Toward fast eventual consistency with performance guarantees

Systems have adopted the notion of eventual consistency which means that the targeted redundancy of data in the system is reached asynchronously, i.e., outside of the critical path of user traffic, so that performance of user traffic is impacted minimally. Here, we propose a scheduling framework that makes decisions about when to schedule the asynchronous tasks associated with new or updated data such that they are completed as soon as possible without violating user traffic quality targets. At the heart of the framework lies a learning methodology that extracts the characteristics of idle periods and infers the average amount of work to be filled during periods of idleness so that asynchronous tasks are completed transparently to the user. Extensive trace-driven evaluation shows the effectiveness and robustness of the proposed framework when compared to common practices.