This is an assessment of the MXNet podling’s maturity, meant to assess the status towards graduation as a top-level Apache project.

It is based on the ASF project maturity model at https://community.apache.org/apache-way/apache-project-maturity-model.html

Status of this document


Last Updated On: 4/5/2022

Maturity model assessment

Mentors and community members are encouraged to contribute to this page and comment on it, the following table summarizes project’s self-assessment against the Apache Maturity Model.




The project produces Open Source software, for distribution to the public at no charge.[1]

YES. The project source code is licensed under the Apache License, version 2.0.


The project's code is easily discoverable and publicly accessible.

YES. Linked from the website, available via GitHub.


The code can be built in a reproducible way using widely available standard tools.

YES. The build uses cmake and instructions are available on the project website at how to build from source. Continuous integration results are available on Jenkins.


The full history of the project's code is available via a source code control system, in a way that allows any released version to be recreated.

YES. Using Git, main repository at https://github.com/apache/incubator-mxnet. Releases are cut from that repository. All releases are tagged.


The provenance of each line of code is established via the source code control system, in a reliable way based on strong authentication of the committer. When third-party contributions are committed, commit messages provide reliable information about the code provenance.[2]

YES.The project uses a git repository, managed by Apache Infra, ensuring provenance of each line of code to a committer. Third party contributions are accepted in accordance with the Contribution Guide only.

Licenses and Copyright


The code is released under the Apache License, version 2.0.

YES. Source distributions clearly state license.


Libraries that are mandatory dependencies of the project's code do not create more restrictions than the Apache License does.[3], [4]

YES. The list of mandatory dependencies have been reviewed to contain approved licenses only. See license.


The libraries mentioned in LC20 are available as Open Source software.

YES. All mandatory dependencies are available as open source software. See 3rdparty dependencies.


Committers are bound by an Individual Contributor Agreement (the "Apache iCLA") that defines which code they are allowed to commit and how they need to identify code that is not their own.

YES. The project uses a repository managed by Apache Infra -- write access requires an Apache account, which requires an ICLA on file.


The copyright ownership of everything that the project produces is clearly defined and documented.[5]

YES. All files in the source repository have appropriate headers (see below). Even further, Software Grant Agreements for the initial donations and Corporate CLAs have been filed.



Releases consist of source code, distributed using standard and open archive formats that are expected to stay readable in the long term.[6]

YESSource releases are distributed via dist.apache.org and on Github.


Releases are approved by the project's PMC (see CS10), in order to make them an act of the Foundation.

YES. All incubating releases have been unanimously approved by the MXNet community and the Incubator, all with at least 3 (P)PMC votes.


Releases are signed and/or distributed along with digests that can be reliably used to validate the downloaded archives.

YES. All releases are signed, and the KEYS file is provided on dist.apache.org.


Convenience binaries can be distributed alongside source code but they are not Apache Releases -- they are just a convenience provided with no guarantee.

YES. Convenience binaries are distributed via pip (Python). R packages are published on and CRAN. Scala packages are published to Apache release repositories. They are not distributed via dist.apache.org.


The release process is documented and repeatable to the extent that someone new to the project is able to independently generate the complete set of artifacts required for a release.

YES. Release process is available describing the entire process. Each MXNet release has been performed by different release manager.



The project is open and honest about the quality of its code. Various levels of quality and maturity for various modules are natural and acceptable as long as they are clearly communicated.

YES. The project records all bugs and feature requests on Github. Github is used to manage projects. In addition Github projects and milestones are used to manage bug fixes, tasks and projects


The project puts a very high priority on producing secure software.[7]

YES. Security issues are treated with the highest priority, according to the CVE/Security Advisory procedure.


The project provides a well-documented channel to report security issues, along with a documented way of responding to them.

YES. Website provides a security page that explains the security vulnerability reporting process.


The project puts a high priority on backwards compatibility and aims to document any incompatible changes and provide tools and documentation to help users transition to new features.

YES. Each release contains a “What’s new” on the website containing the release notes (extracted from github and cwiki). The project follows semantic versioning. The project takes efforts to ensure backward-compatibility for not only the software, but also the artifacts such as pre-trained models.


The project strives to respond to documented bug reports in a timely manner.

