With the right infrastructure to build on, we believe that the common case can be that users don't remember the last time an interactive datacenter service was unresponsive or lost their data.

The Cloud Services Infrastructure Project (CSIP) is targeted at reducing the complexity of building data-center services. To accomplish this, we are building distributed systems infrastructure components and tools to make life easier for the developers building these services.

Previous research efforts have already shown the value of this path: for example, scalable storage systems designed for datacenter workloads free datacenter developers from having to invent their own ad-hoc storage solutions. We believe that current efforts are just the tip of the iceberg; we need significantly more infrastructure to rapidly build scalable and reliable datacenter services.

Based on our prior work building distributed systems, we believe the best way to design infrastructure is to let the needs of real applications drive our exploration of the design space. In past work, we worked with Windows Live operations to design, build and deploy a request tracking and failure management infrastructure for a datacenter environment. In our current work, the design requirements of Windows Live services have guided our research on a partitioning and recovery service for soft state datacenter applications. We plan to continue this pattern of developing both component technologies and end-to-end solutions.

• Atul Adya
• Ronnie Chaiken
• John Dunagan
• Alec Wolman