Optimizing Database Expression Evaluation with JIT Compilation Using Gandiva

This article introduces how Just‑In‑Time (JIT) compilation can be used to efficiently evaluate database expressions, focusing on the Gandiva expression compiler built on the LLVM framework.

It first defines the expression evaluation problem, using examples such as filtering logs where some fields (e.g., IP) are not known in advance, and explains the three traditional evaluation approaches: interpreted execution, virtual‑machine bytecode, and JIT compilation.

The limitations of interpreted execution are discussed, including heavy virtual‑function calls, dynamic type checks, and deep‑first recursion that hinder CPU pipeline performance.

JIT compilation is then described: the SQL parser creates an abstract syntax tree (AST), the expression compiler generates intermediate LLVM IR, and the JIT compiler turns it into native machine code, enabling vectorized SIMD execution.

Gandiva, an Apache project built on LLVM and Arrow columnar format, is presented as a concrete implementation. Its workflow—AST → LLVM IR → Arrow Record Batches → native code—is illustrated, along with recent enhancements such as support for timestamps, array functions, and user‑defined functions.

A short Q&A covers topics like SIMD support in Gandiva, differences between Arrow Compute and Gandiva, and methods for static expression simplification.

The article concludes that JIT‑based expression evaluation, combined with columnar storage, can dramatically improve query performance, especially in modern CPU‑bound workloads.