Amazon DynamoDB, a highly available, horizontally scalable, multitenant NoSQL cloud database service launched in 2012, serves hundreds of thousands of customers and supports Amazon systems like Amazon.com, Alexa, and Amazon fulfillment centers – including on Prime Day when DynamoDB must respond to up to 89.2 million requests per second. In this paper, Elhemali et al. describe how AWS engineers have evolved DynamoDB without compromising consistent, predictable low latency. Specifically, the authors describe how the AWS team altered the physical partitioning scheme of the database in response to customers’ non-uniform traffic patterns and achieve high durability goals through continuous verification of data-at-rest. They also discuss how tools and processes like formal proofs, game days, and upgrade/downgrade tests enable high availability when adding new features to DynamoDB and the importance of designing for predictability over efficiency.