AWS Emulator and the engineer's job when agents write the code

In 2004, Amazon's CTO, Swami Sivasubramanian, made a pivotal decision to hire a brilliant intern, Swaminathan Sivasubramanian, also known as Swami. At the time, a significant challenge faced Amazon was the inability of commercial relational databases to scale to their needs. Swami conducted a deep dive into how key-value databases were actually used at Amazon, leading to the development of Dynamo.

After completing his PhD, Swami joined Amazon full-time, becoming one of the principal architects of AWS's data strategy. His work has also been instrumental in Amazon's AI innovation.

Twenty years ago, the team at Amazon faced the challenge of building a reliable distributed lock service, called "Alf". The service needed to be strongly consistent, working through server failures and packet loss. To test Alf, the team developed a simulation that replicated Alf's environment in a JVM, allowing them to control network conditions and introduce faults.

They also employed formal reasoning techniques such as TLA+ to identify and rectify concurrency bugs. This simulation approach significantly improved the quality of the distributed systems.

In recent years, Amazon has expanded to over 200 services, requiring a test environment for each service used by applications. Building these environments manually would take years, and agents now write code significantly faster than any team can review it in production. To address this, Amazon launched AWS Emulator, a tool that provides coding agents with a local copy of AWS to test against.

This emulator connects to a project from Swami's early career, where he and Al Vermeulen created Alf, a Paxos-based distributed lock service. The emulator enables agents to build and test environments themselves, using written descriptions of service behavior and the real service to check against.