Hatch: Self-distributing systems for data centers

Designing and maintaining distributed systems remains highly challenging: there is a high-dimensional design space of potential ways to distribute system’s sub-components over large-scale infrastructure; the deployment environment for system tends change in unforeseen time. For engineers, this complex prediction problem gauge which may best suit given environment. We present concept self-distributing systems, any local built using our framework can learn, at runtime, most appropriate its perceived operating conditions. Our abstracts distribution list simple actions reward matrix alternatives be used by reinforcement learning algorithms. By doing this, we enable software experiment, live production environment, with different modules placing them hosts throughout infrastructure. implement call Hatch, has three major elements: (i) transparent generalized RPC layer that supports seamless relocation component remote host during execution; (ii) set primitives, including relocation, replication sharding, from create an action/reward possible designs system; (iii) decentralized approach converge towards more optimal real Using example web-serving infrastructure, Hatch able autonomously select suitable among ?700,000 about 5 min.