Velox — Unified Vectorized Execution Engine by Meta

Introduction

Velox is an open-source C++ library from Meta that provides a unified, high-performance vectorized execution engine for data processing. Instead of every query engine rebuilding the same operators, Velox offers a shared set of reusable components — type system, vector buffers, expression evaluation, and operators — that any data system can plug into.

What Velox Does

• Provides a columnar in-memory data format with Apache Arrow compatibility
• Implements vectorized expression evaluation for SQL-like computations
• Offers a library of common operators: filter, project, join, aggregate, sort, and exchange
• Supports adaptive query execution with runtime filter pushdown and dynamic partition pruning
• Enables extensibility through user-defined functions, types, and connectors

Architecture Overview

Velox operates on columnar vectors using a pull-based iterator model where operators request batches from their children. The expression evaluator compiles SQL expressions into a tree of vector functions that process entire columns at once. Memory management uses arena allocation with spill-to-disk support. Connectors abstract data sources (Hive, Iceberg, S3) behind a unified scan interface.

Self-Hosting & Configuration

• Build from source on Linux or macOS using make with dependency setup scripts provided
• Integrate as a C++ library by linking against libvelox in your query engine
• Configure memory limits, spill directories, and thread pool sizes programmatically
• Register custom scalar and aggregate functions via the function registry API
• Use the built-in benchmarking framework to measure operator performance

Key Features

• Vectorized processing delivers 3-10x speedups over row-at-a-time execution
• Adaptive execution reorders and prunes operations based on runtime statistics
• Shared codebase used by Presto, Spark (Gluten), and PyTorch (TorchArrow) at Meta
• Codegen support for JIT-compiling hot expressions
• Comprehensive function library covering 300+ Presto-compatible SQL functions

Comparison with Similar Tools

• Apache DataFusion — Rust-based query engine; Velox is a C++ library designed for embedding into existing engines
• Apache Arrow Acero — Arrow's built-in execution engine; Velox offers more advanced adaptive execution
• DuckDB — Standalone analytical database; Velox is a composable library, not a complete database
• ClickHouse — Full OLAP database; Velox provides building blocks rather than a turnkey solution

FAQ

Q: Is Velox a database? A: No. Velox is an execution library. Query engines like Presto and Spark use Velox as their compute backend.

Q: What languages can use Velox? A: Velox is a C++ library with Python bindings (PyVelox) for experimentation and testing.

Q: Does Velox support GPU acceleration? A: Not natively. Velox focuses on CPU vectorization, though community efforts explore GPU integration.

Q: Who uses Velox in production? A: Meta uses Velox in its Presto deployment. The Gluten project brings Velox to Apache Spark.

Sources

• https://github.com/facebookincubator/velox
• https://velox-lib.io/