Processing a Trillion Rows Per Second on a Single Machine: How Can Nested Loop Joins be this Fast?

This blog post describes our experience debugging a failing test case caused by a cross join query running "too fast." Because the root cause of fail test case spans across multiple layers--from Apache Spark to the JVM JIT compiler-- we wanted to share our analysis in this post. The vast majority of big data SQL or MPP engines follow the Volcano iterator architecture that is inefficient for analytical workloads. Since Spark 2.0 release, the new Tungsten execution engine in Apache Spark implements whole-stage code generation, a technique inspired by modern compilers to collapse the entire query into a single function. This JIT compiler approach is a far superior architecture than the row-at-a-time processing or code generation model employed by other engines, making Spark one of the most efficient in the market.