Technology

Velox

A C++ vectorized execution engine developed by Meta that provides a unified, high-performance data processing backend usable by multiple front-end query engines including Presto, Spark, and custom data systems.

6 connections 3 resources

Summary

What it is

Where it fits

Velox sits beneath query planners as a shared execution layer. For S3-backed workloads, it accelerates scan, filter, aggregation, and join operations against Parquet files on object storage, and is the engine behind Presto's Velox-based execution (Prestissimo).

Misconceptions / Traps

Velox is not a standalone query engine. It is an execution library that must be embedded in a host system (Presto, Spark via Gluten, or a custom application).
Velox's performance gains come from vectorized execution and adaptive filtering, not from caching. It still needs to read data from S3 on cache misses.
Integration with existing query engines (e.g., Spark via Gluten project) is still maturing. Not all Spark operations have Velox equivalents.

Key Connections

scoped_to S3, Lakehouse — accelerates query execution over S3 data
depends_on Apache Arrow — uses Arrow-compatible columnar memory layout
enables Trino — Prestissimo uses Velox as its execution engine
enables Apache Spark — Gluten project integrates Velox with Spark

Definition

What it is

A C++ vectorized database acceleration library created by Meta, designed to be embedded into query engines to provide a unified, high-performance execution layer for data processing on S3-stored data.

Why it exists

Multiple query engines (Spark, Presto, Flink) each implement their own execution runtimes with varying performance characteristics. Velox provides a shared, hardware-optimized execution core that any engine can embed, raising the performance floor for S3-based analytics.

Primary use cases

Accelerating Spark and Presto queries over S3 data, unified vectorized execution for lakehouse queries, hardware-optimized data processing.

Recent developments

Latest signals

Axiom — composable query engines built on Velox (announced April 23, 2026). The Velox project announced Axiom, a framework for composing whole query engines on top of the Velox execution core — the step beyond "embed Velox into your existing engine" toward "use Velox + Axiom as the foundation of a new engine without rebuilding the vectorized execution primitives." Per the facebookincubator/velox repository and the Velox blog.
IBM watsonx.data ships Presto C++ v0.286 on Velox, claims price/performance edge over Databricks Photon. IBM says watsonx.data running the Velox-based Presto C++ engine (v0.286) plus its query optimizer, tested internally on Storage Fusion HCI, delivered better price/performance than Databricks' Photon engine. Per IBM announcements.
Microsoft Fabric's Native Execution Engine for Apache Spark is in public preview, built on Velox. Fabric's native acceleration layer names Velox as one of its two core open-source components (with Gluten), bringing Velox-accelerated Spark to Fabric Data Engineering / Data Science users. Per the Microsoft Fabric blog.
Independent benchmarks continue to favor Velox for native execution. An EDBT 2026 study finds Velox "outperforms or matches" other query engines across all workloads and cardinalities (Understanding the Performance of Native Execution in Big Data Analytics, EDBT 2026); a separate research system, CactusDB, built atop Velox reports up to 441x speedup on co-optimized SQL+ML inference (arXiv:2602.23469).
Project scale, per the official site: ~3.7k GitHub stars and 330+ contributors, with VeloxCon 2026 recordings now published. Per velox-lib.io.