Nimble (formerly known as “Alpha”) is a new columnar file format for large datasets created by Meta. Nimble is meant to be a replacement for file formats such as Apache Parquet and ORC.
Watch this talk to learn more about Nimble’s internals.
Nimble has the following design principles:
-
Wide: Nimble is better suited for workloads that are wide in nature, such as tables with thousands of columns (or streams) which are commonly found in feature engineering workloads and training tables for machine learning.
-
Extensible: Since the state-of-the-art in data encoding evolves faster than the file layout itself, Nimble decouples stream encoding from the underlying physical layout. Nimble allows encodings to be extended by library users and recursively applied (cascading).
-
Parallel: Nimble is meant to fully leverage highly parallel hardware by providing encodings which are SIMD and GPU friendly. Although this is not implemented yet, we intend to expose metadata to allow developers to better plan decoding trees and schedule kernels without requiring the data streams themselves.
-
Unified: More than a specification, Nimble is a product. We strongly discourage developers to (re-)implement Nimble’s spec to prevent environmental fragmentation issues observed with similar projects in the past. We encourage developers to leverage the single unified Nimble library, and create high-quality bindings to other languages as needed.
Nimble has the following features:
-
Lighter metadata organization to efficiently support thousands to tens of thousands of columns and streams.
-
Use Flatbuffers instead of thrift/protobuf to more efficiently access large metadata sections.
-
Use block encoding instead of stream encoding to provide predictable memory usage while decoding/reading.
-
Supports many encodings out-of-the-box, and additional encodings can be added as needed.
-
Supports cascading (recursive/composite) encoding of streams.
-
Supports pluggable encoding selection policies.
-
Provide extensibility APIs where encodings and other aspects of the file can be extended.
-
Clear separation between logical and physical encoded types.
-
And more.
Nimble is a work in progress, and many of these features above are still under design and/or active development. As such, Nimble does not provide stability or versioning guarantees (yet). They will be eventually provided with a future stable release. Use it at your own risk.
Nimble’s CMake build system is self-sufficient and able to either locate its main dependencies or compile them locally. In order to compile it, one can simply:
$ git clone git@github.com:facebookexternal/nimble.git
$ cd nimble
$ make
To override the default behavior and force the build system to, for example, build a dependency locally (bundle it), one can:
$ folly_SOURCE=BUNDLED make
Nimble builds have been tested using clang 15 and 16. It should automatically compile the following dependencies: gtest, glog, folly, abseil, and velox. You may need to first install the following system dependencies for these to compile (example from Ubuntu 22.04):
$ sudo apt install -y \
flatbuffers-compiler \
libflatbuffers-dev \
libgflags-dev \
libunwind-dev \
libgoogle-glog-dev \
libdouble-conversion-dev \
libevent-dev \
liblzo2-dev \
libelf-dev \
libdwarf-dev \
libsnappy-dev \
bison \
flex \
libfl-dev
Although Nimble’s codebase is today closely coupled with velox, we intend to decouple them in the future.
Nimble is licensed under the Apache 2.0 License. A copy of the license can be found here.