Shubham is a software engineer on the Spark Platform team at Bloomberg. He has been a long time user of Hadoop and Spark. He previously contributed to Apache Pig and helped develop the ‘illustrate’ function. He is currently working on improving Spark’s reliability for online analytics.
Apache Spark was designed as a batch analytics system. By caching RDDs, Spark speeds up jobs that iteratively process the same data. This pattern is also applicable to online analytics. We use Bloomberg's Spark Server as a server runtime for online analytics. Our framework implements certain useful patterns applicable to online query processing and is centered on the idea of “Managed” DataFrames that can be refreshed and updated as per user requirements, without violating the immutability of RDDs/DataFrames. However, Spark presents significant challenges with respect to availability and resilience in an online setting where Spark is required to respond to queries with high SLAs. In this talk, we try to identify specific areas where slow-down or failures can result in the largest hits on online-query performance and potential solutions to address these.