Providing Low-Latency Geo-Distributed Transaction Processing for High-Contention Workloads

地理上分布的数据库系统用于管理和处理由许多大型应用程序生成的信息。它们为应用程序提供强大的事务保证，但需要多次广域网往返才能处理访问存储在不同地理区域的数据的事务。这使得它们不适合在延迟敏感型应用程序中使用，导致许多应用程序开发人员将事务替换为他们自己的容易出错的应用程序级解决方案。我们以前的工作通过减少处理地理分布的事务所需的广域网往返次数，解决了这个问题的一部分。但是，争用较多的工作负载仍会导致底层协议协议中的高流产率和路径处理速度变慢。该工作旨在通过引入截止期感知调度和并发控制策略来克服这一限制，该策略利用特定事务类别的属性，并向特定网络消息添加人为延迟，从而减少并发事务之间的争用。

Geographically distributed database systems are used to manage and process information generated by many large-scale applications. They provide strong transactional guarantees for applications, but require many wide-area network roundtrips to process transactions that access data stored in different geographical regions. This makes them unsuitable for use in latency-sensitive applications, causing many application developers to replace transactions with their own error-prone application-level solutions. Our previous work has addressed parts of this problem by reducing the number of wide-area network roundtrips required to process geo-distributed transactions. However, workloads with high contention can still cause high abort rates and slow-path processing in the underlying agreement protocol. This work aims to overcome this limitation by introducing deadline-aware scheduling and concurrency control policies that take advantage of properties of a specific class of transactions, and adding artificial delays to specific network messages, which can reduce contention between concurrent transactions.