Bolt is a C++ acceleration library providing composable, extensible and performant data processing toolkit. It is designed to provide generic and unified interfaces which can be pluggable into “any framework” running on “any hardware” to consume “any data source”.

Initially derived from Velox project, Bolt is created by ByteDance to embrace and unify the contributions from the community. It has been validated on Spark/Flink/Presto/ElasticSearch framework running on x64&ARM CPU/DPU/GPU accessing Parquet/ORC/Text/CSV/Lance file format managed under Hive/Paimon table to provide enterprise-grade cost optimization, results consistency and feature parity

Why Bolt?

“Open Source-First” Philosophy

“Contributions may come in many forms, and all of them are valuable”. The governance model of Bolt community will be in line with the Apache Way and “Community over Code” spirit. While we are working out the detailed governance model on a tree-tier structure of Contributor / Maintainers / Project Management Committee(PMC), we are committed to treating the open source repository as the source of truth, including but not limited to

• Public CI pipelines
• Clear dependency management as code
• Equal code review opportunity for manintainers
• Transparent design discussion

This will ensure the smooth & credible experience for code contribution

Embrace the Analytical Ecosystem

Bolt focuses on the physical execution layer of DBMS while providing first-class and high performance support for popular frameworks and storage formats.

Frameworks:

• Apache Gluten for Apache Spark
• RFC for PrestoDB
• OpenSearch for ElasticSearch
• Flink (Coming Soon)
• …

Storage Formats:

• Parquet
• ORC
• TXT
• CSV
• Paimon
• Lance (Coming Soon)
• …

Enterprise-Grade Performance, Result Consistency & Feature parity

Bolt is designed as a seamless acceleration layer that requires minimum code changes to the existing user jobs. Results/Performance comparison against original frameworks is performed on regular basis to capture regression & corner cases. Key features including

• Adaptive Task Parallelism
• Native Memory Management & Dynamic offheap threshold
• Operator Fusion
• JIT for hotspot expression
• Native Shuffle Support
• …

Getting Started

Get the Bolt Source

git clone https://github.com/bytedance/bolt.git
cd bolt

Set Up the Development Environment

Run the following script to set up your development environment and install dependencies:

scripts/setup-dev-env.sh

# configure your conan profile
# For more detailed usage, see the instructions in configure-conan-profile.py.
python ./scripts/configure-conan-profile.py --compiler clang  --libcxx=libc++
# or
python ./scripts/configure-conan-profile.py --compiler gcc  --libcxx=libstdc++11

Bolt manages its dependencies with Conan, an open-source, cross-platform package manager.

The setup script exports Conan recipes to the local cache. During the first build, dependencies are compiled from source and stored in the local cache. You can set up your own Conan server to speed up builds.

Use the following helper script to configure your Conan profile:

# Auto-detect the compiler, C++ standard library, and linker.
python3 scripts/configure-conan-profile.py

See configure-conan-profile.py for the default toolchain selection behavior and the options that override it.

Building Bolt

Building Bolt for Presto

Run make in the root directory to compile the sources. For development, use make debug to build a non-optimized debug version, or make release to build an optimized version. Use make unittest to build and run tests.

make release

# In main branch, by default, BUILD_VERSION is main.
make release BUILD_VERSION=main

Building Bolt as an Apache Gluten Backend

Bolt is now available as a backend in Apache Gluten. The integration was merged via apache/gluten#12454.

For the latest prerequisites and standard build instructions, see the Bolt Backend section in the Apache Gluten repository.

To validate local Bolt changes with Gluten, first build Bolt in Spark-compatible mode and export it to the local Conan cache:

# In the Bolt repository
make release_spark
make export_release

Then clone the official Apache Gluten repository and build the Bolt backend and Gluten JAR:

git clone https://github.com/apache/gluten.git
cd gluten

make release
make arrow

# Select the target matching your Spark version.
# For example, build the Spark 3.5 package:
make jar_spark35

Building Bolt for other system

You can use the make release && make export_release command to compile and export Bolt, and then use conan to reference Bolt. Below is a conanfile example that references Bolt as a third-party dependency.

# Take gluten for example:
class GluenConan(ConanFile):
  def requirements(self):
    bolt_version="main"
    self.requires(f"bolt/{bolt_version}", transitive_headers=True, transitive_libs=True)

Using Bolt from Python

The bolt-engine package provides two Python APIs:

• boltml: a DataFrame API for reading Parquet and ORC files, filtering, joining, and aggregating data, and applying Python functions to prepare data and ML features. It works with Arrow and NumPy and supports execution on Ray workers.
• pybolt: direct access to Bolt’s native types, vectors, and query plans for building and running execution pipelines from Python.

Start with boltml for data processing; use pybolt to construct query plans directly. See the Python quick start for installation and examples, and the build guide for wheel packaging and Ray deployment.

Contributing

Check our contributing guide to learn about how to contribute to the project.

Community

• Please join our Slack channel and ask in #dev.

License

Bolt is licensed under the Apache 2.0 License. A copy of the license can be found here.