System-aware dynamic partitioning for batch and streaming workloads

When processing data streams with highly skewed and nonstationary key distributions, we often observe overloaded partitions when the hash partitioning fails to balance data correctly. To avoid slow tasks that delay the completion of the whole stage of computation, it is necessary to apply adaptive, on-the-fly partitioning that continuously recomputes an optimal partitioner, given the observed key distribution. While such solutions exist for batch processing of static data sets and stateless stream processing, the task is difficult for long-running stateful streaming jobs where key distribution changes over time. Careful checkpointing and operator state migration is necessary to change the partitioning while the operation is running. Our key result is a lightweight on-the-fly Dynamic Repartitioning (DR) module for distributed data processing systems (DDPS), including Apache Spark and Flink, which improves the performance with negligible overhead. DR can adaptively repartition data during execution using our Key Isolator Partitioner (KIP). In our experiments with real workloads and power-law distributions, we reach a speedup of 1.5-6 for a variety of Spark and Flink jobs.