Cardinality Estimation in Shared-Nothing Parallel Dataflow Engines

Abstract

Shared nothing parallel data ow systems aim to bridge the gap between MapReduce and RDBMSs by combining parallel execution of second order functions with operator based optimizations. In parallel systems, job latency is strongly affected by data shuffling and unbalanced data across nodes, thus the degree of parallelism and the data partition- ing functions must be carefully considered when choosing optimization strategies. However, it is hard to make good optimization choices with- out any information about the distribution of the data. We attempt to overcome this challenge in shared nothing parallel data ows by tracking statistics of data sets during query runtime. We use data streaming algo- rithms to track statistics so as to affect job latency as little as possible. We discuss how collected statistics can potentially be used to improve execution plans during runtime.