Arrow Compute IR (Intermediate Representation)

In the same way that the Arrow format provides a powerful tool for communicating data, Compute IR is intended to provide a consistent format for representing analytical operations against that data. As an arrow-native expression of computation it includes information such as explicit types and schemas and arrow formatted literal data. It is also optimized for low runtime overhead in both serialization and deserialization.

Built-in definitions are included to enable representation of relational algebraic operations- the contents of a “logical query plan”. Compute IR also has first class support for representing operations which are not members of a minimal relational algebra, including implementation and optimization details- the contents of a “physical query plan”. This approach is taken in emulation of MLIR (Multi-Level Intermediate Representation), a system which has had strong successes in spaces of comparable complexity to representation of analytic operations. To borrow terms from that project, there are two mutations of interest:

  • Replacement of representations with semantically equivalent representations which will yield better performance for consumers- an optimization pass.

  • Replacement of abstract or generic representations with more specific and potentially consumer-specific representations- a lowering pass. This modification corresponds to the translation of a logical plan to a physical plan.

Allowing representation of physical plans (and plans which are between logical and physical) in Compute IR enables systems to define incremental optimization and lowering passes which operate on and produce valid Compute IR. This in turn enables communication, manipulation, and inspection at every stage of lowering/optimization by the same tools used for logical-plan-equivalent-IR. This is especially useful for systems where such passes may depend on information only available on every node of a distributed consumer (for example statistics unique to that node’s local data) or may not be universal to all backends in a heterogeneous consumer (for example which optimizations nodes are capable of for non equi joins).