YES. Bug reports in this project are submitted via GitHub issues. The project has resolved 477 bug reports during incubation.



The project has a well-known homepage that points to all the information required to operate according to this maturity model.

YES. The project website has a description of the project with technical details, how to contribute, links to discussion forum and all project related information.


The community welcomes contributions from anyone who acts in good faith and in a respectful manner and adds value to the project.

YES. It’s part of the contribution guide and the current committers are really keen to welcome contributions.


Contributions include not only source code, but also documentation, constructive bug reports, constructive discussions, marketing and generally anything that adds value to the project.

YES. The project accepts a wide range of contributions and consider them merits. Documentation improvement is just as valued as code contribution. The project considers a wide variety of contributions merits towards committer/PPMC membership.


The community is meritocratic and over time aims to give more rights and responsibilities to contributors who add value to the project.

YES. The project committer group grew to 71 people, with 46 PPMC members. New committers and PPMC members are elected based on meritocracy.


The way in which contributors can be granted more rights such as commit access or decision power is clearly documented and is the same for all contributors.

YES. The criteria is documented in the Becoming a committer and PPMC member guide.


The community operates based on consensus of its members (see CS10) who have decision power. Dictators, benevolent or not, are not welcome in Apache projects.

YES. The project works to build consensus. Consensus is built first through discussion. Lazy consensus is employed when applicable.


The project strives to answer user questions in a timely manner.

YES. The project typically provides detailed answers to user questions in a timely manner via dev@ mailing list, user@ mailing list, https://discuss.mxnet.io/, https://discuss.gluon.ai/ and on StackOverflow.

Consensus Building


The project maintains a public list of its contributors who have decision power -- the project's PMC (Project Management Committee) consists of those contributors.

YES. The incubator website contains the list of committers and PPMC members.


Decisions are made by consensus among PMC members [9] and are documented on the project's main communications channel. Community opinions are taken into account but the PMC has the final word if needed.

YES. The project has been making important decisions on the project dev@ mailing lists. 


Documented voting rules are used to build consensus when discussion is not sufficient. [10]

YES. The project uses the standard ASF voting rules. Voting rules are clearly stated before the voting starts for each individual vote.


In Apache projects, vetoes are only valid for code commits and are justified by a technical explanation, as per the Apache voting rules defined in CS30.

YES. The project follows the Apache voting rules. Only vetoes with technical reasons are respected.


All "important" discussions happen asynchronously in written form on the project's main communications channel. Offline, face-to-face or private discussions [11] that affect the project are also documented on that channel.

YES. The project has been making important decisions on the project mailing lists. Minor decisions may occasionally happen during code reviews, which are also asynchronous and in written form, and archived on the commits@ mailing list. Offline discussions are requested to be summarized and brought back to dev@ for further discussion. The project employs GitHub issue RFC mirroring to the dev@ list to bridge the discussions happening on GitHub and dev@.



The project is independent from any corporate or organizational influence. [12]

YES. The project team gathers people from different companies (Amazon, Intel, Nvidia, IBM Canada, Tencent, Cognitect, Cern, Harman, Uber, Airbnb, Mathworks, Cloudera, Samsung, Wolfram Research, Philips, Horizon Robotics, Apple, BIGO Live, NTT, WhiteHat Security, Ant Financial, Alibaba, Microsoft, Facebook, Qualcomm, Dato, Google, Volkwagen, TuSimple, HP, Baidu, Salesforce, JuliaLabs, Lyft, TSMC, Yahoo, Hortonworks, Optum Labs, SigOpt, Curalate, Workday) and universities (Washington, Berlin, ShanDong, Hongkong, Tsinghua, San Francisco, Moscow, Xina Jiaotong, USTC, Yale, Peking, Stanford, Indiana, Fudan, New York, Edinburgh, Renmin, Alicante, Carnegie Mellon, Texas at Arlingon, Alberta, Princeton, Purdue, Nanjing, Zhejiang, Australian National, Los Angelas, Singapore, Johns Hopkins, Cambridge, San Diego, Sichuan, Granada, Rice, Nankai, Chicago, Chiao Tung, Houston, Harvard, Davis, Hannover, Tongji, Toronto, Georgia Tech). No company or organization has significantly more influence than any other. We can note a growth of the contributions coming from different committers.

