Apache Spark is one of the most widely used open source processing engines for big data, with rich language-integrated APIs and a wide range of libraries. Over the past two years, our group has worked to deploy Spark to a wide range of organizations through consulting relationships as well as our hosted service, Databricks. We describe the main challenges and requirements that appeared in taking Spark to a wide set of users, and usability and performance improvements we have made to the engine in response.

Large-scale data analytics frameworks are shifting towards shorter task durations and larger degrees of parallelism to provide low latency. Scheduling highly parallel jobs that complete in hundreds of milliseconds poses a major challenge for task schedulers, which will need to schedule millions of tasks per second on appropriate machines while offering millisecond-level latency and high availability. We demonstrate that a decentralized, randomized sampling approach provides near-optimal performance while avoiding the throughput and availability limitations of a centralized design. We implement and deploy our scheduler, Sparrow, on a 110-machine cluster and demonstrate that Sparrow performs within 12% of an ideal scheduler.

Handling flash crowds poses a difficult task for web services. Content distribution networks (CDNs), hierarchical web caches, and peer-to-peer networks have all been proposed as mechanisms for mitigating the effects of these sudden spikes in traffic to under-provisioned origin sites. Other than a few anecdotal examples of isolated events to a single server, however, no large-scale analysis of flash-crowd behavior has been published to date.

Geo-replicated services need an effective way to direct client requests to a particular location, based on performance, load, and cost. This paper presents DONAR, a distributed system that can offload the burden of replica selection, while providing these services with a sufficiently expressive interface for specifying mapping policies. Most existing approaches for replica selection rely on either central coordination (which has reliability, security, and scalability limitations) or distributed heuristics (which lead to suboptimal request distributions, or even instability). In contrast, the distributed mapping nodes in DONAR run a simple, efficient algorithm to coordinate their replica-selection decisions for clients. The protocol solves an optimization problem that jointly considers both client performance and server load, allowing us to show that the distributed algorithm is stable and effective. Experiments with our DONAR prototype--providing replica selection for CoralCDN and the Measurement Lab--demonstrate that our algorithm performs well "in the wild." Our prototype supports DNS- and HTTP-based redirection, IP anycast, and a secure update protocol, and can handle many customer services with diverse policy objectives.