastore have been designed to support massively parallel reads, and even guarantee strong consistency in updating single data objects. However, strongly consistent updates cannot be implemented arbitrarily fast in large-scale distributed systems. Consequently, data objects that experience high-frequent writes can turn into severe performance bottlenecks. In this paper, we present AutoShard, a ready-to-use object mapper for Java applications running against NoSQL document stores. AutoShard's unique feature is its capability to gracefully shard hot spot data objects to avoid write contention. Using AutoShard, developers can easily handle hot spot data objects by adding minimally intrusive annotations to their application code. Our experiments show the significant impact of sharding on both the write throughput and the execution time.