What is Velox?
Velox is a high-performance, open-source execution engine designed for flexibility and reuse. Distributed as a C++ library, it offers modular, composable components that power a wide range of data processing systems from batch and interactive queries to streaming and AI/ML workloads.
The Velox Ecosystem
Velox powers a growing ecosystem of open-source projects, from Spark acceleration with Apache Gluten to next-gen query engines like Presto C++. It’s also the foundation for ongoing work in hardware-accelerated execution.
Apache Gluten
An accelerator for Apache Spark that integrates Velox for native execution.
Learn More →Key Features
Efficiency
Democratizes runtime optimizations previously only implemented in individual engines.
Consistency
By leveraging the same execution library, compute engines can expose the exact same functions, data types, and semantics to their users.
Reusability
Features and runtime optimizations available in Velox are developed and maintained once, reducing engineering duplication and promoting reusability.
The Velox Community
Latest Community News
4.2k GitHub Stars
488 Contributors
13.5k+ CommitsVeloxCon China 2026
Join us December 5 at Alibaba Xuhui Riverside Park, Shanghai. CFP is now open.
December 2026EventVeloxCon 2026 Recordings
Sessions from VeloxCon 2026 at Meta HQ are now available on-demand on our YouTube channel.
August 2026BlogNative Delta Statistics with Velox Task Barriers
Gluten now evaluates Delta Lake per-file statistics natively using a Velox Task Barrier instead of processing rows one at a time.
August 2026BlogBuild Once, Probe Many: Hash Table Caching in Velox
HashTableCache replaces build-per-task with a build-once, reuse-many protocol for broadcast joins.
August 2026BlogWar of the Allocators
Comparing jemalloc against Velox's custom mmap-based allocator on performance, fragmentation, and RSS control.
July 2026BlogMaking OpenZL Available in Nimble OSS
Nimble's long-standing OpenZL compression integration is now open source for external contributors.
July 2026BlogWhy RIGHT SEMI JOIN Can Be Slower Than LEFT SEMI JOIN in Velox
Why a seemingly equivalent join rewrite caused a 10x regression, and the fix that closed the gap.
June 2026BlogGPU-Accelerated Presto C++ is Here
IBM and NVIDIA bring cuDF-backed GPU operators to Velox, with nightly GPU-accelerated Presto C++ images now published.
June 2026PaperAccelerating Presto with GPUs
IBM Research paper (VLDB 2026) on GPU-aware Velox operators and a GPU-native Presto exchange, up to 6x cheaper than CPU-only Presto.
June 2026BlogFrom copyBits to SIMD: Accelerating Parquet DELTA Decoding in Velox
Targeted bit-copy and SIMD optimizations cut Parquet DELTA decoding CPU time by 4.5x.
June 2026BlogFlatMapVector Adoption for Scaling High-Performance AI/ML Data Pre-Processing
How FlatMapVector improves handling of the map-column features common in AI/ML feature stores.
May 2026BlogNimble Cluster Index: Efficient Indexed Lookups on Columnar Data
A hierarchical index embedded in Nimble files enables O(log n) point lookups without sacrificing columnar scan performance.
April 2026BlogAxiom: Composable Query Engines Built on Velox
Introducing Axiom, a C++ library for building fully composable, high-performance query engines on top of Velox.
April 2026BlogWhy NULLIF Must Be a Special Form
Why NULLIF can't be implemented as a regular function and must be handled as a SQL special form.
April 2026BlogFrom flaky Axiom CI to a Velox bug fix: a cross-repo debugging story
Tracing an intermittent macOS-only CI failure in Axiom back to a bug in its Velox submodule dependency.
March 2026BlogAdaptive Per-Function CPU Time Tracking
A per-function CPU usage breakdown for Velox expression trees, useful for pinpointing costly functions in long-running jobs.
March 2026TalkGTC 2026: GPU-Accelerated Presto and Velox
NVIDIA and IBM Research walk through recent performance improvements for Velox's experimental cuDF-backed GPU execution mode.
March 2026BlogAccelerating Unicode string processing with SIMD in Velox
SIMD-based Unicode length helpers deliver 2-15x speedups across most string length configurations.
March 2026BlogThe hidden traps of regex in LIKE and split
Both LIKE and Spark's split can silently produce wrong results when used with column values instead of constants.
March 2026Blogvelox::StringView API Changes and Best Practices
This post describes in more detail how Velox handles columns of strings.
February 2026BlogTask Barrier: Efficient Task Reuse and Streaming Checkpoints in Velox
How Velox Task Barriers provide a synchronization mechanism.
January 2026BlogWhy Sort is row-based in Velox — A Quantitative Assessment
Rationale behind this design decision and provide experimental evidence for its implementation.
December 2025EventVeloxCon China 2025
Join us in Beijing on December 13 for our first community event in China.
December 2025BlogMulti-Round Lazy Start Merge
The latest Velox update introduces a multi-round spill merge with lazy-start activation.
November 2025BlogEnabling Shared Library Builds in Velox
Learn how we unblocked shared library builds in Velox.
November 2025BlogVelox switches to C++20 standard
Going forward Velox is looking to enhance the codebase by making use of newer compiler functionalities.
August 2025BlogExtending Velox – GPU Acceleration with cuDF
This post describes the design principles and software components for extending Velox with hardware acceleration using NVIDIA's cuDF.
July 2025BlogSEGFAULT due to Dependency Update
How upgrading Folly and Facebook Thrift revealed a surprising issue with coroutine compile flags and vtable structures.
July 2025BlogVelox Primer: Part 3
Diving deeper into Velox internals and optimization strategies.
May 2025EventVeloxCon 2025
Sessions from VeloxCon are now available on-demand on our YouTube channel.
April 2025BlogVelox Primer: Part 2
Exploring modular execution and system reuse with Velox.
March 2025BlogVelox Primer: Part 1
Introduction to Velox and why composable engines matter.
February 2025BlogVelox Query Tracing
More about Velox query tracing including how it works, its framework, and future plans.
December 2024