Amazon's Distributed Computing Manifesto (1998)
The paper begins with a problem that Amazon is facing with its current client-server architecture: applications access the database directly and are tightly coupled with the data model. This makes the application sensitive to changes in the underlying data layer – where and how data is stored.
It proposes two solutions: one, a move towards a service-oriented architecture, and two, to model processes as a workflow.
Service-oriented architectures (SOA) might not seem groundbreaking, but it was a significant departure from thinking about applications in the two-tier client-server model.
In retrospect, it's easier to see how microservices would evolve from the proposed three-tier architecture. Moreover, that thinking of data domains as services set the stage for thinking about infrastructure as shared services a decade later (AWS).
Another interesting thread is the emphasis on workflows. While not as prevalent as SOA, thinking of distributed systems through the lens of a long-running workflow is still common today.
Both workflows are services that emphasize the customer-centricity of Amazon – systems that existed to (1) deliver services to customers and (2) make it as easy as possible for internal developers to add new features and work together.
The proposal anticipated the drawbacks to distributed state synchronization and solved them in ways that would generally resemble how we might architecture it today.