Evicting for the greater good: The case for Reactive Checkpointing in serverless computing

Evictable cloud resources give providers extra flexibility in managing their infrastructure and help reduce resource idleness. Due to its user-transparent and ephemeral nature, serverless computing appears to be a logical fit to make use of this type of resources. However, in its current form, this flexibility comes at the cost of semantics, namely by imposing execution time limits and only guaranteeing at least once semantics. As such, developers must either implement idempotent code or resort to proactive checkpointing or logging approaches, which pessimistically introduce runtime overheads in every invocation. In this paper, we make the case for an alternative approach to handling function interruptions, based on a reactive checkpoint-based mechanism. Following this approach, we present R-Check, a system designed to efficiently migrate function instances running on top of evictable cloud resources when interruptions are imminent. Our evaluation shows that significant resource savings can be achieved using such resources, and that function migration is possible with minimal runtime overhead.