As of today the PPMC has 51 members (incl. mentors). The company with most PPMC members is Amazon, with 12 (23.5%) members affiliated with it. 76.5% are affiliated with different companies and organizations.


Contributors act as themselves as opposed to representatives of a corporation or organization.

YES. The committers and contributors act on their own initiative without representing a corporation or organization.

Dependency and License analysis 

This section analyses project’s direct and transitive dependencies to ensure compliance with Apache Software Foundation’s policies and guidelines.

The following is a list of licenses for all direct and transitive dependencies. The list is annotated where applicable.

The section labelled in Green color is the required dependencies in MXNet. The rest are optional dependencies.

Unless specified, Licenses that contains the name of the dependencies are similar to the standard MIT license.

DependencyOrgLicenseShort Description
Main repository
MXNetApacheApache-2.0The MXNet repo, stats does not include submodules dependencies
Dependencies maintained as git submodule
tvmDMLCApache-2.0Compiler for neural networks, includes NNVM which is used by MXNet
dmlc-coreDMLCApache-2.0 (partial)Various components used across DMLC projects for logging, data, etc.
ps-liteDMLCApache-2.0Parameter server for distributed training
dlpackDMLCApache-2.0Tensor structure definition
cubDMLCBSD 3-ClauseSnapshot of Nvidia’s CUB project. CUB is a library with various utilities for CUDA kernel programming.
openmp-llvmLLVMDualOpenMP C/C++ language extensions in Clang/LLVM compiler.
googletestGoogleBSD 3-ClauseC++ unit test framework
mkl-dnnIntelApache-2.0Intel-optimized Math Kernel Library (MKL) for Deep Neural Networks (DNN)..
Dependencies used for package build but not statically-linked
ATLASATLASBSDOptimized linear algebra library
CUDANvidiaProprietaryGPU programming library
cuDNNNvidiaProprietaryGPU programming library for neural networks
Dependencies used for package build and statically-linked
OpenCVopencvBSD 3-ClauseComputer vision library
OpenBLASxianyiBSD 3-ClauseOptimized linear algebra library

Jean-loup Gailly,
Mark Adler

Zlib license

A Massively Spiffy Yet Delicately Unobtrusive Compression Library

libjpeg-turboVaries developerBSD 2-Clause
IJG 2-Clause

libjpeg-turbo is a JPEG image codec 


The PNG Reference Library Authors

PNG Reference Library License version 2

libpng is the official PNG reference library. 

Silicon Graphics, Inc,

Sam Leffler

libtiff licenseIncluded in this software distribution is a library, libtiff, for reading and writing TIFF, a small collection of tools for doing simple manipulations of TIFF images, and documentation on the library and tools.

The OpenSSL Project

Apache 2.0

The OpenSSL Project is a collaborative effort to develop a robust, commercial-grade, fully featured, and Open Source toolkit implementing the Transport Layer Security (TLS) protocols (including SSLv3) as well as a full-strength general purpose cryptographic library.

libcurlDaniel Stenberglibcurl licenseCurl is a command-line tool for transferring data specified with URL syntax.
eigenvaries ContributorsMPL 2.0

Eigen is a C++ template library for linear algebra: matrices, vectors, numerical solvers, and related algorithms.

protobufGoogle inc.Google LicenseProtocol Buffers (a.k.a., protobuf) are Google's language-neutral, platform-neutral, extensible mechanism for serializing structured data. 
cityhashGoogle incGoogle License

CityHash provides hash functions for strings.

zmqZeroMQ Community

LGPL with Exception:

As a special exception, the Contributors give you permission to link this library with independent modules to produce an executable, regardless of the license terms of these independent modules, and to copy and distribute the resulting executable under terms of your choice, provided that you also meet, for each linked independent module, the terms and conditions of the license of that module. 

The ZeroMQ lightweight messaging kernel is a library which extends the standard socket interfaces with features traditionally provided by specialised messaging middleware products.
lz4Takayuki MatsuokaBSD 2-ClauseLZ4 is lossless compression algorithm, providing compression speed > 500 MB/s per core, scalable with multi-cores CPU. It features an extremely fast decoder, with speed in multiple GB/s per core, typically reaching RAM speed limits on multi-core systems.

  • No labels