Siren Platform User Guide

IO

The shuffling and transfer of data produced by a task is handled by a Collector. A collector will collect data, serialize it into a compact columnar data representation, and transfer it in the form of binary packets. Different collector strategies are implemented that are adapted to different tasks. For example, in case of a Hash Join, a Search/Project task will use a collector with a hash partitioning strategy to create small data partitions and shuffle these partitions uniformly across the cluster.

On the receiver side, when a packet is received, it is stored as is (without deserialization) in an in-memory data store. Tasks, such as the Join task, will directly work on top of these binary data packets in order to avoid unnecessary data copy and deserialization.

The binary data packets are created, stored and manipulated off-heap. This helps to reduce unnecessary loads on the JVM and Garbage Collection when dealing with a large amount of data. We are leveraging the Apache Arrow project for the allocation and management of off-heap byte arrays.